273 Comparing the utility of machine learning algorithms in methane predictions from a global beef cattle database.

Abstract Enteric fermentation accounts for 46% of greenhouse gas emissions from livestock. As feed composition directly influences enteric methane production, it has become a key factor in many prediction models. With the increasing adoption of machine learning (ML) across scientific domains, this study evaluates the effectiveness of tree-based ML models compared to traditional regression methods in predicting methane emissions using a global beef cattle database. This study used a publicly available database of 895 methane mitigation experiments in beef cattle from 1984 to 2023, encompassing 139 features (variables) related to feed composition, methane emission, and experimental conditions. Preprocessing was done based on a previously conducted literature review and eliminating parameters with missing values. The final dataset contained 251 experimental entries and 12 columns of dietary information on the major feed components (Dietary Crude Protein, Acid Detergent Fiber, Neutral Detergent Fiber, Ether Extract, Ash), Experimental Treatments (E.g. Feed additives), Animal parameters (E.g. Metabolic Status, Dry Matter Intake (DMI), Measurement techniques (GreenFeed, Headbox, etc.). Methane yield (CH4g/DMI kg) was designated as the output variable for the models. The descriptive statistics of the dataset were retrieved before data standardization and encoding of categorical features. The Train: Test ratio was assigned to 80:20 and the dataset was subjected to Traditional regression models (Linear Regression - LR, Ridge Regression - RR, Lasso Regression- LsR) and Tree based ML algorithms (Decision Trees - DT, Random Forest- RF, Gradient Boosting- GB, and Extreme Gradient boosting- XGB). Out of the algorithms we explored, RF, GB, and XGB performed well with low Mean Absolute Error (MAE) and Mean Squared Error Values (MSE) and a relatively high Coefficient of Determination (R2) indicating that the advanced tree-based ML algorithms capture the complexities of feed-methane dynamics well across experimental settings. Among the models compared, XGB showed the best performance with an MAE of 2.83 (CH4g/kg DMI), and an MSE of 16.30 (CH4g/kg DMI)2 in methane yield predictions with an R2 of 0.64, while LsR had the weakest predictions (MAE = 4.53CH4g/kg DMI, MSE = 29.62 (CH4g/kg DMI)2, R2 = 0.34). The analysis shows the potential of tree-based ML algorithms to handle high-dimensional (multi-factorial) datasets and highlight non-linear relationships in animal production and nutrition studies. Future work will focus on transforming these models with real-time data streams from Precision Livestock Technology (E.g. sensors) and exploring their application in farm environments to inform decision-making in livestock production systems.