XTRAML: TOWARDS AN EFFECTIVE AND EFFICIENT AUTOML SYSTEM FOR TEMPORAL RELATIONAL DATA

Abstract

Most of the success in current machine learning needs to be attributed to human experts and their manual operations and optimizations. Considering the various fields where ML is applied and the complexity, non-ML-experts always feel difficult to utilize useful ml methods to improve their work effectively. As a solution to narrow the gap and improve efficiency, automated machine learning is becoming a hot topic. We found building effective AutoML can be challenging for temporal relational data is very common in industrial applications using ml. In this thesis, we develop XtraML, which can derive predictive models from raw temporal relational data automatically. We attempt to develop a complete automatic data processing pipeline for the widely used binary classification tasks. Through our experiments with data sets from the real industry, we have shown that XtraML is an efficient and effective AutoML system.