A Random Forest Approach to Model-based Recommendation
Hengru Zhang · Journal of Information and Computational Science · 2014
Model-based recommender systems are popular since models with demographic/item information can highlight the correlation of data and provide an intuitive recommendation. Some algorithms, such as Bayesian classifiers, decision trees, have been used to generate respective models. In this paper, we propose a random forest approach to create model-based recommendations. First, two types of datasets are extracted from the original one for the purposes of demographic-based and content-based recommendation, respectively. Second, average ratings are added into the new datasets as an attribute. Third, a forest is built for each user/item, where each leaf is assigned as a discrete score. Fourth, 6 predicting approaches adopting standard voting, weighted average, etc., are applied to compute evaluation values from the forest. Experimental results on the well-known MovieLens dataset show that some approaches are more reliable than others in terms of mean absolute error.