Development of pipeline feature engineering for building an AutoML service

Abstract

The large-scale implementation of artificial intelligence approaches in applied fields has a number of limitations, one of which is the availability of research competencies, knowledge of data analysis methods, mathematical statistics and machine learning. Automatic machine learning is designed to simplify the methodology of ML application development. Within the framework of this study, a new approach to the construction of pipeline feature engineering for AutoML service is presented, based on the sequential expansion of the feature space and the use of autoencoders to reduce the dimension of input features and reconstruct the final output features. The results of the presented approach are shown by the example of VANET network traffic data when solving the problem of classifying attacks on nodes. The data set was obtained as a result of simulating the real traffic of a certain segment of the VANET network in the OMNET++ environment and subsequent aggregation of data on network flows by means of CICFlowmeter-V4.0. Experiments have shown that machine learning models on the source data have an accuracy of 2% lower on average, which indicates the effectiveness of using the proposed Feature Engineering approach. The highest classification accuracy was demonstrated by Pipeline using the Multi-layered Model autoencoder and the XGBoost classification model - 91.2%. Thus, the presented Feature Engineering approach can be used to build the most effective feature space and improve the quality of machine learning models.