Core Concepts
Directed acyclic graphs (DAGs) enable precise identification and localization of data inconsistencies, improving accuracy and downstream performance.
Abstract
"DAGnosis" is a method developed at the University of Cambridge that leverages directed acyclic graphs (DAGs) to identify and localize data inconsistencies. By focusing on tabular data, DAGnosis provides accurate detection of inconsistencies by leveraging structures to pinpoint the causes. This approach improves downstream performance by deferring predictions on inconsistent samples. The method outperforms traditional approaches like Data-SUITE by providing more detailed insights into flagged samples. DAGnosis offers a systematic and principled data-centric approach, enhancing understanding and guiding future data collection efforts.
Stats
DAGnosis leverages structures modeled as directed acyclic graphs (DAGs).
The method shows empirically that leveraging structural interactions leads to more accurate conclusions in detecting inconsistencies.
DAGnosis provides localized instance-wise conclusions by flagging inconsistencies based on feature-wise analysis.
The approach outperforms the state-of-the-art in accuracy of inconsistency detection and downstream accuracy.