A1580
Title: Non-negative matrix tri-factorization for multi-view biclustering
Authors: Marina Evangelou - Imperial College London (United Kingdom) [presenting]
Ella Orme - Imperial College London (United Kingdom)
Abstract: In a range of application domains, data are collected from different sources (referred to as views) to describe the same objects. This could be the same topics reported on by multiple news outlets or in multiple languages or different types of biological information collected on individuals in a clinical study. These are referred to as multi-view or multi-modal data. Incorporating this multi-view data into models can aid in our ability to extract signals and perform statistical tasks. Multi-view clustering is one such task, seeking to cluster samples using data from multiple sources. The further interest is in identifying the features from each view that are the drivers of the sample clusters. This problem is usually referred to as bi-clustering, where both the rows and the columns of a data matrix are clustered simultaneously. The problem of multi-view biclustering and a novel approach based on matrix factorization, known as restrictive multi-view non-negative matrix tri-factorization (ResNMTF), are discussed. Additionally, an extension of the popular silhouette score allowing for comparisons between biclusterings is discussed.
