Machine Learning in the Wild: The Case of User-Centered Learning in Cyber Physical Systems

Smart environments, such as smart cities and smart homes, are Cyber-Physical-Systems (CPSs) which are becoming an increasing part of our everyday lives. Several applications in these systems, such as energy management through home appliance identification or activity recognition, adopt Machine Learning (ML) as a practical tool for extracting useful knowledge from raw data. These applications are usually characterized by a sequential stream of data, unlike the classical ML scenario in which the entire data is available during training. For such applications, Stream-based Active Learning (SAL) has been designed as a type of supervised ML in which an expert is asked to label the most informative instances as they arrive. Previous SAL techniques assume that the expert is always available and always labels the data correctly. However, in several applications, such as those mentioned above, the SAL activity interweaves with the everyday life of regular residents, who are often not experts, and may also not always be willing to participate in the labeling process. In this paper, we discuss the importance of usercentered ML, and show how taking into account realistic models of user behavior significantly improves the accuracy and reduces the training period of smart environment applications based on SAL. We consider two use cases, namely appliance identification and activity recognition. Results based on real data sets show an improvement in terms of accuracy up to 55.38%.