@inproceedings {pub3209,
	title = {KNN classifier with self adjusting memory for heterogenous concept drift},
	author = {Viktor Losing AND Barbara Hammer AND Heiko Wersing},
	year = {2016},
	month = {December},
	abstract = {Learning from 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 fundamentally different types
of possibly occurring drift undermine classical assumptions such as
i.i.d data. Incremental drift characterizes a continuous change in the
distribution such as the signals of a slowly degrading sensor. A
suddenly malfunctioning sensor on the other hand causes a severe shift
and is defined as abrupt drift. Available algorithms are able to handle
different types of drift, however they target either abrupt or
incremental drift and often incorporate hyperparameter requiring a
priori knowledge about the task at hand.\\
We propose a biological inspired, architecture which partitions the data
into a short-term and long-term memory. The former is a window of
recently seen data-points whose size is adjusted such that the estimated
generalization error is minimized. The latter preserves only those
information from previous concepts which are non-conflicting to the
current one. These memories are combined according to the demands of the
present concept to classify unseen data points. We couple our parameter
free approach with the K-Nearest Neighbor classifier, however, any other
incremental learning algorithm could be used as well. \\
New artificial and real datasets are proposed to evaluate performance on
specific types of drift. Experiments on these as well as on generally
known benchmark datasets compare our approach with state of the art
methods. The highly competitive results throughout all experiments
underline the robustness of our approach.},
	publisher = {IEEE},
	booktitle = {IEEE International Conference on Data Mining}
}