Enhancing Book Recommendation with the use of Reviews

Abstract

Recommendation systems are a major component in current e-commerce websites and applications. There are many studies carried out to ensure that the best recommendations are provided to the user and conversion rate is increased. These techniques usually utilize historical transaction data and user ratings. While most of such websites also provide the capability to review the products bought by the users, the content of these reviews usually does not play a major role in recommendations made to the users. Goodreads is the world’s largest website for readers and book recommendations. A user can keep track of their reading as well as review, rate and recommend books to other users. Book recommendations are also made automatically by Goodreads based on the books a user has already read and rated. As a review is much more expressive than a single rating and tend to explain the user’s decision for a rating, it is reasonable to expect that incorporating reviews will improve the recommendation process. This study attempts to address this by combining sentiment analysis of the user reviews with the recommendation process in Goodreads. To achieve the above goal, the constructed recommender system utilizes LightFM, a Python library facilitating popular recommendation algorithms for implicit and explicit feedback. LightFM enables item and user metadata to be incorporated into traditional matrix factorization algorithms. The item metadata that is utilized in this scenario are the sentiment scores obtained through user reviews of each book. The above recommender system performs better than pure collaborative filtering algorithms such as k-nearest neighbor and SVD for the same Goodreads dataset as evinced by the better AUC score of the LightFM model.