dnn-modeling

DNN Deep Learning Modeling (portable)

Implements MLP (Multi-Layer Perceptron) based on PyTorch for binary classification modeling.

Core Pipeline: Feature standardization → MLP training (BatchNorm + Dropout + Early Stopping) → Probability prediction → Three-stage evaluation

Use Cases：

• High-dimensional feature interaction modeling
• Complex nonlinear relationships exist between features
• Sufficient data volume (>10k samples)
• Pursuit of ultimate model performance (can be ensembled with XGBoost)

---

Parameter Description

Parameter	Required	Default Value	Description
--data_path / -d	✅	-	Data file path (parquet/csv)
--target / -t	✅	-	Target variable column name (0/1 binary classification)
--time_col		busi_dt	Time column name
--train_filter		Automatic split	Training set filter condition (pandas query)
--oot_filter		Split by time	OOT cross-time test set condition
--oot_ratio		0.20	If not provided --oot_filter ratio for splitting OOT by time
--val_ratio		0.25	Ratio for splitting val from train_full
--random_seed		42	Random seed
--exclude_cols		-	Excluded columns, comma-separated
--features		-	Specified feature list, comma-separated; if not provided, inferred automatically
--hidden_dims		128,64,32	Hidden layer dimensions, comma-separated
--dropout		0.3	Dropout ratio
--learning_rate		0.001	Learning rate
--batch_size		512	Batch size
--epochs		100	Maximum training epochs
--patience		10	Early stopping patience epochs
--weight_decay		1e-4	Weight decay (L2 regularization)
--pos_weight		auto	Positive sample weight (auto=automatically calculated)
--model_name		Automatically generated	Model name
--report_output		Automatically generated	Report output path
--output_dir		./outputs/<ts>	Artifact output directory
--config		-	JSON configuration file path

---

Execution Method

Run with default parameters:

python scripts/modeling.py \
  --data_path ./data.parquet --target y_label \
  --time_col busi_dt \
  --exclude_cols "cust_code,busi_dt" \
  --output_dir ./outputs/dnn_run

Custom network architecture:

python scripts/modeling.py \
  --data_path ./data.parquet --target y_label \
  --hidden_dims "256,128,64" \
  --dropout 0.4 \
  --learning_rate 0.0005 \
  --epochs 200 \
  --output_dir ./outputs/dnn_run

---

Common Scenarios

Scenario 1: Quick modeling with default parameters

python scripts/modeling.py --data_path ./data.parquet --target y_label --output_dir ./outputs/dnn_run

Scenario 2: Custom network architecture

python scripts/modeling.py --data_path ./data.parquet --target y_label \
  --hidden_dims "256,128,64" --dropout 0.4 --learning_rate 0.0005 \
  --output_dir ./outputs/dnn_run

Scenario 3: Modeling with specified features

python scripts/modeling.py --data_path ./data.parquet --target y_label \
  --features "feat1,feat2,feat3" --output_dir ./outputs/dnn_run

---

Output Artifacts

1. Modeling report(Markdown) — includes data splitting, network architecture, training curves, three-stage evaluation metrics, stability analysis
2. Model file(.pt) — PyTorch state_dict + model configuration
3. Training log(JSON) — loss/AUC/KS records for each epoch
4. result.json — structured artifact manifest

---

Comparison with Other Modeling Skills

Dimension	dnn-modeling	xgb-modeling	lr-modeling
Algorithm	MLP (PyTorch)	XGBoost	LR + WoE
Feature Processing	StandardScaler	Raw values	WoE binning
Nonlinear capability	Strong (multi-layer activation)	Strong (tree structure)	Weak
Interpretability	Weak	Medium (SHAP)	Strong (coefficients)
Training speed	Slow	Fast	Very fast
Applicable data volume	>10k	Any	Any
Evaluation system	AUC/KS/BCR/PSI	AUC/KS/BCR/PSI	AUC/KS/BCR/PSI

---

Upstream/downstream relationships

• Predecessor：data-profiling → feature-analysis(feature selection)
• Subsequent：dnn-tuning(hyperparameter tuning optimization),model-comparison(multi-algorithm comparison)
• Parallel: with xgb-modeling / lr-modeling can be compared horizontally

---

Notes

1. Data requirements: The target variable must be 0/1 binary classification
2. Data volume: It is recommended that training samples > 10k; DNN is prone to overfitting on small samples
3. Missing values: The script automatically uses median imputation + adds missing indicator columns
4. Standardization: Automatically applies StandardScaler to numerical features
5. Early stopping: When val loss remains unchanged for patience consecutive rounds, it stops automatically
6. No hyperparameter tuning provided: If you need hyperparameter tuning, please switch to dnn-tuning(search network architecture + training parameters)
7. GPU: Automatically detects CUDA, and falls back to CPU when there is no GPU
8. Model saving: Model files are saved to <output_dir>/models/