MAGNOLIA: Matching Algorithms via GNNs for Online Value-to-go Approximation

Authors: Alexandre Hayderi, Amin Saberi, Ellen Vitercik, Anders Wikum

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We train a GNN to estimate VTG and show empirically that this GNN returns high-weight matchings across a variety of tasks. Moreover, we identify a common family of graph distributions... providing theoretical justification for our approach. Empirical analysis. We present MAGNOLIA (Matching Algorithms via GNNs for Online Value-to-go Approximation), a GNN-based online matching framework... trained on graphs with 16 total nodes, MAGNOLIA beats state-of-the-art baseline algorithms...
Researcher Affiliation Academia 1Department of Computer Science, Stanford University, Stanford, CA, USA 2Department of Management Science & Engineering, Stanford University, Stanford, CA, USA. Correspondence to: Anders Wikum <wikum@stanford.edu>.
Pseudocode Yes We include pseudo-code for computing VG(S, t) in Algorithm 1.
Open Source Code Yes We present MAGNOLIA (Matching Algorithms via GNNs for Online Value-to-go Approximation), a GNN-based online matching framework... 1https://github.com/anders-wikum/GNN-OBM
Open Datasets Yes To simulate performance on real-world crowdsourcing tasks, we generate semi-synthetic graphs from OSMnx (Boeing, 2017), a library that encodes road networks for use in ride-sharing applications, and the g Mission dataset (Chen et al., 2014), whose base graph comes from crowdsourcing data for assigning workers to tasks.
Dataset Splits Yes We perform hyperparameter tuning using a validation set of size 300.
Hardware Specification Yes All the training was done on an NVIDIA Ge Force GTX Titan X.
Software Dependencies No The paper mentions software like 'Py Torch Geometric' and 'Optuna library' but does not specify their version numbers or other crucial software dependencies with versions.
Experiment Setup Yes We construct a training set of 2000 instances on 16 nodes. We perform hyperparameter tuning using a validation set of size 300. We perform around 1000 trials, tuning the parameters as described in Table 1.