Descriptors based on elements are too much if the dataset size is limited. It's essential to do feature selection.
The ways for feature selection:

• LASSO algorithm
• Pearson correlation of features and property
• Random Forest (RF)
• Recursive Feature Elimination (RFE)

Optimization

• Format string（using f"" to instead "%s"）-> shorter and more readable
• File writer (using with to instead pd.DataFrame.to_excel) -> safer
• for loop (using enumrate(), item(), iter(), etc. to instead range()) -> more readable
• Length of each line (no more than 80 chars) -> more readable and suit for small screens
• Add docstrings for each class and method -> more readable and easier to generate docs in the future
• File exists (using try...except to instead if os.path.exist()) -> safer and easier to traceback errors
• Adjust vscode code highlight -> more readable
• Function name (using low_underline to instead CaptialFun) -> more readable

Features

• Support multiple nodes calculation (using mpi4py to instead joblib) -> more scalable
• Add some examples -> more guide for users in the future

Hyperparameter optimization can obtain the best performance of ML models. Especially for some limited datasets, it's vital.
For several hyperparameters situations, Grid Search would be the easiest way though time-consuming.
Considering only two hyperparameters (degree & C) here, we need to have some functions to support these features:

• Efficient calculation for a mass of hyperparameter combinations (multiple nodes, multiple threads)
• Collect necessary metrics from calculation results (for figures).
• Plot figures from three views: Total RMSE, Trade-off RMSE, and Test RMSE.
• Plot heatmaps for two hyperparameters with any metrics.
• Big range for C (>5000)
• Normalization for property