Shanghai Jiao Tong University

Spatio-temporal graph learning is a key method for urban computing tasks, such as traffic flow, taxi demand and air quality forecasting. Due to the high cost of data collection, some developing cities have few available data, which makes it infeasible to train a well-performed model. To address this problem, cross-city knowledge transfer has shown its great promise, where the model learned from data-sufficient cities is leveraged to benefit the learning process of data-scarce cities. The challenge for knowledge transfer in urban computing lies in: (1) How to transfer the spatio-temporal knowledge of dynamic feature extraction in source and target cities? (2) How to consider the varied spatial structure within different cities?

Therefore, we propose a model-agnostic few-shot learning framework for spatio-temporal graph called ST-GFSL. Specifically, to enhance feature extraction by transferring cross-city knowledge, ST-GFSL proposes to generate non-shared parameters based on node-level meta knowledge. The nodes in target city transfer the knowledge via parameter matching, retrieving from similar spatio-temporal characteristics. Moreover, we reconstruct the graph structure during meta-learning. The graph reconstruction loss is defined to guide structure-aware learning, avoiding structure deviation among different datasets. Extensive experiments on four traffic speed prediction benchmarks are conducted, and the experimental results demonstrate the effectiveness of our method compared with state-of-the-art methods.

With the vigorous development of data acquisition technologies, traffic data, such as vehicle trajectory, road sensor data, etc., are exploding. Accurate and timely traffic flow forecasting according to historical observations helps road users make better travel plans, alleviate traffic congestion, and improve traffic operation efficiency.

Our research is mainly focused on how to capture the spatiotemporal correlation of the dynamic and complex traffic network, and use deep learning methods to predict the future short-term and long-term urban traffic flow.

We consider both the local and contextual spatial information, and define the spatial neighbors and semantic neighbors of the road nodes. Multi-head graph attention mechanism is utilized to model the road relationship as a dynamic weighted graph. In addition, we propose a novel adaptive graph gating mechanism to selectively update and forget the high-order neighbor information of nodes within the multi-layer stacking. GNN based on adaptive adjacency matrix can identify deviations caused by artificially defined spatial relationships and characterize global spatial correlations.

Traffic management is one of the key issues in building smart city, including traffic signal control, traffic prediction, etc. Our research mainly focuses on modeling traffic patterns and making accurate prediction for the traffic conditions.

Traffic prediction is challenging due to the complex spatial-temporal dependency. Thus, we try to answer these following question: (1) How to capture the time-evolving cascading behaviors in the road network? (2) How to enhance the expressive power of traffic pattern extractor?

To address these problems, we propose a novel deep learning framework to extract the most relevant historical information for prediction by capturing the underlying cascading behavior. An encoder-decoder architecture is adopted, where the historical contextual information of each road is encoded into a sequence of historical embedding. A spatiotemporal attention mechanism is devised to model the cascading behavior in the embedding space so that the most relevant information for prediction is concentrated.

Extensive experiments on both urban and highway traffic datasets verify the effectiveness of our proposed approach.

As an indispensable part of the internet economy, recommender system plays an important role in helping users overcome information explosion. Collaborative filtering is the current main stream for providing recommendation, which can be further divided into a normal one and heterogeneous one, depending on the user-item interaction type.

We mainly focus on heterogeneous collaborative filter, where users exhibit different levels of preference to all items. We try to answer these following question: (1) How to capture the heterogeneous interactions between users and items? (2) How to integrate different preference signals to best characterize user preference?

To address these problems, we propose a deep learning framework to generate high-quality item embedding. The framework first constructs a set of item homogeneous graphs based on the user-item interaction bipartite graph to capture the structural closeness among items under different behaviors. After that, a separate importance-based graph convolutional network is applied to each of the item homogeneous graphs to extract the semantic embeddings under the corresponding behavior. Finally, an attention mechanism is proposed to fuse all the semantic embeddings to get the final item embeddings.

Experiments on a real-world E-commerce platform dataset with millions of interactions indicate that the proposed model can improve recommendation accuracy at least by 10% compared with baselines.