This paper is about integrating machine learning capabilities into database systems to solve the high computational cost incurred by existing automatic neural network model selection. Existing studies suggest a method to query the model records with the highest performance indicators for a specific task by organizing the benchmark repository as a model base (MB), but it has limitations in that it overlooks the fine and evolving relational dependencies between various task queries and model architecture variations, which fails to achieve optimal matching and effectively improves the model. In this paper, we propose M-DESIGN, a curated model knowledge base (MKB) pipeline for neural network improvement, to overcome these limitations. M-DESIGN uses a knowledge weaving engine that adaptively weaves prior insights on model architecture modifications into an adaptive query problem to reframe model improvement. Given a user’s task query, it rapidly matches and iteratively improves candidate models by leveraging a graph relational knowledge schema that explicitly encodes data properties, architecture variations, and pairwise performance differences as joinable relations. This schema supports fine-grained relational analysis of architecture adjustments and drives a predictive query planner that can detect and adapt to out-of-distribution (OOD) tasks. In this paper, we implement M-DESIGN for graph analysis tasks, enriching existing benchmarks with structured metadata that includes three graph tasks and 22 graph datasets, providing 67,760 graph model data records. Experimental results show that M-DESIGN provides optimal models for 26 out of 33 data-task pairs within a limited budget.
