FAT* 2018 Conference Session 2 Summary: Interpretability and Explainability

The Conference on Fairness, Accountability, and Transparency (FAT), which would be held on the 23rd and 24th of February, 2018 is a multi-disciplinary conference that brings together researchers and practitioners interested in fairness, accountability, and transparency in socio-technical systems.

The FAT 2018 conference will witness 17 research papers, 6 tutorials, and 2 keynote presentations from leading experts in the field. This article covers research papers pertaining to the 2nd session that is dedicated to Interpretability and Explainability of machine-learned decisions.

If you’ve missed our summary of the 1st session on Online Discrimination and Privacy, visit the article link for a catch up.

Paper 1: Meaningful Information and the Right to Explanation

This paper addresses an active debate in policy, industry, academia, and the media about whether and to what extent Europe’s new General Data Protection Regulation (GDPR) grants individuals a “right to explanation” of automated decisions.

The paper explores two major papers,

1. European Union Regulations on Algorithmic Decision Making and a “Right to Explanation” by Goodman and Flaxman (2017)
2. Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation by Wachter et al. (2017)

This paper demonstrates that the specified framework is built on incorrect legal and technical assumptions. In addition to responding to the existing scholarly contributions, the article articulates a positive conception of the right to explanation, located in the text and purpose of the GDPR. The authors take a position that the right should be interpreted functionally, flexibly, and should, at a minimum, enable a data subject to exercise his or her rights under the GDPR and human rights law.

Key takeaways:

• The first paper by Goodman and Flaxman states that GDPR creates a "right to explanation" but without any argument.
• The second paper is in response to the first paper, where Watcher et al. have published an extensive critique, arguing against the existence of such a right.
• The current paper, on the other hand, is partially concerned with responding to the arguments of Watcher et al.

Paper 2: Interpretable Active Learning

The paper tries to highlight how due to complex and opaque ML models, the process of active learning has also become opaque. Not much has been known about what specific trends and patterns, the active learning strategy may be exploring.

The paper expands on explaining about LIME (Local Interpretable Model-agnostic Explanations framework) to provide explanations for active learning recommendations. The authors, Richard Phillips, Kyu Hyun Chang, and Sorelle A. Friedler, demonstrate uses of LIME in generating locally faithful explanations for an active learning strategy. Further, the paper shows how these explanations can be used to understand how different models and datasets explore a problem space over time.

Key takeaways:

• The paper demonstrates how active learning choices can be made more interpretable to non-experts. It also discusses techniques that make active learning interpretable to expert labelers, so that queries and query batches can be explained and the uncertainty bias can be tracked via interpretable clusters.
• It showcases per-query explanations of uncertainty to develop a system that allows experts to choose whether to label a query. This will allow them to incorporate domain knowledge and their own interests into the labeling process.
• It introduces a quantified notion of uncertainty bias, the idea that an algorithm may be less certain about its decisions on some data clusters than others.

Paper 3: Interventions over Predictions: Reframing the Ethical Debate for Actuarial Risk Assessment

Actuarial risk assessments might be unduly perceived as a neutral way to counteract implicit bias and increase the fairness of decisions made within the criminal justice system, from pretrial release to sentencing, parole, and probation. However, recently, these assessments have come under increased scrutiny, as critics claim that the statistical techniques underlying them might reproduce existing patterns of discrimination and historical biases that are reflected in the data.

The paper proposes that machine learning should not be used for prediction, but rather to surface covariates that are fed into a causal model for understanding the social, structural and psychological drivers of crime. The authors, Chelsea Barabas, Madars Virza, Karthik Dinakar, Joichi Ito (MIT), Jonathan Zittrain (Harvard),  propose an alternative application of machine learning and causal inference away from predicting risk scores to risk mitigation.

Key takeaways:

• The paper gives a brief overview of how risk assessments have evolved from a tool used solely for prediction to one that is diagnostic at its core.
• The paper places a debate around risk assessment in a broader context. One can get a fuller understanding of the way these actuarial tools have evolved to achieve a varied set of social and institutional agendas.
• It argues for a shift away from predictive technologies, towards diagnostic methods that will help in understanding the criminogenic effects of the criminal justice system itself, as well as evaluate the effectiveness of interventions designed to interrupt cycles of crime.
• It proposes that risk assessments, when viewed as a diagnostic tool, can be used to understand the underlying social, economic and psychological drivers of crime.
• The authors also posit that causal inference offers the best framework for pursuing the goals to achieve a fair and ethical risk assessment tool.