A Review of Off-Line Mode Dataset Shifts
Carla Caldeira Takahashi, Antônio P. Braga · IEEE Computational Intelligence Magazine · 2020
Dataset shifts are present in many real-world applications, since data generation is not always fully controlled and is subject to noise, degradation, and other natural variations. In machine learning, the lack of regularity in data can degrade performance by breaching error constraints. Different methods have been proposed to solve shifting problems; however, shifts in off-line learning mode are not as well examined. Off-line shifts consist of problems where drifts occur only with unlabeled data. Most methods aimed at dataset shifts consider that new labeled data can be received after training, which is not always the case. Here, a review on dataset shift characteristics and causes is presented as a tool for the analysis and implementation of machine learning methods targeting off-line mode dataset shift problems. In this context, a relationship between statistical learning risk functions and error degradation due to variation in data distribution was straightforwardly derived. Moreover, this paper provides a consistent survey of recent popular machine learning methods that address off-line mode dataset shift problems, focusing on the main characteristics of unlabeled data shifts.