Technique Catalog

Every technique in amos, by the stage (worker) that usually uses it. A technique is named in settings by the name in the first column, and its class is in the module of its stage (for example, random_forest is amos.models.RandomForest). The advanced user guide describes how techniques work, their parameters, and how to write your own.

Each technique is listed with the tool it wraps. Optional packages are only imported when a technique that needs them is used (see the extras in the overview). Most models do both tasks. The cat_boost encoder (CatBoost-style target encoding, from category_encoders) is not the catboost model.

Loaders (wrangler)

Name	Tool	Description
download		Downloads a file from a URL and loads it.
load_file		Loads a data file in any format that the clerk knows.
openml	sklearn.datasets.fetch_openml	Downloads a dataset from OpenML (by its name or id) and loads it.

Cleaners (wrangler)

Name	Description
drop_columns	Removes columns.
drop_constant	Removes columns that have only one value (and so tell you nothing).
drop_duplicates	Removes rows that duplicate an earlier row.
drop_missing	Removes rows with missing values.
filter_rows	Keeps the rows that match a query.
keep_columns	Keeps only some columns. The label and groups are always kept.
rename_columns	Renames columns, including the label and groups of the dataset.

Mungers (wrangler)

Name	Description
auto_categorize	Makes columns with few unique values categorical.
coalesce	Takes the first value that is not missing from several columns.
combine_flags	Combines flags into one: whether any or all are true, or how many are.
convert_types	Changes the data types of columns.
count_patterns	Counts the matches of patterns in text.
derive_columns	Makes columns from expressions of other columns, such as "a / b".
extract_all	Keeps every match of a pattern in text.
extract_pattern	Keeps the first match of a pattern in text (or the groups of a match).
flag_patterns	Makes columns that say whether text matches patterns.
map_patterns	Turns text into values by the first pattern that it matches.
map_values	Replaces values with others, such as "N/A" with a missing value.
normalize_text	Normalizes the spaces, case, accents, and punctuation of text.
parse_booleans	Turns text such as "yes" and "no" into booleans.
parse_dates	Turns text into dates.
parse_numbers	Turns text such as "$1,234.50" into numbers.
replace_text	Replaces the matches of patterns in text.
split_text	Splits text into columns where a pattern matches.
strip_text	Trims spaces from text, and optionally makes it lowercase.

Mergers (wrangler)

Name	Description
merge_keys	Adds columns from another table's row that has the same keys.
merge_nearest	Adds columns from another table's row with the nearest date or number.
merge_ranges	Adds columns from another table's row whose range has a date or number.
merge_summary	Adds summaries of the rows of another table that have the same keys.

Shapers (wrangler)

Name	Description
group_rows	Makes one row for each group of rows, with summaries of its columns.
lists_to_rows	Makes a row for each item of a list, such as each judge of a case.
long_to_wide	Spreads the rows of each group into columns, such as one for each year.
wide_to_long	Stacks columns into rows, such as a row for each year.

Describers (explorer)

Name	Description
correlations	Correlations between the numeric columns (including the label).
describe	The pandas description of every column, one row per column.
frequencies	The count and share of each value of the categorical columns.
kaplan_meier	The share of rows without an event over time (a survival curve).
label_balance	The count and share of each value of the label.
missing_values	The count and share of missing values in each column.
summarize	Summary statistics of the numeric columns, as reported in papers.

Splitters (analyst)

Name	Description
group_split	Keeps all of the rows of each group in the same set.
stratified	Splits at random, keeping the share of each class in both sets.
time_split	Uses the latest rows as the test set.
train_test	Splits the rows at random.

Imputers (analyst)

Name	Tool	Description
iterative_impute	sklearn.impute.IterativeImputer	Fills missing values by modeling each feature from the others.
knn_impute	sklearn.impute.KNNImputer	Fills missing values with the average of the most similar rows.
mean_impute	sklearn.impute.SimpleImputer	Fills missing values with the mean of the training rows.
median_impute	sklearn.impute.SimpleImputer	Fills missing values with the median of the training rows.
mode_impute	sklearn.impute.SimpleImputer	Fills missing values with the most common value of the training rows.

Scalers (analyst)

Name	Tool	Description
binarize	sklearn.preprocessing.Binarizer	Makes each feature 1 if it is above a threshold (0 by default).
bins	sklearn.preprocessing.KBinsDiscretizer	Sorts each feature into bins with about the same number of rows.
gauss	sklearn.preprocessing.PowerTransformer	Makes each feature more like a normal distribution (Yeo-Johnson).
max_abs	sklearn.preprocessing.MaxAbsScaler	Divides each feature by its largest absolute value.
min_max	sklearn.preprocessing.MinMaxScaler	Rescales each feature to the range from 0 to 1.
normalize	sklearn.preprocessing.Normalizer	Rescales each row to a length of 1.
quantile	sklearn.preprocessing.QuantileTransformer	Replaces each value with its quantile in the training rows.
robust	sklearn.preprocessing.RobustScaler	Centers on the median and scales by the interquartile range.
standard	sklearn.preprocessing.StandardScaler	Centers each feature on 0 with a standard deviation of 1.

Encoders (analyst)

Name	Tool	Description
backward_difference	category_encoders.BackwardDifferenceEncoder	Compares each category to the one before it.
base_n	category_encoders.BaseNEncoder	Writes the number of each category in base N (4 by default).
binary	category_encoders.BinaryEncoder	Writes the number of each category in binary digits.
cat_boost	category_encoders.CatBoostEncoder	Target encoding in the manner of CatBoost, which limits leakage.
count	category_encoders.CountEncoder	Replaces each category with how often it is in the training rows.
date_parts	skrub.DatetimeEncoder	Splits dates and times into parts: year, month, day, and so on.
gap	skrub.GapEncoder	Encodes messy text as a mix of topics of its substrings.
hashing	category_encoders.HashingEncoder	Hashes the categories into a fixed number of columns.
helmert	category_encoders.HelmertEncoder	Compares each category to the mean of the categories before it.
james_stein	category_encoders.JamesSteinEncoder	Target encoding shrunk toward the overall mean (James-Stein).
leave_one_out	category_encoders.LeaveOneOutEncoder	Target encoding that leaves out each row's own label.
m_estimate	category_encoders.MEstimateEncoder	Target encoding shrunk toward the overall mean by m rows.
min_hash	skrub.MinHashEncoder	Encodes messy text by hashing its substrings, which is fast and robust.
one_hot	sklearn.preprocessing.OneHotEncoder	Makes a column of 0s and 1s for each category (dummy variables).
ordinal	sklearn.preprocessing.OrdinalEncoder	Numbers the categories (0, 1, 2, ...) in sorted order.
polynomial_coding	category_encoders.PolynomialEncoder	Contrasts categories as an ordered (polynomial) sequence.
sum_coding	category_encoders.SumEncoder	Compares each category to the mean of all categories (effect coding).
target	category_encoders.TargetEncoder	Replaces each category with the mean label of its training rows.
tfidf	skrub.StringEncoder	Encodes text by the TF-IDF of its substrings, reduced with SVD.
weight_of_evidence	category_encoders.WOEEncoder	Replaces each category with its weight of evidence (binary labels).

Mixers (analyst)

Name	Tool	Description
interactions	sklearn.preprocessing.PolynomialFeatures	Adds the product of each pair of features.
polynomial	sklearn.preprocessing.PolynomialFeatures	Adds the squares and products of the features (degree 2 by default).
splines	sklearn.preprocessing.SplineTransformer	Replaces each feature with a set of spline curves.

Reducers (analyst)

Name	Tool	Description
k_best	sklearn.feature_selection.SelectKBest	Keeps the k features (10 by default) most related to the label.
pca_reduce	sklearn.decomposition.PCA	Replaces the features with their principal components.
select_percentile	sklearn.feature_selection.SelectPercentile	Keeps the features (50% by default) most related to the label.
variance_threshold	sklearn.feature_selection.VarianceThreshold	Removes features whose variance is at or below a threshold (0).

Samplers (analyst)

Name	Tool	Description
adasyn	imblearn.over_sampling.ADASYN	Adds synthetic rows where the rare class is hardest to learn.
borderline_smote	imblearn.over_sampling.BorderlineSMOTE	Adds synthetic rows near the border between the classes.
near_miss	imblearn.under_sampling.NearMiss	Removes rows of the common classes that are far from the rare rows.
random_over	imblearn.over_sampling.RandomOverSampler	Duplicates random rows of the rare classes.
random_under	imblearn.under_sampling.RandomUnderSampler	Removes random rows of the common classes.
smote	imblearn.over_sampling.SMOTE	Adds synthetic rows between rare rows and their nearest neighbors.
smote_enn	imblearn.combine.SMOTEENN	Applies smote and then removes rows that their neighbors misclassify.
smote_tomek	imblearn.combine.SMOTETomek	Applies smote and then removes Tomek links.
tomek_links	imblearn.under_sampling.TomekLinks	Removes rows of the common class that are paired with rare rows.

Models (analyst)

Name	Classify	Regress	Description
adaboost	sklearn.ensemble.AdaBoostClassifier	sklearn.ensemble.AdaBoostRegressor	AdaBoost: a sequence of small trees, each fixing the last one's errors.
baseline	sklearn.dummy.DummyClassifier	sklearn.dummy.DummyRegressor	Predicts the most common class (or the mean) for every row.
binomial_bayes_mixedglm	amos.models.Statsmodel		A Bayesian logistic regression with a random intercept for each group.
catboost	catboost.CatBoostClassifier	catboost.CatBoostRegressor	CatBoost gradient boosting, which uses categorical features directly.
cox		amos.models.ProportionalHazards	Cox proportional hazards regression of the time until an event.
decision_tree	sklearn.tree.DecisionTreeClassifier	sklearn.tree.DecisionTreeRegressor	A single decision tree.
elastic_net		sklearn.linear_model.ElasticNet	Linear regression with both lasso and ridge penalties.
explainable_boosting	interpret.glassbox.ExplainableBoostingClassifier	interpret.glassbox.ExplainableBoostingRegressor	An Explainable Boosting Machine from InterpretML.
extra_trees	sklearn.ensemble.ExtraTreesClassifier	sklearn.ensemble.ExtraTreesRegressor	An ensemble of extremely randomized trees.
fixest	amos.models.FixedEffects	amos.models.FixedEffects	Regression with fixed effects and clustered standard errors (pyfixest).
gee	amos.models.Statsmodel	amos.models.Statsmodel	Generalized estimating equations from statsmodels, with inference.
generalized_poisson		amos.models.Statsmodel	Generalized Poisson regression of counts, with inference.
glm	amos.models.Statsmodel	amos.models.Statsmodel	A generalized linear model from statsmodels, with inference.
gradient_boosting	sklearn.ensemble.HistGradientBoostingClassifier	sklearn.ensemble.HistGradientBoostingRegressor	Histogram-based gradient boosting from scikit-learn.
knn	sklearn.neighbors.KNeighborsClassifier	sklearn.neighbors.KNeighborsRegressor	Predicts from the k nearest training rows (5 by default).
lasso		sklearn.linear_model.Lasso	Linear regression with a lasso (L1) penalty.
lightgbm	lightgbm.LGBMClassifier	lightgbm.LGBMRegressor	LightGBM gradient boosting.
linear		sklearn.linear_model.LinearRegression	Ordinary least squares regression from scikit-learn.
logit	sklearn.linear_model.LogisticRegression		Logistic regression from scikit-learn.
logit_sm	amos.models.Statsmodel		Logistic regression from statsmodels, with inference.
mixedlm		amos.models.Statsmodel	A linear mixed model from statsmodels, with inference.
mnlogit	amos.models.Statsmodel		Multinomial logistic regression from statsmodels, with inference.
naive_bayes	sklearn.naive_bayes.GaussianNB		Gaussian naive Bayes.
negative_binomial		amos.models.Statsmodel	Negative binomial regression of counts, with inference.
neural_network	sklearn.neural_network.MLPClassifier	sklearn.neural_network.MLPRegressor	A multi-layer perceptron (a simple neural network).
ols		amos.models.Statsmodel	Ordinary least squares regression from statsmodels, with inference.
ordinal_regression	amos.models.Statsmodel		Ordinal regression (ordered logit or probit) from statsmodels.
poisson		amos.models.Statsmodel	Poisson regression of counts (such as the number of arrests).
probit	amos.models.Statsmodel		Probit regression of two classes from statsmodels, with inference.
quantile_regression		amos.models.Statsmodel	Quantile regression from statsmodels, with inference.
random_forest	sklearn.ensemble.RandomForestClassifier	sklearn.ensemble.RandomForestRegressor	A random forest.
ridge		sklearn.linear_model.Ridge	Linear regression with a ridge (L2) penalty.
robust_regression		amos.models.Statsmodel	Robust linear regression from statsmodels, with inference.
svm	sklearn.svm.SVC	sklearn.svm.SVR	A support vector machine (with probabilities for classification).
tabpfn	tabpfn.TabPFNClassifier	tabpfn.TabPFNRegressor	TabPFN, a pretrained model that is often the most accurate on small data.
wls		amos.models.Statsmodel	Weighted least squares regression from statsmodels, with inference.
xgboost	xgboost.XGBClassifier	xgboost.XGBRegressor	XGBoost gradient boosting.
zero_inflated_poisson		amos.models.Statsmodel	Zero-inflated Poisson regression of counts, with inference.

Validators (analyst)

Name	Tool	Description
group_k_fold	sklearn.model_selection.GroupKFold	Folds that keep all of the rows of each group in the same fold.
group_shuffle_split	sklearn.model_selection.GroupShuffleSplit	Repeated random splits of the groups (not the rows) into two sets.
k_fold	sklearn.model_selection.KFold	Divides the rows into folds ("n_splits", 5 by default) at random.
leave_one_group_out	sklearn.model_selection.LeaveOneGroupOut	Scores each group with a copy of the model fitted to the other groups.
leave_one_row_out	sklearn.model_selection.LeaveOneOut	Predicts each row with a copy of the model fitted to every other row.
repeated_k_fold	sklearn.model_selection.RepeatedKFold	k_fold repeated with different random folds ("n_repeats" times).
repeated_stratified_k_fold	sklearn.model_selection.RepeatedStratifiedKFold	stratified_k_fold repeated with different random folds.
shuffle_split	sklearn.model_selection.ShuffleSplit	Repeated random splits of the rows into two sets (Monte Carlo).
stratified_group_k_fold	sklearn.model_selection.StratifiedGroupKFold	Folds that keep each group together and the classes in proportion.
stratified_k_fold	sklearn.model_selection.StratifiedKFold	Folds that keep the share of each class the same in every fold.
stratified_shuffle_split	sklearn.model_selection.StratifiedShuffleSplit	Repeated random splits into two sets, with the classes in proportion.
time_series_split	sklearn.model_selection.TimeSeriesSplit	Scores later rows with copies of the model fitted to earlier rows.

Metrics (critic)

Name	Tool	Task	Better	Description
accuracy	sklearn.metrics.accuracy_score	classify	higher	The share of rows classified correctly.
average_precision	sklearn.metrics.average_precision_score	classify	higher	The area under the precision-recall curve.
balanced_accuracy	sklearn.metrics.balanced_accuracy_score	classify	higher	The average share of each class classified correctly.
brier	sklearn.metrics.brier_score_loss	classify	lower	The mean squared error of the predicted probabilities (lower is better).
cohen_kappa	sklearn.metrics.cohen_kappa_score	classify	higher	Agreement between the predictions and labels beyond chance.
concordance	amos.metrics.concordance_index	regress	higher	How often the model orders pairs of times correctly (Harrell's C).
demographic_parity	fairlearn.metrics.demographic_parity_difference	classify	lower	The largest gap between groups in the share predicted to be positive.
demographic_parity_ratio	fairlearn.metrics.demographic_parity_ratio	classify	higher	The smallest group's share predicted positive over the largest's.
equal_opportunity	fairlearn.metrics.equal_opportunity_difference	classify	lower	The largest gap between groups in the true positive rate.
equal_opportunity_ratio	fairlearn.metrics.equal_opportunity_ratio	classify	higher	The smallest group's true positive rate over the largest's.
equalized_odds	fairlearn.metrics.equalized_odds_difference	classify	lower	The larger gap between groups in true or false positive rates.
equalized_odds_ratio	fairlearn.metrics.equalized_odds_ratio	classify	higher	The smaller ratio between groups of true or false positive rates.
explained_variance	sklearn.metrics.explained_variance_score	regress	higher	The share of the variance of the label that the model explains.
f1	sklearn.metrics.f1_score	classify	higher	The harmonic mean of precision and recall.
log_loss	sklearn.metrics.log_loss	classify	lower	The negative log-likelihood of the true labels (lower is better).
mae	sklearn.metrics.mean_absolute_error	regress	lower	The mean absolute error (lower is better).
mape	sklearn.metrics.mean_absolute_percentage_error	regress	lower	The mean absolute percentage error (lower is better).
matthews	sklearn.metrics.matthews_corrcoef	classify	higher	The Matthews correlation coefficient (phi for two classes).
mse	sklearn.metrics.mean_squared_error	regress	lower	The mean squared error (lower is better).
precision	sklearn.metrics.precision_score	classify	higher	The share of rows predicted to be positive that are positive.
r2	sklearn.metrics.r2_score	regress	higher	The coefficient of determination (R squared).
recall	sklearn.metrics.recall_score	classify	higher	The share of positive rows that are predicted to be positive.
rmse	sklearn.metrics.root_mean_squared_error	regress	lower	The root mean squared error (lower is better).
roc_auc	sklearn.metrics.roc_auc_score	classify	higher	The area under the receiver operating characteristic (ROC) curve.

Evaluators (critic)

Name	Description
classification_report	Precision, recall, f1, and the number of rows of each class.
conformal	Prediction intervals (or sets) with a known rate of coverage.
confusion	How many rows of each class were predicted to be each class.
explain_weights	eli5's explanation of the weights of the model's features.
factor_analysis	The hidden factors that explain the correlations of the features.
fairness	How the model does for each group, and the gaps between groups.
feature_importance	The importance that the model itself gives each feature.
pca	How much of the variance of the features each principal component has.
permutation_importance	How much the model's score drops when each feature is shuffled.
scorecard	The results of every branch of an analysis, ready to publish.
shap_importance	The mean absolute SHAP value of each feature.

Inferers (critic)

Name	Tool	Description
difference_in_differences	doubleml.did.DoubleMLDIDMulti	DoubleML's difference-in-differences: a treatment's effect over time.
distance_matching	dowhy.causal_estimators.distance_matching_estimator.DistanceMatchingEstimator	DoWhy's matching of treated rows with the most similar untreated rows.
doubly_robust	dowhy.causal_estimators.doubly_robust_estimator.DoublyRobustEstimator	DoWhy's doubly robust estimator (augmented inverse propensity weighting).
dr_learner	causalml.inference.meta.BaseDRRegressor	causalml's doubly robust learner of the effect of a treatment.
glm_adjustment	dowhy.causal_estimators.generalized_linear_model_estimator.GeneralizedLinearModelEstimator	DoWhy's generalized linear model of the label, adjusted for the features.
instrumental_variable	dowhy.causal_estimators.instrumental_variable_estimator.InstrumentalVariableEstimator	DoWhy's estimate of an effect through an instrument.
interactive_iv	doubleml.DoubleMLIIVM	DoubleML's effect for those whom an instrument moves to take a treatment.
interactive_regression	doubleml.DoubleMLIRM	DoubleML's average effect of a treatment with two values.
lpcmci	tigramite.lpcmci.LPCMCI	tigramite's LPCMCI: causal discovery in time series with hidden causes.
partially_linear	doubleml.DoubleMLPLR	DoubleML's effect of a treatment that adds the same amount for every row.
partially_linear_iv	doubleml.DoubleMLPLIV	DoubleML's partially linear model of a treatment with an instrument.
partially_linear_panel	doubleml.DoubleMLPLPR	DoubleML's partially linear model of a panel: units seen over time.
partially_logistic	doubleml.DoubleMLLPLR	DoubleML's partially linear model of a label with two classes.
pcmci	tigramite.pcmci.PCMCI	tigramite's PCMCI: which variables cause which others at later times.
pcmci_plus	tigramite.pcmci.PCMCI	tigramite's PCMCI+: causes at later times and at the same time.
potential_outcomes	doubleml.DoubleMLAPOS	DoubleML's effects of each level of a treatment, compared to one of them.
propensity_matching	dowhy.causal_estimators.propensity_score_matching_estimator.PropensityScoreMatchingEstimator	DoWhy's matching of treated and untreated rows by propensity score.
propensity_stratification	dowhy.causal_estimators.propensity_score_stratification_estimator.PropensityScoreStratificationEstimator	DoWhy's comparison of the treated and untreated in strata of propensity.
propensity_weighting	dowhy.causal_estimators.propensity_score_weighting_estimator.PropensityScoreWeightingEstimator	DoWhy's inverse propensity weighting.
quantile_effects	doubleml.DoubleMLQTE	DoubleML's effects of a treatment on quantiles of the label.
regression_adjustment	dowhy.causal_estimators.linear_regression_estimator.LinearRegressionEstimator	DoWhy's linear regression of the label, which adjusts for the features.
regression_discontinuity	doubleml.rdd.RDFlex	DoubleML's regression discontinuity: the effect at a cutoff.
s_learner	causalml.inference.meta.BaseSRegressor	causalml's S-learner: one model of the label, with the treatment in it.
sample_selection	doubleml.DoubleMLSSM	DoubleML's effect of a treatment when the label is only sometimes seen.
t_learner	causalml.inference.meta.BaseTRegressor	causalml's T-learner: one model of the label for each group.
time_series_effect	tigramite.models.LinearMediation	tigramite's causal effect of one series (the treatment) on the label.
tmle	causalml.inference.meta.TMLELearner	causalml's targeted maximum likelihood estimation of an average effect.
x_learner	causalml.inference.meta.BaseXRegressor	causalml's X-learner, which suits a treatment that few rows have.

Plots (artist)

Name	Description
actual_vs_predicted	The label against the model's predictions (regression).
box_plots	Each numeric feature as a box, split by class.
calibration_curve	The share of rows in each class against its predicted probability.
coefficient_plot	Coefficients with their confidence intervals (a forest plot).
confusion_heatmap	The confusion matrix as a heatmap (classification).
correlation_heatmap	The correlations between the numeric columns as a heatmap.
count_plots	The count of each value of each categorical feature, split by class.
det_curve	The detection error tradeoff (DET) curve (classification).
fairness_plot	The model's fairness metrics for each group, as bars.
histograms	The distribution of each numeric feature, in a grid.
importance_plot	The most important features, as horizontal bars.
influence_plot	The influence of each training row on a statsmodels regression.
label_plot	The distribution of the label: a bar for each class, or a histogram.
learning_curve	The model's score as it learns from more of the training rows.
missing_heatmap	Where values are missing, as a heatmap of the rows and columns.
pair_plot	Each pair of numeric features against each other, colored by class.
partial_dependence	The model's average prediction as features change (partial dependence).
precision_recall_curve	Precision against recall at every threshold (classification).
prediction_intervals	Conformal prediction intervals of the test rows (or the set sizes).
qq_plot	The quantiles of the residuals against a normal distribution's.
residual_plot	The model's errors against its predictions (regression).
roc_curve	The receiver operating characteristic (ROC) curve (classification).
search_plot	The cross-validated score of each try in a hyperparameter search.
shap_bar	The mean absolute SHAP value of the most important features.
shap_beeswarm	The SHAP value of each row for the most important features.
shap_decision	How the features move each row's prediction, as lines.
shap_embedding	The rows placed by their SHAP values, colored by one feature's.
shap_force	How each feature pushes one row's prediction up or down.
shap_group_difference	How the SHAP values of the features differ between two groups.
shap_heatmap	The SHAP values of every explained row, as a heatmap.
shap_partial_dependence	The model's prediction as one feature changes, with each row's.
shap_scatter	A feature's SHAP values against its values (a dependence plot).
shap_violin	The distribution of the SHAP values of the most important features.
shap_waterfall	How each feature moves one row's prediction (a waterfall plot).
shape_functions	An explainable boosting model's contribution from each feature.
survival_curves	The share of rows without an event over time (Kaplan-Meier curves).
tree_plot	The splits of a decision tree (or of one tree of a forest).
validation_curve	The model's score as one of its parameters changes.