Mapping the Limits of Sensor-Based Human Activity Recognition

Human Activity Recognition (HAR) performance can be limited by the information available in sensor data, how this information is represented, or the ability of a learning model to estimate the required decision function. This work investigates these limits through a multidimensional sufficiency frontier, where sensor data are progressively transformed across spatial, temporal, spectral, contextual, and signal-quality dimensions. We evaluate the datasets DLA and MHealth under Leave-One-Subject-Out validation using Logistic Regression, SVM-RBF, and MiniRocket as complementary diagnostic probes rather than competing classifiers. Preliminary results reveal substantial redundancy across individual data dimensions: high performance is preserved after reducing sensor channels, sampling rate, spectral bandwidth, temporal context, and signal precision. However, combining these degradations produces clearer performance boundaries, with aggressive configurations reaching balanced accuracies of 0.833 and 0.839 on DLA and MHealth, respectively, before further degradation leads toward an insufficient regime. These findings motivate a systematic framework for mapping sufficient, transitional, and insufficient regions and investigating whether the observed limits arise from information, representation, or estimation. This ongoing work will extend the analysis to additional sensor-based HAR datasets and model-independent measures of data separability.