Graph Structure Inference with BAM: Neural Dependency Processing via Bilinear Attention
Authors: Philipp Froehlich, Heinz Koeppl
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluations demonstrate the robustness of our method in detecting diverse dependencies, excelling in undirected graph estimation and showing competitive performance in completed partially directed acyclic graph estimation via a novel two-step approach. Comprehensive empirical evaluations demonstrate that our approach consistently outperforms state-of-the-art methods across diverse scenarios. |
| Researcher Affiliation | Academia | Philipp Froehlich Heinz Koeppl Department of Electrical Engineering and Information Technology Technische Universität Darmstadt {philipp.froehlich, heinz.koeppl}@tu-darmstadt.de |
| Pseudocode | No | The paper presents neural network architectures in figures (e.g., Figure 1, Figure 7) and mathematical descriptions of its layers, but it does not contain any structured pseudocode or algorithm blocks explicitly labeled as such. |
| Open Source Code | Yes | We provide source code in the supplementary material, together with a conda environment file that specifies the required dependencies and their versions and a README file with instructions to run the code. |
| Open Datasets | No | For training data generation, we utilize Erd os Rényi (ER) graphs, denoted as ER(d, q), where the number of nodes d and the expected degree q are sampled from discrete uniform distributions d U({d, . . . , d}) and q U([q, q]), respectively. For each graph Gi, we generate a data matrix Xi RM d using an SEM... |
| Dataset Splits | No | The paper provides details on training data generation and testing procedures, including hyperparameters, but it does not explicitly specify a distinct validation dataset split or a cross-validation setup for hyperparameter tuning. |
| Hardware Specification | Yes | Training our neural network takes approximately 6 hours on an A-100 GPU with 81,920 Mi B of graphical memory, while inference typically requires less than a few seconds and can be run on a normal computer. |
| Software Dependencies | Yes | We provide source code in the supplementary material, together with a conda environment file that specifies the required dependencies and their versions and a README file with instructions to run the code. |
| Experiment Setup | Yes | The hyperparameters for sampling training data, which are defined in section 2.2, are set to d = 10, d = 100, q = 1, q = min( d /3, 5), M = 50, and M = 1000. Table 3: Hyperparameters for the undirected graph estimation (partial list includes) Number of channels C 100, epochs 1000, Initial learning rate 0.0005, Minibatchsize 1. |