Institute Disambiguation using Author-Institution Co-Occurrence

TL;DR

The paper tackles Institute Name Disambiguation by proposing an unsupervised, context-driven clustering method that relies on author–affiliation co-occurrence rather than string matching. It constructs an Author–Affiliation Matrix and a co-occurrence graph, using thresholded connected components to form institute clusters, with TPU-accelerated matrix operations enabling scalability. The approach yields encouraging results on a 50,000-paper Indian Crossref subset and is complementary to traditional string-matching methods, suggesting practical utility when integrated with existing systems such as ROR. The work demonstrates potential for scalability, branch-level disambiguation, and open-source deployment, with future improvements including larger datasets and hybrid rules from string-matching literature.

Abstract

In this article we propose a novel method to perform unsupervised clustering of different forms of Institute names. We use only author and affiliation metadata to perform the clustering without any string or pattern matching. After analysing only 50000 articles from Crossref database, we see encouraging results which can be scaled up to provide even better results. We compare our clustering with what a well-known method using string matching does and found that the results were complementary. This can help perform institute disambiguation better when integrated with existing systems, especially to provide aliases for cases where traditional string matching fails. The code of this open-source methodology can be found at: https://github.com/Jeet009/Institute-Disambiguation-using-Author-Institution-Co-Occurrence

Figures (4)

• Figure 1: Steps involved to perform the clustering. This is different from most Institute Disambiguation Methods as it does not perform any string matching
• Figure 2: Threshold 3 : In this figure, we see that the different forms of writing AIIMS New Delhi neatly cluster together. However, there are a few errors marked in red. These are clusters of different branch of the same university. AIIMS Jodhpur and AIIMS Rishikesh. This can be mitigated by increasing the threshold to 4, it performs better with many more forms of AIIMS New Delhi identified.
• Figure 3: Threshold 4 : Higher threshold leads to fewer connected nodes, which is to be expected. However, it also leads to no errors. It shows that threshold forms an important parameter to control the performance of the algorithm. Higher Threshold leads to higher precision but lower recall.
• Figure 4: Threshold 4 : In this figure, we see the third biggest connected component. It is another branch of AIIMS located in another city. The algorithm manages to distinguish between the two as it is not based on string matching.