AMANDA : density-based adaptive model for nonstationary data under extreme verification latency scenarios

Abstract

Gradual concept-drift refers to a smooth and gradual change in the relations between input and output data in the underlying distribution over time. The problem generates a model obsolescence and consequently a quality decrease in predictions. Besides, there is a challenging task during the stream: The extreme verification latency (EVL) to verify the labels. For batch scenarios, state-of-the-art methods propose an adaptation of a supervised model by using an unconstrained least squares importance fitting (uLSIF) algorithm or a semi-supervised approach along with a core support extraction (CSE) method. However, these methods do not properly tackle the mentioned problems due to their high computational time for large data volumes, lack in representing the right samples of the drift or even for having several parameters for tuning. Therefore, we propose a density-based adaptive model for nonstationary data (AMANDA), which uses a semi-supervised classifier along with a CSE method. AMANDA has two variations: AMANDA with a fixed cutting percentage (AMANDA-FCP); and AMANDA with a dynamic cutting percentage (AMANDADCP). Our results indicate that the two variations of AMANDA outperform the state-of-the-art methods for almost all synthetic datasets and real ones with an improvement up to 27.98% regarding the average error. We have found that the use of AMANDA-FCP improved the results for a gradual concept-drift even with a small size of initial labeled data. Moreover, our results indicate that SSL classifiers are improved when they work along with our static or dynamic CSE methods. Therefore, we emphasize the importance of research directions based on this approach.