Algorithmic explanations: to become a mockingbird

Thesis

As algorithms are increasingly used to make potentially life-altering decisions the importance of ensuring they meet certain standards of fairness, accountability, and transparency is becoming paramount. Complex models that are poorly understood are susceptible to bias, reliance on protected or irrelevant features, and adversarial examples. This concern has spawned interest in explainable and interpretable algorithms which allow humans to understand algorithmic decisions. Running parallel to this issue is an increasing reliance on algorithmically generated profiles to target advertisements online. Often these profiles are built on sensitive characteristics, without user knowledge or means of recourse. They can influence user perceptions of self, alter the content users are exposed to, and reveal to other parties characteristics which users have chosen not to share. This project aims to increase user autonomy and privacy online by leveraging techniques in explainable machine learning.

This work begins by creating a prototype tool called Mockingbird that builds more than twenty algorithmically generated profiles based on Twitter data. These profiles are created using various techniques, including lexicons, neural networks, and Gaussian Processes. Mockingbird offers explanations for most of these profiles, primarily using a post-hoc explanatory tool. Finally, using synonym suggestions and techniques used to generate adversarial examples, Mockingbird gives users suggestions on how to alter their tweets in order to change their algorithmic profiles. Mockingbird is novel in combining profiles, explanations, and tools to change algorithmic profiles.

The system was tested through several (n = 6) lab experiments to gauge user response to explanations and willingness to alter their data. It appears that users are generally not concerned about their algorithmic profiles when used for advertising, but would be much more concerned in higher stakes applications. Users were also unwilling to change their existing data to control algorithmically generated profiles, viewing the necessary changes as too disruptive to how they use social media. This work points to avenues of further research focusing explicitly on editing social media data to influence high stakes decisions such as job, visa, and loan applications.

Authors: Hare, A

• Alternative title: Providing twitter users with tools to control algorithmic perceptions
• DOI: 10.5287/ora-m8pbegpzx
• Type of award: MSc
• Level of award: Masters
• Awarding institution: University of Oxford
• Language: English
• Keywords: explainable machine learning; transparency; algorithmic profiles
• Subjects: computers in society; Machine learning