Analysis of collaborative filtering algorithms

Abstract

Recommender System is a subclass of information filtering system which predicts the rating given to an item by any user. Collaborative filtering is a key technique in recommender systems. This technique predicts the user rating of an item by collaboration of other users who have similar interests with this user. Collaborative filtering approaches can be categorized as Memory based, Model-based and Hybrid approaches. Memory-based approach can be further classified as Item-based and User-based recommendations. Pearson correlation scheme belongs to user-based scheme and Slope one family of algorithms belong to item-based scheme. Slope one family consists of Normal, Weighted and Bipolar slope one algorithms. Algorithms belonging to model-based approach are Singular value decomposition, Regularized Singular value decomposition and Probabilistic Matrix Factorization. In hybrid approach combination of memory-based and model-based approaches are used for making recommendations. In this thesis we made an attempt to analyze various algorithms in Memory-based and Model-based approaches. In model based algorithms, we analyzed Singular Value Decomposition (SVD) and Regularized Singular Value Decomposition (RSVD). By taking three different dataset sizes, we observed that RSVD outperforms SVD for all three dataset sizes. In memory based algorithms, we analyzed Pearson correlation scheme which takes the correlation between user vectors as similarity measure and Slope one family of algorithms. In slope one algorithms, we proposed an improvement to the existing scheme for determining Threshold value of Bipolar slope one algorithm. We used median and average of min-max ratings which outperforms the existing user average scheme. Finally, we made an analysis of all these algorithms and concluded that RSVD outperforms rest of the algorithms in terms of accuracy of predictions.