Um novo método de AutoML para séries temporais: decomposição, causalidade, otimização evolutiva multitarefa e agentes

Abstract

Time series forecasting plays a fundamental role in several practical domains, supporting planning, decision-making, and the early identification of patterns and anomalies. The effectiveness of these applications depends on the ability to extract relevant information from historical observations, making the continued development of methodologies that enhance model accuracy and interpretability essential. Although automation strategies exist for this process, designing truly universal solutions remains challenging, particularly due to the difficulty of integrating, within a single approach, different types of models and techniques suitable for multiple application domains. In this context, this thesis presents AutoDCE-TS, a novel AutoML approach for multivariate time series forecasting based on Decomposition, Causal Discovery, and Multitask Evolutionary Optimization. AutoDCE-TS adopts a Multiple-Input Single-Output architecture composed of four layers responsible for automating feature extraction and selection, model selection, hyperparameter optimization, and forecasting. The methodology generates explainable pipelines by structuring specific sub-pipelines for each variable in the multivariate series and for the components generated through decomposition. For each variable, a causal graph is constructed and used as input to a regression tree-based model. Variable similarity is leveraged through multitask evolutionary optimization, enabling the simultaneous selection of models and hyperparameters with improved computational efficiency. Experiments were conducted on 18 datasets, involving comparisons with 11 methods. The results demonstrate that AutoDCE-TS is competitive and applicable across different domains, establishing it as a efficient solution for time series forecasting. Furthermore, this thesis introduces AutoDCE-TS-Agentic, a multi-agent system based on large language models (LLMs) designed to extend automation and facilitate the use of AutoDCE-TS. The system coordinates specialized agents to perform tasks such as missing-value imputation, analysis of the input time series, automatic configuration of critical hyperparameters, and generation of explanatory reports. AutoDCE-TS-Agentic provides a more accessible user interface, enhances the interpretability of forecasting pipelines, and makes the process more transparent and reproducible, thereby complementing and extending the capabilities of AutoDCE-TS.