Research map / Label-efficient and robust learning / Recognition and long-tail
ADELA: Accelerating Evolutionary Design of Machine Learning Pipelines with the Accompanying Surrogate Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ADELA, an accompanying surrogate model meta-learned offline on an 11.6M-tuple pipeline corpus to predict pipeline fitness in EA-based AutoML (TPOT) instead of executing pipelines, cutting runtime ~73.66% while retaining ~98.78% of F1.
Problem
EA-based AutoML for ML pipeline design is slowed by costly fitness evaluation of candidate pipelines, and prior surrogate models fail to bridge the semantic gap between heterogeneous dataset meta-features and pipeline representations.
Approach
Use an accompanying pipeline (the best pipeline from the most recent generation, actually executed on the dataset) as a semantic bridge; a siamese GRU-based extractor and dataset attention network feed a meta-trained surrogate that predicts evaluation pipeline performance without executing it.
Key results
- ADELA reaches Relative-F1 of 99.11%, 98.96%, 98.27% for expected pipeline lengths 3, 5, 7, the only method above 98%.
- ADELA reduces runtime by about 72.7% to 74.7% across pipeline lengths (average 73.66%) while retaining 98.78% of final pipeline performance.
- Removing the accompanying mechanism drops Relative-F1 by 4.72 points (to 94.24%) in the ablation at l=5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ADELA to accelerate TPOT pipeline search | TPOT · 16 datasets, l=5 | Time-Reduction: 73.62±10.21% | Helped |
| use ADELA instead of vanilla TPOT | TPOT · 16 datasets, l=5 | Relative-F1: 98.96±1.67% | No clear effect |
| add the accompanying mechanism (ablation) | ADELA · 16 datasets, l=5 | Relative-F1: 94.24±2.98% -> 98.96±1.67% | Helped |
| add the attention-based extractor (ablation) | ADELA · 16 datasets, l=5 | Relative-F1: 96.02±1.86% -> 98.96±1.67% | Helped |
| use the siamese GRU-based extractor instead of average pooling (ablation) | ADELA · 16 datasets, l=5 | Relative-F1: 95.35±2.03% -> 98.96±1.67% | Helped |
| use the code pre-trained model instead of randomized operator embeddings (ablation) | ADELA · 16 datasets, l=5 | Relative-F1: 97.69±1.79% -> 98.96±1.67% | Helped |
Surprising findings
- Selecting the most similar or highest-fitness pipeline from all previous generations performed worse than using the most recent generation, since non-recent pipelines lack semantic correlation with the evaluation pipeline.
Limitations
- Only evaluated on classification tasks with TPOT as the EA-based AutoML testbed.
- Inference speed of the accompanying surrogate is not optimized; authors note future quantization work.
- Relative-F1 declines as pipeline length grows (98.27% at length 7).
Experimental setup
Datasets: OpenML-CC18 spambase phoneme pendigits PhishingWebsites Mushroom Dropout&Success nomao FMA Dota2 Games mfeat-factors isolet
Benchmarks: spambase phoneme pendigits PhishingWebsites Mushroom Dropout&Success nomao FMA Dota2 Games mfeat-factors isolet mfeat-pixel
Metrics: Time-Reduction Relative-F1 macro F1
Baselines: SubStrat IG-KM LTPOT MetaTPOT ML4ML AVATAR
Builds on: TPOT ML4ML MetaTPOT AVATAR UniXCoder
Closest papers by meaning
- Adaptive Data Optimization: Dynamic Sample Selection with Scaling Laws (ICLR 2025 · similarity 0.85)
- EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling (NeurIPS 2026 · similarity 0.84)
- Runtime Analysis of Evolutionary NAS for Multiclass Classification (ICML 2025 · similarity 0.84)
- Decision Tree Induction Through LLMs via Semantically-Aware Evolution (ICLR 2025 · similarity 0.83)
- EVOS: Efficient Implicit Neural Training via EVOlutionary Selector (CVPR 2025 · similarity 0.83)
- Adaptive Prediction-Powered AutoEval with Reliability and Efficiency Guarantees (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.