In this section we compare the semantically enhanced and standard item-based collaborative filtering in the context of two different data sets. In the first case, we focus our attention to the traditional context in which collaborative filtering is used, namely that of item ratings. For this purpose we choose the domain of movies and user's ratings of these movies. Secondly, we apply our approach to Web usage data. Specifically, we have chosen a real estate Web site containing information about various residential properties. While these data sets are quite different, the experimental results in this section demonstrate that the integrated approach yields advantages both in terms of improving accuracy, as well as in resolving some of the shortcomings associated with traditional approaches.
In each case, the data set was divided into random training and test sets. The training sets were used to build the models while the test sets were used to generate and evaluate recommendations. To assure statistical accuracy, this process was repeated five times for different random partitionings of the data. Unless otherwise specified, all of the results reported in this section represent averages over the five folds.