2020/08/26 by Martin Khannouz, Khannouz, Martin, Tristan Glatard +1
Computer Science · Environmental Science · #Air Quality Monitoring and Forecasting #Data Stream Mining Techniques #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Network Security and Intrusion Detection
paper · pdf · doi:10.48550/arxiv.2008.11880
openalex publication_date 2020/08/26 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
This paper evaluates data stream classifiers from the perspective of\nconnected devices, focusing on the use case of HAR. We measure both\nclassification performance and resource consumption (runtime, memory, and\npower) of five usual stream classification algorithms, implemented in a\nconsistent library, and applied to two real human activity datasets and to\nthree synthetic datasets. Regarding classification performance, results show an\noverall superiority of the HT, the MF, and the NB classifiers over the FNN and\nthe Micro Cluster Nearest Neighbor (MCNN) classifiers on 4 datasets out of 6,\nincluding the real ones. In addition, the HT, and to some extent MCNN, are the\nonly classifiers that can recover from a concept drift. Overall, the three\nleading classifiers still perform substantially lower than an offline\nclassifier on the real datasets. Regarding resource consumption, the HT and the\nMF are the most memory intensive and have the longest runtime, however, no\ndifference in power consumption is found between classifiers. We conclude that\nstream learning for HAR on connected objects is challenged by two factors which\ncould lead to interesting future work: a high memory consumption and low F1\nscores overall.\n