Proceedings of the Sixth International Conference on Emerging Trends in Mathematical Sciences & Computing (IEMSC-25)

Table of Contents

1. Frontmatter

2. Impact of Demographic Factors on E-Commerce Sales Events: An Empirical Study of Indian Consumers

   Jyoti Shaw, Anirban Sarkar, Subhabrata Mitra, Subrata Jana

   Abstract

   E-commerce has transformed the retail landscape, with demographic factors playing a crucial role in shaping consumer behaviour during sales events. This study examines the impact of age, gender, income, and education on online shopping experiences among Indian consumers. Using a quantitative research approach, it employs binomial logistic regression and Principal Component Analysis (PCA) to identify key predictors of purchasing behaviour. Findings indicate that younger consumers (aged 24–30) and males are more actively engaged in e-commerce. While gender significantly influences a positive shopping experience, other demographic factors, including income and education, have minimal impact. The study highlights the importance of targeted marketing strategies based on consumer profiles. However, limitations such as gender imbalance in the sample and the model’s moderate explanatory power suggest the need for further research. Future studies should incorporate additional demographic variables to enhance predictive accuracy and optimize e-commerce strategies for diverse consumer segments.

3. Waves at Spring Contact of Two Fiber-Reinforced Thermo-elastic Mediums

   Annu Rani, Aanchal Gaba

   Abstract

   In this manuscript, incidence of a SV wave at spring contact of semi-infinite spaces is examined. The effect of an imperfect interface and initial stress on waves for the incidence of plane waves in fiber-reinforced media is observed in this manuscript. The L-S (Lord-Shulman) theory is used to examine the effects. The incidence of the plane wave is investigated at the interface of fiber-reinforced and prestressed semi-infinite spaces. 2-D and 3-D graphs are plotted to show the impact of initial stress, imperfect parameters and reinforcement. Under initial stress, all refracted waves exhibit mixed behavior. This study is helpful for civil engineers and geologists.

4. Aqua-Sentinels Safeguarding Inland Waters from Plastic Peril

   Biswadip Basu Mallik, Sohini Banerjee, Anshit Mukherjee, Avishek Gupta, Sudeshna Das

   Abstract

   The worst environmental issue of this decade is pollution resulting from marine plastics that significantly affects the ecosystems, marine life, human health, and the economy. In order for this problem to be solved, it becomes compulsory to prevent plastic waste from reaching marine waters through the correct identification and removal of the material from most likely locations on the land. In the past few years, rapid progress has been made in the domain of computer vision and deep learning which has shown strong potential to handle this challenge but hurdles like object recognition in complex scenarios, high preciseness maintenance, and rapid detection still persists. This notable void has motivated us to choose this topic for our research whose aim is to overcome these hurdles stated previously in the literature. This study introduces a novel algorithm based on deep learning which when integrated in underwater drones will be used for detection of floating waste which includes small plastic debris and other debris in inland waterbodies and by using various simulation tools, we have seen that our proposed algorithm successfully overcomes the stated voids. The proposed algorithm is using the FloW dataset, which comprises of two main components: the vision based sub-dataset (FloW-Img) and multimodal sub-dataset (FloW-RI). The algorithm demonstrates improved performance in precisely recognising small objects, such as plastic debris, against complex background. It also has unparalleled accuracy and recall despite challenging environmental situations that make the results functional for real-world use. In addition to it the algorithm is optimized for speed, enabling real-time monitoring and timely responses to pollution events. This study makes a remarkable progress in the enhancement of autonomous waste detection and cleanup models.

5. Predicting Cyclonic Events: A Machine Learning Approach Using Weather Data

   M. Ayyadurai, P. Brijith Manikandan, S. Akshay Kumar, Ashish P. Shaji, Aaditya Partha Sarathy, A. Bernieo Fatim

   Abstract

   As cyclones are among the most destructive natural disasters to human life, anticipating cyclone sensitivity is important in the management of disasters. The purpose of this paper is to advocate for the application of machine learning methods to the forecasting of cyclonic events based on wind speed, air pressure, temperature, and humidity as relevant factors. To counter the challenges of label noise, missing values, and class over-representation, and to estimate the issues present in real data, the acquired meteorological data is pre-processed. On accuracy, precision, recall, and ROC-AUC, for instance, the comparison of the Gradient Boosting and Logistic Regression models shows that the latter outperforms the former. The study focuses on the benefits of ensemble techniques in handling non-linear relationships and class imbalance conditions, and also the need for holistic preprocessing techniques such as noise elimination and feature scaling. The study confirms that machine learning (ML) applications will have a great potential to enhance cyclone forecasting systems, a significant step towards disaster preparedness and risk reduction. To enhance the quality of the model in the future, deep learning techniques will be tried out and blended with current data.

6. Comparative Study of Machine Learning Algorithms for Loan Prediction in Banking

   M. Ayyadurai, R. U. Amirthavarshini, M. N. Chandni, K. Bhuvaneshwari, M. C. Arun, D. Alfred Sam

   Abstract

   This paper entails the comparison of four machine learning algorithms, namely Logistic Regression, Decision Tree, Random Forest, and XGBoost pertaining to loan approvals to facilitate the optimization of credit decisions made by the financial institution that shall be assisted through characteristics of applicants. This paper used the Kaggle dataset and also tried forward feature selection as well as hyperparameter tuning in an alternative way. All models are experimented; the best logistic regression test accuracy is up to 88.73%. Meanwhile, the decision tree as well as the random forest can be seen with obvious overfitting cases. XGBoost also showed a tendency towards overfitting, only a little margin from logistic regression. This Logistics Regression model will achieve overall performance which best strikes a balance between the requirements of high accuracy on the training data and generalization. The paper addresses also an issue, that fits in one area of trade-off, which is model complexity/interpretability—an extremely important issue in finance, because quite obviously, transparency is a necessity. Hence, overfitting has played a very vital role in this model; however, how it affects the former is very simple; Logistic Regression was to be a victim of overfitting at the cost of one being much more complex than the two other models namely, Random Forest and XGBoost models. Hence, the results add to the significance of model selection and validation on loan decisions taken in real-time. These results hint at a growing need for XAI since stakeholders should trust that this will be provided to an automated financial decision-making system with such transparency in place that this holds; further piling on the debate of balance in financial machine learning with better interpretability but potentially greater predictive performance in real applications.

7. Prospects and Challenges in the Adoption of AI—Powered Personal Finance Applications: A User Perspective

   B. Lavanya, D. Srivani

   Abstract

   Money management plays a crucial role in day to day activities of an individual. Effective money management reduces the risk of loses and help individuals to gain wealth. Tracking these personal expenses from day to day activities is a hectic and time consuming task and often has high risk of human errors. These personal finance management applications use Artificial Intelligence to effectively analyse patterns by continuously learning from user interactions and financial habits of various individuals who differ in the age, gender and their income levels from with their user friendly interfaces track people’s expenses and provide customised solutions according to their need thus mitigating errors and saves users efforts and time. This study is based on primary data where responses are collected from various individuals who use these personal finance management applications. The findings demonstrate on how the user demographics plays a role in the extent of personal finance management applications usage and to discover intended purpose among individuals, related challenges they face in adopting these AI powered features for customised solutions and the impact of these AI powered tools in these finance applications.

8. Predictive Modeling of Satellite Collision Risk Using Machine Learning Models

   M. Ayyadurai, M. Darshan, K. Deepak, S. Aravinthaa, D. Pratheba, D. A. Daniel Leve Manickam

   Abstract

   The rapid expansion of space exploration has led to an increase in space debris formation, especially in Low Earth Orbits [LEO]. These space debris poses a significant threat to both ongoing and future space missions. This paper investigates the risk of satellite collisions with space debris and it presents a machine learning model that predicts the collision probability with the help of the orbital data gathered from Space-track.org. The idea is to initially calculate the distance of closest approach between a satellite and nearby debris by splitting the orbit into an interval of points and calculate the distance between the satellite, then we find the minimum value from all those distances which gives us the approximate value of minimum distance. Then, we trained the model with a labeled data which is a collection of minimum distances and their probability of collision through linear regression and random forest algorithms, and the results were promising. With the help of this model, space agencies can determine collision-free satellite orbits and mitigate the threats posed by space debris.

9. BloodLink India: A Smart AWS-Based Cloud Blood Bank System for Real-Time Blood Management

   Dipanjana Chakraborty, Panem Charanarur, Mrinal Kanti Deb Barma

   Abstract

   Indian blood donation still remains plagued by numerous challenges such as insufficient supply, inefficient storage, and no real-time tracking. The infrastructure in place is mostly fragmented, resulting in the challenge of matching demand with supply, particularly during emergency situations. Further, manual recording and the use of conventional techniques make blood banks inefficient, thus making it hard to ensure timely delivery and reduce wastage. To deal with these challenges, we introduce BloodLink India, a sophisticated, cloud-based blood bank solution on AWS technologies. The solution uses AI-based demand forecasting, blockchain for safe and transparent transactions, IoT for real-time monitoring of blood storage temperature, and geolocation-based alerts for the donors. Using AWS cloud infrastructure, BloodLink India offers a scalable, secure, and automated platform to manage blood donation and distribution more efficiently and transparently. This paper discusses the design, methodology, and implementation of BloodLink India along with highlighting its real-world applications. The system would transform blood bank management with real-time information, minimizing wastage, and making blood available where and when it is most needed. Applications range from emergency response coordination to effective donor mobilization via geolocation-based notifications and complementing existing healthcare networks seamlessly, thereby making healthcare more accessible all over India.

10. Ward Guard: An AI-Powered Mask Detection System for Healthcare Safety

    Panem Charanarur, Dipanjana Chakraborty, Madhusudan G. Lanjewar

    Abstract

    The importance of mask compliance in stopping the transmission of airborne illnesses was brought to light by the COVID-19 pandemic, especially in high-risk settings like hospitals and intensive care units (ICUs). Manual enforcement of mask-wearing policies is still ineffective and subject to human error, even with stringent regulations. Ensuring real-time compliance is crucial to lowering infection rates and safeguarding healthcare workers in nations like India, where a large patient intake and a shortage of medical staff pose extra difficulties. Even uneven compliance with mask laws worldwide still impacts public health, necessitating automated solutions. We suggest Ward Guard, an artificial intelligence (AI) powered real-time mask detection system intended for healthcare facilities, to solve these issues. The proposed MobileNetV2-base method achieved an accuracy of 93.5%. It recognizes people who are not wearing masks using computer vision and deep learning, and it promptly alerts specified medical staff through Twilio-based messaging. In contrast to conventional surveillance, Ward Guard improves hospital safety by automating compliance monitoring, drastically lowering the need for physical intervention. The system is perfect for implementation in urban and rural healthcare settings since it is scalable, flexible, and can function in contexts with limited resources. The Ward Guard uses MobileNetV2, which is tailored for edge devices like the Raspberry Pi and NVIDIA Jetson. Hospital administrators may efficiently enforce policies by using the system’s integration of cloud-based analytics to track compliance patterns over time. The suggested approach ensures low latency and real-time processing, guaranteeing a prompt reaction to non-compliance. Ward Guard is a next-generation hospital safety technology that will be enhanced with voice-based warnings, broader PPE detection (gloves, face shields), and connectivity with Electronic Health Records (EHR) for automated reporting.

11. Zero-Shot Learning: Enabling Deep Learning Models to Recognize Unseen Categories

    Madhav Sharma, Pushpendra Sikarwal, Samiksha Agarwal

    Abstract

    Zero-shot studying (ZSL) represents a innovative technique to deep understanding acquisition, permitting models to become aware of and classify studying that has not been encountered across faculties. By the usage of semantic illustration and embedding techniques, ZSL addresses the limitations of conventional supervised gaining knowledge of, which is based closely on categorized information This paper explores ZSL’s theoretical foundations, techniques and applications. The paper identifies states of concern on crucial functions related to ZSL, together with troubles of switch and semantic holes, and explores contemporary trends such as gaining knowledge of and numerous foundational models. The findings spotlight the potential of ZSL to redefine the device to gain information of scalability while simultaneously offering the roadmap for enrichment improvement.

12. Development of Deployment Schedule for Web Browser Based Tasks Framework Using Release Cadence Approach Based on Test Early Methodology

    Nagendra Singh Yadav, Vishal Kumar Goar

    Abstract

    Software delivery in a timely planned manner has been of utmost priority for the software development organization. The new changes to the release schedule can cause further delays in the release causing frustration for end users. As a whole the goal of the software development is to keep the customer base happy but doing so gets difficult as there is no endpoint of introducing new changes in the release to meet the expectations of end users. This causes more expense in software development and delays the schedule of future releases in software development. Our approach is to introduce a generic template of release cadence using test early methodology to support the execution of web browser-based tasks framework. The approach consists of various factors i.e. Time, quality of developed software and timely release which ensures that the release has a hard commit date and at a certain point introducing new changes in the software risks the associated components of feature under development and AUT (application under test).

13. A Novel Approach to Solve Fully Fuzzy Multi Objective Linear Programming Problem

    Pallabi Pal, Sayanta Chakraborty, Apu Kumar Saha

    Abstract

    The present treatise aims to formulate a general fully fuzzy Multi Objective Linear Programming Problem (FFMOLPP) under Intuitionistic Fuzzy Environment (IFE) aiding Triangular Intuitionistic Fuzzy Numbers (TIFNs) to reflect decision maker’s (DMs’) choice. To solve the problem, firstly, the fuzzy parameters have been converted into crisp numbers using the accuracy function defined on TIFN. Secondly, a tolerance based approach has been proposed to convert the fuzzy inequality or equality constraints to crisp constraints. Thirdly, the membership functions corresponding to multiple objectives have been formulated based on optimistic and pessimistic values of the objective functions. Finally, using Bellman-Zadeh’s principle and Zimmermann’s operator, the crisp Linear Programming model has been solved. The proposed method has been illustrated with a numerical example. The proposed approach elicits an efficient solution.The reliability and authenticity of the results have been checked through comparative investigations. It has been observed that the proposed method arouses better objective values in comparison with existing techniques.

14. Optimizing Authenticated Encryption: Enhancements and Performance Analysis of ChaCha20 and Blake3 Algorithms

    Sameeruddin Shaik

    Abstract

    This study investigates the optimization of authenticated encryption algorithms, focusing on ChaCha20 and Blake3. Our analysis demonstrates that the ChaCha20 single-threaded implementation achieves a notable 1.25× speedup over the libsodium library. The parallel version of ChaCha20 exhibits impressive scalability, reaching a peak throughput of 130 GB/s and showing linear speedups up to 32 threads, constrained by CPU memory bandwidth. In comparison, the Blake3 implementation, both single-threaded and parallelized, exhibits performance on par with the reference implementation. The stack approach for Blake3 offers substantial performance improvements over traditional methods, while the divide-and-conquer approach shows scalability with increasing core counts, promising efficiency for large-scale file sizes. Future research will explore advanced vectorization strategies, multi-CPU scalability, and thorough security analysis to further enhance performance and robustness.

15. Multivariate Modeling of Blood Cell Variability for Early Diagnosis of Autoimmune Hemolytic Anemia

    Prasenjit Kundu, Sayani Ghosh

    Abstract

    Autoimmune hemolytic anemia (AIHA) is a rare and challenging autoimmune disorder characterized by the immune system’s attack on red blood cells (RBCs) which is leading to anemia and other health complications. This study investigates the relationships among key hematological parameters which are Hematocrit, Mean Corpuscular Volume (MCV), Mean Corpuscular Hemoglobin (MCH) and Mean Corpuscular Hemoglobin Concentration (MCHC)—and their effects on leukocytes, thrombocytes and erythrocytes. The study uses a large secondary dataset of complete blood count samples. It applies multivariate regression and analysis of variance among various blood cell components with the goal to early screening of AIHA. The study also focuses on creating diagnostic models. The findings show that MCHC has a negative link with platelet counts. Hematocrit and MCH have positive links with RBC counts. Challenges such as multicollinearity and non-normal data distributions were addressed through statistical techniques like Variance Inflation Factor analysis. The results highlight the distinct roles of hematological parameters in diagnosing AIHA and suggest that integrating these parameters can improve early detection and treatment strategies. The study has some limitations in model performance for certain blood components. But it provides a strong base for using blood tests to diagnose autoimmune disorders. It highlights the need for better methods and larger datasets in future research.

16. Generative Adversarial Networks for Early Autism Detection: A Novel Approach Using ML

    Vaibhav C. Gandhi, Nirav V. Patel

    Abstract

    Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by challenges in social interaction, communication, and repetitive behaviors. The earlier the diagnosis, the more interventions and better developmental outcomes can be made. This research explores the capability of Generative Adversarial Networks (GANs) in enhancing machine learning models for the early detection of ASD by synthesizing data that enhances model robustness and accuracy. The dataset was trained on a combination of real and GAN-generated synthetic data to train machine learning models such as SVM, decision trees, and random forests. GANs were used to improve the diversity of the dataset and thereby enhance feature learning and generalization of the model. The model evaluation metrics included accuracy, precision, and recall. The integration of GAN-generated data significantly improved the model's performance. The best accuracy was found for the Random Forest model, which reached 93%, and its precision reached 90% along with a recall of 92%. Feature importance analysis gave the impression that dynamic functional connectivity, electroencephalography, and magnetoencephalography were the most critical in the prediction of ASD. The application of GANs is a promising approach to handling data limitations in the detection of ASD and enhances the performance and generalizability of machine learning models. More validation research is necessary on larger and more diverse datasets to validate the above findings.

17. Exploration of CTGAN for Synthetic Cloud Log Data Generation

    K. Vijay, N. Krissh Shankaran, M. Aashif, S. Santhosh Kumar

    Abstract

    Synthetic data generation, a process where datasets are generated via computer algorithms that don’t compromise privacy are shown to near-perfectly paradigm the abstract and complex relationships between data items. We can generate a diverse and massive quantity of training data with little effort by artificially simulating genuine data sets with synthetic data. By utilizing a conditional tabular generative adversarial network, it is possible to create synthetic data that mimics real-world cloud data while maintaining the network’s integrity. Once the real log data has been preprocessed, identification of key features and their distributions will guide the synthetic generation process. CTGAN can handling imbalanced data, capturing complex distributions while improving the overall stability of the data. In CTGAN, the generator improves by trying to maximize the probability of the discriminator making mistakes, while the discriminator enhances its accuracy in identifying fake samples. The use of these methods for generating realistic data showcases potential in the ability to accurately generate cloud logs datasets which are nearly indistinguishable from real testing data collected from the real world.

18. AyurBotanica—Identification of Medicinal Plants using Deep Learning

    K. Vijay, K. S. J. Sneha, A. Satyanaryana

    Abstract

    Medicinal plants have played an integral role in people’s lives especially with the growing interest in Ayurveda and Natural remedies which bring about little to no side effects. Since ancient times, people have always turned to medicinal plants for therapeutic and medicinal values along with their potential health benefits. Identification of these medicinal plants may bring about a safe and effective way of using these plants to harbor the desired effect. Using deep learning by leveraging the power of transfer learning with EfficientNetV2 architecture, we make this process efficient and accurate. Specifically, Convolutional Neural Networks (CNNs) are utilized to detect these medicinal plants accurately. Unlike traditional methods, which may be laborious and prone to errors, our approach harnesses the efficiency and accuracy of deep learning.