Graph-based spatiotemporal activity classification
Human Activity Recognition (HAR) is a foundational task in wearable computing, healthcare monitoring, smart fitness, and motion analytics. Traditional deep-learning solutions (CNNs, LSTMs, Transformers) struggle when sensor signals contain missing labels, irregular sampling, or high intra-class variability. This project introduces HAR-GCNN, a Graph Convolutional Neural Network architecture designed to model temporal sensor interactions as a graph instead of a flat sequence. This enables the system to learn complex, structured relationships between signals, leading to exceptional robustness even when 66% of labels are missing. The model was benchmarked on the PAMAP dataset and outperformed Meta-MAE (a transformer-based masked autoencoder), CNN, and LSTM baselines—achieving up to 99.99% accuracy.
Python, PyTorch, PyTorch Geometric / DGL, Graph Convolutional Networks, PAMAP2 Dataset, NumPy, Pandas, SciPy, Matplotlib, Seaborn, ONNX, TorchScript