Accelerating Rule-matching Systems with Learned Rankers
Talkat USENIX ATC '19 · on Friday, 2019-07-12· at 12:10· USENIX ATC Track II: Grand Ballroom VII–IX·
summary
A learned ranker classifies which regex rules to try first, implemented as a neural model trained with generated and logged inputs and evaluated on PCRE, PCRE-JIT and RE2; fully connected nets hit the best latency/accuracy tradeoff versus RNNs, SVMs and logistic ensembles. Models use per-field features and generators that mutate regex-derived inputs; larger rule sets use ~400 neurons per layer. Empirical results show top-5 accuracy ≥96.16% and latency reduced up to 78.81% on an 8‑core CPU.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:38Rulematching System
- 2:56Workflow
- 3:49Optimization
- 4:39Customization
- 5:21New Interest Overhead
- 5:53Temporal Locality Heuristics
- 6:44Transition Machine Learning
- 8:24Training
- 9:30Evaluation
- 9:46Optimal Model
- 11:42Conclusion
- 12:00Questions
related talks
13mGenetic Regex
30mOutwitting Deep Learning Models
29mBuilding Learning to Rank models for search using LLMs
16mScalable Synthesis of Regular Expressions From Only Positive Examples
38mPlaying by the Rules-Based-Systems
18mLearning-Enhanced High-Throughput Pattern Matching Based on Programmable Data Plane
32mApisecure, 14 & 15 March 2023 - Meenakshi Kaushik, Leader - Product Management at Cisco.
43mLearning to Rank for Reddit Search: A Project Retro