Organization of Electronic Dance Music by Dimensionality Reduction

Abstract: This thesis aims to produce a similarity metric for tracks of the genre: Electronic Dance Music, by taking a high-dimensional data representation of each track and then project it to a low-dimensional embedded space (2D and 3D) by applying two Dimensionality Reduction (DR) techniques called t-distributed stochastic neighbor embedding (t-SNE) and Pairwise Controlled Manifold Approximation (PaCMAP). A content-based approach is taken to identify similarity, which is defined as the distances between points in the embedded space. This work strives to explore the connection between the extractable content and the feel of a track. Features are extracted from every track over a 30 second window with Digital Signal Processing tools. Three evaluation methods were conducted with the purpose of establishing ground truth in the data. The first evaluation method established expected similarity sub clusters and tuned the DR techniques until the expected clusters appeared in the visualisations of the embedded space. The second evaluation method attempted to generate new tracks with a controlled level of separation by applying various distortion techniques with increasing magnitude to copies of a track. The third evaluation method introduces a data set with annotated scores on valence and arousal values of music snippets which was used to train estimators that was used to estimate the feeling of tracks and to perform classification. Lastly, a similarity metric was computed based on distances in the embedded space. Findings suggest that certain contextual groups such as remixes and tracks by the same artist, can be identified with this metric and that tracks with small distortions (similar tracks) are located more closely in the embedded space than tracks with large distortions.