OAG Entity Alignment

Given two entity sets E1 and E2, the goal is to generate entity linkings L = {(e1, e2) | e1 \in E1, e2 \in E2} such that e1 and e2 represent exactly the same entity in E1 and E2. The subtasks include venue alignment, affiliation alignment, and author alignment. For venue/affiliation alignment, we use their full names for entity descriptions and regard this problem as short text matching. For author alignment, we use heterogeneous information including papers and venues, and regard this subtask as a subgraph matching problem.

Venue dataset

We generate a relatively hard dataset, by excluding trivial cases that can be matched by full name or short name directly. We manually label 1,200 pairs, consisting of 510 positive pairs and 690 negative pairs. We split the dataset randomly and the training/validation/test split is 4 : 1 : 1.

Affiliation dataset

We construct 5,000 pairs for affiliation alignment. For positive pairs, we utilize the former names or alias of affiliations in Wikipedia pages as candidates. For negative pairs, we search each affiliation in Microsoft Academic Graph and select the second-best candidate to form negative pairs (with the first candidate being the exact correct match). The ratio of positive and negative pairs is kept at 1: 1. We split the dataset randomly and the training/validation/test split is 3 : 1 : 1.

Author dataset

We construct 10,000 pairs for author alignment between AMiner and DBLP, all of which are labeled by humans. To construct negative pairs, for each AMiner author, we sample his/her negative counterpart with the same name in DBLP if exists, otherwise, the most frequent co-author of the matched DBLP author is selected. The ratio of positive and negative pairs is kept at 1: 1. The training/validation/test split is 8 : 1 : 1.

We use Precision, Recall, F1, AUC as metrics for classification and use Hit@K for ranking.

SVM: For venues/affiliations: We calculate the Jaccard Index and Cosine Similarity (using tf-idf values) of each pair as input features.

LinKG_C [1]: For each candidate pair, this method first constructs similarity matrices and then uses CNN to map these matrices to similarity vectors. Finally, similarity vectors of different attributes are concatenated and passed into fully-connected layers to output matching scores. For venue/affiliation linking, we use raw sequences and keyword sequences as attributes. For author linking, we just use venue id list and paper id list associated with the authors.

LinKG_L [1]: In this method, each attribute is treated as a word sequence. The content of each attribute is embedded as a vector by LSTM. For each pair, vectors of the same attribute are merged (by concatenation, subtract, etc.) into similarity vectors, which are fed into fully-connected layers to predict labels. We use the same attributes as LinKG_C model for all tasks.

LinKG_G [1]: In this method, we construct a subgraph for each candidate pair and connect them through matched entities. Then, we employ heterogeneous graph attention networks to obtain the matching score. Specifically, for each author, we incorporate his/her papers, published venues, and co-authors as neighbors, and build a two-hop ego network. In addition, papers and venues are merged across AMiner and DBLP since they are less suffered from name ambiguity. Moreover, we concatenate the features of SVM and the hidden embeddings of the merged author pair before the output MLP layers.

Table 1: Results of Venue Alignment (%)

Table 2: Results of Affiliation Alignment (%)

Table 3: Results of Author Alignment (%)

Code & Data: https://github.com/THUDM/OAG-entity-alignment

We kindly request that any published research that makes use of this data cites the following papers.  

@inproceedings{zhang2024oag,
  title={OAG-bench: a human-curated benchmark for academic graph mining},
  author={Fanjin Zhang and Shijie Shi and Yifan Zhu and Bo Chen and Yukuo Cen and Jifan Yu and Yelin Chen and Lulu Wang and Qingfei Zhao and Yuqing Cheng and Tianyi Han and Yuwei An and Dan Zhang and Weng Lam Tam and Kun Cao and Yunhe Pang and Xinyu Guan and Huihui Yuan and Jian Song and Xiaoyan Li and Yuxiao Dong and Jie Tang},
  booktitle={Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={6214--6225},
  year={2024}
}

@article{zhang2022oag,
  title={Oag: Linking entities across large-scale heterogeneous knowledge graphs},
  author={Zhang, Fanjin and Liu, Xiao and Tang, Jie and Dong, Yuxiao and Yao, Peiran and Zhang, Jie and Gu, Xiaotao and Wang, Yan and Kharlamov, Evgeny and Shao, Bin and others},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  volume={35},
  number={9},
  pages={9225--9239},
  year={2022}
}

@inproceedings{zhang2019oag,
  title={OAG: Toward linking large-scale heterogeneous entity graphs},
  author={Zhang, Fanjin and Liu, Xiao and Tang, Jie and Dong, Yuxiao and Yao, Peiran and Zhang, Jie and Gu, Xiaotao and Wang, Yan and Shao, Bin and Li, Rui and others},
  booktitle={Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery \& data mining},
  pages={2585--2595},
  year={2019}
}