Dedicated Memory Models for Continual Learning in the Presence of Concept Drift

Data Mining in non-stationary data streams is gaining more attention recently, especially in the context of Internet of Things and Big Data. It is a highly challenging task since the different types of possibly occurring concept drift undermine classical assumptions such as data independence or stationary distributions. We propose the Self Adjusting Memory (SAM) model, which can deal with heterogeneous concept drift, i.e different types and rates, using biologically inspired memory models and their coordination. The idea is to construct dedicated models for the current and former concepts and apply them according to the given situation. This general approach can be combined with various classifiers meeting certain conditions, which we specify in this contribution. SAM is easy to use in practice since a task specific optimization of the meta parameters is not necessary. We recap the merits of our architecture with the k Nearest Neighbor classifier and evaluate it on artificial as well as real world benchmarks. SAM’s highly competitive results throughout all experiments underline its robustness as well as its capability to handle heterogeneous concept drift.