Improvement of KNN Collaborative Filtering Model in User-based Approach on Anime Recommendation System

Abstract

This research aims to resolve the challenge of finding the list of recommendations that correspond to user preferences. The MyAnimeList dataset is utilized for model evaluation, accessible via Kaggle website. The outcome of this study is the development of a recommendation system based on the preferences of other users (user-based model). The suggested solution employs a collaborative filtering model based on the KNN algorithm and weighted attribute. The dataset consisted of 193,272 user ratings on anime, with the following attributes: username, anime_id, my_score, and my_status. As an extension of the KNN collaborative filtering paradigm, the rating value is weighted based on the user’s status. The determination of the weight is based on the responses of 105 respondents to a questionnaire. my_score and my_status values will be combined and adjusted using MinMaxNormalization in addition to being weighted. This work implemented the KNN algorithm with the following k parameter values: 3, 5, 9, 15, 23, 33, and 45. Variations in parameters are utilized to determine the optimal k value to employ in KNN, which uses the Pearson similarity matrix to calculate user similarity values. The model evaluation indicate that the optimal Mean Absolute Error and Root Mean Square Error values at parameter k = 5 are 0.14726 and 0.19855, respectively. This improved model’s findings further demonstrate that KNN collaborative filtering with an additional weighted parameter can predict ratings with stable and generally low error values for all k values.