报告题目：Towards Explainable Recommendation via the Exploration of Multimodal Data
报告人：程志勇(Postdoctoral Research Fellow, School of Computing, National University of Singapore, Singapore)
Latent factor models (e.g., matrix factorization) achieve good accuracy in rating prediction. However, they suffer from several problems including cold-start, non-transparency, and suboptimal recommendation for individual users.
In this talk, I will introduce our recent works on exploring multimodal data to alleviate these limitations. We employ user text reviews and item images together with the ratings to infer users’ preferences and items’ features on different aspects. To be specific, aspect-aware topic models are applied on text reviews and item images to capture user preferences and item features on different aspects with/without using well-defined aspect labels. The learned aspect preferences are integrated into a designed aspect-aware latent factor model to estimate aspect ratings for each user-item pair. To this end, our model captures a user’s preference on semantic aspects and estimates the importance and ratings of each aspect for an item. Therefore, our model could alleviate the data sparsity problem, gain good interpretability, and achieve more accurate prediction for local user-item pairs.
Comprehensive experimental studies have been conducted on 19 datasets from Amazon and Yelp 2017 Challenge dataset. Results show that our method achieves significant improvement compared with strong baseline methods, especially for users with only few ratings. Moreover, our model could provide reasonable explanation for recommendations.
Zhiyong Cheng is currently a postdoctoral research fellow with the School of Computing at the National University of Singapore. His research interests include information retrieval, recommender systems, multimedia and machine learning. His works have appeared in several top-tier conferences and journals, such as SIGIR, WWW, IJCAI, TOIS, and TKDE. He has served as program committee member of international conferences such as ACM MM and ICDM.