Igor Avbelj (2012) Hybridisation of recommendation systems based on collaborative filtering. EngD thesis.
In this thesis, we present various techniques for recommender systems. We implement the k-Nearest Neighbours method and two matrix factorizations. Recommendations generated with these methods are then combined (hybridization). For the k-Nearest Neighbours method we implemented several similarity measures: Euclidean similarity, cosine similarity and the Pearson correlation coefficient. The first of the matrix factorization methods is BRISMF. It minimizes RMSE (root mean squared error). The second matrix factorization technique is BPRMF, which optimizes rank correlation. Recommendations from these individual methods are combined using average and linear regression blending. Using linear regression, we get weights which are then used in weighted sums. In the end, we were trying to find out in which cases hybridization improves final results. We generate recommendations for users in datasets iTIVI and MovieLens. Both datasets contain data about movies. We experiment with different similarity measures and parameters in order to get the most accurate recommendations. We use hybridization of individual methods, to see which methods together give the best results. On small datasets the most accurate recommendations are generated using the k-Nearest Neighbour method. On big datasets BRISMF gives better results. BPRMF does not give very good results by itself, however when used in hybridization it improves Spearman's rank correlation coefficient significantly. Using hybridization techniques improves results on most datasets.
Actions (login required)