Machine learning based personalized promotion strategy of piglets weaned per sow per year in large-scale pig farms

Abstract

Abstract Background The purpose of this study was to analyze the relationship between different productive factors and piglets weaned per sow per year (PSY) in 291 large-scale pig farms and analyze the impact of the changes in different factors on PSY. We chose nine different algorithm models based on machine learning to calculate the influence of each variable on every farm according to its current situation, leading to personalize the improvement of the impact in the specific circumstances of each farm, proposing a production guidance plan of PSY improvement for every farm. According to the comparison of mean absolute error (MAE), 95% confidence interval (CI) and R2, the optimal solution was conducted to calculate the influence of 17 production factors of each pig farm on PSY improvement, finding out the bottleneck corresponding to each pig farm. The level of PSY was further analyzed when the bottleneck factor of each pig farm changed by 0.5 standard deviation (SD). Results 17 production factors were non-linearly related to PSY. The top five production factors with the highest correlation with PSY were the number of weaned piglets per litter (WPL) (0.6694), mating rate within 7 days after weaning (MR7DW) (0.6606), number of piglets born alive per litter (PBAL) (0.6517), the total number of piglets per litter (TPL) (0.5706) and non-productive days (NPD) (− 0.5308). Among nine algorithm models, the gradient boosting regressor model had the highest R2, smallest MAE and 95% CI, applied for personalized analysis. When one of 17 production factors of 291 large-scale pig farms changed by 0.5 SD, 101 pig farms (34.7%) can increase 1.41 PSY (compared to its original value) on average by adding the production days, and 60 pig farms (20.6%) can increase 1.14 PSY on average by improving WPL, 45 pig farms (15.5%) can increase 1.63 PSY by lifting MR7DW. Conclusions The main productive factors related to PSY included WPL, MR7DW, PBAL, TPL and NPD. The gradient boosting regressor model was the optimal method to individually analyze productive factors that are non-linearly related to PSY.

Keywords

• PSY
• Machine learning
• Correlation coefficient
• Gradient boosting regressor model
• Personalized improvement