Abstract
We present a study on co-authorship network representation based on network embedding together with additional information on topic modeling of research papers and a new edge embedding operator. We use a link prediction model for constructing a recommender system for searching collaborators with similar research interests. By extracting topics for each paper, we construct a keyword co-occurrence network and use its embedding for further generalizing author attributes. Standard graph feature engineering and network embedding methods were combined for constructing a co-author recommender system formulated as a link prediction problem and prediction of future graph structure. We evaluate our survey on a dataset containing temporal information on National Research University Higher School of Economics over twenty-five years of research articles indexed in Russian Science Citation Index and Scopus. Our model of network representation shows better performance for stated binary classification tasks on several co-authorship networks.
Introduction
In the rapidly expanding landscape of academic publishing, identifying suitable collaborators has become a significant challenge. Traditional recommender systems often rely on simplistic metrics like citation counts or centrality, which fail to capture the nuanced structural identity of researchers or the evolving trajectory of their interests. This paper addresses this gap by framing collaboration discovery as a link prediction problem: predicting whether two authors will collaborate in the future based on current network snapshots and research profiles.
Methodology
The researchers propose a dual-embedding framework that integrates topological data with semantic research interests:
- Dual Network Construction: Two distinct networks are built—a standard co-authorship network (where nodes represent authors) and a keyword co-occurrence network (where nodes represent research topics).
- Embedding with node2vec: The
node2vecalgorithm transforms nodes from both networks into low-dimensional vectors, capturing the local and global context of authors and keywords based on their neighborhood in the graph. - Neighbor Weighted-L2 Operator: To represent potential edges between authors, the authors introduced a novel operator that incorporates information from the nodes' neighbors. This captures local community structure more effectively than standard operators like the Hadamard product or simple averaging.
Experimental Results
The model was evaluated across three datasets: a real-world dataset from the National Research University Higher School of Economics (NRU HSE), a curated Scopus subset, and the large-scale AMiner benchmark to test scalability.
Key findings include:
- Superior Edge Representation: The Neighbor Weighted-L2 operator consistently outperformed standard baselines across almost all metrics, suggesting that a researcher's colleagues' connections are strong predictors of their own future links.
- Predictive Accuracy: On the Scopus dataset, the combined model (topology + research interests) achieved a ROC-AUC of 0.878 and an Accuracy of 0.866.
- Feature Synergy: Integrating research interest embeddings with structural embeddings provided a significant performance boost compared to using either feature set in isolation, confirming that "Dual" is better.
Discussion and Conclusion
While the dual network approach shows great promise, challenges remain regarding data sparsity in networks with few publications and the computational complexity of embedding large-scale graphs. The authors found that Support Vector Machines (SVM) performed most consistently on the sparse, high-dimensional feature spaces generated. Future work will explore deep learning solutions, such as graph auto-encoders, to learn node and edge representations simultaneously and capture more complex non-linear relationships within academic networks.
Summary updated