dowhy.causal_refuters.overrule.BCS package
Submodules
dowhy.causal_refuters.overrule.BCS.beam_search module
Beam search utilities for optimization.
This module implements the boolean ruleset estimator from OverRule [1]. Code is adapted (with some simplifications) from https://github.com/clinicalml/overlap-code, under the MIT License.
[1] Oberst, M., Johansson, F., Wei, D., Gao, T., Brat, G., Sontag, D., & Varshney, K. (2020). Characterization of Overlap in Observational Studies. In S. Chiappa & R. Calandra (Eds.), Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics (Vol. 108, pp. 788–798). PMLR. https://arxiv.org/abs/1907.04138
- class dowhy.causal_refuters.overrule.BCS.beam_search.PricingInstance(rp, rn, Xp, Xn, v0, z0)[source]
Bases:
object
Instance of the pricing problem.
For more details, see:
Dash, S., Gunluk, O., and Wei, D. (2018). Boolean decision rules via column generation. In Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa- Bianchi, N., and Garnett, R., editors, Advances in Neural Information Processing Systems 31, pages 4660–4670. Curran Associates, Inc.
- dowhy.causal_refuters.overrule.BCS.beam_search.beam_search(r, X, lambda0: float, lambda1: float, K: int = 1, UB: float = 0, D: int = 10, B: int = 5, wLB: float = 0.5, eps: float = 1e-06)[source]
Beam search to generate solutions to pricing problem.
- Parameters:
r – Cost vector (residuals)
X – Binary features in a DataFrame
lambda0 (float) – Fixed cost of a term
lambda1 (float) – Cost per literal
K (int, optional) – Maximum number of solutions returned, defaults to 1
UB (float, optional) – Initial upper bound on value of solutions, defaults to 0
D (int, optional) – Maximum Degree, defaults to 10
B (int, optional) – Beam width, defaults to 5
wLB (float, optional) – Weight on lower bound in evaluating nodes, defaults to 0.5
eps (float, optional) – Numerical tolerance on comparisons, defaults to 1e-6
dowhy.causal_refuters.overrule.BCS.load_process_data_BCS module
Code for Binarizing Features.
This module implements the boolean ruleset estimator from OverRule [1]. Code is adapted (with some simplifications) from https://github.com/clinicalml/overlap-code, under the MIT License.
[1] Oberst, M., Johansson, F., Wei, D., Gao, T., Brat, G., Sontag, D., & Varshney, K. (2020). Characterization of Overlap in Observational Studies. In S. Chiappa & R. Calandra (Eds.), Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics (Vol. 108, pp. 788–798). PMLR. https://arxiv.org/abs/1907.04138
- class dowhy.causal_refuters.overrule.BCS.load_process_data_BCS.FeatureBinarizer(colCateg: List[str] = [], numThresh: int = 9, negations: bool = False, threshStr: bool = False, threshOverride: Dict = {}, **kwargs)[source]
Bases:
TransformerMixin
Transformer for binarizing categorical and ordinal (including continuous) features.
Note that all features are converted into binary variables before learning Boolean rules.
Initialize transformer for binarizing categorical and ordinal (including continuous) features
- Parameters:
colCateg (List[str], optional) – List of categorical columns, defaults to [], ‘object’ dtype automatically treated as categorical
numThresh (int, optional) – Number of quantile thresholds to binarize ordinal features, defaults to 9
negations (bool, optional) – Include negations, defaults to False
threshStr (bool, optional) – Convert thresholds to strings, defaults to False
threshOverride (Dict, optional) – Dictionary to override quantile thresholds, defaults to {}, formatted as {colname : np.linspace object} to define cuts
- fit(X)[source]
Fit to data, including the learning of thresholds where appropriate.
Sets the following internal variables: * maps = dictionary of mappings for unary/binary columns * enc = dictionary of OneHotEncoders for categorical columns * thresh = dictionary of lists of thresholds for ordinal columns * NaN = list of ordinal columns containing NaN values
- Parameters:
X (pd.DataFrame) – Original features as a Pandas Dataframe
dowhy.causal_refuters.overrule.BCS.overlap_boolean_rule module
OverlapBooleanRule.
This module implements the boolean ruleset estimator from OverRule [1]. Code is adapted (with some simplifications) from https://github.com/clinicalml/overlap-code, under the MIT License.
[1] Oberst, M., Johansson, F., Wei, D., Gao, T., Brat, G., Sontag, D., & Varshney, K. (2020). Characterization of Overlap in Observational Studies. In S. Chiappa & R. Calandra (Eds.), Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics (Vol. 108, pp. 788–798). PMLR. https://arxiv.org/abs/1907.04138
- class dowhy.causal_refuters.overrule.BCS.overlap_boolean_rule.OverlapBooleanRule(alpha=0.95, lambda0=0.01, lambda1=0.01, K=20, D=20, B=10, iterMax=10, eps=1e-06, silent=False, verbose=False, solver='ECOS', rounding='greedy_sweep')[source]
Bases:
object
Overlap Boolean Rule class in the style of scikit-learn
Learn Boolean Rules in Disjuntive Normal Form to describe the positive class.
- Parameters:
alpha (float, optional) – Fraction of the positive samples to ensure are included in the rules, defaults to 0.95
lambda0 (float, optional) – Regularization on the # of rules, defaults to 1e-2
lambda1 (float, optional) – Regularization on the # of literals, defaults to 1e-2
K (int, optional) – Maximum results returned during beam search, defaults to 20
D (int, optional) – Maximum extra rules per beam seach iteration, defaults to 20
B (int, optional) – Width of beam search, defaults to 10
iterMax (int, optional) – Maximum number of iterations of column generation, defaults to 10
eps (float, optional) – Numerical tolerance on comparisons, defaults to 1e-6
silent (bool) – Silence non-optimizer output, defaults to False
verbose (bool, optional) – Verbose optimizer output, defaults to False
solver (str, optional) – Linear programming solver used by CVXPY to solve the LP relaxation, defaults to ‘ECOS’
rounding (str, optional) – Strategy to perform rounding, either ‘greedy’ or ‘greedy_sweep’, defaults to ‘greedy_sweep’
- fit(X: DataFrame, y: Union[ndarray, DataFrame])[source]
Fit model to training data.
- Parameters:
X – Pandas DataFrame containing covariates
y – +1 for Overlap/Support (depending on rules being learned), 0 for non-overlap, and -1 for background samples. Should only contain (+1/0) for overlap rules, or (+1/-1) for learning support rules.
- greedy_round_(X: DataFrame, y: Union[ndarray, DataFrame], xi: float = 0.5, use_lp: bool = False)[source]
Round the rule coefficients to integer values.
For DNF, this starts with no conjunctions, and adds them greedily based on a cost, which penalizes (any) inclusion of negative samples, and rewards (new) inclusion of positive samples, and goes until it covers at least alpha fraction of positive samples.
- Parameters:
X – Pandas DataFrame containing covariates
y – +1 for Overlap/Support (depending on rules being learned), 0 for non-overlap, and -1 for background samples. Should only contain (+1/0) for overlap rules, or (+1/-1) for learning support rules.
xi – Reward for including positive samples, relative to cost (1) for including negative samples
use_lp – Restrict to those conjuctions where the LP coefficients are positive. Note that the LP makes a difference regardless, as we only consider the rules generated by column generation here.
- round_(X: DataFrame, y: Union[ndarray, DataFrame], scoring: str = 'greedy', xi=None, use_lp: bool = True)[source]
Round the rule coefficients to integer values via a greedy approach, either using a fixed reward (scoring=”greedy”) or optimizing the reward for including positive examples according to balanced accuracy on classifying positive vs negative samples (scoring=”greedy_sweep).
- Parameters:
X – Pandas DataFrame containing covariates
y – +1 for Overlap/Support (depending on rules being learned), 0 for non-overlap, and -1 for background samples. Should only contain (+1/0) for overlap rules, or (+1/-1) for learning support rules.
xi – Reward for including positive samples, relative to cost (1) for including negative samples. For scoring=”greedy”, should be a single value, or an array of values for scoring=”greedy_sweep”. For the latter, will default to np.logspace(np.log10(0.01), 0.5, 20).
use_lp – Restrict to those conjuctions where the LP coefficients are positive. Note that the LP makes a difference regardless, as we only consider the rules generated by column generation here.