Meta-algorithm for the process control of complex machine learning model synthesis
Keywords:
Abstract
Introduction: Automated machine learning methods allow automating synthesis of machine learning models adapted to specific data processing. However, these methods require significant time and computational costs. Purpose: To develop a meta-algorithm for the process control of synthesis of machine learning models, that would reduce the computational complexity of automated synthesis of machine learning models. Results: We propose a general meta-algorithm for the process control of synthesis of complex machine learning models and a specific algorithm that allows limiting the search space through meta-learning. The proposed specific algorithm is based on using meta-features of data and an ontology that contains rules for selecting machine learning algorithms depending on the meta-features of the processed data. The ontology is constructed by pre-processing the results of previously synthesized machine learning models. In addition, for the specific algorithm, we develop an algorithm for building a training set and an algorithm for constructing an ontology to reduce the search space. The experiments have shown that the use of the proposed specific algorithm reduces the time of the synthesis of machine learning models by 41.12%. Moreover, the obtained models have increased accuracy (+0.54%), recall (+0.34%) and AUC (+1.85%). Practical relevance: The specific algorithm developed on the base of the meta-algorithm allows reducing the computational complexity of the process of automated machine learning model synthesis and enables the application of machine learning in subject areas that require prompt construction and adaptation of machine learning models to new data and tasks.
