![ALS-仂仗亳仄亳亰舒亳 仄仂亟亠仍亳 SVD
仂亟亠仍:
Tqi
ruui =亮 + bu + bi + pu
argmin
p*q*b*
裡 2
( r亮 b b pTq)
uui u i u
i (u,i )R
2 ( )
+了 pu
2
+ qi
2
+ bu
2 + bi
个仆从亳 仂亳弍从亳:
P-step:
pu = 了nuI + Au ( )1 du
Au =Q[u]TQ[u] = qiqi
T
裡
i:(u,i)R
裡
d =Q[u]T ru = ruiqi
i:(u,i)R
Q-step:
qi = 了niI + Ai ( )1 di
Ai = P[i]T P[i] = pupu
T
裡
u:(u,i)R
裡
di = P[i]T ri = rui pu
u:(u,i)R](https://image.slidesharecdn.com/lecture1-141018015403-conversion-gate01/85/ITMO-RecSys-course-Autumn2014-Lecture1-30-320.jpg)
ITMO RecSys course. Autumn2014. Lecture1
- 2. 亠从仂仄亠仆亟舒亠仍仆亠 亳亠仄 亠从亳 1: 于于亠亟亠仆亳亠 仆亟亠亶 舒仆亳仍亠仆从仂 18 仂从磡 2014
- 3. 弌从舒 Introduction Collaborative filtering Content-based & hybrid methods Evaluation
- 4. Recommender Systems (RSs) are software tools and techniques providing suggestions for items to be of use to a user F. Ricci Introduction
- 5. 仂仍亳亠于仂 舒亠亶 于 仂弍仍舒亳 RS 仗仂 亟舒仆仆仄 google scholar (仂 2014-10-17)
- 6. 亢亳于亠仄 于 仗仂 亠从仂仄亠仆亟舒亠仍仆 亳亠仄!
- 7. 仍舒亳亳从舒亳 RS Available data Tags & Metadata User history Content Collaborative Content-based Hybrid
- 8. 舒仆仆亠 亠亶亳仆亞亳 (explicit feedback) 丕仆舒仆亠 (like) 亳仆舒仆亠 (like/dislike) 丼亳仍仂于亠 (stars) 仂亳 亟亠亶于亳亶 (implicit feedback) 丐亠亞亳, 仄亠舒亟舒仆仆亠 ~ 亰于 亰 (community-based RS)
- 9. 仂舒仆仂于从舒 亰舒亟舒 RS Predict Recommend Similar
- 10. Collaborative filtering Neighborhood methods Matrix factorization methods
- 11. Collaborative filtering Neighborhood methods
- 12. 亟亠 仄亠仂亟舒 (user-based) 舒从 仗仂亟从 仂亠仆亳仍亳 仗仂仂亢亳亠 仗仂仍亰仂于舒亠仍亳? rui = 1 Ni (u) rvi 裡 vNi (u) 亰于亠亳仄 于从仍舒亟 从舒亢亟仂亞仂 rui = wuvrvi 裡 vNi (u) wuv 裡 vNi (u) 仆仂仄舒仍亳亰亠仄 亠亶亳仆亞亳 rui = h1 wuvh rvi ( ) 裡 vNi (u) wuv 裡 vNi (u) $ &&& % ' ))) (
- 13. 舒从仂亠 舒仂礌亳亠 亳仗仂仍亰仂于舒? 仂亳仆仆仂亠 舒仂礌亳亠 cos(u, v) = ruirvi 裡 iIuv 裡 r2 rui vj iIu 2 裡 jIv 仂亠仍亳 亳仂仆舒 PC(u, v) = (rui ru )(rvi rv ) 裡 iIuv 裡 (r r)2 (r r)2 ui u vi v iIu 裡 jIv
- 14. 舒从 仆仂仄舒仍亳亰仂于舒 亠亶亳仆亞亳? Mean centering h rui ( ) = rui ru Z-score h rui ( ) = rui ru u Percentile h rui ( ) = j Iu : ruj rui { } Iu
- 15. Collaborative filtering Matrix factorization methods
- 16. 舒亳仍亠亠 仗亳弍仍亳亢亠仆亳亠 舒仆亞舒 k 丐亠仂亠仄舒: 仍亳 于 仄舒亳亠 了 仂舒于亳 k 仆舒亳弍仂仍亳 亳仆亞仍仆 于亠从仂仂于, 仂 仗仂仍亳仄 仆舒亳仍亠亠 仗亳弍仍亳亢亠仆亳亠 仄舒亳 A 舒仆亞舒 k
- 17. Baseline predictors 仂亟亠仍: ruui =亮 + bu + bi argmin b* 裡 ( 2 r亮 b b) uui u i (u,i )R 裡 2 + 裡 bi +了 bu uU 2 iI $ % & ' ( ) 个仆从亳 仂亳弍从亳:
- 18. SVD 仂亟亠仍: Tqi ruui =亮 + bu + bi + pu argmin p*q*b* 裡 2 ( r亮 b b pTq) uui u i u i (u,i )R 2 ( ) +了 pu 2 + qi 2 + bu 2 + bi 个仆从亳 仂亳弍从亳:
- 19. Neighborhood (item-based) 仂亟亠仍: ruui = bui + sij ruj buj ( ) jSk (u,i ) 裡 sjS ij k (u,i ) 裡 u ruj buj ( ) jSk (u,i ) 裡 = bui + 慮ij
- 20. Neighborhood (optimization) 裡 u ruj buj ( ) jSk (u,i ) 裡 ruui = bui + ij ruj buj ( ) jR(u) ruui = bui + sij ruj buj ( ) jSk (u,i ) 裡 sjS ij k (u,i ) 裡 = bui + 慮ij
- 21. Neighborhood (optimization + implicit) u ruj buj ( ) jSk (u,i ) 裡 裡 + cij ruui = bui + ij ruj buj ( ) jR(u) 裡 jN(u) ruui = bui + sij ruj buj ( ) jSk (u,i ) 裡 sjS ij k (u,i ) 裡 = bui + 慮ij
- 22. Neighborhood (normalization) ruui = bui + R(u) 裡 + cij 1 2 ij ruj buj ( ) jR(u) 裡 裡 + N(u) 1 2 裡 cij jN(u) ruui = bui + ij ruj buj ( ) jR(u) jN(u) ruui = bui + Rk (i,u) 1 2 ij ruj buj ( ) jRk (u) 裡 + Nk (i,u) 1 2 裡 cij jNk (u)
- 23. 弌仆仂于舒 SVD 仂亟亠仍: Tqi ruui =亮 + bu + bi + pu argmin p*q*b* 裡 2 ( r亮 b b pTq) uui u i u i (u,i )R 2 ( ) +了 pu 2 + qi 2 + bu 2 + bi 个仆从亳 仂亳弍从亳:
- 24. Asymmetric-SVD 仂亟亠仍: T R(u) ruui =亮 + bu + bi + qi 1 2 ruj buj ( ) xj jR(u) 裡 + N(u) 1 2 裡 yj jN(u) $ % && ' ( )) argmin p*q*b* 裡 ( r r )2 + 了 quui uui i (u,i )R 2 + bu 2 + bi 裡 2 + xj 2 + yj 2 裡 jN(u) jR(u) $ % && ' ( )) 个仆从亳 仂亳弍从亳:
- 25. SVD++ 仂亟亠仍: T pu + N(u) ruui =亮 + bu + bi + qi 1 2 裡 yj jN(u) $ % && ' ( )) argmin p*q*b* 裡 ( r r )2 + 了 puui uui u (u,i )R 2 + qi 2 + bu 2 + bi 裡 2 + yj 2 jN(u) $ % && ' ( )) 个仆从亳 仂亳弍从亳:
- 26. Integrated model 仂亟亠仍: T pu + N(u) ruui =亮 + bu + bi + qi 1 2 裡 yj jN(u) $ % && ' ( )) + + Rk (i,u) 1 2 ij ruj buj ( ) jRk (u) 裡 + Nk (i,u) 1 2 裡 cij jNk (u)
- 27. 从舒从 于亠 仂 仂仗亳仄亳亰亳仂于舒?
- 28. SGD-仂仗亳仄亳亰舒亳 仄仂亟亠仍亳 SVD 仂亟亠仍: Tqi ruui =亮 + bu + bi + pu argmin p*q*b* 裡 2 ( r亮 b b pTq) uui u i u i (u,i )R 2 ( ) +了 pu 2 + qi 2 + bu 2 + bi 个仆从亳 仂亳弍从亳: 舒于亳仍舒 亟仍 亞舒亟亳亠仆仆仂亞仂 仗从舒: bu bu +粒1 eui 了1bu ( ) bi bi +粒1 eui 了1bi ( ) pu pu +粒 2 euiqi 了2 pu ( ) qu qi +粒 2 eui pu 了2qi ( )
- 29. Ridge regression 仂亟亠仍: yi wT xi wTw0 argmin w n裡 了wTw+ wT xi yi ( )2 i=1 # $ % & ' ( 个仆从亳 仂亳弍从亳: 丐仂仆仂亠 亠亠仆亳亠: w = (了 I + XTX)1 XT y = (了 I + A)1 d A = XTX d = XT y
- 30. ALS-仂仗亳仄亳亰舒亳 仄仂亟亠仍亳 SVD 仂亟亠仍: Tqi ruui =亮 + bu + bi + pu argmin p*q*b* 裡 2 ( r亮 b b pTq) uui u i u i (u,i )R 2 ( ) +了 pu 2 + qi 2 + bu 2 + bi 个仆从亳 仂亳弍从亳: P-step: pu = 了nuI + Au ( )1 du Au =Q[u]TQ[u] = qiqi T 裡 i:(u,i)R 裡 d =Q[u]T ru = ruiqi i:(u,i)R Q-step: qi = 了niI + Ai ( )1 di Ai = P[i]T P[i] = pupu T 裡 u:(u,i)R 裡 di = P[i]T ri = rui pu u:(u,i)R
- 31. 弌舒于仆亠仆亳亠 仄仂亟亠仍亠亶 仗仂 RMSE 仂亟亠仍 50 舒从仂仂于 100 舒从仂仂于 200 舒从仂仂于 亠亠 Item-based 0.9406 kNN Neighborhood 0.9002 SVD 0.9046 0.9025 0.9009 0.9009 Asymmetric 0.9037 0.9013 0.9000 0.9000 SVD SVD++ 0.8952 0.8924 0.8911 0.8911 Integrated model 0.8877 0.8870 0.8868 0.8868 仆舒 亟舒仆仆 Netflix Prize
- 32. 仂亟亠仍 50 舒从仂仂于 100 舒从仂仂于 200 舒从仂仂于 亠亠 Item-based 0.9406 kNN Neighborhood 0.9002 SVD 0.9046 0.9025 0.9009 0.9009 Asymmetric 0.9037 0.9013 0.9000 0.9000 SVD SVD++ 0.8952 0.8924 0.8911 0.8911 Integrated model 0.8877 0.8870 0.8868 0.8868 仆舒 亟舒仆仆 Netflix Prize BULLSHIT!
- 33. 弌ontent-based methods Tag-based methods True content-based methods
- 34. 弌ontent-based methods Tag-based methods
- 36. 弌仗仂仂弍 亞亠仆亠舒亳亳 亞仂于 User-generated Web-mining Expert-generated Metadata
- 37. Similarity by tags (co-occurrence) 舒仆仆亠: 仂弍仍舒从舒 亞仂于 亳 亠 仂亟于舒: 舒从从舒亟舒 舒亶舒 舒亳 Ti Tj Ti Tj Ti Tj 2 Ti Tj Ti + Tj Ti Tj Ti Tj
- 38. Similarity by tags (LSA) 舒亰仍仂亢亳仄 仄舒亳 Items x Tags 仗仂 SVD 亠 仂亟于舒: 从仂亳仆仆仂亠 舒仂礌亳亠 亳 亟. Item features x x Tags Tag features Items 了
- 39. 丐亞仂于亶 于舒仆亟舒仍亳亰仄 丐亞亳 Paris Hillton Last.fm, 仄舒亶 2013
- 40. 丐亞仂于亶 于舒仆亟舒仍亳亰仄 从舒从 弍仂仂? 仗舒于仍亠仆仆亠 亞亳 Paris Hillton User listening habbits Filter tags by similarity
- 41. 弌ontent-based methods True content-based methods
- 42. 亳仄亠 仄亰从舒 Spectral centroid Spectral flatness Spectral skewness Spectral kurtosis Zero-Crossing Rate (ZCR) Mel Frequency Cepstrum Coefficients (MFCCs) Instrumentation Rhythm Harmony Structure Intensity Genre Mood low-level high-level
- 43. Hybrid methods
- 44. 仍舒亳亳从舒亳 仄亠仂亟仂于 Weighted Switching Mixed Cascade
- 45. Evaluation
- 46. 舒从 仄仂亢仆仂 亳亰仄亠亳 从舒亠于仂 RS? Offline test User study Online experiment
- 47. Offline evaluation Prediction accuracy RMSE MAE Usage prediction accuracy Precision/recall @N F1 AUC Ranking accuracy DPM DGC Average Reciprocal Hit Rank (ARHR) Coverage Catalog coverage Sales diversity Gini index Shannon entropy
- 48. User study Confidence Trust Novelty Diversity Serendipity Robustness Adaptivity Scalability
- 49. 弌舒于仆亳于舒 仍亠亞亠!
- 50. Online study methods A-B testing Team-Driven Interleaving (TDI)
- 51. 仆亟亠亶 舒仆亳仍亠仆从仂 舒亰舒弍仂亳从 丕亟舒亳!