Authors - Jishnu Prakash k, Rithish S V, Liz Maria Liyons, Vijval Srinivasan, Deepthi L R Abstract - The rapid spread of misinformation and rumors on social media platforms, particularly Twitter, poses significant risks to public perception and decision-making. This study presents a comprehensive approach to analyzing and mitigating rumor propagation by identifying key influencers and optimizing propagation time within online communities. Our dataset consists of over 800,000 nodes with interactions categorized as retweets, mentions, and replies, each assigned an influence score to quantify user impact. Using the Infomap algorithm, we initially detected 13,500 communities and filtered them to retain 67 influential clusters with a higher number of nodes and stronger influence scores. To analyze the spread of rumors, we developed an algorithm that tracks propagation within these communities, leveraging top influencers as initial spreaders and computing the average propagation time. Furthermore, we introduced a node deletion strategy to iteratively remove high-impact influencers, reducing the overall propagation time and limiting misinformation spread. Finally, we adjusted the propagation times by normalizing them with the earliest influencer timestamps to ensure precise measurement. Our findings highlight that targeted removal of key spreaders significantly disrupts rumor diffusion, providing insights into optimizing influence-based network interventions for misinformation control.
Authors - Yukta, Nishant Yadav, Sukrati Chaturvedi Abstract - The rapid expansion of social media platforms and online news consumption has led to an increased spread of fake news, posing significant challenges to society by influencing public decision- making and causing severe consequences in domains such as politics and healthcare. Traditional methods for identifying fake news are often too slow to combat its swift dissemination. Therefore, identifying and addressing misinformation is crucial for maintaining the accuracy and trustworthiness of information disseminated on social media. Natural language processing (NLP) and artificial intelligence (AI) play a vital role in this endeavor by facilitating the effective examination of extensive datasets to uncover patterns that are suggestive of misinformation. Machine learning models, particularly transformer-based architectures, enhance fake news detection by improving accuracy and interpretability. The integration of Explainable AI (XAI) methods further enhances trasparency and trust in these models
Authors - Praful Sambhare, Nitin Choudhary, Abhay Rahangdale, Atharva Rane, Sahil Raina Abstract - The integration of artificial intelligence (AI) within the fashion industry is becoming increasingly prevalent, with the aim of delivering a shopping experience that is personalized, seamless, and engaging. As an example of this trend, GlamBot is an AI- driven fashion assistant, a sophisticated fashion assistant that employs a range of advanced AI methodologies. These methodologies include Natural Language Processing (NLP), image- based similarity searches, and voice recognition technologies, all of which together transform the interactions that users have with fashion platforms. The functionality of GlamBot significantly improves the user experience by providing tailored fashion recommendations. This is achieved through the analysis of text inputs, the execution of visual similarity searches, and the processing of voice commands. Consequently, fashion discovery is rendered more intuitive and accessible for users, thereby facilitating a more engaging interaction with the fashion domain. . By analyzing and understanding user preferences, GlamBot builds personalized profiles that evolve over time to deliver increasingly accurate recommendations . GlamBot’s image-based search feature allows users to upload pictures of fashion items they like. Using advanced ResNet50 image recognition models, GlamBot analyzes these images and provides visually similar product recommendations, bridging the gap between users’ visual preferences and available fashion products .
Authors - Merlin Priya Jacob, Sukhada Aloni, Hetal Rawat, Shrishti Sakore, Shreya Naik, Shubham Pardhi Abstract - Using Ethereum, IPFS, and the MERN stack, this project offers a transparent and safe blockchain-based examination system. The solution guarantees tamper-proof question paper management by utilising IPFS (via Pinata) for decentralised storage and Ethereum smart contracts. Instructors submit tests to IPFS, ensuring data integrity by storing their cryptographic hashes on the Ethereum blockchain. While MetaMask allows for safe user interaction and authentication, Hardhat makes it easier to design and deploy smart contracts on a testnet. A strong online application is powered by the MERN stack, with Node.js managing database functions and instructor authentication. Exam papers are safely retrieved by authorised superintendents, guaranteeing regulated access. Academic assessment integrity is improved by this decentralised method, which reduces the possibility of paper leaks and unauthorised changes while offering a scalable and effective substitute for conventional test systems.
Authors - Sarbjit Kaur, Jasmeen Gill Abstract - Wireless sensor networks (WSNs) play a vital role in sensing environmental conditions in far-flung areas. However, their energy consumption remains a critical issue, affecting the network's lifetime and coverage area. Clustering has emerged as an efficient strategy to prolong sensor network lifespan, and the Fruit Fly Algorithm (FFA) and Ant Colony Optimization (ACO) are promising techniques for cluster formation and efficient path establishment, respectively. In this study, we propose an innovative approach that combines FFA for cluster formation and ACO for path establishment. This novel algorithm is implemented in MATLAB and evaluated in both homogeneous and heterogeneous environments. We compare our proposed algorithm with the Biogeography-Based Optimization Algorithm (BOA) and the Low Energy Adaptive Clustering Hierarchy (LEACH) algorithm. Our results indicate that the proposed algorithm significantly outperforms both BOA and LEACH in terms of network lifetime and coverage area, particularly in heterogeneous environments.
Authors - M Jayaram, Kodari Madhavi, Amboth Anil Kumar, Pachipala Naveen, Gajula Rithvik Abstract - Human-Computer Interaction prioritizes universally accessible systems, crucial for individuals with physical disabilities. This study introduces an Eyes as Interfaces: A Novel Eye-Tracking Mouse Cursor System, a hands-free solution enabling seamless digital environment interaction. For people those with paralysis, muscular dystrophy, or spinal injuries, this technology provides independent computing access, eliminating dependency on external assistance. A Convolutional Neural Network (CNN) is the system's backbone that provides real-time pupil detection, mapping eye gaze to exact cursor movement. Advanced image processing like this guarantees smooth operation regardless of changing lighting and user conditions. By accurately mapping eye movements to cursor actions, users can navigate and communicate with computer interfaces, opening avenues for information access, communication, and work participation. This encourages independence and enables users to access the web on their own, performing tasks like document creation and web navigation. Beyond personal benefits, this technology promotes inclusivity by bridging the digital divide, allowing for real-time, unrestricted participation in learning, employment, and social activities. Its smooth integration in widespread digital platforms ensures that it carries the highest level of potential in changing lives of people with mobility disabilities Worldwide.
Authors - Kailash Agarwal, Parikshit N. Mahalle, Bhagwan D. Thorat Abstract - Heart failure is a serious medical condition that affects millions worldwide, and early prediction is essential for timely intervention and better patient outcomes. While machine learning models have demonstrated strong predictive capabilities in healthcare, many high-performing models, such as Support Vector Machines (SVM), function as black boxes, making them difficult to interpret in clinical settings. This study examines how Explainable AI (XAI) techniques can enhance transparency in heart failure prediction.Using a publicly available dataset from Kaggle, we preprocess the data with label encoding, feature scaling and Hyperparameter tuning before training various machine learning models for binary classification. Our results indicate that the Support Vector Classifier (SVC) with a Linear kernel achieves the highest predictive accuracy. However, to improve interpretability, we compare its performance with explainable models like Decision Trees and apply post-hoc explanation techniques such as SHAP (SHapley Additive Explanations) and Permutation Importance.Through this comparative analysis, we highlight the trade-off between model accuracy and interpretability, offering insights into the feasibility of XAI-driven models in real-world clinical decision-making. Our findings reinforce the importance of developing AI systems that not only perform well but also provide understandable and trustworthy insights for medical professionals.
Authors - Bhagwan Thorat, Omkar More, Prathamesh Medage, Aditi Mali, Janhavi Maske, Sanika Maind Abstract - Traditional healthcare systems have primarily focused on physical health, often overlooking mental well-being. With the rapid advancement of technology, integrating AI-driven emotion detection into mental health management can offer valuable insights. This paper presents a comprehensive mental health management system that utilizes Haar Cascade classifiers and a Keras deep learning model for real-time emotion recognition via OpenCV. A Flask-based web interface, built using HTML, CSS, and Python, enables users to monitor their emotional states and facilitates therapist booking and automated receipt generation. By leveraging facial expression analysis, the system provides a data-driven approach to mental health assessment, enabling early intervention. The platform also ensures accessibility and efficiency, reducing the burden on healthcare providers. Experimental evaluations demonstrate the system’s effectiveness in accurately detecting emotions and its potential in AI-assisted psychological support. Future enhancements will focus on multi-modal emotion detection, incorporating natural language processing (NLP) and IoT-based physiological monitoring for a more holistic approach to mental health assessment. This research contributes to the growing field of AI-powered mental health solutions, bridging the gap between technology and psychological well-being while promoting early detection and accessible care.
Authors - Sonali Patil, Siddhesh Arun Patil, Ayush Patil, Piyush Pawar, Siddhesh Sandeep Patil Abstract - Around 970 million people face mental health disorders across the world and depression affects 75% of these people because of their sleep disturbances. The connection between persistent sleep problems and depression emerges when affected individuals become twice as likely to develop depression thus establishing sleep as a major sign for mental health forecasting. Current approaches to this problem deal with three key issues which are dataset biases, small available sample sizes along with the reliance on self-reported symptoms instead of actual physiological signals. Our deep learning solution relies on multi-channel Convolutional Neural Networks (CNNs) to analyze wearable sensor data because it tackles existing analysis limitations. DreamT-150 contains heart rate (HR), blood volume pulse (BVP) and electrothermal activity (EDA) measurements from 150 sleep patients. Three models including MultiChannelCNN and MultiChannelEfficientNet and MultiChannelResNet analyzed the signals which appeared as time-series graphs. The best model proved to be EfficientNet-B0 because it demonstrated superior generalization. The pre-trained layers from EfficientNet adjusted the vulnerability of training loss which led to stable model performance. The research demonstrates sleep-derived physiological signals' usefulness for non-invasive mental health predictions which can lead to real-time monitoring systems. The upcoming research aims to boost both dataset range and better models for clinical adoption requirements.
Wednesday August 26, 2026 12:30pm - 2:30pm IST Virtual Room BGOA, India
Authors - Bhoomi C. Parikh, Zankhana Shah Abstract - Stress is any type of mental imbalance that can lead to mental disorders ranging from low to high severities which can be classified as acute and chronic stress conditions. Chronic stress leads to hyperactivation of the sympathetic nervous system, resulting in physical, psychological, and behavioural problems. Currently, there is no recognised standard for stress assessment. Thus depressive disorders leading to stress are a flight or fight response to the stimulus generated by human nervous system caused due to any unacceptable behaviour or circumstance. Throughout this response adrenaline hormones are secreted that leads to increased respiration and heart rates, along with increased muscle activity. Such type of biological alterations prime the organism for a physical response that affects human body mechanisms in terms of sleep abnormalities, digestive disorders or work imbalances in routine lives. Thus WESAD is a multimodal wearable dataset which combines both affective states(baseline, depression and happy) and other sensor modalities such as blood pressure, ECG, skin conductivity , EMG, breathing, and three-axis acceleration. There are also other classification parameters based on physiological changes which are also found in WESAD dataset and by using different types of Machine Learning Classifiers analysis is done . The algorithms with the highest accuracy can be used for developing a novel and a hybrid model which can categorise stress based on Heartrate Variability and stating HRV as a biomarker for stress detection. Both characteristics related to time and frequency of heart rate are categorized in the research study. In the context of the three-class classification based on three affect states , baseline, stress, and amusement result up to 99% was obtained. Use of two affective states like stress and amusement gave an accuracy up to 84% using DT classifier.
