Incremental Factorization Machines for Persistently Cold-starting Online Item Recommendation Takuya Kitazawa Graduate School of Information Science and Technology, The University of Tokyo RecProfile 2016
For 5 Factorization Machines (FMs; general predictor) for context-aware recommendation global bias linear interaction [S. Rendle. ACM TIST, 3(3), 2012.] (Fig. 1) MF FMs
For 6 Factorization Machines (FMs; general predictor) for context-aware recommendation global bias linear interaction [S. Rendle. ACM TIST, 3(3), 2012.] (Fig. 1) MF FMs Persistent cold-start can be seen as: concept drift online algorithms are more effective
9 Incremental Matrix Factorization (iMF) for online item recommendation iMF S can be data stream update (user ID) update (item ID) for each observation Where are auxiliary features? [J. Vinagre et al. UMAP 2014] SGD update
11 Using FMs to item recommendation Solve MF with unique target value [J. Vinagre et al. In Proc. of UMAP 2014, pp. 459–470.] Solve FM as regression (least-square loss) with y = 1.0 1.0 For 1.0
12 Generalizing iMF to incremental FMs (iFMs) SGD update for FMs O( #non-zero in x × k ) S is data stream for each observation SGD update for MF update (user information) update (item information) update update update SGD update SGD update
13 Incremental adaptive regularization FMs have many hyperparameters: Adaptive regularization [S. Rendle. In Proc. of WSDM 2012, pp. 133-142.] Incremental adaptive regularization (“pseudo validation sample” idea) outdated immediately
14 Evaluation of online item recommender: Test-then-learn procedure 0) pre-train with initial 20% of samples [J. Vinagre, et al. In Proc. of REDD 2014.] Evaluation of Recommender Systems in Streaming Environments, (u1 , i1 ) ɾɾɾ timestamped n-samples for evaluation ɾɾɾ (uk , ik ) (uk+3 , ik+3 ) For sample (u, i) 1) top-N recommendation for u 2) check if i is in top-N list 3) compute [email protected] in 4) update model based on (u, i) (un , in )
21 Future work ▶ More competitors/datasets (time-stamped, rich in contexts) ▶ Other approaches for positive-only feedback (e.g. BPR) [J. Vinagre, et al. In Proc. of UMAP 2014.] (Fig. 1)
Incremental Factorization Machines for Persistently Cold-starting Online Item Recommendation Takuya Kitazawa Graduate School of Information Science and Technology, The University of Tokyo RecProfile 2016
takuti
yumakoizumi
sgnm
kei18
gkehr1
yamdan
fullflu
hirosaji
hakubishin3
icttitech
gpeyre
ailaboocu
kitachan_black
pedronauck
rmw
jmmastey
iamctodd
lauravandoore
bkeepers
maltzj
paigeccino
zakiyama
danielanewman
malarkey
morganepeng