Modeling Domain and Feedback Transitions for Cross-Domain Sequential Recommendation

TL;DR

This work addresses cross-domain sequential recommendation by modeling both domain transitions and feedback transitions with Transition^2. The method couples a transition-aware graph encoder (incorporating negative feedback via a transition matrix) with a cross-transition multi-head self-attention mechanism that uses specialized masks to capture diverse transition types. It further employs contrastive losses for domain alignment ($\mathcal{L}_{\text{align}}$) and feedback differentiation ($\mathcal{L}_{\text{cont}}$), achieving improved representations across domains. Experiments on Douban and Amazon CDSR datasets demonstrate that Transition^2 consistently outperforms strong SR and CDSR baselines, validating the importance of negative-feedback-informed cross-domain transitions for accurate recommendations.

Abstract

Nowadays, many recommender systems encompass various domains to cater to users' diverse needs, leading to user behaviors transitioning across different domains. In fact, user behaviors across different domains reveal changes in preference toward recommended items. For instance, a shift from negative feedback to positive feedback indicates improved user satisfaction. However, existing cross-domain sequential recommendation methods typically model user interests by focusing solely on information about domain transitions, often overlooking the valuable insights provided by users' feedback transitions. In this paper, we propose $\text{Transition}^2$, a novel method to model transitions across both domains and types of user feedback. Specifically, $\text{Transition}^2$ introduces a transition-aware graph encoder based on user history, assigning different weights to edges according to the feedback type. This enables the graph encoder to extract historical embeddings that capture the transition information between different domains and feedback types. Subsequently, we encode the user history using a cross-transition multi-head self-attention, incorporating various masks to distinguish different types of transitions. To further enhance representation learning, we employ contrastive losses to align transitions across domains and feedback types. Finally, we integrate these modules to make predictions across different domains. Experimental results on two public datasets demonstrate the effectiveness of $\text{Transition}^2$.

Figures (9)

• Figure 1: A toy illustration and statistics of domain and feedback transitions in cross-domain sequential recommendation. Type 1 transition: when transitioning across domains, the user feedback changed from positive to negative; Type 2 transition: when the user browses the new domain, her feedback improved from negative to positive.
• Figure 2: The overall architecture of $\text{Transition}^2$ comprises five key components: (1) an embedding layer, which initializes item embeddings; (2) a transition-aware graph encoder, which constructs graph-based representations of user sequences while incorporating transition information; (3) a cross-transition multi-head self-attention mechanism, designed to capture diverse transition patterns through specialized masking strategies; (4) a transition alignment module, which employs contrastive learning to align representations across different transitions; and (5) a prediction module, responsible for generating recommendations and optimizing the overall objective.
• Figure 3: Analysis of domain alignment loss $\mathcal{L}_{\text{align}}$.
• Figure 4: Analysis of feedback contrast loss $\mathcal{L}_{\text{cont}}$.
• Figure 5: Sensitivity Analysis of $L$.