Abstract
Selecting high-performance object-relational mappings (ORMs) requires balancing competing objectives, such as insert and query latency and memory usage. In practice, ORM frameworks typically apply fixed heuristics and emit a single schema. At the same time, prior-learning-based approaches often reduce selection to binary Pareto membership classification, thereby limiting fine-grained tradeoff reasoning. We present Y-Map, a neural–symbolic framework that predicts continuous performance metrics for the ORM schema candidates without executing database workloads during inference. Y-Map constructs supervised training data offline via specification-driven schema synthesis and benchmarking, then learns a regression model over fused schema representations that combine interpretable structural descriptors with embeddings from GraphCodeBERT, CodeT5+, and LLaMA. By predicting continuous insert latency, query latency, and memory usage for candidate schemas, Y-Map enables benchmark-free Pareto-style selection, supporting exhaustive scoring in small spaces and budgeted search in larger ones. Across nine benchmark object models, Y-Map improves tradeoff quality and reduces inference cost relative to a representative benchmarking-based baseline, demonstrating the promise of combining symbolic validity guarantees with learned performance prediction for a practical performance-aware schema design.