A0660
Title: A conditional Gaussian process model for ordinal data and its application in predicting herbicidal performance
Authors: Arron Gosnell - University of Bath (United Kingdom) [presenting]
Abstract: With the proliferation of screening tools for chemical testing, it is now possible to create vast databases of chemicals easily. On the other hand, the development of a rigorous statistical methodology that can be used to analyse these large databases is in its infancy, and further development to facilitate chemical discovery is imperative. Current methods employed to analyse these data fail to incorporate the chemical structure of the tested compound, and as a result, this feature is unaccounted for in the model. We will discuss the Tanimoto similarity as a measure of closeness between chemical compounds and its use within a Gaussian process model. We will demonstrate the application of the proposed model for analysing data from agricultural experiments to assess the herbicidal performance of chemical compounds. The response variable is ordinal, so a proportional odds model is used, with the cumulative probabilities being functions of the Gaussian process. We will show that accounting for correlation results in improved model performance over a simple mixed-effects model and an alternative random forests model. We will discuss the tools used to overcome certain hurdles in developing the model and the use of proper scoring rules to evaluate model performance.
