Aspect-Aware Multi-Criteria Recommendation Model with Aspect Representation Learning

Multi-criteria recommendation system refers to a recommendation system that takes multi-criteria or multiaspect ratings into consideration when learning user preferences. The core of the multi-criteria recommendation system is predicting the criteria ratings leading to overall rating estimation. The existing approaches for multi-criteria recommendation system rely on either 1) using collaborative filtering to predict the criteria ratings based on historical criteria ratings, or 2) extracting latent aspects and predicting ratings from the review. The latter approach ignores the explicit ratings and the inferred ratings from the review may not be accurate. In the former approach, historical criteria ratings can be scarce (users may only provide ratings to a few criteria), which could affect the prediction accuracy. In this work, we introduce a novel multi-criteria recommendation model that predicts the criteria ratings from the review and then uses an aggregation function to estimate the overall rating. Multi-criteria ratings are predicted by fine-tuning BERT (Bidirectional Encoder Representations from Transformer) with an added aspect representation layer. Finally, the overall ratings are computed using a deep neural network-based aggregation function. The performance of our proposed model is evaluated with three different datasets including Tripadvisor, RateBeer, and BeerAdvocate. Our method outperforms several competitive baselines.