Evaluating a New Auto-ML Approach for Sentiment Analysis and Intent Recognition Tasks

Abstract

Automated Machine Learning (AutoML) is a research area that aims to help humans solve Machine Learning (ML) problems by automatically discovering good ML pipelines (algorithms and their hyperparameters for every stage of a machine learning process) for a given dataset. Since we have a combinatorial optimization problem for which it is impossible to evaluate all possible pipelines, most AutoML systems use a Genetic Algorithm (GA) or Bayesian Optimization (BO) to find a good solution. These systems usually evaluate the performance of the pipelines using the K-fold cross-validation method, for which the more pipelines are evaluated, the higher the chance of finding an overfitted solution. To avoid the aforementioned issue, we propose a system named Auto-ML System for Text Classification (ASTeC), that uses the Bootstrap Bias Corrected CV (BBC-CV) method to evaluate the performance of the pipelines. More specifically, the proposed system combines GA, BO, and BBC-CV to find a good ML pipeline for the text classification task. We evaluated our approach by comparing it with state-of-the-art systems: in the the Sentiment Analysis (SA) task, we compared our approach to TPOT (Tree-based Pipeline Optimization Tool) and Google Cloud AutoML service, and for the Intent Recognition (IR) task, we compared with TPOT and MLJAR AutoML. Concerning the data, we analysed seven public datasets from the SA domain and sixteen from the IR domain. Four out of those sixteen are composed by written English text, while all of the others are in Brazilian Portuguese. Statistical tests show that, in 21 out of 23 datasets, our system's performance is equivalent to or better than the others.