In modern streaming recommendation services for music and videos, users are reluctant to provide positive feedback, and they tend to generate much more negative feedback such as the skip behavior when they are unsatisfied with the recommended contents. On the other hand, according to the loss aversion theory of behavioral economics, the pain from a loss is psychologically more powerful than pleasure from a similar gain. Therefore, it is crucial to exploit negative feedback for streaming recommendation. Most of the existing works rely on positive feedback to capture user interest, and adapting them to negative feedback faces key challenges such as the constraints of metrics. A few works investigate dissimilarity, including dissimilarity matrix and dissimilarity vector. However, these approaches have only been validated in small-scale problems like writer verification, while the problem scale of recommendation is much larger. In this work, we set out to learn a dissimilarity space for recommendation to capture the unsatisfaction purely based on negative feedback between users and items.

Social recommendation assists the user's interest prediction by using the user's social relationship, thereby improving the performance of the recommendation system. However, in real scenarios, relationships on the web are not necessarily the same as having similar interests in the real world, which means that the user's social relationship is not always true. At the same time, for the same group of users, the interests of friends are not always similar in different scenarios, which also leads to noise in social relationships. The existence of these noisy data leads to a negative impact when directly using social relationships to enhance the recommendation performance in social recommendations. In order to solve these problems, in this work, we designed a social relationship denoising method based on the pre-training and fine-tuning framework. The model first learns general representations in the pre-training process, then, in the fine-tuning stage, we use the user’s interaction history to identify the noise in the social relationship, and finally improve recommendation performance.

In the past few years, micro-videos have become the dominant trend in the social media era. Recent advances in micro-video recommendation have achieved remarkable performance in mining users’ preferences based on historical video-level interactions. However, most of them utilize the video features obtained from other downstream tasks like video classification. These two-stage methods are not fit for the micro-video recommendation task, which usually encodes the whole video to get coarse-grained video features. To address this problem, first, we reprocess video features based on user behavior and validate that more refined video features could benefit the recommendation performance. Furthermore, we propose an end-to-end method to model user interest evolution during the video’s playing process by assessing user interest on fine-grained video clips. The preliminary experiment has shown the effectiveness of our method.

The success of artificial intelligence (AI) is inseparable from large-scale and high-quality data, which is not always available. Involving human forces like crowdsourcing can help provide more training data and improve data quality for AI tasks. But with more privacy concerns and stricter laws, the data isolation problem is becoming worse, just when federated learning (FL) has emerged as a promising solution. In this article, we highlight the core issues in federated learning in the lens of crowdsourcing, including privacy and security, incentive mechanism, communication optimization and quality control. We expect to inspire the design of federated learning systems with existing crowdsourcing techniques. We also discuss emerging future challenges to implement a fully fledged federated learning platform.

This paper investigates the joint power control and spectrum sharing problem for vehicle-to-everything (V2X) communications based on long term evolution unlicensed (LTE-U) technology in a heterogeneous network environment. We classify vehicle user equipments (VUEs) into safety and non-safety VUEs according to their respective service and access categories. To address the joint problem of spectrum allocation, power control, and spectrum sharing, a resource allocation strategy is proposed to guarantee fair coexistence among users. Each periodic access interval over unlicensed spectrum is divided into content period (CP) duration and content free period (CFP) duration. Accordingly, this paper proposes a framework which allows non-safety VUEs to not only contend for unlicensed spectrum in CP duration, but also utilize reserved unlicensed spectrum in CFP duration. Based on the framework, we set our objective to maximize the total throughput of cellular user equipments and non-safety VUEs, and propose a three-step decomposing scheme, which is nested with interior point method and matching algorithm. We also assess the ergodic sum rate with respect to statistical channel state information with reduced computation complexity.

Caching popular content 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.

In the Internet of Vehicles (IoV), data sharing among vehicles for collaborative analysis can improve the driving experience and service quality. However, the bandwidth, security, and privacy issues hinder data providers from participating in the data-sharing process. In addition, due to the intermittent and unreliable communications in IoV, the reliability and efficiency of data sharing need to be further enhanced. This talk mainly discusses the new architectures for IoV based on federated learning to relieve transmission load and address privacy concerns of providers. The first work consists of the permissioned blockchain and the local Directed Acyclic Graph (DAG). Moreover, we propose an asynchronous federated learning scheme by adopting Deep Reinforcement Learning (DRL) for node selection to improve efficiency. The second work proposes the use of Unmanned Aerial Vehicles (UAVs) as wireless relays to facilitate the communications between the IoV components and the FL server and thus improving the accuracy of the FL. However, a single UAV may not have sufficient resources to provide services for all iterations of the FL process. Hence, a joint auction-coalition formation framework to solve the allocation of UAV coalitions to groups of IoV components is proposed. Moreover, possible future directions are also discussed.

Coinciding with this increase of large-scale distributed computing has been the relentless effort to improve the performance of these applications through exploiting the abundance of resources in the hyper-scale cluster and with little attention to resources efficiency. The inefficient use of resources when executing workflows is determined by 1) the excessive amount of resources provisioned and 2) the wastage from unused resources among task runs. Finding a workflow scheduler reducing resource usage is of great practical importance and can lead to (1) improvements in resource utilization, (2) reductions in resource provisioning to save money and (3) savings in energy consumption. To solve the problem, we propose a method to find resource-saving workflow scheduling.

Crowd flow data is of great importance in many urban computing tasks, such as human mobility mining and urban planning, etc. However, the development levels of different cities are unbalanced, and still, many cities suffer from data scarcity issues. Therefore, it’s necessary for us to design a model to generate the crowd flow data for the target city where data are not available based on the data-rich city. Existing transfer learning works learn the correlations between the data-rich source city and data-scarce target city, which still need few flow records or generate static data. To solve the problem, we propose a method to learn mobility knowledge in the source city and generate crowd flow data in the target city.

Covid-19 vaccine inequity among regions continues to stink, undermining the global effort to control the pandemic. Although high-level vaccine sharing guidelines have been proposed by organizations such as the COVAX, it remains elusive why, how, and to what extent a vaccine sharing mechanism can benefit epidemic control. To answer the above questions, we first propose a mobility-aware inter-regional epidemic model to capture the epidemic spread on both global and regional scales. Based on the proposed model, we design and evaluate vaccine sharing mechanisms that redistribute vaccines among vaccine-production regions and non-production regions with different levels of altruism. We find that a proper level of altruism can result in the minimization of accumulated infections, winning over both extreme egoism and extreme altruism. Further, we seek to uncover the source of benefit by analyzing the impacts of imported and exported cases. Finally, we showcase the benefit of vaccine sharing in terms of relaxation of mobility control, which can aid the recovery of global economics.

With the rapid development of the economy and technology, more and more people choose to live in cities, and there are more and more vehicles on the roads. Due to a variety of causes, traffic congestion has greatly affected people's life of quality, people waste a lot of time and money on commuting every day. It is important to understand the nature of congestion caused by different causes, which is the key to traffic control. Working with Amap, we have some labeled data on the categories of congestion causes. These labeled categories data includes high traffic flow from hospitals, schools, scenic spots, exit queues, and toll lines, but in reality, there are many categories of congestion causes that are not labeled. Our goal is to label the unlabeled samples with these known categories or unknown categories. We first used labeled data to mine traffic congestion differences caused by different causes. We found that traffic congestion caused by different reasons presents different features in time, space, and the process of congestion formation and dissipation. Therefore, a congestion event is represented by these different features. Then, a deep clustering method was used to classify all congestion events. To the best of our knowledge, this is the first work to quantitatively classify the causes of traffic congestion and analyze corresponding patterns.

城市规划是对一定时期内城市的经济和社会发展、土地利用、空间布局以及各项建设的综合布局、具体安排和实施管理。基于对最大化公共领域的社会效益的初衷，城市规划这门学科应运而生。而其目标也与我们看城市问题的目标有异曲同工，因而了解城市规划领域的传统方法理论与他们看问题的视角是有必要的。在本次分享中，我将介绍城市规划领域现行的编制与实施的流程，并着重分享该学科调查分析城市的经典方法。希望大家能够在分享中关心思考如下问题： 1、城市规划的目的是什么？ 2、它核心的方法和理念是什么？ 3、现有流程中哪些部分运用到了数据？ 4、现有流程中哪些部分方法有改进空间，可以用data-driven来赋能？

The essence of man is the ensemble of social relations. Thus, learning the formation dynamics of social networks is fundamental. Existing works can be divided into two categories. One is from sociology. Researchers use knowledge-driven methods, summarizing empirical equations from observations, and validating using social experiments. The other is from computer science, which tries to learn the social network formation mechanism from data by formulating it into the dynamic link prediction problem. However, both the knowledge-driven methods and the data-driven methods exhibit fundamental flaws. Knowledge-driven methods cannot accurately model the complex dynamics of network formation, and data-driven methods lack generalizability. In this work, we set out to design a knowledge and data co-driven method to model the dynamic formation process of social networks by using a neural process to learn the social dynamics with the form extracted from classical social formation theories.

As an effective channel for information filtering, the recommender system provides people with many conveniences for access to what they are interested in. However, the ensuing problem of narrowing contents or viewpoints under personalized service (a.k.a. filter bubble) is concerned by society community from various fields. In this discussion, we will first summarize the primary research fields, paradigms, and findings of related studies, then present existing results relevant to the filter bubble on a micro-video platform. At last, some potential research directions are discussed on this open problem.

Cross-domain recommender system is a powerful method to tackle the cold-start and sparsity problem by aggregating and transferring user preferences across different domains. While several cross-domain sequential recommendation models have been proposed to leverage information from a source domain to improve the performance in a target domain, they are based on a too strong assumption, i.e., transferring by overlapped users or items. In addition, they are vulnerable to the variance of data distribution. In this work, we propose a novel approach based on the auto-denoising components that can guarantee robust learning in spite of the various data distribution, extracting 1) the common knowledge across different domains and 2) unique knowledge of each domain simultaneously. In particular, our Auto-denoising Sequential Cross-domain Recommendation (ADSCDR) consists of two novel components. First, with the proposed cross-attention module, where each domain queries the common knowledge from the shared self-attention, we successfully extract the common knowledge without negative transferring. Second, we incorporate the common user-group representation to further extract the common characteristic across different domains, with the proposed group-attention module, by which the probability of the user belonging to several static groups can be obtained. We conduct experiments on two semi-synthetic datasets and three real-world datasets to demonstrate the superiority of our proposed model. Further experiments illustrate the effectiveness of the proposed components.

Traffic signal control is essential for transportation efficiency in road networks. Conventional transportation research suffers from the incompetency to adapt to dynamic traffic situations. Meanwhile, Reinforcement learning (RL) is a promising data-driven approach for adaptive traffic signal control in complex urban traffic networks. While most existing works focus on controlling traffic lights via simplified methods, such as choosing phases only, applicable TSC systems require adaptation to prefixed phase definitions and adjusting cooperation between different signals. We focus on practical arterial TSC development specifically, while developing an RL-based system in TSC is not intuitive due to two major challenges. 1) It is not trivial to handle the cooperation between different signals. 2) Leveraging the extremely high dimension of the joint action space is difficult. In this project, we design a multi-agent reinforcement learning (MARL) based framework for arterial TSC. The decentralized controlling mechanism benefits in reducing the learning difficulty of each local agent and helps in improving observability. Results on manually generated arterial scenarios demonstrate the superiority of such a framework to other baseline methods.

Synthesized user consumption trajectories are crucial for a large number of applications, such as consumption prediction and store recommendation. However, real-world data exhibit certain weaknesses, such as privacy concerns and low quality. Thus, a realistic simulation of a massive amount of user consumption data is of great value. Existing solutions are mainly based on the Generative Adversarial Imitation Learning (GAIL), which is limited due to the lack of ignoring the user’s personalized consumption habits and modeling the impact of new stores on the user’s decision-making process. To address these shortcomings, we propose a Hierarchical Generative Adversarial Imitation Learning （HGAIL) framework, which models the user's universal consumption law and the user's personalized consumption habit. First, we model the user's consumption as a Markov process and propose HGAIL to model user consumption. The upper module models whether users consume or not and the lower module models the user's choice for the stores and contains two parts. One part is based on a novel semantics-based interaction mechanism between the decision-making strategy and visitations to old and new stores, which is used to model universal user consumption law. The other part is based on a data-driven method, which is to model users' personalized consumption habits. On the basis of the HGAIL framework, we then model the user's choice of the store with the user's preference for store features, so as to take into account the impact of new stores on the user's choice.

In the era of information explosion, recommender systems play an indispensable role in what we see, what we purchase, and even what we like. Despite the convenience, recommender systems also create or amplify some societal issues, such as the filter bubble, which is the focus of this discussion. The filter bubbles phenomenon, as first coined by Pariser, describes online personalization services that might potentially confine users to an overwhelmingly narrow set of contents or viewpoints. We make an analysis on a social recommender system, and demonstrate social factors might potentially reinforce filter bubbles in personalized recommender systems. Our findings indicate social factors are not always beneficial, it could create or amplify some issues in the information system.

The COVID-19 global pandemic and resulted lockdown policies have profoundly shaped society in a great many areas, among which is education. Due to the suspension of the traditional co-located face-to-face teaching, formal education refers to online “distance learning from home” for the continuity of students’ learning. Taking as a case study a Chinese leading university where learning takes the form of online distant learning for a whole semester, we (intend to) conduct a large-scale data-driven study together with large-scale surveys to unpack the effects of COVID-19 on education.

Human activity data are useful for many applications, such as activity prediction and recommendations. However, real-world data exhibit certain weaknesses, such as privacy concerns, collection cost, and low quality. Thus, realistic simulation of a massive amount of activity data is of great value. Existing solutions for human activity simulation can be classified into two categories: rule-based and data-driven methods, which are both limited in generating high-quality activity data as they cannot effectively model the inner drive of activities, i.e. human need, which however plays a crucial role in the activity decision-making process. To address these shortcomings, we propose a need-aware activity simulation framework based on imitation learning, with the consideration that the need is the interval driving force of human activities. First, we extend the activity modeling to the need semantics based on well-respected psychology theories. Then, to model the dynamics of human need, which evolves continuously over time and can be changed instantly by happened activities, we design a need state extractor based on Neural Jump Stochastic Differential Equations. Based on the dynamic modeling, we characterize the activity simulation as a continuous-time Markov decision process and solve it with the framework of GAIL. Extensive experiments on two real-life human activity datasets demonstrate that our framework outperforms state-of-the-art baselines significantly on data fidelity in terms of a number of key metrics. Moreover, the data generated by our model can well support practical applications, showing its effectiveness especially when the real data are on small scales.

Site selection determines optimal locations for new stores, which is of crucial importance to business success. However, existing data-driven methods heavily rely on feature engineering, facing the issues of business generalization and complex relationship modeling. To get rid of the dilemma, in this work, we borrow ideas from knowledge graph (KG), and propose a knowledge-driven model for site selection, short for KnowSite. Specifically, motivated by distilled knowledge and rich semantics in KG, we firstly construct an urban KG (UrbanKG) with cities' key elements and semantic relationships captured. Based on UrbanKG, we apply pre-training for semantic representations, which are fed into an encoder-decoder structure for site decisions. With the multi-relational message passing and relation path-based attention mechanism proposed, KnowSite successfully reveals the relationship between various businesses and site selection criteria.

Based on international airline mobility, we propose a meta-population epidemic model that examines the vaccine allocation strategies. We will investigate the efficiency and equity trade-off in the vaccine allocation process, where superior countries may receive greater benefits if they share vaccines with inferior countries. Besides, we will incorporate an economic model to validate the feasibility of the proposed vaccine strategy. Specifically, superior countries will pay extra economic costs when sharing the vaccines. These vaccines may benefit the recovery of global mobility, which will result in stronger economic relations between countries. Superior countries may receive greater economic benefits which can compensate for the cost of sharing vaccines.

The simulation of human mobility in crowds plays a vital role in the control of epidemics, path planning, and urban policy-making. The key challenge of crowds simulation is to model a variety of interactions among pedestrians, and between pedestrians and the environment, such as collision avoidance, walking in a group, and heading for the destination. On the other hand, due to the visual field restriction, pedestrians react to the surrounding pedestrians and environment differently depending on their relative orientation. To tackle these challenges, we propose a directional attentive message-passing network to model the context's different influences on the pedestrians. Our model can learn the inherent law of pedestrian complex interaction, which can benefit the large-scale simulation of pedestrian trajectories. Compared to previous models in the social science domain such as the social force model, our is data-driven to learn the internal laws of human interaction behavior, which is not subject to the assumptions of sociological theory. Thus, our model has better robustness to noises and better generalization to new scenarios and can support sufficiently complex interaction behavior modeling. Our model can be used to simulate human mobility in crowds in many scenarios, such as airports, squares, pedestrian streets.

Mobile cellular network is becoming an indispensable part of urban lives in recent years. Under this circumstance, the network planning problem, which aims to design a cellular network to satisfy the requirements of mobile users, is of increasing significance. However, a critical and challenging problem arises --- how would the planned cellular network behave? To solve this problem, many researchers focus on modeling and simulating the urban cellular traffic, which intuitively represents the activities of cellular network. However, existing methods fail in simulating the long-term temporal behaviors of cellular network traffic, while cannot model the influences of urban environments on the cellular networks. We propose a knowledge-enhanced granular GAN framework to generate cellular traffic in the target area via training on the source area with abundant data. In our framework, we learn the urban knowledge for a cellular network via constructing an urban knowledge graph and modeling long-term temporal behaviors in the cellular traffic via a granular GAN. The granular GAN captures the daily patterns and weekly patterns for the cellular traffic via component-based GANs, and learns long-term temporal correlations via a TCN-based GAN, then generates the cellular traffic step by step. We evaluate our model on a real cellular traffic dataset collected in Shanghai, which is divided into the center, suburb, and outer suburb areas. Our generation model outperforms three state-of-art generation models, and the urban knowledge could improve the performance of our model by 4.71%. Moreover, our model is capable of leveraging the temporal behaviors and urban environment influences from an in-service mobile cellular network in selected areas to the network planning in a target area.