Table 8 Disaggregated wall-clock timings (mean±SD over 5 runs) and peak memory. All numbers are per 5-fold CV on 5046 notes (\(\approx\) 4037 train + 1009 validation). Inference latency measured as mean time to predict one held-out note.
Pipeline stage | PCA+SMOTE | UMAP+SMOTE | t-SNE+SMOTE | No reduction |
|---|---|---|---|---|
(i) DR fit (s) | 3 ± 1 | 71 ± 5 | 95 ± 7 | – |
(ii) Sampler fit (s) | 2 ± 1 | 2 ± 1 | 2 ± 1 | 1 ± 0 |
(iii) Model fit (s) | 37 ± 2 | 37 ± 2 | 37 ± 2 | 72 ± 4 |
Total training (s) | 42 ± 3 | 110 ± 7 | 134 ± 9 | 73 ± 5 |
(iv) Inference (ms/note) | 0.9 ± 0.1 | 0.9 ± 0.1 | 0.9 ± 0.1 | 1.8 ± 0.2 |
(v) Peak memory (GB) | 4.1 ± 0.2 | 4.5 ± 0.3 | 4.6 ± 0.3 | 7.2 ± 0.4 |
