Skip to Main Content
Traditional collaborative filtering (CF) systems widely employing k-nearest neighbor (kNN) algorithms mostly attempt to alleviate the contemporary problem of information overload by generating personalized predictions for items that users might like. Unlike their popularity and extensive usage, they suffer from several problems. First, with increasing number of users and/or items, scalability becomes a challenge. Second, as the number of ratable items increases and number of ratings provided by each individual remains as a tiny fraction, CF systems suffer from sparsity problem. Third, many schemes fail to protect private data referred to as privacy problem. Due to such problems, accuracy and online performance become worse. In this paper, we propose two preprocessing schemes to overcome scalability and sparsity problems. First, we suggest using a novel content-based profiling of users to estimate similarities on a reduced data for better performance. Second, we propose pseudo-prediction protocol to help CF systems surmount sparsity. We finally propose to use randomization methods to preserve individual users' confidential data, where we show that our proposed preprocessing schemes can be applied to perturbed data. We analyze our schemes in terms of privacy. To investigate their effects on accuracy and performance, we perform real databased experiments. Empirical results demonstrate that our preprocessing schemes improve both performance and accuracy.