Transfer learning in heterogeneous collaborative filtering domains

Regularized deep learning for fashion and social media data analysis

“…Because the domain relations of different domains could be automatically discovered and exploited. In order to capture the heterogeneous structures of user feedbacks, Pan et al investigated a series of models where the matrix factorization is regularized by transferring the source domain knowledge in forms of latent features to the target domain [39,38,40]. Specifically, they proposed a Transfer by Collective Factorization (TCF) framework for jointly modeling the target domain numerical ratings with the source domain binary ratings.…”

“…We work on the same scenario as TCF [38,40], where the auxiliary data is binary rating, e.g., like/dislike, while the target is numerical rating, e.g., 5-star gradings. First, in each single layer LSCF, as shown in Figure 1, we jointly factorize the rating matrices R andR in target and source domain into three parts: a user-specific latent feature matrix U , an item-specific latent feature matrix V and two rating patterns B andB in target and source domain respectively.…”

“…Batch based optimization algorithms such as PMF [130] and TCF [40] are more widely used in this problem. It updates model parameters only once after scanning the whole data.…”

“…The common information include but not limited to semantic networks, item attributes, inter-domain correlations and item attributes [33,34]; the second is to share the use or item latent features which could link source and target domains knowledge [35,36,37,38,39,40]; the third is to explicitly or implicitly the common rating patterns in source and target domain [41,42,43,44,45]. …”

“…Pan at el. presented a framework named as Transfer by Collective Factorization (TCF) [38,40] to transfer the shared latent user and item latent factors in both two domains. However, TCF neglects the shared knowledge in rating patterns.…”