Auditing Black-box Models by Obscuring Features

By Philip Adler, Casey Falk, Sorelle A. Friedler, Gabriel Rybeck, Carlos Eduardo Scheidegger, Brandon Smith, Suresh Venkatasubramanian. 2016

26 citations recorded by the source corpus.

Subfields: Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Safety, Robustness & Interpretability

Citations in the corpus

Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations (2017, background): “, 2016], or seeing whether explanations and explanation constraints can further the goals of fairness, accountability, and transparency in machine learning (either by detecting indirect influence [Adler et al., 2016] or by constraining models to avoid it [Dwork et al.”
Iterative Orthogonal Feature Projection for Diagnosing Bias in Black-Box Models (2016, background): “In general, these works can be broadly classified into 3 broad categories: data transformation, algorithm manipulation, and outcome manipulation methodologies (Zemel et al., 2013; Adler et al., 2016).”
A Causal Framework for Discovering and Removing Direct and Indirect Discrimination (2016, methodology): “l massaging (LMSG) and local preferential sampling (LPS) algorithms proposed in (Zliobaitˇ ˙e, Kamiran, and Calders 2011) and disparate impact removal algorithm (DI) proposed in (Feldman et al. 2015; Adler et al. 2016). The causal networks are constructed and presented by utilizing an open-source so”
Fairness in Learning: Classic and Contextual Bandits (2016, background): “[11, 23, 18, 15, 16] and [2] for a study of auditing existing algorithms for disparate impact).”
The ethics of algorithms: Mapping the debate (2016, neutral): “Further work is required to design broadly applicable, low impact auditing mechanisms for algorithms (cf. Adler et al., 2016; Sandvig et al., 2014) that build upon current work in transparency and interpretability of machine learning (e.g. Kim et al., 2015; Lou et al., 2013).”
Explaining Explanations in AI (2018, methodology): “…seek to explain how a fixed model leads to a particular prediction, either by fitting a simpler, local model around a particular decision (Ribeiro et al., 2016), or by perturbing variables to measure how the prediction changes (Adler et al., 2016; Datta et al., 2016; Simonyan et al., 2013).”
On Ensuring that Intelligent Machines Are Well-Behaved (2017, background): “…proposed means for determining how sensitive a black-box supervised learning algorithm is to each feature in a vector of features used to represent the input [Datta et al., 2016; Adler et al., 2016], which can be used to determine the impact of features like race and gender [Datta et al., 2015].”
Posthoc Interpretability of Learning to Rank Models using Secondary Training Data (2018, methodology): “We propose to use a strategy followed by [1, 6, 8] and other model agnostic approaches where we exploit the blackbox to generate labels to train a new interpretable model.”
Identifying Significant Predictive Bias in Classifiers (2016, background): “Topically, the problem of assessing bias in predictive models covers areas including predictive bias, problems in the original training data, disparate impacts of predictions, and adjusting predictions to ensure fairness [1,6,9,17] — in areas like criminal justice [2,5,10,19], but also other various”
Does mitigating ML's impact disparity require treatment disparity? (2017, background): “, 2015, Adler et al., 2016, Johndrow and Lum, 2017]. These methods differ both in terms of what variables are affected by the data processing, and the degree of independence that is achieved. For instance, Kamiran and Calders [2009] propose flipping the negative labels of some observations in the di”
Understanding Black-box Predictions via Influence Functions (2017, background): “Copyright 2017 by the author(s). point (Ribeiro et al., 2016) or by perturbing the test point to see how the prediction changes (Simonyan et al., 2013; Li et al., 2016b; Datta et al., 2016; Adler et al., 2016).”
Preventing undesirable behavior of intelligent machines (2019, background): “For example, research has proposed means for determining how sensitive a black-box supervised learning algorithm is to each feature in a vector of features used to represent the input [105, 106], which can be used to determine the impact of features, such as race and gender [107].”