Hybrid CNN-LSTM with Adaptive Data Augmentation for Subject-Independent Human Activity Recognition
Human Activity Recognition (HAR) from smartphone accelerometer data is widely studied on the WISDM dataset, but random sample-based partitions can allow windows from the same subject to appear in both training and evaluation data, producing optimistic estimates of cross-subject generalization. This paper investigates the DeepConvLSTM architecture on WISDM v1.1 using only tri-axial smartphone accelerometer signals under a subject-disjoint split comprising 25 training subjects and an 11-subject evaluation partition. Nine controlled experiments varied window configuration, model capacity, temporal pooling, and augmentation strategy. Under this single fixed subject split, without repeated random seeds or statistical comparisons, the best-performing configuration among the nine experiments used a 60-sample window (3 s) with 50% overlap and on-the-fly jitter and scaling augmentation, achieving 90.12% accuracy, compared with 89.12% for offline augmentation, 88.75% for class-specific augmentation, and 87.49% with label smoothing. Global Average Pooling showed an observed class-level trade-off relative to last-timestep pooling in the evaluated comparison rather than a general architectural advantage. A persistent 8–10 percentage-point training–evaluation gap remained, with notable confusion among stair-related locomotion classes, which may partly reflect limited subject diversity, class imbalance, and accelerometer-only sensing. Importantly, the same 11-subject evaluation partition was consulted for early stopping, learning-rate scheduling, comparison of all nine experiments, and final model selection; therefore, the reported 90.12% accuracy should not be interpreted as performance on a fully untouched test set. These findings provide a leakage-resistant but selection-sensitive benchmark for subject-independent DeepConvLSTM evaluation on WISDM.