PaperlayerView exact revision on arXiv

Source-linked AI summary

Stratified Transfer Learning for Cross-domain Activity Recognition

Jindong Wang, Yiqiang Chen, Lisha Hu, Xiaohui Peng, Philip S. Yu

arXiv:1801.00820v1cs.CVcs.LG

TL;DR

Scarce activity labels and differing domain distributions make cross-domain activity recognition difficult, while existing methods often overlook class-level affinity. STL generates target pseudo labels, performs intra-class transfer into shared subspaces, and then annotates the target; across three datasets, it improves classification accuracy by 7.68% over the best baseline.

  • Problem

    Activity recognition often lacks sufficient labels, and existing cross-domain methods typically learn global domain shifts while ignoring intra-class affinity.

  • Method

    STL uses majority voting for target pseudo labels, performs intra-class knowledge transfer into shared subspaces, and applies a second target annotation.

  • Results

    7.68% improvement in classification accuracy over the best baseline GFK is reported across OPPORTUNITY, PAMAP2, and UCI DSADS.

  • Takeaways & Limitations

    STL outperforms the comparison methods in most evaluated cases and performs best across the investigated levels of domain similarity.

Abstract

from arXiv · show

In activity recognition, it is often expensive and time-consuming to acquire sufficient activity labels. To solve this problem, transfer learning leverages the labeled samples from the source domain to annotate the target domain which has few or none labels. Existing approaches typically consider learning a global domain shift while ignoring the intra-affinity between classes, which will hinder the performance of the algorithms. In this paper, we propose a novel and general cross-domain learning framework that can exploit the intra-affinity of classes to perform intra-class knowledge transfer. The proposed framework, referred to as Stratified Transfer Learning (STL), can dramatically improve the classification accuracy for cross-domain activity recognition. Specifically, STL first obtains pseudo labels for the target domain via majority voting technique. Then, it performs intra-class knowledge transfer iteratively to transform both domains into the same subspaces. Finally, the labels of target domain are obtained via the second annotation. To evaluate the performance of STL, we conduct comprehensive experiments on three large public activity recognition datasets~(i.e. OPPORTUNITY, PAMAP2, and UCI DSADS), which demonstrates that STL significantly outperforms other state-of-the-art methods w.r.t. classification accuracy (improvement of 7.68%). Furthermore, we extensively investigate the performance of STL across different degrees of similarities and activity levels between domains. And we also discuss the potential of STL in other pervasive computing applications to provide empirical experience for future research.

I. INTRODUCTION

Cross-domain activity recognition addresses scarce labels by transferring annotated data across related sensor domains with different distributions. STL exploits class-level affinity rather than only global domain shift and reports higher accuracy across three public datasets.

  • Motivation: Label scarcity makes cross-domain activity recognition necessary for annotating unlabeled activities using data from an auxiliary domain.HAR applications require sufficient labeled activity data, but obtaining it is expensive and time-consuming.
  • Motivation: Accelerometer readings from different body locations can have different distributions despite representing the same activity, making cross-domain recognition challenging.The example compares chest and leg sensors from one person.
  • Research gap: Existing approaches learn a global domain shift in one subspace but ignore intra-class affinity, producing looser separation between classes.The paper motivates tighter class clustering through class-specific transfer.
  • Proposed framework: STL performs intra-class knowledge transfer by transforming corresponding source and target classes into the same subspaces.The framework first obtains target pseudo labels, performs class-level transfer, and then applies a second annotation.
  • Evaluation: 7.68% improvement in classification accuracy over five state-of-the-art methods is reported across OPPORTUNITY, PAMAP2, and UCI DSADS.The experiments evaluate STL on three large public activity-recognition datasets.
  • Evaluation: STL is also evaluated across different task similarities and activity levels, with potential applications in other pervasive-computing tasks.The paper presents these analyses as experience for future research.

II. RELATED WORK

Related work spans conventional and deep-learning activity recognition, transfer-learning taxonomies, feature-space adaptation, and activity-specific transfer methods. STL is positioned as a feature-based approach that transfers class-specific structure rather than only global alignment.

  • Activity recognition: Conventional HAR methods generally assume training and test data share the same distribution, unlike cross-domain activity recognition.They commonly use preprocessing, feature extraction, model building, and inference.
  • Activity recognition: Deep-learning HAR can extract high-level features automatically, but cross-domain HAR remains an area with open problems.The paper notes that domain-invariant features may improve cross-domain performance.
  • Transfer learning: Transfer learning is categorized into instance-based, parameter-based, and feature-based methods.The categories distinguish reweighting, model-transfer, and shared-feature-space strategies.
  • Feature-based transfer: Feature-based methods transform source and target data into a shared subspace, while STL does so without modeling domain-feature relationships.The paper places STL in the feature-based category.
  • Activity-recognition transfer: Existing HAR transfer methods variously ignore intra-class similarity, omit feature transformation, or learn only a global domain shift.The cited methods include TransEMDT, TransAct, and heterogeneous transfer learning for HAR.
  • Activity-recognition transfer: STL differs from a classwise joint-subspace method by generating pseudo labels through majority voting on both domains rather than clustering the target alone.Using both domains is presented as a way to obtain more reliable candidates.

III. STRATIFIED TRANSFER LEARNING

STL is a cross-domain framework that uses source-trained classifiers to create target candidates, transfers each class into shared subspaces, and supports later target annotation. Its problem formulation assumes equal dimensionality and label spaces but different distributions.

  • Framework overview: STL has three stages: majority-vote candidate generation, intra-class transfer between source and candidates, and second annotation of the target domain.This sequence is the framework’s main pipeline.
  • A. Problem Definition: CDAR uses a labeled source domain to infer labels for an unlabeled target domain whose feature distributions differ from the source.The domains share dimensionality and label spaces, while both marginal and conditional distributions may differ.
  • Framework overview: STL exploits intra-class affinity to transform corresponding source and target classes into the same subspaces instead of learning only a global shift.The framework assumes samples from the same class lie on an intrinsic subspace.
  • B. Majority Voting: Source-trained classifiers collaboratively generate pseudo labels for target samples, and samples without majority consensus are marked as residuals.Candidates are retained for transfer, while residuals are annotated later.
  • B. Majority Voting: Majority voting may use simple voting, weighted voting, or stacked classifiers, and the base classifiers can be of any type.The paper uses majority consensus to determine candidate labels.

C. Intra-class Transfer

STL measures domain divergence in RKHS and performs class-specific transfer, transforming source and target samples of each class into shared subspaces while preserving data structure.

  • Intra-class Transfer: STL uses maximum mean discrepancy (MMD) to measure divergence between source and target distributions in reproducing kernel Hilbert space (RKHS).The original feature space is avoided because features may be distorted there, while RKHS transfer can be more efficient.
  • Intra-class Transfer: Because target labels are unavailable, STL uses majority-voting pseudo labels to calculate MMD distances separately for each class.The class-specific distances are computed for source samples and target candidates grouped by their labels or pseudo labels.
  • Intra-class Transfer: The intra-class objective combines class-wise MMD distances with regularization, while its constraint preserves structural properties after transformation.The regularization term uses λ as a trade-off parameter.
  • Intra-class Transfer: STL solves a generalized eigen-decomposition problem and selects the m smallest eigenvectors to construct the transformation matrix W.This transformation minimizes domain distance while preserving properties, with same-class source and target data mapped into shared subspaces.

D. Second annotation

The second annotation re-labels target candidates after intra-class transfer, then uses those refined labels to classify the remaining target instances.

  • D. Second annotation: After transformation, STL trains a standard classifier on labeled source data and pseudo-labeled candidates to obtain more reliable candidate predictions.The candidates and source domain now share class-specific subspaces after intra-class transfer.
  • D. Second annotation: The labels of residual target instances are obtained by training a classifier on the candidates and their second-annotation labels.This completes target-domain labeling after the candidates receive more concrete labels.

E. Iterative refinement

STL can iteratively reuse second-annotation results to refine target labels, forming an EM-like procedure that is empirically evaluated.

  • E. Iterative refinement: STL can use second-annotation results as the next initial state and repeat intra-class transfer iteratively.The paper describes this procedure as an EM-like algorithm and reports that its effectiveness is validated experimentally.
  • E. Iterative refinement: After the first iteration, STL refines target labels using previous results rather than repeating majority voting.The framework can also tailor each step, including classifiers and voting techniques, to specific applications.

IV. EXPERIMENTAL EVALUATION

The evaluation tests STL for cross-position activity recognition using three public datasets and tasks spanning multiple source-target similarity levels. It covers dataset preparation, feature extraction, and 22 transfer tasks.

  • Experimental setup: The experiments evaluate STL for cross-domain activity recognition on three public datasets: OPPORTUNITY, PAMAP2, and UCI DSADS.The study uses these datasets to assess STL in cross-position activity recognition.
  • Experimental setup: Cross-position recognition labels activities in one body-part domain using labeled data from another related body-part domain.The setup reflects missing activity labels for some body parts and enables transfer from similar labeled body parts.
  • Experimental setup: The evaluation examines transfer across three similarity scenarios: similar body parts of the same person, different body parts of the same person, and similar body parts across people or datasets.These scenarios vary the similarity between source and target domains.
  • Datasets and preprocessing: The preprocessing combines three sensor axes, applies 5s sliding windows, and extracts 81 features per body position from three sensors.The 81 features comprise 27 time- and frequency-domain features per sensor across accelerometer, gyroscope, and magnetometer data.
  • Experimental setup: 22 cross-domain tasks are constructed, using all classes for within-dataset scenarios and four common classes for cross-dataset tasks.The four cross-dataset classes are Walking, Sitting, Lying, and Standing; the least-similar different-body-part/different-person scenario is excluded.

C. Comparison Methods and Implementation Details

The evaluation compares STL with dimensionality reduction and transfer-learning baselines on CDAR tasks, using consistent classifiers, dimensions, and accuracy evaluation.

  • Comparison methods: Five comparison methods are adopted: PCA, KPCA, TCA, GFK, and TKL.PCA and KPCA are dimensionality-reduction methods; TCA, GFK, and TKL are transfer-learning approaches.
  • Implementation details: CDAR tasks use target labels only for testing, with dimensionality-reduction methods evaluated at the same dimension.A classifier with the same parameter is learned using the source domain before target-domain prediction.
  • Evaluation: The experiments report classification accuracy for STL and comparison methods across different CDAR tasks in Table III.The paper uses “T” for Torso / Back / Chest across the three datasets.
  • Implementation details: All six methods use a random forest classifier with #Tree = 30 as the final classifier.STL's majority voting uses SVM (C = 100), kNN (k = 3), and random forest (#Tree = 30).
  • Evaluation: Classification accuracy on the target domain is the evaluation metric.The metric is widely used in existing transfer-learning methods.

D. Classification Accuracy of STL

STL achieves the strongest classification accuracy in most CDAR comparisons and remains best as domain similarity decreases, supporting intra-class transfer beyond global domain alignment.

  • Overall comparison: 7.68% improvement over the best baseline GFK is reported for STL in most CDAR cases.STL improves accuracy by 10% ∼20% over PCA and KPCA and by 5% ∼15% over TCA, GFK, and TKL.
  • Overall comparison: STL strongly outperforms GFK in less similar scenarios, indicating greater robustness across different activity levels.STL performs better than GFK in most cases with significant improvement.
  • Similarity effects: Accuracy drops for all methods as domain similarity decreases, while STL performs best in all tested similarity scenarios.The comparison uses average classification accuracy shown in Fig. 5(a).
  • Similarity effects: Performance is best for similar body parts in the same person, worse across different body parts, and worst across different people.Examples include RA →LA, RA →T, and T →T across datasets.
  • Similarity effects: RUA →T outperforms RLA →T because Right Upper Arm is more similar to Torso than Right Lower Arm is.The passage reports the comparison within the same dataset and degree of similarity.
  • Implications: The experiments identify source–target similarity as important for CDAR and emphasize selecting an appropriate auxiliary domain.Body structure and moving patterns contribute to similarity in the activity-recognition experiments.

F. Performance on Different Levels of Activities

STL is evaluated across activity granularities and implementation factors, with middle-level activities performing best and iterative transfer converging quickly.

  • Activity levels: Middle-level activities achieve the best performance, followed by low-level activities, while high-level activities perform worst.The analysis uses OPP-LL, OPP-ML, and OPP-HL activities from the OPPORTUNITY dataset.
  • Activity levels: Body-part atomic movements support transfer better than high-level activities containing contextual information about environments or objects.OPP-ML can outperform OPP-LL because finer-grained activities better capture similarities between body parts.
  • Effectiveness verification: STL can achieve good performance with fewer pseudo-label candidates and does not largely rely on candidate confidence.Candidate usage is varied from 10% to 100% on LA →RA in DSADS, comparing STL with PCA and TCA.
  • Effectiveness verification: STL iteratively improves accuracy with weak majority-voting classifiers, remains robust to classifier choice, and converges within fewer than 10 iterations.1NN-STL performs slightly worse than STL using random forest; more reliable classifiers are suggested for real problems.
  • Deployment potential: STL supports parallel deployment because majority-voting classifiers can operate independently and intra-class transfer can be performed separately for each class.The passage identifies parallel implementation as important for real applications.
  • Effectiveness verification: Detailed STL results, classifier comparisons, and parameter sensitivity are presented in Fig. 6.The figure covers increasing candidate usage for RA →LA on DSADS, 1NN-STL comparisons, and parameter sensitivity.

H. Parameter Sensitivity

STL remains accurate across tested subspace dimensions, with little sensitivity to reducing dimension and reported robustness to the trade-off parameter. The framework is presented as adaptable to varied pervasive-computing applications and deployment settings.

  • Trade-off parameter λ: The trade-off parameter λ is omitted from evaluation because experiments verified its robustness.The omission is attributed to page limitations.
  • Dimension m: STL achieves the best accuracy across dimensions m ∈{10, 20, 30, 40}.The comparison includes STL and other dimensionality reduction methods.
  • Dimension m: STL accuracy almost does not change as m decreases, indicating robustness across different dimensions.The authors conclude that STL is more effective and robust than the compared methods under different dimensions.
  • Application scope: Each STL step can be tailored to specific applications, supporting use across cross-device, cross-user, and cross-position activity recognition.The paper emphasizes measuring source-target similarity before applying transfer learning.
Loading 1801.00820v1…