Introduction
Musical similarity is notoriously difficult to define objectively. Whether two songs are considered "similar" often depends on the context and the specific listener—one person might focus on a driving rhythm, while another prioritizes melodic contour or harmonic structure. Most music retrieval systems utilize fixed weightings for these features, which fails to capture this inherent subjectivity. This paper proposes a system that allows users to define similarity dynamically through feedback and introduces a rigorous experimental design, Balanced Complete Block Partitioning (BCBP), to calibrate these measures using human studies.
Dynamic Aspect Weighting
To accommodate the contextual nature of music, the authors' system calculates multiple independent distance measures ($D_1, D_2, \dots, D_n$) for various musical features. Instead of requiring users to understand technical music theory, the system employs an intuitive interface based on positive and negative examples.
When a user adds tracks to a "positive" playlist, the system identifies which distance measures are smallest for those tracks and increases their relative weights. If a user marks a suggested song as "negative," the system further refines these weights by seeking measures where positive tracks remain close to each other but are distant from the negative example. This iterative feedback loop allows the algorithm to "learn" the user's specific criteria for similarity in real-time.
Balanced Complete Block Partitioning (BCBP)
A significant challenge in musical similarity research is the "Tversky effect," a psychological phenomenon where the order in which songs are heard biases a listener's judgment. Furthermore, as the number of songs ($n$) in a study increases, the number of possible triadic comparisons (asking which two of three songs are most similar) grows at an $O(n^3)$ rate, making exhaustive testing by a single individual impossible.
The authors developed Balanced Complete Block Partitioning (BCBP) to address these issues. BCBP ensures that every permutation of a triad is tested to neutralize order bias. To manage the massive volume of comparisons, a priority-queue-based algorithm partitions the complete block design into manageable segments for multiple participants while maintaining perfect statistical balance across pieces and positions.
Datasets and Insights
The methodology is applied to a large-scale study involving a corpus of over 14,000 MIDI-encoded western pop songs and more than 40 distinct distance measures (melodic, rhythmic, harmonic, and metadata-based). By using a "find more like these" interface, the system bypasses the need for a specialized musical vocabulary, making advanced retrieval accessible to non-experts.
Results and Future Work
The primary result of this work is the robust methodological framework for combining normalized distance measures with user feedback. While the current model assumes a linear combination of features, future research will explore non-linear perceptions of similarity. Additionally, the authors aim to use the results of their ongoing large-scale study to determine if users naturally cluster into specific profiles, such as those who consistently prioritize rhythm over melody.
Summary updated