Graph-learning-based machine learning improves prediction and cultivation of commercial-grade marine microalgae Porphyridium

Abstract

A graph learning [Binarized Attributed Network Embedding (BANE)] model enhances the single-target and multi-target prediction performances of random forest and eXtreme Gradient Boosting (XGBoost) by learning complex interrelationships between cultivation parameters of Porphyridium. The BANE-XGBoost has the best prediction performance (train R² > 0.96 and test R² > 0.87). Based on Shapley Additive Explanation (SHAP) model, illumination intensity, culture time, and KH₂PO₄ are the most critical factors for Porphyridium growth. The combined facilitating roles of cultivation parameters are found using the SHAP value-based heat map and group. To reach high biomass and daily production rate concurrently, one-way and two-way partial dependent plots models find the optimal conditions. The top 2 critical parameters (illumination intensity and KH₂PO₄) were selected to verify using the graphical user interface website based on the optimized model and lab experiments, respectively. This study shows the graph-learning-based model can improve prediction performance and optimize intricate low-carbon microalgal cultivation.