Incremental mixture learning for clustering discrete data

This paper elaborates on an efficient approach for clustering discrete data by incrementally building multinormal mixture models through likelihood maximization using the Expectation-Maximization (EM) algorithm. The method adds sequentially at each step a new multinomial component to a mixture model based on a combined scheme of global and local search in order to deal with the initialization problem of the EM algorithm. In the global search phase several initial values are examined for the parameters of the multinomial component. These values are selected from an appropriately defined set of initialization candidates. Two methods are proposed here to specify the elements of this set based on the agglomerative and the kd-tree clustering algorithms. We investigate the performance of the incremental learning technique on a synthetic and a real dataset and also provide comparative results with the standard EM-based multinomial mixture model.