📢 New ICML 2021 Paper 📢

Interested in scalable, uncertainty aware, causal sensitivity analyses for personalized treatment recommendations?

Spoiler alert, this one is not just about model size!

paper: arxiv.org/abs/2103.04850
code: github.com/anndvision/quince

1/7

This @icmlconf paper with @sorenmind, @yaringal, and @ShalitUri explores three sources of ignorance about a individual's response to treatment:

a) Insufficient Similarity
b) Insufficient Overlap
c) Insufficient Context (hidden confounding)

2/7

a) Insufficient similarity

We 👏 can’t 👏 know 👏 about 👏 a person’s response to treatment if their data are not properly represented in the study we learn from.

3/7

b) Insufficient overlap

We 👏 can’t 👏 know 👏 about 👏 a person’s response to treatment if similar people are excluded from receiving one of the two treatments.

4/7

c) Insufficient context (hidden confounding)

We 👏 can’t 👏 know 👏 about 👏 a person’s response to treatment if a study excludes factors that are causally linked to both their treatment eligibility and their outcome response.

5/7

Recommending a course of action to someone without considering ignorance due to insufficient similarity, overlap, or context is socially irresponsible and haphazardly harmful.

We propose a principled expression of such ignorance.

6/7

Many thanks to @sorenmind, @yaringal, and @ShalitUri for their contributions to this project!

paper: arxiv.org/abs/2103.04850
code: github.com/anndvision/quince

Stay tuned for work exploring how we can actively reduce such ignorance!

7/7