FairPAN — bringing fairness to neural networks
With the ongoing development of XAI methods, we — data scientists have to take more and more responsibility for the behaviour of our models. It means that we should no longer care only about what are the results and the level of their accuracy, but also how are they obtained. It leads us into the field of fairness, which is best to describe by an example.
Consider the idea of the algorithm that has to predict whether giving credit to a person is risky or not. It is learning on real data of giving credits which were biased against females (historical fact). In that case, the model learns this bias, which is not only included in the simple sex variable but also is hidden inside other variables. Fairness enables us to detect such bias and handles a few methods to fight it.
To learn more about fairness read: ‘Fairmodels: A Flexible Tool For Bias Detection, Visualization, And Mitigation’ by Jakub Wisniewski and Przemysław Biecek
However, everyone who worked with fairness knows that bias mitigation isn’t an easy task and improvement here costs the performance of the model which makes us stand in the situation of balancing the tradeoff between the quality of fairness and performance. Although, with fairpan you can reduce this cost to a minimum and still achieve astonishing results.
What is FairPAN?
In the beginning, we have to understand how GANs work:
Generative Adversarial Networks are two neural networks that learn together. The Generator has to generate new samples that are indistinguishable from original data and the adversarial has to distinguish if the observation is original or generated. The generator is punished whenever the adversarial makes the correct prediction. After such a process generator eventually learns how to make indistinguishable predictions and adversaries’ accuracy drops down to 50% when a model cannot distinguish the two classes. The idea of GANs was proposed in Generative Adversarial Nets, Ian Goodfellow.
Fair Predictive Adversarial Network is a neural network model which mimics GANs by subsetting generator with classifier (predictor) and adversarial has to predict the sensitive value (such as sex, race, etc) from the output of the predictor. This process eventually leads the classifier to make predictions with indistinguishable sensitive values.
When we transfer that statement into fairness metrics, it means that our model specialises in improving the Statistical Parity (STP) ratio metric. STP is so important for us because it describes the similarity of densities of privileged and discriminated classes.
This idea comes from blogs: Towards fairness in ML with adversarial networks, Stijn Tonk and Fairness in Machine Learning with PyTorch, Henk Griffoen however, our implementation in R offers slightly different solutions. And the exact idea behind using GANs for Fairness is described in Achieving Fairness through Adversarial Learning: an Application to Recidivism Prediction, Christina Wadsworth, Francesca Vera, Chris Piech.
Custom loss function
The crucial part of this model is the metric we use to engage the two models into a zero-sum game. This is captured by the following objective function:
So, it learns to minimize its prediction losses while maximizing that of the adversarial (due to lambda being positive and minimizing a negated loss is the same as maximizing it). The objective during the game is simpler for the adversarial: predict sex based on the income level predictions of the classifier. This is captured in the following objective function:
The adversarial does not care about the prediction accuracy of the classifier. It is only concerned with minimizing its prediction losses. Firstly we pretrain classifier and adversarial. Later we begin the proper PAN training with both networks: we train the adversarial, provide its loss to the classifier, and after that, we train the classifier. This method shall lead us to fair predictions of the FairPAN model.
Regular mitigation techniques tend to worsen the performance of the classifier a lot by decreasing accuracy for example, whereas FairPAN has no such drawback and worsening of the performance is small. Moreover, our package is very flexible because it enables you to provide your own neural networks, but also to create one with our functions. The outcomes are also created with the usage of DALEX and fairmodels, so one can use their methods and visualizations. Additionally, the workflow of the package is really simple and clean, because of multiple features available for users, such as preprocess function.
Install the developer version from GitHub:
devtools::install_github("ModelOriented/FairPAN", build_vignettes = TRUE)
The graph below represents how the workflow inside the package looks like. Firstly we have to provide data and use preprocess() which creates all sets needed for this package to work. One can also skip that step, however, it is not advisable to do so. Later we have to create a dataset_loader() which organises our data to be ready for torch usage. The next step is really flexible because we can choose whether we want to create our functions with the package openly via create_model() and pretrain_net() hidden inside pretrain() or we want to provide neural networks create on our own, which can be pretrained or not, depending on our needs. It is extremely powerful because we can provide some well known and pretrained classifiers. Later, we engage the fair_train() process which outcomes we can visualize by setting monitor to true and using plot_monitor(). Although we can finish the process at his spot, we can also analyse the outcomes a bit more with explain_pan() and use all DALEX functions on the returned explainer. This explainer can also be used to apply fairmodels::fairness_check() and other functions from this package.
Achieve fairness and save performance!
library(fairpan) adult <- fairmodels::adult # ------------------- step 1 - prepare data ----------------------- data <- preprocess(data = adult, target_name = "salary", sensitive_name = "sex", privileged = "Male", discriminated = "Female", drop_also = c("race"), sample = 0.02, train_size = 0.6, test_size = 0.4, validation_size = 0, seed = 7 ) dev <- "cpu" dsl <- dataset_loader(train_x = data$train_x, train_y = data$train_y, test_x = data$test_x, test_y = data$test_y, batch_size = 5, dev = dev ) # ------------ step 2 - create and pretrain models ---------------- models <- pretrain(clf_model = NULL, adv_model = NULL, clf_optimizer = NULL, trained = FALSE, train_x = data$train_x, train_y = data$train_y, sensitive_train = data$sensitive_train, sensitive_test = data$sensitive_test, batch_size = 5, partition = 0.6, neurons_clf = c(32, 32, 32), neurons_adv = c(32, 32, 32), dimension_clf = 2, dimension_adv = 1, learning_rate_clf = 0.001, learning_rate_adv = 0.001, n_ep_preclf = 10, n_ep_preadv = 10, dsl = dsl, dev = dev, verbose = TRUE, monitor = TRUE ) # --------------- step 3 - train for fairness -------------------- monitor <- fair_train(n_ep_pan = 17, dsl = dsl, clf_model = models$clf_model, adv_model = models$adv_model, clf_optimizer = models$clf_optimizer, adv_optimizer = models$adv_optimizer, dev = dev, sensitive_train = data$sensitive_train, sensitive_test = data$sensitive_test, batch_size = 5, learning_rate_adv = 0.001, learning_rate_clf = 0.001, lambda = 130, verbose = TRUE, monitor = TRUE ) # --------- step 4 - prepare outcomes and plot them -------------- plot_monitor(STP = monitor$STP, adversary_acc = monitor$adversary_acc, adversary_losses = monitor$adversary_losses, classifier_acc = monitor$classifier_acc) exp_clf <- explain_pan(y = data$test_y, model = models$clf_model, label = "PAN", data = data$data_test, data_scaled = data$data_scaled_test, batch_size = 5, dev = dev, verbose = TRUE ) fobject <- fairmodels::fairness_check(exp_PAN, protected = data$protected_test, privileged = "Male", verbose = TRUE) plot(fobject)
As a result, we can plot two graphs where one of them visualizes metrics during the training process and the second one compares the classifier to PAN models in terms of fairness metrics.
In our example, we can see a drastic improvement in the STP ratio and only a minor loss of models accuracy. The training process is also working correctly because adversarial loss increases and accuracy decreases over time.
The improvement of the Statistical parity ratio is also visible on the second plot. As one can notice, the Equal opportunity ratio has worsened a lot, however, it is not a surprise because in the fairtrain process we generate more True Positives which leads to the improvement of STP and worsening of the aforementioned metric.
For more sophisticated examples with outputs and rich descriptions visit FairPAN Tutorial.
The fairpan package in R is created to proceed with the neural network training process which takes fairness into consideration and mitigates bias visible in the data. It is achieved by creating a model based on the idea of GANs which minimalises the Statistical Parity ratio metric. It is better than other methods because losses in performance due to bias mitigation are much lower. Thefairpan package is also a great tool, because of its flexibility with model creation and data preparation.
If you are interested in other posts about explainable, fair, and responsible ML, follow #ResponsibleML on Medium.
In order to see more R related content visit https://www.r-bloggers.com