FairCORELS, a modified version of CORELS to build fair and interpretable models
Project description
Welcome to FairCorels, a Python library for learning fair and interpretable models using the Certifiably Optimal RulE ListS (CORELS) algorithm!
FairCORELS uses Python, Numpy, GMP, and a C++ compiler. GMP (GNU Multiple Precision library) is not required, but it is highly recommended, as it improves performance. If it is not installed, fairCORELS will run slower.
The fairCORELS
module includes two classifier methods : FairCorelsClassifier
and FairCorelsBagging
.
The FairCorelsClassifier
class implements the fairCORELS method. The FairCorelsBagging
class provides a wrapper to perform the Bagging ensemble method using FairCorelsClassifier
as a base learner.
The currently supported fairness notions are : statistical parity, predictive parity, predictive equality, equal opportunity, equalized odds, and conditional use accuracy equality.
Detail of the classifiers' parameters :
FairCorelsClassifier :
Constructor arguments :

c
: float, optional (default=0.01) Regularization parameter. Higher values penalize longer rulelists. 
n_iter
: int, optional (default=1000) Maximum number of nodes (rulelists) to search before exiting. 
map_type
: str, optional (default="prefix") The type of prefix map to use. Supported maps are "none" for no map, "prefix" for a map that uses rule prefixes for keys, "captured" for a map with a prefix's captured vector as keys. 
policy
: str, optional (default="lower_bound") The search policy for traversing the tree (i.e. the criterion with which to order nodes in the queue). Supported criteria are "bfs", for breadthfirst search; "curious", which attempts to find the most promising node; "lower_bound" which is the objective function evaluated with that rulelist minus the default prediction error; "objective" for the objective function evaluated at that rulelist; and "dfs" for depthfirst search. 
verbosity
: list, optional (default=["rulelist"]) The verbosity levels required. A list of strings, it can contain any subset of ["rulelist", "rule", "label", "minor", "samples", "progress", "mine", "loud"]. "rulelist" prints the generated rulelist at the end.
 "rule" prints a summary of each rule generated.
 "label" prints a summary of the class labels.
 "minor" prints a summary of the minority bound.
 "samples" produces a complete dump of the rules, label, and/or minor data. You must also provide at least one of "rule", "label", or "minor" to specify which data you want to dump, or "loud" for all data. The "samples" option often spits out a lot of output.
 "progress" prints periodic messages as faircorels runs.
 "mine" prints debug information while mining rules, including each rule as it is generated.
 "loud" is the equivalent of ["progress", "label", "rule", "mine", "minor"].

ablation
: int, optional (default=0) Specifies addition parameters for the bounds used while searching. Accepted values are 0 (all bounds), 1 (no antecedent support bound), and 2 (no lookahead bound). 
max_card
: int, optional (default=2) Maximum cardinality allowed when mining rules. Can be any value greater than or equal to 1. For instance, a value of 2 would only allow rules that combine at most two features in their antecedents. 
min_support
: float, optional (default=0.01) The fraction of samples that a rule must capture in order to be used. 1 minus this value is also the maximum fraction of samples a rule can capture. Can be any value between 0.0 and 0.5. 
beta
: float, optional (default=0.0) The weight the unfairness in the objective function 
fairness
: int optional (default=1) The type of fairness metric used. 1 : statistical parity, 2 : predictive parity, 3 : predictive equality, 4 : equal opportunity, 5 : equalized odds, 6 : conditional use accuracy equality 
maj_pos
: int optional (default=1) The position of the rule that defined the majority group If not specified, all individuals not in minority group are in majority group Not used if maj_vect is used 
min_pos
: int optional (default=2) The position of the rule that defined the minority group Not used if min_vect is used 
maj_vect
: int list optional (default=[1]) List of integers in {0,1} indicating if instances belong to the majority group If not specified, this vector is computed using maj_pos 
min_vect
: int list optional (default=[1]) List of integers in {0,1} indicating if instances belong to the minority group If not specified, this vector is computed using min_pos 
mode
: int optional (default=3) Method used for the multiojective framework 1: weigted sum, 2: maximum fairness, 3: epsilonconstraint, 4: maximum accuracy 
useUnfairnessLB
: bool optional (default=False) Use the unfairness lower bound > For some metrics (Sp and EO), an improved filtering using CP is applied For some others, it is not implemented and a simple, less effective lower bound is computed and used 
epsilon
: float optional (default=0.05) max acceptable unfairness 
kbest
: int optional (default=1) Randomly use one of the k best objectives 
forbidSensAttr
: bool optional (default=False) Forbid the use of the sensitive and (if specified) unsensitive attributes in the produced rule lists WARNING : This attribute is not used anymore in the current implementation. Remove columns manually from training data. 
bfs_mode
: int optinal (default=0) BFS strategy 0: original CORELS, 1:FIFO, 2:objective_aware, 3:lower_bound, 4:random 
random_state
: int optional (default=42) Random seed for randomized search
Methods :
.fit(X, y, features=[], prediction_name="prediction", performRestarts=0, initNBNodes=1000, geomRReason=1.5, max_evals=1000000000, time_limit = None):
Method for training the classifier.

X
: arraylike, shape = [n_samples, n_features] The training input samples. All features must be binary, and the matrix is internally converted to dtype=np.uint8. 
y
: arrayline, shape = [n_samples] The target values for the training input. Must be binary. 
features
: list, optional(default=[]) A list of strings of length n_features. Specifies the names of each of the features. If an empty list is provided, the feature names are set to the default of ["feature1", "feature2"... ]. 
prediction_name
: string, optional(default="prediction") The name of the feature that is being predicted. 
max_evals
: int, maximum number of calls to evaluate_children (ie maximum number of nodes explored in the prefix tree) 
time_limit
: int, maximum number of seconds allowed for the model building Note that this specifies the CPU time and NOT THE WALLCLOCK TIME 
performRestarts
: int optional (default=0) 0 : Does not perform any restarts, 1 : Performs geometric restarts, 2 : Performs Luby restarts WARNING : Note that restarts are an oldfeature, not maintained anymore. 
initNBNodes
: int optional (default=1000) Initial limit for the number of nodes in the trie when performing restarts WARNING : Note that restarts are an oldfeature, not maintained anymore. 
geomRReason
: double optional (default=1.5) When performRestarts=True, geomRReason is the reason used for the geometric restart calculation WARNING : Note that restarts are an oldfeature, not maintained anymore.
.predict(X):
Method for predicting using the trained classifier.
X
: arraylike, shape = [n_samples, n_features] The training input samples. All features must be binary, and the matrix is internally converted to dtype=np.uint8. The features must be the same as those of the data used to train the model.
=> Returns : p
: array of shape = [n_samples] > The classifications of the input samples.
.predict_with_scores(X):
Method for predicting using the trained classifier.
X
: arraylike, shape = [n_samples, n_features] The training input samples. All features must be binary, and the matrix is internally converted to dtype=np.uint8. The features must be the same as those of the data used to train the model.
=> Returns : p
: array of shape = [[n_samples],[n_samples]].
The first array contains the classifications of the input samples.
The second array contains the associated confidence scores.
.score(X, y):
Method that scores the algorithm on the input samples X with the labels y. Alternatively, score the predictions X against the labels y (where X has been generated by predict
or something similar).

X
: arraylike, shape = [n_samples, n_features] OR shape = [n_samples] The input samples, or the sample predictions. All features must be binary. 
y
: arraylike, shape = [n_samples] The input labels. All labels must be binary.
=> Returns : a
: float
The accuracy, from 0.0 to 1.0, of the rulelist predictions
.get_params():
Method to get a list of all the model's parameters.
=> Returns : params
: dict
Dictionary of all parameters, with the names of the parameters as the keys
.set_params(params):
Method to set some of the model's parameters.
params
: Set of model parameters. Takes an arbitrary number of keyword parameters, all of which must be valid parameter names (i.e. must be included in those returned by get_params).
.save(fname):
Method to save the model to a file, using python's pickle module.
fname
: string File name to store the model in
.load(fname):
Method to load a model from a file, using python's pickle module.
fname
: string File name to load the model from
.rl(set_val=None):
Method to return or set the learned rulelist
set_val
: RuleList, optional Rulelist to set the model to
=> Returns : rl
: obj
The model's rulelist
.str():
Method to get a string representation of the rule list
=> Returns : rl
: str
The rule list
.repr():
Same behavior as the previous one.
.explain(anEx):
Method to explain a prediction (by providing the matching rule).
anEx
: arraylike, shape = [n_features] The input sample
=> Returns : list l
where
l[0]
is the instance's prediction
l[1]
is the implicant(s) that led to that decision
(both are strings  user friendly)
.explain_api(anEx):
Method to explain a prediction (by providing the matching rule) (shorter output).
anEx
: arraylike, shape = [n_features] The input sample
=> Returns : list l
where
l[0]
is the instance's prediction
l[1]
is the implicant(s) that led to that decision
(both are APIoriented  easy to use by a program)
.explain_long(anEx):
Method to explain a prediction (by providing the matching rule and all the previous unmatched implicants).
anEx
: arraylike, shape = [n_features] The input sample
=> Returns : list l
where
l[0]
is the instance's prediction
l[1]
is the implicant(s) that led to that decision
(both are strings  user friendly)
.explain_long_api(anEx):
Method to explain a prediction (by providing the matching rule and all the previous unmatched implicants) (shorter output).
anEx
: arraylike, shape = [n_features] The input sample
=> Returns : list l
where
l[0]
is the instance's prediction
l[1]
is the implicant(s) that led to that decision
(both are APIoriented  easy to use by a program)
FairCorelsBagging :
This class provides an easy wrapper for the Bagging method that uses the FairCorelsClassifier
class as underlying base learning. Hence, arguments directly passed to the FairCorelsClassifier
object will not be detailed again in this section .
Constructor arguments :
The training sets for the different base learners are automatically computed from the entire provided training set, using the provided parameters.

n_learners
: int, number of base learners of typeFairCorelsClassifier
that will be used 
sample_size
: size of one subsample (used to train exactly one base learner) 
Note that, as training sets will be automatically determined here, the training set
X
,y
, the featuresfeatures
, the textual prediction nameprediction_name
must be provided directly here (NOT IN THE .fit call) 
Other
FairCorelsClassifier
arguments for the base learners :c=0.01
,n_iter=10000
,map_type="prefix"
,policy="lower_bound"
,verbosity=["rulelist"]
,ablation=0
,max_card=2
,min_support=0.01
,beta=0.0
,fairness=1
,maj_pos=1
,min_pos=2
,maj_vect = np.empty(shape=(0))
,min_vect = np.empty(shape=(0))
,mode=4
,useUnfairnessLB=False
,epsilon=0.0
,kbest=1
,forbidSensAttr=False
,bfs_mode=0
,random_state=42
andbaggingVerbose=0
Methods :
.fit(performRestarts=0, initNBNodes=1000, geomRReason=1.5, max_evals=1000000000, time_limit = None, n_workers=1):
Method to train the base learners.

n_workers
: int, maximum number of threads allowed to parallelize the training of the different learners 
 all
FairCorelsClassifier
attributes used to create the base learners
 all
.predict(X):
Predict classifications of the input samples X. Uses majority vote as aggregation function.
X
: arraylike, shape = [n_samples, n_features] The training input samples. All features must be binary, and the matrix is internally converted to dtype=np.uint8. The features must be the same as those of the data used to train the model.
=> Returns : p
: array of shape = [n_samples].
The classifications of the input samples.
.score(X, y):
Score the algorithm on the input samples X with the labels y. Alternatively, score the predictions X against the labels y (where X has been generated by predict
or something similar).

X
: arraylike, shape = [n_samples, n_features] OR shape = [n_samples] The input samples, or the sample predictions. All features must be binary. 
y
: arraylike, shape = [n_samples] The input labels. All labels must be binary.
=> Returns : a
: float
The accuracy, from 0.0 to 1.0, of the rulelist predictions
.explain(anInst):
Explains a prediction (by its matching rules among majority base learners).
anInst
: arraylike, shape = [n_features] The input sample
=> Returns : d
: dictionnary {'implicants':i, 'prediction':p}
where i
is the list of implicants that led majority voters to their prediction and p
is the associated prediction
.explain_complete(anInst):
Explains a prediction (adds complete implications, including antecedents negation for all learners).
anInst
: arraylike, shape = [n_features] The input sample
=> Returns : d
: dictionnary {'implicants':i, 'prediction':p}
where i
is the list of implicants (and unmatched rules) that led majority voters to their prediction and p
is the associated prediction
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for faircorels1.0py3.9macosx10.9x86_64.egg
Algorithm  Hash digest  

SHA256  d9cac207cafc94ff8e05ceabf4eb87197680e2cb160b3908035ded3684836ade 

MD5  58a7f1d77f76fd5d88b5f862c289a70a 

BLAKE2b256  91290cb164d862786c33a0064653bdeb26b7c2f1a559369976b59cdea09c9199 