[Paper Review] Reinforced Neighborhood Selection Guided Multi-Relational Graph Neural Networks

Hao Peng, Ruitong Zhang+4 more|arXiv (Cornell University)|Apr 16, 2021

Advanced Graph Neural NetworksComputer Science126 references20 citations

TL;DR

This paper proposes RioGNN, a reinforcement learning–guided multi-relational GNN that dynamically selects relation-aware neighbors using a label-aware similarity measure and a recursive reinforcement learning framework. It achieves superior performance and explainability in node classification on heterogeneous graphs by adaptively filtering important relations, outperforming SOTA baselines across multiple benchmarks.

ABSTRACT

Graph Neural Networks (GNNs) have been widely used for the representation learning of various structured graph data. While promising, most existing GNNs oversimplified the complexity and diversity of the edges in the graph, and thus inefficient to cope with ubiquitous heterogeneous graphs, which are typically in the form of multi-relational graph representations. In this paper, we propose RioGNN, a novel Reinforced, recursive and flexible neighborhood selection guided multi-relational Graph Neural Network architecture, to navigate complexity of neural network structures whilst maintaining relation-dependent representations. We first construct a multi-relational graph, according to the practical task, to reflect the heterogeneity of nodes, edges, attributes and labels. To avoid the embedding over-assimilation among different types of nodes, we employ a label-aware neural similarity measure to ascertain the most similar neighbors based on node attributes. A reinforced relation-aware neighbor selection mechanism is developed to choose the most similar neighbors of a targeting node within a relation before aggregating all neighborhood information from different relations to obtain the eventual node embedding. Particularly, to improve the efficiency of neighbor selecting, we propose a new recursive and scalable reinforcement learning framework with estimable depth and width for different scales of multi-relational graphs. RioGNN can learn more discriminative node embedding with enhanced explainability due to the recognition of individual importance of each relation via the filtering threshold mechanism. Comprehensive experiments on real-world graph data and practical tasks demonstrate the advancements of effectiveness, efficiency and the model explainability, as opposed to other comparative GNN models.

Motivation & Objective

• Address the limitation of existing GNNs in handling complex, heterogeneous multi-relational graphs with diverse edge semantics.
• Overcome the inefficiency and inflexibility of hand-crafted meta-paths and meta-structures in modeling relational heterogeneity.
• Improve node representation learning by enabling adaptive, relation-specific neighborhood selection that respects semantic diversity.
• Enhance model explainability by identifying and filtering the most relevant relations for each node through a learnable threshold mechanism.
• Develop a scalable, recursive reinforcement learning framework to optimize neighbor selection efficiency across varying graph sizes.

Proposed method

• Construct a multi-relational graph from real-world data to explicitly model node and edge heterogeneity, including attributes and labels.
• Introduce a label-aware neural similarity measure to identify the most relevant neighbors based on node feature and label information.
• Design a reinforced, relation-aware neighbor selection mechanism using deep reinforcement learning to select optimal neighbors per relation.
• Implement a recursive and scalable reinforcement learning framework with estimable depth and width to maintain efficiency on large-scale graphs.
• Apply a filtering threshold mechanism to identify and prioritize the most important relations during message passing for improved explainability.
• Integrate the selected neighborhood information across relations using a learnable aggregation mechanism to produce final node embeddings.

Experimental results

Research questions

• RQ1 How can GNNs effectively model the semantic diversity and complexity of multi-relational graphs without relying on predefined meta-structures?
• RQ2 Can reinforcement learning be used to dynamically and adaptively select the most informative neighbors per relation, improving representation quality?
• RQ3 To what extent does the proposed recursive and scalable RL framework maintain efficiency while enabling deep, relation-aware neighborhood selection?
• RQ4 How does RioGNN enhance model explainability by identifying the importance of individual relations in message passing?
• RQ5 Does RioGNN consistently outperform existing SOTA GNNs in node classification on real-world heterogeneous graph benchmarks?

Key findings

• RioGNN significantly outperforms state-of-the-art GNN models across three real-world benchmark datasets in node classification accuracy.
• The model achieves consistent and stable improvements over baselines, demonstrating robustness across diverse heterogeneous graph structures.
• The reinforcement learning-based neighbor selection mechanism enables more discriminative node embeddings by focusing on relation-specific, high-similarity neighbors.
• The recursive and scalable RL framework maintains computational efficiency even on large-scale multi-relational graphs, supporting estimable depth and width.
• The filtering threshold mechanism enhances model explainability by identifying and highlighting the most influential relations for each node.
• Empirical results show that RioGNN achieves superior performance in practical tasks such as spam review detection and disease diagnosis, validating its real-world applicability.