Towards Decision Focused Learning for Sparse and Weakly Supervised Environments

Raashid Altaf, Pravesh Biyani

Keywords: Decision Focused Learning, Meta Learning, Weakly Supervised Learning, Personalised Trip Planning

Abstract:

Decision-focused learning (DFL) integrates machine learning and optimisation by training predictive models to directly optimise decision quality.
However, DFL typically depends on access to accurate ground-truth targets such as objective function parameters or optimal decisions: This is often infeasible in real-world settings where only sparse, binary feedback on decision outcomes is available.

This work takes first steps towards formalising Decision-Focused Learning for settings where observed data is limited to such binary feedback. We propose a preliminary DFL framework that learns latent user preferences from weakly supervised binary feedback on decision outcomes. The novelty of our approach lies is in a) a ground-truth-free, differentiable surrogate loss that maps binary evaluations to decision outcomes, and b) a novel meta-learning mechanism that learns latent user preference patterns and transfers this knowledge between users to mitigate challenges due to per-user data sparsity.

Our experiments suggest that this framework can reduce decision regret by 20-fold and achieves convergence with $2.4-4\times$ fewer data points than standard predict-then-optimise baselines. On a novel hyper-sparse real-world trip-planning feedback dataset, we show the model's ability to extract user-preference clusters from sparse data ($\approx 1$ interaction/user). We also evaluate our model in cold-start recommendation settings and show that our decision loss correctly prioritises ranking quality, achieving $9.5$\% higher nDCG@5 than the baseline despite $14.6$\% higher MSE.
The aim of this work is to broaden the applicability of DFL and explore its potential in weakly supervised data-sparse regimes, with future work extending non-linear user-preference structure.

Primary Area: optimization

Submission Number: 23285