[Paper Review] Understanding Pooling in Graph Neural Networks

Daniele Grattarola, Daniele Zambon+2 more|arXiv (Cornell University)|Oct 11, 2021

Advanced Graph Neural Networks4 citations

TL;DR

This paper introduces a unified framework, SRC (Selection, Reduction, Connection), to formalize graph pooling in Graph Neural Networks by decomposing pooling into three core operations. It proposes a taxonomy of over 30 pooling methods, evaluates them using three criteria (attribute, structure, and task preservation), and finds no single method universally outperforms others—guiding practitioners to select pooling strategies based on task-specific needs like structure preservation or trainable flexibility.

ABSTRACT

Inspired by the conventional pooling layers in convolutional neural networks, many recent works in the field of graph machine learning have introduced pooling operators to reduce the size of graphs. The great variety in the literature stems from the many possible strategies for coarsening a graph, which may depend on different assumptions on the graph structure or the specific downstream task. In this paper we propose a formal characterization of graph pooling based on three main operations, called selection, reduction, and connection, with the goal of unifying the literature under a common framework. Following this formalization, we introduce a taxonomy of pooling operators and categorize more than thirty pooling methods proposed in recent literature. We propose criteria to evaluate the performance of a pooling operator and use them to investigate and contrast the behavior of different classes of the taxonomy on a variety of tasks.

Motivation & Objective

• To address the lack of a unifying formalism for graph pooling operators in GNNs, which are widely used but poorly standardized in the literature.
• To formalize graph pooling as a composition of three operations: selection (node grouping), reduction (attribute aggregation), and connection (edge reconstruction).
• To develop a comprehensive taxonomy of over 30 existing pooling methods based on learnability, complexity, output size, and pooling strategy (hierarchical vs. global).
• To define and apply three evaluation criteria—attribute preservation, structural preservation, and task-specific adaptability—to compare pooling operators across diverse benchmarks.
• To provide practical guidelines and a framework-agnostic Python API for implementing and evaluating pooling operators in GNNs.

Proposed method

• Propose a modular, formal framework where graph pooling is decomposed into three functions: selection (grouping nodes into supernodes), reduction (aggregating node features per supernode), and connection (reconstructing edges between supernodes).
• Formalize the SRC functions as node- and graph-embedding operations, enabling universal approximation of any pooling operator using recent GNN universality theory.
• Categorize pooling methods along four dimensions: learnability (learned vs. non-learned), complexity (linear vs. quadratic), output size (fixed vs. variable), and pooling strategy (hierarchical vs. global).
• Define three evaluation criteria: (1) preservation of node attribute information, (2) preservation of original graph topology (via spectral and Laplacian alignment), and (3) adaptability to downstream tasks (measured by downstream accuracy).
• Implement a standardized, framework-agnostic Python API (SRC) to enable reproducibility and integration of pooling operators into GNN libraries.
• Conduct extensive experiments on synthetic and real-world datasets (e.g., PROTEINS, MUTAG, ModelNet10) to evaluate and contrast the performance of different pooling classes under the defined criteria.

Experimental results

Research questions

• RQ1 How can graph pooling operators be formally unified under a single, modular framework?
• RQ2 What are the key structural and functional differences among the diverse graph pooling methods proposed in the literature?
• RQ3 Which pooling operator characteristics (learnability, complexity, output size, strategy) most influence performance across different graph types and tasks?
• RQ4 To what extent do different pooling operators preserve node attributes, graph structure, and task-specific information?
• RQ5 Can a systematic evaluation framework identify the most suitable pooling method for a given application scenario?

Key findings

• No single pooling operator consistently outperforms all others across all datasets and criteria; performance varies significantly by task and graph type.
• Non-trainable, sparse methods like Graclus and NDP achieve the highest average accuracy on several benchmarks, including Enzymes and Mutagenicity, suggesting strong generalization.
• On large-scale datasets like ModelNet10, spectral-based methods (MinCut, NDP) and trainable dense methods (DiffPool) achieve top performance, with MinCut reaching 83.9% accuracy.
• Dense, trainable pooling operators (e.g., DiffPool, MinCut) show superior performance in preserving structural information, as measured by Laplacian quadratic loss and spectral alignment.
• Trainable sparse methods often discard entire graph substructures, leading to information loss and reduced performance, making them less advisable overall.
• The no-pooling baseline outperforms some pooling methods on certain datasets (e.g., Mutagenicity), indicating that graph pooling is not always beneficial and should be applied with caution.