Prediction of Body Weight of Turkish Tazi Dogs using Data Mining Techniques: Classiication and Regression Tree (CART) and Multivariate Adaptive Regression Splines (MARS) Senol Celik 1, * and Orhan Yilmaz 2 1 Department of Animal Science, Faculty of Agriculture, Bingol University, Bingol, Turkey 2 Vocational High School of Posof, Ardahan University, Posof, Ardahan, Turkey Article Information Received 02 April 2017 Revised 27 July 2017 Accepted 27 September 2017 Available online 26 February 2018 Authors’ Contribution OY provided the data and SC made statistical calculations and wrote the article. OY reviewed the article. Key words CART, MARS, Body size, Body weight. Body weight of dogs is crucial trait for breeding, racing and housekeeping. However, variables and factors that correctly estimate this trait are lacking. Here, we applied classiication and regression tree (CART) and multivariate adaptive regression splines (MARS) approaches to estimate the most important variables in predicting the body weight of Turkish Tazi dogs. Using various body measurements, the CART algorithm proposed that withers height (WH), abdominal width (AW), rump height (RH) and chest depth (CD) can signiicant effect the body weight. Quantitatively, it was identiied that values of WH > 62.500 cm and RH > 67.500 cm can positively correlated with the highest body weights. On the other hands, MARS model’s inding showed that the dogs which had the values of WH > 51 cm can be expected to have the highest body weights. The calculated model evaluation criteria of CART algorithm was R2=0.6889, Adj. R2=0.6810, r=0.830, SD ratio=0.5549, RMSE=1.1802, RRMSE=6.3838 and ρ=3.4884, respectively, whereas the calculated model evaluation criteria of MARS method were R2=0.9193, Adj. R2=0.8983, r=0.9588, SD ratio=0.2840, RMSE=0.6041, RRMSE=3.2635 and ρ=1.6661. Taken together, the MARS algorithm appeared to be eficient compared to CART algorithm since the MARS algorithm’s goodness- of-it criteria yielded better results. Using MARS algorithm, the body weight of animals (dogs) can be predicted and exploited in different performances. INTRODUCTION T urkish Tazi dogs have been bred in Turkey for centuries (Tepeli and Cetin, 2003). Recently, this breed is mainly raised in provinces of Konya and Sanliurfa (Yilmaz and Ertugrul, 2011). The Turkish Tazi (Sight Tazi) is a hunting breed and has been used for racing and hunting for decades (Serpell, 1996; Palika, 2007; Yilmaz, 2008). The average weight of mature Turkish Tazi dog is 19.0±0.25 kg for males, and 17.8±0.28 kg for females, while its average shoulder height is 3.1±0.47 cm for males, and 61.0±0.48 cm for females (Yılmaz, 2008; Yılmaz and Ertuğrul, 2011). The average withers height of a Turkish Tazi dog is 70 cm, and its average body weight is 24 kg (Tepeli, 2003). These body parameters including light weight and slim body structure favor the hunting capabilities of this breed. Several data mining practices are being practiced in various ields of livestock to estimate the body, which is one of the most important traits for selection. The multivariate * Corresponding author: senolcelik@bingol.edu.tr 0030-9923/2018/0002-0575 $ 9.00/0 Copyright 2018 Zoological Society of Pakistan adaptive regression splines (MARS) has been proposed for livestock, however, this approach has not yet been implemented to predict the body weight in dog husbandry. The data-mining MARS has been applied for the detection of artiicial insemination problems in cattle (Grzesiak et al., 2010). Of the different prediction methodologies applied so far, following variables have the greatest contribution to the determination of an insemination class: the length of calving interval, body condition score, pregnancy duration, artiicial insemination age, milk yield, milk fat, protein content, and lactation number (Grzesiak et al., 2010). CART (Classiication and Regression Tree) and NBC (Naïve Bayesian Classiier) methods, applied for the detection of cows with conception problems, also yielded useful results. Applying these approaches, most important input variables for CART included the duration between calving and conception, calving interval and the difference between the mean body condition score and condition score at artiicial insemination (Grzesiak et al., 2011). Topal et al. (2010) have identiied factors affecting birth weight in Swedish red cattle using regression tree analysis. According to their obtained outcome, the most signiicant variables affecting birth weight were birth type, birth ABSTRACT Pakistan J. Zool., vol. 50(2), pp 575-583, 2018. DOI: http://dx.doi.org/10.17582/journal.pjz/2018.50.2.575.583