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
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
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 )