Leveraging meta-path and co-attention to model consumer preference stability in fashion recommendations

Abstract

With countless outfit combinations available, consumers often experience choice overload. Two key challenges that significantly impact the quality of recommendation systems are recommendation accuracy and fluctuations in consumer preferences. Previous works primarily extracted generic product features and modeled the compatibility of fashion items, overlooking the relationships hidden in user-product interactions and the evolution of consumer preferences. Unfortunately, this evolution of consumer preferences has not received much attention in the RS studies. To address these limitations, we propose a GPA-BPR (General compatibility and Personalized preference with co-Attention mechanism) framework, which integrates multimodal insights for practical outfit evaluation and utilizes item-user-item meta-paths to capture consumers' stable preferences. Experiments demonstrate significant performance improvements. The co-attention mechanism in our framework effectively enhances recommendations based on meta-path contexts compared to similar previous studies.