The movement of individuals is influenced by various factors, which can impact the collective behavior of crowds, making it difficult to predict. Knowledge-driven approaches, such as crowd dynamics, provide fundamental rules and models with good interpretability and generalizability across different scenarios, but they cannot capture the complexity and randomness of data. Deep learning methods have shown good performance in data fitting, but their generalization ability may suffer from extrapolation or observation bias, leading to physically inconsistent or unreliable predictions. Our research aims to combine crowd dynamics and deep learning for micro-level pedestrian simulation. Firstly, we generate crowd trajectories based on a diffusion probability model, where the model takes the current movement data as input and denoises the acceleration distribution of the next timestep from Gaussian noise. Secondly, we embed social force knowledge as conditional information into the probabilistic model and decompose the acceleration into repulsive and attractive forces, allowing the model to focus on learning the dynamic interaction between pedestrians and the environment. Finally, we propose a new diffusion model training mechanism for long-term simulation. Further more, I will briefly introduce our future exploration of incorporating geometric equivariance in the current task.

How to use reinforcement learning technology to optimize wireless network coverage is a topic that we are very interested in. In this seminar, two works on wireless network coverage optimization using multi-agent reinforcement learning will be shared. Remote Electrical Tilt (RET) optimization is an efficient method for adjusting the vertical tilt angle of Base Stations (BSs) antennas in order to optimize Key Performance Indicators (KPIs) of the network. The first work [1] demonstrates how to use coordination graphs and reinforcement learning for RET optimization in a complex application involving hundreds of cooperating agents. The work shows how mobile networks can be modeled using coordination graphs and how network optimization problems can be solved efficiently using multiagent reinforcement learning.The second work [2] focuses Unmanned Aerial Vehicles-BaseStations (UAV-BSs). UAV-BSs can provide communication services for ground devices (GDs) in various scenarios, such as emergency communication and traffic offloading in hotspots. The work characterizes the fairness at user-level based on proportional fairness scheduling and formulate a weighted-throughput maximization problem via designing UAV-BSs’ trajectory. The approach adopts the framework of centralized training with distributed execution. Simulation results show that the method can improve fairness of communication service by sacrificing a small amount of throughput. [1] Bouton M, Farooq H, Forgeat J, et al. Coordinated reinforcement learning for optimizing mobile networks[J]. arXiv preprint arXiv:2109.15175, 2021. [2] Z. Qin, Z. Liu, G. Han, C. Lin, L. Guo, and L. Xie, "Distributed UAV-BSs Trajectory Optimization for User-Level Fair Communication Service With Multi-Agent Deep Reinforcement Learning," IEEE Transactions on Vehicular Technology, vol. 70, no. 12, pp. 12290-12299, Dec. 2021

The widespread spread of COVID-19 and H1N1 has had a huge impact on public health. To accurately predict the spread of infectious diseases using machine learning models, we proposed a data-driven macro-micro collaborative model to predict disease dynamics. We first simulate the dynamics of disease transmission based on a human mobility dataset. Secondly, the macro graph and micro graph were generated. Then, the prediction of macro and micro graph collaboration is realized, and the epidemic dynamics in the future prediction window can be realized on the granularity of the epidemic dynamic prediction region according to the limited history window. Our innovation is not limited to separate macro or micro data, but to use a macro-micro collaborative approach to capture the changes of the micro graph under different dynamics of the disease by considering the dynamics of the user granularity along with the regional granularity prediction.

Base station energy saving aims to reduce total energy consumption by optimizing and adjusting the operation of base station equipments. This paper considers how to adjust the control statuses of base stations and cells to meet predicted traffic demands and simultaneously minimize energy consumption. First, this paper analysis the relationship between the energy consumption of different devices in the base station and their traffic load, and then set up a compatible energy consumption model. Then, this paper decomposes the optimization problem and proposes a reinforcement learning method to adjust the cell control statuses, and then adapt the base station control statuses and traffic load conditions to near-optimal based on the cell control statuses.Then, this paper decomposes the optimization problem and proposes a reinforcement learning method to adjust the cell control statuses, and then based on them adapt the base station control statuses and traffic load conditions to near-optimal. Finally, the algorithm is evaluated under realistic base station distribution and traffic demand distribution. The experimental results show that the algorithm can save about 55% of energy consumption compared to the original control scheme and has a significant improvement compared to other baselines.

Meituan is the Groupon e-commerce site in China. It helps consumers find the most trustworthy businesses and low discount products. We introduce two recommendation tasks in Meituan: 1) knowledge graph based recommendation and 2) recommendation on biased data. The knowledge graph is a natural and general representation of data in modern large commercial recommender systems which involve heterogeneous types of data. Knowledge graph based recommenders faces two problems: how to represent the semantic relations of knowledge graph and how to fuse the knowledge graph information to make recommendations. Recommendation algorithms typically build models based on user-item interactions (e.g., clicks, likes, or ratings) to provide a personalized ranked list of items. These interactions are often distributed unevenly over different groups of items due to varying user preferences. The main challenge is how to adapt the recommendation model to handle a new biased data.

Inequality is a long-standing issue existing across various scenarios, such as gender inequality of scientific impact, infrastructure inequality in cities. However, inequality of people's online consumption has not recieved much attention in related researches. In this work, we aim to reveal widespread inequality of vulnerable users' content accessibility and consumpotion in a short-video platform. Furthermore, we will explore the essential mechanism of how this inequality forms with large-scale and long-term interaction records, from both perspectives of user behaviors and algorithmic effects. Preliminary results have shown that boys and women are consuming more homogeneous content, even for new users, demonstrating the inequality from recommendation algorithms.

The onset of the Covid-19 pandemic in February 2020 prompted the U.S. federal government and states “stay-at-home” orders that led to a surge in telecommuting adoption The purpose of this project is to explore the influence of telecommuting on human mobility, and to study how the change of human mobility influence urban center, community structure and the rationality of infrastructure distribution. We first verify the correlation between telecommuting data and human mobility data, and find that there is a negative correlation between their amount of change. This is consistent with our hypothesis that an increase in telecommuting will cause a decrease in human mobility. Then, we move on to explore the impact of human mobility on urban structure. Most of the existing researches on urban structure are based on static indicators, such as population density and height of urban buildings. The innovation of our project is to use dynamic metrics, human movement, to study urban structure and its changes. We build a human mobile network before and after the epidemic based on the safegraph data set, and then (1) divide urban centers by calculating the centrality index of the network, (2) divide communities by community detection algorithm, and (3) study the impact on infrastructure distribution. I will present our preliminary conclusions in the presentation.

With the deployment of 5G networks, the huge energy consumption and carbon emissions caused by 5G networks have become a difficult problem. To solve urgent development for sustainable traditional energy-saving methods have acceptable performance on 4G networks, but they fail to solve the problem of energy saving in heterogeneous networks consisting of both 5G and 4G base stations. In this paper, we propose a data-driven method to mitigatecarbon emissions of heterogeneous 4G and 5G networks. Specifically, to realize large-scale network energy saving, we propose to divide the grid based on the co-coverage relationship of cells, use historical data to predict traffic and select active cells for the grid, and model energy efficiency profiles for the grid. Then, we offload 5G network traffic to its corresponding equivalent grid with the highest energy efficiency Profiles to save energy and reduce carbon emissions through 4G and 5G joint shutdown control. We conduct an experiment to verify the energy-saving effect of our method in Nanchang City. The experiment results show that our method can save 3442.7 MWh of energy consumption and reduce 3276.5 tons of CO2 emissions in Nanchang in a week, and its energy saving rate is 18.15% higher than the greedy method and 38.85% higher than threshold-based method.

With the development of mobile communication technologies, mobile networks have become more complicated, while there is an increasing requirement for the quality of communication services. The digital twin for communication networks can abstractly represent a real communication environment and be used as a foundation for testing services provided under actual or simulated conditions. Mobile networks consist of multiple base stations (BSs), and various parameters could be tuned to provide adequate services. Automatically optimizing such parameters remains challenging since they are sensitive to environmental fluctuations and might impact each other. Reinforcement learning (RL) techniques are potent solutions for learning configuration strategies from interactive data, but they are typically challenging to adapt to multiple agents. In this paper, we illustrate how to employ digital twin technique and reinforcement learning in mobile communication networks consisting of many collaborating agents. We explain how mobile networks may be modeled in the form of digital twins and how to handle network optimization issues effectively using multi-agent reinforcement learning (MARL). Especially, in our proposed MARL algorithm, multiple features are extracted elaborately from historical knowledge through UNet and allows to explicitly learn coordinating behaviors among all antennas through value functions. We demonstrate experimentally that the proposed MARL algorithm outperforms alternative approaches. The utilization of the optimized MARL can well generalize to a high number of agents without reducing cooperation.

Diversity in recommendations has become a growing area of interest, but attempts to improve diversity typically come at a cost of accuracy. In this work, we propose a model for improving diversity without hurting accuracy by adapting to users' individual diversity sensitivity: namely, whether certain users respond positively or negatively to more diverse recommendations. We first validate the presence of variation in users' diversity sensitivity in a representative dataset. With the help of psychology literature, we define two levels of diversity preferences based on intrinsic user traits: 1) an overall preference for variety in recommendations, and 2) a preference for diversity per item based on the strength of category interests. Then we design a method for inferring the two types of sensitivities for each user so we can adapt recommendations accordingly. To do so, we propose DivSense, a method to adaptively sample diverse neighbors in a graph convolutional network according to users' inferred diversity preferences. Our proposed approach consists of two major components: 1) entropy adjusted neighbor sampling, which addresses overall diversity sensitivity, and 2) cluster based sub-interest neighbor sampling, which addresses category based diversity sensitivity. Through extensive experiments, we show that our solution exceeds the state of the art in both diversity and accuracy. We further evaluate results for users at different extremes of sensitivity to determine that the model does, indeed, adapt to users' sensitivities to in a more optimal way than the state of the art method.

Polarization is a ubiquitous phenomenon in social systems. Empirical studies document substantial evidence for opinion polarization across social media, showing a typical bipolarized pattern devising individuals into two groups with opposite opinions. While coevolving network models have been proposed to understand polarization, existing works cannot generate a stable bipolarized structure. Moreover, a quantitative and comprehensive theoretical framework capturing generic mechanisms governing polarization remains unaddressed. In this paper, we discover a universal scaling law for opinion distributions, characterized by a set of scaling exponents. These exponents classify social systems into bipolarized and depolarized phases. We find two generic mechanisms governing the polarization dynamics and propose a coevolving framework that counts for opinion dynamics and network evolution simultaneously. Under a few generic assumptions on social interactions, we find a stable bipolarized community structure emerges naturally from the coevolving dynamics. Our theory analytically predicts two-phase transitions across three different polarization phases in line with the empirical observations for the Facebook and blogosphere datasets. Our theory not only accounts for the empirically observed scaling laws but also allows us to predict scaling exponents quantitatively.

To better understand how the wireless environment characterizes the electromagnetic signal, it is inevitable to study channel modeling. The ultimate purpose of channel modeling is to achieve high accuracy and efficiency and to support the system design and 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, approaches based on machine learning revealed the feasibility of modeling channels in a data-driven way. In this work, we jointly consider the physical knowledge and the measured data. Specifically, we solve the integral form of Maxwell's equations to roughly derive the electric field distribution and then leverage the neural network to learn to correct the error. Moreover, we apply a neural network to replace the integral operation in the physical knowledge to accelerate the computation. By comparing with representative baselines, the results reveal the superiority of our proposed method.

Spatio-temporal point process (STPP) is a stochastic collection of events accompanied with time and space. Due to computational complexities, existing solutions for STPPs compromise with conditional independence between time and space, which consider the temporal and spatial distributions separately. The failure to model the joint distribution leads to limited capacities in characterizing the spatio-temporal entangled interactions given past events. In this work, we propose a novel parameterization framework for STPPs, which leverages diffusion models to learn complex spatio-temporal joint distributions. We decompose the learning of the target joint distribution into multiple steps, where each step can be faithfully described by a Gaussian distribution. To enhance the learning of each step, an elaborated spatio-temporal co-attention module is proposed to capture the interdependence between the event time and space adaptively. For the first time, we break the restrictions on spatio-temporal dependencies in existing solutions, and enable a flexible and accurate modeling paradigm for STPPs. Extensive experiments from a wide range of fields, such as epidemiology, seismology, crime, and urban mobility, demonstrate that our framework outperforms the state-of-the-art baselines remarkably, with an average improvement of over 50%. Further in-depth analyses validate its ability to capture spatio-temporal interactions, which can learn adaptively for different scenarios.

The growing complexity of next-generation networks exacerbates the modeling and algorithmic flaws of conventional network optimization methodology. In this paper, We propose a mobile network digital twin (MNDT) architecture for 6G networks. To address the modeling and algorithmic shortcomings, the MNDT uses a simulation-optimization structure. The feedback from the network simulation engine, which serves as validation for the optimizer's decision outcomes, is specifically used by an AI-based optimizer to iteratively optimize. We develop a network digital twin prototype system in practice using data-driven technology to accurately model the behavior of mobile network elements (e.g., mobile users and base stations), wireless environments, and network performance. An artificial intelligence (AI)-powered network optimizer has been developed based on the deployed MNDT prototype system. The results of the experiments demonstrate that the proposed MNDT infrastructure can provide practical network optimization solutions while being able to adapt to the more complex environment that mobile network systems are operating in.

The two-dimensional bin packing problem (2DBP) is a critical optimization problem in the industry of furniture production, glass cutting, and so on, where the objective is to cut smaller-sized items from a minimum number of large standard-sized raw materials. In practice, hundreds of customer orders (sets of items) are manufactured every day in a factory. To relieve the pressure in management, a common practice is grouping the orders into batches for production so that the items from one order are in the same batch rather than scattered all over the production line. In this work, we formulate it as the grouped 2D bin packing problem, a bi-level problem where the upper level partitions orders into groups and the lower level solves 2DBP for items in each group. The main challenges are (1) the coupled optimization of upper and lower levels, and (2) the high computational efficiency requirement for practical application. To tackle these challenges, we propose an iteration-based hierarchical reinforcement learning framework, which can learn to solve the optimization problem in a data-driven way and provide fast online performance after offline training. Extensive experiments demonstrate that our method not only achieves the best performance compared to all the baselines but is also robust to the change in dataset distribution and problem constraints. Finally, our method is deployed in the factory of ARROW Home, China, reducing the cost of raw materials by 4.1%.

Vehicle trajectory boosts various city service and management applications, and also scientific researches. The city-scale and full-amount vehicle trajectory data is especially valuable and vital in some critical applications, such as city simulation with city-scale OD generation, vehicle pollution emission calculation, and road network science with knowing the flow of each road. My talk consists of two parts. The first part introduces a city-scale vehicle trajectory recovery system based on traffic camera video data towards the release of a city-scale vehicle trajectory dataset. In this system, we address the computation resource restriction problem in this city-scale task by dividing city into a few regions. We perform recovery in each region and then carefully piece up the trajectories across regions with a trajectory merging algorithm. The second part introduces my personal brain storming and some investigations on the promising and important research directions which can especially benefit from the city-scale and full-amount vehicle trajectory data. I’m looking forward to your kind suggestions on any other interesting research directions on trajectory applications.

How to use simulation technology to promote urban science and applied research is a topic of great interest to us. In this seminar, two works on simulation technology-assisted urban science research will be shared. The first work [1] simulates the traffic congestion process in several cities with the help of microscopic traffic simulation, and the analysis finds that in the congestion process the maximum number of congested clusters (the key parameter in percolation theory) and the maximum macroscopic fundamental diagram (MFD) flow occur at the same moment. The work provides a new perspective to understand the propagation of traffic congestion. The second work [2] focuses on the evaluation of the resilience of urban transport networks. It first simulates road traffic in 40 U.S. cities under everyday conditions and calibrates the delay time calculations using real-world statistics. Second, it simulates changes in the average delay time for urban access with different percentages of road damage, giving an evaluation of the resilience of transport networks in 40 U.S. cities.

People’s beliefs about COVID-19 are constructed by not only their own knowledge and experience with COVID but also other people’s attitudes and shared information. However, not all people’s words have the same impact on us. Those that are closer to us often have more influence on us than others. In this project, through conducting surveys and experiments, we investigate how people’s beliefs toward COVID are influenced by other people, with whom they have different strengths of social ties. In this presentation, I will share our research design and preliminary findings related to the experiments. Specifically for the design, we let participants, who self-confirmed COVID infection, share their COVID experience with five people separately, with whom they are related differently, i.e., a family member, a friend, a classmate or colleague, an acquaintance, and a stranger, and measure these five people’s change of COVID beliefs. I will also showcase our preliminary findings during the presentation.

For walking pedestrians, when they are blocked by obstacles or other pedestrians, they adjust their speeds and directions to avoid colliding with them, which is called collision avoidance behavior. This behavior is the most complex part of pedestrians' walking processes and its modeling and simulation are the keys to realistic crowd simulation, which serves as the foundation for various applications. However, most existing methods either lack the representation power to accurately model the complex collision behavior or do not model it explicitly, which leads to a poor level of realism of the simulation. To realize realistic crowd simulation, we propose to analyze, understand and model the collision avoidance behavior in a data-driven way. First, to automatically detect collision avoidance behavior for further analysis, we propose a domain transformation algorithm that detects it by transforming the trajectories in the spatial domain into a new domain where the behavior is much more apparent and is thus easier to detect. The new domain also provides a new perspective for understanding collision avoidance behavior. Second, since there are no mature metrics to evaluate the level of realism, we propose a new evaluation metric based on the least-effort theory, which evaluates the realism of collision avoidance behavior by its physical and mental consumption. This evaluation metric also provides the foundation of modeling. Third, for realistic crowd simulation, we design a reinforcement learning model. It trains agents with our proposed reward function that models pedestrians' intrinsic needs of "reducing effort consumption" and thus can guide agents to behave realistically when avoiding collisions. Extensive experiments show our model is 55.9% and 52.5% more realistic in collision avoidance behavior than the best baselines on two real-world datasets.

Sequential recommendation predicts the next item based on users' historical behaviors. Longer historical behaviors often improve recommendation accuracy but bring efficient problems. Existing works for long sequential behaviors only consider the relation between historical items and target items, failing to capture the relation between historical items. Such a manner indeed fails to make full use of the historical information because there is interest drifting within the historical sequence that is helpful to predict the target item. As sequences get longer, the following two main challenges have not been addressed: (1) efficient modeling under increasing sequence length and (2) interest drifting within historical items. In Wechat's Channels, we deploy Sparse Attentive Diffusion for Long-sequence Recommendation (DiffuRec) with Sparse Attention Layer and Diffusion Attention Layer. Specifically, we first propose three sparse attention modules - global attention, window attention, and random attention - to capture the global interest drifting for short-term items, local interest drifting, and casual interest drifting for historical items. Then to expand the interest drifting of the sparse attentions, we further propose Diffusion Attention Layer to expand the receptive field of each historical item by multi-step Personalized PageRank iteration. The theoretical analysis illustrates that our proposed diffusion method can approximate full attention efficiently. Experiments on two real-world datasets show the superiority of our proposed method against state-of-the-art baselines. Further ablation study and visualization also sustain its effectiveness.

With recent advances in reinforcement learning, we have witnessed countless successes of intelligent agents from game-playing to industrial applications. Especially, multi-agent reinforcement learning (MARL) is suitable for many real-world scenarios and has vast potential applications in the real world. However, typical MARL methods can only handle tens of agents, leaving scenarios with up to hundreds or even thousands of agents almost unexplored. There exist two key challenges in scaling up the number of agents: (1) agent-agent interactions are critical in multiagent systems while the number of interactions grows quadratically with the number of agents, causing great computational complexity and difficulty in strategies-learning; (2) the strengths of interactions vary among agents and over time, making it difficult to precisely model such interactions. In this paper, we propose a Graph Attention Mean Field (GAT-MF) method. By converting agent-agent interactions into interactions between each agent and a weighted mean field, which precisely models the dynamics of the interactions, we greatly reduce the computational complexity. We mathematically prove the correctness of this approach. We design a graph attention mechanism to automatically capture the different and time-varying strengths of interactions, ensuring that our method can precisely model interactions among the agents.We conduct extensive experiments in both manual and real-world scenarios with more than 3000 agents. The results demonstrate that our method achieves superior performance with an improvement of 42.7%, and saves 86.4% training time and 19.2% GPU memory comparing the best baseline method.

Millions of slum dwellers suffer from poor accessibility to urban services due to inadequate road infrastructure within slums, and road planning for slums is critical to the sustainable development of cities. Existing re-blocking or heuristic methods are either time-consuming which cannot be generalized to different slums, or yield sub-optimal road plans in terms of accessibility and construction costs. In this paper, we present a deep reinforcement learning based approach to automatically layout roads for slums. We propose a generic graph model to capture the topological structure of a slum, and devise a novel graph neural network to select locations for the planned roads. Through masked policy optimization, our model can generate road plans that connect places in a slum at minimal construction costs. Extensive experiments on real-world slums in different countries verify the effectiveness of our model, which can significantly improve accessibility by 14.3% against existing baseline methods. Further investigations on transferring across different tasks demonstrate that our model can master road planning skills in simple scenarios and adapt them to much more complicated ones, indicating the potential of applying our model in real-world slum upgrading.

The Origin-Destination~(OD) matrix provides an estimation of number of individuals traveling between regions, i.e., mobility flow in the city, which is widely-used in urban planning, transportation, etc. Given various city characteristics of urban regions, generating the city-wide OD matrix without using historical flow information has become increasingly appealing to both researchers and practitioners. However, existing works are limited in independent generation of each element, i.e., flow, in OD matrix, overlooking the element relations within the matrix that can be well formulated as a network. In this paper, we instead propose to generate the city-wide OD matrix from the network perspective, and design a graph denoising diffusion method to learn the conditional joint probability distribution of all elements in the OD matrix given city characteristics at region level. To overcome the learning difficulty of the city-wide OD matrix covering over thousands of regions, we decompose the original one-shot generative modeling of the diffusion model into two cascaded stages, corresponding to the generation of network topology and mobility flow, respectively. To further reproduce important network properties contained in city-wide OD matrices, we design an elaborated graph denoising network structure including a node property augmentation module and a graph transformer backbone. Empirical experiments on data collected in two large US cities have verified that our method can generate OD matrices for new cities with network statistics remarkably similar with the ground truth, further achieving superior outperformance over competitive baselines in terms of the generation realism.