README_EN

Code Function Description

This project is a machine learning training and prediction tool based on KAN (Kolmogorov-Arnold Network) for identifying optimal metabolites from metabolite expression data. It supports both Chinese and English language modes as well as dual-platform (Windows and Linux) execution. By integrating base learners such as XGBoost, Random Forest, SVM, and Gradient Boosting, combined with the KAN network for model fusion, it enables efficient classification tasks.

Main features include:

• Model Training: Hyperparameter tuning with Optuna, training of base learners, and fusion via KAN.
• Data Prediction: Loading pre-trained models for classification on new datasets, outputting probabilities and prediction results.
• Multi-language Support: Chinese (KANMB_CN.py) and English (KANMB_EN.py) versions provide language-specific logging output.

---

Environment Setup

A conda environment is required:

# Windows
conda env create -f KANMB_Win_environment.yaml
# Linux
conda env create -f KANMB_Linux_environment.yaml

Or use manual installation:

conda create -n KANMB python=3.10
pip install tqdm matplotlib PyYAML numpy pandas joblib scikit-learn xgboost torch torchvision pykan optuna

---

Usage Steps

1. Model Training

python KANMB_CN.py --mode train \
  --num_folds 5 \
  --n_trials 30 \
  --train_file "./Data/TrainandVaild.csv" \
  --model_output_dir "./test"

2. Data Prediction

python KANMB_CN.py --mode pre \
  --pred_file "./Data/TestData.csv" \
  --output_file "./test/KAN.csv" \
  --model_dir "./test"

---

Parameter Description

Category	Parameter	Type	Required	Default	Description
General	--mode	str	Yes	train	Running mode: train (model training) / pre (prediction)
	--num_folds	int	No	5	Number of cross-validation folds
	--n_trials	int	No	30	Number of hyperparameter tuning trials
Training Parameters	--train_file	str	Yes	-	Path to training dataset (CSV)
	--model_output_dir	str	Yes	-	Directory to store trained models
Prediction Parameters	--pred_file	str	Yes	-	Path to input CSV file for prediction
	--output_file	str	Yes	-	Path to save prediction results
	--model_dir	str	Yes	-	Path to directory containing pre-trained models

---

Data Format Requirements

Training File (train_file)

• Format: CSV file with header row
• Required column: TARGET (classification label, 0/1)
• Other columns: Numeric features (no preprocessing required)

Prediction File (pred_file)

• Format: CSV file with header row
• Feature columns: Must match the training dataset columns (excluding TARGET)
• Index column: Recommended to include a unique identifier column (e.g., ID)

---

Notes

1. Path Rules: Use absolute or relative paths (e.g., ./Data/), avoid Chinese characters in file paths.
2. Directory Setup: It is recommended to create model_output_dir in advance for training mode.
3. Dependencies: The KAN library should be installed via pip install kan (latest version) to ensure compatibility with PyTorch.
4. CUDA Support: For large datasets, it is recommended to configure CUDA according to the official PyTorch documentation. For most cases, CPU computation is sufficient.
5. Required Parameters:
   • Training mode requires --train_file and --model_output_dir
   • Prediction mode requires --pred_file, --output_file, and --model_dir
6. Log Files: Training process outputs kan_training.log, prediction process outputs predict.log, both useful for troubleshooting.