Research map / Optimization / Optimizers
Locally Convex Global Loss Network for Decision-Focused 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 LCGLN, a global surrogate loss for decision-focused learning built from a partial input convex neural network that is convex in chosen inputs, and evaluates it on three stochastic decision problems.
Problem
Decision-focused learning requires gradients of optimal decisions w.r.t. predicted parameters, which are hard to compute for non-smooth or discrete optimization; existing surrogate losses are either local per-instance, restricted to certain problem classes, or require hand-chosen parametric forms.
Approach
Learn a single global surrogate of the regret with a partial input convex neural network (PICNN) that is convex in the prediction for chosen inputs (encouraging local minima at the true label) while remaining non-convex globally, trained on model-based samples plus (y, y, 0) anchors, and then used to backpropagate into the predictive model.
Key results
- On the inventory stock problem with 32 samples, LCGLN reached normalized regret 0.174 ± 0.002, versus 0.182 for LANCER and 0.228 for DFL.
- On budget allocation with 500 fake targets and 32 samples, LCGLN achieved 0.468 ± 0.009, the best among all methods (LANCER 0.490).
- On portfolio optimization LCGLN achieved 0.185 ± 0.000 at 32 samples, slightly better than DFL (0.187) and EGL-WMSE (0.187), and LCGLN was best at every sample size tested.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LCGLN instead of the strongest baseline, LANCER | LCGLN · Inventory Stock | normalized test regret: 0.182 $\pm$ 0.004 -> 0.174 $\pm$ 0.002 | Helped |
| use LCGLN instead of the strongest baseline, LANCER | LCGLN · Budget Allocation (500 Fake Targets) | normalized test regret: 0.490 $\pm$ 0.010 -> 0.468 $\pm$ 0.009 | Helped |
| use LCGLN instead of the strongest baseline, DFL | LCGLN · Portfolio Optimization | normalized test regret: 0.187 $\pm$ 0.002 -> 0.185 $\pm$ 0.000 | Helped |
| increase the sample count from 2 to 32 (ablation) | LCGLN · Inventory Stock | normalized test regret: 0.237 $\pm$ 0.005 -> 0.174 $\pm$ 0.002 | Helped |
| increase the sample count from 2 to 32 (ablation) | LCGLN · Budget Allocation (500 Fake Targets) | normalized test regret: 0.529 $\pm$ 0.016 -> 0.468 $\pm$ 0.009 | Helped |
| increase the sample count from 2 to 32 (ablation) | LCGLN · Portfolio Optimization | normalized test regret: 0.186 $\pm$ 0.000 -> 0.185 $\pm$ 0.000 | No clear effect |
Surprising findings
- LANCER reaches 0.182 on inventory regardless of sample size, outperforming LCGLN until 32 samples, indicating LCGLN needs more samples to become competitive.
- Naive MLP surrogates were reported to perform poorly, motivating the PICNN structure.
Limitations
- Achieving high decision quality requires careful selection of samples, as stated by the authors.
- LCGLN underperforms LANCER on inventory with fewer than 32 samples and does not beat all baselines in every setting (e.g., DFL best on budget with 0 fake targets).
Experimental setup
Datasets: portfolio daily stock returns 2004-2017 (50 stocks) synthetic inventory stock instances synthetic budget allocation instances
Benchmarks: inventory stock budget allocation (0/5/50/500 fake targets) portfolio optimization
Metrics: normalized test regret standard error mean
Baselines: PFL DFL LODL-DQ LANCER EGL-WMSE EGL-DQ
Builds on: partial input convex neural network LODL EGL LANCER model-based sampling
Closest papers by meaning
- Decision-Focused Learning with Directional Gradients (NeurIPS 2024 · similarity 0.88)
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (NeurIPS 2025 · similarity 0.87)
- Leaving the Nest: Going beyond Local Loss Functions for Predict-Then-Optimize (AAAI 2024 · similarity 0.85)
- Contextual Optimization Under Model Misspecification: A Tractable and Generalizable Approach (ICML 2025 · similarity 0.85)
- Feasibility-Aware Decision-Focused Learning for Predicting Parameters in the Constraints (NeurIPS 2025 · similarity 0.85)
- DFF: Decision-Focused Fine-Tuning for Smarter Predict-Then-Optimize with Limited Data (AAAI 2025 · similarity 0.84)
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.