<![CDATA[Large-scale urban sensing platforms generate millions of environmental observations, making full-data DNN training computationally costly. This paper presents a controlled empirical study of balanced cluster sampling (BCS), which partitions preprocessed feature space and draws equal quotas under a fixed instance budget. BCS is compared with uniform random sampling, label-stratified sampling, and two established coreset selectors using a two-million-row synthetic archive modeled on the Array of Things Chicago deployment. The pipeline applies log1p transformation, train-only z-score normalization, cyclical and one-hot calendar encoding, and a fixed ten-layer DNN. On a stratified test set of 300,000 rows, the full-data baseline achieves 98.01 ± 0.27% accuracy. BCS with K = 8 achieves 96.72 ± 0.55% accuracy and 96.29 ± 0.60% macro-F1 using 14.3% of training data, compared with 96.52 ± 0.64% for uniform random sampling; the 0.19-point difference is not statistically significant (p = 0.532). Label-stratified sampling performs at least as well (97.05 ± 0.41%), while Moderate Coreset and EL2N achieve 89.03 ± 1.35% and 55.54 ± 11.91%, respectively. Although pruning reduces DNN training time by 71.0%, end-to-end time decreases by only 43.3%. These results indicate that, for this workload, uniform or label-stratified sampling is sufficient, while additional selection machinery provides no practical benefit. ]]>
Copyrights © 2026
