All models¶

Every model aspective has considered, and what became of it: the ones in the package, the ones the benchmark gate removed, the ones dropped before a benchmark could run, shortlisted ones not built yet, and the ones left out because they already have a maintained home or are out of scope. Negative results are kept on purpose: a published method that does not beat its vanilla version on dozens of public datasets is worth knowing about, and the code of every removed model stays in the git history at the commit named.

Generated by uv run aspective-bench report from catalog.py and the committed benchmark results; do not edit by hand.

Status	Models
In aspective	28
Removed by the benchmark gate	13
Dropped before benchmarking	3
Shortlisted, not built yet	7
Left out: a maintained home exists, or out of scope	13

In aspective¶

Model	Task	Paper	Improves on	Gate result
`RegularizedDiscriminantAnalysis`	classification	Friedman, JASA 1989	LDA, QDA	superior vs LDA (median accuracy +0.009, 68 datasets); non-inferior vs QDA (median accuracy +0.003, 33 datasets)
`MixtureDiscriminantAnalysis`	classification	Hastie & Tibshirani, JRSS-B 1996	LDA	superior vs LDA (median accuracy +0.002, 68 datasets)
`AODEClassifier`	classification	Webb, Boughton & Wang, Machine Learning 2005	Naive Bayes	superior vs Naive Bayes (5 bins) (median accuracy +0.021, 68 datasets)
`A2DEClassifier`	classification	Webb, Boughton, Zheng, Ting & Salem, Machine Learning 2012	Naive Bayes	superior vs Naive Bayes (5 bins) (median accuracy +0.022, 68 datasets)
`TANClassifier`	classification	Friedman, Geiger & Goldszmidt, Machine Learning 1997	Naive Bayes	superior vs Naive Bayes (5 bins) (median accuracy +0.013, 68 datasets)
`KDBClassifier`	classification	Sahami, KDD 1996	Naive Bayes	superior vs Naive Bayes (5 bins) (median accuracy +0.022, 68 datasets)
`HiddenNaiveBayesClassifier`	classification	Zhang, Jiang & Su, AAAI 2005	Naive Bayes	superior vs Naive Bayes (5 bins) (median accuracy +0.020, 68 datasets)
`FlexibleNaiveBayes`	classification	John & Langley, UAI 1995	Gaussian naive Bayes	superior vs Gaussian NB (median accuracy +0.023, 68 datasets)
`DANNClassifier`	classification	Hastie & Tibshirani, IEEE TPAMI 1996	k-NN	non-inferior vs k-NN (k=5) (median accuracy +0.006, 68 datasets)
`RotationForestClassifier`	classification	Rodríguez, Kuncheva & Alonso, IEEE TPAMI 2006	Random forest	non-inferior vs Random forest (median accuracy +0.000, 68 datasets)
`LogitBoostClassifier`	classification	Friedman, Hastie & Tibshirani, Annals of Statistics 2000	AdaBoost	superior vs AdaBoost (depth 3) (median accuracy +0.002, 68 datasets)
`MultiBoostClassifier`	classification	Webb, Machine Learning 2000	AdaBoost	non-inferior vs AdaBoost (depth 3) (median accuracy +0.003, 68 datasets)
`FirthLogisticRegression`	classification	Firth, Biometrika 1993; Puhr et al., Statistics in Medicine 2017	Logistic regression	non-inferior vs ML logistic (median Brier score -0.000, 31 datasets)
`CauchyLogisticRegression`	classification	Gelman, Jakulin, Pittau & Su, Annals of Applied Statistics 2008	Logistic regression	superior vs ML logistic (median Brier score -0.000, 31 datasets)
`LoessRegressor`	regression	Cleveland & Devlin, JASA 1988	k-NN regression	superior vs k-NN (tuned k) (median R² +0.030, 52 datasets)
`LocalLinearForestRegressor`	regression	Friedberg, Tibshirani, Athey & Wager, JCGS 2020	Random forest	superior vs Random forest (leaf 5) (median R² +0.006, 52 datasets)
`M5ModelTreeRegressor`	regression	Quinlan 1992; Wang & Witten, ECML 1997	CART	superior vs CART (tuned leaf size) (median R² +0.052, 52 datasets)
`RelaxedLasso, RelaxedLassoCV`	regression	Meinshausen, CSDA 2007	Lasso	non-inferior vs LassoCV (median R² -0.000, 52 datasets)
`AdaptiveLasso, AdaptiveLassoCV`	regression	Zou, JASA 2006	Lasso	non-inferior vs LassoCV (median R² -0.000, 52 datasets)
`NonconvexPenaltyRegressionCV (MCP)`	regression	Zhang, Annals of Statistics 2010	Lasso	non-inferior vs LassoCV (median R² -0.000, 52 datasets)
`NonconvexPenaltyRegressionCV (SCAD)`	regression	Fan & Li, JASA 2001	Lasso	non-inferior vs LassoCV (median R² -0.000, 52 datasets)
`Bolasso`	regression	Bach, ICML 2008	Lasso	non-inferior vs LassoCV (median R² -0.002, 52 datasets)
`L2BoostRegressor`	regression	Bühlmann & Yu, JASA 2003; Bühlmann, Annals of Statistics 2006	Lasso	non-inferior vs LassoCV (median R² -0.000, 52 datasets)
`MARSRegressor`	regression	Friedman, Annals of Statistics 1991	Linear regression, CART	superior vs LassoCV (median R² +0.026, 52 datasets); superior vs CART (tuned leaf size) (median R² +0.025, 52 datasets)
`PSplineGAMRegressor`	regression	Hastie & Tibshirani, Statistical Science 1986; Eilers & Marx, Statistical Science 1996	Linear regression	superior vs LassoCV (median R² +0.011, 52 datasets)
`RSHashDetector`	anomaly detection	Sathe & Aggarwal, ICDM 2016	Isolation Forest	non-inferior vs Isolation Forest (median ROC AUC +0.006, 46 datasets)
`LeSiNNDetector`	anomaly detection	Pang, Ting & Albrecht, ICDM Workshops 2015	Isolation Forest	non-inferior vs Isolation Forest (median ROC AUC +0.006, 46 datasets)
`ExtendedIsolationForestDetector`	anomaly detection	Hariri, Carrasco Kind & Brunner, IEEE TKDE 2021	Isolation Forest	non-inferior vs Isolation Forest (median ROC AUC +0.002, 46 datasets)

Removed by the benchmark gate¶

Model	Task	Paper	Improves on	Why
`RandomRotationClassifier`	classification	Blaser & Fryzlewicz, JMLR 2016	Random forest	Inferior on the larger suites: median accuracy −0.006, 11 severe losses against 1 severe win (earlier inconclusive with min-max and rank scaling). Code: `2fecda7`.
`RandomRotationRegressor`	regression	Blaser & Fryzlewicz, JMLR 2016	Random forest	Inferior: median R² −0.023, worse on 15 of 25 datasets. Code: `2fecda7`.
`IDKDetector (Isolation Distributional Kernel)`	anomaly detection	Ting, Xu, Washio & Zhou, KDD 2020	Gaussian KDE, one-class SVM	Inconclusive on the larger suites vs Gaussian KDE (median ROC AUC −0.012) and one-class SVM (−0.003); the hypersphere kernel was inconclusive or worse at ψ = 16, 64, 256. Code: `2fecda7`.
`RobustKDEDetector`	anomaly detection	Kim & Scott, JMLR 2012	Gaussian KDE	Inconclusive on the larger suites: median ROC AUC +0.000, p = 0.08; also with the paper's median nearest-neighbour bandwidth. Code: `2fecda7`.
`SupervisedPCARegressor`	regression	Bair, Hastie, Paul & Tibshirani, JASA 2006	PCR	Inferior: median R² −0.010, worse on 13 of 25; built for many more features than samples. Code: `2aa9e33`.
`LoOPDetector`	anomaly detection	Kriegel, Kröger, Schubert & Zimek, CIKM 2009	LOF	Inconclusive on the larger suites: median ROC AUC −0.012, worse on 25 of 46. Code: `2aa9e33`.
`DARTClassifier`	classification	Rashmi & Gilad-Bachrach, AISTATS 2015	Gradient boosting	Inferior on the larger suites: median accuracy −0.005, worse on 28 of 68. Code: `b2bf66c`.
`DARTRegressor`	regression	Rashmi & Gilad-Bachrach, AISTATS 2015	Gradient boosting	Inconclusive on the larger suites: median R² −0.001, worse on 12 of 31. Code: `b2bf66c`.
`SqrtLasso`	regression	Belloni, Chernozhukov & Wang, Biometrika 2011	LassoCV	Inferior: the pivotal penalty over-shrinks, median R² −0.012, worse on 14 of 25. Code: `b2bf66c`.
`SCiForestDetector`	anomaly detection	Liu, Ting & Zhou, ECML PKDD 2010	Isolation Forest	Inconclusive on the larger suites: median ROC AUC −0.002, worse on 20 of 46. Code: `b2bf66c`.
`INFLODetector`	anomaly detection	Jin, Tung, Han & Wang, PAKDD 2006	LOF	Inferior: median ROC AUC −0.019, worse on 21 of 37. Code: `b2bf66c`.
`AntiHubDetector (AntiHub2)`	anomaly detection	Radovanović, Nanopoulos & Ivanović, IEEE TKDE 2015	k-NN distance detector	Inferior: median ROC AUC −0.078, worse on 28 of 37; built for very high-dimensional data. Code: `b2bf66c`.
`LDOFDetector`	anomaly detection	Zhang, Hutter & Jin, PAKDD 2009	k-NN distance detector	Inferior: median ROC AUC −0.121, worse on 27 of 37. Code: `b2bf66c`.

Dropped before benchmarking¶

Model	Task	Paper	Improves on	Why
`KDEOS`	anomaly detection	Schubert, Zimek & Kriegel, SDM 2014	LOF	The exact definition could not be checked against the paper, and the version built from secondary descriptions lost to LOF on a planted-outlier sanity check.
`MM robust regression`	regression	Yohai, Annals of Statistics 1987	Least squares	Dropped at the design stage of the second wave, before any code or benchmark.
`Non-negative garrote`	regression	Breiman, Technometrics 1995	Lasso	Dropped at the design stage of the second wave, before any code or benchmark.

Shortlisted, not built yet¶

Model	Task	Paper	Improves on	Why
`Logistic Model Trees`	classification	Landwehr, Hall & Frank, Machine Learning 2005	C4.5 / CART, logistic regression	Needs a tree builder with LogitBoost carried down the tree; only Weka has it.
`Conditional inference forests`	classification, regression	Hothorn, Hornik & Zeileis, JCGS 2006	CART, random forest	Large effort (permutation-test splits); its gain is unbiased variable importance, which the accuracy-based gate cannot measure.
`Distributional random forests`	regression	Ćevid, Michel, Näf, Bühlmann & Meinshausen, JMLR 2022	Random forest, quantile regression forest	Large effort; it predicts distributions, so it would be judged on the probabilistic track (CRPS).
`Twin SVM`	classification	Jayadeva, Khemchandani & Chandra, IEEE TPAMI 2007	SVM	Not reached yet.
`Relevance vector machine`	classification, regression	Tipping, JMLR 2001	SVM, SVR	Not reached yet.
`PIDForest`	anomaly detection	Gopalan, Sharan & Wieder, NeurIPS 2019	Isolation Forest	Not reached yet.
`HiCS`	anomaly detection	Keller, Müller & Böhm, ICDE 2012	LOF, feature bagging	Not reached yet.

Left out: a maintained home exists, or out of scope¶

Model	Task	Paper	Improves on	Why
`NGBoost`	classification, regression	Duan et al., ICML 2020	Gradient boosting	Maintained in `ngboost`. Its strength is predictive distributions, which the probabilistic track (CRPS, NLL, log loss) can now score if it is ever reconsidered.
`Sparse projection oblique forests (SPORF)`	classification, regression	Tomita et al., JMLR 2020	Random forest	Maintained in `treeple`.
`Quantile regression forests`	regression	Meinshausen, JMLR 2006	Random forest	Maintained in `quantile-forest`.
`Generalized and causal random forests`	regression	Athey, Tibshirani & Wager, Annals of Statistics 2019	Random forest	Maintained in `EconML`.
`Hierarchical shrinkage, FIGS, GOSDT`	classification, regression	ICML 2022; PNAS; ICML 2020	CART	Maintained in `imodels`.
`BART`	regression	Chipman, George & McCulloch, Annals of Applied Statistics 2010	Tree ensembles	Maintained in `stochtree` and `pymc-bart`.
`Large-margin nearest neighbours (LMNN)`	classification	Weinberger & Saul, JMLR 2009	k-NN	Maintained in `metric-learn` (scikit-learn-contrib).
`Regularized greedy forest`	classification, regression	Johnson & Zhang, IEEE TPAMI 2014	Gradient boosting	Maintained in `rgf-python`.
`Deep Isolation Forest, iNNE, LODA, ECOD, COPOD, GLOSH`	anomaly detection	IEEE TKDE, Machine Learning, ICDM, TKDD	Isolation Forest, LOF	Maintained in PyOD.
`Aggregated Mondrian forests`	classification, regression	Mourtada, Gaïffas & Scornet, JRSS-B 2021	Random forest	Online learning, out of scope.
`Robust random cut forest`	anomaly detection	Guha, Mishra, Roy & Schrijvers, ICML 2016	Isolation Forest	Streaming, out of scope.
`xStream`	anomaly detection	Manzoor, Lamba & Akoglu, KDD 2018	Isolation Forest	Streaming, out of scope.
`Online Isolation Forest`	anomaly detection	Leveni, Weigert Cassales, Pfahringer, Bifet & Boracchi, ICML 2024	Isolation Forest	Online learning, out of scope.