Personalized Review-Oriented Explanations For Recommender Systems

Explainable recommender systems aim to provide clear interpretations to a user regarding the recommended list of items. The explanations present different formats to justify the recommended list of items such as images, graphs or text. We propose to use review-oriented explanations to help users in their decision since we can find crucial detailed feature in the reviews given by users. The model uses advances of natural language processing and incorporates the helpfulness score given in previous reviews to explain the recommended list of items provided by a latent factor model prediction. We conducted empirical experiments in the Yelp and Amazon datasets, proving that our model improves the quality of the explanations. The model outperforms baselines models by 13% for NDCG@5, 83% for HitRatio@5, 13% for NDCG@10, and 55% for HitRatio@10 in the Yelp dataset. For the Amazon dataset, the observed improvement was 9% for NDCG@5, 83% for HitRatio@5, 9% for NDCG@10, and 22% for HitRatio@10.