aspective

Naive Bayes, LDA, k-NN, CART, random forests, AdaBoost, the lasso, logistic regression and Isolation Forest are the workhorses of applied machine learning. The literature has published variants of each that fix a known weakness and compared them carefully with the original, but most never reached a mainstream Python library. aspective collects them.

Why aspective?

• From the literature. Every estimator comes from a paper in a strong venue that compared it with its vanilla version in a principled way. Its docstring cites that paper and says when to prefer it.
• Benchmark-gated. On dozens of public datasets, every estimator is statistically non-inferior to the vanilla model it improves on (protocol, results). Candidates that fail the gate are removed, and the benchmark README says which.
• Drop-in. Supervised estimators pass scikit-learn's check_estimator, so pipelines, grid search and cross-validation just work. Anomaly detectors follow PyOD's BaseDetector contract.

Installation

pip install aspective          # classifiers and regressors
pip install "aspective[pyod]"  # detectors then subclass PyOD's BaseDetector

The anomaly detectors work without PyOD through a small stand-in base class with the same contract. With PyOD installed they are genuine PyOD detectors and work inside PyOD's ensembles (LSCP, SUOD), thresholding and benchmarks.

Quickstart

from sklearn.datasets import load_wine
from sklearn.model_selection import GridSearchCV
from aspective.discriminant_analysis import RegularizedDiscriminantAnalysis

X, y = load_wine(return_X_y=True)
search = GridSearchCV(
    RegularizedDiscriminantAnalysis(),
    {"lam": [0, 0.25, 0.5, 0.75, 1], "gamma": [0, 0.1, 0.5]},
).fit(X, y)

from aspective.anomaly import RSHashDetector

detector = RSHashDetector(contamination=0.05, random_state=0).fit(X)
detector.decision_scores_  # higher means more anomalous
detector.predict(X)  # 0 inlier, 1 outlier, as in PyOD

Estimators

Start from the model you already use; the guide says which variant helps when.

Estimator	Improves on	Paper
discriminant_analysis.RegularizedDiscriminantAnalysis	LDA, QDA	Friedman, JASA 1989
discriminant_analysis.MixtureDiscriminantAnalysis	LDA	Hastie & Tibshirani, JRSS-B 1996
naive_bayes.AODEClassifier	Naive Bayes	Webb, Boughton & Wang, Machine Learning 2005
naive_bayes.A2DEClassifier	Naive Bayes	Webb, Boughton, Zheng, Ting & Salem, Machine Learning 2012
naive_bayes.TANClassifier	Naive Bayes	Friedman, Geiger & Goldszmidt, Machine Learning 1997
naive_bayes.KDBClassifier	Naive Bayes	Sahami, KDD 1996
naive_bayes.HiddenNaiveBayesClassifier	Naive Bayes	Zhang, Jiang & Su, AAAI 2005
naive_bayes.FlexibleNaiveBayes	Gaussian naive Bayes	John & Langley, UAI 1995
neighbors.DANNClassifier	k-NN classification	Hastie & Tibshirani, IEEE TPAMI 1996
neighbors.LoessRegressor	k-NN regression	Cleveland & Devlin, JASA 1988
ensemble.RotationForestClassifier	Random forest	Rodríguez, Kuncheva & Alonso, IEEE TPAMI 2006
ensemble.LogitBoostClassifier	AdaBoost	Friedman, Hastie & Tibshirani, Annals of Statistics 2000
ensemble.MultiBoostClassifier	AdaBoost	Webb, Machine Learning 2000
ensemble.LocalLinearForestRegressor	Random forest	Friedberg, Tibshirani, Athey & Wager, JCGS 2020
tree.M5ModelTreeRegressor	Regression tree (CART)	Quinlan 1992; Wang & Witten, ECML 1997
linear_model.FirthLogisticRegression	Logistic regression	Firth, Biometrika 1993; Puhr et al., Statistics in Medicine 2017 (FLIC/FLAC)
linear_model.CauchyLogisticRegression	Logistic regression

Citing

If you use aspective, please cite it via CITATION.cff (or GitHub's "Cite this repository" button) and the original paper of each estimator you use.

Contributing

Proposals for new estimators are welcome: open an estimator proposal. See CONTRIBUTING.md for the bar an estimator has to clear and the development setup:

uv sync --extra pyod                 # package, tests and the bench/ testbed
uv run pytest                        # unit tests and the benchmark gate check
uv run aspective-bench run my-model  # benchmark a model, see bench/README.md