Source-linked AI summary
Cross-Subject Transfer Learning in Human Activity Recognition Systems using Generative Adversarial Networks
Elnaz Soleimani, Ehsan Nazerfard
TL;DR
HAR models can lose performance on unseen users, and collecting sufficient labeled data for each newcomer is impractical. SA-GAN uses a GAN framework for semi-supervised cross-subject transfer from labeled source data to unlabeled target data. It outperformed other methods in more than 66% of experiments and reached up to 90% of supervised-model performance in some cases.
Problem
HAR models may perform dramatically worse on unseen users, while collecting and labeling sufficient data for each newcomer is time-consuming and sometimes impractical.
Method
SA-GAN performs semi-supervised instance-based cross-subject transfer using labeled source data, unlabeled target data, and adversarial generator–classifier training.
Results
SA-GAN improved over No Transfer in all reported cases, ranked first in more than 66% of experiments, and reached up to 90% of supervised-model performance in some cases.
Takeaways & Limitations
The results support using SA-GAN to transfer knowledge across subjects when target-user labels are unavailable or limited.
Takeaways & Limitations
Future work identifies GAN mode collapse as a problem to address and proposes more stable GAN versions and multiple-source transfer.
Abstract
from arXiv · showhide
Application of intelligent systems especially in smart homes and health-related topics has been drawing more attention in the last decades. Training Human Activity Recognition (HAR) models -- as a major module -- requires a fair amount of labeled data. Despite training with large datasets, most of the existing models will face a dramatic performance drop when they are tested against unseen data from new users. Moreover, recording enough data for each new user is unviable due to the limitations and challenges of working with human users. Transfer learning techniques aim to transfer the knowledge which has been learned from the source domain (subject) to the target domain in order to decrease the models' performance loss in the target domain. This paper presents a novel method of adversarial knowledge transfer named SA-GAN stands for Subject Adaptor GAN which utilizes Generative Adversarial Network framework to perform cross-subject transfer learning in the domain of wearable sensor-based Human Activity Recognition. SA-GAN outperformed other state-of-the-art methods in more than 66% of experiments and showed the second best performance in the remaining 25% of experiments. In some cases, it reached up to 90% of the accuracy which can be obtained by supervised training over the same domain data.
1. Introduction
Human Activity Recognition models often lose accuracy on unseen users, while collecting enough labeled data for each newcomer is impractical. Transfer learning is introduced as a way to adapt knowledge from training users to new users, motivating this study’s investigation of a proposed solution.
- 1. Introduction: HAR models can achieve promising accuracy with enough labeled data but may perform dramatically worse on data from new users.The shift may involve feature-space or label-space distributions.
- 1. Introduction: Collecting and labeling sufficient data for every new user is time-consuming and sometimes impractical.Human-subject observation creates a practical barrier to retraining for newcomers.
- 1. Introduction: Transfer learning adapts knowledge from source-domain training users to target-domain new users.The approach is presented as a response to performance loss across user domains.
- 1. Introduction: The study investigates how to address these limitations and analyzes the results of its proposed solution.
2. Related Works
Prior HAR and transfer-learning research spans handcrafted and deep representations, several knowledge-transfer categories, and adversarial domain adaptation. However, GAN-based knowledge transfer for improving classification in wearable-sensor HAR remains sparsely studied.
- 2. Related Works: Traditional HAR models such as KNN, HMM, SVM, RF, and Naive Bayes often rely on handcrafted features and may generalize poorly to abstract activities.The passage also notes specialist expertise demands and limited feature abstraction.
- 2. Related Works: Deep HAR approaches use data-driven or model-driven representations, including autoencoders and recurrent architectures, but some incur high resource consumption or low learning rates.
- 2. Related Works: Knowledge-transfer methods include instance, feature-representation, parameter, and relational transfer.These categories differ in whether they transform instances, representations, model knowledge, or data relationships.
- 2. Related Works: Existing transfer approaches include source-instance weighting, common-subspace transformation, TCA, GFK, and semi-supervised cross-domain activity translation.
- 2. Related Works: GAN research in HAR has mainly focused on generating artificial wearable-sensor data, while knowledge transfer for classification remains sparsely studied.The paper contrasts this gap with stronger GAN activity in vision-based problems.
3. Proposed Model: SA-GAN
SA-GAN transforms labeled source-domain samples toward the unlabeled target domain using a generator, discriminator, and classifier. Its adversarial and classification objectives jointly encourage target-like, class-discernible transferred data while addressing mode collapse.
- 3. Proposed Model: SA-GAN: SA-GAN transforms source-domain samples into labeled data resembling the target domain, enabling classifier training for target-domain data.The setting uses labeled source data and unlabeled target data.
- 3. Proposed Model: SA-GAN: The generator creates target-like artificial data, the discriminator distinguishes artificial from real target data, and the classifier assigns labels.
- 3. Proposed Model: SA-GAN: The classifier helps prevent generator mode collapse by evaluating the discernibility of generated outputs during generator updates.Mode collapse occurs when many source inputs are mapped to the same fake output.
- 3. Proposed Model: SA-GAN: Model complexity is intended to match source–target subject distance, with architecture choices including convolutional discriminator layers and residual-block generator and classifier components.
- 3. Proposed Model: SA-GAN: Training alternates mini-batch updates for the discriminator, classifier, and generator while keeping the other two components fixed in each step.Separate optimization steps allow different optimizers, learning rates, and loss functions.
- 3. Proposed Model: SA-GAN: The generator is updated using a gradient that combines discriminator and classifier outputs according to their task factors.The task factors determine how strongly each output contributes to the generator update.
4. Experiments
Experiments evaluate SA-GAN for cross-subject transfer on wearable-sensor HAR data, comparing it with alternative transfer methods and supervised or no-transfer baselines. The results show broad improvements across subject transformations, with performance approaching supervised training in some cases.
- Experimental Setup: The experiments used the Opportunity Challenge dataset, containing wearable-sensor recordings from four subjects performing predefined activities.The evaluation selected the dataset’s most abstract activity level and used separate files for training, validation, and testing.
- Results and Analysis: The evaluation compared SA-GAN with GFK, STL, and KNN+PCA using Weighted F1-measure across source and target subject pairs.Each experiment specified a source subject, target subject, and Wasserstein distance; Table 2 reports the model comparison.
- Results and Analysis: SA-GAN produced the predominant results in more than 66% of experiments and the second-best classification performance in the remaining 25%.The comparison used Weighted F1-measure across source–target subject transformations.
- Results and Analysis: In three experiments, SA-GAN reached up to 90% of the accuracy achieved by supervised training on labeled target-domain data.Supervised target-domain performance was treated as a reference boundary for feasible performance.
- Results and Analysis: For Subject 1 to Subject 2, SA-GAN showed appreciable enrichment despite slight mode collapse on the Relaxing and Sandwich Time classes.Table 3 compares the proposed transfer model with supervised models trained on Subject 2 and Subject 1 data.
- Results and Analysis: For Subject 1 to Subject 2 transfer, SA-GAN achieved W-F1 0.73, near supervised training at 0.75 and above no transfer at 0.45.The comparison used the target subject’s labeled data for the supervised reference.
5. Conclusions and Future Work
The paper presents SA-GAN as a semi-supervised, instance-based approach for cross-subject knowledge transfer in wearable-sensor HAR, addressing scarce labels for newcomers. It reports promising transfer performance, while identifying mode collapse and multi-source transfer as future concerns.
- 5. Conclusions and Future Work: SA-GAN performs semi-supervised instance-based transfer to provide target-domain data for training a HAR classifier.The method applies a Generative Adversarial Network framework to cross-subject knowledge transfer.
- 5. Conclusions and Future Work: SA-GAN reached up to 90% of the supervised model’s performance in some cases.
- 5. Conclusions and Future Work: Future work targets more stable GAN versions to prevent mode collapse and improve recognition results.The authors also intend to examine multiple-source transfer and combinations of models from different source domains.