Sparse Causal Model: A Novel Approach for Causal Discovery and Attributions on Sparse Dataset

kalyani pakhale, MADHU KRISHNA R, Ravindra Patil

Keywords: Causality, Causal Inference

TL;DR: This study introduces a causal modeling framework for sparse data, improving model selection and attribution, with a 53% boost in R-squared accuracy over traditional methods.

Abstract: This paper introduces a novel approach to tackle the challenges of causal modeling and attribution in sparse and non-continuous data with limited feature knowledge. Traditional methods rely on static inputs and lack adaptability to dynamic changes in causal relationships, resulting in a limited understanding and goodness-of-fit. We introduce a unique causal discovery framework on real-world sparse datasets to address this challenge. We leverage a Directed Acyclic Graph (DAG) by discovering causal relationships between the variables by identifying confounder-treatment pairs that make the variable selection process robust and efficient. We propose a three-stage causal model that uses multiple distinct regressors such as likelihood-based, tree-based, and Generalized Additive Models (GAMs). Furthermore, we introduce a Model Score by including the sensitivity analysis involving random shuffling confounders and treatments to select the best optimal model. We implement a partial dependency approach to understand the attribution of variables, contributing by adding a 53% increase in the R2 score compared to traditional methods. This research underscores the limitations of conventional approaches in addressing real-world challenges to address practical scenarios effectively.

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10292