Human Activity Recognition (HAR) systems—the technology powering fitness trackers and health monitors—face a persistent "last mile" problem. While a deep learning model trained on thousands of individuals might achieve high average accuracy, it often fails when deployed on a new user. This failure stems from the fact that human movement is deeply idiosyncratic; factors like age, gait, height, and even device placement create a significant distribution shift between the training population and the individual wearer.
Researchers are moving away from "one-size-fits-all" models toward adaptive frameworks that can bridge this inter-subject gap through personalization.
The Challenge of Inter-Subject Variability
The core difficulty in HAR is that the variation within a single activity class across different people—how you run versus how I run—can exceed the variation between different activity classes. As noted in The Empirical Study On Deep Convolutional Network Transferring Among Users Within Activity Recognition, this high intra-class variance causes models to misidentify activities entirely when transitioning to a new user.
This is not just a software issue but a biomechanical one. Md. Fokhrul Islam and colleagues highlight a "simulation-to-reality gap" in Fusing Biomechanical and Spatio-Temporal Features for Fall Prediction, where models trained on stunt actors fail to generalize to the diverse kinematic profiles of older adults. To be effective, a model must account for these "behavioral fingerprints."
Bridging the Gap with Domain Adaptation
To address this, researchers use Unsupervised Domain Adaptation (UDA) to align the data distribution of a "source" population with a "target" individual without requiring the new user to manually label their own data.
One prominent approach is SWL-Adapt, developed by Ling Chen and Rong Hu. In SWL-Adapt: An Unsupervised Domain Adaptation Model with Sample Weight Learning, the authors use meta-optimization to learn which samples from the source population are most relevant to the target user. Similarly, the MUCT framework by Qi Jia and Jing Guo combines deep domain adaptation with a "consistency filter" to generate high-confidence pseudo-labels for the new user's data, allowing the model to self-correct.
Personalization with Minimal Data
If a model requires hours of labeled data to adapt, it becomes impractical for real-world use. The research direction has therefore pivoted toward few-shot learning and active learning.
- Few-Shot Learning: In Matching networks for personalised human activity recognition, Nirmalie Wiratunga and Stewart Massie demonstrate that a "Matching Network" can achieve high accuracy using just 30 seconds of data from a new user. This avoids the "averaging effect" of traditional machine learning that often obscures individual nuances.
- Active Learning: Frameworks like ActiveHARNet use Bayesian uncertainty estimation to identify the most "confusing" data points for a specific user. By asking the user to label only these specific points, the system reduces the human effort required for personalization by over 60%.
Decentralized and Heterogeneous Personalization
As privacy concerns grow, the field is exploring Federated Learning (FL), where models are trained across multiple devices without sharing raw sensor data. However, standard FL assumes all users are similar, which contradicts the reality of inter-subject variability.
To solve this, François Portet and Sannara Ek introduced FedDist in Evaluation and comparison of federated learning algorithms for Human Activity Recognition on smartphones. Instead of simply averaging model weights, FedDist identifies and preserves unique neural network features specific to diverse users. Other researchers, such as Qiang Shen in Federated Multi-Task Attention for Cross-Individual Human Activity Recognition, use attention mechanisms to help the global model focus on the most relevant features for each individual's unique gait and habits.
Continual Adaptation
Finally, personalization is not a one-time event. A user's movement patterns may change due to injury, aging, or even changing footwear. CLAD-Net, proposed by Edison Thomaz and Hassan Ghasemzadeh, uses a self-supervised Transformer to learn general movement patterns while employing a distilled CNN to adapt to specific activity labels over time. This allows the system to evolve alongside the user, ensuring that the "last mile" of accuracy remains closed throughout the device's lifecycle.