Predictions of dairy related phenotypes using milk mid infrared spectral data under different degree of information availability

Through this thesis work, we built equations from milk mid-infrared (MIR) spectra to assess 3 phenotypes estimated at cow level (bodyweight (BW), bodyweight change (BWC), and dry matter intake (DMI)) as well as 2 phenotypes estimated on a herd basis (grazing intensity (GRASS) and gradient of production intensification (GPI)). Predictive models have limitations as they cannot fully explain reference variability. Factors influencing the model's predictive capacities include the choice of algorithm, model calibration, and data quality and variability. In addition, care must be taken to ensure that data is not corrupted during collection, processing, transformation, or transfer. In our case, the reference data was a crucial aspect that required special attention as some of it was missing or suspected of being unrepresentative. This limitation prompted us to develop approaches beyond traditional supervised modeling methods. To address this issue, we categorized the equations into three groups based on their degree of information availability: complete, semi-complete, and information shortage references. Each category implied specific directions in the modeling strategy. We had enough reference records to use algorithms with supervised learning for the complete category in which we had the BW, BWC, and DMI phenotypes. This pipeline allowed obtaining a prediction error of around 50kg for BW and around 3 kg for DMI. The information shortage category included the GRASS phenotype. We had no reference sample for modeling, requiring unsupervised processes as the learning could not be based on optimizing a cost function. Moreover, the algorithm's direction had to be externally validated to ensure consistency with observations made elsewhere, as the model could not rely on a structure of knowledge based on observations. The obtained predictions observed the grazing evolution managed by the farmers. For the semi-complete category, which included the GPI phenotype, we had under-representative reference samples regarding the potential spanning of the population, requiring a combination of supervised and unsupervised techniques to train and validate the model. However, this approach helped overcome the limitation of a lack of reference values and allowed for the development of more robust and accurate predictive models. The GPI phenotype distinguished extensive and intensive farmers with moderate accuracy. In conclusion, this thesis showed the possibility of designing a different model framework following the degree of reference information. The built equations developed in this thesis will also be helpful for farmers to help them in their daily management and breeding decisions.