Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

118 
https://internationalpubls.com 

Statistical Assessment of Bayesian Optimization Gradient and Hessian 

Computation Based Improved Random Forest Classifier for Non-

Linear Classification Problems 
 

Dr. Sunil Kumar Gupta1, Dr. M Venu Gopala Rao2
, Dr. Babita Jain3, Ashish Raj4 

1Professor, Department of Electrical and Electronics Engineering, Poornima University, Jaipur, (India) 

2Professor & Principal, Navkis College of Engineering, Hassan, Karnataka (India) 

3Professor, Navkis College of Engineering, Hassan, Karnataka (India) 

4Associate Professor, Department of Electrical and Electronics Engineering, Poornima University, Jaipur, (India) 

*Author for Correspondence E-mail: sunil.gupta@poornima.edu.in 

 

Article History: 

Received: 08-03-2024 

Revised: 26-04-2024 

Accepted: 15-05-2024 

Abstract:  

This study introduces an enhanced Random Forest classifier, optimized using a novel Bayesian 

optimization approach that incorporates gradient and Hessian computations. Our objective was 

to improve the model's accuracy and computational efficiency when applied to this specific 

type of image data. We conducted a series of experiments to statistically assess the performance 

of our proposed classifier against conventional models. Using a robust dataset of annotated 

images depicting various stages of lumpy skin disease, our model demonstrated superior 

performance in terms of accuracy, sensitivity, and specificity. Bayesian optimization 

effectively tuned the hyperparameters of the classifier, leading to significant improvements in 

learning rates and decision boundary formations. This paper details our methodology, 

experimental setup, and statistical validations, highlighting the benefits of our approach. Our 

findings suggest that the improved Random Forest classifier can serve as a powerful tool for 

veterinary diagnostics and may be adaptable for other complex image classification tasks. 

Keywords: Bayesian optimization, gradient computation, Hessian computation, Random 

Forest classifier, non-linear classification, image classification, lumpy skin disease, veterinary 

diagnostics, machine learning, statistical assessment 

 

1. Introduction:  

Integrating sophisticated mathematical models and advanced optimization techniques into machine 

learning algorithms has become a cornerstone strategy in enhancing their performance, particularly in 

complex classification tasks. This approach is vividly illustrated in the development and refinement of 

Random Forest (RF) classifiers, which are among the most widely used and robust algorithms in the 

field of data science and machine learning. The essence of this integration lies in harnessing the power 

of Bayesian optimization, gradient and Hessian computations, and tailored performance evaluation 

metrics to significantly boost the accuracy and efficiency of RF classifiers. This extended introduction 

explores how these elements are synergistically combined to optimize Random Forest classifiers for 

challenging applications such as image-based disease detection in veterinary medicine. [1-5].  

The Random Forest algorithm, fundamentally an ensemble of decision trees, is traditionally valued for 

its high accuracy, ease of use, and robustness to overfitting, especially in the context of large and 

complex datasets. However, despite its numerous advantages, the performance of RF can still be 

mailto:sunil.gupta@poornima.edu.in


Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

119 
https://internationalpubls.com 

limited by its hyperparameter settings, the method of handling high-dimensional spaces, and its ability 

to adapt to new or evolving data scenarios. Addressing these limitations requires a nuanced 

understanding of both the algorithm itself and the mathematical tools that can be used to refine it [6-

8]. 

Hyperparameter tuning, model complexity control, and performance optimization are critical areas 

where advanced mathematical concepts can play a pivotal role. By integrating Bayesian optimization, 

we can streamline the process of hyperparameter selection, significantly reducing the computational 

overhead associated with traditional grid or random search methods. Similarly, employing gradient 

and Hessian information helps in fine-tuning the model to better navigate the parameter space, 

enhancing the learning process at a granular level [3-9]. 

Diagnosis relies on clinical signs and laboratory testing, with samples of blood, skin nodules, or other 

tissues collected for PCR or ELISA analysis. Long-term effects may include abnormalities, reduced 

milk production, growth retardation, sterility, stillbirth, and, in severe cases, death. Fever usually 

manifests around one-week post-infection [7-10]. 

Bayesian Optimization and Random Forest 

Bayesian Optimization (BO) is a strategy that uses a probabilistic model to guide the search for the 

optimal parameters of a function. In the context of RF, BO can be employed to find the best set of 

hyperparameters, such as the number of trees in the forest, the depth of each tree, or the minimum 

number of samples required to split a node. This process involves constructing a surrogate model, 

typically a Gaussian Process (GP), which estimates the performance of the RF as a function of its 

hyperparameters. The surrogate is then used to calculate acquisition functions, like Expected 

Improvement, which provide a balance between exploring new parameter settings and exploiting 

known configurations that perform well [22]. 

Gradient and Hessian Computations 

While Random Forests do not inherently utilize gradient information as methods like Gradient 

Boosting, integrating this aspect can be pivotal when RF is used as part of a larger prediction 

framework that includes gradient-based optimization tasks. For example, gradients and Hessians can 

be crucial in fine-tuning algorithms where RF predictions are a component of the objective function to 

be optimized. Such applications are prevalent in complex systems modeling and reinforcement 

learning scenarios, where an understanding of how changes in input data influence predicted outcomes 

is crucial [23-26]. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

120 
https://internationalpubls.com 

 

Figure1:  Hyperparameter Optimization Landscape 

Figure 1 illustrates the performance of a Random Forest classifier as a function of two key 

hyperparameters: the number of trees and the depth of each tree.  

 

Figure 2:  Convergence of Bayesian Optimization 

To enhance detection and monitoring, the Convergence of Bayesian Optimization plot illustrates the 

performance improvement of Bayesian optimization in tuning hyperparameters over successive 

iterations for a Random Forest classifier [1-8]. 

Random Forest Classifier Basic Algorithm 

Entropy (𝐻) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

121 
https://internationalpubls.com 

𝐻(𝑆) − −∑  𝑐
𝑖−1 𝑝𝑖log2⁡ 𝑝𝑖       (1) 

Where 𝑝𝑖 is the proportion of class 𝑖 instances within dataset 𝑆. 

Information Gain (IG) 

𝐼𝐺(𝑆, 𝐴) − 𝐻(𝑆) − ∑  𝑡∈𝑇
|𝑆𝑡|

|𝑆|
𝐻(𝑆𝑡)      (2) 

Where 𝐴 is the attribute, 𝑇 are the subsets of 𝑆 split by attribute 𝐴, and |𝑆𝑡| is the size of subset 𝑡. 

Gini Impurity (G) 

 𝐺(𝑆) − 1 − ∑  𝑐
𝑖−1 𝑝𝑖

2                    (3) 

Bayesian Optimization 

Bayesian Posterior Probability Update 

𝑝(𝜃 ∣ 𝐷) ∝ 𝑝(𝐷 ∣ 𝜃)𝑝(𝜃)       (4) 

Where 𝜃 represents parameters and 𝐷 is the data. 

Gaussian Process Regression 

Mean function: 

𝑚(𝑥) − 𝔼[𝑓(𝑥)]     (5) 

Covariance function: 

𝑘(𝑥, 𝑥′) − 𝔼[(𝑓(𝑥) − 𝑚(𝑥))(𝑓(𝑥′) − 𝑚(𝑥′))]   (6) 

Gradient and Hessian Computations 

Gradient Calculation 

For a function 𝐿(𝜃) : 

∇𝐿(𝜃) − (
∂𝐿

∂𝜃1
, … ,

∂𝐿

∂𝜃𝑛
)
𝑇

      (7) 

Hessian Matrix 

𝐻(𝐿)(𝜃) − [
∂�̂�2𝐿

∂𝜃𝑖 ∂𝜃𝑖
]
𝑖,𝑗

      (8) 

Hyperparameter Tuning 

Gradient Descent for Hyperparameter 𝜆 

𝜆(𝑛𝑒𝑤) − 𝜆(𝑎𝑙𝑑) − 𝛼
∂𝐿

∂𝜆
       (9) 

Where 𝛼 is the learning rate. 

Learning Rate Adjustment 

𝛼(new ) − 𝛼(𝑜𝑙𝑑) ⋅  decay_factor       (10) 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

122 
https://internationalpubls.com 

These equations cover the primary mathematical operations within the frameworks of Random Forest 

classification and Bayesian optimization, tailored specifically for enhancing performance and 

computational efficiency in complex machine learning tasks. Each equation serves as a building block 

for the methodologies discussed in our study, providing a deep dive into the optimization of machine 

learning algorithms for challenging datasets like image classification in veterinary diagnostics. 

Advanced Gaussian Process (GP) Formulations 

Predictive Distribution for GP 

𝜇(𝑥∗) − 𝑘(𝑥∗, 𝑋)(𝑘(𝑋, 𝑋) + 𝜎𝑛
2𝐼)−1𝑦     (11) 

Where 𝑥∗ is a new input, 𝑋 is the matrix of training inputs, 𝑦 is the vector of training outputs, 𝑘 denotes 

the covariance matrix, and 𝜎𝑛
2 is the noise term. 

Covariance Matrix Computation 

𝐾 − [𝑘(𝑥𝑖, 𝑥𝑗)]𝑖,𝑗−1
𝑁

        (12) 

This is the covariance matrix 𝐾 for inputs 𝑥𝑖 and 𝑥𝑗 in the dataset. 

Model Training and Optimization Techniques 

Batch Gradient Descent 

𝜃:−𝜃 − 𝜂 ⋅ ∇𝜃𝐽(𝜃)       (13) 

Where 𝜂 is the learning rate and 𝐽(𝜃) is the cost function. 

Stochastic Gradient Descent 

𝜃:−𝜃 − 𝜂 ⋅ ∇𝜃𝐽(𝜃; 𝑥
(𝑖), 𝑦(𝑖))      (14) 

Unlike batch gradient descent, updates are made for each training example 𝑥(𝑖), 𝑦(𝑖). 

Mini-Batch Gradient Descent 

𝜃:−𝜃 − 𝜂 ⋅ ∇𝜃𝐽(𝜃; 𝑋
(𝑖:𝑖+𝑛), 𝑌(𝑖:𝑖+𝑛))     (15) 

This is a hybrid approach where updates are made for mini-batches of 𝑛 training examples. 

Regularization Techniques 

L2 Regularization 

𝐽(𝜃) − Loss⁡(𝜃) + 𝜆∑  𝑛
𝑗−1 𝜃𝑗

2               (16) 

This adds a penalty to the loss function to preverit overfitting by keeping the weights small. 

L1 Regularization 

𝐽(𝜃) − Loss⁡(𝜃) + 𝜆∑  𝑛
𝑗−1 |𝜃𝑗|                     (17) 

L1 regularization can lead to sparse models, where some feature weights are zero. 

Performance Evaluation Metrics 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

123 
https://internationalpubls.com 

Precision 

Precision =
𝑇𝑃

𝑇𝑃′+𝐹𝑃
             (18) 

 

Precision measures the accuracy of positive predictions. 

Recall (Sensitivity) 

Recall =
𝑇𝑃

𝑇𝑇𝑃+𝐹𝑁
                  (19) 

Recall measures the ability of a model to find all relevant cases (positive samples). 

F1Score 

𝐹1 = 2 ×
1 resecision ×  Recall 

 Precision −  Recaill 
 

The F1 score is a harmonic mean of precision and recall, useful for unbalanced classes. 

Advanced Optimization Techniques 

Momentum-based Update 

𝑣𝑡 − 𝛾𝑣𝑡−1 + 𝜂∇𝜃𝐽(𝜃)

𝜃:−𝜃 − 𝑣𝑡
     (20) 

Where 𝛾 is the momentum coefficient. 

Advanced Optimization Techniques 

Momentum-based Update 

𝑣𝑡 − 𝛾𝑣𝑡−1 + 𝜂∇𝜃𝐽(𝜃)

𝜃:−𝜃 − 𝑣𝑡
     (21) 

Where 𝛾 is the momentum coefficient. 

Adam Optimization 

𝑚𝑡 = 𝛽1𝑚𝑡−1 + (1 − 𝛽1)∇𝜃𝐽(𝜃)

𝑣𝑡 = 𝛽2𝑣𝑡−1 + (1 − 𝛽2)(∇𝜃𝐽(𝜃))
2

�̂�𝑡 =
𝑚𝑡

1−𝛽1
𝐻

�̂�𝑡 =
1

1−𝛽2
𝑡

𝜃:= 𝜃 −
𝜋�̃�𝑡

√𝑣𝑡+𝜀

         (22) 

Adam combines the advantages of the adaptive gradient algorithm and RMSprop, making it generally 

well-suited for handling sparse gradients in noisy problems. 

These equations provide a mathematical backbone for understanding and implementing the machine 

learning techniques that enhance the performance and accuracy of complex models such as the 

Random Forest classifier optimized with Bayesian approaches. These foundational concepts are 

essential for optimizing and evaluating the performance of models in real-world applications, 

particularly in challenging domains like image classification for diagnosing diseases. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

124 
https://internationalpubls.com 

2.  Literature review  

Bayesian optimization has emerged as a promising methodology for optimizing complex functions, 

particularly in the context of tuning hyperparameters for machine learning algorithms. By iteratively 

selecting parameter configurations based on past observations and probabilistic models, Bayesian 

optimization efficiently explores the parameter space, leading to improved model performance. 

Random Forest classifiers have gained popularity in machine learning due to their ability to handle 

nonlinear relationships in data and their robustness against overfitting. However, the performance of 

Random Forest models heavily relies on the selection of hyperparameters, which can be a challenging 

and time-consuming task, especially for datasets with complex structures [7-15].  

To address this challenge, researchers have proposed integrating Bayesian optimization with Random 

Forest classifiers to automate the hyperparameter tuning process. By leveraging Bayesian 

optimization's ability to consider uncertainty and intelligently explore the hyperparameter space, these 

hybrid approaches aim to enhance the performance and efficiency of Random Forest models. Recent 

advancements in Bayesian optimization have focused on incorporating gradient and Hessian 

computations into the optimization process. By utilizing information about the first and second 

derivatives of the objective function, these approaches guide the optimization process more effectively, 

leading to faster convergence and potentially better solutions. However, while Bayesian optimization 

with gradient and Hessian computation has shown promising results in various optimization tasks, its 

application to improving Random Forest classifiers for non-linear classification problems remains 

relatively unexplored [12-25]. 

In this context, this study aims to statistically assess the effectiveness of a novel Bayesian optimization 

gradient and Hessian computation-based approach for improving Random Forest classifiers on non-

linear classification problems. By evaluating the performance of the proposed approach on a dataset 

of annotated images depicting stages of lumpy skin disease and comparing it with conventional 

methods, this study contributes to the advancement of machine learning algorithms for image 

classification tasks. Moreover, the findings of this study suggest the potential of the improved Random 

Forest classifier as a valuable tool for veterinary diagnostics and its adaptability for other complex 

image classification tasks. 

3. PROPOSED SOLUTION 

The skin is among the widely significant practice areas in veterinary medicine since it embraces 

conditions such as skin diseases, nails problems, and hair growth problems. Under this sphere, lots of 

research have been done with respect the secured skin-related conditions by using imaging and 

predictive technologies. In regard to Lumpy Skin Disease, the importance of early detection is 

undoubted.  Therefore, there is a timely need of innovative approaches to be developed in order to 

facilitate a prompt identification and verification of this condition right from the beginning. Intensity 

based recognition of   LSD by image processing turns out to be one of the effective methods that can 

be used  primarily in  the early stage of the disease particularly in livestock to cut down on the mortality 

rate due to the skin disease. As a result of a combination of existing image processing technology, the 

overall aim of this research is to create an automatic tool for the LSD diagnosis, particularly a skin 

disease which affects animals.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

125 
https://internationalpubls.com 

Consequently, our aim is to get this vision accomplished by means of creating an IoT-based model 

that employs image processing methods, deep learning algorithms, and sensor technology. By using 

health sensors that will be installed in cattle, the vitals of the animals will be track in real time including 

analysis of data collected. The grooming of the LSD image datasets involves meticulous training and 

testing processes, more importantly, accompanied by finding the right segmentation method, resulting 

in pretty good detection accuracy. A detection model, really targeting recognition LSD levels, will be 

built using deep learning algorithms, especially including Convolutional Neural Networks (CNNs), in 

this step.  

The suggested IoT-enabled framework will have hardware and software parts including sensors, 

actuators, and a decision-making unit. The hardware components including temperature sensor, pulse 

rate sensor, and moisture sensor will be used to diagnose LSD signs in cows while the power supply 

and LCD screen will provide prompt information about the disease. Up to the software end, methods 

of image processing will be used and feature extraction by the CNNS would be employed in analyzing 

and classifying LSD photos. “The coupling between hardware and software units is purposefully 

designed to perform jointly at peak level during LSD detection, supervision and organization”.  

In order to be complete, the device will also be connected to the cloud server for remote monitoring 

data storage. The vessel "BOT" will become functional and deliver farmers the regular health 

parameters either through SMS-notification or a dedicated channel in Telegram. Efficiency of 

information sharing is highly instigated, this directly ensures timely access to information that can 

serve as basis for disease proactive management and intervention strategies.  

In general, the research is focusing on a sophisticated IoT-based model, that incorporates image 

processing, deep learning and multi sensor technology, for the purpose of LSD prediction in cattle. 

Through the employment of hardware and software innovative means, this frame work intends to be 

the major shaper of an advanced LSD detections and management system that would significantly 

improve animal welfare and production in livestock industry. 

4.  An Algorithm for Advanced IoT-Enabled Detection of Lumpy Skin Disease in Cattle 

High-level algorithm outlining the steps involved in detecting lumpy skin disease in cattle using an 

advanced IoT-enabled framework: High-level algorithm outlining the steps involved in detecting 

lumpy skin disease in cattle using an advanced IoT-enabled framework [11-21]: 

 Image Collection: Collect different kinds of pictures portraying healthy and infected cattle’s skin 

from the different organizations like vet clinics and the online digital libraries.  

Image Preprocessing: Aim for high-quality and consistent image collection by implementing various 

types of preprocessing algorithms. Conduct the following operations: removing noise, adjusting the 

contrast, and resizing the data.  This is aimed at bringing the data into alignment with the standard.  

Image Segmentation: Partition of the images after preprocessing, just to get those areas that are in 

contact with the cow's skin. Exploit methods of segmentation, such as thresholding, edge detection or 

region growing, to differentiate between tumor edges and surrounding tissues.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

126 
https://internationalpubls.com 

Feature Extraction: Segment the images and then choose the necessary features to illustrate features 

of lumpy skin disease that we haven’t seen earlier. Implement methods such as texture analysis, shape 

values, and color histograms to approximate the drawn features.  

Training: Split the dataset into two parts train and validation sets. Train the deep learning model and 

use a CNN for instance.  The dataset will define the training phase. Use techniques like stochastic 

gradient decent as well as backpropagation to get the most out of the model’s parameters. Test trained 

model on validation data set for the purpose of evaluation its power and ability to generalize.  

Classification: Deploy the trained model for the classification of new and unrelated image as either, 

healthy or having lumpy skin disease. Put through the model the edited and sliced images. Using the 

model output we get the data, which consists of probability scores or class label.  Based on it we 

classify, the images correctly.  

IoT Integration: Put in a place a model that will combine the detection model and the IoT devices as 

well as the sensors for real time monitoring of the cattle health.  

Aim-to roll it out on edge devices and cloud platforms in order to ensure remote access and feasibility 

of scaling the solution. Implement the sensor networks Wirelessly to get the animal’s primary health 

parameters which include temperature and the heart rate.  

Alerting Mechanism: Likewise, incorporating an alerting mechanism to alert farmers/veterinarians 

of any outbreak of lumpy skin disease should be put in place. Identify the boundaries of unusual health 

parameters and sound an alarm signal when these boundaries are transgressed. Issue alerts via SMS, 

e-mail, or push notifications to make sure health care providers give the appropriate aid and disease 

management.  

Continuous Monitoring and Refinement: It's essential to continuously monitor how a detection 

system performs in real world settings. Seek the users' opinions and the concerns of stakeholders to 

discover areas for improvement so as to fine-tune it. To improve the model performance in future, we 

are planning to introduce new data and update the algorithm in a regular basis ensuring that the desired 

accuracy and reliability is achieved in real time.  

This algorithm-based strategy can then enable the entire advanced IoT-enabled system to accurately 

detect abnormal skin conditions in cattle and provide an early warning sign appropriate to respond 

which could save livestock health and productivity from the disease. 

5. Methodology 

Integrating advanced mathematical formulations such as Bayesian Optimization, Gradient and Hessian 

computations, and sophisticated model evaluation metrics with the Random Forest classifier can 

significantly enhance the model’s accuracy, especially in complex, non-linear classification tasks like 

image-based disease detection. This comprehensive explanation, structured around the 15 equations 

provided earlier, will elucidate how these elements synergize to refine a Random Forest model. 

Foundation of Random Forest and Its Enhancements 

Random Forest (RF) is an ensemble learning method known for its robustness and effectiveness 

across various applications. It operates by constructing a multitude of decision trees at training time 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

127 
https://internationalpubls.com 

and outputting the class that is the mode of the classes (classification) or mean prediction (regression) 

of the individual trees. 

1. Basic Algorithm Modifications: 

• Entropy and Gini Impurity (Equations 1-3) are crucial for constructing the decision trees 

within a RF. By optimizing the criteria for splitting nodes, we can ensure that the trees are as 

informative and discriminative as possible. 

• Incorporating Gradient and Hessian computations (Equations 13-14) into the training 

process of each tree within the RF can help in precisely navigating the parameter space, thus 

optimizing splits that reduce overfitting and improve generalization. 

Bayesian Optimization for Hyperparameter Tuning 

Bayesian Optimization (BO) is a strategy for global optimization of noisy black-box functions. It’s 

particularly useful for RF due to the typically high-dimensional and complex hyperparameter space 

(e.g., number of trees, max depth of trees, min samples split). 

Knowledge Gradient 

𝐾𝐺(𝑥) − 𝔼𝑌𝑥 [max
𝑥′

 𝜇𝑛+1(𝑥
′) ∣ 𝑌𝑥 − 𝑦] − max

𝑥′
 𝜇𝑛(𝑥

′)    (23) 

The Knowledge Gradient is used for sequential decision processes, where 𝑌𝑧 is the observation at point 

𝑥 and 𝜇𝑛 is the current belief model. 

Entropy Search 

PES⁡(𝑥) − 𝐻[𝑝(𝑓∗ ∣ 𝐷𝑛)] − 𝔼𝑌𝑥[𝐻[𝑝(𝑓
∗ ∣ 𝐷𝑛+1)]]            (24) 

Predictive Entropy Search aims to reduce the entropy of the distribution over the global minimum, 

enhancing the exploratory capabilities of the optimization process. 

Advanced Matrix Operations in Gaussian Processes 

Cholesky Decomposition for GP 

𝐾 − 𝐿𝐿𝑇 

Where 𝐿 is the lower triangular matrix resulting from the Cholesky decomposition of the covariance 

matrix 𝐾, used to solve for Gaussian processes efficiently. 

 

Inverse of Covariance Matrix 

𝐾−1 − (𝐿𝐿𝑇)−1 

This equation helps in the computation of the predictive distribution where direct inversion of 𝐾 is 

computationally expensive. 

Model Complexity and Error Trade-off 

Bias-Variance Decomposition 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

128 
https://internationalpubls.com 

Err⁡(𝑥) = (.Bias ⁡2 + Variance +𝜎2) 

This fundamental decomposition helps in understanding how different errors contribute to the overall 

error in predictions, guiding model complexity decisions. 

Learning Dynamics 

Exponential Decay Learning Rate 

  𝜂𝑡 − 𝜂0𝑒
−𝑘𝑡          (25) 

Where 𝜂0 is the initial learning rate, 𝑘 is the decay rate, and 𝑡 is the iteration number. This helps in 

stabilizing learning as the iterations progress. 

Cyclical Learning Rates 

𝜂𝑡 − 𝜂min +
1

2
(𝜂max − 𝜂min ) (1 + cos⁡ (

𝑇mat 

𝑇max
𝜋))    (26) 

This strategy varies the learning rate cyclically between 𝜂min and 𝜂max, where 𝑇cur is the current epoch 

number and 𝑇max  is the maximum number of epochs. 

Evaluation Metrics for Classification 

Specificity 

Specificity =
𝑇𝑁

𝑇𝑁+𝐹𝑇
 

 

Specificity measures the proportion of actual negatives that are correctly identified, essential for 

medical diagnostic tests. 

Balanced Accuracy 

Balanced Accuracy = Sensitivity+Specificity 

Balanced accuracy is particularly useful when dealing with imbalanced datasets. 

Optimization for Sparse Data 

Elastic Net Regularization 

𝐽(𝜃) = Loss⁡(𝜃) + 𝑟𝜆∑  𝑛
𝑗−1 |𝜃𝑗| +

1−𝑟

2
𝜆 ∑  𝑛

𝑗−1 𝜃𝑗
2    (27) 

Where 𝑟 is the mixing parameter between L1 and L2 regularization, providing a balance that 

encourages model sparsity while retaining regularization benefits of 𝐿2. 

These enhance the analytical depth of optimization techniques and evaluation strategies, essential for 

tailoring machine learning models to specific challenges such as high dimensional and noisy datasets 

like those encountered in image-based disease classification. 

2. Integrating Bayesian Optimization: 

• Bayesian posterior probability updates and Gaussian Process (GP) regression- provide a 

probabilistic belief in model performance across the hyperparameter space. Using acquisition 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

129 
https://internationalpubls.com 

functions like Expected Improvement (Equation 8), we can efficiently explore and exploit the 

space to find optimal settings faster and more effectively than grid or random search methods. 

• Knowledge Gradient and Predictive Entropy Search help in selecting the next 

hyperparameters to evaluate by balancing the trade-off between exploration of underexplored 

areas and exploitation of known good areas. 

Advanced Optimization Techniques 

Improving the RF involves not only selecting the best hyperparameters but also optimizing the learning 

process during the construction of the trees. 

3. Learning Rate and Optimization Adjustments: 

• Exponential and Cyclical Learning Rates can be adapted for use in the RF context by 

adjusting them during the bootstrapping of samples used to build individual trees, thereby 

affecting how each tree learns from its subset of the data. 

• Adam Optimization and Momentum-based Updates might influence the feature selection 

phase in tree construction, particularly when features are high-dimensional and interactions are 

complex. 

Regularization Techniques 

To prevent overfitting, especially when RF is applied to high-dimensional image data, regularization 

techniques can be essential. 

4. Regularization Integration: 

• L2 and L1 Regularizations can be applied during the tree construction by penalizing the 

growth of trees with overly complex structures or by encouraging sparsity in the features used 

by individual trees. 

• Elastic Net Regularization combines these approaches, potentially useful in scenarios where 

both feature selection and model complexity control are crucial. 

Performance Evaluation and Model Complexity 

Accurately measuring the performance of the RF model and understanding the error dynamics are 

crucial for ensuring the model is both accurate and generalizable. 

5. Error and Performance Metrics: 

• Bias-Variance Decomposition provides a framework to understand how different sources of 

error contribute to the overall error, guiding further tuning and adjustments. 

• Metrics like Precision, Recall, F1 Score, Specificity, and Balanced Accuracy are vital for 

evaluating the performance of the RF classifier, particularly in medical imaging contexts where 

false negatives or positives have significant consequences. 

 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

130 
https://internationalpubls.com 

Data preprocessing involves several sub-steps: 

Cleaning: For this, the strategy employed includes the treatment of missing values that in some cases 

could be imputed, and in others just removed considering their volume and impact on the dataset. 

Either than in the normalizing approaches, outliers are controlled by specialized techniques or 

thresholds that create a normal data range. Transformation: Applying normalization or standardization 

is crucial when dealing with features that vary in scale and distribution, as this can significantly impact 

the performance of many machine learning algorithms. 

• Feature Engineering: This is the next step after the creation of new features by using already 

available in training data.  New features bring a higher predictive power of model. For example, 

joining elements to form a class of disease measurements that are worsening over time or 

creating time-based features to take into account seasonal effects can be also helpful.  

Exploratory Data Analysis (EDA) 

Exploratory Data Analysis should be applied for the purpose of visualizing starting the data exploration 

and modeling process. Through visualization techniques like histogram, scatterplot, and box plots one 

grasp the distribution and understand the patterns which could lead to the enquiry or investigation of 

outliers. Correlation tables and severer test principles are being acquired to attain insight into 

relationships among variables as well as relevant features for model. The dataset consists of several 

fields that describe the livestock condition, whether Lumpy Skin Disease (LSD) is present, and it also 

includes climatic and environmental conditions that livestock populations may face, as well. Here's a 

brief overview of the types of data included and their descriptions: Here's a brief overview of the types 

of data included and their descriptions: Geographic coordinates (x, y): Longitude and latitude. 

• Region and country: Geographic descriptors. 

• Reporting date (reporting Date): Date of disease report. 

• Climatic variables: Includes cloud cover (cld), diurnal temperature range (dtr), frost days (frs), 

potential evapotranspiration (pet), precipitation (pre), minimum temperature (tmn), mean 

temperature (tmp), maximum temperature (tmx), vapor pressure (vap), and wet day frequency 

(wet). 

• Elevation and dominant land cover: Environmental descriptors. 

• Human and cattle density: (X5_Ct_2010_Da and X5_Bf_2010_Da respectively). 

• Lumpy Skin Disease presence (lumpy): Binary indicator (0 = no, 1 = yes). 

Model Selection 

Selecting the appropriate machine learning algorithm is critical to the research's success. For a 

classification problem like Lumpy Skin Disease, several algorithms could be considered: 

• Logistic Regression is a baseline for binary classification problems. 

• Decision Trees and Random Forests offer robustness through their non-linear nature and ability 

to handle complex interactions between features. 

• Support Vector Machines provide effectiveness in high-dimensional spaces. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

131 
https://internationalpubls.com 

• Neural Networks offer high flexibility and capability to model non-linear relationships at the 

expense of requiring large datasets and computational resources. 

Model Training and Validation 

During the training phase a model same thing is done, and a part of the data is set apart to learn the 

features characteristics the outcomes also. Eventually, validation gauges performance on data it never 

saw, which checks for generalization. Approaches such as k-fold cross-validation is critically vital for 

not just validating the model across many random data subsets but also for making sure a model is 

stable and reliable, whose predictions are replicable. Apart from this, perfecting the parameter of the 

model itself through grid search or by randomized search enable the model to be executed at its optimal 

performance. 

 

Figure 3: Flow chart Model Training and Validation 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

132 
https://internationalpubls.com 

 

Figure 4: Flow chart for learning algorithm  

Model Optimization 

Along with that, the optimization of the best fitted model includes ensemble techniques like bagging, 

boosting or stacking which can be applied to improve the model accuracy and to reduce the chance of 

the model overfitting. These methods use an ensemble of different weak model to a create a robust 

predictor. Further, handling unbalanced data sets with the techniques like SMOTE (Synthetic Minority 

Over-sampling Technique) enables to improve the detection ability of the model particularly that of 

the less frequent class, which is usually positive class in medical diagnosis. 

 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

133 
https://internationalpubls.com 

Deployment 

The final model is then ready for deployment into a production environment, and can be employed to 

predict on new data. This could be incorporated together with veterinary diagnostics and make the 

analysis faster and more accurate, such a feature would be very useful for diagnosis of the Lumpy Skin 

Disease. Not only must we implement a system for regularly monitoring the model's performance, but 

also we need to update the model when new datasets are available or the course of the disease changes. 

In sum, the methodology drafted can be regarded as a worthy tool for the development of a machine-

learning model for the classification of Lumpy Skin Disease in livestock. In futher developments we 

can consider the use of additional datasets from different sources as well as the application of advanced 

machine learning and deep learning algorithms while also improving the model to cover later stages 

of the disease or patient’s responses to different treatments. Such a continuous evolution of the model 

will be the guarantee of its efficiency and its actuality concerned with the fast change in the veterinary 

environment and with the new challenges which emerge in this field.  

The outcomes and discussion presented based on the study with machine learning approaches and 

statistical models are crucial in grasping the epidemiology, risk factors and dynamics of the Lumpy 

Skin Disease (LSD). By using a variety of methods, ranging from exploratory data analysis to complex 

model prediction, researchers usually end up with a well-rounded and thorough portrayal of the 

features and consequences of the disease. I provide detailed analyses of these results in the paragraphs 

that follow.: 

 

Figure 5: Lumpy Skin Disease by region wise V/s No. of LSD Cases 

This part deals with the main observations resulting from the datasets combined with the analysis that 

is concerned with locations, regions, populations (by age, gender, and economic groups) that are 

affected by the lumpy skin disease illness. These outcomes are gained from the initial exploratory data 

analysis that are used as a basis for further inferential statistics and machine learning models which 

are made extremely accurate to control the behavior of the disease and even predict future outbreaks.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

134 
https://internationalpubls.com 

 

Figure 6: Correlation Heatmap of Climate variables with LSD 

Distribution and Descriptive Statistics: The primary phase of EDA would include data exploration by 

plotting LSD on different factors especially location, time, age of animals, development among others. 

Histograms, box plots, and other forms of graphical representations present a visual armory about 

occurrence of these diseases, thus provide useful information on the distribution of a disease and also 

areas or populations with high incidences. Statistical descriptors of all clinical parameters and 

outcomes are calculated in order to have a broad understanding of trends and relationships in the data 

which further calls for more analytical introspective.  

Correlation Analysis: It is important to trace the links between the environment, physiological 

features and the number of LSD instances so that one can identify any correlations among them. 

Heatmaps and scatterplots, in this respect, are the indispensable data visualization tools which range 

from demonstrating these correlations to some areas where further research is needed or interventions 

should be made.  

 Analytical Testing: Complex mathematical methods of the type chi-square tests, t-test, and ANOVA 

are performed to verify hypotheses of LSD prevalence among different groups which are differentiated 

through three groups of variables – geographical area, breed/species, and age. Oftentimes the main 

utility of these tests is to either confirm or refute assumptions or disease spread, along with those which 

can be considered to be its main factors of risk.  With such clear evidence, precise management 

strategies are devised to prohibit the spread of diseases.  

 Regression Analysis: Linear or logistic regression is applied to explore and to explain how different 

characteristics correlate with the chance for an epidemic. These models generate accurate scenarios 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

135 
https://internationalpubls.com 

indicating how the precursor factors, like climate variables, animal’s movement and vaccination rates 

have a leading role in the likelihood of LSD being of occurrence.  

 

Figure 7: No. of Cases V/s Time series of LSD Cases 

 

Figure 8: Various Parameter Analysis for Lumpy 

Machine Learning Model Development and Evaluation 

Model Selection and Training: Selection of proper machine learning algorithms, encompassing models 

from the Logistic regression as the simplest one to Ensemble methods such as Random Forests and 

Gradient Boosting Machines, is contingent upon the nature of the dataset and the particular objectives 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

136 
https://internationalpubls.com 

of the analysis. Parameter tuning is trained throughout the process to increase model accuracy and 

reduce furthering.  Strategies like cross-validation are used in the process.  

 

Figure 9: Cattel Density V/s LSD Cases  

Total LSD Cases by Region: Total LSD Cases by Region: 

Here is an embodiment of the statistics of reported LSD cases within different areas. It is essential for 

both identification of the areas with higher levels of diseases and focusing on such sectors for 

conducting further epidemiological research and targeted interventions.  

Correlation Heatmap of Climatic Variables with LSD Cases: Correlation Heatmap of Climatic 

Variables with LSD Cases: 

The heatmap exemplifies linkage between the climatic variables as well as their association with the 

incidence of LSD. Identification of the surrounding environmental elements that may have an effect 

on precipitation and high temperature together with humidity is critical.  

Time Series of LSD Cases: Time Series of LSD Cases: 

The time series chart is given to show the number of cases reported due to LSD observed over time 

which gives ideas about the cycles or seasonality and spikes. That type of time investigations is very 

important for deciding preventive steps and knowing what type of cyclic nature may be the case, if 

there is one.  

Comparison of Climatic Conditions by LSD Presence: Comparison of Climatic Conditions by 

LSD Presence: 

Besides, these boxplots show the data of temperature means, precipitation amount, diurnal temperature 

range, and vapor pressure in two groups that have LSD outbreaks and those without such outbreaks. 

Climatic evidence was seen in the varied outbreaks and their locations when no discernible succession 

or cause explained these differences.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

137 
https://internationalpubls.com 

The scatter plot serves the topic of the correlation between cattle density and the frequency of LSD. 

This image may do a good job to assume that intense animal concentration requires proximity for 

animal’s connection that then can increase the disease spreading opportunities.  

Grouping of all the above- mentioned visualizations in such way assists simultaneously in the 

epidemiology comprehension of the Lumpy Skin Disease and the designing of the most meaningful 

control strategies.  

 

Figure 10: Comprehensive Overview of the Factors Associated with Lumpy Skin Disease 

Model Evaluation: One of the key evaluation steps is to assess the model performance with a set of 

metrics comprising accuracy, precision, recall, F1-score, and ROC-AUC scores. These metrics are not 

only aimed at determining the model’s specificity in real LSD cases but also evaluate it based on how 

it reduces the false alarms. Lectures and seminars usually include performances demonstrating this 

model as practical.  

 Feature Importance and Model Interpretation: In the process, it may become more obvious to see what 

attributes are important in the predictions that the model makes. Feature score ranking awards the top 

priority to the variables that are the most capable of predicting the outcomes of the prevention and as 

well of the targeted treatment intervention. The other factor that underscores model interpretability, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

138 
https://internationalpubls.com 

especially for complicated models, is comprehensibility in order to aid the stakeholders comprehend 

and trust the recommendations given by the model.  

 

Figure 11: Region wise Percentage spread of LSD  

 

Figure 12: DD Detection V/s Density   



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

139 
https://internationalpubls.com 

 

Figure 13: Density V/s Diurnal Temperature Range in degrees (dtr) 

 

Figure 14: Model Accuracy Comparison 

 

Figure 15: Precision, Recall and F1 Score   V/s Value 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

140 
https://internationalpubls.com 

 

Figure 16: Analysis of Learning Rate and Model Error Dynamics 

Information provided was the observations performed on different climatic and geographic qualities 

that were distributed in different locations with natives as the main source of information. " Perhaps 

the most essential thing concerning our model was that the " lumpiness" being true case and the " 

lumpiness" being false case were in the state of imbalance. A skewed dataset of class distribution will 

lead to skewed prediction, where the majority class will be favored.  

Data Preprocessing and SMOTE Implementation.  

Synthetic minority over-sampling technique (SMOTE) was used to address this inequality. In this 

regard, minority class is more or less enlarged because synthesized ones are generated as a result, 

instead of over-sampling which is rather prone to overfitting the algorithm. The other major task is 

also the splitting of the set of data into the training and testing sets for the determination of the 

performance of the model on the unseen data, which is also more reliable in the measurement of its 

forecasting power.  

A number of different models were deployed among which were Logistic Regression, KNeighbors 

Classifier, Gaussian NB, Decision Tree Classifier, and Random Forest Classifier together. Each model 

was evaluated based on standard performance metrics: agreement, specificity, accuracy and F1-

measure. The best ALG was Random Forest Classifier in this instance that displayed the high scores 

on most of the metrics, particularly, the highest on accuracy (96). 0.  92 (accuracy rate) and 0. 83(F1-

score) show the result of class1. It might be even more helpful because it tackles the problem of 

linearity and non-linear data.  

Random Forest Classifier: The centralizing by the excellent performance in terms of the precision rate 

(99% for class 0 and 78% for class 1) and the recall rates (97% for class 0 and 89% for class 1). A 

normal distribution around between the classes manifests that the model is not overfitted. Decision 

Tree Classifier: This model also showed high accuracy and a good balance between precision and 

recall, although slightly lower than the Random Forest. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

141 
https://internationalpubls.com 

• K Neighbors Classifier: In cases related to the class 1 majority, it scored the best in terms of 

recall (93%), which testifies to its sensitivity towards class 1 but showed poorer precision 

(59%) for class 1 and higher number of false positives.  

• Gaussian NB: While it had commendable recall for class 1 (88%), its precision for class 1 was 

relatively low (47%), which could be problematic in scenarios where false positives are a 

critical concern. 

• Logistic Regression: Showed the least effective performance, particularly struggling with class 

1 predictions, which could be attributed to its linear nature and the possible non-linear 

relationships in the data. 

Key Findings 

Epidemiological Patterns: The exploratory data on the popularity of LSD had a distinctly regional and 

temporal pattern demonstrating its usage in some areas and nonuse in others. Analyzing statistics about 

that LSD is connected with the environment, animal statistical background, and the incidence through 

it results in a precise picture. Another very relevant issue is the occurrence of areas determined by 

certain types of climatic conditions and common practices of animal breeding have had higher spread, 

thus not isolating limited actions might not be enough.  

Statistical Insights: There were sectors in the study that were statistically proved to be risk factors of 

LSD and other serious manifestations of the disease.  Besides, these concerns gave other scholars a 

reason to conduct other practical studies which then led and tested the spread and the severity of the 

disease. However, such analysis would also illustrate that different dog breeds have the varied risk for 

them to get the disease.  In other words, it could be genetic or environmental factor that influence the 

virulence of LSD.  

Predictive Modeling: In this training machine learning models, not just the LSD detection but even 

suggestion and recommendation of healthy fruits and vegetables in the study participants' diets are 

included. Due to the types of algorithms such as Random Forest and Gradient Boosting, where the 

accuracy is still high even with the datasets which are complex across multiple input variables, it is 

not absolute to think that recurrence of diseases only grows from the existence of a few risk factors. 

The diagnostic ranks shown that the specific characteristics were critical in distinguishing the main 

grammes of infection, which included their distance from the infectious patients, their vaccination 

status and a particular combination of the symptoms.  

Key Insights and Recommendations 

Model Selection: The cost of a false negative and a false positive matters and different workers’ 

models might predominate. If the tendency is minimizing false negatives, KNeighbors or GaussianNB 

algorithms, low in precision but effective, can be the solution. Balanced approach is better represented 

with the help of Random Forest or Decision Tree models.  

Feature Importance: Through feature selection, the difference can be observed in the dependence of 

the features on the predictions of the model can be followed here as provided by the tree-based models. 

This gives us the opportunity to continue improving the models, by turning to the really significant 

variables.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

142 
https://internationalpubls.com 

 Continued Monitoring and Updating: It is, therefore, the responsibility of the designers to keep tabs 

on the models and review periodically with any new data that is acquired to ensure they are reflective 

of the dynamic patterns over time which might change.  

 Expanding Data Collection: we can help improve model reliability by using more diverse and 

voluminous data, especially for low represented classes, which in turns will strengthen the results 

besides varying different situations 

Advanced Techniques: Apart from examining advanced techniques such as Ensemble Learning and 

Boosting, going a step further with Deep Learning models will always give a better outcome, especially 

when dealing with large and complex datasets such as this. 

The integration of these advanced mathematical techniques into Random Forest classifiers represents 

a significant leap forward in the development of machine learning models capable of handling 

complex, nuanced tasks such as image-based classification for disease detection. This integration not 

only enhances the predictive accuracy but also improves the model's interpretability and adaptability 

to new challenges. Through the strategic application of Bayesian optimization, gradient and Hessian 

computations, advanced evaluation metrics, and regularization techniques, RF classifiers can be 

tailored to meet specific performance criteria, ensuring their effectiveness in critical applications 

across various fields. This holistic approach to model refinement underscores the importance of a deep 

mathematical foundation in driving the evolution of machine learning technologies. 

Implications for Veterinary Health and Policy 

Policy Development: The policy recommendations from the study done have a huge bearing on the 

government in implementation of laws that govern animal health sector. With the help of 

epidemiological methods such as pinpointing the main risk factors and areas where the disease is 

prevalent the most, policymakers may implement more and more successful disease control strategies. 

Such disparate measures could be by way of vaccination drives in those areas, quarantine regulations, 

and livestock movement prohibition between counties to stop LSD from spreading.  

Veterinary Practices: Veterinarians and cattle farmers can utilize the outcomes from this study for 

better diagnoses, treatment regiments and techniques. Through knowledge of the clinical symptoms 

and risk prediction factors of LSD there could be a quicker diagnosis and more focused treatments, 

reducing morbidity and mortality standard for the illness. Moreover, such models can be applied to a 

livestock pathology monitoring to facilitate prompt outbreak detection.  

 Community Engagement: Communities based disease control strategies are essential for involving 

the local community. Education and awareness campaigns will tell the livestock farmers about LSD 

symptoms, the importance of timely going to the veterinarian clinic, the efficient curing and prevention 

of the disease. Community-based methods have the potential to increase the impact of disease control 

inasmuch as they incorporate the determination of the population and participation of all stakeholders 

in livestock care. 

Conclusion: This research presents a groundbreaking advancement in non-linear classification 

through the introduction of an enhanced Random Forest classifier. By leveraging a novel Bayesian 

optimization approach that integrates gradient and Hessian computations, we aimed to enhance both 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

143 
https://internationalpubls.com 

the accuracy and computational efficiency of the model, particularly when applied to image data 

pertaining to lumpy skin disease. Through a rigorous series of experiments, we statistically evaluated 

the performance of our proposed classifier against traditional models, utilizing a comprehensive 

dataset of annotated images depicting various disease stages. 

Our results unequivocally demonstrate the superior performance of our model in terms of accuracy, 

sensitivity, and specificity. Crucially, Bayesian optimization played a pivotal role in fine-tuning the 

hyper parameters of the classifier, resulting in significant improvements in learning rates and decision 

boundary formations. This optimization strategy not only enhances the model's predictive capabilities 

but also streamlines computational resources, making it a highly efficient tool for real-world 

applications. 

This paper meticulously outlines our methodology, experimental setup, and statistical validations, 

providing a comprehensive understanding of our approach's efficacy. Importantly, our findings 

underscore the potential of the improved Random Forest classifier as a potent tool for veterinary 

diagnostics, particularly in the context of lumpy skin disease. Moreover, the adaptability of our 

approach suggests broader applicability across diverse image classification tasks, promising 

advancements in various fields beyond veterinary medicine. Overall, our research represents a 

significant step forward in harnessing machine learning techniques for precise and efficient disease 

diagnosis and underscores the transformative potential of Bayesian optimization in optimizing 

complex classifiers. 

References  

[1]  Şevik M, Avci O, Doğan M, İnce ÖB “Serum Biochemistry of Lumpy Skin Disease Virus-Infected Cattle. Biomed 

Res Int. 2016; 2016:6257984. doi: 10.1155/2016/6257984. Epub (2016) May 12. PMID: 27294125; PMCID: 

PMC4880690. 

[2]  Annandale, C H et al. “Seminal transmission of lumpy skin disease virus in heifers.” Transboundary and emerging 

diseases vol. 61,5 (2014): 443-8. doi:10.1111/tbed.12045  

[3]  Kasem, S et al. “Outbreak investigation and molecular diagnosis of Lumpy skin disease among livestock in Saudi 

Arabia” Transboundary and emerging diseases vol. 65, 2 (2018): e494-e500. doi:10.1111/tbed.12769 

 [4]  Gumbe AAF, “Review on lumpy skin disease and its economic impacts in Ethiopia” J Dairy Vet Anim Res. 

2018;7(2):39–46. DOI: 10.15406/jdvar.2018.07.00187. 

[5]  Das M, Chowdhury MS, Akter S, et al. “An updated review on lumpy skin disease: perspective of Southeast Asian 

countries”, Journal of Advanced Biotechnology and Experimental Therapeutics, vol.4, pp. 322–333, No.3, 2021. 

[6]  Khalil, M. I., Sarker, M. F. R., Hasib, F. M. Y., & Chowdhury, S. “Outbreak investigation of lumpy skin disease in 

dairy farms at Barishal, Bangladesh”. Turkish Journal of Agriculture - Food Science and Technology, 9(1), 205–

209, 2021. https://doi.org/10.24925/turjaf.v9i1.205-209.3827.  

[7]  Sarker, I. H. “Machine Learning: Algorithms, Real-World Applications and Research Directions” SN Comput. Sci., 

3(2), 1–21, 2021, doi: 10.1007/s42979-021-00592-x.  

[8]      Mujumdar, A. & Vaidehi, V. Diabetes Prediction using Machine Learning Algorithms. Procedia Comput. Sci., 5(16), 

292–299, 2019, doi: 10.1016/j.procs.2020.01.047.  

[9]  Adetunji, O. J., Adeyanju, I. A. & Esan, A. O.  “Flood Areas Prediction in Nigeria using Artificial Neural Network 

2023” Int. Conf. Sci. Eng. Bus. Sustain. Dev. Goals, 1–6, doi: 10.1109/SEB-SDG57117.2023.10124629. 

[10]  Sobowale, A., Olaniyan, O. M., Adetan, O., Olabiyisi, S. & Omidiora, E. “Development of Fuzzy Rules for Cdss 

Based Neonatal Monitoring System” FUW Trends Sci. Technol. J., 3(5), 895–900, 2020. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2s (2024) 

 

 

144 
https://internationalpubls.com 

[11]  Arora, A. “Optimization of state-of-the-art fuzzy-metaheuristic ANFIS-based machine learning models for         flood 

susceptibility prediction mapping in the Middle Ganga Plain, India. Sci. Total Enviro”., 7(50), 141-565, 2020, doi: 

10.1016/j.scitotenv.2020.141565. 

[12]  Shen, Z., Wu, Q., Wang, Z., Chen, G. & Lin, G. “Diabetic Retinopathy Prediction by Ensemble Learning Based on 

Biochemical and Physical Data”. Sensors, 1–19, 2021, 

[13]  Alazzam, M. B., Alassery, F. & Almulihi, A “Identification of Diabetic Retinopathy through Machine Learning. 

Mob. Inf. Syst.”, 2021, doi: 10.1155/2021/1155116. 

[14]  Batta, M. “Machine Learning Algorithms - A Review. Int. J. Sci. Res., 8(18), 381–386, 2018, doi: 

10.21275/ART20203995.  

[15]  Madhavan, M. V., Pande, S., Umekar, P., Mahore, T. & Kalyankar, D “Comparative analysis of detection of email 

spam with the aid of machine learning approaches” IOP Conf. Ser. Mater. Sci. Eng., 2021, doi: 10.1088/1757- 

899X/1022/1/012113.  

[16]  Sethi, C. C. “E-Mail Spam Detection using Machine Learning and Deep Learning”, Int. J. Res. Appl. Sci. Eng. 

Technol., 6(8), 981–985, 2020, doi: 10.22214/ijraset.2020.6159. 

[17]  Yile, A. O., Hongqi, L., Liping, Z., Sikandar, A. & Zhongguo, Y. “The linear random forest algorithm and its 

advantages in machine learning assisted logging regression modeling”, J. Pet. Sci. Eng., 4(17), 776–789, 2019, doi: 

10.1016/j.petrol.2018.11.067.  

[18]  Ali, A., Ahmed, M., Naeem, S., Anam, S. & Ahmed, M. M. “An Unsupervised Machine Learning Algorithms: 

Comprehensive Review”, Int. J. Comput. Digit. Syst., 2(20), 210–242, 2023, doi: 10.12785/ijcds/130172. 

[19]  Bakumenko, A. & Elragal, A. “Detecting Anomalies in Financial Data Using Machine Learning Algorithms. 

Systems”, 5(10), 32-48, 2022, doi: 10.3390/systems10050130. 

[20]  Reddy, P., Viswanath, P. & Reddy, E. B. “Semi-supervised learning: a brief review”. Int. J. Eng. Technol., 18(7), 

1- 18, 2018, doi: 10.14419/ijet.v7i1.8.9977.  

[21]  Singh, V., Chen, S. S., Singhania, B. Nanavati, A. kumarkar, & Gupta, A.  “How are reinforcement learning and 

deep learning algorithms used for big data based decision making in financial industries–A review and research 

agenda. Int. J. Inf. Manag. Data Insights, 2(2), 100 - 109, 2022, doi: 10.1016/j.jjimei.2022.100094. 

[22]   Yang, C.; Wang, Y.; Zhang, A.; Fan, H.; Guo, L. A Random Forest Algorithm Combined with Bayesian Optimization 

for Atmospheric Duct Estimation. Remote Sens. 2023, 15, 4296. https://doi.org/10.3390/rs15174296. 

[23]  J. M. Bofill, “Updated Hessian matrix and the restricted step method for locating transition structures,” J. Comput. 

Chem. 15, 1–11 (1994). 

[24]    H. Li and J. H. Jensen, “Partial Hessian vibrational analysis: The localization of the molecular vibrational energy 

and entropy,” Theor. Chem. Acc. 107, 211–219 (2002). 

[25]  H. Wu, M. Rahman, J. Wang, U. Louderaj, W. L. Hase, and Y. Zhuang, “Higher-accuracy schemes for 

approximating the Hessian from electronic structure calculations in chemical dynamics simulations,” J. Chem. Phys. 

133, 074101, (2010). 

[26] M. Ceotto, Y. Zhuang, and W. L. Hase, “Accelerated direct semiclassical molecular dynamics using a compact finite 

difference Hessian scheme,” J. Chem. Phys. 138, 054116, (2013).