Эксперимент: Missingness-as-a-Label Signal (#1168)

{ "worked": true, "confidence": 8, "verdict": "Built the entropy-logistic missingness likelihood with trainable a and positive b, including a detached-entropy ablation and CPU/CUDA fallback. The math check matched dJ/da=mean(q-m) and dJ/db=mean((q-m)H) to numerical precision, with nonzero entropy-to-logit gradients. Across three seeds, joint training modestly improved informative-missingness test NLL but not accuracy; random missingness produced no meaningful gain.", "metrics": { "baseline": "Informative missingness: accuracy 0.8885 +/- 0.0058, NLL 0.2670 +/- 0.0043. Random missingness: accuracy 0.8859 +/- 0.0059, NLL 0.2738 +/- 0.0147.", "idea": "Informative joint: accuracy 0.8882 +/- 0.0055, NLL 0.2631 +/- 0.0063; learned b was positive in all runs and mean predicted/true missingness correlation was about 0.93. Random joint: accuracy 0.8855 +/- 0.0063, NLL 0.2722 +/- 0.0137. Detached entropy matched the supervised classifier metrics.", "math_check": "a gradient -0.0441457297454092 versus expected -0.0441457297454092; b gradient 0.034629446878288096 versus expected 0.03462944687828812; gradient_match=true." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 missingness_signal.py && /home/maxwelhelp/main/bin/python3 repeat_check.py", "files": [ "missingness_signal.py", "repeat_check.py", "results.json", "repeat_results.json" ], "limitations": "Only a small two-dimensional synthetic Gaussian binary task was tested; no MNIST/CIFAR, multiclass setting, label-fraction sweep, FLOP/speed study, or formal calibration metric was evaluated. The repeat experiment used three seeds and 500 full-batch optimization steps." }