NeuroSketch: Fast and Approximate Evaluation of Range Aggregate Queries with Neural Networks

Area of Science:

• Computer Science
• Machine Learning
• Database Systems

Background:

• Range aggregate queries (RAQs) are crucial for real-world applications, often requiring fast, approximate answers.
• Existing machine learning (ML) models for RAQs lack theoretical understanding and do not leverage query-specific information.
• Current ML approaches focus on modeling data, limiting performance optimization.

Purpose of the Study:

• To develop a theoretically grounded ML approach for answering RAQs.
• To investigate modeling queries rather than data for improved RAQ performance.
• To introduce a practical framework, NeuroSketch, for efficient RAQ answering.

Main Methods:

• Developed a novel ML approach by training neural networks to learn query answers directly.
• Focused on modeling queries, enabling theoretical analysis and error bounds.
• Created the NeuroSketch framework, a neural network-based system for RAQs.

Main Results:

• Established a theoretical framework providing distribution and query-dependent error bounds for neural networks in RAQs.
• NeuroSketch demonstrates significant performance improvements over state-of-the-art methods.
• Achieved multiple orders of magnitude speedup and enhanced accuracy on diverse datasets.

Conclusions:

• Modeling queries offers a promising direction for ML-based RAQ systems.
• NeuroSketch provides a theoretically sound and practically efficient solution for RAQs.
• The approach advances the state-of-the-art in approximate query answering.