Abstract Computer vision systems are employed in a variety of high-impact applications. However, making them trustworthy requires methods for the detection of potential biases in their training data, before models learn to harm already disadvantaged groups in downstream applications. Image data are typically represented via extracted features, which can be hand-crafted or pre-trained neural network embeddings. In this work, we introduce a framework for bias discovery given such features that is based on optimal transport theory; it uses the (quadratic) Wasserstein distance to quantify disparity between the feature distributions of two demographic groups (e.g., women vs men). In this context, we show that the Kantorovich potentials of the images, which are a byproduct of computing the Wasserstein distance and act as “transportation prices", can serve as bias scores by indicating which images might exhibit distinct biased characteristics. We thus introduce a visual dataset exploration pipeline that helps auditors identify common characteristics across high- or low-scored images as potential sources of bias. We conduct a case study to identify prospective gender biases and demonstrate theoretically-derived properties with experiments on the CelebA and Biased MNIST datasets.

In this article the statement after Equation 1 had an error in the published version. Please refer the correction as follows: “where ν = T#µ and T# is the push-forward of µ along the function T : X → Y” was incorrectly written as “where T# is the push-forward of µ along the function T : X → Y. Furthermore, Equation 1 itself was incorrectly formulated. Namely, the integral should have been over X and not over X × Y. The original article has been corrected.

Personalized PageRank (PPR) is a popular scheme for scoring the relevance of network nodes to a set of seed ones through a random walk with restart process. Calculating the scores of all network nodes often involves the power method, which iterates the PPR formula until convergence to an empirically selected numerical tolerance. However, finding a tolerance that is not so lax as to impact pairwise node comparisons but not so strict as to require a high number of iterations to converge requires time-consuming empirical investigation. In this work we aim to avoid this investigation by stopping power method iterations when node score order is robust against subsequent changes. To do this, we analyse the expected fraction of random walks considered at a given iteration and identify a potential stopping point that depends on a (fixed) confidence level of future iterations preserving node order. Experiments on four real-world networks show that a confidence level of 98% runs in a fraction of the time and yields more than 0.999 Spearman correlation with the node order of 10-20numerical tolerance. Furthermore, that stopping point is comparable to empirically selecting a numerical tolerance that yields robust node order.

Local community detection is an emerging topic in network analysis that aims to detect well-connected communities encompassing sets of priorly known seed nodes. In this work, we explore the similar problem of ranking network nodes based on their relevance to the communities characterized by seed nodes. However, seed nodes may not be central enough or sufficiently many to produce high quality ranks. To solve this problem, we introduce a methodology we call seed oversampling, which first runs a node ranking algorithm to discover more nodes that belong to the community and then reruns the same ranking algorithm for the new seed nodes. We formally discuss why this process improves the quality of calculated community ranks if the original set of seed nodes is small and introduce a boosting scheme that iteratively repeats seed oversampling to further improve rank quality when certain ranking algorithm properties are met. Finally, we demonstrate the effectiveness of our methods in improving community relevance ranks given only a few random seed nodes of real-world network communities. In our experiments, boosted and simple seed oversampling yielded better rank quality than the previous neighborhood inflation heuristic, which adds the neighborhoods of original seed nodes to seeds.

Machine learning bias and fairness have recently emerged as key issues due to the pervasive deployment of data-driven decision making in a variety of sectors and services. It has often been argued that unfair classifications can be attributed to bias in training data, but previous attempts to 'repair' training data have led to limited success. To circumvent shortcomings prevalent in data repairing approaches, such as those that weight training samples of the sensitive group (e.g. gender, race, financial status) based on their misclassification error, we present a process that iteratively adapts training sample weights with a theoretically grounded model. This model addresses different kinds of bias to better achieve fairness objectives, such as trade-offs between accuracy and disparate impact elimination or disparate mistreatment elimination. We show that, compared to previous fairness-aware approaches, our methodology achieves better or similar trades-offs between accuracy and unfairness mitigation on real-world and synthetic datasets.