Boundary-Aware Multi-Behavior Dynamic Graph Transformer for Sequential Recommendation

TL;DR

This work tackles dynamic, multi-behavior sequential recommendation by proposing MB-DGT, a boundary-aware dynamic graph transformer. It unifies three components: (i) Transformer-based Multi-behavior Correlation Modeling (TMCM) to capture behavior-specific dependencies, (ii) Time-Aware Aggregation (TAA) to encode temporal and higher-order interactions on an $L$-hop graph, and (iii) Boundary-Aware Personalized Prediction (BAPP) to learn user-specific boundaries $b_{u,k}(t)$ governing different behaviors via a boundary-adjusted loss. Empirical results on Yelp, Taobao, and Tmall show MB-DGT outperforms traditional, multi-behavior, and multi-behavior sequential baselines, with ablations confirming the contributions of each module and the importance of auxiliary behaviors and interpretability of cross-behavior dependencies. The approach provides improved recommendations, better handling of cold-start via higher-order connectivity, and interpretable behavior relationships, with potential impact on real-world recommender systems that must adapt to evolving user interests across multiple action types.

Abstract

In the landscape of contemporary recommender systems, user-item interactions are inherently dynamic and sequential, often characterized by various behaviors. Prior research has explored the modeling of user preferences through sequential interactions and the user-item interaction graph, utilizing advanced techniques such as graph neural networks and transformer-based architectures. However, these methods typically fall short in simultaneously accounting for the dynamic nature of graph topologies and the sequential pattern of interactions in user preference models. Moreover, they often fail to adequately capture the multiple user behavior boundaries during model optimization. To tackle these challenges, we introduce a boundary-aware Multi-Behavioral Dynamic Graph Transformer (MB-DGT) model that dynamically refines the graph structure to reflect the evolving patterns of user behaviors and interactions. Our model involves a transformer-based dynamic graph aggregator for user preference modeling, which assimilates the changing graph structure and the sequence of user behaviors. This integration yields a more comprehensive and dynamic representation of user preferences. For model optimization, we implement a user-specific multi-behavior loss function that delineates the interest boundaries among different behaviors, thereby enriching the personalized learning of user preferences. Comprehensive experiments across three datasets indicate that our model consistently delivers remarkable recommendation performance.

Figures (8)

• Figure 1: This figure illustrates multi-behavior sequential recommendation, depicting interactions across timestamps for three users.
• Figure 2: Two examples of 2-hop multi-behavior temporal interaction graphs with user and item as target nodes respectively at moment $T_4$
• Figure 3: The model architecture of MB-DGT with users as target nodes. The overall structure is depicted in the left panel, while the right panel provides a detailed view of the model's two main modules. The lower right corner illustrates the serialized representation encoding process for behavioral association perception, and the upper right corner demonstrates the time-aware aggregation process for one-hop neighbors.
• Figure 4: The performence of auxiliary behavioral data ablation experiments on HR@10 and NDCG@10.
• Figure 5: Boundary distribution on different behaviors for one thousand users on Taobao dataset.