Compact web browsing profiles for click-through rate prediction

In real time advertising we are interested in finding features that improve click-through rate prediction. One source of available information is the bipartite graph of websites previously engaged by identifiable users. In this work, we investigate three different decompositions of such a graph with varying degrees of sparsity in the representations. The decompositions that we consider are SVD, NMF, and IRM. To quantify the utility, we measure the performances of these representations when used as features in a sparse logistic regression model for click-through rate prediction. We recommend the IRM bipartite clustering features as they provide the most compact representation of browsing patterns and yield the best performance.