Chaining Data - A Novel Paradigm in Artificial Intelligence Exemplified with NMF based Clustering

Norman J Mapes, Sumeet Dua

Keywords: NMF, clustering, linking data, chaining data

Abstract: In the era of artificial intelligence there is an acceleration of high quality inference from the fusion of data and we have overcome the linking challenge associated with higher order features. We have fundamentally linked together tables of databases for clustering algorithms and expect this paradigm and those related to it to produce many new insights. We propose linked view clustering that is an extension of multi-view clustering by adding complementary and consensus information across linked views of each datapoint. While there are many methods, we focus on non-negative matrix factorization combined with the fusion of linking data in a manner that corresponds to extracting knowledge from the multiple tables of a relational database. It is commonplace to identify hashtag communities on social media by word usage, however there exists troves of data not included but could be. We can incorporate locations by hashtag to improve community detection, this is multiNMF or multiview clustering, but we extend this method to beyond the first link. A general artificial intelligence method to incorporate any table that can be chained backwards has not been done before to our knowledge. We call this linked view NMF or chained view clustering and give the algorithms to perform multiplicative updates and the general solution that can be solved using automatic differentiation such as JAX. We demonstrate how the equations can be interpreted on synthetic data as well as how information flows through the links and as a proof of concept on real data we incorporate word vectors using the method on an authorship clustering dataset.

One-sentence Summary: We have fundamentally linked together tables of databases for clustering algorithms and expect this paradigm and those related to it to produce many new insights.