Aspect-level item recommendation based on user reviews with variational autoencoders

• Published in: Information sciences 2024-06, Vol.671, p.120655, Article 120655
• Main Authors: Ou, Wei, Huynh, Van-Nam
• Format: Article
• Language: English
• Subjects: Aspect-based recommender systems; Aspect-level sentiment prediction; Item ranking; Variational autoencoder
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2024.120655

In this paper we propose an aspect-based recommendation model based on variational autoencoders, that provides not only coarse predictions about what items users may like, but also finer-grained predictions of specific aspects users may be interested in of the recommended items. The proposed model first employs convolution operations to learn probability distributions of aspect-level embeddings for users and items from user aspect-level sentiments on seen items. Then it samples from these distributions and feeds the samples into transpose convolution operations to ‘reconstruct’ missing aspect-level sentiments for unseen items, thereby serving as signals for generating recommendations. To prevent overfitting, we impose a prior on the representation distributions and penalize the model if the learned distributions diverge from the prior. Based on the output of the proposed model, we propose a two-stage ranking scheme that combines aspect-level and overall sentiment signals to rank items. Experiment results show that the proposed model outperforms state-of-the-art aspect-based recommendation models, and the two-stage ranking scheme improves the traditional ranking by overall sentiment predictions.