Human activity recognition (HAR) aims to classify sensory time series into
different activities, with wide applications in activity tracking, healthcare,
human computer interaction, etc. Existing HAR works improve recognition
performance by designing more complicated feature extraction methods, but they
neglect the label semantics by simply treating labels as integer IDs. We find
that many activities in the current HAR datasets have shared label names, e.g.,
"open door" and "open fridge", "walk upstairs" and "walk downstairs". Through
some exploratory analysis, we find that such shared structure in activity names
also maps to similarity in the input features. To this end, we design a
sequence-to-sequence framework to decode the label name semantics rather than
classifying labels as integer IDs. Our proposed method decomposes learning
activities into learning shared tokens ("open", "walk"), which is easier than
learning the joint distribution ("open fridge", "walk upstairs") and helps
transfer learning to activities with insufficient data samples. For datasets
originally without shared tokens in label names, we also offer an automated
method, using OpenAI's ChatGPT, to generate shared actions and objects.
Extensive experiments on seven HAR benchmark datasets demonstrate the
state-of-the-art performance of our method. We also show better performance in
the long-tail activity distribution settings and few-shot settings.

本论文提出了一个基于序列到序列的神经网络框架，以学习共享活动标签名称的语义，而不是将标签仅视为整数 ID，这有助于提高人类活动识别的性能。该方法将学习活动分解为学习共享标记，通过在七个基准人类活动识别数据集上的广泛实验表明其优于现有方法。