Abstract
Link prediction is a fundamental task in network analysis, particularly within large-scale document repositories like Wikipedia. While traditional methods rely heavily on graph topology, this research proposes a text-centric approach utilizing Part-of-Speech (POS) tagging and statistical feature selection. By leveraging a Random Forest classifier, the proposed method achieved a near-perfect F1 score of 0.99999 in the DSAA 2023 Link Prediction Challenge, demonstrating that linguistic structure is a highly effective proxy for network connectivity.
Introduction
In document-based networks, link prediction involves determining the likelihood of a connection—such as a hyperlink—existing between two nodes. Most conventional approaches focus on the structural properties of the graph, such as common neighbors or path lengths. However, these methods often overlook the rich semantic and linguistic information contained within the nodes themselves.
Researchers from Vietnam National University addressed this gap by developing a pipeline that prioritizes textual features over graph topology. Their work demonstrates that the linguistic composition of Wikipedia articles provides sufficient signal to predict links with extreme accuracy, even without complex deep learning architectures.
Methodology
The authors developed a streamlined pipeline focused on high-quality feature engineering rather than model complexity. The process is divided into four primary stages:
1. Preprocessing
To ensure the integrity of the linguistic analysis, the raw Wikipedia text was cleaned to remove special characters and noise that could interfere with natural language processing tools.
2. POS Tagging
Using the NLTK library, the researchers performed Part-of-Speech (POS) tagging on the text of every article. This process identifies the grammatical role of each word (e.g., noun, verb, adjective), transforming raw text into a structured linguistic profile.
3. Statistical Feature Selection
A key innovation in this study was the use of a statistical t-test to refine the feature set. Rather than utilizing all 37 available POS tags, the authors compared linked pairs against non-linked pairs to identify which tags were statistically significant indicators of a relationship. This process narrowed the focus to 18 specific tags, including nouns (NN), adjectives (JJ), determiners (DT), and coordinating conjunctions (CC).
4. Classification
The researchers evaluated several machine learning models, including Logistic Regression, K-Nearest Neighbors (KNN), XGBoost, and Artificial Neural Networks (ANN). Ultimately, the Random Forest algorithm was selected as the optimal model due to its superior ability to capture non-linear relationships between the selected linguistic features.
Experimental Setup and Dataset
The study was conducted using the dataset provided by the DSAA 2023 Competition, which consisted of:
- node.tsv: A collection of 837,834 Wikipedia nodes containing unique IDs and article text.
- train.csv: A balanced dataset of 948,232 samples, comprising approximately 46% positive (linked) pairs and 54% negative (unlinked) pairs.
The dataset represented a sparsified subgraph of the broader Wikipedia network, providing a rigorous environment for testing the generalizability of the text-based approach.
Results and Discussion
The Random Forest model achieved an exceptional F1 score of 0.99999 on both the public and private leaderboards of the DSAA 2023 Link Prediction Challenge, securing 7th place overall.
Key Findings:
- Linguistic Commonality: The frequency and density of shared POS tags between two articles proved to be an incredibly strong predictor of a link. High commonality in specific tags like nouns and adjectives serves as a reliable proxy for topical relevance.
- Model Stability: The model exhibited virtually no performance degradation between the public and private test sets, suggesting that the feature selection process successfully avoided overfitting.
- Computational Efficiency: Unlike heavy Transformer-based models (e.g., BERT), this POS-tagging approach is computationally lightweight. It provides a high-performance alternative for environments where training time and hardware resources are limited.
Conclusion and Future Work
This research highlights the power of traditional machine learning and statistical pruning in the era of Large Language Models. By focusing on the underlying linguistic structure of documents, the authors achieved state-of-the-art results in link prediction.
Future research directions include:
- Dynamic Prediction: Adapting the model to handle evolving networks where links are added or removed over time.
- Advanced Taggers: Exploring more modern NLP libraries like spaCy or Stanza to determine if higher tagging accuracy further improves prediction.
- Hybrid Models: Integrating these textual features with graph-based embeddings (such as node2vec) to create a multi-modal prediction system.
Summary updated