Diffusion-Based Synthetic Data Pretraining for Enhancing Activity Recognition

E. Riveros, D. Vega-Oliveros, A. Soriano-Vargas, A. Rocha

Summary

Diffusion model generates synthetic smartwatch accelerometer/gyroscope/magnetometer windows; CABiGRU (conv + BiGRU + multi-head attention + residual) is pre-trained on synthetic data then fine-tuned on real. On DEO (drinking/eating/other) reaches 90.6% balanced accuracy, beating a strong supervised baseline, with gains on the minority class.

Reported benchmarks