νμ
(μνΈνλ) νν°λ§, Recommender Systemμ΄λΌκ³ λ λΆλ¦Ό. ProjectPrometheusμμ μ¬μ©νλ€.
Approaches to Collaborative Filtering ¶
problem spaceκ° 2μ°¨μ matrix μ ννλ₯Ό μκ°ν΄λ³Έλ€. νμ λν΄μλ itemμ, μ΄μ λν΄μλ userλ₯Ό λκ³ , κ·Έμ λ°λ₯Έ rating μ κ°μΌλ‘ λλ€. μ΄ matrix λ₯Ό μ΄μ©, CollaborativeFiltering μ νΉμ μ¬μ©μ(user) i μ λν΄μ rating μ μμΈ‘νκ³ , item λ€μ μΆμ²νλ€.
λ³΄ν΅ λ€μμ κ³Όμ μ κ°μ§λ€.
- userμΈ i μ λΉμ·ν ν₯λ―Έλ μ νΈλλ₯Ό κ°μ§λ μ¬μ©μ μ§ν©(user set)μ μ ννλ€.
- 1 μμ μ νλ user set μΌλ‘λΆν° user i μκ² μ ν©ν μΆμ² item μ μμΈ‘ν΄λΈλ€.
ex) μ΄ user set μμ item j μ λν΄μ λμ μ μ (rating)μ μ£Όμμ κ²½μ°, user i μκ² item j λ₯Ό μΆμ²νλ€.
Algorithms ¶
- νμ¬ μ΄μ©μ€μΈ user μ λΉμ·ν μ·¨ν₯μ μ¬μ©μ μ§ν©μ μ ν - calculate user correlation
- 1μμ μ νλ μ¬μ©μ μ§ν©μ€ μμΈ‘μ μν λΆλΆμ§ν© μ ν - neighbourhood selection
- μ νλ neighbours λ€κ³Ό μλ£λ₯Ό κ·Όκ±°λ‘ μμΈ‘ - generate a prediction
Calculate user correlation ¶
- Pearson correlation
- Constrained Pearson correlation
- The Spearman rank correlation
- The Vector similarity
- Emtropy-based uncertainty measure
- Mean-square difference algorithm
Neighbourhood Selection ¶
- Correlation thresholding
- Best-n correlations
Generate a prediction ¶
Metrics ¶
CollaborativeFiltering μ μ μ©μ±μ νκ°νλ κΈ°μ€.
- Coverage - μμ€ν
μ΄ μΆμ΄μ§ itemμΌλ‘λΆν° μΆμ² item μ μ 곡ν΄μ£Όλ λ₯λ ₯. (μλ§λ μΆμ²itemνλͺ©μ/μ 체itemνλͺ©μ μ λν νΌμΌνΈμ΄λ €λ. μμ νμ)
- Accuracy - μμ€ν
μ΄ μΆμ²ν item μ λν μ νμ± (μΆμ² itemμ λν μ¬μ©μμ μ μ)
κ΄λ ¨ μλ£λ€ ¶
κ°λ‘
ꡬν μμ€ν μ/λ Όλ¬Έ
PrincipiaCyberneticaμ μλ μμ£Ό κ°λ¨ν κ°λ‘ (μ²μ 보λ μ¬λμκ² μΆμ²) http://pespmc1.vub.ac.be/COLLFILT.html
- Overview on various CF algorithms (recommended) http://www.research.microsoft.com/users/breese/cfalgs.html
- νκΈ κ°λ‘
- http://www.ecminer.com/m3_webBrain.html
- CACM 1997 Mar
- http://citeseer.nj.nec.com/483304.html
- Personalization on the Web
ꡬν μμ€ν μ/λ Όλ¬Έ