TinyNARM: Simplifying Numerical Association Rule Mining for Running on Microcontrollers.

This paper presents a tinyNARM which is an experimental effort in approaching/tailoring the classical Numerical Association Rule Mining to limited hardware devices, precisely on ESP32 microcontrollers so that devices do not need to depend on in-cloud remote servers. The tinyNARM reduces the number of attributes in the transaction database by discretizing the continuous numeric attributes, and replacing the stochastic evolutionary algorithm for association rule mining with its deterministic counterpart. The preliminary results of the comparative study, in which the in-cloud NiaARM and the on-device tinyNARM were included by mining several UCI ML datasets, revealed that the quality of mined association rules and the time complexity were good enough for continuing the research in the direction of the tinyML.