Fig. 2.

Covariate relevance assessment comparison with other methods and demonstration of our method’s scalability. (a) Comparison between lgpr and linear mixed effect modelling using the lme4 and lmerTest packages. The panels show ROC curves for the problem of classifying covariates as relevant or irrelevant, when the total number of data points is N = 100, 300 and 600, respectively. (b) Comparison against LonGP. The bar plots show the fraction of times each covariate was chosen in the final model over 100 simulated datasets. The red underlined text indicates the covariates that were relevant in generating the data. The left panel shows results for 100 datasets that includes the disease-related age (diseaseAge) as a relevant covariate. The centre panel shows results for 100 simulations where the disease-related age was not a relevant covariate. The right panel shows distribution of runtimes over the total 200 datasets for both methods. The bar lengths are average runtimes, and the turnstiles indicate runtime standard deviations