Abstract
Human activity recognition (HAR) models suffer significant performance degradation when faced with data heterogeneity (device, users, environments) at test time. Current approaches to this problem using domain adaptation or transfer learning attempt to improve performance in one specific target domain, often using data from said domain. Requiring access to data from the target domain is limiting and cannot be generally assumed. In addition, there is often no single target domain, but rather multiple ones arising from different sources of data heterogeneity. One way to achieve good performance in this setting would be to gather data from all potential domains the model may encounter at deployment - this is generally infeasible.
This work presents the case for training models which are domain-agnostic, i.e., that generalise to unseen test domains. This requires a new way to evaluate models; we discuss a regime called leave-datasets-out, and present a starting benchmark for HAR using binary classification. Two state-of-the-art deep models in the literature are tested; they significantly under-perform in unseen domains when compared to their performance on seen domains. It is shown that under this new evaluation regime, a simple model with an appropriate inductive bias performs at least as well as two current deep models on the benchmark, with a p-value of 5.75x10−4 and 0.13 when testing for a difference in mean accuracy, whilst being at least 10 times faster to train. Additionally, we provide evidence that domain diversity under certain conditions improves performance on both seen and unseen domains. We hope this work provides useful insights to further develop HAR models suitable for real world deployment.
This work presents the case for training models which are domain-agnostic, i.e., that generalise to unseen test domains. This requires a new way to evaluate models; we discuss a regime called leave-datasets-out, and present a starting benchmark for HAR using binary classification. Two state-of-the-art deep models in the literature are tested; they significantly under-perform in unseen domains when compared to their performance on seen domains. It is shown that under this new evaluation regime, a simple model with an appropriate inductive bias performs at least as well as two current deep models on the benchmark, with a p-value of 5.75x10−4 and 0.13 when testing for a difference in mean accuracy, whilst being at least 10 times faster to train. Additionally, we provide evidence that domain diversity under certain conditions improves performance on both seen and unseen domains. We hope this work provides useful insights to further develop HAR models suitable for real world deployment.
Original language | English |
---|---|
Title of host publication | Proceedings of the 2022 ACM International Joint Conference on Pervasive and Ubiquitous Computing |
Place of Publication | New York |
Publisher | Association for Computing Machinery (ACM) |
ISBN (Electronic) | 9781450394239 |
DOIs | |
Publication status | Published - 24 Apr 2023 |
Event | UbiComp/ISWC '22: The 2022 ACM International Joint Conference on Pervasive and Ubiquitous Computing - Cambridge, United Kingdom Duration: 11 Sept 2022 → 15 Sept 2022 |
Publication series
Name | Proceedings of UbiComp: Ubiquitous Computing |
---|---|
Publisher | Association for Computing Machinery (ACM) |
Conference
Conference | UbiComp/ISWC '22 |
---|---|
Country/Territory | United Kingdom |
City | Cambridge |
Period | 11/09/22 → 15/09/22 |
Keywords
- Human activity recognition
- Generalization
- Robustness