An efficient latent-factor-based approach to social relationship recommendation

Social relationship recommenders aim at predicting potential useful relationships with high accuracy and efficiency, which is critically important in social network services for addressing information overload. Existing relationship recommenders mostly emphasize on friend recommendation in online social networks, which can not satisfy the requirements of industrial information systems. This work proposes an efficient latent-factor (LF)-based approach to predict multicategory relationships rather than only friendship in a social network. The main idea is to construct multicategory relationship data and develop the corresponding recommenders. To do so, two dimensions are designed for social relationship data, i.e., a category dimension built on real social relationship types, and an extended dimension built on the involved persons. Depending on the two dimensional relationship data, we construct a rating matrix by analyzing user preferences to each social relationship category one belonged to. For analyzing the resultant rating matrix with high accuracy, the hill-climbing and extended-linear-biases-enhanced latent factor (HC-ELBLF) model is proposed. The original grid-search-based learning process in the original ELBLF is substituted by the hill-climbing algorithm, which is an efficient and practical greed algorithm for parameter selection. The experimental results on two industrial datasets show effectiveness of the proposed HC-ELBLF approach.