Research map / Label-efficient and robust learning / Recognition and long-tail
Fine-grained Control of Generative Data Augmentation in IoT Sensing
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 fine-grained controllable generative data augmentation for IoT sensing by conditioning diffusion and VAE generators on interpretable STFT spectrogram statistical metrics and augmenting within that metric space.
Problem
IoT sensing models overfit under domain shift, and existing augmentation is either hand-crafted transformations with artifacts or label-only conditional generative models that offer no fine-grained control.
Approach
Define a metric space of spectrogram statistics (temporal amplitude range, F0 amplitude, contrast, flatness, entropy), condition a generative model on these metrics with an added metric loss, and augment by interpolating, disturbing, or domain-knowledge-guided editing of metric values.
Key results
- Diffusion-based Disturb augmentation with DeepSense reaches 0.8307 accuracy on HAR versus 0.7752 with no augmentation and 0.7531 for cGAN-style label-only augmentation (cGAN 0.7531 with DeepSense).
- The Reconst upper bound (validation metrics as conditions) reaches 0.8618 accuracy on HAR with DeepSense using the diffusion backbone.
- In the HAR running-style case study, domain-knowledge instruction substantially improves over no augmentation and closely approaches the Reconst upper bound.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use fine-grained metric disturbance instead of label-only conditional generation | Diffusion augmentation with DeepSense · Human Activity Recognition | accuracy: 0.8034 -> 0.8307 | Helped |
| use fine-grained metric disturbance instead of label-only conditional generation | Diffusion augmentation with Transformer · Human Activity Recognition | accuracy: 0.8155 -> 0.8205 | Helped |
| use fine-grained metric disturbance instead of transformation augmentation | Diffusion augmentation with DeepSense · Vehicle Detection | accuracy: 0.8024 -> 0.8200 | Helped |
| use fine-grained metric disturbance instead of label-only conditional VAE | Diffusion augmentation with Transformer · Vehicle Detection | accuracy: 0.7720 -> 0.8008 | Helped |
| use fine-grained metric disturbance instead of label-only conditional diffusion | Diffusion augmentation with DeepSense · Harmful Brain Activity Recognition | accuracy: 0.7374 -> 0.7511 | Helped |
| use fine-grained metric disturbance instead of label-only conditional diffusion | Diffusion augmentation with Transformer · Harmful Brain Activity Recognition | accuracy: 0.7225 -> 0.7335 | Helped |
| add time-domain metric (ablation) | Diff-disturb with DeepSense · Human Activity Recognition | accuracy: 0.8290±0.0202 -> 0.8307±0.0171 | Helped |
| add frequency-domain metric (ablation) | Diff-disturb with DeepSense · Human Activity Recognition | accuracy: 0.8177±0.0155 -> 0.8307±0.0171 | Helped |
| add time-frequency-domain metrics (ablation) | Diff-disturb with DeepSense · Human Activity Recognition | accuracy: 0.8112±0.0166 -> 0.8307±0.0171 | Helped |
| add time-domain metric (ablation) | Diff-disturb with DeepSense · Vehicle Detection | accuracy: 0.7745±0.0122 -> 0.7925±0.0053 | Helped |
| add frequency-domain metric (ablation) | Diff-disturb with DeepSense · Vehicle Detection | accuracy: 0.7910±0.0068 -> 0.7925±0.0053 | Helped |
| add time-frequency-domain metrics (ablation) | Diff-disturb with DeepSense · Vehicle Detection | accuracy: 0.7880±0.0173 -> 0.7925±0.0053 | Helped |
Surprising findings
- The VAE backbone yields less crisp spectrograms and weaker augmentation gains than the diffusion backbone, so augmentation effectiveness depends heavily on generator quality.
Limitations
- Metrics are selected empirically rather than via a quantitative importance analysis, which may introduce bias.
- Deciding the proper domain-knowledge augmentation strategy relies heavily on practitioner expertise.
- The VAE backbone generates less realistic spectrograms, limiting augmentation benefit with that generator.
Experimental setup
Datasets: RealWorld-HAR Vehicle detection (self-collected seismic and acoustic) Harmful Brain Activity Classification (EEG)
Benchmarks: RealWorld-HAR Vehicle detection dataset Harmful Brain Activity Classification
Metrics: accuracy F1 score
Baselines: mixup (data space interpolation) feature space interpolation time and frequency domain transformation cDiff cVAE cGAN
Builds on: mixup NVAE conditional diffusion model harmonic product spectrum
Closest papers by meaning
- MeRino: Entropy-Driven Design for Generative Language Models on IoT Devices (AAAI 2025 · similarity 0.80)
- Real-Time Calibration Model for Low-Cost Sensor in Fine-Grained Time Series (AAAI 2025 · similarity 0.79)
- Mechanistic Circuit Identification for Controllable Data Generation (NeurIPS 2026 · similarity 0.79)
- UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation (NeurIPS 2025 · similarity 0.78)
- ScatterAD: Temporal-Topological Scattering Mechanism for Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.77)
- Synthetic Series-Symbol Data Generation for Time Series Foundation Models (NeurIPS 2025 · similarity 0.77)
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.