Gaussian Mixture Flow Matching with Domain Alignment for Multi-Domain Sequential Recommendation

TL;DR

This work tackles multi-domain sequential recommendation by addressing exponential domain transitions, domain heterogeneity, and sparse data through GMFlowRec, a unified generative framework. It combines a dual-masked Transformer to separate domain-invariant and domain-specific intents, a domain-aligned prior to guide cross-domain generation, and a Gaussian Mixture Flow Matching module to capture multi-modal transition trajectories. Empirical results on JD and Amazon demonstrate state-of-the-art performance, with substantial improvements in NDCG and strong robustness in few-shot and high-transition scenarios, all while using a single unified backbone for scalability. The approach offers practical impact by enabling efficient, scalable cross-domain recommendations in real-world platforms with many domains and diverse user behaviors.

Abstract

Users increasingly interact with content across multiple domains, resulting in sequential behaviors marked by frequent and complex transitions. While Cross-Domain Sequential Recommendation (CDSR) models two-domain interactions, Multi-Domain Sequential Recommendation (MDSR) introduces significantly more domain transitions, compounded by challenges such as domain heterogeneity and imbalance. Existing approaches often overlook the intricacies of domain transitions, tend to overfit to dense domains while underfitting sparse ones, and struggle to scale effectively as the number of domains increases. We propose \textit{GMFlowRec}, an efficient generative framework for MDSR that models domain-aware transition trajectories via Gaussian Mixture Flow Matching. GMFlowRec integrates: (1) a unified dual-masked Transformer to disentangle domain-invariant and domain-specific intents, (2) a Gaussian Mixture flow field to capture diverse behavioral patterns, and (3) a domain-aligned prior to support frequent and sparse transitions. Extensive experiments on JD and Amazon datasets demonstrate that GMFlowRec achieves state-of-the-art performance with up to 44\% improvement in NDCG@5, while maintaining high efficiency via a single unified backbone, making it scalable for real-world multi-domain sequential recommendation.

Figures (7)

• Figure 1: Performance of SASRecSASrec on the JD multi-domain dataset. Bars report NDCG@10 across domains grouped by the number of domains within a user sequence. The line chart represents the domain transition rate, illustrating the increasing complexity of cross-domain interactions as the number of domains increases.
• Figure 2: GMFlowRec integrates a Dual-Masked Transformer to encode multi-domain user sequences into domain-invariant and domain-specific priors. This is achieved via two complementary attention masks: a domain-invariant mask that captures global user preferences across domains, and a domain-specific mask that preserves intra-domain intent. A Domain Alignment operation further refines the domain-specific prior into a domain-aligned prior, which guides the reverse generative process modeled by Gaussian Mixture Flow Matching.
• Figure 3: Performance comparison under domain transition sequence conditions across JD and Amazon datasets. We report NDCG@10 for each method under four settings: JD (w/o and w/ transition) and Amazon (w/o and w/ transition). GMFlowRec (Our) consistently achieves top performance, especially under transition settings. Blue shading indicates higher performance, while red indicates lower performance.
• Figure 4: Ablation and Component Analysis: (a) Performance comparison under the Amazon scenario, showing the impact of removing $\ell^{\text{GMM}}$ and $\ell^{\text{Prior}}$ from the full model. (b) Effect of varying the number of Gaussian components $K$ in the GMM prior.
• Figure 5: Performance comparison under few-shot domain sequence conditions across JD and Amazon datasets. We report NDCG@10 for each method under four settings: JD (w/o and w/ few-shot) and Amazon (w/o and w/ few-shot). GMFlowRec (Our) consistently achieves top performance, especially in few-shot scenarios. Blue shading indicates higher performance, while red indicates lower performance.