Modeling Multi-factor and Multi-faceted Preferences over Sequential Networks for Next Item Recommendation

• Published in: Machine Learning and Knowledge Discovery in Databases. Research Track Vol. 12976; pp. 516 - 531
• Main Authors: Du, Yingpeng, Liu, Hongzhi, Wu, Zhonghai
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Multi-faceted preference; Multi-factor preference; Preference learning; Sequential networks; Sequential recommendation
• ISBN: 3030865193; 9783030865191
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-86520-7_32

Attributes of items carry useful information for accurate recommendations. Existing methods which tried to use items’ attributes relied on either 1) feature-level compression which may introduce much noise information of irrelevant attributes, or 2) item- and attribute- level transition modeling which ignored the mutual effects of multi-factor for users’ behaviors. In addition, these methods failed to capture multi-faceted preferences of users, therefore, the prediction for the next behavior may be affected or misled by the irrelevant facets of preferences. To address these problems, we propose a Sequential Network based Recommendation model, named SNR, to extract and utilize users’ multi-factor and multi-faceted preferences for next item recommendation. To model users’ multi-factor preferences, we organize the item- and attribute- level sequences of users’ behaviors as unified sequential networks, and propose an attentional gated Graph Convolutional Network model to explore the mutual effects of the preference factors contained in sequential networks. To capture users’ multi-faceted preferences, we propose a multi-faceted preference learning model to simulate the decision-making process of users with the Gumbel sotfmax trick. Finally, we fuse the multi-factor and multi-faceted preferences in a unified latent space for next item recommendation. Extensive experiments on four real-world data sets show that the proposed model SNR consistently outperforms several state-of-the-art methods.