Optimization Of Online Music Recommendation Based On Serendipity Calculations

Until recently,most research on music recommendation system have focused on improving the accuracy of recommendation so as to recommend songs that best suit the musical tastes of users.However,recommender systems based on accuracy often suffer from the so-called “over-specialization” problem,recommending songs cater to user tastes too much,while excluding songs differ greatly from user profile and may be liked by users.User tastes are not static.Users may become bored with recommendation results that "accurately cater to their tastes" but "lack of novelty" and their satisfaction drops in the meantime.Some scholars are also worried that recommendations based on accuracy will shape users’ cognition and have a negative impact on user growth.In order to broaden the user’s interest boundary,alleviate the negative impact of over-specialization,and improve users’ satisfaction with recommendations,new research directions based on serendipity have emerged in the recommendation field.According to existing literature,we believe that serendipity includes three dimensions,namely relevance,novelty,and unexpectedness.A song recommendation list based on serendipity contains songs that users might like(relevant),did not expect to be recommended(unexpected),and have not seen before(novel).Based on the existing classic recommendation algorithms,we propose an optimization method for online music recommendation that considers the components of serendipity.Specifically,taking the KKBox music dataset as an example,we use the classic collaborative filtering algorithm ALS to generate an initial recommendation list based on accuracy for each user.In order to alleviate the problem of over-specialization and broaden the user’s interest boundary,we constructed two reranking methods with different purpose,increasing the preference for songs which are surprised and have rich information.We’d like to trade a small decrease in accuracy for a big increase in serendipity,so we determine the weight parameters through offline evaluation performance.In addition to offline evaluation,we also conduct a user study when evaluating our recommendation method.User study proved that our method could increase serendipity by59% with a minor cost of reducing the accuracy by 17%.At the same time,the user’s satisfaction with the recommendation results is 26% higher.Our work provides an optimization idea for online music recommendation based on serendipity.Also,we provide the platform with a feasible and low-cost reranking method based on serendipity.