[Paper Review] Node Classification Meets Link Prediction on Knowledge Graphs

Ralph Abboud, İsmail İlkan Ceylan|arXiv (Cornell University)|Jun 14, 2021

Advanced Graph Neural Networks36 references4 citations

TL;DR

This paper unifies transductive node classification and link prediction on incomplete knowledge graphs by proposing MLP-X, a framework that extends shallow embedding models with multi-layer perceptrons to jointly leverage node features and relational structure. The key contribution is a new benchmark dataset, WikiAlumni, where MLP-X models outperform GNNs in both node classification and link prediction, especially under relational incompleteness, demonstrating the benefits of open-world reasoning over the closed-world assumption of GNNs.

ABSTRACT

Node classification and link prediction are widely studied in graph representation learning. While both transductive node classification and link prediction operate over a single input graph, they have so far been studied separately. Node classification models take an input graph with node features and incomplete node labels, and implicitly assume that the graph is relationally complete, i.e., no edges are missing. By contrast, link prediction models are solely motivated by relational incompleteness of the input graphs, and do not typically leverage node features or classes. We propose a unifying perspective and study the problems of (i) transductive node classification over incomplete graphs and (ii) link prediction over graphs with node features, introduce a new dataset for this setting, WikiAlumni, and conduct an extensive benchmarking study.

Motivation & Objective

• To address the gap in existing node classification and link prediction methods by studying transductive node classification on incomplete graphs and link prediction with node features.
• To propose a unified framework that jointly optimizes for both node classification and link prediction using node features and relational structure.
• To introduce WikiAlumni, a new knowledge graph benchmark with real-world features, rich relations, and balanced classes, suitable for both tasks.
• To evaluate the performance of MLP-X models against GNNs across varying degrees of relational incompleteness and feature availability.
• To investigate the interplay between node classification and link prediction, particularly how clustering effects from model inductive biases affect performance.

Proposed method

• Propose MLP-X, a framework that extends any shallow embedding model X by adding a multi-layer perceptron (MLP) to process node features, enabling joint optimization of node classification and link prediction.
• Construct WikiAlumni as a heterogeneous, multi-relational knowledge graph derived from Wikipedia alumni data, enriched with entity features and unary facts for class prediction.
• Train models using cross-entropy loss for node classification and negative sampling or cross-entropy for link prediction, depending on the model’s expressiveness.
• Use a unified training objective that combines node classification and link prediction, allowing the model to learn from both labeled classes and missing relations.
• Evaluate models under controlled relational incompleteness by systematically masking edges and measuring performance on both tasks.
• Compare MLP-X models against standard GNNs and base shallow models (e.g., TransE, RotatE, BoxE) under varying degrees of graph incompleteness and feature availability.

Experimental results

Research questions

• RQ1 Can node classification models be made robust to relational incompleteness by jointly learning from missing edges and node features?
• RQ2 Does incorporating node features improve link prediction performance in knowledge graphs, especially when combined with node classification?
• RQ3 How do the inductive biases of different embedding models (e.g., clustering in TransE and RotatE) affect the trade-off between node classification and link prediction performance?
• RQ4 To what extent does the open-world assumption (OWA) in MLP-X models outperform the closed-world assumption (CWA) of GNNs in incomplete graph settings?
• RQ5 Can a unified framework like MLP-X achieve state-of-the-art performance across both node classification and link prediction tasks on a real-world knowledge graph?

Key findings

• MLP-X models outperform GNNs in transductive node classification on WikiAlumni, especially as relational incompleteness increases, demonstrating robustness to missing edges.
• MLP-X models significantly improve link prediction performance over their base models by leveraging node features, with gains particularly notable in models like MLP-BoxE that are fully expressive.
• MLP-RotatE and MLP-TransE exhibit clustering behavior due to their inductive biases, which improves node classification but degrades fine-grained link prediction, highlighting a trade-off.
• MLP-BoxE achieves the best overall performance across all incompleteness levels because its full expressiveness avoids clustering, and cross-entropy loss better aligns with its scoring mechanism than negative sampling.
• The WikiAlumni dataset is well-suited for both node classification and link prediction, offering real-world features, balanced classes, and rich relational structure for benchmarking.
• The study demonstrates that joint optimization of node classification and link prediction via MLP-X leads to stronger, more generalizable representations than isolated task training.