Handling of Labeling Uncertainty in Smart Homes using Generalizable Fuzzy Features
Cédric Demongivert, Kévin Bouchard, Sébastien Gaboury, Maxime Lussier, Hubert Kenfack Ngankam, Mélanie Couture, Nathalie Bier, Sylvain Giroux · 2021
Smart homes research is now entering a phase of real deployment and of early commercialization. For the type of smart homes used to monitor the daily life of residents, activity recognition is one of the key artificial intelligence components necessary. In labs, it is mostly based on machine learning methods, but in real deployments, due to the difficulty to build labeled datasets, it still usually depends largely on logical systems and inference rules. In this work, we try to leverage generalizable fuzzy features to evaluate the quality of the label inferred by commonsense inference. The fuzzy rules are built from annotated instances in CASAS's dataset and by transferring them to our own infrastructure. The data exploited include 11 of our deployed smart homes and shows promising results. Our experiments shows that it is likely possible to exploit those rules to evaluate the quality of our data labeling.