Graph-based method for autonomous adaptation in online learning of non-stationary data

This work introduces a structural approach to addressing online learning problems by leveraging dataset relationships to represent the problem and estimate the likelihoods for a Bayesian classifier. The basic assumption behind such an approach is that the independent dominating set of a Gabriel graph contains the skeleton of the data and, therefore, conveys the information required to estimate drifts and density functions. The dominating set is a property of the Gabriel graph, which is deterministic and does not require hyperparameters to be set in advance. To accommodate the dynamic nature of streaming data, a method is proposed for updating the graph efficiently without recalculating all edges. The KDE estimator and its parameters are then directly derived from the dominating set, allowing the process to operate autonomously based on the spatial relationships in the data and the properties of the graph. The implementation involves constructing the Gabriel graph and its independent dominating set, from which the KDE estimator and Bayesian classifier are created to classify data streams. Results indicate that this method effectively handles various types of concept drifts, demonstrating its robustness and adaptability in online learning scenarios.