Improving Recommender Systems Performance with Cross-domain Scenario: Anime and Manga Domain Studies

Recommendation systems have been an integral part of supporting various applications used in daily life by presenting information according to users' preferences and solving information overload problems. Most of the time, users only interact with a small portion of information in the system, which leads to sparsity in the user and item interaction matrix. This sparsity leads to performance degradation of the recommenders. Cross-domain recommendation system is one technique that can be used to alleviate the impact of sparsity. It works by transferring useful knowledge from additional domains that have some relation to users or items in the system. However, its implementation depends on several aspects that are not commonly fulfilled by common datasets that are available. This study tries to implement cross-domain recommendation techniques in a new domain of anime and manga. These domains have close relationships and support the use of cross-domain recommendations. This study compares performance in single and cross-domain scenarios, along with sensitivity analysis. The use of cross-domain scenarios for manga recommendations resulted in a decrease in RMSE of 0.075 and relatively stable accuracy above 90% even at the sparsity level of 0.01%. The result shows that cross-domain recommenders have lower RMSE and better accuracy than single-domain recommenders in cases when anime domains are used as auxiliary domains for manga recommendation, but not otherwise.