A Dynamic MOLMAP approach for pattern classification in three-way data

In many pattern recognition methods, numerical features for each sample should be represented as a vector to the learning algorithm and generally the data can be arranged in a two dimensional array. This could be a challenging issue if we have an array of features, say a matrix per sample which results in a three dimensional data array. The MOLMAP (MOLecular Map of Atom-level Properties) approach was originally introduced to deal with three dimensional data arrays and calculate molecular descriptors. The MOLMAP approach is based on selforganizing map (SOM) and needs to have a predefine network structure which is not easily decidable. We presented a dynamic MOLMAP approach based on Growing Self-Organizing Map (GSOM) for classification of three-way data set. The proposed approach produces an informative MOLMAP-score which can used to learn a classifier. The potential of the proposed method was evaluated using two analytical datasets, electronic-nose and fluorescence. The final classification models were built using XYfused neural network and evaluated by10-fold cross validation. The results show that the Dynamic MOLMAP outperforms the classical one at the same number of neurons in term of classification accuracy. The proposed approach not only have less tunable parameters but also can be used to exploratory analysis and inspecting feature space.