Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
On Logic-based Self-Explainable Graph Neural Networks
Authors: Alessio Ragno, Marc Plantevit, Céline Robardet
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Logi X-GIN across several graph-based tasks and show that it achieves competitive predictive performance while delivering clear, logic-based insights into its decision-making process. In this section, we perform extensive experiments to evaluate our proposal. We begin by comparing the classification performances of Logi X-GIN with respect to its black-box counterpart. Successively, we perform an analysis of the interpretability of Logi X-GIN also in comparison with state-of-the-art approaches. We mainly focus on a logic-based analysis by identifying the global logic rules and layer-wise rules. For an additional analysis, we also report an experiment on node attributions in the Appendix E, comparing with other self-interpretable models and post-hoc approaches. In Table 1, we report the accuracy of the Logi X-GIN model compared to that of the black-box model across all datasets. Table 2, instead, reports the accuracy scores of several SE-GNNs on graph classification. |
| Researcher Affiliation | Academia | Alessio Ragno INSA Lyon, CNRS, LIRIS UMR 5205, F-69621 Villeurbanne, France EMAIL Marc Plantevit EPITA Research Laboratory (LRE), F-94276, Le Kremlin-Bicêtre, France EMAIL Céline Robardet INSA Lyon, CNRS, LIRIS UMR 5205, F-69621 Villeurbanne, France EMAIL |
| Pseudocode | No | The paper describes the Logi X-GIN architecture using mathematical equations (e.g., Equation 1, 3, 4, 5, 7) and textual descriptions of its components and operation. It does not contain a dedicated pseudocode block or an algorithm section with structured, code-like steps. |
| Open Source Code | Yes | Overall, the main contributions of our work are as follows: ... we provide an open-source implementation of our proposed approach and the conducted experiments1. 1Public Git Hub repository: https://github.com/spideralessio/Logi X-GIN |
| Open Datasets | Yes | We perform our experiments on the following 7 graph and 3 node classification datasets spanning synthetic graphs (BA2Motifs [15], BAMulti Shapes [2], BAShapes, BACommunity and Tree Grid [31]), molecular graphs (MUTAG [11], Mutagenicity [14], NCI1 [28], and BBBP [16]), and protein graphs (PROTEINS [12]). |
| Dataset Splits | No | We use grid-search to find the optimal hyperparameter combinations and report the test set results of the best performing hyperparameters on the validation set. Specifically, to ensure statistical validity, we report mean and standard deviation values over 10 different seeds. |
| Hardware Specification | Yes | The experiments were performed on a machine equipped with an Intel(R) Core(TM) i9-10900K CPU @ 3.70GHz and an NVIDIA Ge Force RTX 4090. |
| Software Dependencies | No | The implementation is done in Py Torch and Py Torch-Geometric and all the library versions are detailed in the environment specifics available in the supplementary material. |
| Experiment Setup | Yes | In Table 4 and Table 5 we detail the hyperparameters of the models. |