Multi-Domain Active Learning For Recommendation

Recently,with the recommendation techniques widely used in the Internet applications,people can quickly choose what they want from a large amount of things.However,existing recommendation techniques still have several problems.On one hand,Data in many website involves more than one domain.For example,in Douban Movie website,it contains comedy film,romance film and action film.When designing recommendation algorithm for these websites,both domain shared knowledge and domain specific knowledge should be considered.On the other hand,with the development of the recommendation techniques,the problem of data sparsity has been the bottleneck of further improving them.Therefore,how to effectively get ratings become a key problem.And in recent years,active learning has been proposed to solve the data sparsity in recommendation.Based on the above problems,we proposed an active learning strategy for multi-domain recommendation.The detailed works are as follows:1.We proposed a new problem,which is multi-domain active learning for recommendation.The challenge of this problem is that existing active learning strategies for recommendation can only consider domain specific knowledge and ignore domain independent knowledge,which leads to the waste of labeling efforts.For this problem,we propose a novel multi-domain active learning strategy,which can consider not only the knowledge of each domain but also the knowledge between each domain when querying ratings.2.For the domain specific features and domain independent features,we design an expected entropy based active learning strategy and a variance based active learning strategy to measure the generalization error of these two parts,which makes our multi-domain active learning strategy can be applied to specific multi-domain recommendation model.3.Design and develop a multi-domain recommendation experiment system.Experiments,including performance comparison,significance test and parameter sensitivity,are conducted on five multi-domain recommendation tasks,which is constructed by real-world datasets.Experimental results demonstrate our proposed active learning strategy is much better than current active learning strategies,can save much more labeling efforts and works very stably.