Recommending Composite Items Using Multi-Level Preference Information: A Joint Interaction Modeling Approach

TL;DR

Recommender systems struggle with composite-item suggestions where user preferences for outfits may diverge from preferences for individual items. The authors propose JIMA, a neural, multi-task framework that jointly models latent factors across $Y^{(1)}$ (user-top-bottom outfits), $Y^{(2)}$ (user-top), $Y^{(3)}$ (user-bottom), and $Y^{(4)}$ (top-bottom fit), while incorporating two-way and higher-order interactions to capture personalized and domain-expert compatibility. Through simulations and offline/online experiments in fashion, JIMA consistently outperforms baselines, with ablations confirming the value of both joint modeling and interaction terms. The work demonstrates that a single unified model can predict preferences at multiple granularities and enable recommendations for outfits and individual items even when data for one level is sparse. This approach has practical impact for fashion platforms and can generalize to other domains requiring multi-level, contextually aware composite-item recommendations.

Abstract

With the advancement of machine learning and artificial intelligence technologies, recommender systems have been increasingly used across a vast variety of platforms to efficiently and effectively match users with items. As application contexts become more diverse and complex, there is a growing need for more sophisticated recommendation techniques. One example is the composite item (for example, fashion outfit) recommendation where multiple levels of user preference information might be available and relevant. In this study, we propose JIMA, a joint interaction modeling approach that uses a single model to take advantage of all data from different levels of granularity and incorporate interactions to learn the complex relationships among lower-order (atomic item) and higher-order (composite item) user preferences as well as domain expertise (e.g., on the stylistic fit). We comprehensively evaluate the proposed method and compare it with advanced baselines through multiple simulation studies as well as with real data in both offline and online settings. The results consistently demonstrate the superior performance of the proposed approach.

Figures (4)

• Figure 1: The overview of the proposed method. LF stands for latent factors, and $M$ is the number of fully connected layers in the feedforward neural networks.
• Figure 2: Example interface for recommendation rating using different techniques
• Figure A1: Example interface for rating the presented individual clothing items
• Figure A2: Example interface for rating the presented outfits