Research And Application Of Personalized Recommendation Algorithm For Multi-domain Overlapping Users

For recommendation systems,data sparsity and user cold-start is one of the important challenging problems.Facing the problem of data sparsity,a common solution is to use cross-domain recommendation methods to collect data from other auxiliary domains outside the target domain to increase the density of the target domain data,but if some users who are in the lack of data in the auxiliary domain have rich interactive data in the target domain,this traditional cross-domain recommendation method cannot increase the density of the auxiliary domain data.For this problem,this thesis proposes a multi-domain recommendation method using multi-task learning technology that is cross-stitch network based on overlapping user scenarios between domains.This method can not only improve the accuracy of the target domain recommendation,but also improve the accuracy of the auxiliary domain recommendation.When user uses the system for the first time,the recommendation system cannot produce accurate recommendation results because of the lack of user interest information.In order to solve the user cold-start problem,we use MV-DNN to map the user vector and item vector to the shared semantic space,and then calculates the similarity between the user and the item,and proposes a cross-domain recommendation method based on a deep neural network,that is,using deep neural network transfers user interests in the auxiliary field to the target field,and solves the cold-start problem of users in the target field.The main work of this thesis is as follows:1)In a single domain,the implicit behavior data is converted into a score,and then filled into the existing display score matrix of the recommendation system.And the extract the user vector and item vector from the rating matrix.2)A multi-domain recommendation algorithm based on multi-task learning is proposed,which uses multi-task learning technology to share user interests in various fields,enrich user vectors in various fields,and improve the accuracy of recommendations in various fields.3)A cross-domain recommendation algorithm based on deep neural networks is proposed.The user vectors extracted from various fields are merged first,and the merged user vectors are mapped to the item vectors in each field to share the semantic space to solve the user problem.Cold start problem.4)Based on the IPTV scenario,the algorithm proposed in this thesis is implemented and tested with real environment data.Experiments show that the two algorithms proposed in this thesis achieve good performance in improving the recommendation accuracy and solving the user's cold start problem.