I Did Not Smoke 100 Cigarettes Today! Avoiding False Positives in Real-World Activity Recognition

Activity recognition (AR) systems are typically built and evaluated on a predefined set of activities. AR systems work best if the test data contains and only contains these predefined activities. In real world applications, AR systems trained in this manner generate serious false positives, for example if "smoking" is one of the activities in the training data but "lifting weights" is not. Due to the similarity of two activities, an AR system may report a user smoking 100 times a day but he actually did a bicep workout 100 times. In this work, we propose a new approach to train an AR system leveraging the large quantity of unlabeled data which reflects activities users perform in real life. The proposed mPUL (Multi-class Positive and Unlabeled Learning) approach significantly reduces the false positives. We argue that mPUL is a much more effective training method for real-world AR applications.