Research map / Label-efficient and robust learning / Recognition and long-tail
Generalizable Sensor-Based Activity Recognition via Categorical Concept Invariant Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CCIL, a regularizer for domain-generalized sensor-based activity recognition that aligns per-sample concept matrices (classifier weights times features) with a momentum-updated class-wise mean, and reports gains over ERM and DG baselines on four HAR benchmarks.
Problem
Sensor-based human activity recognition degrades on unseen subjects, positions, or datasets because of distribution shift from inter-subject variability; existing domain-invariant methods only regularize penultimate features.
Approach
Builds a concept matrix from the elementwise contributions of features times classifier weights, and penalizes deviation of same-class samples' concept matrices from a dynamically momentum-updated class mean, jointly capturing feature- and logit-invariance; added to cross-entropy with a few lines of code.
Key results
- Cross-person setting: average accuracy 90.7% on DSADS vs 88.2% for DIVERSIFY (ERM 80.3%), 77.0% on USC-HAD, 84.5% on PAMAP2.
- Cross-position on DSADS: 38.7% average vs 37.4% for DIVERSIFY (ERM 30.6%).
- One-person-to-another: 63.8% average vs 61.7% for DIVERSIFY; cross-dataset: 52.6% vs 51.1% for DIVERSIFY.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CCIL rather than DIVERSIFY | CNN for cross-person HAR · DSADS | accuracy (AVG): 88.2 -> 90.7 | Helped |
| Apply CCIL rather than DIVERSIFY | CNN for cross-person HAR · USC-HAD | accuracy (AVG): 74.0 -> 77.0 | Helped |
| Apply CCIL rather than DIVERSIFY | CNN for cross-person HAR · PAMAP2 | accuracy (AVG): 80.8 -> 84.5 | Helped |
| Apply CCIL rather than DIVERSIFY | CNN for cross-position HAR · DSADS | accuracy (AVG): 37.4 -> 38.7 | Helped |
| Apply CCIL rather than DIVERSIFY | CNN for cross-dataset HAR · Cross-dataset | accuracy (AVG): 51.1 -> 52.6 | Helped |
| Apply CCIL rather than DIVERSIFY | CNN for one-person-to-another HAR · One-person-to-another | accuracy (AVG): 61.7 -> 63.8 | Helped |
| Add concept-matrix invariance instead of using feature-invariance alone (ablation) | CCIL for cross-person HAR · DSADS | accuracy (AVG): 88.6 -> 90.7 | Helped |
| Add concept-matrix invariance instead of using logit-invariance alone (ablation) | CCIL for cross-person HAR · DSADS | accuracy (AVG): 84.5 -> 90.7 | Helped |
| Use dynamic momentum update rather than setting λ=0 (ablation) | CCIL for cross-person HAR · DSADS | accuracy (AVG): 85.5 -> 90.7 | Helped |
| Use dynamic momentum update rather than fixing the mean concept matrix from the initial pretrained model (λ=1) (ablation) | CCIL for cross-person HAR · DSADS | accuracy (AVG): 89.3 -> 90.7 | Helped |
Surprising findings
- Most existing DG methods fail to consistently improve over the ERM baseline under these rigorous HAR settings.
- Logit-invariance alone performs only slightly better than ERM, while feature-invariance alone is stronger in the DSADS ablation.
Limitations
- Gains over the best baseline are modest (about 1-3 points on several settings), with reported averages over three trials only.
- Evaluated mainly with a small CNN backbone; Transformer results only in supplementary material on DSADS.
- Hyperparameter sensitivity (momentum lambda and alpha) relies on choosing values on the benchmark sets.
Experimental setup
Datasets: DSADS USC-HAD PAMAP2 UCI-HAR
Benchmarks: DSADS USC-HAD PAMAP2 UCI-HAR
Metrics: accuracy
Baselines: ERM DANN CORAL Mixup GroupDRO RSC ANDMask GILE AdaRNN DIVERSIFY
Builds on: empirical risk minimization DIVERSIFY GILE AdaRNN
Closest papers by meaning
- HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors (AAAI 2024 · similarity 0.84)
- Scalable Hierarchical Deep Neural Network for Time Series Analysis in Wearable Sensor-based Human Activity Recognition (AAAI 2025 · similarity 0.83)
- Limitations in Employing Natural Language Supervision for Sensor-Based Human Activity Recognition - And Ways to Overcome Them (AAAI 2025 · similarity 0.83)
- MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition (NeurIPS 2025 · similarity 0.83)
- ZeroHAR: Sensor Context Augments Zero-Shot Wearable Action Recognition (AAAI 2025 · similarity 0.83)
- Domain-Informed Label Fusion Surpasses LLMs in Free-Living Activity Classification (Student Abstract) (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.