A Data-Driven Investigation Into Similarity Measures For Global Chemical Products

Croda is one of the largest chemical companies in the United Kingdom, producing and distributing products across the globe. It is the aim of this research to provide Croda a means for determining the similarity and therefore interchangeability of their products between manufacturing sites. To do this, numerous analytical approaches including the Bhattacharyya distance, Mahalanobis distance, hierarchical clustering, distribution modelling and separation are investigated. Novel approaches to outlier detection and exploratory analysis are also examined. These analyses are applied to three data sets, each corresponding to a chemical product - Tween20, BrijCS20 and Glycerox HE. These data sets consist of the mass charge ratios and their abundance obtained via MALDI-TOF mass spectrometry. Of the analyses conducted, hierarchical clustering as well as distribution fitting yielded the most promise, although both methods were susceptible to outliers. The Gaussian model, for example, fits the data for the products quite accurately but is less accurate for higher masses. In conclusion, it is found that finding the desired similarity measure is extremely challenging.