Performance comparison of collaborative-filtering approach with implicit and explicit data

Abstract

Challenge in developing a collaborative filtering (CF)-based recommendation system is the problem of cold-starting of items that causes the data to sparse and reduces the accuracy of the recommendations. Therefore, to produce high accuracy a match is needed between the types of data and the approach used. Two approaches in CF include user-based and item-based CFs, both of which can process two types of data; implicit and explicit data. This work aims to find a combination of approaches and data types that produce high accuracy. Cosine-similarity is used to measure the similarity between users and also between items. Mean Absolute Error is also measured to discover the accuracy of a recommendation. Testing of three groups of data based on sparseness results in the best accuracy in an explicit data-based approach that has the smallest MAE value. The result is that the average MAE value for user based (implicit data) is 0.1032, user based (explicit data) is 0.2320, item based (implicit data) is 0.3495, and item based (explicit data) is 0.0926. The best accuracy is in the item-based (explicit-data) approach which is the smallest average MAE value.