Multi-tasking Learning:Research And Application In Implementing Multi Re-commenders Ensemble

With the booming of Internet, various content/commertial service providers are trying to introduce new ways of entertaining and shopping, while actually brought chaos of information overload. Users have to spend more time than before on alternatives, suffering from lack of sufficient processing abilily, they face the danger of poor decision. Search engine helps with the overload, but fails in decision making due to the fact that one has to know what to look for before searching. Recommender systems are brought in as a decision-making consultant.In this paper, we study to further improve the accuracy of traditional recommender systems. Defined as ensemble of several collaborative filtering systems, our ensembled recommender mixed diversified advantages of different CF recommender, resulting in better accuracy in predicting. We built several CF models, with adjustment to satisfy our environment and emploit extra information provided, such as rating date. Then we build a universal linear combination of those predictors. Observed the fact that it out-performed single predictors, we also noticed that some users benefit a lot while others actually receive poor recommendations. We introduce multi-task learning mechanism to fully consider the impact of different user profile. We build a new user-dependent linear combination model, giving different user a different weight vector. To reduce computing complexity, we have the users clustered, and observe the impact on performance.We deliberately described our experiment approaches, and the result indicates that user-dependent ensemble model brings higher accuracy on prediction. But unforturnately, the more business-wise criteria of recall/precision proved us lack of real improvement. Thus further research is needed in such area.