Multimodal Variational Disentangled Knowledge Alignment for Cross-domain 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: Variational Encoding, 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: Multimodal recommendation systems have been widely used in e-commerce and short video platforms. Due to the large differences in data volume and data distribution in different business scenarios, cross-domain recommendation is studied to improve the effect of target domain by using rich source domain data. Some studies use encoders to represent domain information and design knowledge alignment to achieve cross-domain knowledge transfer. However, simple information representation and alignment methods are easily affected by noisy information and lead to negative transfer problems. The distribution of features in different domains also has a large deviation, which affects the effective transfer of knowledge. Therefore, we propose a Variational Disentangled Cross-domain Knowledge Alignment Method (VDKA) for multimodal recommendation. Specifically, we propose a variational multimodal graph attention encoder, which consists of variational autoencoder and graph attention encoder. Variational encoder can learn domain sharing and domain specific representations under multimodal data utilization. Then we introduce variational optimization objectives and disentangled representation objectives to improve the accuracy of domain representation. Furthermore, in order to solve the problem of domain knowledge distribution drift, adversarial learning is designed to realize cross-domain knowledge alignment. We conducted comprehensive experiments on four real-world multimodal data sets, and the experimental results show that our proposed VDKA method outperforms other state-of-the-art models. Ablation experiments have verified the effectiveness of our various designs.
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: 2505
Loading