Tracking Recurrent Concepts Using Context in Memory-constrained Devices

The dissemination of ubiquitous devices with data analysis capabilities motivates the need for resource-aware approaches able to learn in reoccurring concept scenarios with memory constraints. The majority of the existing approaches exploit recurrence by keeping in memory previously learned models, thus avoiding relearning a previously seen concept when it reappears. In real situations where memory is limited it is not possible to keep every learned model in memory, and some decision criteria to discard such models must be defined. In this work, we propose a memory-aware method that associates context information with stored decision models. We establish several metrics to define the utility of such models. Those metrics are used in a function that decides which model to discard in situations of memory scarcity, enabling memory-awareness into the learning process. The preliminary results demonstrate the feasibility of the proposed approach for data stream classification problems where concepts reappear and memory constraints exist.