In this tutorial you will get to know when, why and how to use fairmodels
. fairmodels
is a tool for bias detection, visualization and mitigation. It is compatible with DALEX
(Biecek (2018)) and DALEXtra
which are model agnostic explainers. Some knowledge of how to use those explainers will be needed but in this tutorial you should grasp the idea.
For this tutorial we will use compas
data (Angwin et al. (2016)) to see if someone will become recidivist in next 2 years.
Let’s say you are building algorithm for court system that predicts if someone will become recidivist in the future. First you gather information then you build a model and predict outcomes. You get satisfying accuracy. It is pretty good, but it appears that the model is more likely to say that African Americans will become recidivists. It was trained on data that where discrimination of certain ethnic groups was visible. Racist algorithms are no good for the court system. So now we have some options. First one is to gather more reliable data, and the second one is trying to understand bias and choose model with least amount of it. We will choose the second option.
We will use modified ProPublica’s compas data to represent our problem.
library(fairmodels)
data("compas")
head(compas)
For fairmodels
package to work properly we want to flip factor levels in target variable, so positive outcome (not being a recidivist) is being predicted by models. It is only needed for one specific function but more on it later.
compas$Two_yr_Recidivism <- as.factor(ifelse(compas$Two_yr_Recidivism == '1', '0', '1'))
We train a ranger
(Wright and Ziegler (2017)) model and create an explainer with DALEX
.
library(DALEX)
library(ranger)
# train
rf_compas <- ranger(Two_yr_Recidivism ~., data = compas, probability = TRUE)
# numeric target values
y_numeric <- as.numeric(compas$Two_yr_Recidivism)-1
# explainer
rf_explainer <- explain(rf_compas, data = compas[,-1], y = y_numeric, colorize = FALSE)
#> Preparation of a new explainer is initiated
#> -> model label : ranger ( default )
#> -> data : 6172 rows 6 cols
#> -> target variable : 6172 values
#> -> predict function : yhat.ranger will be used ( default )
#> -> predicted values : No value for predict function target column. ( default )
#> -> model_info : package ranger , ver. 0.13.1 , task classification ( default )
#> -> predicted values : numerical, min = 0.1590342 , mean = 0.5449543 , max = 0.8705104
#> -> residual function : difference between y and yhat ( default )
#> -> residuals : numerical, min = -0.8505374 , mean = -7.418443e-05 , max = 0.7780252
#> A new explainer has been created!
Than we create call function fairness_check()
This function aggregates many explainers so you may compare many models. We assign object to name fobject
which is short version of fairness_object
- object returned by fairness_check()
,
fobject <- fairness_check(rf_explainer, # explainer
protected = compas$Ethnicity, # protected variable as factor
privileged = "Caucasian", # level in protected variable, potentially more privileged
cutoff = 0.5, # cutoff - optional, default = 0.5
colorize = FALSE)
#> Creating fairness classification object
#> -> Privileged subgroup : character ([32m Ok [39m )
#> -> Protected variable : factor ([32m Ok [39m )
#> -> Cutoff values for explainers : 0.5 ( for all subgroups )
#> -> Fairness objects : 0 objects
#> -> Checking explainers : 1 in total ( [32m compatible [39m )
#> -> Metric calculation : 13/13 metrics calculated for all models
#> Fairness object created succesfully
Let’s see if our ranger model has bias.
print(fobject, colorize = FALSE)
#>
#> Fairness check for models: ranger
#>
#> ranger passes 2/5 metrics
#> Total loss : 3.7895
plot(fobject)
Of course the protected
parameter can be changed, for example to (here binary) compas$Sex
.
In many metrics ranger
exceeds fairness threshold (which can be changed by epsilon
parameter). If bars reach red field on the left it means that there is bias towards certain unprivileged subgroup. If they reach one on the right it means bias towards privileged (Caucasian - in all metrics here Caucasian subgroup is referenced as base - 1) subgroup. This values mean the proportion (ratio) of certain unprivileged subgroup to privileged subgroup (eg. Asian/Caucasian). By default it represents 80% rule (for eg. women should get credit at rate at least 80% of that of men) (Code of Federal Regulations (1978)) . Metrics measure how equal treatment and mistreatment among subgroups is. More on those metrics: wikipedia and in Verma and Rubin (2018)
Why do we have this bias? Model did learn from biased data. We can see it on plot below
plot_density(fobject)
As we can see it is more likely that model will categorize African Americans as being recidivists than for example Asians.
Sometimes it is worth it to see what the raw (unscaled) metrics are. User could use fobject$groups_data
but the easier way is to plot metric_scores
object.
plot(metric_scores(fobject))
To really see what fairness_object
is about, we need to make some more models and explainers.
library(gbm)
rf_compas_1 <- ranger(Two_yr_Recidivism ~Number_of_Priors+Age_Below_TwentyFive,
data = compas,
probability = TRUE)
lr_compas_1 <- glm(Two_yr_Recidivism~.,
data=compas,
family=binomial(link="logit"))
rf_compas_2 <- ranger(Two_yr_Recidivism ~., data = compas, probability = TRUE)
rf_compas_3 <- ranger(Two_yr_Recidivism ~ Age_Above_FourtyFive+Misdemeanor,
data = compas,
probability = TRUE)
df <- compas
df$Two_yr_Recidivism <- as.numeric(compas$Two_yr_Recidivism)-1
gbm_compas_1<- gbm(Two_yr_Recidivism~., data = df)
explainer_1 <- explain(rf_compas_1, data = compas[,-1], y = y_numeric)
explainer_2 <- explain(lr_compas_1, data = compas[,-1], y = y_numeric)
explainer_3 <- explain(rf_compas_2, data = compas[,-1], y = y_numeric, label = "ranger_2")
explainer_4 <- explain(rf_compas_3, data = compas[,-1], y = y_numeric, label = "ranger_3")
explainer_5 <- explain(gbm_compas_1, data = compas[,-1], y = y_numeric)
Now we create one object with all explainers
fobject <- fairness_check(explainer_1, explainer_2,
explainer_3, explainer_4,
explainer_5,
protected = compas$Ethnicity,
privileged = "Caucasian",
verbose = FALSE)
As we can see there is some parameters in fairness_check such as:
1. x, … - list of DALEX
explainers, and other fairness_object
objects
2. protected - factor, containing subgroups as levels. Protected stands for protected variable (or sensitive attribute)
3. privileged - character, level in protected, it is subgroup suspected of having better results
4. cutoff - numeric, vector of cutoffs values matching the order of levels in protected variable. It affects only explainers so if fairness_object
is passed it’s cutoff vector won’t be changed.
5. label - character, vector of labels for explainers only. Very convenient for fairness_check()
iterative approach - having explainer, checking for bias, mitigating bias, passing both explainer and fairness object and comparing fairness.
6. epsilon - numeric, boundary position in fairness_check()
. Fairness metrics are satisfied if parity loss values are between (-epsilon, epsilon)
fairness_object
is output value of fairness_check()
It is S3 object consisting of: * Parity loss metrics Popular confusion matrix metrics with parity loss - sum of absolute logarithms of ratios of unprivileged subgroups and privileged one (more on that in documentation ?fairness_check
. If model would have 0 in certain parity loss metric it would mean that it treats all subgroups equally.
fobject$parity_loss_metric_data
Fairness object gets metrics based on confusion matrix and checks them over the groups.
# for the first model
fobject$groups_data$ranger$TPR
#> African_American Asian Caucasian Hispanic
#> 0.6697490 0.8695652 0.8149883 0.8406250
#> Native_American Other
#> 0.8333333 0.8858447
It is simply metrics for certain subgroup.
What is relation between $groups_data
and $parity_loss_metric_data
?
If we were going only to take score from certain metric (Let’s say fpr and 0.3) we wouldn’t know if it is good or bad. But we are aiming for equal treatment over all groups so if this metric score would be the same in all groups it would be very good. But the metrics wouldn’t be comparable between each others (fpr - 0.3 in all groups and accuracy - 0.9 in all groups, both are good in terms of parity). That is why we use privileged - to set benchmark. And for example Caucasian in fpr had score of 0.3 and African American 0.6. After setting privilieged = Caucasian
Caucasian would have score 0, and African American log(0.6/0.3)
. It enables comparing and adding scores.
Note: When dealing with plots we use formula sum(abs(log(Mu/Mp))) for all unprivileged metrics (Mu - metric unprivileged, Mp - metric privileged) to represent aggregated score in subgroups. In short is how much it differs from ideal scores. The closer to zero the better
group_confusion_matrices Similarly to groups data it stores confusion matrix for each subgroup
explainers
list of DALEX
explainers
# for first model
fobject$cutoff$ranger
#> $African_American
#> [1] 0.5
#>
#> $Asian
#> [1] 0.5
#>
#> $Caucasian
#> [1] 0.5
#>
#> $Hispanic
#> [1] 0.5
#>
#> $Native_American
#> [1] 0.5
#>
#> $Other
#> [1] 0.5
list of cutoff values for each model
fairness_check_data data used in print and plot of fairness_object
. It is already processed data and ready to plot. If someone were to use abs()
metrics there would be equal to particular metrics in $parity_loss_metric_data
. It means that it allows negative values. So when value is negative it means that score of privileged group in this metric was better.
… - other parameters passed to fairness_check()
We now have a few models in our fairness_object
Let’s see how they perform in different metrics.
sm <- stack_metrics(fobject)
plot(sm)
It displays accumulated (Stacked) metric scores for each model. The least metric score the better.
cm <- choose_metric(fobject, "TPR")
plot(cm)
With this task we should use PCA
. We call create_fairness_pca()
to create fairness pca object.
fair_pca <- fairness_pca(fobject)
print(fair_pca)
#> Fairness PCA :
#> PC1 PC2 PC3 PC4 PC5
#> [1,] 0.5812848 1.4182294 0.5735680 -0.35552270 -7.806256e-18
#> [2,] -2.5382280 -0.5612499 0.7226884 0.28522944 4.440892e-16
#> [3,] 1.0757097 1.2296514 -0.5833416 0.37621550 8.604228e-16
#> [4,] 3.1715509 -1.5289220 0.0907467 -0.03198912 5.551115e-16
#> [5,] -2.2903174 -0.5577090 -0.8036615 -0.27393312 -5.551115e-16
#>
#> Created with:
#> [1] "ranger" "lm" "ranger_2" "ranger_3" "gbm"
#>
#> First two components explained 93 % of variance.
Let’s plot!
plot(fair_pca)
Another way to deal with grouped data is using heatmap.
fheatmap <- fairness_heatmap(fobject)
plot(fheatmap, text_size = 3)
For both models and metrics dendograms are created. This way through hierarchical clustering we can look on similarities between models/metrics. It should give similar but more detailed information than PCA Now we know what those scores are and how “similar” models are to each other
Sometimes we would like to know how good are models in performance metrics and in fairness metrics at the same time, to see the tradeoff between them.
fap <- performance_and_fairness(fobject, fairness_metric = "STP")
#> Performace metric is NULL, setting deafult ( accuracy )
#>
#> Creating object with:
#> Fairness metric: STP
#> Performance metric: accuracy
plot(fap)
When we have narrowed down our search for the best model we can use group_metric
to check once again metrics within groups and decide which model to use.
fobject2 <- fairness_check(explainer_1,explainer_2,
protected = compas$Ethnicity,
privileged = "Caucasian",
verbose = FALSE)
gm <- group_metric(fobject2, fairness_metric = "FPR")
#> Performace metric not given, setting deafult ( accuracy )
#>
#> Creating object with:
#> Fairness metric: FPR
#> Performance metric: accuracy
plot(gm)
fradar <- fairness_radar(fobject2)
plot(fradar)
We may see how cutoff affects parity loss of metrics
All cutoffs measures where metrics exist (are not NA) and how they change if we modify cutoffs in all subgroups.
In this plot NA
values are natural, so warnings are to be expected.
ac <- all_cutoffs(fobject2)
plot(ac)
This function shows how parity loss metrics would change if we modified cutoff only for one subgroup (here African American) with other cutoffs fixed.
cpc <- ceteris_paribus_cutoff(fobject2, subgroup = "African_American")
plot(cpc)
fairmodels
is powerful and flexible tool for detecting bias and asserting fairness in models. Many metrics are available (there are tools to create own metrics based on confusion matrix). Besides bias detection fairmodels
offers also bias mitigation. It is set of pre-processing and post-processing algorithms working either on data or on explainer which function is to lower parity loss in certain metrics. If you are interested in this and adult data case study, please check Advanced Tutorial. If you encountered a bug or you have cool idea for a new feature please write and issue here.