Source-linked AI summary
Multi-Source Deep Domain Adaptation with Weak Supervision for Time-Series Sensor Data
Garrett Wilson, Janardhan Rao Doppa, Diane J. Cook
TL;DR
Time-series domain adaptation lacks robust, efficient methods despite costly labeling and domain variation. This paper proposes CoDATS, extends it to multiple sources and target-label proportions, and reports improved accuracy and training efficiency, with gains depending on label-distribution differences and domain similarity.
Problem
Time-series sensor labels are expensive or infeasible to collect, while prior domain-adaptation work has focused mainly on images and limited time-series methods can be slow.
Method
CoDATS uses a convolutional time-series architecture with domain-invariant adaptation, and DA-WS incorporates known target-domain label proportions as a regularizing constraint.
Results
CoDATS improves over prior time-series adaptation work in accuracy and training-time efficiency, with additional accuracy gains from multiple sources and target-label proportions.
Takeaways & Limitations
CoDATS supports single- and multi-source time-series adaptation and can use weak target-domain distribution information when such information is available.
Takeaways & Limitations
DA-WS helps only when target and source label distributions differ, while forcing invariance across many distinct domains can reduce discriminative ability through negative transfer.
Abstract
from arXiv · showhide
Domain adaptation (DA) offers a valuable means to reuse data and models for new problem domains. However, robust techniques have not yet been considered for time series data with varying amounts of data availability. In this paper, we make three main contributions to fill this gap. First, we propose a novel Convolutional deep Domain Adaptation model for Time Series data (CoDATS) that significantly improves accuracy and training time over state-of-the-art DA strategies on real-world sensor data benchmarks. By utilizing data from multiple source domains, we increase the usefulness of CoDATS to further improve accuracy over prior single-source methods, particularly on complex time series datasets that have high variability between domains. Second, we propose a novel Domain Adaptation with Weak Supervision (DA-WS) method by utilizing weak supervision in the form of target-domain label distributions, which may be easier to collect than additional data labels. Third, we perform comprehensive experiments on diverse real-world datasets to evaluate the effectiveness of our domain adaptation and weak supervision methods. Results show that CoDATS for single-source DA significantly improves over the state-of-the-art methods, and we achieve additional improvements in accuracy using data from multiple source domains and weakly supervised signals. Code is available at: https://github.com/floft/codats
1 INTRODUCTION
Time-series sensor labels are costly, while existing domain-adaptation research has focused largely on images and slower recurrent approaches. The paper introduces CoDATS and extends adaptation to multiple sources and target-domain label proportions.
- Time-series sensor data support applications including activity, sleep-stage, gesture, speech, and medical classification, but labeling can be expensive or infeasible.
- Prior time-series adaptation methods use recurrent networks that can be slow to train, despite convolutional networks offering comparable accuracy with faster training and evaluation.
- CoDATS is a convolutional domain-adaptation model designed to improve efficiency, accuracy, and compatibility for time-series data.
- CoDATS extends beyond single-source adaptation to multiple source domains, especially for datasets with high variability between domains.
- Domain Adaptation with Weak Supervision uses target-domain label proportions, which may be easier to collect than instance-level time-series labels.
- Experiments on diverse real-world benchmarks report improvements from CoDATS, multiple sources, and the proposed weak-supervision method.
2 RELATED WORK
Related work includes broad single-source image adaptation, less-developed multi-source methods, and several specialized time-series approaches. The paper positions VRADA and R-DANN as its closest directly comparable time-series baselines and distinguishes its weak-supervision setting from prior work.
- Prior domain-adaptation research has developed numerous strategies for single-source image data, while multi-source adaptation has received less attention.
- The paper provides code for the proposed work at the stated project repository.
- VRADA and R-DANN apply recurrent feature extractors with adversarial domain-invariance objectives to time-series adaptation.
- Other time-series approaches address specialized tasks using instance weighting, autoencoders, GANs, RNNs, or combined CNN-RNN architectures.
- VRADA and R-DANN are included as baselines because they were developed for time-series classification and can be directly compared with CoDATS.
- The proposed weak-supervision approach is inspired by posterior regularization but targets raw time-series sensor data rather than natural-language processing.
3 PROBLEM SETUP
The setup reuses labeled source-domain data and unlabeled target-domain time series for classification under domain shift. The weak-supervision extension additionally supplies target-domain label proportions as a training constraint.
- Domain adaptation transfers labeled data from one domain to a related target domain where labels may be unavailable.
- Domain shifts can arise from differences in activity patterns, sensor positions, or sampling rates, motivating domain-invariant feature extraction.
- Self-reported activity time can provide target-domain label proportions that regularize a model toward domain invariance and distributional consistency.
- Single-source adaptation uses n = 1 source domain, whereas multi-source adaptation uses n > 1 source domains.
- Evaluation uses held-out target samples drawn independently from the target distribution, with ground-truth labels used to compute accuracy.
- For DA-WS, the input is time-series data plus a discrete target-domain label distribution giving probability p_y for each label y.
4 CODATS FOR DOMAIN ADAPTATION
CoDATS combines domain-invariant adaptation with a convolutional architecture for time-series data, supporting both single-source and multi-source settings. Its adversarial training aligns domains while preserving source-task classification, and its fully convolutional design handles temporal data efficiently and with variable lengths.
- Adaptation Method: CoDATS adapts domain-invariant training to time-series data and supports both single-source and multi-source domain adaptation.Its method is designed around a feature extractor, task classifier, and domain classifier.
- Adaptation Method: The feature extractor learns features that support source-task classification while making domain classification difficult through adversarial optimization.The domain classifier is trained to predict domains, whereas the feature extractor is updated to produce domain-invariant representations.
- Adaptation Method: The gradient reversal layer combines the competing feature/task and domain-classifier objectives into one optimization step.Its forward behavior is the identity, while its backward behavior reverses and scales gradients using λ.
- Adaptation Method: Equation 8 yields an adaptation method capable of handling both single-source (n = 1) and multi-source (n > 1) domain adaptation.The domain labels distinguish the target domain from the n source domains in the multi-class domain-classification objective.
- Model Architecture: CoDATS uses a fully convolutional feature extractor and global average pooling to process time-series data, including variable-length sequences.The architecture is proposed as an alternative to recurrent networks, which pose training challenges involving vanishing and exploding gradients.
5 DA WITH WEAK SUPERVISION
DA-WS extends CoDATS by using known target-domain label proportions as weak supervision during training. It regularizes predictions so their aggregate distribution matches the target proportions while preserving source-label learning and domain-invariant representations.
- DA-WS method: DA-WS incorporates known target-domain label proportions by searching for model parameters whose predicted target labels approximately match them.The method introduces a differentiable regularization term into the training objective.
- Predicted distribution: The predicted target-domain distribution is formed from the task classifier’s softmax probabilities for each label.For target instance x, C(F(x))_y represents the probability assigned to label y.
- Distribution alignment: DA-WS matches the aggregate predicted distribution to the known target distribution using a distance measure such as Kullback-Leibler divergence.The alignment is computed over multiple predictions rather than requiring every individual prediction to match the target proportions.
- Training objective: The combined objective learns source-task labels, domain-invariant features across domains, and target predictions aligned with known label proportions.These objectives are jointly incorporated into the DA-WS training objective.
- Scope of benefit: DA-WS provides additional information only when target and source label distributions differ; otherwise, its KL regularizer becomes zero under empirical risk minimization.The experiments corroborate this condition for improvement.
6 EXPERIMENTAL SETUP
The experiments evaluate CoDATS and DA-WS on diverse time-series domain-adaptation benchmarks, beginning with single-source comparisons and extending to multiple sources. The setup also tests whether target label proportions provide additional performance information.
- CoDATS evaluation: CoDATS is compared with prior methods and baselines on diverse time-series adaptation benchmarks using accuracy as the evaluation metric.The initial comparison focuses on single-source adaptation, the setting studied by prior work.
- Multi-source adaptation: The evaluation tests whether multi-source adaptation yields additional performance gains beyond single-source adaptation.The model is designed to support both settings.
- Weak supervision evaluation: DA-WS experiments test whether coupling weak supervision with CoDATS further improves accuracy.The experiments compare DA-WS against methods that do not use the additional weak supervision information.
- Datasets and domains: Datasets are split into multiple domains using participant identifiers, representing differences among data-collection participants.The selected datasets include participant-level domain structure for adaptation experiments.
- Datasets and domains: HAR contains sensor data from 30 participants, while HHAR contains accelerometer data from 31 smartphones with varied manufacturers, models, and orientations.HHAR is included as a more diverse activity-recognition dataset than HAR.
- Weak supervision evaluation: Target label-proportion information is simulated by estimating label proportions from the target domain’s training data.Both single-source and multiple-source settings are evaluated.
7 RESULTS AND DISCUSSION
Experiments show that CoDATS improves single-source adaptation accuracy, consistency, and training efficiency, while multiple source domains and weak supervision can provide further gains. Performance generally improves as more source domains are added, although negative transfer may appear in some settings.
- Single-Source CoDATS: CoDATS outperformed both prior methods on all but 4 of 40 adaptation problems, including one tie.It also outperformed the lower bound on all but 8 problems, including two ties.
- Single-Source CoDATS: CoDATS typically produced more consistent results across random initializations, with lower standard deviations than R-DANN and VRADA on all but 14 of 40 problems.
- Single-Source CoDATS: CoDATS required only 20-22% of R-DANN’s training time and 6% of VRADA’s training time across the evaluated datasets.It was faster than both baselines on all datasets and took less than twice the lower-bound training time.
- Multi-Source CoDATS: Negative transfer may cause CoDATS performance to decline with many source domains, especially on WISDM AR after more than 20 source domains.The authors suggest this effect may be less noticeable when activity or gesture-recognition settings involve fewer than 20 participants.
- Multi-Source CoDATS: Multi-source adaptation increased average performance over single-source adaptation, with CoDATS curves rising as additional source domains were included.Domain adaptation still outperformed the lower bound at the highest source-domain count on HHAR and uWave.
- Domain Adaptation with Weak Supervision: CoDATS-WS improved accuracy most on WISDM, by 5.8%, while changes on balanced datasets ranged from -1.5% to +1.8%.The method maintained the same training efficiency and could further improve multi-source adaptation.
8 CONCLUSIONS AND FUTURE WORK
The paper concludes that CoDATS improves time-series domain adaptation in accuracy and training-time efficiency, with further gains from multiple sources and weak supervision.
- CoDATS improves time-series domain adaptation in both accuracy and training-time efficiency over prior work.
- Using multiple source domains can provide additional accuracy gains beyond single-source adaptation.
- Weak supervision based on known target-domain label proportions can provide additional accuracy gains with CoDATS.
- Future work includes heterogeneous feature sets, additional forms of weak supervision, and model improvements informed by InceptionTime.
A.1 Experimental Setup
The experiments use controlled dataset splits, repeated random initializations, validation-based model selection, and domain-specific batching for single-source, multi-source, and weakly supervised adaptation.
- 30,000 iterations, Adam with learning rate 0.0001, batch size 128, and the DANN learning-rate schedule define model training.
- 80%-20% training-test splits are further divided into training-validation sets, while target test data remains unavailable during evaluation.
- Single-source experiments sample 10 adaptation problems, and results average three random initializations with standard deviations on holdout test data.
- Models are selected using holdout validation performance from evaluations every 4,000 iterations and at the end of training.
- Multi-source batches weight domains uniformly, while DA-WS requires enough target predictions per batch to estimate label proportions.
- For multi-source training, half the batch is allocated across sources and half to the target to avoid performance degradation as source count increases.
A.2 Datasets
The study uses sensor datasets spanning human activity and gesture recognition, with participant-based domains, fixed-window preprocessing, and varying label balance across datasets.
- HAR uses normalized three-axis sensor data from 30 participants, sampled at 50 Hz and segmented into 128-step windows.
- HHAR experiments use accelerometer data from 31 smartphones, segmented into non-overlapping windows of 128 time steps.
- WISDM AR includes 33 participants with sufficient labeled data, sampled at 20 Hz and segmented into non-overlapping 128-step windows.
- uWave data cover 8 participants over 7 days, with gestures zero-padded to 315 time steps for batch consistency.
- UCI HAR, UCI HHAR, and uWave are relatively balanced between domains, whereas WISDM AR has large cross-domain label-proportion differences.