Flexible tool for bias detection, visualization, and mitigation. Uses models explained with DALEX and calculates fairness metrics based on confusion matrix for protected group. Allows to compare and gain information about various machine learning models. Mitigate bias with various pre-processing and post-processing techniques. Make sure your models are classifying protected groups similarly.
Install it from CRAN:
or developer version from GitHub:
Checking fairness is easy!
library(fairmodels) library(ranger) library(DALEX) data("german") # ------------ step 1 - create model(s) ----------------- lm_model <- glm(Risk~., data = german, family=binomial(link="logit")) rf_model <- ranger(Risk ~., data = german, probability = TRUE, num.trees = 200) # ------------ step 2 - create explainer(s) ------------ # numeric y for explain function y_numeric <- as.numeric(german$Risk) -1 explainer_lm <- explain(lm_model, data = german[,-1], y = y_numeric) explainer_rf <- explain(rf_model, data = german[,-1], y = y_numeric) # ------------ step 3 - fairness check ----------------- fobject <- fairness_check(explainer_lm, explainer_rf, protected = german$Sex, privileged = "male") print(fobject) plot(fobject)
fairness_check parameters are
* x, … -
fairness_objects (products of fairness_check).
* protected - factor with different subgroups as levels. Usually specific race, sex etc…
* privileged - subgroup, base on which to calculate parity loss metrics.
* cutoff - custom cutoff, might be single value - cutoff same for all subgroups or vector - for each subgroup individually. Affecting only explainers.
* label - character vector for every explainer.
Models might be trained on different data, even without protected variable. May have different cutoffs which gives different values of metrics.
fairness_check() is place where
fairness_objects are checked for compatibility and then glued together.
So it is possible to to something like this:
fairness_object <- fairness_check(explainer1, explainer2, ...) fairness_object <- fairness_check(explainer3, explainer4, fairness_object, ...)
even with more
If one is even more keen to know how
fairmodels works and what are relations between objects, please look at this diagram class diagram
There are 12 metrics based on confusion matrix :
|Metric||Formula||Full name||fairness names while checking among subgroups|
|TPR||true positive rate||equal opportunity|
|TNR||true negative rate|
|PPV||positive predictive value||predictive parity|
|NPV||negative predictive value|
|FNR||false negative rate|
|FPR||false positive rate||predictive equality|
|FDR||false discovery rate|
|FOR||false omission rate|
|STP||statistical parity||statistical parity|
|ACC||accuracy||Overall accuracy equality|
and their parity loss.
How is parity loss calculated?
i denotes the membership to unique subgroup from protected variable. Unprivileged subgroups are represented by small letters and privileged by simply “privileged”.
some fairness metrics like Equalized odds are satisfied if parity loss in both TPR and FPR is low