Distribution-Based User Purchase Characteristics Network for Sequential Recommendation

Sequential recommendation aims to predict the next item a user will interact with based on an analysis of their historical behaviors and interactions. Many Transformer-based methods embed items as dense vectors and utilize scaled dot-product attention to establish item relationships, resulting in impressive performance. However, most existing Transformer-based methods ignore the shopping intention, which is crucial for e-commerce platforms to improve customers' shopping experience. In addition, since the user behavior sequences often exhibit diversity and uncertainty, it is difficult for dot-product self-attention to capture this uncertainty. To address the aforementioned problems, we propose a Distributed-based User Purchase Characteristics Network (DUPCN) that embeds each item and it's attribute as a Gaussian distribution to mine distributed user purchase characteristics. Unlike existing Gaussian/distributional SR models that only model the uncertainty of intra-sequence item transitions based on user behavior sequences, DUPCN constructs an intent-oriented dual-source distribution by fusing user behavior sequences and multimodal item attributes (such as price, category, color). This enables DUPCN to mine potential shopping intentions from the distribution. Specifically, to obtain the feature representation of item sequences and their attributes, strengthen users' personal purchasing preferences, and weaken uninteresting items and attributes, the Wasserstein Transformer based on Item Relationship Graph which measures attentions as scaled Wasserstein distances between items to discover potential associations between different items (item attributes). Furthermore, the Feature Fusion module is designed to align learned shopping intent with user preferences through JS-divergence minimization. The experiments on four large normal benchmark datasets show that DUPCN has excellent performance compared with several state-of-the-art methods, especially in solving the short sequences. The source code is publicly available at https://github.com/WFU-Ge/DUPCN.