Accuracy vs. traffic trade-off of Learning IoT Data Patterns at the Edge with Hypothesis Transfer Learning

Right now, the dominant paradigm to supportknowledge extraction from raw IoT data is through global cloudplatform, where data is collected from IoT devices, and analysed.However, with the ramping trend of the number of IoT devicesspread in the physical environment, this approach might simplynot scale. The data gravity concept, one of the basis of Fog andMobile Edge Computing, points towards a decentralisation ofcomputation for data analysis, whereby the latter is performedcloser to where data is generated. Along this trend, in this paperwe explore the accuracy vs. network traffic trade-off when usingHypothesis Transfer Learning (HTL) to learn patterns fromdata generated in a set of distributed physical locations. HTLis a standard machine learning technique used to train modelson separate disjoint training sets, and then transfer the partialmodels (instead of the data) to reach a unique learning model.We have previously applied HTL to the problem of learninghuman activities when data are available in different physicallocations (e.g., areas of a city). In our approach, data is not movedfrom where it is generated, while partial models are exchangedacross sites. The HTL-based approach achieves lower (thoughacceptable) accuracy with respect to a conventional solution basedon global cloud computing, but drastically cuts the networktraffic. In this paper we explore the trade-off between accuracyand traffic, by assuming that data are moved to a variablenumber of data collectors where partial learning is performed.Centralised cloud and completely decentralised HTL are thetwo extremes of the spectrum. Our results show that there isno significant advantage in terms of accuracy, in using fewercollectors, and that therefore a distributed HTL solutions, alongthe lines of a fog computing approach, is the most promising one.