Incorporating Interventional Independence Improves Robustness against Interventional Distribution Shift
Gautam Sreekumar [email protected], Vishnu Naresh Boddeti [email protected] · 2025
Read Incorporating Interventional Independence Improves Robustness against Interventional Distribution Shift by Gautam Sreekumar [email protected], Vishnu Naresh Boddeti [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-05412.txt.
Description
Gautam Sreekumar and Vishnu Naresh Boddeti show, in a 2025 paper, that teaching a model the independence structure created by interventions makes it sturdier when the data distribution shifts. Their setting is learning representations of hidden variables linked by a known causal graph. Training data mixes passive observations with data from targeted interventions on some variables.
Standard practice pools both data types and ignores the independence relations the interventions create. The result is a representation that predicts far worse on intervened data than on observed data, and the gap grows when intervention data is scarce. The authors first tie that gap to violations of the intervention induced independence. For linear models, they derive how much interventional data suffices for a cure. Building the causal structure into training, rather than hoping the model finds it, is the paper's message for robust machine learning.
Catalog record
| CREATOR | Gautam Sreekumar [email protected], Vishnu Naresh Boddeti [email protected] |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 120 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-05412