Machine learning approach for higher-order interactions detection to ecological communities management

Ecological communities may present complex interactions, such that the standard Lotka-Volterra mathematical model may not describe the system adequately. Here, we address the question of how to detect automatically the presence of a particular set of ecological nonlinear interactions known as higher-order interactions (HOI). Our proposal is based on estimating standard Lotka-Volterra model parameters by fitting experimental data, and then on generating synthetic HOI samples by stochastic differential equations. These samples are then regarded as input of machine learning classifiers to train the models and predict if the experimental time series present HOI or not. We present a case study using experimental data for single-, pairwise- and three-species of Drosophila. We compare the classical statistical tests from literature against our proposed machine learning approach. For the testing sets, the classical statistical tests (additive test and multiplicative test) showed a lack of robustness, with best results of 0.75 accuracy, 0.50 sensitivity and 1.0 specificity. In contrast, our proposed machine learning HOI recognition approach showed robustness across the testing sets, achieving mean results of 0.93 accuracy, 0.88 sensitivity and 0.99 specificity. The advantage of our machine learning HOI recognition is that we are able to train the classifiers with noisy data that resembles experimental data, producing a more robust automatic tool for HOI detection than classical statistical tests.