Optimizing Personalized E-Commerce Micro-Video Recommendation with Self-Adaption Generative Gating Graph

In the current digital age, recommending micro-videos (MVs) on e-commerce sites has emerged as a pivotal challenge. These platforms often grapple with severe sparsity issues since users typically engage in browsing products rather than videos. Enhancements in Graph Neural Networks (GNN) through Knowledge Graphs (KG) have shown widespread effectiveness in mitigating issues of sparsity and initiating recommendations for new users. Yet, prevailing KG-boosted GNN approaches seldom incorporate multi-domain user activities and overlook the sequential dependence of initial nodes when assimilating their adjacent nodes. In this work, we introduce an innovative Self-adapting Generative Gating Graph model (SGGG) that operates across both MV and product domains, acknowledging the sequential dependencies during neighbor aggregation. We employ a unified KG to establish links between MVs and products, enabling the sampling of subgraphs for each product interaction, thereby creating a heterogeneous behavior graph. A novel sequence generation mechanism, the Self-adapting Generative Graph Attention (SGGA), is developed for neighbor aggregation, capturing the sequential dynamics of each subgraph. Additionally, we unveil a distinctive salient feature selection technique, KG gating, equipped with a KG-driven hierarchical gating mechanism, to extract high-level semantic information from the KG. Our method’s effectiveness is validated through experiments on two publicly available datasets, demonstrating substantial advancements over leading recommendation techniques, encompassing those based on MVs and KGs. Furthermore, the deployment of our model on a major e-commerce platform, corroborated by online A/B testing, underscores its practical efficacy.