agraspqfs 1.0.0

Adaptive GRASP-based feature selection with self-tuning subset size

A-GRASPQ-FS

Adaptive GRASPQ Feature Selection with self-tuning subset size

A-GRASPQ-FS is a scikit-learn compatible feature selector based on a reactive GRASP metaheuristic. It jointly solves two decisions that are usually treated separately:

1. how many features should be selected for a given dataset and classifier; and
2. which features should compose that subset.

Instead of requiring a fixed value of k, A-GRASPQ-FS searches over a cardinality interval and adapts the probability of choosing each subset size according to the validation performance observed during the search.

The original command-line research prototype used in the ISCC 2026 paper is still available in main.py. Version 1.0 adds a production-oriented Python package in src/agraspqfs/, with a plug-and-play API for scikit-learn workflows.

Associated paper

This repository accompanies the ISCC 2026 paper:

Vagner E. Quincozes, Silvio E. Quincozes, Célio Albuquerque, Diego Passos, and Daniel Mossé. Adaptive Feature Selection with Self-Tuning Subset Size for Intrusion Detection. ISCC 2026, Algarve, Portugal, 2026.

If you use this software in academic work, please cite the paper and/or the software release. Citation files are provided in CITATION.cff and docs/CITATION.md.

Highlights in v1.0

• scikit-learn compatible transformer: fit, transform, fit_transform, get_support, get_feature_names_out.
• Works with any scikit-learn compatible classifier as the subset evaluator.
• Joint optimization of subset size and feature composition.
• Presets for different budgets: fast, balanced, robust, and paper.
• In-memory cache to avoid repeated evaluations of the same subset.
• Explicit computational controls: max_evaluations, time_budget, and early_stopping_rounds.
• Exploratory evaluation on a row sample via evaluation_sample_size.
• Final full-data re-evaluation of elite candidate subsets.
• CLI command: agraspqfs.
• Unit tests, examples, GitHub Actions workflow, citation metadata, and release checklist.

Installation

From the repository root:

pip install -e .

For development and tests:

pip install -e ".[dev]"
pytest

Optional XGBoost support:

pip install -e ".[xgboost]"

Minimal usage

from agraspqfs import AGraspQFeatureSelector
from sklearn.naive_bayes import GaussianNB

selector = AGraspQFeatureSelector(
    estimator=GaussianNB(),
    preset="fast",
    min_features=2,
    max_features=15,
    random_state=42,
)

X_selected = selector.fit_transform(X, y)

print(selector.selected_features_)
print(selector.n_features_selected_)
print(selector.best_score_)

Pipeline usage

from agraspqfs import AGraspQFeatureSelector
from sklearn.naive_bayes import GaussianNB
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline

pipe = Pipeline([
    ("fs", AGraspQFeatureSelector(
        estimator=GaussianNB(),
        preset="balanced",
        min_features=2,
        max_features=20,
        random_state=42,
    )),
    ("clf", LogisticRegression(max_iter=2000)),
])

pipe.fit(X_train, y_train)
y_pred = pipe.predict(X_test)

The estimator inside AGraspQFeatureSelector is used to evaluate feature subsets. The classifier after the selector in the pipeline can be the same model or a different one.

CLI usage

agraspqfs   --csv data/hibrid_dataset_GOOSE_train.csv   --target class   --estimator nb   --preset fast   --min-features 2   --max-features 15   --output results/agraspqfs_selection.json   --history-csv results/agraspqfs_history.csv

The CLI prints and saves a JSON report containing selected features, internal validation score, external holdout score, number of evaluations, cache hits, and elapsed time.

Presets

Preset	Intended use	Behavior
fast	notebooks, smoke tests, quick exploration	fewer evaluations, holdout validation, row sampling
balanced	default practical use	moderate budget, 3-fold CV, partial row sampling
robust	slower experiments	larger budget, 5-fold CV, full exploratory data
paper	compatibility with the research-style workflow	conservative budget inspired by the original prototype

Explicit parameters override preset values.

Example:

selector = AGraspQFeatureSelector(
    preset="fast",
    max_evaluations=100,
    time_budget=60,
    early_stopping_rounds=30,
    random_state=42,
)

Main parameters

Parameter	Description
estimator	Classifier used to evaluate candidate subsets. Defaults to GaussianNB.
scoring	Any scikit-learn scoring name, e.g., f1_weighted, accuracy, roc_auc_ovr.
min_features, max_features	Minimum and maximum subset size.
rcl_size	Number of ranked candidate features considered by the metaheuristic.
ranking_method	mutual_info, variance, random, none, or callable.
evaluation_sample_size	Fraction of rows used in exploratory evaluations.
cv	Internal CV. Use None for holdout validation.
max_evaluations	Maximum number of non-cached subset evaluations.
time_budget	Time budget in seconds.
cache	Reuse repeated subset scores.
refit	If True, fit estimator_ on the selected features after selection.

Main outputs

After fit, the selector exposes:

selector.selected_features_      # names of selected features
selector.selected_indices_       # integer indices
selector.n_features_selected_    # selected subset size
selector.support_                # boolean mask
selector.best_score_             # best internal score among final candidates
selector.baseline_score_         # internal score with all features
selector.history_                # pandas DataFrame with evaluation trace
selector.size_weights_           # final adaptive cardinality weights
selector.n_evaluations_          # expensive evaluations actually computed
selector.n_cache_hits_           # repeated evaluations avoided by cache
selector.stop_reason_            # completed, budget_exhausted or early_stopping

Examples

Run:

python examples/quickstart.py
python examples/pipeline_example.py
python examples/benchmark_small.py

A minimal notebook is available at examples/agraspqfs_demo.ipynb.

Legacy research prototype

The original scripts are still available:

python main.py -d ereninho -a nb -rcl 10 -is 5 -pq 10 -lc 50 -cc 50

Baseline scripts are also preserved:

python baselines.py -d ereninho -m rfecv -a nb

For new users, prefer the package API or the agraspqfs CLI.

Repository structure

.
├── src/agraspqfs/              # Production package
│   ├── selector.py             # AGraspQFeatureSelector
│   ├── cli.py                  # agraspqfs command
│   └── _version.py
├── tests/                      # Unit tests
├── examples/                   # Quickstart, pipeline, benchmark and notebook examples
├── docs/                       # Documentation and release checklist
├── .github/workflows/          # CI and PyPI publishing workflows
├── main.py                     # Legacy research prototype
├── baselines.py                # Baseline methods
├── utils.py                    # Legacy utilities
├── priority_queue.py           # Legacy queue implementation
├── data/                       # Example/research datasets
├── results/                    # Runtime outputs, ignored except .gitkeep
├── CITATION.cff                # Citation metadata
├── LICENSE                     # MIT license
├── pyproject.toml              # Packaging metadata
└── README.md

Notes on data preprocessing

A-GRASPQ-FS expects the feature matrix passed to fit to be numeric, as is common for scikit-learn estimators and mutual information ranking. For categorical features, use a preprocessing pipeline before the selector, for example OneHotEncoder or ColumnTransformer.

The CLI includes automatic preprocessing for CSV files: numeric columns are scaled and categorical columns are one-hot encoded.

Performance note

A-GRASPQ-FS is a wrapper method and can be computationally expensive because it repeatedly evaluates machine learning models. Use preset="fast", max_evaluations, time_budget, early_stopping_rounds, and evaluation_sample_size to control cost. See docs/PERFORMANCE.md.

License

This project is released under the MIT License. See LICENSE.

The license permits reuse, modification and distribution. For academic use, please cite the associated ISCC 2026 paper and/or the software release as described in CITATION.cff.

Citation

@inproceedings{quincozes2026agraspqfs,
  title     = {Adaptive Feature Selection with Self-Tuning Subset Size for Intrusion Detection},
  author    = {Quincozes, Vagner E. and Quincozes, Silvio E. and Albuquerque, Célio and Passos, Diego and Mossé, Daniel},
  booktitle = {Proceedings of ISCC 2026},
  address   = {Algarve, Portugal},
  year      = {2026}
}

Update the BibTeX with DOI/pages/publisher once the final proceedings metadata is available.