Social event detection with Reinforced Deep Heterogeneous Graph Attention Network

With the rapid development of Graph Neural Networks, social event detection has become increasingly important as a way of letting the public know about significant global events in a clear and timely fashion. However, current methods are still unsatisfactory for detecting events. We argue there are three main reasons for this. First, any redundant, overlapping, or noisy information in a node’s neighbourhood will damage the semantic representations of final node embeddings. Second, during the aggregation process, most current detection methods predominantly focus on the interactions with auxiliary models in the horizontal direction, ignoring the extraction and integration of the rich semantic and structural information available in the vertical direction. In addition, the time interval between different messages in social networks is also not fully taken into account. Third, GNN embedding models are not robust to some nodes with limited neighbours, and with insufficient structural information, performance suffers. What is needed is an approach that can improve the embedding ability of GNN models by extracting and aggregating more semantic and structural information from the same graph to generate more representative and robust message embeddings, rather than constructing a richer message graph by adding additional attributed data. Hence, in this thesis, we propose a novel Reinforced Deep Heterogeneous Graph Attention Network (Re-DHAN) for offline and incremental social event detection tasks. The framework first relies on the multi-agent reinforcement learning algorithm (A2C) to select the most meaningful neighbours from the dense social networks within various meta-path graphs. Then, to extract and aggregate more temporal, semantic, and structural information from both the horizontal and vertical directions, we present a new DHAN embedding model that effectively integrates mini-batch sampling, time decay weights, adjacency matrices, and bias matrices via Deep Deep Temporal Node Attention and Semantic Aggregation Attention. In terms of time elements, the exponential time decay coefficient based on the time interval between different messages is used to penalize the standard attention weights. Finally, to ensure that the embeddings produced by the DHAN encoder are both robust to few neighbours and are highly representative, we developed a new Dual Graph Contrastive Learning model, called (DGCL), to simulate the missing structural information. The DGCL model includes a triplet loss function and an unsupervised GraphCL model with multi-scaled subgraph augmentations. This model also mitigates the semantic perturbations from any potentially noisy data. In offline settings, the K-Means clustering algorithm reveals the social events. In incremental detection scenarios, the process involves incremental social message graph update, dynamic social event detection, and incremental detection model maintenance. A series of experiments in offline and incremental social event detection demonstrate ReDHAN as superior to the existing state-of-the-art approaches. Furthermore, the results also show the ReDHAN has powerful and robust abilities to capture, extract, and aggregate the semantic and structural information in graphs.