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 Incremental adaptive regularization (“pseudo validation sample” idea) outdated immediately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 recall@N 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
yudai00
rpc
shuntaros
chokkan
muraoka7
mickey_kubo
apostoloskritikos
kuehara
speed1313
tmats
kyoun
yukiar
sferik
moore
morganepeng
addyosmani
brad_frost
carmenintech
tammyeverts
philnash
michaelherold
aki_iinuma
lauravandoore
smashingmag