Interactive evolution: interactive genetic algorithms for addressing popularity bias in music recommender systems

Abstract

Within recommender systems, there is a well-known bias called popularity bias, where these systems tend to recommend more popular items, over items in the so-called long-tail, which refers to the vast number of less popular items that collectively make up a significant portion of the total data. Popularity bias causes problems for both users and artists, such as limited exposure for niche or undiscovered artists and a lack of variety in users' music consumption habits. To address this issue, we developed an interactive genetic algorithm (IGA) for music recommendations, which evolves a population of recommendations based on user feedback. Our method improves on previous approaches by incorporating mutation, as well as dynamic crossover and mutation rates. We benchmarked our method against a previous approach using simulated users. Results show that our method shows similar feedback scores across all users as the benchmark. However, the convergence rate was higher, meaning optimal solutions were found more quickly. Moreover, our method improves feedback scores significantly for users with more niche interests, showing a 150.85\% improvement from the initial to the final generation, whereas the benchmark shows a statistically lower increase of 107.27\%. For users with traditional preferences, our system showed similar performance to the benchmark. The results suggest potential for the real-world applications of IGAs for music recommendations, as well as show the impact of incorporating mutation and dynamic genetic operation rates into IGAs.