The paper deals with a generative probabilistic graphical model with hidden states based on nonlinear principal manifolds specified as a grid of nodes to solve the problem of classification of sequences or time-series data. Kohonen's self-organizing map is used to approximate the training data as the grid nodes. The model is presented in factor-graph form the used factor-functions description. Method of learning and probabilistic inference is developed on the proposed model. Evaluated quality of the classification of the proposed model is compared with existing models (HMM, HCRF) on different sets of data from the UCI repository, including a comparative evaluation in the case when small amount of the training data is available.
1 - 1 of 1 items