Forgetting by Pruning: Data Deletion in Join Cardinality Estimation

2025-11-25

AI Summary

Machine unlearning in multi-table learned cardinality estimation (CE) systems faces three key challenges upon data deletion: attribute-level sensitivity, cross-table error propagation, and join overestimation due to domain value disappearance. Method: This work introduces pruning—previously unexplored in machine unlearning—for efficient parameter updates without full retraining. It proposes two novel pruning strategies: distribution-sensitivity pruning and domain-value pruning. The approach integrates distribution-sensitivity analysis, semi-join-based deletion-result construction, parameter sensitivity scoring, and complete domain-value removal, and is compatible with NeuroCard and FACE architectures. Results: Evaluated on IMDB and TPC-H, the method achieves the lowest query error (Q-error), significantly reduces convergence iterations, and incurs only 0.3%–2.5% of the computational overhead of fine-tuning. It outperforms full retraining in both accuracy and efficiency.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Learning to Search
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Security and Privacy: Security and privacy of machine learning and AI applications

Abstract

Machine unlearning in learned cardinality estimation (CE) systems presents unique challenges due to the complex distributional dependencies in multi-table relational data. Specifically, data deletion, a core component of machine unlearning, faces three critical challenges in learned CE models: attribute-level sensitivity, inter-table propagation and domain disappearance leading to severe overestimation in multi-way joins. We propose Cardinality Estimation Pruning (CEP), the first unlearning framework specifically designed for multi-table learned CE systems. CEP introduces Distribution Sensitivity Pruning, which constructs semi-join deletion results and computes sensitivity scores to guide parameter pruning, and Domain Pruning, which removes support for value domains entirely eliminated by deletion. We evaluate CEP on state-of-the-art architectures NeuroCard and FACE across IMDB and TPC-H datasets. Results demonstrate CEP consistently achieves the lowest Q-error in multi-table scenarios, particularly under high deletion ratios, often outperforming full retraining. Furthermore, CEP significantly reduces convergence iterations, incurring negligible computational overhead of 0.3%-2.5% of fine-tuning time.

Problem

• Addresses data deletion challenges in learned cardinality estimation systems
• Solves attribute sensitivity and domain disappearance in multi-table joins
• Mitigates overestimation issues caused by inter-table dependency propagation

Innovation

• Pruning parameters based on sensitivity scores
• Removing support for eliminated value domains
• Achieving low error with minimal computational overhead