I. Introduction
In various in-home health monitoring applications, data are widely collected via the IoT for supervised learning tasks. For example, data can be collected to train a classifier predicting the status of an old man living alone. In most cases, the performance of supervised learning is superior to unsupervised learning in in-home health monitoring applications. Therefore, many researchers focus on it and have made much progress. Most existing studies simply assume that the training data are perfect, sufficient and cost free. However, in real applications, these assumptions might be false. For example, the number of training examples might be insufficient, obtaining the labels of training examples is expensive, and only positive and unlabeled examples are available. In essence, the above cases are in the scope of weakly supervised learning [1], which has been a popular research topic in recent years. It covers a broad range of techniques, such as semisupervised learning [2], [3], active learning [4], [5], inaccurate supervision [6], [7], and positive and unlabeled learning [8], [9]. In this work, we focus on inaccurate supervision to study.