Traditional mixed linear modelling versus modern machine learning to estimate cow individual feed intake

Three modelling approaches were used to estimate cow individual feed intake
(FI) using feeding trial data from a research farm, including weekly recordings
of milk production and composition, live-weight, parity, and total FI.
Additionally, weather data (temperature, humidity) were retrieved from the
Dutch National Weather Service (KNMI). The 2014 data (245 cows; 277
parities) were used for model development. The first model (M1) applied an
existing formula to estimate energy requirement using parity, fat and protein
corrected milk, and live-weight, and assumed this requirement to be equal to
energy intake and thus FI. The second model used ‘traditional’ Mixed Linear
Regression, first using the same variables as in M1 as fixed effects (MLR1), and
then by adding weather data (MLR2). The third model applied Boosted
Regression Tree, a ‘modern’ machine learning technique, again once with the
same variables as M1 (BRT1), and once with weather information added
(BRT2). All models were validated on 2015 data (155 cows; 165 parities) using
correlation between estimated and actual FI to evaluate performance. Both
MLRs had very high correlations (0.91) between actual and estimated FI on 2014
data, much higher than 0.46 for M1, and 0.73 for both BRTs. When validated on
2015 data, correlations dropped to 0.71 for MLR1 and 0.72 for MLR2, and
increased to 0.71 for M1 and 0.76 for both BRTs. FI estimated by BRT1 was, on
average, 0.35kg less (range: -7.61 – 13.32kg) than actual FI compared to 0.52kg
less (range: -11.67 – 19.87kg) for M1. Adding weather data did not improve FI
estimations.