A Music Recommendation System Based on Music Data
A Music Recommendation System Based on Music Data Grouping and User Interests Hung-Chen and Arbee L. P. Chen Department of Computer Science National Tsing Hua/Chengchi University
Outline • Introduction • Music Recommendation System – System Architecture • Recommendation Mechanisms • Experiments • Implementation
Introduction • For music recommendation system – a preliminary recommendation of the notification service • textual descriptions – our music recommendation system (MRS) • perceptual properties • The approaches for recommendation system – the content-based filtering approach – the collaborative filtering approach – the statistics approach
Architecture A polyphonic music object • one melody track • other accompaniment tracks
Track Selector • A polyphonic music object – one track for melody – the others tracks for accompaniment • The track for melody contains much more distinct notes. • The method proposed in [Uitd 98] – considers all tracks – chooses the notes with the highest pitch – may result in an extracted melody containing the notes which belong to the tracks of accompaniment.
Track Selector (Cont. ) • We use a measure of pitch density to select a representative track. • where NP is the number of distinct pitches in the track AP is the number of all distinct pitches in MIDI standard, i. e. , 128. • The track with the highest pitch density is selected as the representative track.
Feature Extractor • Six features are extracted from the perceptual properties of the representative track – Mean (MP) of the pitch values – Standard deviation (SP) of the pitch values – Pitch density (PD) – Pitch entropy (PE) – Tempo degree (TD) – Loudness (LD)
Feature Extractor (pitch entropy) where Pj is defined as follows: where Nj is the total number of notes with the corresponding pitch in the representative track T is the total number of notes in the representative track
Feature Extractor (tempo degree) • The tempo degree is defined as a ratio of the number of fast measures to the number of measures in the representative track. • A measure is a fast measure if the average note duration in the measure is shorter than one.
Classifier • There is an example of the classification using two features, PD and PE. • For each incoming music object, there are two situations to consider when performing the classification. – no music group exists in the database. – some music groups exist in the database. • Some music groups exist in the database. – The distances between the new feature point and each group centroid are computed. – The group with the minimum distance is selected for consideration. – There are two cases to consider.
Classifier (Cont. ) The new feature point falls into the area of a group (case 1).
Classifier (Cont. ) The new feature point does not fall into the area of any group (case 2).
Profile Manager Access History Access Time Object ID Music Group Transaction 2001/4/06 AM 11: 47: 03 1 B T 1 2001/4/06 AM 11: 47: 03 23 C T 1 2001/4/12 AM 10: 11: 25 7 D T 2 2001/4/12 AM 10: 11: 25 5 C T 2 2001/4/12 AM 10: 11: 25 32 B T 2 2001/4/16 AM 09: 51: 33 16 A T 3 2001/4/16 AM 09: 51: 33 19 B T 3 2001/4/16 AM 09: 51: 33 42 A T 3 2001/4/20 AM 08: 31: 12 31 D T 4 2001/4/20 AM 08: 31: 12 63 C T 4 2001/4/20 AM 08: 31: 12 26 A T 4 2001/4/22 AM 10: 24: 49 53 B T 5 2001/4/22 AM 10: 24: 49 12 A T 5
Recommendation Mechanisms • The CB method – based on content-based filtering approach • The COL method – based on collaborative filtering approach • The STA method – based on statistics approach
The CB Method • Content-based filtering approach – similarity between documents and profiles • The CB method is to recommend the music objects that belong to the music groups the user is recently interested in. • To capture the recent interesting music groups of the user according to group weights – time (TWj) – number (MOj, i)
The CB Method (Cont. ) • The group weight GWi of music group Gi – n is the number of latest transactions used for analysis – TWj is the weight of transaction Tj (assigned by the system) – Moj, i is the number of music objects which belong to music group Gi in transaction Tj • The number of music objects Ri from each music group for recommendation - N is the number of music objects in the recommendation list - GWi is the group weight of the target group - M is the total number of music groups
Access Time Object ID Music Group (Gi) Transaction (Tj) Transaction weight (TWj) 04/06 11: 47 1 B T 1 0. 4096 04/06 11: 47 23 C T 1 0. 4096 04/12 10: 11 7 D T 2 0. 512 04/12 10: 11 5 C T 2 0. 512 04/12 10: 11 32 B T 2 0. 512 04/16 09: 51 16 A T 3 0. 64 04/16 09: 51 19 B T 3 0. 64 04/16 09: 51 42 A T 3 0. 64 04/20 08: 31 31 D T 4 0. 8 04/20 08: 31 63 C T 4 0. 8 04/20 08: 31 26 A T 4 0. 8 04/22 10: 24 53 B T 5 1 04/22 10: 24 12 A T 5 1 Music Group (Gi) Group Weight (GWi) Number of Recommended Music Objects (Ri) A 3. 08 8 B 2. 5616 6 C 1. 7216 4 D 1. 312 4 A B Insert Time Object ID 03/11 11: 33 73 02/18 14: 25 29 Insert Time Object ID 03/15 09: 18 92 03/02 18: 03 66
The COL Method • Collaborative filtering approach – similarity between profiles • The COL method is to provide unexpected findings due to the information sharing between relevant users. – Derive user interests and behaviors as user profiles – Users with similar profiles will be identified as relevant users.
The COL Method (Cont. ) • Incremental Mining for User Interests – Interest Table • First and Count columns: support measure • Last column: data update Transactions T 1 {a, c, e} T 2 {b, c, e, f} T 3 {d, e, f} T 4 {b, c, d} a b c d e f Music Group First Last Count Support T 1 T 2 T 1 T 3 T 1 T 2 T 1 T 4 T 4 T 3 1 2 3 2 N/A 67% 75% 100% 75% 67% 7
The COL Method (Cont. ) • Mining Thresholds – Minimum support threshold =75% – Minimal count =2 – Expired time =4 Music Group Transactions T 1 {a, c, e} T 2 {b, c, e, f} T 3 {d, e, f} T 4 {b, c, d} T 5 {b, c, e, f, g} a b c d e f g First Last Count Support T 1 T 2 T 1 T 3 T 1 T 2 T 5 T 1 T 5 T 4 T 5 T 5 1 3 4 2 4 3 1 N/A 75% 80% 67% 80% 75% N/A 8
The COL Method (Cont. ) • Incremental Mining for User Behaviors – Behavior Table • 2 -groups set column • Minimum support threshold =60% Transactions T 1 {a, c, e} T 2 {b, c, e, f} T 3 {d, e, f} T 4 {b, c, d} 2 -Group Set [a, c] [a, e] [b, c] [b, d] [b, e] [b, f] First T 1 T 2 T 4 T 2 … Last T 1 T 4 T 2 Count 1 1 2 1 1 1 Support N/A 67% N/A N/A … 9
The COL Method (Cont. ) • User grouping by user interests and behaviors – Distance Measure • I-B matrix – number of dimension • I-B vector – User 1: 000000100110110 – User 2: 100100000010110 User 1 User 2 c e a e b c a c e f abcef a 00000 b 0100 c 110 e 11 f 0 abcef a 10010 b 0000 c 010 e 11 f 0 11
The COL Method (Cont. ) • After user grouping – The average group weights from other relevant users in the same user group • The average group weight indicates the average preference degree of the music group of other relevant users. – The weight differences between the average group weights and the user’s group weights. • The music group with positive weight difference indicates other users accessed the music group more often than the user himself.
Music Group Weight (GWj) Music Group Weight (GWj) A 3. 08 A 2. 824 A 2. 6896 B 2. 5616 B 2. 64 B 2. 024 C 1. 7216 C 0. 64 C 3. 2096 D 1. 312 D 0. 8 D 1. 8 E 0 E 1. 4096 E 0 F 0. 4096 F 0. 8 User X Music Group Number of Recommended Music Objects (Ri) C 2 E 10 F 8 User Y User Z Music Group Weight Difference Music Group Average Group Weight A -0. 3232 A 2. 7568 B -0. 2296 B 2. 332 C 0. 2032 C 1. 9248 D -0. 012 D 1. 3 E 0. 7048 F 0. 6048 Difference Table Average Table
The STA Method • long-term hot music group – The music group contains the most music objects in the access histories of all users. • short-term hot music group – The music group contains the most music objects in the latest five transactions of all access histories
Experiments • Effectiveness of the track selector – Select the representative tracks from 100 MIDI files by expert. – Apply our method on the same testing data set. – An 83% correctness rate is achieved.
Experiments (Cont. ) • Effectiveness of the feature selection • Pitch: Mean (MP), Standard Deviation (SP) Pitch Density (PD), Pitch Entropy (PE) • Duration: Tempo Degree (TD) • Loudness: Mean (LD) Feature MP SP PD PE TD LD Error Rate 65% 60% 56% 62% Feature Sets (PE, TD) (PE, LD) (TD, LD) (PE, TD, LD) Error Rate 43% 47% 39% 44%
Experiments (Cont. ) • Quality of recommendations Single Feature PE TD LD The CB Method 37% 39% 35% The COL Method 19% 23% 18% The STA Method 24% 27% 21% Feature Set (PE, TD) (PE, LD) (TD, LD) The CB Method 59% 51% 62% The COL Method 17% 23% 22% The STA Method 29% 31% 26%
Implementation
Implementation (Cont. )
Implementation (Cont. )
Issues • Music Classification for More Precise Recommendation • Collaborative Recommendation with User Feedback
- Slides: 32