Reproduce “Semi-supervised Adversarial Learning for Complementary Item Recommendation” on OpenResearch

About this paper

Koby Bibas, Oren Sar Shalom, Dietmar Jannach

Researchers from Meta, Amazon, and the University of Klagenfurt present ALCIR, a semi-supervised adversarial learning model for complementary item recommendations, specifically designed for cold items lacking interaction data. ALCIR leverages item side-information and a CycleGAN-inspired architecture to learn category-specific item representations, achieving superior accuracy and catalog coverage compared to established baselines on Amazon e-commerce datasets.

Key insights

• Maintaining distinct latent spaces for each item category and learning explicit translation functions between them is effective for generating contextually relevant complementary recommendations.
• A CycleGAN-inspired adversarial component, with cycle consistency and a category classifier, significantly enhances semi-supervised learning by improving item representation translation and preserving important item traits across category latent spaces.
• Explicitly defining target categories for complementary items improves recommendation precision, reduces inference time, and aligns with practical e-commerce strategies for product bundling.

Summary by alphaXiv.