
Research Article
Hybrid CF on Modeling Feature Importance with Joint Denoising AutoEncoder and SVD++
@INPROCEEDINGS{10.1007/978-3-030-67537-0_18, author={Qing Yang and Heyong Li and Ya Zhou and Jingwei Zhang and Stelios Fuentes}, title={Hybrid CF on Modeling Feature Importance with Joint Denoising AutoEncoder and SVD++}, proceedings={Collaborative Computing: Networking, Applications and Worksharing. 16th EAI International Conference, CollaborateCom 2020, Shanghai, China, October 16--18, 2020, Proceedings, Part I}, proceedings_a={COLLABORATECOM}, year={2021}, month={1}, keywords={Denoising AutoEncoder Collaborative Filtering Attention unit}, doi={10.1007/978-3-030-67537-0_18} }
- Qing Yang
Heyong Li
Ya Zhou
Jingwei Zhang
Stelios Fuentes
Year: 2021
Hybrid CF on Modeling Feature Importance with Joint Denoising AutoEncoder and SVD++
COLLABORATECOM
Springer
DOI: 10.1007/978-3-030-67537-0_18
Abstract
AutoEncoder is an unsupervised learning approach that can maps inputs to useful intermediate features, which can be used to build recommendation. Intermediate features of different entities obtained by AutoEncoder may have different weight for predicting users behavior. However, existing research typically uses a uniform weight on intermediate features to make a fast learning algorithm, this general approach may lead to the limited performance of the model. In this paper, we proposes a novel approach by using SGD to dynamically learn the intermediate features importance, which can integrate the intermediate features into matrix factorization framework seamlessly. In the previous works, the entities intermediate features learned by AutoEncoder are modeled as a whole. On this basis, we proposes to use attention parameters in entity intermediate feature to dynamically learn the intermediate features importance and build fine-grained model. By learning unique attention unit for each entity intermediate feature, the entities intermediate features are integrated into the matrix factorization framework better. Extensive experiments conducted over two real-world datasets demonstrate our proposed approach outperforms the compared models.