dnn-tuning

What problem does it solve?

This Skill helps you systematically tune a deep neural network (DNN) for binary classification by diagnosing overfitting/underfitting and then running Optuna-based search over network architecture and training hyperparameters to improve validation performance and generalization on OOT.

Core Features & Use Cases

• DNN Hyperparameter Optimization: Uses Optuna TPE Bayesian optimization to search both model structure (n_layers, layer_width, dropout) and training parameters (learning_rate, weight_decay, batch_size).
• Diagnosis-Driven Constrained Search Space: Builds a dynamic search space guided by train/val gaps to reduce overfitting or address underfitting.
• Two Execution Modes (Safe by Default): Interactive mode pauses each round for user feedback, while AUTO mode iterates automatically until convergence criteria are met.
• Workflow Fit for Finance Risk Modeling: Works on the project’s classic three-split setup (train/val/OOT), where val drives early decisions and OOT is used for final reporting. Use cases include credit scoring / churn risk models where AUC/KS quality matters.

Quick Start

Ask the agent to run DNN tuning with your data path and target column in interactive mode, assuming you have already trained a baseline model with dnn-modeling.