Classification And Regression Of Tree-Structured Data In The Model Space

As a kind of important structured data, tree is widely used in biology’, molecular chemical, and text processing. However, traditional machine learning methods are not enough for dealing with tree data since the tree contains node information and structure information. Currently, the widely adopted approach based on the subtree kernel demonstrated the idea that the similarity of the trees is determined by the shared subtree. The main idea of learning in the model space is to compose an abstract model layer beyond the experimental data and learning methods so that we can use the model to represent the original data to process the data. We propose a new approach to learning tree-structured data in the model space. Our method uses the tree echo state network to represent tree-structured data as a fixed-size vector. In other words, tree-structured data is mapped into points in the model space. Furthermore, we combine the model with the kernel methods to improve the performance of classification.In this paper, the main contribution of our work is:(1) We propose the tree echo state network model based on the echo state network. This method builds a model for the tree-structured data from the perspective of considering tree as a multi-branch, sequence-limited structure. Then we compute state variables for each node and use two kinds of state mapping function (one based on the root and the other based on the average) to represent models for tree-structured data.(2) After getting the model of tree-structured data, learning is then performed directly in the model space, instead of the original data space.(3) The proposed method is evaluated on three real-world data sets and compared with the other methods. The experimental results on the proposed method can effectively classify and predict the tree-structured data.