[Paper Review] Unsupervised Statistical Feature-Guided Diffusion Model for Sensor-based Human Activity Recognition

Si Zuo, Vítor Fortes Rey+3 more|arXiv (Cornell University)|May 30, 2023

Anomaly Detection Techniques and Applications4 citations

TL;DR

This paper proposes an unsupervised statistical feature-guided diffusion model (SF-DM) that generates high-quality synthetic sensor time-series data for human activity recognition (HAR) without requiring labeled data. By conditioning the diffusion process on statistical features like mean, standard deviation, Z-score, and skewness, SF-DM produces diverse, realistic synthetic data that significantly improves HAR classifier performance, outperforming SMOTE, SVM-SMOTE, and TimeGAN across multiple benchmark datasets.

ABSTRACT

Human activity recognition (HAR) from on-body sensors is a core functionality in many AI applications: from personal health, through sports and wellness to Industry 4.0. A key problem holding up progress in wearable sensor-based HAR, compared to other ML areas, such as computer vision, is the unavailability of diverse and labeled training data. Particularly, while there are innumerable annotated images available in online repositories, freely available sensor data is sparse and mostly unlabeled. We propose an unsupervised statistical feature-guided diffusion model specifically optimized for wearable sensor-based human activity recognition with devices such as inertial measurement unit (IMU) sensors. The method generates synthetic labeled time-series sensor data without relying on annotated training data. Thereby, it addresses the scarcity and annotation difficulties associated with real-world sensor data. By conditioning the diffusion model on statistical information such as mean, standard deviation, Z-score, and skewness, we generate diverse and representative synthetic sensor data. We conducted experiments on public human activity recognition datasets and compared the method to conventional oversampling and state-of-the-art generative adversarial network methods. Experimental results demonstrate that this can improve the performance of human activity recognition and outperform existing techniques.

Motivation & Objective

• To address the challenge of scarce and costly-to-annotate labeled sensor data in human activity recognition (HAR).
• To develop a data generation method that preserves temporal and statistical properties of real sensor time-series data without relying on class labels.
• To improve HAR model performance by leveraging synthetic data generated from unlabeled real data.
• To evaluate the effectiveness of statistical feature-guided conditioning in enhancing the quality and diversity of generated sensor data.

Proposed method

• The method employs a diffusion model conditioned on statistical features—mean, standard deviation, Z-score, and skewness—extracted from unlabeled sensor time-series data.
• A two-stage training framework is used: first, the SF-DM is pre-trained on a large set of unlabeled sensor data; second, a separate HAR classifier is fine-tuned using a small amount of real labeled data and the synthetic data.
• The diffusion model uses an encoder-decoder architecture to denoise corrupted time-series data, guided by statistical feature embeddings.
• Statistical features are computed over sliding windows of sensor data (e.g., 400, 200, 200 for MM-FIT, PAMAP2, and Opportunity) and used as conditioning signals during generation.
• The model is trained to reverse a forward diffusion process that gradually adds noise to real sensor sequences, learning to reconstruct them from noise using statistical priors.
• Performance is evaluated using accuracy and macro F1 score on HAR tasks, with comparisons to SMOTE, SVM-SMOTE, and TimeGAN under identical training protocols.

Experimental results

Research questions

• RQ1 Can an unsupervised diffusion model conditioned on statistical features generate synthetic sensor data that improves HAR performance without labeled data?
• RQ2 How does the statistical feature-guided diffusion model compare to conventional oversampling and GAN-based methods in generating diverse and representative sensor data?
• RQ3 To what extent does the inclusion of statistical features (e.g., skewness, Z-score) enhance the quality and utility of generated synthetic time-series data?
• RQ4 How does the performance of the HAR classifier scale with varying amounts of real labeled data when pre-trained on synthetic data from SF-DM?

Key findings

• SF-DM achieved up to a 5% improvement in macro F1 score over the baseline on the MM-FIT and Opportunity datasets when using only 20% real data.
• On the same datasets, SF-DM outperformed TimeGAN by 7% to 13% in macro F1 score, demonstrating superior generalization and data quality.
• On the PAMAP2 dataset, SF-DM maintained performance comparable to the baseline and slightly outperformed it, with no degradation observed.
• The ablation study showed that SF-DM[Corresp.P]—using statistical features instead of labels—consistently outperformed CC-DM (class-conditioned diffusion model), especially in low-data regimes.
• When training on limited unlabeled data, SF-DM performance degraded, indicating that sufficient data is necessary for stable diffusion model training.
• The method demonstrated robustness in handling class imbalance, with macro F1 being a more reliable metric than accuracy, and SF-DM consistently improved it across datasets.