Açiklanabilir tavsiye sistemleri

Recommender systems generally produce successful recommendations using rating data. However, recommendation systems that rely solely on rating data cannot capture users' feelings and thoughts in detail, so they may exhibit limited performance in some cases. To overcome this limitation, the use of textual data, especially user reviews, in recommender systems provides greater insight into individual preferences and sentiments. In this way, the accuracy and level of personalization of recommendations can be significantly increased. Explainability of recommender systems is another critical issue. Explaining the decision-making mechanism and the reasoning behind recommendations enhances the fairness and transparency of these systems, which increases trust in the system and helps reveal any biases that may be present in the system. This study introduces a novel recommendation algorithm built on user-generated comments. Language models determine the similarity between users' comments. Instead of relying on explicit user ratings, sentiment scores derived from these comments reflect users' opinions about products. Recommendations for the target user are generated by leveraging the sentiment scores of the most similar users. Ten different language models are employed to calculate similarity scores, and their performances are evaluated based on recommendation accuracy. We selected Sentence-BERT (SBERT), the best-performing language model, to calculate similarity scores for our algorithm. Using the attention mechanisms of text representation models and the SHAP method, the system provides reasons behind each recommendation, resulting in an explainable recommendation system.