Modeling Musical Influence Through Data

Abstract

Musical influence is a topic of interest and debate among critics, historians, and general listeners alike, yet to date there has been limited work done to tackle the subject in a quantitative way. In this thesis, we address the problem of modeling musical influence using a dataset of 143,625 audio files and a ground truth expert-curated network graph of artist-to-artist influence consisting of 16,704 artists scraped from AllMusic.com. We explore two audio content-based approaches to modeling influence: first, we take a topic modeling approach, specifically using the Document Influence Model (DIM) to infer artist-level influence on the evolution of musical topics. We find the artist influence measure derived from this model to correlate with the ground truth graph of artist influence. Second, we propose an approach for classifying artist-to-artist influence using siamese convolutional neural networks trained on mel-spectrogram representations of song audio. We find that this approach is promising, achieving an accuracy of 0.7 on a validation set, and we propose an algorithm using our trained siamese network model to rank influences.