Skip to content

Pipeline Chart

Every named pipeline in amos, step by step. A pipeline fills in the settings of a project's three workers: the analyst prepares the data and fits the models, the critic cross-validates and evaluates them, and the artist draws the figures. Name one with pipeline in the "general" section (see pipeline), and every setting that you write yourself wins over the pipeline's:

[general]
label = target
seed = 43
pipeline = classification

A step whose technique is auto adapts to the data when the project runs (see the choosers in the technique catalog). The catalog's pipelines table describes each pipeline in a sentence.

Analyst
split→impute→scale→encode→select→sample→model
Dashed steps are applied only when the model needs them.
Critic
validator→auto_evaluator
Cross-validates on the training rows, then applies the evaluators that suit the fitted model.
Artist
auto_plot+extra plots
Draws the figures that suit the model, and any that the pipeline adds.

Pipelines for studies

8 pipelines are named for a kind of study. Those that predict compare several models in an experiment. Those for inference fit one model and keep its coefficients in the units of the features, so they neither scale nor resample.

AnalystCriticArtist
pipelinesplitimputescaleencodesamplemodelvalidateevaluatedraw
classificationclassifyPredicts classes, comparing a logit model, a random forest, and gradient boosting by balanced accuracy.stratifiedautostandardautoautologitrandom_forestgradient_boostingbaselinecompared by balanced_accuracystratified_k_foldauto_evaluatorauto_plot
regressionregressPredicts a number, comparing ridge regression, a random forest, and gradient boosting by RMSE.train_testautostandardauto—ridgerandom_forestgradient_boostingbaselinecompared by rmsek_foldauto_evaluatorauto_plot
count_regressionregressPredicts counts, comparing Poisson and negative binomial regression with a random forest and gradient boosting fitted for counts, by RMSE.train_testautostandardauto—poissonnegative_binomialrandom_forestgradient_boostingbaselinecompared by rmsek_foldauto_evaluatorauto_plot
classification_inferenceclassifyEstimates the effects of the features on classes with a logit, multinomial logit, or ordinal model.stratifiedauto—auto—ordinal_regressionlogit_smmnlogitauto_model chooses onestratified_k_foldauto_evaluatorauto_plot
regression_inferenceregressEstimates the effects of the features on a number with ordinary least squares, or robust regression for errors with outliers.train_testauto—auto—robust_regressionolsauto_model chooses onek_foldauto_evaluatorauto_plot
count_inferenceregressEstimates the effects of the features on counts with the Poisson, negative binomial, or zero-inflated Poisson model that suits their dispersion.train_testauto—auto—poissonnegative_binomialzero_inflated_poissonauto_model chooses onek_foldauto_evaluatorauto_plot
panel_classificationclassifyneeds "groups" in [general]Estimates the effects of the features on classes in panel data with a fixed effects logit model and standard errors clustered by group.stratifiedauto—auto—fixestgroups as fixed effects, clustered by the firststratified_k_foldauto_evaluatorauto_plot
panel_regressionregressneeds "groups" in [general]Estimates the effects of the features on a number in panel data with a fixed effects model and standard errors clustered by group.train_testauto—auto—fixestgroups as fixed effects, clustered by the firstk_foldauto_evaluatorauto_plot

Pipelines of each model

53 pipelines, one for each model and each task that it can do. Each prepares the data only as its model needs, so they are grouped by the family of the model.

— step skippedgrid or random: a hyperparameter search, scored by balanced accuracy or RMSE

AnalystCriticArtist
pipelinesplitimputescaleencodeselectsamplesearchvalidatorextra plots
Trees Not scaled; classes weighted (or resampled, for AdaBoost). Importances, SHAP values, partial dependence, and a tree plot.
adaboost_classificationclassifystratifiedauto—auto—autogridstratified_k_fold—
adaboost_regressionregresstrain_testauto—auto——gridk_fold—
decision_tree_classificationclassifystratifiedauto—auto——gridstratified_k_fold—
decision_tree_regressionregresstrain_testauto—auto——gridk_fold—
extra_trees_classificationclassifystratifiedauto—auto———stratified_k_fold—
extra_trees_regressionregresstrain_testauto—auto———k_fold—
random_forest_classificationclassifystratifiedauto—auto———stratified_k_fold—
random_forest_regressionregresstrain_testauto—auto———k_fold—
Boosted trees that handle missing values Missing values left in; not scaled. The evaluators and plots of trees.
catboost_classificationclassifystratified——auto———stratified_k_fold—
catboost_regressionregresstrain_test——auto———k_fold—
gradient_boosting_classificationclassifystratified——auto——randomstratified_k_fold—
gradient_boosting_regressionregresstrain_test——auto——randomk_fold—
lightgbm_classificationclassifystratified——auto——randomstratified_k_fold—
lightgbm_regressionregresstrain_test——auto——randomk_fold—
xgboost_classificationclassifystratified——auto—autorandomstratified_k_fold—
xgboost_regressionregresstrain_test——auto——randomk_fold—
Penalized linear models Standardized features. Coefficients as importances; a Q-Q plot.
elastic_net_regressionregresstrain_testautostandardauto——gridk_fold—
lasso_regressionregresstrain_testautostandardauto——gridk_fold—
logit_classificationclassifystratifiedautostandardauto——gridstratified_k_fold—
ridge_regressionregresstrain_testautostandardauto——gridk_fold—
Least squares Not scaled; categories encoded against a reference level.
linear_regressionregresstrain_testauto—auto———k_fold—
Distances, kernels, and networks Standardized features. Permutation importance and partial dependence.
knn_classificationclassifystratifiedautostandardautok_bestautogridstratified_k_fold—
knn_regressionregresstrain_testautostandardautok_best—gridk_fold—
neural_network_classificationclassifystratifiedautostandardauto—autogridstratified_k_foldlearning_curve
neural_network_regressionregresstrain_testautostandardauto——gridk_foldlearning_curve
svm_classificationclassifystratifiedautostandardauto——gridstratified_k_fold—
svm_regressionregresstrain_testautostandardauto——gridk_fold—
Gaussian Features made more normal (Yeo-Johnson); a calibration curve.
naive_bayes_classificationclassifystratifiedautogaussauto—auto—stratified_k_fold—
Additive Missing values left in; the shape function of each feature.
explainable_boosting_classificationclassifystratified——auto———stratified_k_fold—
explainable_boosting_regressionregresstrain_test——auto———k_fold—
Pretrained TabPFN fills, scales, and weighs the features itself.
tabpfn_classificationclassifystratified——auto———stratified_k_fold—
tabpfn_regressionregresstrain_test——auto———k_fold—
Models for inference Categories encoded against a reference level; not scaled or resampled. A regression table and Q-Q, influence, and coefficient plots.
binomial_bayes_mixedglm_classificationclassifyneeds "groups" in [general]group_splitauto—auto———stratified_group_k_fold—
gee_classificationclassifyneeds "groups" in [general]group_splitauto—auto———stratified_group_k_fold—
gee_regressionregressneeds "groups" in [general]group_splitauto—auto———group_k_fold—
generalized_poisson_regressionregresstrain_testauto—auto———k_fold—
glm_classificationclassifystratifiedauto—auto———stratified_k_fold—
glm_regressionregresstrain_testauto—auto———k_fold—
logit_sm_classificationclassifystratifiedauto—auto———stratified_k_fold—
mixedlm_regressionregressneeds "groups" in [general]group_splitauto—auto———group_k_fold—
mnlogit_classificationclassifystratifiedauto—auto———stratified_k_fold—
negative_binomial_regressionregresstrain_testauto—auto———k_fold—
ols_regressionregresstrain_testauto—auto———k_fold—
ordinal_inferenceclassifystratifiedauto—auto———stratified_k_fold—
poisson_regressionregresstrain_testauto—auto———k_fold—
probit_classificationclassifystratifiedauto—auto———stratified_k_fold—
quantile_inferenceregresstrain_testauto—auto———k_fold—
robust_inferenceregresstrain_testauto—auto———k_fold—
wls_regressionregressneeds "weights" in [wls_parameters]train_testauto—auto———k_fold—
zero_inflated_poisson_regressionregresstrain_testauto—auto———k_fold—
Survival Hazard ratios and Kaplan-Meier survival curves.
cox_regressionregressneeds "event" in [cox_parameters]train_testauto—auto———k_foldsurvival_curves
Baseline Ignores the features: the score that every other model must beat.
baseline_classificationclassifystratified——auto———stratified_k_foldlabel_plot
baseline_regressionregresstrain_test——auto———k_foldlabel_plot