Reconstructing and Querying ML Pipeline Intermediates


Software systems that learn from user data with machine learning (ML) are used in many critical decision-making processes. Unfortunately, real-world experience shows that their predictions are often very brittle with respect to issues in the data processed by them. In practice, these issues are hard to detect and address because real world ML systems contain complex data preparation pipelines which consume multiple heterogenous inputs and apply a mix of relational and ML-specific operations on the data. We propose to enable data scientists to automatically reconstruct and query the intermediate data from such pipelines to reduce the level of expertise and manual effort required to debug this data.

Conference on Innovative Data Systems Research (CIDR, abstract)