The Neural Network with Automatic Feature Selection for Solving Problems with Categorical Variables

Abstract

The article deals with the problem of feature selection in data analysis. It is critically important for any tasks which are solved by machine learning algorithms. The difficulty is that features depend on the domain area, so the feature selection requires a lot of manual work. In this paper, the regression task with categorical input data is considered. The problem is that when using simple methods such as logistic regression, it is necessary to combine features into separate levels to obtain an acceptable result. This requires expert work or computationally expensive full search of all variants. In the present paper, the neural network for solving this task is proposed. Its architecture and details of learning are discussed. The advantage of this architecture is that it doesn’t require selection of complicated levels. For its operation, it is sufficient to list all input features. Experiments have shown that such neural network with raw features works better than logistic regression with handcrafted chosen levels. Also in this article, the metagraph representation of proposed neural networks is discussed. This makes it more convenient to work with this neural net.