Study on pig feed conversion ratio prediction based on machine learning

Authors:
  • DPDi Pan
    Di Pan
  • CZCaiyun Zhang
    Caiyun Zhang
  • TLTuowu Li
    Tuowu Li
  • ZZZhe Zhang
    Zhe Zhang
  • QWQishan Wang
    Qishan Wang
  • YPYuchun Pan
    Yuchun Pan
Open Access
Publisher website
Google Scholar
Abstract
Objective This paper explores the feasibility of using partial data to assess the pig feed conversion ratio (FCR) and identify the optimal method to predict the overall pig FCR throughout the entire growth period. Methods Growth performance data, feed intake data, and the pedigree information of a total of 1084 pigs from 2021 to 2022 were used for the study. Four different machine learning methods (Random Forests (RF), Extreme Gradient Boost (XGBoost), Support Vector Regression (SVR), and linear regression (LR)) were compared with the baseline method that simply utilized the known portion of FCR as a proxy for the target prediction. The analysis utilized single-week data as well as different ranges of early-stage and late-stage growth period data. The Pearson correlation coefficient between the overall pig FCR estimated breeding values (EBVs) and the predicted EBVs was used as an indicator to evaluate the prediction accuracy under different scenarios. Results The correlation between the FCR of each week and the entire duration FCR was weak, and the entire duration FCR may be affected by a combination of factors rather than just a single week. When using data from a single week to predict the entire duration FCR, RF showed the most stability and accuracy of all models. When using data from the first and second halves to predict the FCR over different time horizons, the predictive performance of all models declined as the prediction span increased. Among them, RF and SVR had better prediction performances across the time windows. Notably, when using the latter half of the data for prediction, the LR model had the best prediction performance. Conclusion For the mid-term measurements, when only accuracy is considered, using the latter half of the data is recommended due to its higher representativeness. When the available data cover less than half of the growth period, RF performs better for the prediction, regardless of whether early or late data are used. In contrast, when the measurement data extend beyond the halfway point, a linear model is preferable to predict the overall FCR, as it better captures the representativeness of incomplete FCR data for the overall assessment.

This publication has 15 references indexed in Scilit: