[Paper Review] Active Learning for Node Classification: The Additional Learning Ability from Unlabelled Nodes.

Juncheng Liu, Yiwei Wang+3 more|arXiv (Cornell University)|Dec 13, 2020

Advanced Graph Neural Networks32 references4 citations

TL;DR

This paper proposes LSCALE, a novel active learning method for node classification that leverages unlabelled nodes by combining unsupervised and supervised features in a K-Medoids clustering framework. By dynamically integrating both feature types and using incremental clustering to avoid redundancy, LSCALE significantly outperforms state-of-the-art methods across five benchmark datasets, demonstrating that unlabelled nodes contain untapped learning potential beyond label acquisition alone.

ABSTRACT

Node classification on graph data is an important task on many practical domains. However, it requires labels for training, which can be difficult or expensive to obtain in practice. Given a limited labelling budget, active learning aims to improve performance by carefully choosing which nodes to label. Our empirical study shows that existing active learning methods for node classification are considerably outperformed by a simple method which randomly selects nodes to label and trains a linear classifier with labelled nodes and unsupervised learning features. This indicates that existing methods do not fully utilize the information present in unlabelled nodes as they only use unlabelled nodes for label acquisition. In this paper, we utilize the information in unlabelled nodes by using unsupervised learning features. We propose a novel latent space clustering-based active learning method for node classification (LSCALE). Specifically, to select nodes for labelling, our method uses the K-Medoids clustering algorithm on a feature space based on the dynamic combination of both unsupervised features and supervised features. In addition, we design an incremental clustering module to avoid redundancy between nodes selected at different steps. We conduct extensive experiments on three public citation datasets and two co-authorship datasets, where our proposed method LSCALE consistently and significantly outperforms the state-of-the-art approaches by a large margin.

Motivation & Objective

• To address the limitation of existing active learning methods in node classification, which fail to fully exploit information from unlabelled nodes beyond label acquisition.
• To investigate whether unlabelled nodes can contribute additional learning capacity when combined with unsupervised representation learning.
• To design an active learning strategy that effectively selects informative nodes by clustering in a hybrid feature space of supervised and unsupervised embeddings.
• To reduce redundancy in node selection across multiple active learning rounds using an incremental clustering mechanism.

Proposed method

• LSCALE constructs a latent feature space by dynamically combining unsupervised node representations (e.g., from node2vec or DeepWalk) with supervised embeddings from a previously trained model.
• It applies the K-Medoids clustering algorithm to identify representative nodes in the hybrid feature space, selecting cluster medoids as the next nodes to label.
• The method incorporates an incremental clustering module that updates the clustering structure across active learning rounds to avoid selecting nodes already labeled or highly similar to previously selected ones.
• The selection process is iterative: after each labeling round, the model is retrained, and the feature space is updated to reflect new supervised signals.
• The dynamic combination of features ensures that both structural patterns (from unsupervised learning) and predictive signals (from supervised learning) guide node selection.
• The approach avoids reliance on uncertainty sampling or diversity-based heuristics, instead using clustering in a hybrid feature space to identify informative, representative nodes.

Experimental results

Research questions

• RQ1 Can unlabelled nodes provide additional learning capacity beyond their role in label acquisition during active learning for node classification?
• RQ2 Does combining unsupervised and supervised node features in a latent space improve active learning performance compared to methods that use only supervised signals?
• RQ3 Can K-Medoids clustering on a hybrid feature space effectively identify informative nodes for labeling, outperforming uncertainty-based or diversity-based selection strategies?
• RQ4 How does the incremental clustering mechanism prevent redundancy in node selection across multiple active learning rounds?
• RQ5 To what extent does LSCALE generalize across diverse graph structures, such as citation networks and co-authorship graphs?

Key findings

• LSCALE consistently outperforms state-of-the-art active learning methods for node classification across three citation datasets (PubMed, Cora, Citeseer) and two co-authorship datasets (DBLP, ACM).
• The performance gain of LSCALE is significant and large-scale, with improvements exceeding 10% in accuracy on multiple datasets under the same labeling budget.
• Ablation studies show that the dynamic combination of unsupervised and supervised features is critical to LSCALE’s success, as removing either component leads to performance degradation.
• The incremental clustering module effectively reduces redundancy, ensuring that selected nodes are diverse and representative across active learning iterations.
• The empirical results confirm that unlabelled nodes contain untapped information that can be leveraged through proper feature engineering and clustering, challenging the assumption that only labeled nodes contribute to learning.
• Even a simple baseline using random sampling with unsupervised features outperforms existing active learning methods, highlighting the underutilization of unlabelled data in prior work.