Transfer Learning for Activity Recognition
Derek Hao Hu
Activity recognition aims to recognize and predict human activities based on series of sensor readings. In recent years, using machine learning methods to solve activity recognition problems becomes popular. One special difficulty for adopting machine learning methods, or more specifically, supervised learning methods, is the workload to annotate a large number of sensor readings. However, in reality, labeling sensor readings to their corresponding activities is a time-consuming task. In practice, we usually have a set of labeled training instances for an activity recognition task and hopefully we can transfer such knowledge to a new recognition scenario. In this paper, we propose a transfer learning method based on distribution matching to solve this problem. Next, by validating our method on different datasets and comparing our method to previous approaches of activity recognition, we had demonstrated the effectiveness of our algorithm and that our method can indeed relieve the need to label new sensor readings.