Link Prediction between Group Entities in Knowledge Graphs (Student Abstract)

Jialin Su, Yuanzhuo Wang, Xiaolong Jin, Yantao Jia, Xueqi ChengAAAI 2020 · OpenReview

Abstract

Link prediction in knowledge graphs (KGs) aims at predicting potential links between entities in KGs. Existing knowledge graph embedding (KGE) based methods represent individual entities and links in KGs as vectors in low-dimension space. However, these methods focus mainly on the link prediction of individual entities, yet neglect that between group entities, which exist widely in real-world KGs. In this paper, we propose a KGE based method, called GTransA, for link prediction between group entities in a heterogeneous network by integrating individual entity links into group entity links during prediction. Experiments show that GTransA decreases mean rank by 5.4%, compared to TransA.

Introduction

Knowledge Graphs (KGs) are massive networks of facts represented as triples: (head entity, relation, tail entity). Link prediction is the task of guessing missing facts—for example, predicting that "Company A" might "Collaborate with" "Company B."

Most existing KGE methods (like TransE or TransA) treat every entity as a single point in a vector space. However, real-world networks often contain a hierarchy consisting of individual entities (e.g., specific researchers) and group entities (e.g., institutions). A relationship between two groups is often the sum of many individual interactions between their members. Standard models often ignore this internal structure, treating a group as a "black box," which leads to suboptimal performance when member-level data is available.

Methodology

The proposed method, GTransA, builds upon the TransA model—a locally adaptive translation method. The core innovation is the introduction of a dynamic weight function, $w(r)$, into the standard scoring function:

$$f_r(h, t) = ||h + w(r) - t||$$

Instead of a static relation vector, $w(r)$ adjusts based on the relationship between the members of the two groups. This weight is determined by two factors:

  1. Link Density ($\alpha_{U,V}$): Measures the frequency of individual links between members of Group U and Group V.
  2. Group Significance ($\beta_{U,V}$): Accounts for the total connectivity of each group to prevent bias toward larger groups.

By integrating these member-level statistics, the model learns that a group-level link is more likely if supported by a robust foundation of individual-level interactions.

Experimental Setup

The researchers evaluated GTransA using a heterogeneous network constructed from the DBLP and Aminer academic datasets. The graph included entities such as researchers, papers, institutions, and venues. The primary task was to predict future collaboration links between institutions (group entities) based on the historical interactions of their researchers (individual entities).

Results and Discussion

GTransA was compared against traditional KGE methods and graph-based baselines.

Experimental results comparing GTransA with baseline methods across Mean Rank and Hits@10 metrics.

As shown in the results, GTransA outperformed all baselines. It achieved a Mean Rank of 4103.78 (Filtered), representing a 5.4% improvement over the previous state-of-the-art, TransA (4340.75). In the Hits@10 metric, GTransA reached 0.245, surpassing TransA’s 0.234. These results suggest that leveraging member-level data makes the model more robust against data sparsity.

Key insights include:

  • Groups as Aggregates: Group entity representations should be functions of their constituents, reflecting collective behavior.
  • Hierarchical Integration: GTransA successfully bridges the gap between micro-links (individuals) and macro-links (groups).
  • Effective Weighting: Logarithmic scaling in the weight functions prevents large groups from dominating the model while rewarding high-density connections.

Conclusion

GTransA demonstrates that incorporating hierarchical member information into KGEs significantly improves group-level link prediction. While the model assumes known group memberships, it offers a powerful approach for modeling complex organizational relationships in heterogeneous networks. Future research may explore handling dynamic group memberships and reducing computational overhead for extremely large groups.

Summary updated

Abstract

Link prediction in knowledge graphs (KGs) aims at predicting potential links between entities in KGs. Existing knowledge graph embedding (KGE) based methods represent individual entities and links in KGs as vectors in low-dimension space. However, these methods focus mainly on the link prediction of individual entities, yet neglect that between group entities, which exist widely in real-world KGs. In this paper, we propose a KGE based method, called GTransA, for link prediction between group entities in a heterogeneous network by integrating individual entity links into group entity links during prediction. Experiments show that GTransA decreases mean rank by 5.4%, compared to TransA.
BibTeX
@inproceedings{su2020link,
  author = {Jialin Su and Yuanzhuo Wang and Xiaolong Jin and Yantao Jia and Xueqi Cheng},
  title = {{Link Prediction between Group Entities in Knowledge Graphs (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=Go0xSuJ2bL},
}
Figure