new_benchmark

[Alice0416]

Description

This pull request adds a random forest algorithm utilizing features from the Sine Coulomb Matrix and MagPie featurization algorithms. Here are the key details of the algorithm:

• Sine Coulomb Matrix: Creates structural features based on Coulombic interactions within a periodic boundary condition (suitable for crystalline materials with known structures).

• MagPie Features: Weighted elemental features derived from elemental data such as electronegativity, melting point, and electron affinity.

Both algorithms were executed within the Automatminer v1.0.3.20191111 framework for convenience, although no auto-featurization or AutoML processes were applied.

Data Processing

• Data Cleaning: Features with more than 1% NaN samples were dropped. Missing samples were imputed using the mean of the training data.

• Featurization:

1. For structure problems: Both Sine Coulomb Matrix and MagPie features were applied.

2. For problems without structure: Only MagPie features were applied.

Model Details

• Random Forest: Utilizes 500 estimators.

• Hyperparameter Tuning: None performed. A large, constant number of trees were used in constructing each fold's model, using the entire training+validation set as training data for the random forest.

Additional Information

Raw Data and Example Notebook: Available on the matbench repository.

Included files

-- benchmarks
---- matbench_v0.1_RFSCM/Magpie
------ results.json.gz             # required filename
------ my_python_file.py            # required filename
------ info.json                   # required filename