Abstract
Music recommendation systems often rely on emotion-based tags to categorize content. However, individual tags can be ambiguous. This paper proposes a method to rank pairs of tags based on their "acoustic similarity," identifying combinations that describe a cohesive sound. By calculating a utility metric derived from acoustic feature variance, the researchers demonstrate how specific tag pairs can reduce ambiguity and improve the precision of music categorization.
Introduction
Music Emotion Recognition (MER) is essential for modern streaming services, allowing users to search by mood (e.g., "happy" or "calming"). However, single tags often cover a broad range of acoustic profiles. For instance, "happy" might apply to both high-energy pop and gentle folk. Using multiple tags (e.g., "happy" and "upbeat") can refine results, but the vast number of possible combinations—the "curse of dimensionality"—makes manual selection impractical. This research introduces an automated ranking system to identify the most useful tag pairs for recommendation.
Methodology
The proposed system ranks tag pairs based on the acoustic similarity of the music they describe. The core idea is that a "high utility" tag pair should describe a group of songs that are more acoustically similar to each other than songs described by either tag individually.
Feature Extraction and Filtering
The researchers utilized 68 acoustic features, including dynamics, timbre (MFCCs), harmony, and rhythm. To ensure statistical significance, they filtered the initial pool of 15,000 tag pairs from the CAL500 dataset down to 1,000 pairs that each covered at least 10% of the music database.
The Utility Metric ($U(R)$)
The utility of a tag set $R$ is defined by its acoustic dissimilarity $D(R)$, measured as the average distance of songs from their centroid in the feature space. The utility metric is defined as the ratio of the pair's dissimilarity to the minimum dissimilarity of its individual tags:
$$U(R) = \frac{D(R)}{\min(D(T_1), D(T_2))}$$
- $U(R) < 1$: The pair is more specific (acoustically cohesive) than the individual tags, indicating high utility.
- $U(R) > 1$: The pair results in a more diverse, less cohesive set of sounds, indicating low utility.
Bridging Emotion Models
The paper connects two major psychological frameworks for music emotion: the Hevner Adjective List and Thayer’s Emotional Circumflex Model. While Thayer's 2D model (Arousal vs. Valence) is scientifically robust, it is less intuitive for general users than categorical adjectives. The proposed ranking method allows systems to leverage user-friendly adjectives while achieving the data-driven precision of dimensional models.
Results and Discussion
The study successfully identified tag pairs that significantly reduced acoustic ambiguity. For example, the pair "Not-Emotion-Calming-Soothing" and "Usage-Driving" achieved a utility score of 0.89, making it 11% more specific than its component tags. Conversely, some intuitive pairs failed; "Not-Emotion-Angry-Aggressive" combined with "Not-Emotion-Powerful-Strong" had a $U(R)$ of 1.11, increasing acoustic diversity.
These results highlight that simply adding more tags does not guarantee better recommendations. The relationship between the tags and the underlying audio signal is the critical factor in reducing ambiguity.
Conclusion
This research provides a quantitative framework for optimizing music tags. By using acoustic similarity as a proxy for semantic clarity, recommendation systems can suggest more precise mood categories. Future work involves scaling this method to larger datasets and exploring more complex tag combinations beyond simple pairs.
Summary updated