Research · TU Delft

Questions worth a semester

Two strands I chased at TU Delft: teaching models to make good decisions, not just good predictions. And graph neural networks for the graphs where similar things don’t connect.

Learning Feasible Decisions Directly from Features for Combinatorial Optimization

Cristian Turcan · first author

A solver-free way to predict decisions that are feasible by construction, trained to minimise regret, not prediction error.

Proposes LR-STE, which projects a neural predictor’s output onto the problem’s linear relaxation and applies a feasibility-correction layer, trained end-to-end with a straight-through estimator to minimise decision regret rather than prediction error. Benchmarked on 0/1 Knapsack and the Travelling Salesperson Problem against SPO+, two-stage predict-then-optimize, and projection-free (STE) baselines.

decision-focused learning predict-then-optimize LR-STE knapsack · TSP

Read the paper

Heterophilic Methods on Multi-Label Graphs

Cristian Turcan · sole author

On multi-label graphs, how tightly a graph “sticks together” predicts accuracy more than how clever the model is.

Benchmarked 8 models (MLP and DeepWalk baselines plus 6 heterophily GNNs: H2GCN, GPR-GNN, LINKX, FAGCN, ACM-GCN, Ordered GNN) across 5 real and 2 synthetic multi-label graphs in PyTorch Geometric. Controlled homophily and label-cardinality sweeps showed that node-classification performance tracks graph homophily more than model sophistication. Built gnnbench, a config-driven pipeline with a synthetic multi-label graph generator, hardened with split-leakage tests and parallelised as SLURM array jobs on TU Delft’s DelftBlue cluster.

heterophilic GNNs multi-label graphs PyTorch Geometric gnnbench

Read the thesis