MINDS CDT doctoral researcher
Understanding how recommendation systems learn from network patterns
Formal project title: Latent Node Class Effects in Link Prediction Optimisation
Kieran Maguire is investigating how patterns across networks affect link-prediction models, the systems behind many product, content and social recommendations. His work aims to help make recommendations more relevant, transparent and fair.
Research at a glance
| Research focus | Link prediction and recommendation systems |
| MINDS theme | Agent-Based Adaptive Systems |
| Key concepts | Networks and graphs, link prediction, graph neural networks, model optimisation |
| Wider applications | Product, content and social-network recommendations |
| Current status | MINDS CDT doctoral researcher |
| Cohort | MINDS CDT, 2021 |
How online systems predict new connections
Many online services try to predict what a person is likely to connect with next, such as a product to buy, an article to read or someone to follow. Behind many of these suggestions is a task called link prediction.
Link prediction treats data as a network. Each person, product or other item is a node, and each connection or interaction between two of them is a link. A link-prediction model estimates whether a link is likely to exist or form, then ranks the possible connections from most to least likely.
These predictions can be shaped by patterns across the whole network, including how different groups of users are connected and how their interests are distributed. Some of these patterns are latent classes: underlying groups that exist in the data without being directly labelled. A model may appear accurate overall while giving less relevant results to some users, and it can be difficult to see why it ranks one connection above another.
Kieran is investigating how latent classes of nodes affect the way link-prediction models learn and rank possible connections. His research looks inside the training process itself, alongside the quality of the final predictions.
Why the research matters
More relevant recommendations
A better understanding of how network patterns shape model learning could help researchers develop recommendation systems that serve different patterns of user interest more consistently.
Greater transparency
Clearer insight into how models learn to rank connections may make recommendation systems easier to interpret and examine.
Examining fairness
The research could help researchers identify and examine cases where a model serves some parts of a network less well than others.
Research approach
Kieran’s work focuses on optimisation: how link-prediction models adjust during training to improve their predictions. He compares what models learn about the underlying structure of a network with how well they perform on standard link-prediction tests.
In his accepted GCLR 2026 workshop paper, Kieran and his co-author found that standard training can allow some widely used link-prediction models to learn a shortcut. These models are usually trained on small batches of examples containing a fixed mix of real and false connections. Some models can exploit that fixed mix, which may limit how much they learn about the graph as a whole. Varying the mix during training reduced the shortcut and led the models to reflect more of the network’s underlying class structure, with some reduction in standard prediction scores. The findings suggest that standard evaluation may overstate how well link-prediction models learn about the networks they are trained on.
Research keywords: link prediction, recommendation systems, graph neural networks, network structure, latent node classes, model optimisation, fairness, transparency
Research and outputs
Featured output: GCLR 2026 workshop paper
Mini-Batch Class Composition Bias in Link Prediction, by Kieran Maguire and Srinandan Dasmahapatra.
Accepted workshop paper, GCLR 2026: the 5th Workshop on Graphs and more Complex Structures for Learning and Reasoning, co-located with AAAI 2026.
Read “Mini-Batch Class Composition Bias in Link Prediction” on arXiv (arXiv:2604.25978)
Connect with Kieran’s research
Interested in link prediction, recommendation systems or network-based machine learning? Contact the MINDS team about Kieran’s research or a wider collaboration enquiry.
MINDS CDT email: mindscdt@soton.ac.uk