Content-Based Music Retrieval Using Query Integration for Users with Diverse Preferences

Keiichiro Hoashi, Hiromi Ishizaki, Kazunori Matsumoto, Fumiaki SugayaIsmir 2007

Abstract

Standard content-based music retrieval systems often struggle to accommodate users with eclectic tastes. By averaging multiple preferred songs into a single user profile, these systems frequently fail to capture distinct musical interests, leading to poor recommendation quality. This paper proposes Query Integration (QI), a method that employs hierarchical clustering to dynamically generate multiple query vectors from a user's sample set. Experimental results demonstrate that QI significantly improves retrieval accuracy, particularly for users with highly diverse preferences, while maintaining computational efficiency through Cluster Selection (CS) techniques.

Introduction: The Averaging Problem

Content-based Music Information Retrieval (MIR) systems typically represent a user's preference by aggregating a "sample set" of liked songs into a single vector. While computationally efficient, this approach suffers from a fundamental flaw: the averaging problem. For instance, if a user enjoys both aggressive heavy metal and solo piano, the vector average of these genres often lands in a mathematical "no man's land," resulting in recommendations for mid-tempo pop or soft rock—genres the user may not enjoy.

Proposed Method: Query Integration (QI)

To address this, the authors introduce Query Integration (QI). Instead of forcing a single average, QI uses hierarchical clustering to group similar songs within a user's sample set into distinct clusters.

Figure 1: Conceptual diagram of Query Integration (QI) clustering song vectors into integrated queries.

As shown in Figure 1, song vectors ($S_1$ to $S_5$) are grouped based on a similarity threshold ($\tau_q$). Songs that exceed this threshold are merged into an "integrated query" vector. This process continues until no further merges are possible, resulting in a small set of queries that accurately represent the distinct facets of a user's taste. To manage the computational cost of multiple queries, the authors also implement Cluster Selection (CS), which restricts the search to the most relevant database clusters using a vector quantizer (VQ) tree.

Evaluation and Results

The system was tested on a collection of 6,863 songs. Performance was measured by the ratio of "Good" songs ($C_g$, rated 4 or 5 by human subjects) appearing in the top 50 retrieval results.

Handling Diverse Preferences

The most significant advantage of QI is its performance for users with diverse tastes. The study categorized users based on their "Average Query Similarity" (AvgQSim) into LowSim (diverse), MidSim, and HiSim (consistent) groups.

Figure 4: Performance comparison ($C_g$ ratio) across different query similarity classes.

As illustrated in Figure 4, the conventional averaging method (Conv) performs worst for the LowSim group (0.562). In contrast, QI (at $\alpha = 1.0$) boosts accuracy to 0.715 for these users. This confirms that dynamic query generation effectively captures the nuances of eclectic listening habits.

Efficiency vs. Accuracy

While treating every song as an individual query (No-QI) yields the highest accuracy, it increases computational load (CalcRate) by 500%. Query Integration achieves comparable accuracy at a much lower cost. Furthermore, by applying Cluster Selection (CS), the CalcRate can be reduced to as low as 31.3% while still outperforming the conventional method's accuracy.

Conclusion

Query Integration offers a robust solution for personalized music retrieval by acknowledging and preserving the diversity of user preferences. By moving away from simplistic averaging and toward dynamic, cluster-based querying, MIR systems can provide more relevant recommendations for eclectic listeners without sacrificing system performance.

Summary updated

BibTeX
@inproceedings{hoashi2007content,
  author = {Keiichiro Hoashi and Hiromi Ishizaki and Kazunori Matsumoto and Fumiaki Sugaya},
  title = {{Content-Based Music Retrieval Using Query Integration for Users with Diverse Preferences}},
  booktitle = {Ismir},
  year = {2007},
}
Figure