Use of Multivariate Adaptive Regression Splines (MARS) and Classification and Regression Tree (CART) Data Mining Algorithms to Predict Live Body Weight of Tswana Sheep

TL;DR

This study uses data mining algorithms to predict the live body weight of Tswana sheep based on body measurements, helping farmers without scales make better decisions.

Contribution

The study introduces MARS as a more effective algorithm than CART for predicting body weight in Tswana sheep using heart girth measurements.

Findings

Heart girth has a strong and significant correlation with body weight in Tswana sheep.

MARS outperformed CART in predicting body weight based on heart girth measurements.

Using heart girth as a predictor can improve body weight and meat production in Tswana sheep.

Abstract

Indigenous Tswana sheep play an important role in household food security and socio-cultural obligations for resource-poor farmers. For these farmers, weighing scales are not readily accessible; hence, sales of sheep are mostly dependent on physical appearance. Therefore, as was the purpose for this study, determining the relation between body weight (BW) and other linear body measurements and using multivariate adaptive regression splines (MARS) and classification and regression tree (CART) data mining algorithms to predict BW in Tswana sheep is key for resource poor farmers, particularly in places where there is a lack of weighing scales. Heart girth (HG) showed strong and significant correlations with BW. However, the correlation does not show the influence of HG on BW; hence, MARS and CART were used to determine the effect of HG on the BW of Tswana sheep. The MARS algorithm was the easier and more precise algorithm to use in predicting BW in Tswana sheep than the CART model. The high correlation of heart girth and body weight could also be used as an indirect selection criterion, as selecting for sheep with larger heart girth would result in a concurrent improvement in body weight in Tswana sheep, leading to increased meat production.

This study was conducted to (i) determine the association between live body weight (BW) and biometric traits, (ii) examine the effect of biometric traits on BW of Tswana sheep using MARS and CART data mining algorithms, (iii) compare the performance of the algorithms and, finally, select the best algorithm for predicting BW in Tswana sheep. BW and sixteen biometric traits were measured from 392 Tswana sheep (males = 85 and females = 307) aged three to four years. Pearson’s correlation coefficients were used to establish the relationship between BW and biometric traits. The goodness of fit criteria were computed to assess the predictive performance of the data mining algorithms and select the best-fit model for predicting BW. The results showed that BW had a positive and significant correlation with heart girth (HG) (r = 0.99); thus, HG was used as a sole predictor of BW. The goodness of fit results indicated that MARS has a higher predictive performance than the CART algorithm, suggesting that the MARS algorithm can be used to predict BW Tswana sheep. These findings are an important statistical tool for the selection and concurrent improvement of useful biometric traits in genetic improvement programs to improve BW in Tswana sheep.

Figures