Music Recommendation A Data Mining Approach

1
Music RecommendationA Data Mining Approach

• Daniel McEnnis
• 2nd year PhD

2
Overview

• High level overview
• Toolkit Improvements
• Experiments
• Evaluation
• Algorithms research
• Data
• Future work

3
Project Goals

• Integrate social information
• Make algorithms culturally aware
• Implement existing algorithms
• Systematic evaluation framework

4
Similarity Algorithms

• Create new relations based on some aspect of
  similarity
• 6 different varieties of similarity
• Each algorithm can use one of 6 distance functions

5
Aggregator Algorithms

• Takes data from one set of actors and moves it to
  another
• 6 different varierties
• Each variety uses one of 7 aggregator functions
• Basic building block of Graph-RAT applications

6
Graph Triples Census

• Probable novel algorithm
• Proof of Correctness Completed
• Proof of Time Complexity Completed
• Literature review in progress

7
SUCCESS!

• Graph-RAT programming language now functioning
• Graph-RAT integrates social, cultural, personal,
  and audio data into algorithms
• Includes most commercial algorithms
• Contains primitives for existing academic systems
• Evaluation is entirely automated

8
PROBLEMS
9
Evaluation Exploration

• 9 types of music recommendation
• Personalized versus generic
• Open query versus targeted query
• Dynamic versus static data
• New music versus all music

10
Personalized Radio

• Open query with personalized presentation
• Static data vs dynamic data
• New items prediction vs predict anything

11
Targeted Search

• Not personalized
• Similarity queries
• Automatically generating targeted lists for a
  browsing hierarchy
• New music vs all music
• Static vs dynamic data

12
Personalized Tag Radio

• Create a personalized play list matching a given
  query
• New music vs all music
• Static vs dynamic data

13
Excluded Types

• Top 40 prediction
• Rendered obsolete by other types

14
Existing Algorithms

• Item-to-Item collaborative filtering
• 7 variations
• User-to-user collaborative filtering
• 7 variations
• Associative mining collaborative filtering
• Direct machine learning playlist data
• Direct machine learning audio data

15
Novel Algorithms

• Machine learning over profile data
• Machine learning over cultural and profile data
• Machine learning on different concatenations
• Audio
• Playlist
• Profile
• Cultural

16
Initial Data

• LiveJournal
• Separating music data is difficult
• No tag info or audio content
• No enough musical data
• LastFM by User
• No audio content
• Data cleaning is an issue

17
Current Data

• 40s Jazz Recordings
• 1800 annotated recordings from 70 CDs
• Covers nearly all 40s popular music
• LastFM by Song
• Retrieves tag and user info by song
• Data cleaning on user playcounts needed

18
Data Cleaning Tags

• Polysemy
• Synonomy
• Disjoint
• Hypersomny
• Hyposomny
• Initial algorithms developed

19
Future Work Programming

• Radically different programming environment
• SQL
• LINQ library package in C

20
Future Work Scalability

• Distributed SQL database implementation
• Just-in-time compilation
• Event-based recalculation of algorithm results
• Parallel execution of algorithms
• Multi-threaded algorithms