2013313990 이현성 졸업작품 제안서

19
Mimic Collaborative filtering without user’s behavior data 이이이 , 이이이 , Khawaji Faisal Mohammed

Upload: ita9naiwa

Post on 20-Mar-2017

60 views

Category:

Engineering


3 download

TRANSCRIPT

Page 1: 2013313990 이현성 졸업작품 제안서

Mimic Collaborative fil-tering without user’s be-

havior data

이현성 , 윤규원 , Khawaji Faisal Mohammed

Page 2: 2013313990 이현성 졸업작품 제안서

contents backgrounds

Collaborative Filtering Matrix Factorization

Model description Model structure Input

Page 3: 2013313990 이현성 졸업작품 제안서

Background

Page 4: 2013313990 이현성 졸업작품 제안서

Collaborative filtering From user’s behavior, calculate

similarity between users or be-tween items.

From that similarity, we can rec-ommend to user A

1. Items that User B who have high similarity with user A likes.

2. Items that is very similar with what user A likes.

Page 5: 2013313990 이현성 졸업작품 제안서

Example (cosine similarity)

Similarity between items

Similarity between users

Page 6: 2013313990 이현성 졸업작품 제안서

a user listens only a small fraction of songs.

One solution for this problem is to calculate similarity from factorized matrix, not from original song by user matrix.

Page 7: 2013313990 이현성 졸업작품 제안서

Matrix factorization

dimensionality reduction. SVD, or other methods can be used to achieve this.

Music by User Matrix(each elements are how many times user j

listened music i)Music by # Feature matrix

Matrix M Factorized Matrix F

Page 8: 2013313990 이현성 졸업작품 제안서

Matrix factorization

Page 9: 2013313990 이현성 졸업작품 제안서

Matrix factorization itself has good features. It reduces noise

Calculation become simpler with lower number of columns

Matrix factorization reveals latent(hidden) factors from user’s behav-ior patterns.

Page 10: 2013313990 이현성 졸업작품 제안서

Limitation of Matrix factorization(also collaborative filter-ing) It cannot recommend to newbies

It cannot recommend just released music

This is called the Cold Start Problem

Page 11: 2013313990 이현성 졸업작품 제안서

Our goal : alleviate Cold Start Prob-lem Develop a model transform song into factorized vec-

tors without user behavior data.What this song’s row vector in F will be if it….

Music by # Feature matrix

New music

Page 12: 2013313990 이현성 졸업작품 제안서

Assumption User preference to music is proportional to

Audio But not only audio!

Page 13: 2013313990 이현성 졸업작품 제안서

Very similar song, but different pref-erences

User’s preferences are affected not only audio itself, but also other aspects of music.

such as who created it, popularity of the singer, released recently…

Page 14: 2013313990 이현성 졸업작품 제안서

Model description

Merging(NN)

Output

CNN

MFCC features Other features available from tags

Neural network

Page 15: 2013313990 이현성 졸업작품 제안서

Input MFCC data extracted from music

audio 그밖에 더 넣을 정보는… ?

Metadata( 넣을 까 말까 생각하는 것들… ) Tags

Bag of word representation Popularity of music/artist.

We think Youtube view count is good to capture music’s popularity.

Page 16: 2013313990 이현성 졸업작품 제안서

MFCC

Page 17: 2013313990 이현성 졸업작품 제안서

Bag of words representation for tags There are other models like word2vec or

kind of…

For tags, bag of words representation is good because tags have no context infor-mation (i.e. order of words is not given)

Page 18: 2013313990 이현성 졸업작품 제안서

Example

Music A 신나는 여름음악 OST …

Music B 차분한 헤어졌을 때 … …

Tags 신나는 차분한 OST 여름음악 헤어졌을Music A 1 0 1 1 0

Music B 0 1 0 0 1

Page 19: 2013313990 이현성 졸업작품 제안서

Thanks.Any questions?