discover

Getting Started: A Step-by-Step Tutorial

This tutorial will guide you through your first analysis using DISCOVER. We will go through installation, data preparation, configuration, execution, and interpretation of the output. We will take config.json as an example.

Step 1: Installation

Ensure you have Python 3.8+ installed. It is strongly advised to use a virtual environment.

1. Clone the Repository:

   git clone <your-repository-url>
   cd discover-project
   

2. Activate a Virtual Environment:

   python -m venv venv
   source venv/bin/activate  # On Windows, use `venv\\\\Scripts\\\\activate`
   

3. Install Core Dependencies:

   pip install pandas numpy sympy scikit-learn matplotlib seaborn joblib
   

4. Install Optional Dependencies (Optional):
   • For GPU Acceleration (NVIDIA):

     # Put XX in place of your CUDA version (e.g., 118 for CUDA 11.8)
     pip install cupy-cudaXX
     

     *   **For GPU Acceleration (Apple Silicon):** ```bash pip install torch torchvision torchaudio ```
     

   • For Unit-Awareness:

     pip install pint
     

Step 2: Prepare Your Data

DISCOVER expects your data in a single CSV file.

• Each row should be an example (e.g., a material or a data point).
• Each column should be a primary feature or the target property.

For our case, sample_dataset.csv contains columns for main features like R_A (ionic radius), ELN_A (electronegativity), and target property E_a (migration barrier).

Step 3: Prepare the Analysis

The workflow is managed by config.json. Let’s break down the critical parts for an initial run.

{
    // 1. DATA & I/O
    "data_file": "manuel_data.csv",
    "property_key": "E_a",
    "non_feature_cols": ["Name", "Delta_E", "."],
},

"workdir": "manuel_data",

    // 2. FEATURE CONSTRUCTION
    "depth": 3,
    "op_rules": [
      {"op": "add"}, {"op": "sub"}, {"op": "mul"}, {"op": "div"},
      {"op": "sqrt"}, {"op": "sq"}
    ],

    // 3. SEARCH & SELECTION
    "max_D": 2,
    "search_strategy": "sa",
"selection_method": "bic",

    // 4. MODEL & VALIDATION
    "task_type": "regression",
    "fix_intercept": false,

    // 5. COMPUTATIONAL
    "n_jobs": -1,
    "device": "cpu",
    "random_state": 42
}

• data_file: Path to your input CSV.
• property_key: The exact name of the target column in your CSV.
• workdir: Current working directory where output files are written.
• depth: Maximum complexity of the generated features. 2 or 3 is a good default.
• op_rules: Mathematical operations with which features are merged.
• max_D: Maximum number of features (dimension) in the final model.
• search_strategy: The strategy used to find the best feature combination. greedy is fast, but sisso++ or sa (Simulated Annealing) could be more thorough.
• selection_method: The criterion to select the best dimension. cv (Cross-Validation) is stable, but bic (Bayesian Information Criterion) penalizes model complexity.
• task_type: The machine learning task type. regression is used for continuous values.

Step 4: Run the Analysis

With your configuration set, execute the run_discover.py script in your command line:

python run_discover.py config.json

DISCOVER will start, giving live output on how it’s proceeding:

1. Loading data and config.
2. Creating features iteratively, reporting at each depth.
3. Printing the top-screened features at each iteration.
4. Executing the selected search strategy for each dimension D.
5. Checking the models and selecting the best dimension.
6. Printing the final summary report to the console.

Step 5: Understanding the Output

All output is saved to the directory specified by workdir (e.g., manuel_data/).

Key Console Output: The final part of the console output is the summary report, the most important takeaway:

========================= FINAL MODEL REPORT =========================
DISCOVER Summary
=================
TASK_TYPE: regression
MODEL_SELECTION_METHOD: BIC
SELECTED_DIMENSION: 2
SEARCH_STRATEGY: SA
.

******* BEST MODEL (D=2) ********
Score BIC Score: -123.456
R2/Accuracy on full data: 0.9512

============================================================
Final Symbolic Model: regression
============================================================
P = -0.1234 + 0.5678*sqrt(R_A/R_B) + 0.9012*Abs(ELN_A - ELN_B)

Important Files in the workdir:

• discover.out: A text file with the same exhaustive summary report to the console.
• final_models_summary.json: A machine-readable JSON file providing exhaustive detail regarding the top discovered model for each dimension, including coefficients and feature expressions.
• top_sis_candidates.csv: A CSV file ranking the best individual features (1D models) discovered, and worth examining which top features are most effective.
• plots/: A folder of automatically generated plots, such as the model selection score plot (selection_scores.png).
• models/: Contains detailed text files for each model dimension (model_D01.dat, model_D02.dat, etc.).

What’s Next?

Congratulations! You’ve successfully executed your first DISCOVER analysis. You now possess an interpretable, symbolic model that describes your data.

For exploring more sophisticated features, check out our How-To Guides, where you can observe how to:

• Execute DISCOVER programmatically inside a Python program.
• Employ GPU acceleration for massive speed improvements.
• Employ sophisticated search strategies and operators.
• Classify tasks.

This site is open source. Improve this page.