2016 IEEE 2nd International Conference on Collaboration and Internet Computing (CIC) (2016)
Pittsburgh, Pennsylvania, United States
Nov. 1, 2016 to Nov. 3, 2016
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/CIC.2016.051
Data mining techniques are playing an increasing role in making crucial decisions in our daily lives ranging from credit card approvals to employment decisions. Typically algorithms used to build decision models remain as a black-box to the end user. Therefore the process of decision making appears to be opaque. At the same time, increasing the transparency of a black-box decision making model allows us to discover hidden discrimination, and hold entities accountable. Although algorithmic transparency with respect to black box classifiers requires addressing many challenges, our main objective in this paper is to investigate whether a black-box classification model is biased against certain subgroups. Specifically, we study the indirect discrimination of hidden protected features. Protected features, such as race, gender, and religious beliefs, are those that are prohibited to be legally used for making decisions. Simply removing the protected features is not enough to eliminate discrimination because there could be strong correlations between protected features and non-protected features, such as race versus zip code. In this paper, we present two techniques to measure discrimination of a black-box model as a result of data bias or algorithmic weakness. Data bias is investigated further by introducing artificial bias to the dataset under consideration. Our experimental results demonstrate the effectiveness of our bias measures where bias comes from different sources.
Correlation, Data mining, Feature extraction, Data models, Itemsets, Decision making, Bayes methods
Y. Alufaisan, M. Kantarcioglu and Y. Zhou, "Detecting Discrimination in a Black-Box Classifier," 2016 IEEE 2nd International Conference on Collaboration and Internet Computing (CIC), Pittsburgh, Pennsylvania, United States, 2016, pp. 329-338.