TrustNet: Deep Abstaining Classifier
Selective Prediction & Fraud Detection with Deep Abstaining Classifiers
Role
Lead ML Researcher & Author: PyTorch model, DAC loss function, multi-seed validation, Flask REST API
Primary Stack
Python · PyTorch · NumPy · Scikit-Learn · Flask · Matplotlib
“Traditional ML classifiers are forced to output a prediction even on borderline, high-uncertainty transactions, causing high false negatives and undetected financial fraud.”
Must introduce an abstention option ('I don't know') to defer uncertain cases to human review
Rigorous loss formulation: L_DAC = -log(p_true + p_abstain) + alpha * p_abstain
All findings must be validated across 3 independent random seeds with paired t-tests
Why add an explicit Abstention Neuron over post-hoc thresholding?▼
Softmax confidence thresholding is notoriously overconfident on out-of-distribution inputs. TrustNet incorporates abstention directly into the loss objective (L_DAC) during gradient descent with an alpha penalty (alpha = 0.3), forcing the network to learn genuine uncertainty representations in safety-critical domains.
Why paired t-tests across 3 random seeds?▼
Machine learning research papers frequently overclaim from a single lucky seed. TrustNet evaluates standard MLP against DAC across seeds 42, 123, and 999 with paired t-tests to prove statistical significance in selective risk reduction.
Distributed System Architecture
Explore the multi-tier topology below. Switch between the interactive blueprint canvas, standard Mermaid.js flowcharts, and the step-by-step request simulator.
Deep Abstaining Classifier (DAC) pipeline with 30-feature vector normalization, 3-class MLP architecture (Legitimate, Fraud, Abstain), and calibrated uncertainty evaluation.
Engineering Post-Mortem & Next Iteration
I would extend the DAC loss function to multi-modal transformer architectures for joint transaction and image verification.