Semi-Supervised Learning Combining Phenotyped and Non-phenotyped individuals for Enhancing Prediction in Residual Feed Intake

Abstract Text: Genomic prediction is challenging for residual feed intake (RFI), because the costly measurement on individual feed intake limits the size of reference population. To improve the genomic prediction accuracy in RFI, we introduced self-training model (one of semi-supervised learning strategies), as a novel method combining phenotyped and non-phenotyped individuals. It trained the model using its own predictions on non-phenotyped animals. The results suggested that self-training wrapped around support vector machine increased the prediction accuracy up to 3% using about 1,000 non-phenotyped animals. The improvement increased as the number of non-phenotyped animals included increased, but may approach a plateau. This method can be particularly helpful for enhancing the genomic prediction on new traits such as RFI at the early stage, when the size of reference population is limited. The extension to other traits needs to be further studied.

Keywords:

genomic prediction

semi-supervised learning

residual feed intake

dairy cattle