Mutual proximity graphs for improved reachability in music recommendation.

Area of Science:

• Machine Learning
• Data Visualization
• Recommender Systems

Background:

• Hubness is a challenge in high-dimensional machine learning, leading to biased recommendations.
• In music recommendation systems, hubness causes popular items to be over-recommended and niche items to be ignored.
• This results in poor music catalogue reachability and limited user discovery.

Purpose of the Study:

• To investigate the impact of hubness on a k-nearest neighbour (knn) graph-based music recommendation system.
• To propose and evaluate mutual proximity graphs as an alternative to knn graphs for mitigating hubness.
• To assess the effectiveness of mutual proximity graphs in improving music catalogue reachability.

Main Methods:

• Utilized a real-world music recommendation system employing k-nearest neighbour (knn) graph visualization.
• Introduced and implemented mutual proximity graphs as a novel approach to graph construction.
• Compared the performance of mutual proximity graphs against knn graphs, mutual knn graphs, and mutual knn graphs with minimum spanning trees.

Main Results:

• Mutual proximity graphs significantly reduce the negative effects of hubness by avoiding highly connected hub vertices.
• The proposed graphs demonstrate superior graph connectivity compared to traditional knn-based approaches.
• Improved music catalogue reachability was observed with mutual proximity graphs, enhancing item discovery.

Conclusions:

• Mutual proximity graphs offer a robust solution to the hubness problem in high-dimensional recommender systems.
• This approach enhances the diversity and reachability of recommendations in music platforms.
• The findings suggest a promising direction for developing more effective and balanced recommender systems.