際際滷

際際滷Share a Scribd company logo

iTALS: implicit tensor factorization for context-aware recommendations (ECML/PKDD 2012 presentation)

Bal叩zs Hidasi
Bal叩zs Hidasi

This presentation is about the context-aware recommender algorithm iTALS. iTALS is a context-aware recommender algorithm for implicit feedback data. The user-item-context(s) setup is modelled in a binary tensor. Weights are also assigned to the cells based on the certainity of their information. An ALS-based algorithm is proposed that is capable of efficiently factorizing this tensor. Additionally a novel context information is introduced: sequentiality. This context allows us to incorporate association rule like information into the factorization framework and to differentiate between items with different repetetiveness patters and thus to make recommendations more accurate. This presentation was originally given at ECML/PKDD 2012 in Bristol.

1 of 17
iTALS Fast ALS-based tensor factorization for context-aware recommendation from implicit feedback Bal叩zs Hidasi balazs.hidasi@gravityrd.com Domonkos Tikk domonkos.tikk@gravityrd.com ECML/PKDD, 25TH SEPTEMBER 2012, BRISTOL
Overview Implicit feedback problem Context-awareness Seasonality Sequentaility iTALS Model Learning Prediction Experiments
Feedback types Feedback: user-item interraction (events) Explicit: Preferences explicitely coded E.g.: Ratings Implicit: Preferences not coded explicitely E.g.: purchase history
Problems with implicit feedback Noisy positive preferences E.g.: bought & disappointed No negative feedback available E.g.: had no info on item Usually evaluated by ranking metrics Can not be directly optimized
Why to use implicit feedback? Every user provides Some magnitudes larger amount of information than explicit feedback More important in practice Explicit algorithms are for the biggest only
Context-awareness Context: any information associated with events Context state: a possible value of the context dimension Context-awareness Usage of context information Incorporating additional informations into the method Different predictions for same user given different context states Can greatly outperform context-unaware methods Context segmentates items/users well
Seasonality as context Season: a time period E.g.: a week Timeband: given interval in season Context-states E.g.: days Assumed: aggregated behaviour in a given timeband is similar inbetween seasons User Item Date Context and different for different 1 A 12/07/2010 1 timebands 2 B 15/07/2010 3 E.g.: daily/weekly routines 1 B 15/07/2010 3 1 A 19/07/2010 1
Sequentiality Bought A after B B is the context-state of the users event on A Usefullness Some items are bought together Some items bought repetetively Some are not Association rule like information incorporated into model as context Here: into factorization methods Can learn negated rules If C then not A
M3 iTALS - Model Binary tensor D dimensional User User item context(s) M1 T Importance weights Lower weights to zeroes (NIF) Higher weights to cells with Item more events M2 Cells approximated by sum of the values in the elementwise product of D vectors Each for a dimension Low dimensional vectors
iTALS - Learning
iTALS - Prediction Sum of values in elementwise product of vectors User-item: scalar product of feature vectors User-item-context: weighted scalar product of feature vectors Context-state dependent reweighting of features E.g.: Third feature = horror movie Context state1 = Friday night third feature high Context state2 = Sunday afternoon third feature low
Experiments 5 databases 3 implicit Online grocery shopping VoD consumption Music listening habits 2 implicitizied explicit Netflix MovieLens 10M Recall@20 as primary evaluation metric Baseline: context-unaware method in every context- state
Scalability Running times on the Grocery dataset 350 300 iALS epoch Running time (sec) 250 iTALS (time bands) epoch 200 iTALS (sequence) epoch 150 iALS calc. matrix 100 iTALS (time bands) calc. matrix 50 iTALS (sequence) calc. matrix 0 10 20 30 40 50 60 70 80 90 100 Number of factors (K)
Results Recall@20 0.1200 0.1000 0.0800 0.0600 0.0400 0.0200 0.0000 VoD Grocery LastFM Netflix MovieLens iALS (MF method) iCA baseline iTALS (seasonality) iTALS (sequence)
Results Precision-recall curves 0.25 Recall-precision (Grocery) 0.20 0.15 0.10 iALS 0.05 0.00 iCA baseline 0.00 0.05 0.10 0.15 0.20 Recall-precision (VoD) iTALS 0.08 (season) 0.06 iTALS (sequence) 0.04 0.02 0.00 0.00 0.05 0.10 0.15 0.20
Summary iTALS is a scalable context-aware factorization method on implicit feedback data The problem is modelled by approximating the values of a binary tensor with elementwise product of short vectors using importance weighting Learning can be efficiently done by using ALS and other computation time reducing tricks Recommendation accuracy is significantly better than iCA-baseline Introduced a new context type: sequentiality association rule like information in factorization framework
Thank you for your attention! For more of my recommender systems related research visit my website: http://www.hidasi.eu T

More Related Content

iTALS: implicit tensor factorization for context-aware recommendations (ECML/PKDD 2012 presentation)

  • 1. iTALS Fast ALS-based tensor factorization for context-aware recommendation from implicit feedback Bal叩zs Hidasi balazs.hidasi@gravityrd.com Domonkos Tikk domonkos.tikk@gravityrd.com ECML/PKDD, 25TH SEPTEMBER 2012, BRISTOL
  • 2. Overview Implicit feedback problem Context-awareness Seasonality Sequentaility iTALS Model Learning Prediction Experiments
  • 3. Feedback types Feedback: user-item interraction (events) Explicit: Preferences explicitely coded E.g.: Ratings Implicit: Preferences not coded explicitely E.g.: purchase history
  • 4. Problems with implicit feedback Noisy positive preferences E.g.: bought & disappointed No negative feedback available E.g.: had no info on item Usually evaluated by ranking metrics Can not be directly optimized
  • 5. Why to use implicit feedback? Every user provides Some magnitudes larger amount of information than explicit feedback More important in practice Explicit algorithms are for the biggest only
  • 6. Context-awareness Context: any information associated with events Context state: a possible value of the context dimension Context-awareness Usage of context information Incorporating additional informations into the method Different predictions for same user given different context states Can greatly outperform context-unaware methods Context segmentates items/users well
  • 7. Seasonality as context Season: a time period E.g.: a week Timeband: given interval in season Context-states E.g.: days Assumed: aggregated behaviour in a given timeband is similar inbetween seasons User Item Date Context and different for different 1 A 12/07/2010 1 timebands 2 B 15/07/2010 3 E.g.: daily/weekly routines 1 B 15/07/2010 3 1 A 19/07/2010 1
  • 8. Sequentiality Bought A after B B is the context-state of the users event on A Usefullness Some items are bought together Some items bought repetetively Some are not Association rule like information incorporated into model as context Here: into factorization methods Can learn negated rules If C then not A
  • 9. M3 iTALS - Model Binary tensor D dimensional User User item context(s) M1 T Importance weights Lower weights to zeroes (NIF) Higher weights to cells with Item more events M2 Cells approximated by sum of the values in the elementwise product of D vectors Each for a dimension Low dimensional vectors
  • 11. iTALS - Prediction Sum of values in elementwise product of vectors User-item: scalar product of feature vectors User-item-context: weighted scalar product of feature vectors Context-state dependent reweighting of features E.g.: Third feature = horror movie Context state1 = Friday night third feature high Context state2 = Sunday afternoon third feature low
  • 12. Experiments 5 databases 3 implicit Online grocery shopping VoD consumption Music listening habits 2 implicitizied explicit Netflix MovieLens 10M Recall@20 as primary evaluation metric Baseline: context-unaware method in every context- state
  • 13. Scalability Running times on the Grocery dataset 350 300 iALS epoch Running time (sec) 250 iTALS (time bands) epoch 200 iTALS (sequence) epoch 150 iALS calc. matrix 100 iTALS (time bands) calc. matrix 50 iTALS (sequence) calc. matrix 0 10 20 30 40 50 60 70 80 90 100 Number of factors (K)
  • 14. Results Recall@20 0.1200 0.1000 0.0800 0.0600 0.0400 0.0200 0.0000 VoD Grocery LastFM Netflix MovieLens iALS (MF method) iCA baseline iTALS (seasonality) iTALS (sequence)
  • 15. Results Precision-recall curves 0.25 Recall-precision (Grocery) 0.20 0.15 0.10 iALS 0.05 0.00 iCA baseline 0.00 0.05 0.10 0.15 0.20 Recall-precision (VoD) iTALS 0.08 (season) 0.06 iTALS (sequence) 0.04 0.02 0.00 0.00 0.05 0.10 0.15 0.20
  • 16. Summary iTALS is a scalable context-aware factorization method on implicit feedback data The problem is modelled by approximating the values of a binary tensor with elementwise product of short vectors using importance weighting Learning can be efficiently done by using ALS and other computation time reducing tricks Recommendation accuracy is significantly better than iCA-baseline Introduced a new context type: sequentiality association rule like information in factorization framework
  • 17. Thank you for your attention! For more of my recommender systems related research visit my website: http://www.hidasi.eu T