A N APPROACH TO AUTOMATIC MUSIC PLAYLIST GENERATION USING I T UNES AND BEHAVIORAL DATA By Darrius Serrant, Undergraduate Supervised by Mitsunori Ogihara,

AN APPROACH TO AUTOMATIC MUSIC PLAYLIST GENERATION USING ITUNES AND BEHAVIORAL DATABy Darrius Serrant, Undergraduate

Supervised by Mitsunori Ogihara, PhD

CSC410: Computer Science Project Planning

AT A GLANCE

Motivation Automatic Playlist Generation Problem Related Work Scope of Project System Features Process Overview Testing and Evaluation

MOTIVATION

Music: food for the soul! Smorgasbord of expressions, emotions, and

representations Binds us to friends, memories, experiences, etc… Marketable, available and consumable

The typical music library 1,000+ titles Diverse in features Difficult to organize, explore, and experience

AUTOMATIC PLAYLIST GENERATION PROBLEM

Manual playlist creation Burdensome and time consuming Subjective

Automatic playlist creation: Create music playlists fulfilling arbitrary

requirements # of titles Permutation Measure of variety

An NP-hard problem

RELATED WORK

Scalable search algorithms1

Search algorithms based on skipping behavior2

Reduction to the traveling salesman problem3

Local search CSP algorithm4

Case-base approach to playlist generation5

Song selection via a network flow model6

The Music Genome Project7

RELATED WORK (CONTINUED)

Commonalities: Assumes limited knowledge of music library Assumes usage of audio feature extraction

techniques Requires explicit specification of playlist

constraints

SCOPE OF PROJECT

A unique approach to the automatic playlist generation problem Eliminates explicit user specifications Adapts to users’ listening preferences More expressive than audio features extraction

Research objectives Analyze contents of users’ music library Monitor and learn users’ listening habits Generate playlists of twelve songs by request

SYSTEM FEATURES

iTunes Library Data Extraction Extract music titles and their characteristics

Song Characteristics Aggregator Collect metadata from Internet sources

Machine Learning Statistically model users’ music listening habits

Playlist Generation Build a playlist from a “playlist” state space

SYSTEM FEATURES (CONTINUED)

User Feedback Evaluation of generated playlists Periodical mood assessments Software application monitoring

PROCESS OVERVIEW

PROCESS OVERVIEW

1. User listens to music through iTunes1. Monitor systems’ active processes2. Monitor local weather forecasts3. Receive user’s mood updates

2. User closes down iTunes3. Begin pre-playlist generation tasks

1. Collect data from user’s iTunes Music Library2. Collect data from Internet sources3. Update user’s listening pattern

PROCESS OVERVIEW (CONTINUED)

4. Automatically generate a new playlist1. Extract search heuristics from listening pattern.2. Build a new playlist from the search space.

5. User evaluates the generated playlist6. Incorporate user feedback into listening

pattern

TESTING AND EVALUATION

Phase One: Theoretical Testing Under simulated conditions Tasks:

Evaluate scalability of search algorithms Verify production of desired playlists for “naïve” users

Phase Two: Live Testing Deliver product to actual users Tasks:

Evaluate scalability of search algorithms for Mac and PC users

Verify production of desired playlists for “actual” users Test effects of volatile mood and environmental

changes on playlist generation.

CURRENT AND FUTURE WORK

Version 1.0 in development iTunes Data Extractor

Apache Xerces 2.7 XML Parser Data Collectors

Mood Collection System Process Collection Listening Pattern Assembly

Machine Learning Weka 3.6 Supervised Learning Algorithms Decision Tree Learning

Search Algorithms Breadth-first search Local beam search Genetic algorithm

CURRENT AND FUTURE WORK (CONTINUED)

Version 1.0 in development (continued) Data Storage

Oracle Berkeley DB Java Edition Testing

Theoretical testing Evaluation of developed search algorithms

Future Work International Symposium on Music Information

Retrieval The complete concept