Listener-Aware Music Recommendation from Sensor and Social Media Data

Music recommender systems are lately seeing a sharp increase in popularity due to many novel commercial music streaming services. Most systems, however, do not decently take their listeners into account when recommending music items. In this note, we summ

  • PDF / 109,369 Bytes
  • 5 Pages / 439.37 x 666.142 pts Page_size
  • 62 Downloads / 203 Views

DOWNLOAD

REPORT


Abstract. Music recommender systems are lately seeing a sharp increase in popularity due to many novel commercial music streaming services. Most systems, however, do not decently take their listeners into account when recommending music items. In this note, we summarize our recent work and report our latest findings on the topics of tailoring music recommendations to individual listeners and to groups of listeners sharing certain characteristics. We focus on two tasks: context-aware automatic playlist generation (also known as serial recommendation) using sensor data and music artist recommendation using social media data.

1

Introduction

The importance of incorporating user characteristics and contextual aspects into recommender systems has been acknowledged many times [1,16,17]. Research that looks into this matter in the domain of music recommendation is scarce, though. Addressing this issue, we summarize our latest work on the tasks of (i) automatic music playlist generation incorporating contextual aspects of the listener and (ii) music artist recommendation tailored according to various user and listening characteristics. Both tasks are highly related to machine learning and data mining. In fact, we approach the former task by gathering a wide variety of listener-centric sensor data from a smart phone app and exploiting machine learning techniques to learn relationships between these features and music metadata (e.g. artist or track name). The latter task is related to data mining as we acquire and analyze huge amounts of listening events produced by users of social media and build recommendation algorithms that consider personal characteristics of the listeners and their listening behavior, also using novel features mined from user-generated data.

2

Automatic Playlist Adaptation Based on Sensor Data

Addressing the task of automatic playlist generation, we developed an Android app for smart devices, dubbed “Mobile Music Genius” (MMG) [6,8], which collects a variety of user-specific features during playback, ranging from time, location, and weather to ambient noise, light level, and motion. In addition, MMG gathers music metadata (artist, track, mood, and genre) and records c Springer International Publishing Switzerland 2015  A. Bifet et al. (Eds.): ECML PKDD 2015, Part III, LNAI 9286, pp. 213–217, 2015. DOI: 10.1007/978-3-s319-23461-8 16

214

M. Schedl

player interaction (e.g. play, skip, pause events). Integrating a decision tree algorithm that is trained and retrained on the contextual features using track names as classes, MMG continuously monitors the context feature values of the listener and uses the classifier to suggest tracks suited to a given context, whenever the changes in context features exceed a threshold. These tracks are subsequently inserted into the playlist after the currently played one. During a pilot study, we collected 7,628 data points (context features, music metadata, and interaction data) created by 48 students at JKU Linz. Based on this dataset, we investigated a variety o