Title

EXPLAINABLE RECOMMENDATIONS USING EXTRACTED TOPICS FROM ITEM REVIEWS AND WORD MATCHING

Abstract

Abstract

Explanation in the recommendations is a crucial aspect in many applications to share reasoning and context with the users in addition to the recommended item. In this thesis, an innovative method for generating explainable recommendations is designed, implemented, and tested. The proposed design consists of extracting some phrases from the user's written review texts, assigning them to the users as preferences and items as their features, and then generating recommendations using the similarities between these assigned phrases. In such a design, since the recommendations are made using phrases that are understandable by people, the exact same phrases can be used to explain the reasoning behind the recommendations. Not many studies, however, uses keyword extraction techniques and word vectorizers to generate recommendations. Due to the lack of work in the area, it is decided to study such an algorithm that use keyword extraction and word vectorizers to uncover its capabilities. The expectation was to have a recommender that performs worse than the state-of-the-art deep learning models. Still, we were expecting to have a recommender that can generate sane results and also have the ability to explain its reasoning. To evaluate the proposed recommender design, alongside of calculating numerical results for the quality of the recommender, a user study with 15 people is conducted. These experiments showed that people like 55\% of the recommendations generated by the proposed method, while 58\% of the explanations for the recommended items are found meaningful.

Supervisor(s)

Supervisor(s)

MERT TUNC

Date and Location

Date and Location

2022-08-31 11:00:00

Category

Category

MSc_Thesis