Fig. 1

Overview of the first data pass. MeShClust v3.0 is based on the mean shift algorithm, which is an instance of unsupervised learning. The scaled-up MeShClust v3.0 is also an instance of out-of-core learning [35], in which the learning algorithm is trained on separate batches of the training data consecutively. The algorithm requires multiple passes through the input data. In the first data pass, the tool reads a batch of input sequences. Then the mean shift algorithm (all of the four steps) is run on the batch until convergence. Sequences that cannot be assigned to any center are kept in the reservoir. Next, a new batch is read. The main mean shift is run on this batch but without the initialization step and for one iteration only, i.e., already found centers are shifted and merged on the new batch and no new centers are discovered. Sequences that cannot be assigned to any of the centers are added to the reservoir. When the reservoir has enough sequences (more than the batch size), sequences in it are shuffled and a batch of them is clustered using an independent instance of the mean shift algorithm. This instance is run until convergence. The resulting centers (if any) are merged with the centers accumulated by the main mean shift. This procedure is repeated until all sequences are read and the reservoir is empty. In subsequent passes, the algorithm rereads input sequences batch by batch. The main mean shift algorithm is run for one iteration on each batch. If the number of clusters does not change during a pass, the algorithm converges. In the final data pass, all sequences are reread batch by batch, and each sequence is assigned to the cluster with the closest center to it