Hierarchical Latent Knowledge Analysis for Co-Occurrence Data

Abstract

Two-mode and co-occurrence data have been frequently seen in the real world. We address the issue of predicting unknown cooccurrence events from known events. For this issue, we propose a new method that naturally combines observable co-occurrence events with their existing latent knowledge by using tlierarchical latent variables. This method builds a space-efficient hierarchical latent variable model and estimates the probability parameters of the model with an EM algorithm. In this paper, we focus on a data set of protein-protein interactions, a typical co-occurrence data set, and apply our method to the problem of predicting unknown protein-protein interactions, an important research issue in current computational biology. Using a real data set of protein-protein interactions and latent knowledge, we tested the performance of our method, comparing it with other recent machine learning approaches, including support vector machines. Our experimental results show that our method clearly outperforms the predictive performance obtained by other approaches. The results indicate that both our idea of using existing knowledge as a latent variable and our inethodology impleinenting it are effective for drastically improving the predictive performance of existing unsupervised learning approaches for cooccurrence data. ICML Proceedings of the Twentieth International Conference on Machine Learning