Probabilistic knowledge infusion through symbolic features for context-aware activity recognition

In the general machine learning domain, solutions based on integration of deep models with knowledge-based approaches are emerging. Indeed, such hybrid systems have advantage improving recognition rate and model’s interpretability. At same time, they require a significantly reduced amount labeled data to reliably train model. However, these techniques been poorly explored in sensor-based Human Activity Recognition (HAR) domain. The common-sense knowledge about activity execution can potentially improve purely data-driven approaches. While few infusion proposed for HAR, rely rigid logic formalisms that do not take into account uncertainty. this paper, we propose P-NIMBUS, novel approach HAR relies probabilistic reasoning. A ontology is charge computing symbolic features combined automatically extracted by CNN model from raw sensor high-level context data. particular, encode activities consistent user’s surrounding context. These infused within before classification layer. We experimentally evaluated P-NIMBUS dataset mobile devices includes 14 different performed 25 users. Our results show outperforms state-of-the-art neuro-symbolic approaches, requiring limited training reach satisfying rates (i.e., more than 80% F1-score only 20% data).