Research And Implementation Of Cross-Domain Recommendation Algorithm Based On Transfer Learning From Social Network

Recommended system is considered to be the most effective method of information filtering technology in the big data era. Recommended system not only succeeds in commercial field, but also produces crucial social influence. Although the successful application Recommendation System has achieved in the internet, there are bottlenecks limit its development: the data sparseness and the "cold start" problem. However, the researches on cross-domain recommendation which aims to transfer the similar "knowledge" from available auxiliary domain to the target system data, has provided a new trend to solve these two problems.This thesis is to analysis the behavior of cross-domain users and improved the transfer learning methods. Besides, the impact of cross-domain users leave on the cross-domain recommendation is also investigated. Specific research work is as follows:1. Firstly, this thesis proposes a cross-domain collaborative filtering recommendation algorithm. The core idea of the algorithm is to transfer the cross-user behavior information in the field of auxiliary systems to the target recommendation system. Thereby the user information in the target system is enriched and become easier for recommendation. Simulations of this algorithm are made on real data sets. Experimental results show that in some case the algorithm is better than the classic recommendation algorithm in enhancing accuracy rate.2. Then, this thesis implements a cross-domain algorithm which can transfer learning the common user rating patterns between auxiliary domain and target domain. The algorithm is simulated on real data sets with certain amount of cross-domain users. This algorithm has improved the accuracy compared with traditional methods.3. Finally, this thesis proposes two indicators that are used for selecting the suitable auxiliary domains. The indicators are considered to measure the user ratings distribution and cross-domain users’ impact, which are called KL divergence and user confidence. The simulations verify that the KL divergence and user confidence has positive correlation with the accuracy of cross-domain recommendation algorithms.