[Paper Review] Graph Convolutional Networks for Classification with a Structured Label Space

Meihao Chen, Zhuo-Ru Lin+1 more|arXiv (Cornell University)|Oct 12, 2017

Machine Learning and Data Classification28 references17 citations

TL;DR

This paper proposes a Graph Convolutional Network (GCN)-augmented neural network classifier that leverages a known label graph structure to improve multi-class classification, particularly in enhancing the semantic relevance of top-k predictions. By stacking GCN layers on concatenated input and class representations, the model learns label-wise features that respect structural relationships, outperforming baseline models in graph-theoretic metrics like prediction diameter and distance to ground truth, especially in document classification and object recognition tasks.

ABSTRACT

It is a usual practice to ignore any structural information underlying classes in multi-class classification. In this paper, we propose a graph convolutional network (GCN) augmented neural network classifier to exploit a known, underlying graph structure of labels. The proposed approach resembles an (approximate) inference procedure in, for instance, a conditional random field (CRF). We evaluate the proposed approach on document classification and object recognition and report both accuracies and graph-theoretic metrics that correspond to the consistency of the model's prediction. The experiment results reveal that the proposed model outperforms a baseline method which ignores the graph structures of a label space in terms of graph-theoretic metrics.

Motivation & Objective

• To address the limitation of standard multi-class classifiers that ignore structural relationships among labels.
• To improve the relevance of top-k predictions by incorporating known label relationships into the classification process.
• To develop a deep learning framework that approximates structured inference without iterative probabilistic inference.
• To evaluate the model on tasks with hierarchical or correlated labels, such as document classification and object recognition.
• To introduce graph-theoretic metrics to assess the semantic consistency of model predictions beyond standard accuracy.

Proposed method

• The model uses a GCN to process a graph-structured label space, where nodes represent classes and edges encode known relationships.
• Input features and class embeddings are concatenated and passed through multiple GCN layers to produce label-aware representations.
• The final classifier uses a score function based on dot product between input representation and GCN-processed class embeddings.
• The entire network is trained end-to-end via standard backpropagation, avoiding complex inference procedures used in CRFs or MRFs.
• The method incorporates message passing across the label graph to propagate structural information into class representations.
• Graph-theoretic metrics such as prediction diameter and distance to ground truth are used to evaluate the semantic coherence of top-k predictions.

Experimental results

Research questions

• RQ1 Can a GCN-augmented neural network improve the semantic relevance of top-k predictions in multi-class classification by leveraging label structure?
• RQ2 How does the proposed model compare to standard MLPs and CRF-based models in terms of both accuracy and graph-theoretic metrics?
• RQ3 To what extent does the label graph structure influence prediction quality when the model is trained with and without structural awareness?
• RQ4 Does message passing through the label graph lead to more coherent and semantically consistent predictions compared to independent class embeddings?
• RQ5 Can the proposed method generalize across different types of label graphs, such as WordNet hierarchies or definition-based similarity graphs?

Key findings

• The GCNTD model significantly outperforms standard MLP and MLP-CRF baselines in top-k precision and recall, especially as k increases, indicating better clustering of semantically relevant predictions.
• On the document classification dataset, GCNTD achieved 83% top-1 accuracy and 95% top-10 accuracy, with a prediction diameter of 2.40 and mean distance to ground truth of 1.35.
• The GCNTD-FC variant, which assumes a fully connected label graph, achieved the highest accuracy (84%) and best graph-theoretic metrics, confirming the benefit of full message passing.
• In object recognition, GCNTD improved graph-theoretic metrics like diameter and distance compared to MLP, though MLP slightly outperformed in standard accuracy.
• Ensemble methods boosted GCNTD performance more than MLP, particularly in top-1 and top-10 accuracy, indicating stronger generalization under uncertainty.
• The model’s predictions were semantically closer to the ground truth, as evidenced by lower prediction diameter and smaller distance to true labels, especially in the document classification task.