Satellite imagery has been a promising source to infer regional socioeconomic status. Over the years, various methods have been proposed to predict the economic activity for irregular-shape regions, such as the administrative districts. However, they treat the region as a set of separate grids and merely aggregate the grid embeddings for downstream tasks, ignoring the geographical and semantic relationships between grids. In this work, we propose to represent the girds in a region as a graph structure and introduce a graph-based architecture to extract graph-level feature for downstream tasks. Specifically, we first split the irregular region to a number of patches which are viewed as nodes and construct a graph by connecting the nearest neighbors. Based on the graph representation of grids, we build a graph model to transform and exchange information among all the nodes. This is primarily a work proposal, and any suggestion or comment is welcome.

The next-generation mobile networks are evolving into very complex systems because of super-heterogeneous network architectures and various service requirements. Traditional networking approaches, i.e., reactive, centrally-managed approaches are of limited capability, and therefore, cannot satisfy the future complex networks. On the other hand, data-driven intelligent techniques offer as alternative options and possibly superior over traditional methods. With the development and application of 5G network, the communication capability of mobile network has been greatly enhanced, but at the same time, the complexity of the network has also been greatly increased. In order to propose an improvement plan for a series of problems such as mobile network optimization, energy saving, carbon reduction, etc., on the one hand, we will learn a traffic driven 3G mobile network traditional energy saving method and apply it to the complex 4G and 5G hybrid network in Nanchang. On the other hand, we will adopt some deep learning methods to improve each link of this energy saving method, and improve some shortcomings of the original method, to achieve better optimization effect on mobile network.

Effective urban planning plays a critical role in the sustainable development and competitiveness of cities. With the rapid urbanization process, tens of millions of people are migrating from rural areas to cities every year, requiring enormous number of urban communities to be planned and designed in a short time, which is an impossible task using only human labor. However, despite the convenience brought by geographic information system (GIS) and computer-aided design (CAD), the core and fundamental part of determining the spatial layout of land use and roads is still far from automation, relying heavily on human experts. Due to the large solution space, it is pretty difficult for human experts to obtain an optimal spatial plan with respect to multiple quantitative objectives. In this work, we propose an artificial intelligence (AI) urban planning framework, which utilizes the power of deep reinforcement learning (DRL) to automatically generate spatial plans for urban communities. As a general framework, we construct a graph to describe the topology of urban geographical elements in arbitrary irregular forms, and formulate urban planning as a sequential decision making problem on the graph. To tackle the challenge of the huge solution space, we then develop a general reinforcement learning model which conducts efficient search with value and policy networks based on graph neural network. Through massive number of trials, our method learns to lay out land use functionalities and roads effectively, achieving superior performance to plans designed by human experts in objective metrics, with significant improvements on the accessibility to urban services. Extensive experiments demonstrate that our framework can generate spatial plans for different community forms and in different styles. Finally, we propose a human-AI collaborative workflow of urban planning, in which our framework can make professional human designers more productive. Our framework demonstrates the great potential of data-driven urban design and paves the way for more explorations in leveraging the power of DRL in spatial planning.

Understanding mobile traffic data and predicting future traffic volumes plays an import role in the resource allocation and providing high quality services. However, predicting the mobile traffic data in a large scale is still challenging due to the complicated temporal variation and complex spatial relation. We proposed a knowledge graph driven approach, which is divided into a static pathway and a dynamic pathway, and we use urban knowledge graph to mine the different relationships between base stations. In the static pathway, we use a RGCN to mine the spatial relations between base stations. And we use Transformers Encoders and RGCN in the dynamic pathway to mine the temporal features of base station traffic sequences.

The base station is the main equipment for mobile network to provide services to users, distributed in various areas of physical space.It is the basis for users to access mobile network anytime and anywhere. Generate fine-grained changes in the aggregated traffic of the base station, which helps support digital twin networks, and design, test and improve mobile network planning and optimization strategies at a more macro level strategy and mechanism. We propose a multi-period base station aggregation traffic generation method for urban knowledge enhancement. We model the impact of entities such as POI, blocks, business circles, etc. on the base station through urban knowledge atlas. On this basis, we design a generation antagonism network model for urban knowledge enhancement, and propose a cross-city knowledge joint modeling method. The base station is used as a node to build a graph, the base station graph embedding vector is learned through GCN, and the base station graph embedding vector similarity monitoring rules are designed, Transfer urban knowledge and traffic characteristics of base stations, and realize cross-city migration of base station aggregation traffic generation method. The validity of the model is verified by using the data sets of three cities, and the model is expanded and explained by supplementary experiments.

Complex network is the point-edge system composed of some individuals with interactions. Traditional research is basically oriented to single networks, such as social networks, transportation networks, biological structure and so on. In real life, many networks are connected to each other. The change of points on one network often leads to the change of another network point, and the change is sometimes more significant than that on a single network. Here, we will use the graph neural network and the reinforcement learning method to select the important nodes in the urban interdependent network, based on Urban Simulator.

Urban public transitsystems are essential for citizens to access various resources, opportunities, and services in the city, especially for people not owning a private car. However, most public transportation networks in the United States are conceived to serve the masses and are financed by public investments, neglecting the mobility needs of the marginalized minor ethnicity populations. Here we take New York, the largest and most public transit-dependent city in the US as an example to demonstrate the severe racial inequailty of the public transit system, its implication on mobility gap, and potential cause by urban policies. Our analyses suggest among urban neighborhoods with low private car ownership, majority-white neighborhoods have significantly higher accessibility range, jobs, and facilities by public transit than the majority-black and majority-Hispanic neighborhoods. Furthermore, we associate this accessibility gap with behavioural gaps in gyration radius, employment status, and average distance take to visit urban facilities. By aligning the accessibility gap with urban policies, we discover that public housing developments have significant public transit disadvantage. These results altogether reveal a systematic path of the inequality in New York City’s public transit service.

Understanding the social fabric of urban life requires a deep understanding of urban mobility, especially how it reacts to external shocks such as Covid-19-induced health crises. However, various urban venues bear different levels of impact, making it challenging to reveal universal patterns. To uncover the mystery, we build a POI sector network based on the visitation preferences of residents and conceptualize a Well-and-Ditch model to bridge node-level mobility resilience with network centrality. We verify the conceptual model with large-scale mobility data covering multiple phases of the pandemic. Next, we reveal the source of heterogeneity by fitting a mobility preference function based on CBG mobility level and POI sector centrality. Finally, we analyze the social and epidemiological implications of such mobility changes.

Crowd flow represents the number of people entering and leaving an urban region within a given time interval. Effective crowd flow generation can benefit a wide a range of applications for public safety and transportation system. Existing works mostly depend on historical crowd flow information, which is not always available. In this work, we aim to generate crowd flow for regions without information about real flow. Diffusion models, which use a Markov chain to gradually convert pure noise into complicated data distribution, has shown great capability to model high dimensional data distributions and achieved considerable performance in many tasks such as image and audio synthesis. Inspired by this, we introduce diffusion model to crowd flow generation task, where we use an urban knowledge graph to model complex relationships between regions and leverage transformer to capture temporal flow patterns. This work is still ongoing and your valuable advice would be appreciated.

Population mobility in the city represented by OD (origin-destination) flows is of great value for many urban applications, such as urban planning and urban simulation. Existing works relay on a Unreasonable assumption, which suppose that the topology of the OD network is already known for regions of interest. When the city scale becomes large, the OD network size grows huge to make it difficult for the model to capture its whole structure. We divide the difficult large-scale OD network generation into two steps. The method first does topology generation, and then generates flows based on the topology. Diffusion model, which models the data generation process as a iterative denoise procedure from pure noise, has shown excellent performance in many fields, e.g., computer vision and nature language process. We introduce diffusion model to generate city-wide OD networks for new cities without any population mobility information

A well-known problem in deep learning is that deep neural network is difficult to train and this predicament becomes worse as the depth increases. To handle this problem, a variety of special neural network structures has been proposed, e.g. Resnet and Densenet. Besides, recently it is shown that some randomly-wired networks surprisingly show competitive performance while some others are mediocre. Currently it is not well understood how design choices on neural network architecture affect training performance. Towards this problem we studied how neural network architecture affects its trainability and performance and further provided some hypothesis to explain aforementioned phenomenon.

As an important objective beyond recommendation accuracy, diversity has attracted more and more attention in recommender systems. Despite recent algorithm advances in improving recommendation diversity and capturing user preferences accurately, how diversity influences user engagement is still underexplored. In this work, together with user preferences, we will treat diversity as another factor, and investigate their relationships with short-term (e.g., consumption time in a session) and long-term (e.g., weekly activeness) user engagement. Relevant findings are expected to provide new insight for diversified algorithms, which are usually optimized for recommendation as diverse as possible.

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 propose a specially designed NLP method to classify AI and Non-AI paper. We analyze the changes happening along with the development of AI on the paper production patterns as well as the career life of the involved researchers. We try to uncover the impact of AI on researchers and found AI tends to accelerate the process of young researcher to become leading authors, namely democratizing science participation. We hope to provide enlightenment to the future researchers through our work.

Street view data is one of the most common data sources for urban prediction tasks, such as estimating socioeconomic status, sensing physical urban changes, and identifying urban villages. Typical research in this field consists of two steps: acquiring a dataset with a street view image sampling algorithm and designing a prediction algorithm for urban prediction tasks. However, most of the previous research focuses on the prediction algorithms, leaving the sampling algorithms underexplored. To fill this gap, we set out to investigate how different street view image sampling algorithms affect the performance of the follow-up tasks and develop an effective street view image sampling algorithm for urban prediction. Through a comprehensive analysis of the performance of different sampling algorithms in three of the most common urban prediction tasks, including commercial activeness prediction, urban liveliness prediction, and urban population prediction, we provide solid empirical evidence that the sampling algorithm significantly affects the performance of the prediction model. Specifically, the performance differences of different sampling algorithms can reach over 25\%. Further, we revealed that the sampling step size and the sampling quality are two important factors that affect the performance of a sampling algorithm, while the sampling angle has little influence. Inspired by our analysis results, we propose an effective street view image sampling algorithm, DAS, which contains a denoising module and an adaptive sampling module. It can dynamically adjust the sampling step size to adapt to the optimal size for each region and get rid of the impact of noise images in the meantime. Experiments on three large-scale datasets demonstrate its superior performance over multiple state-of-the-art baselines, and further ablation study shows the effectiveness of each module. Finally, through a thorough discussion of our findings and experimental results, we provide insights into the street view image sampling algorithm design, and we call for more researches in this blank area.

Diffusion models are the generative models of interest today. Diffusion models are now showing remarkable results in various fields, including image generation, voice synthesis, etc. Diffusion models are also known to outperform other generative models, such as generative adversarial networks (GANs) and variational self-encoders (VAEs), in terms of sampling diversity and quality. Inspired by these outstanding performances, we apply the diffusion model to synthesize trajectory data. This is a work in progress and I will briefly introduce the related work and preliminary model design.

The speed of microscopic traffic simulation for cities is still an essential problem, which limits researchers from using it to do interesting research. At present, the field of computing acceleration for microscopic traffic simulation mainly focuses on multithreading and distributed computing. However, a lot of applications now shows that GPU is much faster than CPU. In order to improve the efficiency of microscopic traffic simulation with the help of GPU, we move most of the processing tasks of microscopic traffic simulation to GPU and design suitable parallelism solutions for the GPU architecture. Also, in order to solve the conflict between complex interactions between vehicles and GPU architecture, a set of indexes suitable for GPU is proposed. Based on such design and implementation, in preliminary tests, the GPU version of the microscopic traffic simulation program ran faster than 100 steps/second.

Over the years, various formulas for pedestrians have been proposed, but they are often based on human experience and limited to specific scenarios, such as the social force model and centrifugal force model. Inspired by AI for scientific discovery, we propose to automatically learn the formula of human flow through machine learning, which is significant different from manually designing formulas. Specifically, we first search the Pareto-optimal message-passing flow of GNN and then learn the symbolic model (i. e., formulas) for each message function to obtain the entire formula describing pedestrian dynamics. Our learned model can consider individual, interactive, and collective behaviors, corresponding to node-level, edge-level, and global-level variables on graphs describing pedestrian interactions, which are more accurate than existing models and with great interpretability.

Cities built environment greatly shapes citizens’ health outcomes. Previous studies point out that the walkability of a city correlates with the dementia incidence, and the severity of COVID-19 transmission is also influenced by the urban resilience. With the recent advances of deep learning and big data processing, geo-tagged imagery provides us a new perspective to assess the built environment of cities and how it affects public health. In this talk, we will introduce literature that uses street view and satellite view images in several prediction tasks, and propose a possible solution to accurately assess the COVID-19 incidence in urban spaces.

High-resolution carbon emission accounting usually requires large-scale time-consuming counting, costing high expense and resulting in release lag. To overcome these weaknesses, we leverage multi-source urban imagery as proxies to predict urban-scale grid-level carbon emission. While satellite imagery depicts the land use distribution of the city, the nighttime imagery provides information of the intensity of human activities. Combining these two sources of urban imagery, the sector and the intensity of the carbon emission can be inferred, laying a good foundation for carbon emission prediction. In today's talk, I will briefly introduce the problem formulation, the literature review, and the preliminary model design of this project.

Nowadays, natural disasters and human actions are considered uncertain, unpredictable urban threats, and it is expected that the city can respond quickly and minimize the associated consequences and dangers. Therefore, the concept of urban resilience has increasingly attracted the attention of urban researchers and authorities. This talk would like to review the origin, evolution, dimensions, and challenges of urban resilience study. It can be concluded that urban resilience covers four basic pillars: resisting, recovering, adapting, and transforming. And its measurements involve many characteristics like redundancy, robustness, adaptation, or integration. Due to the malleable definition boundaries, the evaluation methods of urban resilience lack consensus and are rudimentary, providing a main challenge and opportunity for developing future holistic studies. Finally, this talk would discuss the advantages, feasibility, and potential challenges of introducing resilience methods from complex network studies into urban studies.

Humanity’s fate is tightly linked with sustainable development, which recently is identified as seventeen sustainable development goals (SDGs), to end poverty, protect the planet, and ensure prosperity for all. Nevertheless, monitoring and achieving SDGs still face serious challenges in timeliness, completeness, efficiency, etc. To get of the dilemma, increasingly more researchers turn to artificial intelligence (AI) for SDGs built upon the advancement of sensing techniques and computational capabilities. In this talk, we will briefly review the related studies of AI for SDGs. Specifically, we will first identify several concepts related to this area, and further present an overview of AI-driven SDG applications. Finally, we will focus on typical works of using mobility, phone and imagery data for poverty and welfare, which might shed light on our future research.

With the fast development of the mobile internet, every user can create and upload content for other users to access and review, which is widely known as user-generated content (UGC). For example, on TikTok, there are millions of new videos produced on a daily basis. Although machine learning algorithms for traditional services of platform-generated content (PGC), such as personalized search or recommendation, have been well studied, they focus more on the interest and benefits of consumers (better user experience) or the platform itself (higher business profit). Yet, the benefit of content producers is neglected by algorithms, and especially for the underrepresented producers such as elder or low-income people, whether their content is assigned the same-level exposure traffic is less considered. In this work, we take on an unusual route to identify and analyze the under-researched problem and propose an easy-to-use graph neural network based method to achieve better inclusivity of underrepresented producers while keeping high accuracy.

The rising population in many big cities and the increased demands in transport have exacerbated the traffic congestion problems over the years. Commuters are experiencing longer traveling time and having difficulty in planning their journey efficiently. The considerable impact of congestion on transportation networks is reflected by the vast amount of research papers dedicated to congestion identification and prediction. However, the task of forecasting congestion duration has not been well studied. First, since traffic congestion will propagate to adjacent road segments over time, it exhibits complex spatio-temporal correlations. Second, the dynamic of traffic congestion is a continuous changing process, which is unsatisfactory to model in a widely used discrete-time manner. To address them, we propose a co-evolving graph neural ODE network for traffic congestion duration forecasting. To be specific, we capture spatio-temporal dynamics during traffic congestion by combining ODE and graph neural network. Since traffic congestion is closely related to other factors such as speed, we propose to model them as a hidden time-evolving process and integrate it to the derivation of congestion dynamics by designing a co-evolving ODE network. We plan to conduct the experiments to verify the effectiveness of our proposed model in multiple real-world datasets.

Traffic forecasting is one of the most popular spatio-temporal tasks in the field of machine learning. A prevalent approach in the field is to combine graph convolutional networks and recurrent neural networks for the spatio-temporal processing(STGNNs). However, the existing approaches focus on the tasks of predicting traffic volume or traffic speed, which does not directly address the challenges on traffic congestion prediction. Previous works focused on microscopic prediction which aims at guessing the next influenced road, leaving macroscopic information unused. To the best of our knowledge, no previous works have suggested a unified model for both microscopic and macroscopic scales. In this paper, we propose a novel multi-scale traffic congestion prediction model based on a simple contagion process. On the macro scale, we use two macroscopic characteristics for network traffic dynamics, namely congestion propagation rate β and congestion dissipation rate μ. We describe the dynamics of congestion spread using these new parameters embedded within a system of ordinary differential equations, similar to the well-known susceptible-infected-recovered (SIR) model. On the micro scale, the simple contagion model provides physics information for short-term time series input to the conventional STGNNs. Experiments on three real-world datasets demonstrate that our framework is capable of significantly enhancing STGNNs.

Order combination is a vital optimization problem in industrial production, and a reasonable order combination method can significantly improve the utilization of raw materials and reduce economic costs. However, there are many difficulties in finding an optimal order combination method quickly in the production process. First, the production specification constraints are diverse, and the order combination results must satisfy all rules. In addition, the order combination is an NP-hard problem, and it is difficult to find the optimal solution in a limited time. To solve the above difficulties, we model the order combination problem as a nonlinear multi- knapsack problem and use a combination of heuristics and neural networks to solve the problem. The work is in progress, and a solution based on the heuristic approach is available, and subsequent work will combine the deep learning approach to optimize the solution further.

With the improvement of people's demand for network quality, the sixth-generation (6G) aims to realize the physical entity in the real environment Digitization and Intelligence and improve the quality of information and communication services, satisfying humans fully immersive interactive scenes. However, the growing complexity of next-generation networks exacerbates the modeling and algorithmic flaws of conventional network optimization methodology. To this aim, this paper proposes a simulation system focusing on modelling essential network elements (e.g., base stations and users) and complicated interactions between these entities. Moreover, reinforcement learning could interactively optimize diverse real-world deployments, channels and power consumption with the feedback from the simulation engine, which overcomes intractable NP-hard optimization question. To show the efficiency of proposed system, we build an application case about base station coverage area to present the optimization process and performance.

Spatio-temporal point processes (STPPs) are the dominant paradigm for modeling sequences of spatio-temporal events happening at irregular intervals. The standard way of learning in such models is by decomposing the conditional intensity function into the ground intensity and conditional density, and then optimizing it with maximum likelihood estimation. However, parameterizing the intensity function or directly the density function usually incurs several trade-offs and is also limited by certain-structure approximators. In this work, to overcome existing limitations, we propose a new parametrization for STPPs which leverages diffusion probabilistic models and enables flexible, high-fidelity modeling of discrete events that are localized in continuous time and space.

With the rapid development of wireless communication technologies, we have experienced the advent of 5G commercialization with services based on the principle of wireless transmission. To better understand how the electromagnetic signal is characterized by the wireless environment, it is inevitable to study the way to model the channel. The ultimate purpose for channel modeling is to achieve both high accuracy and efficiency, to support the system design and the performance evaluation. The conventional methods for channel modeling are classified into two categories, the stochastic model and the deterministic model. However, they have the shortcomings of low fidelity and high computational time consumption, respectively. Recently, the approaches based on machine learning reveal the feasibility of doing channel modeling in a data-driven way. In this work, we attempt to jointly consider the physical knowledge and the measured data. Specifically, we involve the frequency domain Maxwell’s equations as a physical constraint in the loss function, and utilize the neural network to construct the distribution of the electromagnetic field accurately and efficiently. By comparing the electromagnetic energy with the signal source, we are able to derive the path loss. This is an ongoing work and suggestions and comments are appreciated.

Sharing parameters in multi-agent spatio-temporal deep reinforcement learning has played an essential role in allowing algorithms to scale to a large number of agents. Parameter sharing between agents significantly decreases the number of trainable parameters, shortening training times to tractable levels, and has been linked to more efficient learning. However, having all agents share the same parameters can also have a detrimental effect on learning. The existing work only considers the fixed parameter sharing and does not consider the update of the agent strategy, which will damage the final optimization performance. To solve this problem, a hierarchical reinforcement learning framework was introduced to give agents the ability to automatically identify agents that may benefit from parameter sharing and dynamically share parameters them during learning. For the upper level agent, it first uses CVAE to extract the agent's spatiotemporal information from the agent trajectories as the basis for agent parameter sharing and outputs which agents in the lower level should share parameters. For the lower level agents, they will output specific actions to interact with the environment and feed back rewards to the upper level agent. Experiments done in ride-hailing platform show that the proposed method not only can improve the efficiency of parameter sharing, but also ensure the quality of policy learning in multi-agent spatio-temporal system.

Couriers’ salary for their labor in the last-mile delivery service accounts for 70% of the cost in JD logistics. In order to reduce cost and increase efficiency, we propose to study the problem of courier’s working process recovery. This enables the scientific management of couriers and the last-mile delivery task by inferring the specific behaviors during couriers’ working process, so that manager can know exactly what are they doing and can analyze their working metrics in quantity. There are several related works on last-mile delivery data mining, however, we are the first to focus on courier’s whole working process and the fine-grained behavior semantics behind the data. I will review some related works, draw our contributions, formulate the problem of courier’s working process recovery, analyze the main challenges and depict our intended methods on overcoming the spatial-temporal issues in data and considering the correlation between behaviors.

Nowadays with the continuous improvement of computing power, modeling and simulating the mechanisms of the complex urban system are increasingly in the spotlight. In the past year, we designed and implemented an efficient and extensible urban simulation system that simulates the evolution and interdependence of different city network layers including traffic, power, water, and communication. The system is designed to own strong scalability in both simulation functions and layers, and it provides researchers with tools to easily control the simulation process and get the results. This talk will show you the basic framework of the urban simulation system, our ongoing engineering and research work based on it, and some future works about perfecting the system architecture.

Directed acyclic graphs (DAGs) have broad applications to understand numerous real-world complex systems across many disciplines, ranging from evolutionary biology to social science. A common property of many evolving DAGs is that the node degrees are strongly correlated. Despite the importance of correlated DAGs, the theoretical framework to account for its evolution is missing. Here we develop a generic framework for the evolving correlated DAGs characterized by a stochastic point process. A minimal model based on the theoretical framework reproduces the critical statistical characteristics of real-world citation graphs, known as a typical DAG. The model not only analytically predicts the empirically observed assortativity, but also allows us to capture other important properties of real-world evolving DAGs.

With the growing demand for mobile networks in recent years, it is important to allocate the available wireless network resources more efficiently. In this talk, we consider the following resource allocation scenario: allocating wireless network resources to users in a certain region at a certain time period to maximize the overall throughput of the network while ensuring that the demands of each user are met. We decompose the resource allocation problem into multiple constrained optimization problems and solve them using safe reinforcement learning. Compared with traditional numerical optimization methods, deep reinforcement learning has no specific requirements for the form of optimization objectives, and the learned policy networks have certain generalization capabilities. In a real-world base station distribution scenario, the algorithm can maximize throughput while meeting user demands.

Socioeconomic indicators, such as population density and commercial activeness play an important role in Sustainable Development Goal (SDG) monitoring of poverty, health, quality, etc. including However, traditional collection ways like in-person survey suffer issues from timeliness, resolution as well as cost. Moreover, such statistical activities have been cut back in many countries recently due to COVID-19, particularly those that need them most. Despite the challenges, easily accessible urban imagery (including satellite imagery and street view imagery) becomes an important tool for socioeconomic indicator prediction in recent studies. Most existing methods extract visual features from a certain type of urban imagery with either geographic information or points of interest information considered, which however fail to fuse various types of urban imagery nor cover rich semantic information in urban environment. Therefore, in this study, we leverage urban knowledge graph to capture enrich semantic information in urban environment, which further guides the fusion of two types of urban imagery in a contrastive learning framework. Experiments on three cities across various socioeconomic indicators validate the effectiveness of the proposed model.

In the era of big data, fully releasing the value of data elements and playing the supporting key role of data in the development of digital economy is the inevitable law of development in the era of digital society. Currently, there is a strong demand for data applications, data gradually penetrate into various fields, and various data applications develop rapidly, but the effective supply of data is insufficient, and the support for municipal governance, economic development and livelihood services needs to be further strengthened. For this reason, a complete market-based trading platform is needed to be formed within the scope of tradable data elements, which are interconnected and acted by supply and demand, price, competition, risk and other elements. In this paper, we focus on data elements as the background, and conduct transformative research on network architecture, core mechanisms, and circulation protocols for the efficient delivery and reliable circulation of data elements. The research content includes the following four points: the concept of data components and circulation requirements, the general design and concrete implementation of the network using CDN and blockchain.

Over the years, various conversational recommendation methods have been proposed, but they often treat the conversational task as two separate parts, which are the conversational part and the recommendation part. The conversational model is responsible for generating a response for users, and recommendation model will also generate an item for users. However, such kind of paradigm is hard to unify two kinds of information, failing to generate good responses. Different from the traditional way to treat two parts separately, we propose to unify two parts as a whole model. Specifically, we first use recommendation data and natural language data together to improve the generalization of word embedding. Then, based on T5, our model using different prompt to train the various data in one paradigm. In this way, our learned model can directly generate items within a response.

Mobile edge computing (MEC) has been envisioned as a promising technology for enhancing the computational capacities, by enabling computational task offloading. We employ massive multiple-input multiple-output methods (MIMO) to facilitate base station task offloading in small cell networks (SCNs) with multiple mobile users. The SCN BS can relief the computation workload by utilizing the idle resources of network devices. However, the feasible offloading choices vary over the time due to mobility driven changes. Moreover, intercell interference (ICI) can significantly impact the transmission latency in SCNs. One way to control the ICI in the vertical plane is through exploiting the vertical angle of the BS antenna pattern, commonly referred to as tilt. By appropriately selecting the tilt, it is possible to increase the desired signal level at an intended user. With aim to minimize the offloading latency, we jointly investigate the impact of the elevation angle of the BS antenna pattern (tilt) and downlink power allocation in a conventional uncoordinated transmission mode.

Caching popular contents at the network edge has been considered as a promising enabler to relieve the pressure on networks due to the fact that a substantial portion of global data traffic is repeatedly requested by many subscribers and thus redundantly generated. Recommendation, on the other hand, has attracted spiraling attention for its capability of reshaping users’ contents demand patterns. The practicability of recommendation in boosting the gains of edge caching with uncharted users’ feature information. It is essential to characterize the average system cost for a generic network model, disclosing its dependence on the recommendation and caching strategies. To develop a joint caching and recommendation system it is essential to take into account constraints of each content provider’s cache capacity budget, each individual user’s recommendation size and recommendation quality. However, the implicit information regarding users’ preference makes the problem inextricable. To address this issue, a versatile long short term memory (LSTM) network assisted prediction paradigm is required to attain the preference schema of users with the assistance of their historical behavior data. This approach has the potential to significantly improve users' data viewing experience by optimizing data caching through recommendation system.

Diversity and the corresponding equality and inequality induced by different types of diversity have been a heatedly discussed topic over the years and are identified as crucial for not only scientists but also policymakers and funders. Among these dimensions of diversity is interdisciplinarity, which has been identified as essential for addressing the increasingly complex key societal and scientific challenges contemporarily. In this talk, I will provide an overview of the related field and discuss some of our preliminary discoveries.

Urban public transportation systems are essential for citizens to access various resources, opportunities, and services in the city, especially for people not owning a private car. However, most public transportation networks in the United States are conceived to serve the masses and are financed by public investments, neglecting the mobility needs of the marginalized poor and minor ethnicity populations. Here we use travel time data and large-scale mobility data in three large US cities to depict how marginalized neighborhoods are struggling in the trap of the public transportation system. Our results suggest among urban neighborhoods with low private car ownership, those with lower income, higher poverty rates, lower shares of white residents, and higher shares of black residents experience significantly longer travel time by public transportation than their counterparts. Further analyses show that this inequality in travel time is associated with the inequality in accessing job opportunities and enjoyment-centric consumption venues.

Although recommender systems have been around for a while, their utility has become more expansive and their necessity more widespread in recent years, especially in popular social applications like Tiktok (Douyin) and Kwai (Kuaishou). With the elevated importance of content recommenders in these apps, it is more pressing than ever that the recommender systems capture user preferences as closely and in a fine-grained manner as possible, while also preventing the occurrence of negative effects such as filter bubble and addictive behavior. Meanwhile, recommender models enhanced with cognitive information such as user personality are of increasing interest to researchers, but they have not yet been applied to the short video domain. To bridge this gap, we present a user personality-aware graph based model to increase diversity of short-video recommendations while maintaining high performance. Upon preliminary investigation, our proposed model shows potential in achieving the intended goals.

Population mobility in the city represented by OD (origin-destination) flows is of great value for many urban applications, such as urban planning and urban simulation. Existing works focus on the region-pair wise OD flow but didn’t consider the whole structure of the population mobility in the city. We propose to generate the OD flows between every two region as a mobility network to capture the topological features of population mobility on one-shot. Diffusion model, which models the data generation process as a iterative denoise procedure from pure noise, has shown excellent performance in many fields, e.g., computer vision and nature language process. We introduce diffusion model to generate city-wide OD networks for new cities without any population mobility information. The project just starts and the method is in its infancy. Your valuable suggestions are expected.

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: unevenness 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. We then extend our framework to more general and classical network structures, e.g. Resnet. We hope that the findings could offer some beneficial insights for designing more effective neural network-based models in artificial intelligence.

The real world often requires large amounts of mobility data to support applications, and real trajectories are limited by user privacy or device acquisition lead to problems with both quality and quantity. So the importance of trajectory generation is self-evident. However, existing trajectory generation models ignore user-level modeling and lack universality in the temporal modeling of trajectory points that can take into account different travel modes. In this paper, we propose variational human trajectories generator which is a novel probabilistic model based on the variational temporal point process to synthesize human trajectories. The model can generate user-level features, obtain accurate simulations by decoupling travel time and dwell time, and also generate continuous-time distributed, variable-length trajectories. Our model has been extensively experimented on two real datasets, and the results show superiority over state-of-the-art baselines, demonstrating the effectiveness of our model.

Fifth-generation (5G) telecommunications would alter contemporary lifestyles; however, 5G networks require a large number of base stations, resulting in increased carbon emissions. China has deployed over 800,000 5G base stations since 2021. Here, we demonstrate that 5G deployment will substantially increase carbon emissions. We develop a theoretical model to quantify mobile network carbon emissions. We discovered that after the deployment of 5G, the overall network carbon efficiency decreases significantly, resulting in the potential emission of an additional 7 megatons of CO2. By implementing sleep control mechanisms on base stations to reduce the misalignment of information flow and energy flow in space and time, a 70% reduction is possible. Moreover, we illustrate a feasible path to carbon neutrality for 5G Mobile Networks and conduct techno-economic analyses. This study provides a clear method for assessing the environmental impact of 5G base stations and will assist mobile networks in achieving their sustainable development objectives.

Over the past four decades, rapid urbanization around the world has brought significant amount of people to cities, the most important settlement in human history. The concentration of resources and talent greatly influenced the development of human society, yet significant problems are also emerging in over-aggregated, congested cities. Non-communicable diseases and rising health inequalities hurt the foundation of health and wellbeing in cities, which are responsible for 77.3% of all-cause years of life lost in Chinese cities in 2015. Here, we systematically collect a comprehensive dataset that captures important environmental determinants in 29 England cities to facilitate further understand of how city environment affects citizens’ health. Specifically, we summarize vision information (fine-grained street view images and satellite images), environment information (air quality and POI distribution), along with microscopic physical & mental health status from prescription data. Large-scale, comprehensive dataset of environmental determinants in cities are important missing pieces to understand the cause of non-communicable diseases in urban spaces, which may shed light on the design of healthy cities.

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 the paper production patterns as well as the career life of the involved researchers. We try to uncover the impact of AI on researchers and provide enlightenment on the future research works.

Understanding the intrinsic mechanism of pedestrian interaction is of great scientific value and application value, which can benefit public safety, architectural design, and epidemic control, etc. Over the years, various formulas for pedestrians have been proposed, but they are often based on human experience and limited to specific scenarios, such as the social force model and centrifugal force model. Different from the traditional way to extract formulas by expert knowledge, we propose to automatically learn the formula of human flow through machine learning and analyze the similarities and differences of the formula of human flow in different scenarios. Specifically, we first search the Pareto-optimal message-passing flow of GNN and then learn the symbolic model (i. e., formulas) for each message function in a human-machine teaming way. Extensive experimental results demonstrate that our model can learn diverse formulas with explicit physical meanings in diverse scenarios that is more accurate than existing expert-designed formulas.

Recent decades have witnessed intensive investigations in genetic algorithms (GAs) that have been adopted to solve vehicle routing problems (VRPs). GAs are generally regarded as nondeterministic problem solvers inspired by natural selection and genetic variation. They could find good-quality solutions within a reasonable computational time. However, those algorithms may be less efficient because they are designed manually without intelligent parameter configuration to support various evolutionary processes. In this work, we focus on the most influential process in GAs, i.e. crossover, and conduct a preliminary attempt on incorporating GAs with reinforcement learning. The automatic gene locus selection for crossover is formulated as a Markov Decision Process (MDP). Specifically, dual-aspect collaborative transformer (DACT) and proximal policy optimization (PPO) have been utilized to intelligently select the appropriate gene locus for the Order Crossover operator, which leads to an efficient search and improves solution quality and convergence rate. Experiments are conducted to validate the effectiveness and efficiency of the proposed Neural-Genetic algorithm.

Transportation is one of the major sources of carbon emission in cities. In recent years, many new technologies have been applied to vehicles, such as autonomous driving and new energy. Among these new vehicle technologies, there have been numerous studies discussing the carbon reduction benefits generated by electric vehicles at different scopes and from different perspectives. However, there seems to be a lack of research work on the carbon reductions that autonomous driving strategies can produce at the urban scale. Therefore, we are going to use our microscopic traffic simulator to simulate three autonomous driving strategies (eco-driving, convoy driving, global route scheduling) and then analyze the impact of these strategies and their penetration rate on urban carbon emissions.

In the logistics system, each delivery station has jurisdiction over several road districts. Several couriers in this delivery station will serve road districts to complete the pickup and delivery of express tasks. Generally, the terminal-arborization of the logistics system refer to the interaction stage between couriers and customers, which is also a reflection of the quality of logistics services. At present, the management and optimization of logistics terminals are still insufficient, and there are problems such as insufficient service efficiency. With the deployment of the logistics database, designing a logistics terminal simulation system can help the management to understand the terminal service quality and optimize the terminal operation mode. In the simulation of the logistics terminal scene, we first analyze the real data to understand the real working mode of the courier. Furthermore, based on the agent-based modeling method, we design a hierarchical framework to simulate the working status of couriers. Finally, we conducted a courier simulation in a real delivery station and compared it with the real courier working status.

In a category of multi-agent systems such as ride-sharing system, individual agent pursues its own goal and interacts with each other in its neighbors. The complex cooperative or competitive behaviors vary with agents in different neighbors. In such system remains two challenges. One is how to define an appropriate cooperation scope of agents. The cooperation scope of agents is critical to performance but manually-designed scope is time-consuming and hard to scale. The other one is how to deal with the complex relationship of agents in different neighbors. In this work, we develop a novel MARL method, which is a bi-level framework to automatically learn the cooperation scope and the cooperation pattern of neighbors. We conduct experiments in large-scale ride-sharing platform, which incorporate hundreds of agents.

User-Generated Content (UGC) is at the core of online short video platforms where users can both produce and consume content. This differs from traditional e-Commerce domains where content producers and consumers are usually from two separate groups. In this work, we propose a GNN based method for recommending content on UGC-based platforms. Specifically, we learn a core embedding for each user, and ultilize multi-view and attention technique to project the user’s core embedding into two ‘role’ embeddings (i.e., a producer and consumer role). We design propagation mechanisms on User-Item, User-Producer and Follows graphs to further learn representations for users and items. Empirical studies on two large-scale UGC applications show that our method outperforms standard collaborative filtering methods as well as recent graph-based recommendation methods and UGC recommendation methods.

To understand the social fabric of urban life requires a deep understanding of urban mobility, especially how it changes in reaction to external shocks such as pandemics, storms, and wildfires. However, various urban venues bear different levels of impact, making it challenging to reveal the universal patterns. To uncover the mystery, we build a venue sector network based on the visitation preferences of residents and conceptualize a Well-and-Ditch model to bridge node-level mobility resilience with network centrality. Taking COVID-19 as an example, we verify the conceptual model with large-scale mobility data covering multiple phases of the pandemic, and also reveal the changes in network-level robustness. Understandings obtained from this work will be applied to improve the classic Explore-and-Preferential-Return (EPR) model in capturing mobility changes during crises.

From critical infrastructure to physiology and urban systems, complex systems rarely occur in isolation. Instead, the functioning of nodes in one system often promotes or suppresses the functioning of nodes in another and even alters the topology of networks in systems. At the same time, external shocks, such as the global pandemic and the war, also pose great challenges to the resilience of urban systems. For example, governments have discussed policies to improve the resilience of supply chains by reducing dependence on foreign suppliers to counteract the adverse effects of shocks on the economy. In this talk, I will review relevant papers and offer some potential directions to understand the resilience and recovery of some urban systems under external shocks through the lens of convolving multiplex networks.

Social media users face a tension between presenting themselves in an idealized or authentic way. Here, we explore how prioritizing one over the other impacts users’ well-being. We estimate the degree of self-idealized vs. authentic self-expression as the proximity between a user’s self-reported personality and the automated personality judgments made on the basis Facebook Likes and status updates. Analyzing data of 10,560 Facebook users, we find that individuals who are more authentic in their self-expression also report greater Life Satisfaction. This effect appears consistent across different personality profiles, countering the proposition that individuals with socially desirable personalities benefit from authentic self-expression more than others. We extend this finding in a pre-registered, longitudinal experiment, demonstrating the causal relationship between authentic posting and positive affect and mood on a within-person level. Our findings suggest that the extent to which social media use is related to well-being depends on how individuals use it.

In the past few years, logistics services grows rapidly with the fast development of e-commerce. A courier is asked to record the delivery time of the parcel after delivering it, and each delivery task will be accounted on time if the courier finish the delivery before required finish time. Therefore, delivery on time 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 whether or not to send more men to delivery stations with low predicted delivery on time rate. In this work, we first investigate factors that will affect on time rate, and then reveal the spatial-temporal correlations between blocks. We first develop a transformers architecture model, which encode the features of each block. Then we provide a static graph derived from the real world road map and develop our model based on graph nearual network.Finally, we conduct some experiments on real world data to validate the effectiveness of the proposed models.

Economic inequality, emphasized by the Sustainable Development Goals, is difficult to quantify as reliable data on household incomes or economic status are missing for most of the world. Although researchers have developed a proxy for measuring the economic inequality with the aid of remotely sensed nighttime light (NTL) data, it is acquired on a relatively low spatial resolution and rarely holds information about the ground objects, which makes the proxy difficult to interpret. To bridge the gap, in the study, we utilize an interpretable computational framework to predict economic inequality at the local level by applying object detections on high-resolution day-time satellite imagery. Further, we study the effectiveness of the dataset in consecutive years, which compares favorably against the NTL proxy.

The novel coronavirus pandemic hits the world in late 2019, and people around the world have been more or less affected by the pandemic in countless ways. It also happens in the academic world. This pandemic is one of the few incidents in history that can cause impact to almost all areas of research. Ever since the outbreak of Covid-19, research on this pandemic has been one of the most popular topics in numerous disciplines. Undoubtedly, we can feel the influence it brings to the academic world. Covid has gathered researchers of different disciplines, working together to produce novel ideas. No doubt covid has influenced the academic world, but to our best knowledge, questions of A. How impactful it is? B. In what aspect does it affect? C. What features do those impacts have, compared to other topics? remain widely unexplored. In this special era, it would be very critical to answer: how did the pandemic affect interdisciplinary research, so that researchers of different disciplines can better understand Covid-19-related research? This ongoing work aims to answer the above questions by examining the features of Covid-19, as an academic topic, in 2 aspects: A. Interdisciplinary Collaborations B. Impact Factors of Interdisciplinary Publications. Then concludes its impact done to the development of interdisciplinary research by comparing it with its counterparts of identical nature in both the short and long term.

Accurate user interest modeling is important for today's streaming recommendation. One of intuitive solutions is sequential recommendation which relays on click behaviors. However, there is no active click behavior in streaming recommendation. Instead, users are passive to receive the streaming contents and return skip or no skip behaviors, which can be treat as negative or positive feedbacks. Besides, skip and no skip are not simply positive or negative correlated. Thus it is challenging for us to capture the contrary feedbacks in one sequence. To capture the sequential pattern of heterogeneous feedbacks, FeedRec has exploited shared vanilla Transformer and grouped each feedback into different Transformers. Indeed, such task may be challenging for vanilla Transformer as there are contrary relationships of different feedbacks. To further capture the contrary relationships in streaming recommendation, we propose Factorization-heads Attention with Bayesian Contrastive Learning based on feedback-aware mask and pair-wise loss. Specifically, we first suppose each head can capture specific feedback relationship, and improve multi-heads attention by talking-heads attention to break the barrier between different heads (feedbacks). Then we further propose factorization-heads attention to decompose the relationships between different heads, before exploiting a feedback-aware mask to inject more feedback information into the head interaction. Finally, we propose Bayesian Contrastive Learning to attract the same feedback and repel the contrary feedback by feeding contrary feedbacks into positive and negative towers and perform BPR loss on them to capture the pair-wise relationship. Experiment results on two real-world datasets show the superiority of our proposed method against state-of-the-art baselines. Further ablation study and visualization also confirms its effectiveness.

With the increasing adoption of GPS modules, there are a wide range of urban applications based on trajectory data analysis, such as vehicle navigation, travel time estimation, and driver behavior analysis. The effectiveness of urban applications relies greatly on the high sampling rates of trajectories precisely matched to the map. However, a large number of trajectories are collected under a low sampling rate in real-world practice, due to certain communication loss and energy constraints. To enhance the trajectory data and support the urban applications more effectively, many trajectory recovery methods are proposed to infer the trajectories in free space. In addition, the recovered trajectory still needs to be mapped to the road network, before it can be used in the applications. However, the two-stage pipeline, which first infers high-sampling-rate trajectories and then performs the map matching, is inaccurate and inefficient. In this paper, they propose a Map-constrained Trajectory Recovery framework, MTrajRec, to recover the fine-grained points in trajectories and map match them on the road network in an end-to-end manner. MTrajRec implements a multi-task sequenceto- sequence learning architecture to predict road segment and moving ratio simultaneously. Constraint mask, attention mechanism, and attribute module are proposed to overcome the limits of coarse grid representation and improve the performance. Extensive experiments based on large-scale real-world trajectory data confirm the effectiveness and efficiency of their approach.

With the pledge of global carbon neutrality by 2050, carbon issues are gaining increasing attention worldwide. The transportation sector accounts for a large share of this, taking up about 29% of the total emission. Therefore, it is of great significance to understand transportation-oriented emissions and develop countermeasures. However, despite extensive research on carbon emission accounting, they fail to precisely calculate high-resolution and real-time transportation carbon emissions. Also, the existing methods can hardly accurately estimate the effect of the carbon reduction policies. Therefore, in this sense, we propose a mobility-based urban transportation emission simulation framework that can solve the above two dilemmas. Specifically, we first introduce massive, passively-collected mobile phone data to recover the individual mobilities of the city. Pairing mobilities with vehicles, we input them into Urban Simulator (or Car Following Model) to obtain the instantaneous speed and acceleration of the vehicles with which we can calculate the instantaneous carbon emission. This way, transportation emissions with high spatial and temporal resolution can be obtained. By restricting vehicle accessibility and other aspects, we can calculate the corresponding carbon emission under these restrictions. We can then precisely assess the emission reduction policies and adopt the most effective policy in urban governance.

Recommender systems are prone to be misled by biases in the data. Models trained with biased data fail to capture the real interests of users, thus it is critical to alleviate the impact of bias to achieve unbiased recommendation. In this work, we focus on an essential bias in micro-video recommendation, duration bias. Specifically, existing micro-video recommender systems usually consider watch time as the most critical metric, which measures how long a user watches a video. Since videos with longer duration tend to have longer watch time, there exists a kind of duration bias, making longer videos tend to be recommended more against short videos. In this paper, we empirically show that commonly-used metrics are vulnerable to duration bias, making them NOT suitable for evaluating micro-video recommendation. To address it, we further propose an unbiased evaluation metric, called WTG (short for Watch Time Gain). Empirical results reveal that WTG can alleviate duration bias and better measure recommendation performance. Moreover, we design a simple yet effective model named DVR (short for Debiased Video Recommendation) that can provide unbiased recommendation of micro-videos with varying duration, and learn unbiased user preferences via adversarial learning. Extensive experiments based on two real-world datasets demonstrate that DVR successfully eliminates duration bias and significantly improves recommendation performance with over 30% relative progress.

Effective urban region embeddings can reveal characteristics of regions and help downstream tasks such as predicting population and crime. Although many existing works leverage data from various sources to better characterize urban regions, they fail to organize heterogeneous data in a comprehensive way. Inspired by the capability of knowledge graph in modeling complex relationship, in this work, we propose an urban region embedding model based on KG.

The COVID-19 outbreak has greatly impacted the economy, particularly on consumption behaviors. Revealing how consumption responses to COVID-19 can be a powerful aid for urban consumption recovery and assessing the economic situation. In this paper, we examine how COVID-19 affected citizens’ consumption behaviors by revealing recession-recovery patterns over the pandemic period. Specifically, we first explore the recession-recovery pattern of national consumption and the effects of various policies during the period. Then, recession-recovery patterns across cities are widely studied. We also investigate how consumption structures change during the pandemic and explore the association between change patterns in citizens’ consumption and cities’ socioeconomic characteristics. In general, national consumption represents a V-shaped pattern during the pandemic, experiencing a dramatic decline and a rapid rebound. Consumption is significantly inhibited by lockdown while it is stimulated positively but gradually by easing policies. Consumption patterns at the city level are associated with socioeconomic characteristics. Cities with high-income groups experience a greater decline, and cities with a high share of the secondary sector have a higher recovery rate in consumption. The consumption structure redistributes but does not fundamentally change. During the recession and early recovery phase, consumption related to basic living saw a significant rise, whereas leisure-related consumption dropped dramatically and recovered slowly. This study can assist policymakers in implementing diversified market provisions and targeted lockdown policy adjustments for consumption recovery in cities with different socioeconomic backgrounds.

Dynamic link prediction is essential for a wide range of domains, including social networks, bioinformatics, and knowledge bases. Existing works either focus on modeling structural information or temporal information while neglecting the complex relationship between them. In this paper, we propose to model the inherent correlations among the evolving dynamics of different structural scales for dynamic link prediction. Following this idea, we propose an Attentional Multi-scale Co-evolving Network (AMCNet). Specifically, We model multi-scale structural information by a motif-based graph neural network with multi-scale pooling. Then, we design a hierarchical attention-based sequence-to-sequence model for learning the complex correlations among the evolution dynamics of different structural scales. Extensive experiments on four real-world datasets with different characteristics demonstrate that AMCNet significantly outperforms the state-of-the-art in both single-step and multi-step dynamic link prediction tasks.

The time-dependent vehicle routing problem (TDVRP) is formulated based on the original VRP, where the travel cost between nodes are time-varying, defined by several cost matrices. The goal is to find the routing strategy to visit all nodes spatially with a minimum cost calculated via these matrices. Generating high-quality TDVRP solutions is of great significance in real-world applications, such as logistic systems. Even though deep reinforcement learning (DRL) has shown effectiveness in solving standard VRPs in recent years, directly utilizing it to solve TDVRPs still faces two major challenges: a) how to handle the heterogeneity between nodewise features and pair-wise cost matrices as raw feature inputs of a given problem, and b) how to capture the relationship and dependency along both spatial and temporal dimensions of these features. In this paper, we propose a novel end-to-end DRL based framework named spatial-temporal transformer (S2T). Within S2T, a heterogeneous feature embedding module is developed to process raw feature inputs into unified embeddings, while a context encoder based on a spatial-temporal attention structure further encodes the embeddings from both spatial and temporal dimensions in a joint perspective. We conduct extensive experiments on both synthetic and real-world datasets. Results show that S2T can generate solutions with significantly faster speed than previous TDVRP methods with comparable solution quality.

The last-mile pickup-and-delivery problem is a core bottleneck in modern logistic systems, which includes a bi-level problem structure. In the upper level, packages are transported by shuttle trucks between the central depot and contact stations. In the lower level, couriers inside each sub-region are responsible for picking up packages from customers to contact stations and delivering packages from contact stations to customers. The goal is to find the optimal routes for couriers and shuttle trucks that minimize the total distance of both while satisfying capacity and time window constraints. Besides the challenges of large problem scale and complex constraints, the bi-level structure adds an additional challenge of the coordination of the upper and lower levels, which cannot be solved by existing methods. To tackle these challenges, We propose a learning-based iterative method with deep reinforcement learning (DRL) to solve this problem hierarchically. First, the DRL model is employed to implicitly learn from rewards how to coordinate the two levels, and find the best assignment of customers to contact stations. Then, the reduced sub-problem of finding routes for the couriers and shuttle trucks can be solved using heuristic algorithms. Extensive experiments demonstrate that our method significantly reduces the cost of solutions compared to the baseline methods.