Sum-Product Network-Based Cardinality Estimation Research

As data distribution is non-uniform and attribute correlations vary in strength, challenges are encountered when using learning-based cardinality estimation methods in database query optimization. An improved algorithm based on the sum-product network was proposed to address these issues. The algorithm has two improvements: a self-adaptive clustering selection algorithm (SCS) for the clustering algorithm module and an RDC-based Automated Search for a Hyperparameter Tuning method (RDC-ASHT) for the attribute partitioning module. SCS selects the best clustering method, improves the ability to partition and merge clusters, and validates the performance of the grid-based clustering method. RDC-ASHT searches for optimal threshold intervals, reducing computational resources and time costs and enabling faster discovery of the optimal solution. Experimental results demonstrate that the improved model has better speed and generalization performance than the original and can accurately estimate cardinalities.