Information-theoretic measures of music listening behaviour

Abstract

We present an information-theoretic approach to the mea- surement of users’ music listening behaviour and selection of music features. Existing ethnographic studies of mu- sic use have guided the design of music retrieval systems however are typically qualitative and exploratory in nature. We introduce the SPUD dataset, comprising 10, 000 hand- made playlists, with user and audio stream metadata. With this, we illustrate the use of entropy for analysing music listening behaviour, e.g. identifying when a user changed music retrieval system. We then develop an approach to identifying music features that reflect users’ criteria for playlist curation, rejecting features that are independent of user behaviour. The dataset and the code used to produce it are made available. The techniques described support a quantitative yet user-centred approach to the evaluation of music features and retrieval systems, without assuming objective ground truth labels.