This paper evaluates data stream classifiers from the perspective of connected devices, focusing on use case Human Activity Recognition. We measure both classification performance and resource consumption (runtime, memory, power) five usual algorithms, implemented in a consistent library, applied to two real human activity datasets three synthetic datasets. Regarding performance, results show overall superiority Hoeffding Tree, Mondrian forest, Naïve Bayes over Feedforward Neural Network...

The Internet of Things could benefit in several ways from mining data streams on connected objects rather than the cloud.In particular, limiting network communication with cloud services would improve user privacy and reduce energy consumption devices.Besides, applications leverage computing power for improved scalability.

Mondrian Forests are a powerful data stream classification method, but their large memory footprint makes them ill-suited for low-resource platforms such as connected objects. We explored using reduced-precision floating-point representations to lower consumption and evaluated its effect on performance. applied the Forest implementation provided by OrpailleCC, C++ collection of algorithms, two canonical datasets in human activity recognition: Recofit Banos et al. Results show that precision...

This paper evaluates data stream classifiers from the perspective of connected devices, focusing on use case HAR. We measure both classification performance and resource consumption (runtime, memory, power) five usual algorithms, implemented in a consistent library, applied to two real human activity datasets three synthetic datasets. Regarding performance, results show an overall superiority HT, MF, NB over FNN Micro Cluster Nearest Neighbor (MCNN) 4 out 6, including ones. In addition, some...

Supervised learning algorithms generally assume the availability of enough memory to store their data model during training and test phases. However, in Internet Things, this assumption is unrealistic when comes form infinite streams, or are deployed on devices with reduced amounts memory. In paper, we adapt online Mondrian forest classification algorithm work constraints streams. particular, design five out-of-memory strategies update trees new points limit reached. Moreover, trimming...

Supervised learning algorithms generally assume the availability of enough memory to store data models during training and test phases. However, this assumption is unrealistic when comes in form infinite streams, or are deployed on devices with reduced amounts memory. Such constraints impact model behavior assumptions. In paper, we show that under constraints, increasing size a tree-based ensemble classifier can worsen its performance. particular, experimentally existence an optimal for...

Supervised learning algorithms generally assume the availability of enough memory to store data models during training and test phases. However, this assumption is unrealistic when comes in form infinite streams, or are deployed on devices with reduced amounts memory. Such constraints impact model behavior assumptions. In paper, we show that under constraints, increasing size a tree-based ensemble classifier can worsen its performance. particular, experimentally existence an optimal for...

Mondrian Forests are a powerful data stream classification method, but their large memory footprint makes them ill-suited for low-resource platforms such as connected objects. We explored using reduced-precision floating-point representations to lower consumption and evaluated its effect on performance. applied the Forest implementation provided by OrpailleCC, C++ collection of algorithms, two canonical datasets in human activity recognition: Recofit Banos \emph{et al}. Results show that...