Analysis of Netflix Data (Independent Study)

Today Recommendation systems [6] have become indispensable because of the sheer overload of information made available to a user from web-services(Netflix, IMDB, Amazon, Yelp and many others). Recommendation systems are a well studied research area. In the following work, we present our study on the Netflix Challenge [3]. The Netflix Challenge can be summarized in the following way : Given a movie m, predict the rating of a particular user u, given a list of user-movie ratings which may not contain the u;m pair. The performance of all such approaches is measured using the RMSE (root mean-squared error) of the submitted ratings from the actual ratings. Currently, the best system has an RMSE of 0.8585 [4]. In our attempt, we tried a a variety of enhancements to the approach that we followed in CS345a class project. Our previous approach initially seemed like a very promising apporach since we obtained an RMSE of 0.8312 for close to 70% of the validation set data. Unfortunately, none of our variants lead to an RMSE value below the required value of 0.85626. Section 2 gives some background on our approach last quarter by summmarizing our results for last quarter. Section 3 contains a detailed description of our current approach with Section 3.2 focussing on the two variants of collaborative filtering that we experimented with. Finally, Section 4 contains design choices that motivated us to choose parameters, error analysis and our final results.