Benchmarks for Data Clustering Algorithms in Data Science

About the Project

Data Science problems focus on processing and analysing so-called Big Data to discover insights and patterns. Big Data examples include massive sensor data, unstructured and semi-structured historical data, and many types of social-media, or biological, or chemical data. The storage requirements for Big Data make activities like retrieval, processing, and analysis challenging and time-consuming. Clustering these big data sets unravels these challenging issues.

The problem of Data Clustering is concerned with categorising unlabelled data into groups. The expected output of clustering is a labelling of the data such that members in the same group would be similar. Data Clustering is an established unsupervised learning technique; the earliest algorithms date back to the 1960s. Today, there is a variety of clustering approaches: density-based, partition-based, hierarchical, parallel and distribute, etc. Under each paradigm there are typically tens of algorithms.

In this PhD, the student will investigate the key paradigms of data clustering, they will seek to create benchmark data sets to compare algorithms against one another, and they will aim to apply the best clustering techniques to a problem domain of their choice. There is a plethora of data clustering algorithms but rarely are they benchmarked against a standard data set that can indicate how good they are. The project will update previous work done by the supervisor as well as new work from others. The work programme will include extending 2D variable-sized benchmarks to 3D and further, extending two-cluster benchmarks to more than two clusters, and extending circular-cluster benchmarks to other shapes.

Successful completion of the work will result in research publications at the highest standard and a track record for the student in the field of Data Science.

Funding Notes

there is no funding for this project