With the rapid increasing of smartphones and versatile new applications, the mobile data grows exponentially in recent years. Upon the establishment of 5G wireless networks, massive channel data are collected in various scenarios under high frequency band. With the development of computer science, the machine learning based methods have been introduced into the field of channel modeling. Unlike the conventional stochastic model which is represented in the equation form, or the deterministic model which requires intensive computation like raytracing, the machine learning based model intend to fit the channel data for discovering the regularities with a lower complexity. Up to now, researchers have achieved superior results on channel modeling in indoor and urban scenarios, with the help of techniques such as ANN and CNN. However, the performances are highly depended on the collected data. The Physics-Informed Neural Networks (PINN) consider both data and physical laws when designing the network architecture and loss functions, which show its superiority in areas like computational fluid dynamics, epidemic model and traffic flow estimation. In this talk, we discuss how to combine the propagation law of electromagnetic wave with the machine learning technique, to form the ‘knowledge plus data driven’ paradigm in channel modeling. This is an ongoing work, and suggestions are welcome.

Most of the epidemic models only focus on one aspect of disease spreading, either be macroscopic, or microscopic. Macroscopic epidemic model, such as SEIR model, leverage ordinary differential equations (ODEs) to depict aggregated transmission patterns of a large group, while some newly proposed microscopic models use GCN to capture individual level transmission probabilities. However, there is a missing link between macroscopic and microscopic models, where the advantages of these models are segregated. In this work, we hope to propose a multiscale epidemic model that combines the advantages of both macroscopic and microscopic models, where the GCN model captures the individual level transmission pattern under the macroscopic constraint provided by expert knowledge based ODEs. We will introduce several important works on microscopic transmission models, neural network supported ODEs, and recent effort on multiscale simulation. Based on the above discussion, we propose a prototype framework to accurately predict both microscopic and macroscopic epidemic spreading, breaking the barrier between different scales.

In recent years, as new algorithms, models and applications in the field of artificial intelligence (AI) are widely proposed, AI has implications for countless aspects of the society, such as the future of work, the stock market, medicine, transportation, the future of warfare and the governance of society. On the other hand, research shows that the development of AI has the potential to significantly disrupt labor markets of traditional workforce, leading to increasing unemployment. However, how AI impact the career of scientific workforce, namely researchers, remains unexplored. In this work, we use the Microsoft Academic Graph (MAG) dataset and analyze the changes happening along with the development of AI on paper publication and citation patterns. We try to uncover the impact of AI on researchers and provide enlightenment on the future research works.

Resilience has long been recognized as a defining property of many dynamical complex systems. It refers to a system’s ability to adjust its activity to retain its basic functionality when errors, failures and environmental changes occur. Although hard to characterize, previous studies have systematically separated the roles of the system’s dynamics and topology in deciding its resilience[1]. Theoretically, the patterns of resilience depend only on the systems’ intrinsic dynamics, but the specific level of the current system’s resilience is affected by the network topology, including density, heterogeneity and symmetry. Therefore, the above theory motivates us to explore an interesting problem of generating network topology for achieving resilience in complex systems. Notably, the ability to generate resilient networks is vital in 1) critical infrastructure planning scenarios like power grids, road networks, etc. and 2) network attack&defence scenarios requiring highly resilient training instances. However, it is notoriously difficult for traditional statistical physics approaches to solve the above generation problem. Instead, existing works tend to enhance the system’s resilience by slight modifications of network topologies[2,3]. With the capability of coping with nonlinearity and multi-dimensionality in complex systems, machine learning based approaches could be a promising solution. However, it still needs to overcome the following two technical challenges 1) generating graphs with the network dynamics underlying nodes’ self-evolution and pairwise interactions, and 2) conditional generation with both resilience and original utility [4,5]. This talk will mainly focus on the definition, background and possible solutions to the above problem.

Origin-destination (OD) flow data, which indicates population mobility in the city, is of great value in many urban applications, such as urban planning, public resource allocation and traffic simulation. Since the OD data is not always available for particular region of interest, generating OD data based on easily obtained information is a necessary task. There are works focus on generating static OD patterns, while ignoring the temporal distribution over the course of a day. However, the distribution of OD over time is very critical information, for example, a large amount of OD between two regions that are evenly distributed throughout the day will not put a lot of tension on transportation system. Therefore, we propose an OD generation framework that combines forward prediction and reverse inference to generate a comprehensive temporal OD flow in a day, taking into consideration the spatio-temporal characteristics of OD. Specifically, we use an OD predictor to generate temporal OD flows based on static city attributes, and then use the generated OD flows to predict traffic status on the road network, to enable the framework to optimize the forward prediction process and the reverse inference process using gradients.

Realistic simulation of a massive amount of human mobility data is of great use in many fields. However, existing solutions for mobility simulation are limited in either modeling variable-length trajectories with continuous temporal distribution or incorporating multi-dimensional context information. Our model consists of two layers of Variational Auto-Encoder (VAE). The first layer is designed as a VAE module for modeling user commuting attributes, establishing a unified mapping of user commuting attributes to Gaussian distribution, realizing support for sampling directly from Gaussian distribution and generating a large number of user trajectories. The second layer is the VAE module for modeling the overall trajectory level, fully modeling the relationship between visiting location attributes and stay time, further modeling the stay time and travel time modules respectively, together forming the time module to more accurately model the trajectory time distribution.

Large ride-hailing platforms, such as DiDichuxing, connect thousands of vehicles to millions of ride-demands in a day, improving transportation efficiency through the tasks of order dispatching and vehicle repositioning. To provide efficient strategy, there are two major challenges: a) due to the inter-influence between order dispatching task and vehicle repositioning task, a joint method is required to tackle both tasks jointly. b) The cooperation between different vehicles, which is highly related to strategy performance, is difficult to model and tackle due to the large number of drivers. To solve the challenges above, we design a cooperative multi-agent reinforcement learning-based hierarchical framework. On one hand, in order to make joint action, we model the action as the preference of driver features for order features or repositioning features. On the other hand, in order to achieve efficient cooperation between drivers, we design a community-based cooperative method. First, we use community-detection method to divide the city into several communities. Second, we use cooperative reinforcement learning method in each community to achieve direct cooperation for drivers in the same community. Finally, we design a hierarchical framework to coordinate the driver repositioning between different communities. This is still an ongoing work. Your suggestions and questions are important for our work.

Urban planning designs and regulates the uses of space in a city, which aims at improving the sustainability and competitiveness of the city. From the view of algorithm design, urban planning can be regarded as a sequential decision making process, where the planner successively arrange an object to a location in the city. As reinforcement learning (RL) has shown great power in other sequential decision making problems, such as Go and chip design, we attempt to tackle the problem of urban planning with the help of RL. Specifically, we formulate the problem as an MDP, where at each step, the agent (planner) decides a location for an object based on the current plan, and the environment (simulator) updates the urban plan given the new arrangement. We use both CNN and GNN as the policy’s main network architecture, to capture the relationship between different land uses. We use a widely-adopted policy gradient framework PPO to optimize our policy, which conducts safe exploration without catastrophic failure. To evaluate the performance of urban planning, we focus on static metrics including traffic efficiency, public service accessibility and ecological landscaping. We use the proposed RL method to perform urban planning on the community scale, where experiments show that the proposed approach can generate reasonable plans of different land uses. The performance with respect to traffic, public service and greenness can be effectively improved with the proposed method. As this is still an ongoing project, I will also introduce the roadmap and future plans of this work.

Urban vibrancy is largely characterized by frequent human mobility to various types of urban venues, which is however prone to external shocks such as the COVID-19 pandemic, making it critical to understand the heterogeneous resilience associated with different venues. In our previous work, we analyze a large-scale mobility dataset that records visitations to 2.08 million urban venues in the U.S. from March 2019 to February 2022. From the data, we categorize heterogeneous urban venues into 5 types by jointly considering the mobility drop in the lockdown phase and mobility recovery in the reopening phase, and reveal that life-related venues endure smaller mobility reduction and better recovery than work-related venues. Based on these observations, in our current work, we conceptualize a Well-and-Ditch model to describe the relationship between mobility level and proportion of visits to different types of urban venues. We examine the model by analyzing the relationship between the centrality of venues in the venue sector network and the mobility change associated with these venues in COVID-19.

With the increase of computing power and the development of data science, modeling and simulation are becoming indispensable tools in urban science research. Cities, as complex systems made up of many aspects such as mobility, infrastructure, have complex interactions and relationships among multiple elements. In order to provide researchers with tools to model and simulate complex urban systems, we first propose a city model that focuses on three key concepts: human, thing and space. According to the city model, we design and develop Mirage, an efficient and extensible city simulation framework and also implement an efficient mobility module as Mirage's necessary module. To show the extensibility of Mirage, we build an application case about urban vulnerability to present the cascading impact among urban infrastructures and other application cases about decision making to test Mirage's ability to support decision making algorithms like reinforcement learning. We also conduct extensive experiments to verify the efficiency of Mirage and its mobility module.

With the rapid development of e-commerce, couriers are required to offer efficient and reliable logistics services. A courier is asked to record the delivery time of the parcel after delivering it, and each delivery task will be accounted timely if the courier finish the delivery before required finish time. Therefore, delivery timely rate is computed, which is vital for many applications in JD.com, e.g., customers’ experience evaluations, efficiency of package dispatching, delivery performance evaluations. Especially, it can help schedulers make decisions on sending more men to delivery stations with low delivery time rate predictions. In this work, we first investigate affctors that will affect timely rate, and then reveal the spatial-temporal correlations between blocks. We further develop a transformers architecture model, which encode the features of each block. Finally, we conduct some experiments on real world data to validate the effectiveness of the proposed models.

In past decades, the knowledge graph (KG) has been widely investigated, which represents real-world knowledge into entity-relation-entity triple facts. Moreover, recent studies propose the hyper-relational knowledge graph (HKG) to represent more complex information, by associating triples with additional relation-entity pairs (a.k.a., qualifiers). Due to the incompleteness of HKG, the basic HKG completion task is to complete the missing links between entities therein. Especially, the best performing baselines adopt the famous Transformer architecture to model the interactions among entities and relations for HKG completion, which however largely ignore the structure information. Therefore, in this work, we investigate structure-aware Transformers for HKG Completion. Specifically, we first transform the HKG into homogenous graph which turns relations to edge-nodes, and then extract pairwise shortest paths between entities for structure information. We further develop two categories of structure-aware Transformers on HKG, which encode the path structures as input and attention bias of Transformer, respectively. Finally, we conduct some experiments to validate the effectiveness of the proposed models.

Living in cities affords expanded access to various resources, infrastructures, and services at reduced travel cost. However, recent research shows that the accessibility to POIs is unevenly distributed across the urban space, especially among citizens of different income levels. This inequality is rarely been viewed from the perspective of transportation network, particularly in car-oriented US cities. Here, we first examine the visiting patterns of citizens to various categories of urban POIs by the SafeGraph mobility data. Significant socioeconomic gaps in the accessibility to specific POI categories are observed. For instance, people living in high income and vehicle ownership rate communities have higher chances of visiting nature parks, museums, and fitness centers than those with lower income and vehicle ownership rates, especially when the distance to POIs is less than 10 kilometers. This income gap in short-distance travel implies the potential inequality in urban transportation networks, where the disadvantage of public transit networks may constrain the poor people visiting desired places. Then we use Open Trip Planner to query millions of trips in the five largest US cities to obtain the travel time and distance by public transit and by car. We observe the scaling relation between travel time and distance and a universal law that the time cost of public transit compared to driving first decreases and then increases with distance, partially explaining the POI accessibility gap at a short distance. We further analyze the proportion of motorways in queried driving routes and the walking time in public transit routes to explain the universal inequality.

Traffic congestion not only greatly influences people's travel experience, but also brings huge economic losses and serious air pollution. It has become one of most serious problems in the transportation systems. The spread of traffic congestion in urban networks has been viewed as a complex spatio-temporal phenomenon. Understanding the traffic congestion propagation mechanism can benefit the design of congestion control strategies. Although most existing approaches have studied to model traffic flow at a microscopic level, its microscopic dynamic that reflects the realistic traffic congestion propagation process is still less explored. To address it, by virtue of the reverse modeling theory of complex network, we aim to construct the microscopic dynamics from real-world traffic data. This talk will introduce some representative works about this problem, and discuss how to construct the microscopic dynamics for traffic congestion.

In the era of information explosion, recommender systems play an indispensable role in what we see, what we purchase, and even what we like. Despite the convenience, recommender systems could potentially isolate people from a diversity of viewpoints or content, forming information cocoons. To delineate the mechanisms behind the phenomenon, we provide a complex adaptive-systems perspective on the effects of recommender systems and reveal the emergence of information cocoons. We observe that the interplay between similarity-based recommendations and users’ exploration determines the formation of information cocoons and the population will converge into three different states of information cocoons with different degrees of similarity-based recommendations and exploration. We find that users’ susceptibility to recommendations has moderating effects on the emergence of information cocoons. We also point out that the utilization of negative samples can play a complementary role to the effects of exploration. Overall, our study sheds light on the impacts of similarity-based personalized recommendation algorithms and unveils avenues to steer the emergence of information cocoons.

Dynamic vehicle routing problem (DVRP) and its variants arise naturally in real-world applications like logistics. Here "dynamic" refers to the fact that not all customer demands are known beforehand but are dynamically generated as we are planning the routes. Indeed, there are a lot of previous works on solving these problems. Existing methods mainly tackle DVRP by periodically re-solving when new demands come in or incorporating predicted future demands using an additional predictive module. However, due to the NP-hard nature of combinatorial problems, it is hard or very time-consuming to reach a fairly good solution. And the dynamic generation of demands further adds to the complexity of the problem. Inspired by the recent trend of "learning to solve", we seek to devise learning-based approach that can achieve a better performance than methods with hand-crafted and domain-knowledge-centric rules.

Recently, mobile networks have gained global popularity while mobile user traffic facilitates diverse applications. However, large-scale mobile user traffic is hardly available, and privacy concerns impede high-quality data sharing and utilization. Existing generation models fail to simulate the multi-scale temporal dynamics in mobile user traffic on individual and aggregate levels. In this work, we propose a multi-scale hierarchical generative adversarial network (MSH-GAN) containing multiple generators and a multi-class discriminator. Specifically, our observations show that the mobility traffic usage behavior exhibits the characteristic of being a mixture of multiple behavior patterns, which are referred to as micro-scale behavior patterns and modeled by different pattern generators in our model. Further, we discover that the traffic usage behavior of different users exhibits strong clustering characteristics with the co-existence of users with similar and different traffic usage behavior, respectively. Thus, we model each cluster of users as a class in the output of the discriminator, which is referred to as macro-scale user clusters. Then, the gap between micro-scale behavior patterns and macro-scale user clusters is bridged by introducing the switch mode generators, which describe the traffic usage behavior in terms of switching between different patterns and cooperate with pattern generators to form users' traffic usage behavior. The pattern generators are shared by all users, while the switch mode generators are only shared by a specific cluster of users, which models the multi-scale hierarchical structure of the traffic usage behavior of massive users. Finally, we urge MSH-GAN to learn the multi-scale temporal dynamics via a combined loss function, including adversarial loss, clustering loss, aggregated loss, and regularity terms. Extensive experiment results demonstrate that MSH-GAN outperforms state-of-art baselines by at least 118.17% in terms of critical data fidelity and usability metrics. Moreover, observations show that MSH-GAN is capable of simulating the traffic patterns and pattern switch behaviors.

Edge-side caching-assisted device-to-device (D2D) communication has been recognized as a viable solution for reducing network latency and alleviating the high burden of backhaul transmission links. Caching mechanisms at the network edge are strongly dependent on the distribution of individual users' content preferences. Some proactive measures should be addressed in order to fully realize the benefits of edge caching. Reshaping content request probabilities of different users has a substantial impact on the cache decision. This presentation is about how recommendation can be used to improve the caching effectiveness of D2D enabled wireless content caching networks. The objective is to maximize cache hit ratio for a generic network while taking into account the personalized recommendation quality, quantity, and cache capacity requirements of each user. A time-efficient sub-optimal method will be discussed, which operates iteratively and has a verifiable convergence guarantee and polynomial time complexity.

Women have been recognized as disadvantaged in science: past literature has demonstrated the prevalence of gender inequality in science, where women have fewer publications, fewer collaborators, less funding, and are identified as penalized in hiring. However, other research has underlined the privileges of diversity in science, where ethnic diversity is reported to be associated with higher impact and underrepresented groups are identified as higher-rate producers of scientific novelty. We are wondering how, despite this prevalent gender inequality in science, gender mixture can benefit science. Leveraging publication records of more than 240 million papers from Microsoft Academic Graph, we show that gender-mixed teams produce more interdisciplinary science. Not only do publications of gender-mixed teams cite more venue communities, refer to more distant knowledge combinations, and unite more distant authors compared with same-gender teams, but higher Gini impurity in terms of gender is associated with higher interdisciplinarity. Our results further highlight the importance of promoting gender diversity and equality in science.

As an important part of the exploration of smart city construction, research on digital twin cities has attracted attention from all walks of life. As an important step of digital twin city construction, the ability of simulate complex urban system has also become an academic hot research topic. To this research hotspot, my project team has cooperated to complete a multi-level urban simulation system combining power grids, water supply and drainage networks, communication networks, and traffic networks, providing data and platform support for studies on urban vulnerability and intelligent decision-making. Under the large-scale urban simulation scenario, the microscopic traffic flow simulation part in the simulation module will become the performance bottleneck of the overall simulation system due to the number of agents and the complexity of the calculation. And the architecture of GPU is specially designed for such large-scale computing tasks with high consistency of computing logic. Based on the GPU architecture and the CUDA computing platform provided by NVIDIA, and the existing CPU-based microscopic traffic flow simulation system in the laboratory, aiming at its shortcomings in efficiency, combining the architecture of GPU and the optimization strategy provided by CUDA, this paper refactors the design of the original system. The efficiency of data reading and writing is improved by the effective use of GPU on-chip storage, the more efficient computing task allocation is achieved through the task flow mechanism of CUDA, the efficient interaction interface and the strategy of CPU and GPU coordination are realized through the design of snapshots and buffers. Also, the access to the traffic simulation part has been completed for the overall city simulation system. Performance tests in various simulation scenarios show that compared with the original simulation system, the system has achieved a 2.18 times improvement in the simulation efficiency. At the same time, through the visualization results, the correctness of the simulation results of our system is confirmed.

Real-world recommender systems always suffer from the long-tail problem, which leads to the inferior recommendation performance, especially for tail items. Most existing works approach this problem by leveraging auxiliary data, which is not practical. In this work, we address the problem by proposing the SELT method (short for SELf-Training for Long-Tail recommendation). In short, there are two critical challenges including 1) how to generate and 2) how to exploit pseudo labels (i.e., user-item interactions). To this end, we first propose strategies of per-dimension normalization and adaptive generation, to address the first challenge. We then propose a teacher-student self-learning framework, under which the student model is supervised by pseudo labels with the teacher model knowledge integrated. As a result, information within pseudo and original labels can compensate for each other, addressing the second challenge. Extensive experimental results on three benchmark datasets demonstrate the effectiveness of our proposed method. Our SELT can significantly and steadily outperform state-of-the-art methods, no matter for head or tail.

Nowadays, researchers are increasingly applying data-driven methods to solve interdisciplinary issues. This talk will introduce two representative works that effectively utilize mobility data to handle novel but challenging interdisciplinary problems. Specifically, the first work proposed a building occupancy estimation method with mobile phone data, which is meaningful for sustainable city planning. The second work introduced mobility data to scheduling plug-in electric vehicles' charging so that we can shave the pronounced peak in power demand. In conclusion, both works carefully dig into the implications of known data and apply these implications to solve some interdisciplinary topics that lack direct data sources. Such a way of thinking can inspire us in dealing with problems that lack direct data sources.

Traffic simulator forms the foundation of intelligent transportation systems, supporting important applications, including traffic analysis, control, and optimization. Existing state-of-the-art simulators typically use physical models or empirical rules based on expert knowledge to simulate the car-following and lane-changing behaviors separately, which leads to two drawbacks. First, they lack simulation accuracy because the car-following and lane-changing behaviors are inherently related. Second, they fall short in characterizing heterogeneous driver behaviors due to the constraint representation power of the models. Some recent work tries to solve this problem by deep learning models, yet these models typically have poor generalizability and thereby are not qualified for application. To bridge this gap, we propose a Hybrid Physics-informed Traffic Simulator (HPTS) that jointly models the car-following and lane-changing behaviors. Specifically, to equip our model with long term generalizability, we propose to introduce physical models as prior knowledge to enhance the neural network model. In our car following model, we take a residual learning framework and let the neural network learn the residuals of the physical model calibrated on the real-world data. In our lane-changing model, we introduce the expected utility theory to guide the design. To jointly model the two behaviors, we design a dynamic interaction mechanism to break the isolation between the car-following and lane-changing model, enabling us to optimize the two models together.

In wireless twin communication network , heterogeneous network is a novel network architecture proposed in long-term-evolution, which highly increases the capacity and coverage compared with the conventional networks. However, in order to provide the best services, appropriate resource management must be applied. In this talk,we consider the joint optimization problem of user association,subchannel allocation and power allocation for downlink transmission in multi-association access heterogeneous networks. To solve the optimization problem, we first divide it into two subproblems: 1) user association and subchannel allocation for fixed power allocation and 2) power allocation for fixed user association and subchannel allocation. Subsequently, we obtain a locally optimal solution for the joint optimization problem by solving these two sub problems alternately. For the first subproblem, we derive the globally optimal solution based on graph theory. For the second subproblem, we obtain solver solution on the difference of two convex functions approximation method. Simulation results demonstrate that the proposed algorithms can significantly enhance the overall network throughput.For the large-scale and high-speed problems required by the industry, an algorithm is proposed to classify and process base stations according to frequency bands in parallel, and then jointly process them globally. Under the condition of constant performance, the time complexity is significantly reduced and the scale is improved.

Vehicular fog computing (VFC) has emerged as a promising paradigm to alleviate the computation workload of static infrastructure. In this context, exploiting the underutilized resources of congested vehicles is particularly attractive, while the respective offloading mechanisms have been less explored. With this motivation, we leverage jammed vehicles to serve the computation tasks of base station (BS) and propose a novel task offloading framework which facilitates parallel execution at each jammed vehicular cloudlet (JVC) by segmenting task data. Keeping in view the time varying computation capacity of JVCs and heterogeneity of computation tasks, we formulate task assignment and resource allocation to maximize the total profit of the considered vehicular edge computing network over a short horizon. We obtain the optimal solution using Genetic Algorithm (GA), and further propose a low complexity greedy scheme named Profit-Aware Task Assignment (PATS). PATS handles the optimization as a fractional-knapsack problem and assigns offloading tasks to JVCs in each slot by taking account of time varying communication and computation costs. Extensive performance evaluation shows that, GA outperforms other baseline schemes while PATS achieves a comparable performance to GA. Although PATS is sub-optimal theoretically, it provides acceptable performance and is suitable for large scale implementation.

In real-world express systems, couriers need to satisfy not only the delivery demands but also the pick-up demands of customers. Delivery and pickup tasks are usually mixed together within integrated routing plans. Such a mixed routing problem can be abstracted and formulated as Vehicle Routing Problem with Mixed Delivery and Pickup (VRPMDP), which is an NP-hard combinatorial optimization problem. To solve VRPMDP, there are three major challenges as below. a) Even though successive pickup and delivery tasks are independent to accomplish, the inter-influence between choosing pickup task or delivery task to deal with still exists. b) Due to the two-way flow of goods between the depot and customers, the loading rate of vehicles when leaving the depot affects routing decisions. c) The proportion of deliveries and pickups will change due to the complex demand situation in real-world scenarios, which requires robustness of the algorithm. To solve the challenges above, we design an encoder-decoder based framework to generate high-quality and robust VRPMDP solutions. First, we consider a VRPMDP instance as a graph and utilize a GNN encoder to extract the feature of the instance effectively. The detailed routing solutions are further decoded as a sequence by the decoder with attention mechanism. Second, we propose a Coordinated Decision of Loading and Routing (CDLR) mechanism to determine the loading rate dynamically after the vehicle returns to the depot thus avoiding the influence of improper loading rate settings. Finally, the model equipped with a GNN encoder and CDLR simultaneously can adapt to the changes in the proportion of deliveries and pickups. We conduct the experiments to demonstrate the effectiveness of our model. The experiments show that our method achieves desirable results and has great generalization ability. In addition, I will introduce the further application and research of this method in the real express system.

Location-aware recommendation leverages the location context information of user interactions to make recommendations. Existing works have mostly overlooked the location context's impact on user preference which is complex and implicit; this impact is moderated by the characteristics of the user, the item, and the location. To fill the gap, we propose our DALE (short for Disentangled preference-learning for location-aware recommendation), which explicitly captures the complex interdependence between users’ physical visitation and online content consumption with special designs of graph propagation, disentangled embedding, self-supervised learning, and multi-task learning. Specifically, we model location-aware user preference via location edge-based user-item embedding propagation. We disentangle users’ preferences that are dependent on specific locations from those independent ones via a self-supervised learning method. We design a multi-task learning workflow jointly optimizing two tasks of location visitation prediction and recommendation. Extensive experiments on two datasets show that our DALE model outperforms the state-of-the-art baselines by 6.09% and 3.96% on average on the two datasets. Further studies also validate the effectiveness of designs in our proposed model. Our work offers new insights on understanding and modeling the complex dependency of users’ online and offline behaviors.

In recent years, short-form video platforms emerged rapidly and attracted a large and wide variety of users. While great efforts have been invested to develop advanced recommendation algorithms, another perspective of users is still unexplored. In this work, we focus on information cocoon to measure user states of video consumption on the platform, i.e., the homogeneity of video contents. Specifically, an exploratory study is conducted on this phenomenon on Kuaishou with a large-scale dataset of half a million new users' behavioral records that lasted for one year. Firstly, we evaluate the evolution of users' information cocoons and find the limitation of recommender on user requirement for diverse video content. In addition, the anatomy of information cocoons is further explored via correlation analysis from three aspects, including user demographics, video categories, and their interactions driven by recommender and user preferences. Correspondingly, we observe that user demographics are not significant factors affecting their cocoons in general. However, certain video contents do make users addicted and lead to more heterogeneous consumption. At last, although recommender takes effect on cocoons formation, more accurate personalization does not contribute to more severe cocoons. In contrast, users with narrow preferences are more likely to be trapped. In summary, our study illuminates the current worrying of overwhelmingly heterogeneous video consumption due to over-personalization on online short-video platforms. Besides, we provide potential directions for mitigating information cocoons implied by correlation analysis.

Epidemic simulation is an important part of risk evaluation in the urban simulator. This research aims to construct an agent-level network-based simulator. Moreover, we are attempting to model the complexity of the transmission process to assist the simulation. Firstly, to model the effect of networks’ community structure in the spreading process, we designed a novel information theory-based community detection algorithm to group vertices that are related in the view of causality. Secondly, to model the circular structure in a complex network, we adopt Persistent Homology, which is a topological data mining approach, which could measure the network’s clustering feature and hollow feature. Finally, in addition to the above two feature extraction methods, we are designing Graph

Collective intelligence(CI), defined as the ability of a group to perform a wide variety of tasks, is critical to solving many scientific, business, and other problems. There has been much research to explore the determinants of collective intelligence, among which the structure of group collaboration and communication is strongly correlated with CI. Similarly in the neural network, the model performance is influenced by its structure and relies on the information flow between nodes/agents as in the social collaboration. Inspired by this we introduce a new framework to model the group collaboration in the neural network. In this framework we concentrate on two factors affecting the structure of agent communication: inequlity of speaking turns and modularity. The experiment results demonstrate the consistent pattern shown in social collaboration and information flow in neural network, providing evidence for linking connectionism in neural network and collective intelligence in social science. Simultaneously this finding could offer some beneficial insights for designing more effective neural network-based models in aritificial intelligence.

Comparing the topological structures of urban road networks (URNs) can benefit various studies such as understanding urban growth. Existing approaches to characterize URNs fail to quantify the similarity between URNs, while the widely used graph neural networks shed light on this problem. In this talk, I’m going to introduce a recent work that leverages graph neural networks to measure the spatial homogeneity of URNs. This study claims that the spatial homogeneity is closely related to the predictability of road links. Specifically, the authors train a GNN model on part of the road networks to predict the existence of the rest of road links, and use the prediction accuracy to measure the spatial homogeneity. Furthermore, experiments within a city and between different cities around the world show that the method can help understand city development, indicate intercity similarity and reveal historical urban insights.

With the prevalence of smart appliances, smart meters, and Internet of Things (IoT) devices in smart grids, artificial intelligence (AI) built on the rich IoT big data enables various energy data analysis applications and brings intelligent and personalized energy services for users. In conventional AI of Things (AIoT) paradigms, a wealth of individual energy data distributed across users’ IoT devices needs to be migrated to a central storage (e.g., cloud or edge device) for knowledge extraction, which may impose severe privacy violation and data misuse risks. Federated learning, as an appealing privacy-preserving AI paradigm, enables energy data owners (EDOs) to cooperatively train a shared AI model without revealing the local energy data. Nevertheless, potential security and efficiency concerns still impede the deployment of federatedlearning-based AIoT services in smart grids due to the low-quality shared local models, non-independently and identically distributed (non-IID) data distributions, and unpredictable communication delays. In this article, we propose a secure and efficient federated-learning-enabled AIoT scheme for private energy data sharing in smart grids with edge-cloud collaboration. Specifically, we first introduce an edge-cloud-assisted federated learning framework for communication-efficient and privacy-preserving energy data sharing of users in smart grids. Then, by considering non-IID effects, we design a local data evaluation mechanism in federated learning and formulate two optimization problems for EDOs and energy service providers. Furthermore, due to the lack of knowledge of multidimensional user private information in practical scenarios, a two-layer deep reinforcement-learning-based incentive algorithm is developed to promote EDOs’ participation and high-quality model contribution. Extensive simulation results show that the proposed scheme can effectively stimulate EDOs to share high-quality local model updates and improve the communication efficiency.

Routing problems, from the fundamental traveling salesman problem (TSP), to more complicated vehicle routing problems (VRP) with practical constraints, play an important role in many real-world applications. Generating solutions both effectively and efficiently is critical in reducing industrial costs and improve urban life efficiency. Besides extensive research on utilizing exact or heuristic/meta-heuristic methods to solve these NP-hard problems, recent works also focus on using deep learning-based technologies. However, these attempts usually stops at basic problem settings, including TSP and capacitated VRPs (CVRP), and ignores more practical constraints, and are limited in generalizability. In this project, we aim at developing a general routing solvers that considers different constraints automatically, which may have a great potential in real-world usage.

Recent years have witnessed an explosion of extensive geolocated datasets related to human movement, enabling scientists to quantitatively study individual and collective mobility patterns, and to generate models that can capture and reproduce the spatio-temporal structures and regularities in human trajectories. The study of human mobility is especially important for applications such as estimating traffic forecasting, urban planning, and epidemic modeling. I will review the approaches developed to reproduce various mobility patterns, with the main focus on recent developments. The related work can be presented in two main aspects: individual mobility and population mobility. This topic can be used both as an introduction to the fundamental modeling principles of human mobility, and as a collection of technical methods applicable to specific mobility-related problems.

Nowadays, people have increasing demands for accurate and real-time navigational information through the complex road network. To better facilitate the travel of users, fastest route recommendation (FRR) has become an important task in urban computing. The challenge of this work is how to accurately predict the traffic condition in the future and find the fastest path under dynamic traffic condition. In this seminar, I will introduce some related papers I have studied on FRR, including the dynamic improvement of the original heuristic path planning algorithm and the combination of reinforcement learning and heuristic algorithm to solve the route planning problem. Current research usually divides the FRR problem into two parts: prediction and planning, and models them independently, which ignores the correlation between prediction and planning, and limits the model accuracy. It is expected to design an end-to-end network to solve this problem.

Mobility simulation is essential for applications ranging from location-based services to urban planning. Yet many existing models either only simulate macroscopic mobility using aggregated flow data or model individual mobility choices without considering population dynamics. However, macroscopic patterns are important as they are a fundamental level of mobility organization, and thereby influence individual decisions, meanwhile individual preferences determine fine-grained mobility choices, which in turn affect population patterns. In this work, we propose a cooperative framework for mobility simulation by integrating individual preference and population dynamics. It is expected that the framework can model the influence of social interactions and individual preference in the simulation process, and thus achieve better performance over both microscopic and macroscopic metrics.

Nowadays, city is a complex couple of systems interacting and cooperating with each other. For example, traffic system, energy(electricity) system, communication system, water supply and drainage system. Obviously, these systems can exchange substances in physical space, what’s more, in modern efforts of constructing and managing smart cities, we are interested in the exchange of information among systems in cyber space. I’m going to introduce people’s recent rethinking on city from the perspective of system theory, which is about smart city as Cyber-Physical Systems. Several papers will be reviewed about the main concepts and potential research problems of how to understand, simulate and decide when faced with the modern smart cities from the complex system theory. I will also combine with our current efforts at our simulation platform of simulating the coupling effects of different systems in city as examples. Besides, the deeper interaction among city systems also brings the higher risk of the cascading failure, so the issues of city resilience or city vulnerability will also be introduced.

One of the basic values of Internet companies is data. Every day, a huge number of big data tasks will be executed to assure data available in real time. As a result, daily scheduling projects are characterized for large scale, complex and fixed dependencies, and consistent executions. Project scheduling can make better use of your computing resources and save money. Poor scheduling strategies can lead to resource waste and task failure. In this paper, we study a resource estimation and scheduling strategies for large-scale time-constrained projects. First, by combining project decomposition with an operational methodology, a method for estimating the resource consumption of large-scale projects is given. Then, a priority-based scheduling technique is provided to improve the project completion rate in the case of exceptions. We create a large-scale dataset including 300,000 tasks based on a real-world scenario of Internet companies, then do strategy simulation experiment. The experimental results show that the proposed scheduling strategy reduces the resource consumption by 40% and the completion rate is improved.

The Transformer architecture has become a dominant choice in many domains, such as natural language processing and computer vision. Recently, some works have also investigated the potential of Transformer on graph representation learning. Therefore, from the perspective of knowledge graph (KG), a natural question is, can Transformer perform well on KG representation learning? In this seminar, I will firstly give a brief review of various types of KGs including KG, temporal KG and hyper-relational KG. Then I will introduce the basic design of Transformer, further cover related works leveraging Transformer for graph and KG learning. Finally, I will identify our research problem and summarize some early tries.

Urban digital twin water network simulation can help the management to better understand the operation of water infrastructure in the city and make better management. The water system in a city includes two major parts, the water supply network and the drainage network, both of which are the lifeline of the city. In our urban water system simulation, we include the simulation of the water supply network as well as the simulation of the drainage network. The water supply network mainly collects water from water sources and transports it to the water storage points for domestic production; the drainage network mainly collects and discharges wastewater from production and domestic use into the city or surrounding waters, as well as discharges of precipitation.We have implemented the coupling of water systems with population activities and other infrastructure networks in cities, providing an integrated simulation platform for the study of urban complex systems.

In the era of information explosion, recommender systems play an indispensable role in what we see, what we purchase, and even what we like. Despite the convenience, recommender systems could potentially isolate people from a diversity of viewpoints or content, forming information cocoons. To delineate the mechanisms behind the phenomenon, we provide a complex adaptive-systems perspective on the effects of recommender systems and intend to reveal the dynamics of the formation of information cocoons. Then we (intend to) validate our framework with empirical results.

Traffic congestion has a negative impact on our daily lives. Predicting the trend of traffic congestion provides the valuable guideline to address the problem of the traffic congestion. Most existing approaches focus on the tasks of predicting traffic volume and traffic speed, which does not effectively address the challenges on traffic congestion prediction. First, traffic congestion exhibits daily and weekly temporal patterns, but these patterns are not strictly the same. This indicate the complicated long-term periodicity. Second, traffic congestion sparsely distributes over different periods of the time, which leads to the complex short-term and mid-term temporal dependencies. Third, since traffic congestion would propagate to adjacent road segments over time, which indicates that it has complex spatio-temporal correlations. To address them, we propose a periodic spatio-temporal graph convolutional network for traffic congestion prediction. Specifically, we propose to capture the peculiarity and commonality of long-term periodic temporal patterns to handle the complicated long-term periodicity. To effectively capture short-term and mid-term temporal dependencies, we regard a continuous time sequence of the congested conditions as a traffic congestion event, and then utilize the long short-term memory model to learn the sequential dependencies of traffic congestion events. A graph convolutional network is further incorporated to explore the complex spatio-temporal correlations. Extensive experiments demonstrate the superiority of our model. In addition, we deploy our model in production at Amap, and it achieves great performance improvement in terms of F1-score compared to the production baseline. This confirms that our model is a practical solution for the real-world congestion prediction services.

Most of the epidemic models we have seen focus on the aggregated level of epidemic curves, which can only predict the cumulative infections of disease. However, what we more care about is the underlying microscopic epidemic patterns behind the simple curve. It is at critical importance for many real-world tasks, such as source detection of the epidemic outbreak and fine-grained lockdown measures. In this work, we try to build a diffusion based GCN model to capture the microscopic epidemic patterns that has a strong correlation with physical process, providing intuitions for understanding how the disease spread in a social network.

City is a dynamic complex system, which contains many aspects such as transport, water, energy, communications, economy, society, etc. The complexity of city systems poses great challenges for city simulation. It is not easy to organize such aspects of city together and keep the whole simulation system scalable. Based on a basic understanding of city and existing engineering practice, we believe that human is the key to connecting various elements in city and then we propose a human-centered city simulation framework that connects our traffic simulation, water simulation, power simulation and communication simulation together. The framework focuses on three concepts, human, thing and space. Human has some attributes, like position, demand, schedule. Generating algorithms are responsible for generating part of attributes of human. Each simulation module maintains part of attributes of human in specific spaces, which reflects the relationship between human and thing. Data exchange between different simulation modules reflects the relationship between things.

Recent years have witnessed the rapid development of online micro-video platforms, in which the recommender system plays an essential role in overcoming the information overloading problem and providing personalized content for users. Some business scenarios of micro-video platforms have UGC characteristics and videos contain geographic information, however, the application of UGC characteristics and geographic information in micro-video recommendation is still blank. Making recommendations based on the nature of UGC and the introduction of geographic information is conducive to enhancing user engagement, increasing user loyalty, and also promoting the transformation of users from online content consumption to offline actual consumption. To make good use of the UGC characteristics and geographic information in micro-video recommendation, we propose a solution based on graph neural network. Specifically, we design a heterogenous graph with three type of nodes: user, video and region. Customized node role differentiation and embedding aggregation methods are employed to get good representation of the nodes. The experimental results on a real-world dataset demonstrate the effectiveness of our proposed model.

The development of emerging network applications (e.g., AR/VR, telesurgery, real-time communications) is increasing the challenge of maintaining current communication networks. These applications usually have high criteria (e.g., ultra-low deterministic latency), making it more challenging for network operators to manage their network resources effectively. In this session, I will introduce the Network Digital Twin (NDT) as a critical enabler for effective network administration in contemporary networks. I outline the overall architecture of the NDT and suggest that current advancements in Machine Learning (ML) allow developing an NDT that efficiently and accurately replicates real-world networks. In addition, I analyze the primary ML technologies that allow constructing the components of the NDT architecture. Finally, I outline the open issues that the research community needs to solve in the following years in order to allow the application of the NDT in real-world settings.

Epidemics, including SARS, H1N1, and most recently, the prolonged COVID-19, pose substantial threats to human society, especially in urban environments where various locations are closely connected by intense human mobility. Since the collection and processing of large-scale, fine-grained human mobility data have been enabled by recent years' rapid development of Information and Communication Technologies (ICTs), researchers have proposed numerous new ways of modeling epidemics via leveraging such data, which accelerate our understanding of transmission patterns and improve public health policies. In this survey, we first provide an overview of human mobility data (by collection method, spatial & temporal resolution, duration, etc.). Then we categorize state-of-the-art studies into four topics, based on the role human mobility plays: (1) real-time monitoring/tracing; (2) simulation/forecasting (further, for policy evaluation); (3) identifying and explaining epidemic-related inequalities; and (4) informing post-pandemic planning/urban sustainability impacts. Finally, we discuss existing challenges and future research directions.

Urban development influences people’s travel options, therefore changing the OD flow distribution of the city. Knowing how urban development impacts OD flow is crucial for city development and road planning. However, urban development, typically in the form of POI evolutions, involves all kinds of POIs, and therefore is complicated to model. In this work, we present a framework to model the impact of urban development on OD flow distribution from a causal perspective.

The disadvantaged population is often underserved and marginalized in technology engagement: prior works show they are generally more reluctant and experience more barriers in adopting and engaging with mainstream technology. Here, we contribute to the HCI4D and ICTD literature through a novel “counter” case study on Chinese social commerce (e.g., Pinduoduo), which 1) first prospers among the traditionally underserved community from developing regions ahead of the more technologically advantaged communities, and 2) has been heavily engaged by this community. Through in-depth interviews with social commerce users from the traditionally underserved community in Chinese developing regions, we demonstrate how social commerce, acting as a “virtual bazaar”, brings online the traditional offline socioeconomic lives the community has lived for ages, fits into the community's social, cultural, and economic context, and thus effectively promotes technology inclusivity. Our work provides novel insights and implications for building inclusive technology for the “next billion” population.

In nature, networks rarely appear in isolation. They are typical elements in larger systems and can have non-trivial effects on one another. For example, network elements such as base stations of a communication network often depend on the supply of the power grid network. These networks, referred to as interdependent networks, have been studied for more than a decade, where the most famous research case is the 2003 Italian blackout. In this example, the failure of a single node of the power grid network caused the failure of the connected nodes in its interdependent communication network, resulting in a change in connectivity between the nodes of the communication network. In turn, it further influences the transmission of control signals of the power grid network, which eventually leads to the breakdown of the two networks. This cascading failure process is also an important research problem of interdependent networks. A series of follow-up works have used the percolation theory to study the phase transition process, attack strategy, and recovery process in interdependent networks. In this talk, we will review relevant papers published in the past 10 years, and think of the potential research problems based on our laboratory's urban vulnerability simulation platform.

In mobile edge computing (MEC), the computation offloading of massive users could cause the task uploading congestion to deteriorate the users’ offloading delay. The non-orthogonal multiple access (NOMA) enabled MEC is envisioned to address this issue by allowing multiple users to simultaneously upload their tasks on one subchannel. However, the differentiated uploading delay of users may make task uploading completion inconsistent with NOMA decoding order, which complicates the co-channel interference and restricts NOMA to reducing the uploading delay. In this paper, we characterize the interaction between the differentiated uploading delay and co-channel interference for a pair of NOMA users. Furthermore, we propose a computation offloading scheme to reduce the users’ average offloading delay by jointly optimizing offloading decision and resource allocation. Specifically, the proposed scheme first obtains the optimal power allocation based on the characterized interaction and the closed-form solution of computation resource allocation by convex programming. Then, the NOMA user pairing and offloading decision are iteratively determined by semidefinite relaxation and convex-concave procedure. Simulation results show that the proposed scheme effectively mitigates co-channel interference under differentiated uploading delay of users and outperforms in reducing the users’ average offloading delay and increasing the number of users to offload tasks.

Traditional symbolic regression aims to discover the underlying equation from a set of input-output pairs. We generalize this task on a graph, which widely exists in the real world, such as rediscovering the law of universal gravitation, Electric Potential energy formula. Specifically, given the input consisting of a graph and some variables on the node, edge, and graph levels, and output consisting of a dependent variable on the node, edge, or graph level, we aim to distill graph-structured physical mechanism described by equations. For example, given several objects’ mass, position, and velocity (node variable), their relative position (edge variable), and the graph of spring linking them, we aim to find the symbolic model to describe the relationship between the acceleration and the above variables. However, it is challenging because such calculation involves not only a lot of symbols but also a message-passing structure to be searched. To tackle this problem, we first design a search space of message-passing structures and propose to search for it by simplifying the structure while maintaining accuracy, and then distill equations from the component functions of the message-passing network.

Social recommendation assists users' interest prediction by using social relationships, thereby improving the performance of the recommender system. However, in real scenarios, relationships on the web are not necessarily the same as having similar interests in the real world, which means that the user's social relationship is not always true, which leads to noises. The existence of these noisy data leads to a negative impact when directly using social relationships to enhance the recommendation performance in social recommendations. In order to solve the problem, in this work, we designed a social relationship denoising method. Specifically, we propose to explicitly measure the confidence of social relationships through the user's behavioral history. We design a transformer-based neural network module and train the denoising model via self-supervised learning. We also propose to assign different denoising strategies to different users based on social theories. The experimental results on several classic social recommendation models show that the recommendation performance is effectively improved by removing the noisy relations in the social network.

Data caching in D2D networks is a promising technique to reduce communication overhead. However, the efficiency of caching techniques is dependent on users’ data viewing preferences. To reap the full benefits of data caching, recommendation systems appear to be a promising approach to reshape the data request probability of users based on their data preferences, which can improve content caching decisions significantly. It is essential to quantitatively investigate how recommendation can be applied to maximize the cache hit ratio. It requires a time-efficient algorithm joint recommendation and caching policies to maximize the efficiency of scarce cache resources.

Network traffic data facilitates understanding the Internet of Things (IoT) behaviors and improving IoT service quality in the real world. However, large-scale IoT traffic data is rarely accessible, and privacy issues also impede realistic data sharing even with anonymous personal identifiable information. Researchers propose to generate synthetic IoT traffic but fail to cover the multiple services provided by widespread real-world IoT devices. In this work, we take the first step to generate large-scale IoT traffic via a knowledge-enhanced generative adversarial network (GAN) framework, which introduces both the semantic knowledge (e.g., location and environment information) and the network structure knowledge for various IoT devices via a knowledge graph. We use a condition mechanism to incorporate the knowledge and device category for IoT traffic generation. Then, we adopt LSTM and a self-attention mechanism to capture the temporal correlation in the traffic series. Extensive experiment results show that the synthetic IoT traffic datasets generated by our proposed model outperform state-of-art baselines in terms of data fidelity and applications. Moreover, our proposed model is able to generate realistic data by only training on small real datasets with knowledge enhanced.

Satellite imagery depicts the earth’s surface remotely and provides comprehensive information for many applications, such as land use monitoring and urban planning. Existing studies on unsupervised representation learning for satellite images only take into account the images’ geographic information, ignoring human activity factors. To bridge this gap, we propose using Point-of-Interest (POI)data to capture human factors and design a contrastive learning-based framework to consolidate the representation of satellite imagery with POI information. Also, we design an attention model that merges the representations from the geographic and POI perspectives adaptively. On the basis of real-world datasets collected from Beijing, we evaluate our method for predicting socioeconomic indicators. The results show that the representation containing POI information outperforms the geographic representation in estimating commercial activity-related indicators. Our proposed framework can estimate the socioeconomic indicators with an R2 of 0.874 and outperforms the baseline methods.

Modeling users’ long-term and short-term interests is crucial for accurate recommendation. However, since there is no manually annotated label for user interests, existing approaches always follow the paradigm of entangling these two aspects, which may lead to inferior recommendation accuracy and interpretability. In this paper, to address it, we propose a Contrastive learning framework to disentangle Long and Short-term interests for Recommendation (CLSR) with self-supervision. Specifically, we first propose two separate encoders to independently capture user interests of different time scales. We then extract long-term and short-term interests proxies from the interaction sequences, which serve as pseudo labels for user interests. Then pairwise contrastive tasks are designed to supervise the similarity between interest representations and their corresponding interest proxies. Finally, since the importance of long-term and short-term interests is dynamically changing, we propose to adaptively aggregate them through an attention-based network for prediction. We conduct experiments on two large-scale real-world datasets for e-commerce and short-video recommendation. Empirical results show that our CLSR consistently outperforms all state-of-the-art models with significant improvements: GAUC is improved by over 0.01, and NDCG is improved by over 4%. Further counterfactual evaluations demonstrate that stronger disentanglement of long and short-term interests is successfully achieved by CLSR. The code and data are available at https://github.com/tsinghua-fib-lab/CLSR.

Rapidly developing location acquisition technology has provided us with massive mobile trajectories, which supports a fantastic way of understanding and modeling individuals’ mobility. However, existing data-driven methods either fail to capture the long-range dependency or suffer from a high computational cost. To overcome these issues, we propose a knowledge-driven framework for mobility prediction, which leverages knowledge graph (KG) to formulate the mobility prediction task into the KG completion problem through integrating the structured “knowledge” from the mobility data. However, most related mobility prediction works only focus on the structured information encoded in existing triples, which ignores the rich semantic information of paths. In this paper, we apply a dedicated module to extract the supplementary semantic structure of paths in KG, which contributes to the interpretability and accuracy of our model. Specifically, the extracted rules are applied to capture the dependencies between relational facts. Moreover, incorporating user information in the entity-relation space with the corresponding hyperplane, our method could capture diverse user mobility patterns and model user characteristics to improve the accuracy of mobility prediction. Extensive evaluations illustrate that our proposed model beats state-of-the-art mobility prediction algorithms, which verifies the superiority of utilizing the logical rules and user hyperplanes.

Mixed Integer Programs (MIPs) are a class of NP-hard problems where the goal is to minimize a linear objective subject to linear constraints, with some or all of the variables constrained to be integer-valued. Significant research and engineering effort have gone into developing practical solvers, such as SCIP, CPLEX, Gurobi, and Xpress. These solvers use sophisticated heuristics to direct the search process for solving a MIP. Meanwhile, the rapid development of deep learning techniques brings new opportunities in augmenting neural networks for more efficient and effective MIP solving. We thus summarize and review several major advanced methods in recent years for solving MIPs via learning, and hope the review could offer new research insights to those who are working on optimization-related projects in our lab.

With the proliferation of personalized mobile applications, high-quality user profiling is becoming more and more important. For example, it can improve the user experience in the recommender system. The correlations between user profiles(e.g., gender, age, occupation) and their mobility patterns have been widely observed, which makes it possible to leverage mobility data to infer user’s profiles. However, existing methods mostly depend on feature engineering, which is time-consuming. Apart from that, some methods based on deep learning often ignore the semantics of user mobility. In this work, we propose a knowledge-driven model, which aims to learn effective user embeddings for user profiling. Specifically, we first construct an urban KG with a city’s key elements, relationships, and users’ trajectories, which contain distilled knowledge and rich semantics. Then the urban KG is fed into a KG embedding model to generate user’s KG embedding. Besides, in order to better characterize the user’s trajectory, we further use the relationships in KG to guide a self-supervised model for user’s trajectory representation learning. Finally, the user’s KG embedding and trajectory embedding are combined to get the user’s embedding. Experiments show that the learned user embeddings can help user profiling task.

Recent years have witnessed the rapid development of online micro-video platforms, in which the recommender system plays an essential role in overcoming the information overloading problem and providing personalized content for users. Although some progress has been achieved in the micro-video recommendation, there are still some limitations in learning the representations of user interests and video features. Specifically, the user modeling in existing works is performed at a coarse-grained level, i.e., video level. However, in micro-video recommendation, the user feedback is at a continuous form—users can skip over a video at each frame—which reveals fine-grained user preferences. In this work, we approach the problem of learning fine-grained user preferences for micro-video recommendation by first collecting two real-world datasets. To address the challenges of preference modeling and weak supervision signal, we propose a solution named FRAME (short for FinegRAined preference-modeling for Micro-video rEcommendation). Specifically, we first adopt visual feature extraction and transformation to maintain the fine-grained video embeddings. We then propose graph convolution layers to learn the user preference from complex and fine-grained user-clip relations and hybridsupervision objectives for enhancing the supervision signal. The experimental results on two collected real-world datasets demonstrate the effectiveness of our proposed model.

Crowd simulation acts as the basic component in traffic management, urban planning, and emergency management. Most existing approaches use physics-based models due to their robustness and strong generalizability, yet they fall short infidelity since human behaviors are too complex and heterogeneous for a universal physical model to describe. Recent research tries to solve this problem by deep learning methods. However, they are still unable to generalize well beyond training distributions. In this work, we propose to jointly leverage the strength of the physical and neural network models for crowd simulation by a Physics-Infused Machine Learning (PIML) framework. The key idea is to let the two models learn from each other by iteratively going through a physics-informed machine learning process and a machine-learning-aided physics discovery process. We present our realization of the framework with a novel neural network model, Physics-informed Crowd Simulator (PCS), and tailored interaction mechanisms enabling the two models to facilitate each other. Specifically, our designs enable the neural network model to identify generalizable signals from real-world data better and yield physically consistent simulations with the physical model's form and simulation results as a prior. Further, by performing symbolic regression on the well-trained neural network, we obtain improved physical models that better describe crowd dynamics. Extensive experiments on two publicly available large-scale real-world datasets show that, with the framework, we successfully obtain a neural network model with strong generalizability and a new physical model with valid physical meanings at the same time. Both models outperform existing state-of-the-art simulation methods in accuracy, fidelity, and generalizability, which demonstrates the effectiveness of the PIML framework for improving simulation performance and its capability for facilitating scientific discovery and deepening our understandings of crowd dynamics.

Consumption intent, defined as the decision-driven force of consumption behaviors, is crucial for improving the explainability and performance of user-modeling systems, with various downstream applications like recommendation and targeted marketing. However, consumption intent is implicit, and only a few known intents have been explored from the user consumption data in Meituan. Hence, discovering new consumption intents is a crucial but challenging task, which suffers from two critical challenges: 1) how to encode the consumption intent related to multiple aspects of preferences, and 2) how to discover the new intents with only a few known ones. In Meituan, we designed the AutoIntent system, consisting of the disentangled intent encoder and intent discovery decoder, to address the above challenges. Specifically, for the disentangled intent encoder, we construct three groups of dual hypergraphs to capture the high-order relations under the three aspects of preferences and then utilize the designed hypergraph neural networks to extract disentangled intent features. For the intent discovery decoder, we propose to build intent-pair pseudo labels based on the denoised feature similarities to transfer knowledge from known intents to new ones. Extensive offline evaluations verify that AutoIntent can effectively discover unknown consumption intents. Moreover, we design the downstream application experiments to verify the effectiveness of the deployment strategy of AutoIntent.

In the fight against COVID-19 pandemic, vaccines are the most critical resource but are still in short supply around the world. Therefore, efficient vaccines allocation strategies are urgently called for, especially in large-scale metropolis where uneven health risk is manifest in nearby neighborhoods. However, there exist several key challenges in solving such problem: (1) great computation complexity in the large scale scenario adds to the difficulty in experts’ vaccines allocation decision making; (2) heterogeneous information from all aspects in the metropolis’ contact network makes information utilization difficult in decision making; (3) when utilizing the strong decision-making ability of reinforcement learning (RL) to solve the problem, poor explainability limits the credibility of the RL strategies. In this paper, we propose a reinforcement learning enhanced experts method. We deal with the great computation complexity via a specially designed algorithm aggregating blocks in the metropolis into communities and we hierarchically integrate RL among the communities and experts’ solutions within each community. We design a self-supervised contact network representation algorithm to fuse the heterogeneous information for efficient vaccines allocation decision making. We conduct extensive experiments in three US metropolises with real-world data and the results show that our method outperforms the best baseline, reducing 9.01% infections and 12.27% deaths. We further demonstrate the explainability of the RL model, which not only adds to its credibility and feasibility in the real world but also enlightens the experts in turn.

Widely applied in today’s recommender systems, sequential recommendation predicts the next interacted item for a given user via his/her historical item sequence. However, sequential recommendation suffers data sparsity issues like most recommenders. To extract auxiliary signals from the data, some recent works exploit self-supervised learning to generate augmented data via dropout strategy, which, however, leads to sparser sequential data and obscure signals. In this paper, we propose Dual Contrastive Network (DCN) to boost sequential recommendation, from a new perspective of integrating auxiliary user-sequence for items. Specifically, we propose two kinds of contrastive learning. The first one is the dual representation contrastive learning that minimizes the distances between embeddings and sequence-representations of users/items. The second one is the dual interest contrastive learning which aims to self-supervise the static interest with the dynamic interest of next item prediction via auxiliary training. We also incorporate the auxiliary task of predicting next user for a given item's historical user sequence, which can capture the trends of items preferred by certain types of users. Experiments on benchmark datasets verify the effectiveness of our proposed method.

Synthesized user online store consumption sequence are crucial for a large number of applications. Generative adversarial imitation learning (GAIL) is promising for sequence generation. However, there are three main challenges existing in deploying GAIL to generate synthetic user online store consumption sequences: (1) the consumption decision process of users is complex and difficult to model; (2) user's decision-making space will change as the new stores go online, which cannot by handled by the existing GAIL framework; (3) sparse consumption decisions in the collected real consumption is hard for discriminator to judge whether they are real or not. In this paper, we propose a knowledge enhanced hierarchical GAIL for user consumption sequence synthesizing, named KnowledgeGAIL, to tackle them. For the first challenge, it designs a hierarchical policy function to model the complex consumption decision process of users from a psychological point of view. For the second challenge, it designs a key-value action layer structure to convert the number of changing stores into batch size, so that changes in the number of stores will not affect GAIL training. For the third challenge, it proposes a knowledge enhanced reward function, which combines the data-driven NN discriminator and the knowledge-based discriminator. The knowledge-based discriminator is general and able to cover most consumption decisions, therefore assisting the NN discriminator to judge those less frequent consumption decisions. Extensive experiments conducted on two real-life user consumption datasets demonstrate that the proposed framework outperforms all the state-of-the-art baselines significantly. Furthermore, the significant improvement in experimental performance on the practical application by utilizing our generated data validates that our approach is practical in real life.

Living in cities affords expanded access to various resources, infrastructures, and services at reduced travel cost, which improves urban life and promotes systemic gains in economic productivity, social interaction, and collective innovation. However, recent research shows that urban development also leads to inequality in accessing urban resources, which manifests in distinct travel and visitation patterns for residents with different demographic backgrounds. Here, we utilize large-scale open-access mobility data containing detailed records of US residents' daily interaction with more than one million urban venues to identify mobility inequality among urban residential neighborhoods. To tackle limitation of correlation analysis in previous literature, we propose a propensity score matching method to identify the causal treatment effect of demographic features on access to urban resources. Moreover, we design a novel Counterfactual RANdom-walks based Embedding (CRANE) method to disentangle the potential confounding effects on urban mobility network, which can efficiently approximate the causal effect of demographic feature on urban resource accessibility with continuous embedding vectors. Our analysis reveals that neighborhoods with higher average incomes display more extensive mobility and neighborhoods with higher portion of bachelor degree holders yield greater reductions in mobility during the COVID-19 pandemic. We also uncover significant racial and income inequality in visitation rates to sports and education venues. We demonstrate that our representation learning method can improve the explanatory capacity and robustness of our findings by extending it from aggregate statistics to individual neighborhoods and allowing for cross-city knowledge transfer.

Traffic congestion incurs long delays in travel time, which seriously affects our daily travel experience. Exploring why traffic congestion occurs is significantly important to effectively address the problem of traffic congestion and improve user experience. Traditional approaches to mine the congestion causes depend on human efforts, which is time consuming and cost-intensive. Hence, we aim to discover the known and unknown causes of traffic congestion in a systematic way. However, to achieve it, there are three challenges: 1) traffic congestion is affected by several factors with complex spatio-temporal relations; 2) the amount of congestion data with known causes is small due to the limitation of human label; 3) more unknown congestion causes are unexplored since several factors contribute to traffic congestion. To address above challenges, we design a congestion cause discovery system consisting of two modules: 1) congestion feature extraction, which extracts the important features influencing congestion; and 2) congestion cause discovery, which utilize a deep semi-supervised learning based method to discover the causes of traffic congestion with limited labeled causes. Specifically, it first leverages a few labeled data as prior knowledge to pre-train the model. Then, the k-means algorithm is performed to produce the clusters. Extensive experiments show that the performance of our proposed method is superior to the baselines. Additionally, our system is deployed and used in the practical production environment at Amap.

Public transportation is an important mode of travel for urban residents, and its route design has a greater impact on the quality of service provided to passengers and the interests of operators. With the continuous advancement of urbanization, the existing urban layout is quite different from the past, which leads to the need to optimize the existing public transportation routes. Urban Transit Routing Problem is an NP-hard, multi-constrained combinatorial optimization problem with a vast search space for which the evaluation of candidate solutions route sets can be both challenging and time-consuming, with many potential solutions rejected due to infeasibility. We propose a new method based on deep reinforcement learning to solve the Urban Transit Routing Problem.