A Hybrid Recommendation Method with Multilayer Perception Applied on Real World Data
Liyang Bi, Yongji Wang, Dacheng Qu, Bei Guan · 2018
Collaborative filtering (CF) is an approach widely used in recommender systems (RS). Traditional CF-based methods simply utilize user-item rating matrix which implies interactions between users and items to make recommendation. However, rating matrix is often very sparse in real world, resulting in these methods a significant degradation in recommendation performance. Therefore, some employ side information of users and items to address the sparse problem. Nevertheless, when it comes to the interactions between user and item latent factors, they still utilize linear inner product. Neural Collaborative Filtering (NeuMF) is an appealing recent method employing Multilayer Perception (MLP) to learn interaction function between user and item latent factors. However, this method does not integrate side information of users and items. To tackle problems above, we generalize effective learning ability of multilayer perception and propose a hybrid recommendation method with multilayer perception which jointly learns deep representations of users and items from side information as well as rating matrix and interaction function between user and item latent factors. We also propose to add a data normalization layer after combining two vectors coming from different magnitudes. Extensive experiments on two real world datasets show that our model outperforms state-of-the-art algorithms.