Fig. 10
From: Human limits in machine learning: prediction of potato yield and disease using soil microbiome data
Boxplots with the weighted F1 scores (y-axis) by random forest and Bayesian neural network (Bayesian NN) models for pitted scab disease. The range of each box plot depicts the weighted F1 scores for normalized datasets at each taxonomic level. There are 6 normalization methods for alpha diversity and environmental factors (Soil and DS), and one normalization method for microbiome (OTU-S3). See Tables 2 and S4 for description of models and number of normalization methods for each predictors. The models including both types of predictors outperform other models, yet models including microbiome data alone (OTU-S3) are comparable which suggests that the microbial information indeed contains signal to predict the disease outcome on its own. However, microbiome data is more expensive to collect, and perhaps not necessary, given that the model without microbiome data (Soil) performs just as accurately (blue dashed line)