Summary of Understanding Disparities in Post Hoc Machine Learning Explanation, by Vishwali Mhasawade et al.
Understanding Disparities in Post Hoc Machine Learning Explanation
by Vishwali Mhasawade, Salman Rahman, Zoe Haskell-Craig, Rumi Chunara
First submitted to arxiv on: 25 Jan 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computers and Society (cs.CY); Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper investigates how data generation processes and black box models contribute to explanation disparities across sensitive attributes such as race and gender. Current post-hoc explanation methods show significant discrepancies in explanation fidelity, with most research focusing on improving explanation metrics rather than examining the role of data and model properties. Through simulations and experiments on real-world datasets, this study assesses how limited sample size, covariate shift, concept shift, omitted variable bias, and model properties like inclusion of sensitive attributes or functional form affect explanation disparities. The results show that increased covariate shift, concept shift, and omission of covariates exacerbate explanation disparities, with neural networks being more susceptible to these issues due to their ability to capture underlying functional forms. The study also finds consistent effects of concept shift and omitted variable bias on explanation disparities in the Adult income dataset. Overall, the paper highlights that disparities in model explanations can be influenced by data and model properties, providing recommendations for designing explanation methods that mitigate undesirable disparities. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at why certain groups have different explanations when using artificial intelligence models to understand what makes them similar or different. Right now, most AI explanation methods don’t work well across different groups like race and gender. The researchers wanted to know if the problem lies in how we collect data or use these models. They ran simulations and tested their ideas on real-world datasets to see how different factors affect explanation disparities. They found that when there’s not enough data, when the data is shifted away from what it was originally, or when important information is missing, explanations get worse for certain groups. They also discovered that some AI models are better at capturing underlying patterns and relationships than others. Overall, the paper shows that we need to consider how we collect data and use these AI models if we want fairer explanations. |
