The Research On Recommendation Via Cross-domain Transfer Learning

With the development of internet,the information online is over-load.In order to help users search information quickly and effectively,recommender systems has been designed,which recommends items to users according to users’ history behaviors.Collaborative filtering,targeting at making recommendations or predictions for users by collecting preferences from others,is one of the most popular techniques.Lots of previous works focus on the sole source of rating data or check-in data.Furthermore,external information(e.g.,users’ social relations and items’attributes)is considered to improve the recommendation performance.However,external information sometimes is not easy to obtain and the recommendation algorithms are sensitive to the noisy data.Recently,transfer learning has been applied into the recommender systems to boost the recommendation performance.Transfer collaborative filtering exploits the knowledge extracted from the source domain to improve the recommendation performance in the target domain.Most previous TCF methods only focus on transferring knowledge from one single source domain to one target domain.Recently,some efforts have been made to learn knowledge from multiple sources to utilize much available information.In general,there are two strategies to exploit the knowledge from multiple sources.The first one is to joint multiple sources according to the overlapping users or items.This strategy may limit the applicability of the proposed method.Data of overlapping users or items in different applications are few and hard to be collected in real world.The second one is to re-weight different sources to boost the recommendation performance.However,it is not uncommon to get inconsistent information from different sources.The existing multi-source TCF methods simply re-weight domains to merge them together while ignoring the difference of data distribution among various domains.In order to address the limitations of the existing TCF methods,we propose two novel multi-source TCF methods.Firstly,we propose a transfer collaborative filtering framework from multiple sources via consensus regularization,called TRACER for short.Based on the assumption that knowledge learnt from different sources is used to make predictions for the same target domain so that their predicted results should be as similar as possible,we design a novel consensus regularization to enforce the predicted results to be similar.Then,we propose the TCF method via consensus regularization,which is based on the CBT framework,to make transfer learning from multiple sources for collaborative filtering.The TRACER framework handles the information inconsistency with a consensus regularization,which drives outputs from multiple sources to make the same predictions.Experiments conducted on two real-world datasets validate the effectiveness of the proposed method.In addition,we propose a novel local ensemble framework across multiple source domains for collaborative filtering(called LOEN for short),where weights of multiple sources for each missing rating in the target domain are determined according to their corresponding local structures.Compared with the previous TCF methods,LOEN does not require overlapping data and considers the divergence of sources through exploiting the local structures of ratings,which allows LOEN to be more general and effective.Experiments conducted on real datasets validate the effectiveness of LOEN,especially for knowledge transfer across unrelated source domains.