An adaptive learning system for time series forecasting in the presence of concept drift

• Autor(a) principal: CAVALCANTE, Rodolfo Carneiro
• Data de Publicação: 2017
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFPE
• dARK ID: ark:/64986/001300000ssvk
• Texto Completo: https://repositorio.ufpe.br/handle/123456789/25349
• topic: Inteligência computacional; Previsão de séries temporais

A time series is a collection of observations measured sequentially in time. Several realworld dynamic processes can be modeled as time series. One of the main problems of time series analysis is the forecasting of future values. As a special kind of data stream, a time series may present concept drifts, which are changes in the underlying data generation process from time to time. The concept drift phenomenon affects negatively the forecasting methods which are based on observing past behaviors of the time series to forecast future values. Despite the fact that concept drift is not a new research area, the effects of concept drifts in time series are not widely studied. Some approaches proposed in the literature to handle concept drift in time series are passive methods that successive update the learned model to the observations that arrive from the data stream. These methods present no transparency to the user and present a potential waste of computational resources. Other approaches are active methods that implement a detect-and-adapt scheme, in which the learned model is adapted just after the explicit detection of a concept drift. By using explicit detection, the learned model is updated or retrained just in the presence of drifts, which can reduce the space and computational complexity of the learning system. These methods are generally based on monitoring the residuals of a fitted model or on monitoring the raw time series observations directly. However, these two sources of information (residuals and raw observations) may not be so reliable for a concept drift detection method applied to time series. Residuals of a fitted model may be influenced by problems in training. Raw observations may present some variations that do not represent significant changes in the time series data stream. The main contribution of this work is an active adaptive learning system which is able to handle concept drift in time series. The proposed method, called Feature Extraction and Weighting for Explicit Concept Drift Detection (FW-FEDD) considers a set of time series features to detect concept drifts in time series in a more reliable way, being trustworthy and transparent to users. The features considered are weighted according to their importance to define concept drifts at each instant. A concept drift test is then used to detect drifts in a more reliable way. FW-FEDD also implements a forecasting module composed by a pool of forecasting models in which each model is specialized in a different time series concept. Several computational experiments on both artificial and real-world time series showed that the proposed method is able to improve the concept drift detection accuracy compared to methods based on monitoring raw time series observations and residual-based methods. Results also showed the superiority of FW-FEDD compared to other passive and active adaptive learning systems in terms of forecasting performance.