On-line Active Learning in Data Stream Regression using Uncertainty Sampling based on Evolving Generalized Fuzzy Models

In this paper, we propose three criteria for efficient sample selection in case of data stream regression problems within an on-line active learning context. The selection becomes important whenever the target values, which guide the update of the regressors as well as the implicit model structures, are costly or time-consuming to measure and also in case when very fast models updates are required to cope with stream mining real-time demands. Reducing the selected samples as much as possible while keeping the predictive accuracy of the models on a high level is thus a central challenge. This should be ideally achieved in unsupervised and single-pass manner. Our selection criteria rely on three aspects: 1.) the extrapolation degree combined with the model’s non-linearity degree, which is measured in terms of a new specific homogeneity criterion among adjacent local approximators, 2.) the uncertainty in model outputs which can be measured in terms of confidence intervals using so-called adaptive local error bars — we integrate a weighted localization of an incremental noise level estimator and propose formulas for on-line merging of local error bars; 3.) the uncertainty in model parameters which is estimated by the so-called A-optimality criterion which relies on the Fisher information matrix. The selection criteria are developed in combination with evolving generalized Takagi-Sugeno (TS) fuzzy models (containing rules in arbitrarily rotated position), as it could be shown in previous publications that these outperform conventional evolving TS models (containing axis-parallel rules). The results based on three high-dimensional real-world streaming problems show that a model update based on only 10%-20% selected samples can still achieve similar accumulated model errors over time to the case when performing a full model update on all samples. This can be achieved with a negligible sensitivity on the size of the active learning latency buffer (ALLB).