Paper Session 9
AbstractAsk a Question?
Towards Fair Deep Anomaly Detection
Anomaly detection aims to find instances that are considered unusual and is a fundamental problem of data science. Recently, deep anomaly detection methods were shown to achieve superior results particularly in complex data such as images. Our work focuses on deep one-class classification for anomaly detection which learns a mapping only from the normal samples. However, the non-linear transformation performed by deep learning can potentially find patterns associated with social bias. The challenge with adding fairness to deep anomaly detection is to ensure both making fair and correct anomaly predictions simultaneously. In this paper, we propose a new architecture for the fair anomaly detection approach (Deep Fair SVDD) and train it using an adversarial network to decorrelate the relationships between the sensitive attributes and the learned representations. This differs from how fairness is typically added namely as a regularizer or a constraint. Further, we propose two effective fairness measures and empirically demonstrate that existing deep anomaly detection methods are unfair. We show that our proposed approach can remove the unfairness largely with minimal loss on the anomaly detection performance. Lastly, we conduct an in-depth analysis to show the strength and limitations of our proposed model, including parameter analysis, feature visualization, and run-time analysis.
Evaluating Fairness of Machine Learning Models Under Uncertain and Incomplete Information
Training and evaluation of fair classifiers is a challenging problem. This is partly due to the fact that most fairness metrics of interest depend on both the sensitive attribute information and label information of the data points. In many scenarios it is not possible to collect large datasets with such information. An alternate approach that is commonly used is to separately train an attribute classifier on data with sensitive attribute information, and then use it later in the ML pipeline to evaluate the bias of a given classifier. While such decoupling helps alleviate the problem of demographic scarcity, it raises several natural questions such as: how should the attribute classifier be trained?, and how should one use a given attribute classifier for accurate bias estimation? In this work we study this question from both theoretical and empirical perspectives. We first experimentally demonstrate that the test accuracy of the attribute classifier is not always correlated with its effectiveness in bias estimation for a downstream model. In order to further investigate this phenomenon, we analyze an idealized theoretical model and characterize the structure of the optimal classifier. Our analysis has surprising and counter-intuitive implications where in certain regimes one might want to distribute the error of the attribute classifier as unevenly as possible among the different subgroups. Based on our analysis we develop heuristics for both training and using attribute classifiers for bias estimation in the data scarce regime. We empirically demonstrate the effectiveness of our approach on real and simulated data.
Detecting Discriminatory Risk through Data Annotation based on Bayesian Inferences
Thanks to the increasing growth of computational power and data availability, the research in machine learning has advanced with tremendous rapidity. Nowadays, the majority of automatic decision making systems are based on data. However, it is well known that machine learning systems can present problematic results if they are built on partial or incomplete data. In fact, in recent years several studies have found a convergence of issues related to the ethics and transparency of these systems in the process of data collection and how they are recorded. Although the process of rigorous data collection and analysis is fundamental in the model design, this step is still largely overlooked by the machine learning community. For this reason, we propose a method of data annotation based on Bayesian statistical inference that aims to warn about the risk of discriminatory results of a given data set. In particular, our method aims to deepen knowledge and promote awareness about the sampling practices employed to create the training set, highlighting that the probability of success or failure conditioned to a minority membership is given by the structure of the data available. We empirically test our system on three datasets commonly accessed by the machine learning community and we investigate the risk of racial discrimination.