Browsing by Author "Ser, G."

Now showing 1 - 8 of 8

Comparing Covariance Structures Using Different Optimization Techniques in Glmm on Some Sexual Behaviors of Male Lambs
(Pakistan Agricultural Scientists Forum, 2013) Ser, G.; Yesilova, A.; Yilmaz, A.
This study is concerned with use of generalized linear mixed models (GLMM) to analyse the repeated measurements based on count data obtained from the sexual behaviors of male lambs. A combination of different optimization techniques and covariance structures were applied to four constructed models. These models were defined in terms of random effect specifications. Therefore, residuals was assumed to be random (Model A), intercept assumed to be random effect (Model B), time (slope) assumed to be random effect (Model C) and both intercept and time assumed to be random effects (Model D). Five different techniques quasi-newton (QUANEW), newton-raphson (NEWRAP), trust region (TRUREG), newton-raphson ridge (NRRIDG) and double-dogleg (DBLDOG) optimization techniques were used for analyzing these models. Three different covariance structures compound symmetry (CS), unstructured (UN) and first-order autoregressive (AR(1)) were used. In conclusion, based on likelihood criteria, the Model A with CS structure outperformed other models for the repeated measurement data of sexual behavior characteristics.
Determining the Best Model With Deep Neural Networks: Keras Application on Mushroom Data
(Centenary University, 2019) Ser, G.; Bati, C.T.
This study was conducted to reveal the best classifying model with deep neural networks. For this purpose, 20 different candidate models of optimization method (Sgd, Adagrad, Rmsprop, Adam and Nadam), activation function (Tanh and ReLU) and combinations of neurons were studied. By comparing the performance of these candidate models, the best model for classification was determined. The present results indicated that the performance of the models varied according to the parameters, the most successful model has 64 neurons in the hidden layer, the activation function was ReLU and the Rmsprop was used as the optimization method (92% accuracy). In addition, it was determined that the model with the lowest success rate was the model with 32 neurons, ReLU activation function and Sgd optimization method (70% accuracy). Also considering all results; Rmsprop, Adam and Nadam optimization methods were found to be more successful than the other two methods and ReLU activation function produced more successful results than Tanh. © 2019, Centenary University. All rights reserved.
Effects of Data Augmentation Methods on Yolo V5s: Application of Deep Learning With Pytorch for Individual Cattle Identification
(Centenary University, 2023) Bati, C.T.; Ser, G.
In this paper, we investigate the performance of the YOLO v5s (You Only Look Once) model for the identification of individual cattle in a cattle herd. The model is a popular method for real-time object detection, accuracy, and speed. However, since the videos obtained from the cattle herd consist of free space images, the number of frames in the data is unbalanced. This negatively affects the performance of the YOLOv5 model. First, we investigate the model performance on the unbalanced initial dataset obtained from raw images, then we stabilize the initial dataset using some data augmentation methods and obtain the model performance. Finally, we built the target detection model and achieved excellent model performance with an mAP (mean average precision) of 99.5% on the balanced dataset compared to the model on the unbalanced data (mAP of 95.8%). The experimental results show that YOLO v5s has a good potential for automatic cattle identification, but with the use of data augmentation methods, superior performance can be obtained from the model. © 2023, Centenary University. All rights reserved.
Evaluation of Beagle Genotype Imputation Method and an Application
(Centenary University, 2017) İzdeş Baransel, S.; Ser, G.
This study describes the genotype imputation method using the Beagle program and was completed with the aim of evaluating the imputation performance in three different scenarios. A population-based imputation method, the Beagle program uses linkage disequilibrium (LD) information between missing single nucleotid polimorfizm (SNP) and observed SNPs, the local haplotype cluster (Browning) model to offer high rates of imputation accuracy for estimation of deficient genotypes. With this aim, the study used 1356 SNP data from the 22nd chromosome of 191 individuals in the 1000 Genome project, cutting the dataset at random rates of 20%, 50% and 70% to create three different scenarios. The Allelic-R2 values obtained as a result of the three scenarios were larger than 90% and imputations with high degree of accuracy were made. In this study the differences in the size of the reference datasets in the Beagle program were not identified to have a clear effect on imputation accuracy. In conclusion, in terms of results obtained from samples with different sizes using the Beagle program, imputations with high accuracy could be made. © 2017, Centenary University. All rights reserved.
Evaluation of Overdispersed Data Set by Using Generalized Linear Mixed Model Approach
(Centenary University, 2016) Ser, G.; Yeşilova, A.
The aim of this study is to investigate the problem of overdispersion frequently observed in the data sets having Poisson distribution, in which the variation is larger than the mean. Data taken Turkish Statistical Institute (TUIK) covered between 2010 and 2015 were consist of kids reared in eighteen city of Turkey. Three different model algorithm are generated in the generalized linear mixed model approach to eliminate the problem of overdispersion in the data set. The study was conducted in two steps. In the first step, we specified the model algorithm showing overdispersion case. In the second step, however, we used the two model algorithm to overcome the elimination of this overdispersion problem. The standard errors estimated in the case of overdispersion were smaller than the case where overdispersion was eliminated. Nevertheless, in case of overdispersion in the data set, it was determined that there were statistically significant differences among factors of years for population size of kid (P<0.0001), whereas these differences among years for population size of kids were not significant when overdispersion were eliminated statistically. Consequently, the present results showed that overdispersion in data set can led to important misunderstandings, if the case of this overdispersion that data sets is ignored. In generalized linear mixed model approach, as an alternative, the use of negative binomial distribution instead of Poisson distribution or adding the random effects under Poisson distribution assumption in model algorithms occurred presents the effective alternative solutions to overcome the overdispersion case. © 2016, Centenary University. All rights reserved.
Evaluation of Predictive Ability of Two Artificial Neural Network Algorithms and Multiple Regression Model for Meat Quality Traits Affected by Pre-Slaughter Factors
(Pakistan Agricultural Scientists Forum, 2021) Ser, G.; Bati, C. T.; Arik, E.; Karaca, S.
Recently, Artificial Neural Network (ANN) has been developed as an alternative to classical statistical methods in animal production. The methods can do classification or prediction by analyzing the information in the data set with the help of the neural network without requiring any preconditions (for example, distribution of data, non-linear data, highly correlated variables, etc.). In this context, we hypothesized that ANN, which is not only used in large and complex data sets but also estimates better in small data sets compared to classical statistical methods. The ability of ANN of Bayesian Regularization (BR-ANN) or Levenberg-Marquardt (LM-ANN) algorithms and Multiple Regression (MR) model to predict meat quality traits were assessed in a comparative study. The multilayer ANN algorithms obtained prediction data of meat quality measurements from pre-slaughter information using 1, 2, 4, 6 and 8 neurons in the hidden layer applied 10 times for each model. The performance of the methods was assessed according to the coefficient of determination (R-2) criteria, root mean squared error (RMSE) and residual prediction deviation (RPD). The comparison of the findings of BR-ANN and ML-ANN algorithms showed a similar ability to predict meat quality traits (error of prediction and R-2 values between 0.32-2.72 and 0.19-0.49, respectively). However, MR model predictions had lower performance than ANN algorithms, resulting in a wider error of prediction interval (0.4-3.44) and low R-2 (0.16-0.44). The RPD for meat quality traits was fair for BR-ANN and LM-ANN but was poor for MR. Based on our results, the ANN algorithms produced more reasonable prediction values than the MR model. ANN algorithms can be used as an acceptable alternative method for simple physical measurements of meat quality. ANN algorithms can be used reliably in small data sets.
Modelling Overdispersed Seed Germination Data: Xgboost's Performance
(Pakistan Agricultural Scientists Forum, 2023) Ser, G.; Bati, C. T.
Depending on the extent of variability in germination count data, the problem of overdispersion arises. This problem causes significant problems in estimation. In this study, gradient boosting algorithms are used as a new approach to support precision agriculture applications in estimating overdispersed germination counts. The database consisting of germination count data of weed (Amaranthus retroflexus L. and Chenopodium album L) and cultural plants (Beta vulgaris L. and Zea mays L.) with white cabbage seedlings, known for their allelochemical effects, was created. Accordingly, gradient boosting (GB) and extreme gradient boosting (Xgboost) algorithms were first developed for default values to estimate the germination counts of each plant; then, different combinations of hyperparameters were created to optimize the performance of the models. Root mean square error (RMSE), mean poisson deviation (MPD) and coefficient of determination (R2), were used as the statistical criteria for evaluating the performance of the above algorithms. According to the experimental results, the Xgboost algorithm showed superior performance compared to GB in both the default and hyperparameter combinations in the germination counts of A. retroflexus, C. album, B. vulgaris and Z. mays (RMSE: 0.725-2.506 and R2: 0.97-0.99). Our results indicate that the Xgboost made successful predictions of germination counts obtained under experimental conditions. Based on these results, we suggest the use of Xgboost optimal models for larger count data in precision agriculture.
Using the Poisson and Negative Binomial Regression Modeling of Zooplankton Aquatic Insect Count Data
(Centenary University, 2017) Erdinç, S.; Yeşilova, A.; Ser, G.
The aim of this study was to use for Poisson and negative binomial regressions in the modelling of zooplankton aquatic insect counts. Poisson regression is frequently used to analyze for dependent variable based on count data. In data sets, generally overdispersion is seen. In such cases, applying Poisson regression causes biased parameter estimations and standart errors. When there is overdispersion in data set, it is better to use negative binomial regression model. In negative binomial regression model, parameter estimations are obtained by considering the effect that stems from overdispersion. The overdispersion and zero-inflated parameter levels range was obtained to be quite high. All of the dependent variables were statistically significant on zooplankton aquatic insect counts (p<0.01) in the negative binomial regression. In the case of station of Haraba was taken as the reference level, most zooplankton aquatic insect counts was at the Yolcati station (7.972 times more), while at least zooplankton aquatic insect counts was at the Çarpanak station (99.59% less) (p<0.01). In the case of month of september was taken as the reference level, zooplankton density in August was found to be higher compared to other months (p<0.01). Because of the overdispersion had a significant effect, negative binomial regression was better results than the Poisson regression. © 2017, Centenary University. All rights reserved.