Research map / Learning theory and algorithms / Prediction and decision losses
Interactive Hyperparameter Optimization in Multi-Objective Problems via Preference Learning
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 interactive HPO for multi-objective ML that learns a Pareto-front quality indicator from user pairwise comparisons of Pareto fronts via RankSVM, then optimizes with SMAC, evaluated on LCBench/YAHPO energy-accuracy tradeoffs.
Problem
Hyperparameter optimization for multi-objective ML algorithms requires choosing a Pareto front quality indicator, which users often cannot select correctly for their desired front shape.
Approach
Learn a latent utility function over Pareto fronts from pairwise user comparisons (object ranking with linear RankSVM on feature representations of fronts), and use it as the loss in a standard BO-based HPO tool (SMAC) instead of a pre-selected indicator.
Key results
- Preference-based SMAC performs better or equal in 11 of 16 cases versus indicator-based SMAC, and comparable when the correct indicator is chosen (diagonal).
- Preference learning ranking performance (Kendall's Tau) increases with number of pairwise comparisons; HV and R2 easier to learn than SP and MS.
- With only 28 pairwise comparisons the end-to-end approach reaches performance comparable to 140 comparisons (robustness check).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a learned utility from HV preferences instead of optimizing with SP | SMAC for MO-ML hyperparameter optimization · LCBench 32 datasets, 56 pairwise comparisons; HV preference | HV: 0.52 -> 0.75 | Helped |
| Use a learned utility from SP preferences instead of optimizing with HV | SMAC for MO-ML hyperparameter optimization · LCBench 32 datasets, 56 pairwise comparisons; SP preference | SP: 0.03 -> 0.02 | Helped |
| Use a learned utility from MS preferences instead of optimizing with HV | SMAC for MO-ML hyperparameter optimization · LCBench 32 datasets, 56 pairwise comparisons; MS preference | MS: 0.19 -> 0.6 | Helped |
| Use a learned utility from R2 preferences instead of optimizing with SP | SMAC for MO-ML hyperparameter optimization · LCBench 32 datasets, 56 pairwise comparisons; R2 preference | R2: 0.47 -> 0.24 | Helped |
| Use preference-based SMAC rather than indicator-based SMAC with the matching indicator | SMAC for MO-ML hyperparameter optimization · LCBench 32 datasets, correct-indicator cases | indicator-specific quality | No clear effect |
| Train the preference model with 28 rather than 140 pairwise comparisons | Interactive preference-based HPO · LCBench 32 datasets | indicator-specific quality: 28 pairwise comparisons; 140 comparisons | No clear effect |
| Replace linear RankSVM with non-linear RankSVM (ablation) | Interactive preference-based HPO · LCBench 32 datasets | end-to-end HPO performance | No clear effect |
| Replace linear RankSVM with non-linear RankSVM (ablation) | Pareto-front preference ranking · LCBench preference-ranking evaluation | Kendall's Tau | Hurt |
Surprising findings
- Non-linear RankSVM yields worse Kendall's Tau yet comparable end-to-end HPO performance.
Limitations
- Limited to MO problems with two objectives since users must visually compare Pareto fronts.
- Feature representation depends on an ordering of models and imputation, making it assumption-heavy.
- Sampled Pareto fronts may not cover the Pareto front space, risking generalization issues.
Experimental setup
Datasets: LCBench (35 OpenML CC-18 datasets) KDDCup09_appetency covertype Amazon_employee_access
Benchmarks: LCBench via YAHPO-Gym surrogate
Metrics: Kendall's Tau hypervolume spacing maximum spread R2
Baselines: SMAC with hypervolume indicator SMAC with spacing indicator SMAC with maximum spread indicator SMAC with R2 indicator ParEGO
Builds on: SMAC RankSVM LCBench YAHPO-Gym pymoo
Closest papers by meaning
- User Preference Meets Pareto-Optimality in Multi-Objective Bayesian Optimization (AAAI 2025 · similarity 0.89)
- Multi-Objective Bayesian Optimization with Active Preference Learning (AAAI 2024 · similarity 0.88)
- Preference-Guided Diffusion for Multi-Objective Offline Optimization (NeurIPS 2025 · similarity 0.88)
- IOP: An Idempotent-Like Optimization Method on the Pareto Front of Hypernetwork (AAAI 2025 · similarity 0.88)
- Pareto Front-Diverse Batch Multi-Objective Bayesian Optimization (AAAI 2024 · similarity 0.87)
- Improving Pareto Set Learning for Expensive Multi-objective Optimization via Stein Variational Hypernetworks (AAAI 2025 · similarity 0.87)
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.