Recommendation Based on Latent Topics and Social Network Analysis
Jian-hua Yeh, Meng-Lun Wu · 2010
In 2007, Netflex provided a large training dataset describing user ratings of movies for KDD Cup contest. Many competitors proposed various kinds of data mining model trying to achieve the best prediction performance. The first place winner among the competitors got the best root mean square error (RMSE) of 0.256. Most of the models applied statistical machines learning techniques with collaborative mining approach to achieve their best performance. In this paper, a hybrid recommendation model is proposed to get better prediction result which combines both content-based and collaborative recommendation approaches with latent topic discovery and social network analysis. This model was tested using 2007 KDD Cup movie dataset and found that either with single content-based approach or single collaborative approach is hard to get better RMSE result than hybrid models. By combining both kinds of approaches, the latent-topic-only approach observed in our experiment achieves only RMSE=0.274, while with Bonacich power centrality in social network get better improvement to 0.252, which proved that our model is better than all of the competitors in the contest.