Multi-interest Disentangled Representation Learning for Multimodal Recommendation
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Multiple Interests, Disentangled Representation, Multimodal Recommendation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In recent years, multimodal recommendation systems have been widely used in e-commerce and short video platforms. How to effectively utilize multimodal data and avoid the interference of multimodal noise information has become the key research direction of researchers. Many studies add multimodal data as auxiliary features to the model, which brings positive effects. Pictures, text and audio signals in short videos are more likely to attract users' interest than basic attributes. The user's multiple personalized interests largely determine the user's behavioral preferences. In order to effectively utilize user interest to improve model effect, We propose a new Multi-interest Disentangled Representation Learning method for multimodal recommendation (MIDR). Specifically, we first introduce the expected maximum to describe the relationship between interest and predicted target, and establish the optimization object based on multi-interest recommendation. Then, considering the relationship between user interest and multiple modalities, we introduce disentangled representation learning to extract modal sharing and modal specific interest representations. Furthermore, we introduce multi-interest contrast module to help model learning interest representation based on self-supervised learning. We conducted experiments on three real-world data sets, and our proposed MIDR outperformed other state-of-art models. The effectiveness of the disentangled interest representation module and interest contrast module was verified by the ablation experiment.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2495
Loading