Bipartite graph representation learning aims to obtain node embeddings by compressing sparse vectorized representations of interactions between two types of nodes, e.g., users and items. Incorporating structural attributes among homogeneous nodes, such as user communities, improves the identification of similar interaction preferences, namely, user/item embeddings, for downstream tasks. However, existing methods often fail to proactively discover and fully utilize these latent structural attributes. Moreover, the manual collection and labeling of structural attributes is always costly. In this article, we propose a novel approach called Dirichlet Max-margin Matrix Factorization (DM3F), which adopts a self-supervised strategy to discover latent structural attributes and model discriminative node representations. Specifically, in self-supervised learning, our approach generates pseudo group labels (i.e., structural attributes) as a supervised signal using the Dirichlet process without relying on manual collection and labeling, and employs them in a max-margin classification. Additionally, we introduce a Variational Markov Chain Monte Carlo algorithm (Variational MCMC) to effectively update the parameters. The experimental results on six real datasets demonstrate that, in the majority of cases, the proposed method outperforms existing approaches based on matrix factorization and neural networks. Furthermore, the modularity analysis confirms the effectiveness of our model in capturing structural attributes to produce high-quality user embeddings.

The cross-device user matching task is to identify the behavior-logs (i.e., behavior sequences) on multiple devices that belong to one real person. Due to its anonymous and long-term properties, most previous methods of learning behavior embeddings cannot effectively capture two important features in the sequences, namely high-order connections and long-range dependencies. To this end, we propose a novel framework called Two-tier Graph Contextual Embedding (TGCE) to solve the above problems simultaneously. In the first tier, we construct behavior evolutionary graphs (BEGs) for behavior sequences and design an order-preserving neighbor aggregation network to collectively model transitions of behaviors with their neighbors. As repeated behaviors can be grouped into single nodes, our model joints neighboring environments around behaviors in a collective way, and behavior embeddings can be enriched. In the second tier, we further build scaled shortcut graphs (SSGs) by refining BEGs with random walk-based edge addition, then a position-aware graph attention network is further imposed on SSGs to facilitate fast information propagation. As distant graph nodes can be directly connected by shortcut edges, we can further capture long-range dependencies. By stacking two graph tiers, our approach can obtain graph contextual embeddings for behaviors to further improve user matching. Experimental results on the benchmark dataset show that our model outperforms various baselines in the user matching task. Our code is released on https://github.com/13061051/TGCE_2021.

Script event prediction (SEP) aims to choose a correct subsequent event from a candidate list, given a chain of ordered context events. Event representation learning has been proposed and successfully applied to this task. Most previous methods learning representations mainly focus on coarse-grained connections at event or chain level, while ignoring more fine-grained connections between events. Here we propose a novel framework which can enhance the representation learning of events by mining their connections at multiple granularity levels, including argument level, event level and chain level. In our method, we first employ a masked self-attention mechanism to model the relations between the components of events (i.e. arguments). Then, a directed graph convolutional network is further utilized to model the temporal or causal relations between events in the chain. Finally, we introduce an attention module to the context event chain, so as to dynamically aggregate context events with respect to the current candidate event. By fusing threefold connections in a unified framework, our approach can learn more accurate argument/event/chain representations, and thus leads to better prediction performance. Comprehensive experiment results on public New York Times corpus demonstrate that our model outperforms other state-of-the-art baselines. Our code is available in https://github.com/YueAWu/MCer.