Optimal Inference in Contextual Stochastic Block Models

Authors: O Duranthon, Lenka Zdeborova

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that there can be a considerable gap between the accuracy reached by this algorithm and the performance of the GNN architectures proposed in the literature. This suggests that the CSBM, along with the comparison to the performance of the optimal algorithm, readily accessible via our implementation, can be instrumental in the development of more performant GNN architectures.
Researcher Affiliation Academia O. Duranthon and L. Zdeborová Statistical Physics of Computation laboratory, École polytechnique fédérale de Lausanne (EPFL), Switzerland EMAIL
Pseudocode Yes Algorithm 1: The AMP BP algorithm. Algorithm 2: The AMP AMP algorithm. Algorithm 3: The multi-community AMP BP algorithm.
Open Source Code Yes We provide a simple-to-use implementation of the algorithm (attached in the Supplementary Material) and in our repository.1 1gitlab.epfl.ch/spoc-idephics/csbm
Open Datasets No Our motivation to study the CSBM is due to the interest this model has recently received in the community developing and analyzing graph neural networks (GNNs). Indeed, this model provides an idealized synthetic dataset on which graph neural networks can be conveniently evaluated and benchmarked.
Dataset Splits Yes We define Ξ as the set of revealed training nodes, that are observed. We set ρ = |Ξ|/N; ρ = 0 for unsupervised learning. We assume Ξ is drawn independently with respect to the group memberships. We define PU,i an additional node-dependant prior. It is used to inject information about the memberships of the observed nodes: We consider the same task as before: on a single instance of CSBM a fraction ρ of node labels are revealed and the GNN must guess the hidden labels.
Hardware Specification Yes Computationally, running the code for a single experiment, N = 3 × 10^4, α = 1 and d = 5 takes around one minute on one CPU core.
Software Dependencies No We use the architectures implemented by the Graph Gym package (You et al., 2020).
Experiment Setup Yes Intra-layer parameters: we take the internal dimension h = 64; we use batch normalization; no dropout; PReLU activation; add aggregation; convolution operation in {generalconv, gcnconv, gatconv} (as defined in the config.py file). Inter-layer design: we take K ∈ {1, 2, 3, 4} layers of message-passing; no pre-process layer; one postprocess layer; stack connection. Training configuration: The batch size is one, we train on the entire graph, revealing a proportion ρ of labels; the learning rate is 3 × 10^−3; we train for forty epochs with Adam; weight decay is 5 × 10^−4.