Reproduce “CIRP: Cross-Item Relational Pre-training for Multimodal Product Bundling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunshan Ma, Yingzhi He, Wenjun Zhong, Xiang Wang, Roger Zimmermann, Tat-Seng Chua
CIRP, a novel framework from the National University of Singapore, integrates explicit cross-item co-purchase relations directly into a multimodal encoder to learn rich item representations for product bundling. The method achieves over 25% relative improvement on product bundling tasks compared to strong baselines, while effectively handling cold-start items and enhancing pre-training efficiency through relation pruning.
Key insights
- Jointly optimizing Image-Text Contrastive (ITC) and Cross-Item Contrastive (CIC) losses is crucial, as CIC alone can degrade the model's semantic modeling capability if not balanced by ITC.
- Integrating explicit cross-item relations directly into the multimodal encoder enables the model to generate relation-aware representations, even for cold-start items without direct interaction history.
- The relation pruning module effectively removes noisy or redundant graph connections, improving relational data quality, enhancing performance, and drastically reducing pre-training convergence time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.