Multi-Model Lactation Curve Prediction Based on Triple-Cross-Missing Algorithm

Genomic prediction has revolutionized dairy cattle breeding, significantly advancing genetic progress through improved accuracy and reduced generation intervals. The success relies heavily on the availability of accurate lactation phenotypes, often estimated from test-day milk yields (TDMY). In this study, a triple-cross-missing algorithm is proposed to simulate data missing in real scenarios, and the performance of five models in filling the missing data is evaluated.The results showed that, while BP and LR maintained balanced performance across different missing rates, the neural network models (MLP and RNN) provided superior accuracy. In contrast, Wood lactation curve predictions suffered from a relatively larger loss of accuracy, especially during early lactation stages, due to its limited flexibility. The RNN model, which incorporates a recursive causal structure, showed the potential to maintain higher accuracy and lower bias than MLP. Increased missing rates led to larger mean squared errors and smaller correlations between actual and imputed TDMY. More significant biases were observed at a higher missing rate ($\gt 50 \%$), negatively affecting projected lactation yields. The implications on lactation yield estimation and genetic evaluations are also discussed.