================================================================================================ FACULTY: khoa_cong_nghe_thong_tin | strategy=expanding | event y<5.0 DEVICE: cuda (Tesla V100-SXM2-32GB) | model=MultiTaskNet(reg+cls+blend) | stable_temporal_cls: H=4, strength=0.5, bounds=[0.5,1.5] DEV folds=[1, 2, 3, 4, 5] | HOLDOUT folds=[6, 7] | tune=F1 (chon signal theo AP dev) ================================================================================================ >>> DEVELOPMENT (tune tren fold dev) [Class weights] n_low=613, n_high=19112 | raw_inverse_ratio=31.178 | CB_low=10.1327, CB_high=0.7071 | CB_ratio=14.330 [Temporal policy] fold 1 H=4 meaningful_threshold=100 strength=0.5 bounds=[0.5,1.5] | raw_range=(0.2973,1.0000) final_range=(0.7060,1.2867) temporal_ESS=19316/19725 semester ordinal used age rows fail% raw final mass% 20181 50 yes 7 635 5.83% 0.2973 0.7076 2.28% 20182 51 yes 6 624 6.57% 0.3536 0.7471 2.36% 20183 52 no 6 37 29.73% 0.3536 0.7548 0.14% 20191 53 yes 5 1609 4.91% 0.4204 0.7932 6.47% 20192 54 yes 4 1617 5.75% 0.5000 0.8491 6.96% 20193 55 no 4 3 0.00% 0.5000 0.8464 0.01% 20201 56 yes 3 3115 3.08% 0.5946 0.9137 14.43% 20202 57 yes 2 2815 2.74% 0.7071 0.9917 14.15% 20203 58 no 2 1 0.00% 0.7071 0.9899 0.01% 20211 59 yes 1 5277 1.19% 0.8409 1.0836 28.99% 20212 60 yes 0 3992 2.91% 1.0000 1.1956 24.20% [Combined weights] reg/blend=BASELINE mean=1.0000 range=(0.8894,4.4472) | cls mean=1.0000 range=(0.4992,13.0377) ESS=5205/19725 | pos/neg mass baseline=0.4596 new=0.4596 [Class weights] n_low=802, n_high=24145 | raw_inverse_ratio=30.106 | CB_low=8.7669, CB_high=0.7420 | CB_ratio=11.815 [Temporal policy] fold 2 H=4 meaningful_threshold=125 strength=0.5 bounds=[0.5,1.5] | raw_range=(0.2500,1.0000) final_range=(0.6815,1.2762) temporal_ESS=24299/24947 semester ordinal used age rows fail% raw final mass% 20181 50 yes 8 635 5.83% 0.2500 0.6822 1.74% 20182 51 yes 7 624 6.57% 0.2973 0.7168 1.79% 20183 52 no 7 37 29.73% 0.2973 0.7203 0.11% 20191 53 yes 6 1609 4.91% 0.3536 0.7575 4.89% 20192 54 yes 5 1617 5.75% 0.4204 0.8064 5.23% 20193 55 no 5 3 0.00% 0.4204 0.8052 0.01% 20201 56 yes 4 3115 3.08% 0.5000 0.8637 10.78% 20202 57 yes 3 2815 2.74% 0.5946 0.9324 10.52% 20203 58 no 3 1 0.00% 0.5946 0.9316 0.00% 20211 59 yes 2 5277 1.19% 0.7071 1.0137 21.44% 20212 60 yes 1 3992 2.91% 0.8409 1.1116 17.79% 20221 62 yes 0 5222 3.62% 1.0000 1.2277 25.70% [Combined weights] reg/blend=BASELINE mean=1.0000 range=(0.8861,4.4303) | cls mean=1.0000 range=(0.5057,11.1883) ESS=8000/24947 | pos/neg mass baseline=0.3924 new=0.3924 [Class weights] n_low=954, n_high=28591 | raw_inverse_ratio=29.970 | CB_low=7.9554, CB_high=0.7679 | CB_ratio=10.360 [Temporal policy] fold 3 H=4 meaningful_threshold=148 strength=0.5 bounds=[0.5,1.5] | raw_range=(0.2102,1.0000) final_range=(0.6631,1.3134) temporal_ESS=28621/29545 semester ordinal used age rows fail% raw final mass% 20181 50 yes 9 635 5.83% 0.2102 0.6635 1.43% 20182 51 yes 8 624 6.57% 0.2500 0.6945 1.47% 20183 52 no 8 37 29.73% 0.2500 0.6967 0.09% 20191 53 yes 7 1609 4.91% 0.2973 0.7311 3.98% 20192 54 yes 6 1617 5.75% 0.3536 0.7750 4.24% 20193 55 no 6 3 0.00% 0.3536 0.7742 0.01% 20201 56 yes 5 3115 3.08% 0.4204 0.8266 8.72% 20202 57 yes 4 2815 2.74% 0.5000 0.8883 8.46% 20203 58 no 4 1 0.00% 0.5000 0.8878 0.00% 20211 59 yes 3 5277 1.19% 0.5946 0.9615 17.17% 20212 60 yes 2 3992 2.91% 0.7071 1.0492 14.18% 20221 62 yes 1 5222 3.62% 0.8409 1.1534 20.39% 20222 63 yes 0 4598 3.31% 1.0000 1.2769 19.87% [Combined weights] reg/blend=BASELINE mean=1.0000 range=(0.8856,4.4281) | cls mean=1.0000 range=(0.5092,10.4487) ESS=10843/29545 | pos/neg mass baseline=0.3457 new=0.3457 [Class weights] n_low=1172, n_high=34650 | raw_inverse_ratio=29.565 | CB_low=6.9857, CB_high=0.7975 | CB_ratio=8.759 [Temporal policy] fold 4 H=4 meaningful_threshold=180 strength=0.5 bounds=[0.5,1.5] | raw_range=(0.1768,1.0000) final_range=(0.6421,1.3241) temporal_ESS=34449/35822 semester ordinal used age rows fail% raw final mass% 20181 50 yes 10 635 5.83% 0.1768 0.6423 1.14% 20182 51 yes 9 624 6.57% 0.2102 0.6692 1.17% 20183 52 no 9 37 29.73% 0.2102 0.6702 0.07% 20191 53 yes 8 1609 4.91% 0.2500 0.7012 3.15% 20192 54 yes 7 1617 5.75% 0.2973 0.7393 3.34% 20193 55 no 7 3 0.00% 0.2973 0.7389 0.01% 20201 56 yes 6 3115 3.08% 0.3536 0.7844 6.82% 20202 57 yes 5 2815 2.74% 0.4204 0.8381 6.59% 20203 58 no 5 1 0.00% 0.4204 0.8379 0.00% 20211 59 yes 4 5277 1.19% 0.5000 0.9020 13.29% 20212 60 yes 3 3992 2.91% 0.5946 0.9782 10.90% 20221 62 yes 2 5222 3.62% 0.7071 1.0688 15.58% 20222 63 yes 1 4598 3.31% 0.8409 1.1764 15.10% 20231 65 yes 0 6277 3.47% 1.0000 1.3044 22.86% [Combined weights] reg/blend=BASELINE mean=1.0000 range=(0.8843,4.4214) | cls mean=1.0000 range=(0.5121,9.2494) ESS=15460/35822 | pos/neg mass baseline=0.2963 new=0.2963 [Class weights] n_low=1416, n_high=39480 | raw_inverse_ratio=27.881 | CB_low=6.0754, CB_high=0.8180 | CB_ratio=7.428 [Temporal policy] fold 5 H=4 meaningful_threshold=205 strength=0.5 bounds=[0.5,1.5] | raw_range=(0.1487,1.0000) final_range=(0.6250,1.3598) temporal_ESS=39082/40896 semester ordinal used age rows fail% raw final mass% 20181 50 yes 11 635 5.83% 0.1487 0.6277 0.97% 20182 51 yes 10 624 6.57% 0.1768 0.6518 0.99% 20183 52 no 10 37 29.73% 0.1768 0.6510 0.06% 20191 53 yes 9 1609 4.91% 0.2102 0.6806 2.68% 20192 54 yes 8 1617 5.75% 0.2500 0.7147 2.83% 20193 55 no 8 3 0.00% 0.2500 0.7150 0.01% 20201 56 yes 7 3115 3.08% 0.2973 0.7555 5.75% 20202 57 yes 6 2815 2.74% 0.3536 0.8038 5.53% 20203 58 no 6 1 0.00% 0.3536 0.8040 0.00% 20211 59 yes 5 5277 1.19% 0.4204 0.8614 11.12% 20212 60 yes 4 3992 2.91% 0.5000 0.9296 9.07% 20221 62 yes 3 5222 3.62% 0.5946 1.0109 12.91% 20222 63 yes 2 4598 3.31% 0.7071 1.1076 12.45% 20231 65 yes 1 6277 3.47% 0.8409 1.2225 18.76% 20232 66 yes 0 5072 4.81% 1.0000 1.3589 16.85% 20233 67 no 0 2 0.00% 1.0000 1.3598 0.01% [Combined weights] reg/blend=BASELINE mean=1.0000 range=(0.8784,4.3918) | cls mean=1.0000 range=(0.5135,8.1464) ESS=20205/40896 | pos/neg mass baseline=0.2664 new=0.2664 signal blended : AP=0.192 | F1-opt tau=0.317 flag_rate=0.054 (ratio=1.30) -> P=0.257 R=0.333 F1=0.290 signal P(y P=0.274 R=0.322 F1=0.296 <= CHON (AP cao nhat) ### DEVELOPMENT (validation) — BLENDED [signal=blended] fold test base% | AP AUC | ACC@t P@t R@t F1@t | rate% ACC@r P@r R@r F1@r 1 5222 3.6% | 0.113 0.656 | 0.963 0.364 0.042 0.076 | 5.4% 0.927 0.159 0.238 0.191 2 4598 3.3% | 0.376 0.854 | 0.960 0.398 0.434 0.415 | 4.7% 0.951 0.329 0.467 0.386 3 6277 3.5% | 0.158 0.801 | 0.914 0.173 0.390 0.239 | 4.3% 0.939 0.197 0.243 0.218 4 5072 4.8% | 0.271 0.830 | 0.911 0.273 0.504 0.354 | 4.5% 0.939 0.355 0.332 0.343 5 6398 5.5% | 0.205 0.756 | 0.920 0.280 0.291 0.285 | 6.2% 0.916 0.263 0.299 0.280 ------------------------------------------------------------------------------------------------------------ POOLED 27567 4.2% | 0.192 0.760 | 0.932 0.257 0.333 0.290 | 5.1% 0.933 0.254 0.308 0.279 ### DEVELOPMENT (validation) — CLASSIFICATION HEAD P(y=5 Failure <5 ------------------------------------------------------------------ Normal >=5 4,863 225 Failure <5 365 198 ------------------------------------------------------------------ Normal correctly rejected: 95.58% Failures successfully detected: 35.17% Confusion matrix — fold 6 — drift-aware @rate=6.461% ------------------------------------------------------------------ Actual / Predicted Normal >=5 Failure <5 ------------------------------------------------------------------ Normal >=5 4,904 184 Failure <5 382 181 ------------------------------------------------------------------ Normal correctly rejected: 96.38% Failures successfully detected: 32.15% 7 6926 15.1% | 0.379 0.776 | 0.838 0.448 0.313 0.369 | 11.7% 0.833 0.433 0.336 0.378 Confusion matrix — fold 7 — fixed threshold @t=0.28213 ------------------------------------------------------------------ Actual / Predicted Normal >=5 Failure <5 ------------------------------------------------------------------ Normal >=5 5,474 404 Failure <5 720 328 ------------------------------------------------------------------ Normal correctly rejected: 93.13% Failures successfully detected: 31.30% Confusion matrix — fold 7 — drift-aware @rate=11.733% ------------------------------------------------------------------ Actual / Predicted Normal >=5 Failure <5 ------------------------------------------------------------------ Normal >=5 5,417 461 Failure <5 696 352 ------------------------------------------------------------------ Normal correctly rejected: 92.16% Failures successfully detected: 33.59% ------------------------------------------------------------------------------------------------------------ POOLED 12577 12.8% | 0.382 0.801 | 0.864 0.455 0.327 0.380 | 9.4% 0.863 0.452 0.331 0.382 Confusion matrix — POOLED — fixed development threshold ------------------------------------------------------------------ Actual / Predicted Normal >=5 Failure <5 ------------------------------------------------------------------ Normal >=5 10,337 629 Failure <5 1,085 526 ------------------------------------------------------------------ Normal correctly rejected: 94.26% Failures successfully detected: 32.65% Confusion matrix — POOLED — drift-aware warning policy ------------------------------------------------------------------ Actual / Predicted Normal >=5 Failure <5 ------------------------------------------------------------------ Normal >=5 10,321 645 Failure <5 1,078 533 ------------------------------------------------------------------ Normal correctly rejected: 94.12% Failures successfully detected: 33.09% ### HOLDOUT (test) — CLASSIFICATION HEAD — EQUAL WARNING BUDGETS [signal=P(y