Online bi-clustering with sparsity priors
Résumé
Given a deterministic matrix M = (m ij), we want to cluster the row i and the column j before to predict m ij , where the values of the matrix are observed sequentially. The proposed algorithms are PAC-Bayesian procedures with new sparsity priors. Sparsity regret bounds are stated without any assumption on the matrix. These results are based on [13] where online clustering algorithms are suggested. Eventually, we also state minimax lower bounds for these problems, using a classical probabilistic reduction scheme. It shows the minimax optimality of the proposed algorithms in a worst case scenario.
Origine : Fichiers produits par l'(les) auteur(s)