Portal de Periódicos da CAPES

NASM: Nonlinearly Attentive Similarity Model for Recommendation System via Locally Attentive Embedding

2019; Institute of Electrical and Electronics Engineers; Volume: 7; Linguagem: Inglês

10.1109/access.2019.2916938

2169-3536

Zhenpei Shan, Yunqi Lei, Defu Zhang, Jian Zhou,

Expert finding and Q&A systems

Recommendation system, as a core service for many customer-oriented online services, is employed to predict the personalized rating of users on their potentially preferable items. In modern industrial settings, an item-based collaborative filtering (item-based CF) method has been long popular owing to its excellent interpretability and high efficiency in the real-time personalized recommendation. In this model, the current target item is recommended according to the interacted similarity from the user's profile, which implies that the key of item-based CF is in the estimation of historical item similarity. Early studies usually utilize statistical measures including cosine similarity and Pearson correlation coefficient to estimate similarity with low accuracy caused by lacking optimization tailed. Recently, there are some learning-based models attempting to learn item similarity by optimizing a recommendation-aware loss function. However, these efforts are mainly concentrated on the application of the shallow linear model, and relative works that deploy some deep learning components for item-based CF are scarce. In this paper, we propose a nonlinearly attentive similarity model (NASM) for item-based CF via locally attentive embedding by introducing local attention and novel nonlinear attention to capture local and global item information, simultaneously. The NASM is based on a neural attentive item similarity (NAIS) model and further achieves significantly superior performance. The experimental results demonstrate that the NASM achieves more competitive recommendation performance in terms of hit ration (HR) and the normalized discounted cumulative gain (NDGC) in comparison with other state-of-the-art recommendation models.