• 34,123 publications
• 2008 to 2013
• >14,000 no abstract
• >4,000 no authors
• ~2,000 duplicated titles
• ~3,000 Researchers
• ~450 Organizations
• UoB with 7 levels of hierarchy

From 2008 to 2013
34,123 Publications: 14,192 without Abstract,4,459 without authors, and 1987 using duplicate titles. 3263 Researchers from 448 nested Organisations.

The organisation hierarchy of University of Bristol has 7 levels with 'UNIV' in level 0, and the only level 7 organisation is 'TSRC' (Transplantation Sciences (URC)).

• Removed without authors
• Removed duplicates
• Organizations at 3rd hierarchical level
• In the PURE system 68 out of 133

We filtered out publications without authors and removed duplicates.
We also choose organisations at the 3rd level of the hierarchy to represent the individual researchers, as we want to compare organisations at the same level.
There are 133 organizations in the 3rd level, but only 68 of them have submitted papers in the PURE system.

• Undirected graph
• Edge == Organization
• Vertex == #publications in common
• Self-loops == no collaboration

We are interested in identifying the organizations that cooperate more with the rest.
The following chord diagram shows the number of publications coauthored between different organisations and the publications with no cooperation.

• We propose two simple methods
• Absolute number of coauthored publications
• Org_1 = 10 + 3
• Ration of coauthored / no-coauthored
• Org_1 = (10 + 3) / 6

• Different rankings

We show the cooperation relations between organisations in the chord diagram, but how can we rate their contributions to the cooperation atmosphere? How to evaluate the tendency of organisations working with others?

The absolute number of publications coauthored with other organisations is one simplest way, another is the ratio of coauthored publications to those solely worked inside the organisation.

These 2 criteria give us very different rankings, as we show the results in the next slides. The rank 1 organisation is 'SOCS' (School of Clinical Sciences) and 'VICH' (Senior Team) respectively. (The full name of 'VICH' is very mysterious...)

• Clustering: titles and abstracts
• Tokenization and removed stopwords
• Vocabulary size ~15.000 words
• Latent Dirichlet Allocation Topic Models

We use Latent Dirichlet Allocation Topic Models to cluster publications of each year by their titles and abstracts .

The discovered topics are visualised in the next slide, the X coordinate represents how many researchers contributed to the topic, the Y coordinate represents how many organisations contributed, the size of a topic circle represents how many publications assigned to it. We also show top words and top organisations of a topic.

• Tokenization with scikit learn API
• Default Englist stopwords
• Vocabulary size ~15.000 words