April 26 ~ 27, 2025, Copenhagen, Denmark
Venkata Duvvuri, Chetan Kulkarni, Sritha Gogineni
Abstract – COVID-19 pandemic has created a major impact around the world. Governments and businesses small or big around the world are facing unprecedent decisions to either close up or reopen or drive other policies based on the sentiment of people. While, understanding this sentiment and accompanying emotions has been researched especially in social media channels like Twitter, we propose a novel way to capture sentiment and emotions using intelligent chatbots (EmoBot) that reduces the participants biases inherent in prior analysis. We devise Emotion Extraction Layers (EEL) based on latest deep learning techniques like BERT (Bidirectional Encoder Representations from Transformers) and compare these models with traditional machine learning models. We show for a variety of emotions that the new deep learning models predict 1-5% (Sad, Fearful & Angry) better than traditional machine learning techniques. Further, we showcase that leveraging retail sentiment data using transfer learning techniques can help cross the cold start chasm of having no chatbot data initially, and this technique achieves -8% closer in performance when compared to having enough COVID sentiment data.
COVID, sentiment analysis, chatbots, BERT, deep learning, transfer learning
Owen Arnold Brue1, Samuel Silverberg2 , 1Ransom Everglades School, 3575 Main Highway, Miami, FL 33133 , 2Computer Science Department, California State Polytechnic University, Pomona, CA 91768.
Proper rowing training should utilize professional trainers, but not everyone who wants to get into rowing has access to that. This application will solve this problem by using pose analysis applied to videos of users’ rowing forms [1]. After the algorithm analyzes their form, A.I. will use that data to compose a personalized message that can give advice on how to improve their form. A major design challenge is creating the user feedback on how to improve form using pose analysis. Instead of showing the user how much their pose differs from the ideal form based on coordinate data, we can enlist A.I. language models [2]. This will help translate the differences from numerical data into more natural sounding advice. For rowers, they should use my application as a way to improve their form. Novices will benefit if they’re unable to train with a professional, so they’ll be able to learn proper form early and avoid injury [8]. Experts benefit too if they desire to maintain good form and continue to stave off possible injuries.
Rowing Form Analysis, Pose Estimation, AI-Powered Coaching, Personalized Feedback
Michelle Chen1, Laurie Delinois2 , 1Weston High School, 444 Wellesley St, Weston, MA 02493 , 2Computer Science Department, California State Polytechnic University, Pomona, CA 91768
Both Parkinson’s disease and Alzheimer’s disease have become more globally prevalent as the years pass. This project aims to mitigate/slow down their effects by providing a guided way for those afflicted to practice their motor function and cognitive abilities through a drawing app that integrates AI-generated prompts for users to work with. Challenges included image clarity when processing both 2D and 3D images to create line art with and ensuring that AI prompts were not overly repetitive so as not to frustrate users. It was also a challenge to ensure that aesthetic issues such as lack of brush types and UI would not deter those that the application is meant for by causing frustration and disinterest. These issues were fixed through experimentation and testing to determine the best values for each component of the program. For example, the experimentation done for the image processing scene offered insight as to the best edge depth value for high clarity in processed 3D images. Through our second experiment we found that it is also important how specific we are with the prompt given to the chat-gpt API, as a lack of specifics can cause redundancy in the prompts given to users. Ultimately, this application will provide those with AD and PD an opportunity to experience a form of art therapy in a calm, individual environment where they can go at their own pace in maintaining their motor skills, providing an alternative approach to more traditional therapy routes.
Art therapy, AI-generated prompts, cognitive training, motor function rehabilitation
Muhammad Uzair Akmal1, Selvine George Mathias1, Saara Asif1, Leonid Koval1, Simon Knollmeyer1, and Daniel Grossmann2, 1AImotion Bavaria Technische Hochschule Ingolstadt, Germany, 2Faculty of Computer Science and Data Processing Technische Hochschule Ingolstadt, Germany
The lack of trust and fairness in artificial intelligence (AI) systems driven by biases, misclassified data, lack of transparency, and limited interoperability, raises significant ethical concerns and socioeconomic impacts. This study presents a reference architecture for an AI pipeline aligned with Industry 5.0 principles, focusing on human-centered design, sustainability, social responsibility, and resilience. It enhances human-AI collaboration by involving four user types (data scientists, domain experts, organizations, and end users) who share decision-making responsibilities during the AI system development process. The architecture incorporates Active Learning (AL) to address data bias and misclassification issues and Transfer Learning (TL) to ensure model reusability in resource-constrained environments. Post-modeling Explainability gives stakeholders insight into model behavior and outcomes, fostering transparency and trust. Additionally, two user-ranked custom validation metrics evaluate the architecture and calculate Mean Average Precision (MAP) for Rankings. These metrics ensure the architecture design and outcomes adhere to ethical AI principles while promoting collaborative, responsible, and sustainable AI development.
Artificial intelligence, Human-centric AI, Active learning, Transfer learning, Explainable AI, Intelligent systems, Industry 5.0.
Saritha Kondapally, Senior Member IEEE
The growing reliance on data-driven innovation in healthcare often collides with the critical need to protect patient privacy, creating a tension between progress and compliance. This study bridges that gap by introducing a Variational Autoencoder (VAE)-based framework to generate synthetic healthcare data that mirrors real-world datasets while ensuring privacy preservation. By leveraging synthetic EHRs created using the Synthea tool, the framework achieves a balance between statistical fidelity and data utility, enabling secure sharing and collaboration without compromising sensitive information. Through rigorous evaluation of distributional alignment and predictive performance, this work demonstrates the promise of synthetic data in unlocking the full potential of AI-driven healthcare solutions, offering a path to innovation that respects both privacy and progress.
Privacy-Preserving Data Generation, Variational Autoencoders (VAEs), Synthetic Healthcare Data, Generative AI in Healthcare, HIPAA & GDPR Compliance, Electronic Health Records (EHRs), Data Privacy and Utility Trade-off, Machine Learning for Healthcare, AI, Federated Learning & Differential Privacy, Data Sharing & Secure Collaboration, Feature Engineering, FHIR Standard for Interoperability
Leonard Knapp1 Sven Nitzsche1, Matthias B¨orsig1, Alexandru asilache1,Ingmar Baumgart1, and J¨urgen Becker2, 1FZI Research Center for Information Technology, Karlsruhe, Germany, 2Karlsruhe Institute of Technology, Karlsruhe, Germany
In computer networks, protection against potential threats is paramount, requiring robust security measures. However, traditional rule-based Intrusion Detection Systems (IDSs) often fail to adapt to dynamic environments, prompting the exploration of innovative solutions such as Neural Network (NN)- based approaches. This research explores the efficacy of Spiking Neural Networks (SNNs) as the sole data processor in IDSs, which differentiates this approach from previous work. Through extensive experimentation on the NSL-KDD, CIC-IDS-2017, CIC-IOT-2023, and AWID3 datasets, various SNN configurations were examined alongside conventional Artificial Neuronal Networks (ANNs). The results highlight the promising performance of SNNs, which achieved remarkable accuracies of up to 99.22% on these datasets using rate and density encoding. Furthermore, a comparative analysis reveals the competitive advantage of SNNs over their ANN counterparts in generating fewer false positives at equivalent accuracy, emphasizing their adaptability to time-dependent data. This study thoroughly evaluates the achievable accuracy of an IDS built from spiking- and artificial neurons within a feed-forward fully connected topology. For the spiking neurons, the Leaky Integrate-and-Fire (LIF) model is selected. The results obtained by this approach support a paradigm shift towards SNN-based IDSs to strengthen network security, although further research is essential to ensure broader applicability and scalability.
Spiking Neural Network, Artificial Neural Network, Intrusion Detection System, Computer Network.
Ngo Dinh Luan, Nguyen Van Thanh Thong, Hoang Ngoc Dung, FPT University, Vietnam
Alcohol intoxication is a major cause of traffic accidents and a potential threat to public safety, especially in Vietnam. Traditional detection methods, such as breathalyzers or blood tests, are invasive to privacy, time-consuming, and require the subjects cooperation. This study provides an innovative, non-invasive method for detecting alcohol-induced cognitive impairment on the street using thermal imaging technology. Our method provides real-time drunkenness recognition by analyzing specific facial temperature variations, features, and physiological patterns using deep learning such as convolutional neural network (CNN) models. Experimental evaluations demonstrate the advantages of this method, providing a non-contact, real-time solution that is important for law enforcement, healthcare, and traffic safety.
Drunk identification, Convolutional Neural Network, Thermal image, Image Processing.
Shakthi I. Weerawansa, Uthayasanker Thayasivam, Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka
Depression is a major global health concern, and early identification is crucial for effective intervention and therapy. This study explores machine learning techniques to predict depression risk using a dataset from a mental health survey. Three models—XGBoost, Random Forest, and Logistic Regression—were evaluated based on accuracy, precision, recall, and F1-score. XGBoost achieved the highest accuracy (93.51%) and recall (80%) for the minority class, demonstrating its superiority in handling imbalanced datasets. Our findings revealed that financial stress, work pressure, and sleep duration were the most influential predictors. These findings highlight the potential of machine learning in developing automated mental health screening tools.
Wilson Caldeira and José Manuel Simões, Department of Media, Art & Technology, USJ University, Macau, China
As artificial intelligence (AI) technologies have advanced to produce images that are indistinguishable from those created by humans, understanding how these images are interpreted is crucial. This study investigates perceptions of realistic attributes of visuals generated by generative AI. It further examines the influence of these synthetic images on trust in the visual media. Employing a theoretical analysis grounded in the existing literature, this study elucidates the interplay between realism, memory distortion, and trust within contemporary visual culture.
AI-generated imagery, realism, memory, trust, visual representations
Maxwell Palance1, Austin Amakye Ansah2, 1Stanford Online High School, 415 Broadway Academy Hall, Floor 2, 8853, 415 Broadway, Redwood City, CA 94063, 2The University of Texas at Arlington, 701 S Nedderman Dr, Arlington, TX 76019
Students with Autism Spectrum Disorder (ASD) face unique challenges in traditional and online coding education [5]. While existing platforms offer authentic programming exercises, they often fail to address the neurodiverse learning needs of these students. This paper proposes an AI-powered e-learning platform, CODEversity, designed to offer an engaging and gamified coding education experience [6]. Built using Flutter and Firebase, the platform integrates real-world coding exercises with interactive visual and auditory feedback to foster positive reinforcement. A chatbot assistant powered by large language models helps guide students in course selection based on their interests and progress. The system also incorporates a dynamic quiz program, utilizing a tree-based course roadmap and practical coding exercises executed via a sandboxed Python environment [7]. A robust course builder ensures seamless course creation, validation, and updates without data loss. By prioritizing engagement and inclusivity, CODEversity aims to make coding education accessible and enjoyable for neurodiverse learners.
E-learning, Neurodiversity, Autism, Programming, AI Assistant
Qinwen Zhong1, Nahil Memon2 , 1Huasiong College of Iloilo, Ledesco Village, La Paz, Iloilo City, Philippines , 2Computer Science Department, California State Polytechnic University, Pomona, CA91768.
ShuttleFit aims to improve access to affordable badminton training by providing AI-assisted biomechanical feedback through smartphone video analysis [1]. Traditional coaching is often costly, and ShuttleFit offers an alternative by integrating video upload and preprocessing, pose estimation and feedback, and progress tracking. Video handling is managed via Flask with cloud processing, while MediaPipe-based pose estimation and K-Means clustering enable movement analysis, though challenges exist in rapid motions and cluttered environments [2]. Initial experiments showed 64% accuracy, highlighting limitations in complex movement segmentation. Potential improvements include adopting DBSCAN for clustering and exploring edge computing to enhance performance. Unlike existing solutions requiring wearables or 3D motion capture, ShuttleFit’s smartphone-centric approach aims to lower costs and increase accessibility [3]. However, further research and development are required to fully validate its effectiveness, reliability, and real-world impact on badminton training.
AI-Assisted Coaching, Biomechanical Feedback, Pose Estimation, Badminton Training.
Ronald Kiprotich Tarus1 and Silas Lagat2 , 1Department of Management science, University of Strathclyde, Glasgow City, UK , 2Ministry of health Nandi County Eldoret -Kapsabet Road, Kapsabet, Kenya , 2Ministry of health Nandi County Eldoret -Kapsabet Road, Kapsabet, Kenya.
Software quality becomes the most important dimension to measure the success of any engineering firm, keeping in view the current software industry scenario, which is fast growing. With the effect of more advanced and complex software features, conventional quality management techniques quickly become ineffective. Discussed in this paper are how modern methodologies such as Agile, DevOps, and Six Sigma combined with business analytics will be bringing the revolution in software-quality management (SQM). Integrating business intelligence, predictive analytics, and machine learning, source-code detection rate limiting issues, improving decision-making processes, and constantly improving the software quality will become possible. The application of case studies and empirical research will demonstrate the conditions necessary for optimizing testing, performance, and ultimately high-quality software product delivery. The paper outlines how organizations may formulate robust frameworks integrating principles of software engineering with advanced analytics to deal with the challenges of fast- evolving software development environments.
Software Quality Management, Business Analytics, Agile, DevOps, Six Sigma, Predictive Analytics, Machine Learning, Continuous Improvement, Software Testing.
Shengtian Hong1 , Yu Sun2 , 1Milton Academy, 170 Centre Street, Milton, MA 02186 , 2California State Polytechnic University, Pomona, CA, 91768, Irvine, CA92620.
SingSense addresses the gap in traditional VR singing applications by integrating real-time biometric feedbacktocreate an entertaining and immersive musical experience. Utilizing heart rate variability data, the systemdynamically adjusts the kind of songs recommended to better align the virtual performance space with the users physiological state. Key components include accurate real-time HRV data capture, low-latency data transmission, and creating a responsive VR environment. Experiments demonstrated the systems ef ectiveness in enhancinguseremotional state and immersion through personalized interactions. By harmonizing technology with human emotion, SingSense of ers a novel platform that redefines interactive music experiences.
VR, Singing, Health
Abol Froushan, Fellow of the Royal Society of Arts, London, UK
Poetry presents unique challenges for natural language processing (NLP) due to its fragmented structure, intertextuality, and multimodal nature. Conventional NLP models struggle to capture its evolving semantic relationships, particularly across translations, historical contexts, and interpretative traditions. This paper introduces the Poetic Ontology Dataset (POD), a structured resource designed to embed poetic meaning as a dynamic, topological construct rather than a static textual entity. By applying sheaf theory, functorial mappings, and graph embeddings, we model poetic motifs and metaphors as interdependent structures within a meta-body—a network of poetic relations spanning time and cultures. Empirical validation compares AI-assisted meta-body analysis, derived through structured motif tracking and graph clustering, against traditional NLP embeddings. The results demonstrate that ontology-aware NLP models preserve semantic continuity and intertextual depth more effectively than conventional approaches. This work establishes a foundation for meaning-aware NLP architectures, bridging computational poetics, multimodal embeddings, and topological data analysis.
Computational Poetics, Ontology-Based NLP, Sheaf Theory in NLP, Functorial Mapping in AI, Poetic Meaning Representation, Graph-Theoretic Poetic Analysis
Jingyi Liu1, Jonathan Sahagun2, 1Tarbut V’Torah Community Day School, 5200 Bonita Canyon Dr, Irvine, CA 92603, 2Computer Science Department, California State Polytechnic University, Pomona, CA 91768
Exercise has become an increasingly popular trend, but injuries caused by overexertion remain a concern. To address this issue, we developed an AI-powered application that utilizes face detection to prevent users from overexercising. Our application integrates Vertex AI and Bluetooth Low Energy (BLE) technology to enhance connectivity and performance. During development, we encountered challenges in detecting user fatigue and establishing a reliable connection between the application and the treadmill. To validate our system, we conducted experiments to assess the accuracy of AI face detection across different users. Our findings indicate that the AI successfully detects faces regardless of age; however, it struggles to recognize individuals with darker skin tones due to limited training data. Compared to other methods, our application offers a proactive approach to injury prevention during exercise. Additionally, it includes a mechanism to forcibly stop the user if they ignore fatigue warnings, ensuring a safer workout experience.
AI, Facial Recognition, Fitness Equipment, Exercise Safety
Zhengyang He1, Rodrigo Onate2, 1St. Andrews School, 63 Federal Road, RI 02806, United States, 2Computer Science Department, California State Polytechnic University, Pomona, CA 91768
This research presents an AI-powered basketball shot analysis system that integrates YOLO object detection and pose estimation to evaluate shooting mechanics [1]. The system detects key components such as basketball, hoop, and player movement, while tracking elbow, shoulder, and knee angles to assess shot accuracy and provide actionable feedback. The backend processes upload videos, detecting whether a shot was made and analyzing player movements, while the frontend displays AI-generated insights and stores feedback in Firebase for progress tracking [2]. Two experiments were conducted to evaluate system performance. The shot detection accuracy test showed an 89% overall accuracy, correctly identifying 86% of made shots and 92% of missed shots. The pose estimation test measured a mean absolute error of 4.2° for elbow angles, 5.1° for shoulder angles, and 4.8° for knee angles, confirming high reliability. However, low-light conditions and extreme camera angles introduced detection errors, suggesting improvements through data augmentation, real-time processing, and optimized model training. By providing automated, AI-driven shooting feedback, this system offers a cost-effective alternative to personal coaching, making basketball training more accessible, efficient, and data-driven for players at all skill levels.
Basketball Shot Analysis, YOLO Object Detection, Pose Estimation, AI-Powered Sports Training, Automated Shooting Feedback
Sahana, Irin Prima, Department of Finance, Aalborg University, Denmark
This project explores the potential of personalized meal recommendation systems, focusing on user preferences and nutritional goals. The aim is to improve user satisfaction and dietary adherence by offering tailored food suggestions. We analyze data collection methods, recommendation algorithms, and challenges such as user diversity and food data complexity.
Meal Recommendation, Personalization, Nutrition, Data Science, AI, NLP.
Tyler Kaiyang Chen1, Yen-Hao Wang2, 1Northwood High School, 4515 Portola Parkway, Irvine, CA 92620, 2Computer Science Department, California State Polytechnic University, Pomona, CA 91768
This paper presents the design, implementation, and evaluation of a Golf Swing Analyzer, a low-cost, accessible system that delivers real-time feedback on golf swing mechanics [1]. Our system leverages MediaPipe for pose estimation, and a rule-based machine learning model training on labeled golf swing images to assess the swing based on parameters like elbow stability and shoulder posture. The backend, built with Python Flask, processes user inputs and runs swing analysis while the frontend provides an intuitive interface for ease of use [2]. To validate our approach, we conducted an experiment with 20 diverse swing images, which highlighted issues such as image blur, incorrect camera angles, and background distractions that impacted prediction accuracy. Compared to existing methods using expensive motion capture systems, or deep neural network-based analysis, our approach is faster, more accessible, and does not require expensive equipment or large training datasets. By improving accessibility and affordability, Perfect Pivot enables golfers of all skill levels to refine their swing, making golf coaching more inclusive and improving their technique more conveniently.
Golf Swing Analysis, Pose Estimation, Machine Learning, Real-Time Feedback
Baopu Tai1, Mirna Shabo2 , 1Maranatha High School, 169 S. Saint John Avenue, Pasadena, CA 91105 , 2Computer Science Department, California State Polytechnic University, Pomona, CA 91768.
ShotTrainer is an AI-powered basketball training application designed to improve shooting accuracy by utilizing YOLO AI-based video analysis [1]. Traditional basketball training methods rely on manual observation, which can be inaccurate and time-consuming [2]. ShotTrainer automates shot detection and performance tracking, providing real-time feedback to players and coaches. Unlike previous research focusing on kinematic analysis, AI method reviews, or strategy-based coaching tools, ShotTrainer is a fully implemented, player-centric solution that allows users to analyze their shot attempts, makes, and misses effortlessly. The system was tested in an experiment analyzing 20 basketball shot videos under different conditions to measure its accuracy. The AI’s expected outputs were compared to actual predictions, resulting in 90% accuracy—with 18 correctly classified shots and 2 misclassified ones. The misclassifications were primarily due to lighting variations and camera angles, where overexposure (bright sunlight) and low-light conditions affected the AI’s ability to track the ball trajectory accurately. These issues suggest that the model was trained on a dataset that lacked diverse lighting and background conditions, impacting its performance in extreme scenarios. Future improvements will focus on expanding the training dataset and fine-tuning detection thresholds to enhance the model’s robustness and accuracy. By offering a cost-effective, accessible, and practical solution for basketball training, ShotTrainer bridges the gap between AI and real-world sports analytics, making it an essential tool for players, coaches, and enthusiasts aiming to enhance their shooting performance.
AI-based Training, Basketball Performance Analytics, YOLO Object Detection, Shot Accuracy Tracking
Theodora-Stavroula Korma Department of Communication and Information studies, Rijkuniversiteit Groningen, Groningen, The Netherlands
Predictive policing, an algorithm-driven crime prevention initiative, claims to render the criminal justice system more effective and neutral. Yet, this essay argues that these algorithmic models reinforce system-level prejudices and unfairly focus on over marginalized populations while amplifying injustice.As these models draw from historical data covering four decades shaped by biased police operations, they can magnify racial profiling and harden social hierarchies. Furthermore,Furthermore, these systems lack of transparency and accountability has ethical consequences on surveillance, due process, and civil rights violations. In line with Design Justice principles, this paper calls for a redesign of predictive policing that is not about control by systems but the empowerment of communities. Instead of being used as enforcement tools, these algorithms must be redesigned to address root causes of social harm, promote equitable resource allocation, and engage communities in decision-making. Through participatory governance and moral algorithmic design, predictive technologies can serve justice rather than subvert it, so that communities are protected, not monitored.
Predictive policing, algorithmic bias, systemic injustice, racial profiling, Design Justice.
Ahmad Mahmood1, Zainab Ahmad,1 Iqra Ameer2, and Grigori Sidorov1 , 1Instituto Politecnico Nacional (IPN), Centro de Investigación en Computacion(CIC), Mexico City, Mexico ,2Division of Science and Engineering, The Pennsylvania State University, Abington, PA, USA
The development of medical question-answering (QA) systems has predominantly focused on high-resource languages, leaving a significant gap for low-resource languages like Urdu. This study proposed a novel corpus designed to advance medical QA research in Urdu, created by translating the benchmark MedQuAD corpus into Urdu using the Generative AI-based translation technique. The proposed corpus is evaluated using three approaches: (i) Information Retrieval (IR), (ii) Cache-Augmented Generation (CAG), and (iii) Fine-Tuning (FT). We conducted two experiments, one on a 500-instance subset and another on the complete 3,152-question corpus, to assess retrieval effectiveness, response accuracy, and computational efficiency. Our results show that JinaAI embeddings outperformed other IR models, while OpenAI 4o mini, FT achieved the highest response accuracy (BERTScore: 70.6%) but is computationally expensive. CAG eliminates retrieval latency but requires high resources. Findings suggest that IR is optimal for real-time QA, Fine-Tuning ensures accuracy, and CAG balances both. This research advances Urdu medical AI, bridging healthcare accessibility gaps
Information retrieval, retrieval-augmented generation, cache-augmented generation, fine-tuning, Urdu medical question-answering
Gianni Jacucci, Department of Information Engineering and C.S., University of Trento, Italy
This focused and dedicated review essay examines the architectural design work of a Trustworthy Governable Platform (TGP), that redefines the information paradigm for institutional communication platforms—particularly in relational enterprises—where complex social dynamics are at play. Examples include addressing accountability and ensuring contextual integrity. We adopt a design-oriented sociological perspective to uncover how the platforms creator ensured that its social use and meaning align with the specific situational needs from the outset. The study begins by examining the key features of social interactions that characterise relational enterprises, as identified and articulated by the author of the design process through selected theoretical approaches. Next, we analyse the communicative affordances of technology in action that are necessary to realise these social interaction features. Finally, we explore the design elements of the supporting information and communication infrastructure that the author deemed essential to enable the appropriate options for action. These elements ensure users experience the intended social usage and meaning of the technology in the context of relational enterprises. The purpose of the essay is to stimulate discussion on what we deem a most promising enterprise. TGP has been under active technical development for a number of years and, at the time of writing, is undergoing initial trials for deployment and appropriation.
Relational Enterprises, Communications on Information, Entrepreneurial, Management & Organisation Cybernetics, Socio-cybernetics, Complex Adaptive System and Theory.
Raeanne Li1, Carlos Gonzalez2, 1The Harker School, 500 Saratoga Ave, San Jose, CA 95129, 2California State Polytechnic University, Pomona, CA, 91768, Irvine, CA 92620
For many students, mathematics is a challenging subject because of its abstract nature. Through my own learning experience, I realized how visual representations can simplify complex problems, making them intuitive and engaging. These moments sparked my passion for exploring how visualization can empower math learners, especially those with diverse learning needs and styles, to overcome the barriers associated with traditional teaching methods. Students with dyslexia, ADHD, or autism often face challenges with text-heavy explanations and abstract concepts, but visualization tools, such as charts, diagrams, and interactive models, can make a difference, offering alternative approaches that better support their learning [1]. This project focuses on developing an AI-powered visualization platform designed to generate visual representations for math word problems. With Retrieval-Augmented Generation (RAG), the smart system retrieves relevant data from external sources and generates content-specific math problems, ensuring high accuracy and alignment with user queries. A key contribution of this research is the integration of a dual-LLM architecture with RAG to enhance diagram creation [2]. The first LLM generates clear, concise, and imperative instructions from natural language queries, while the second LLM translates these instructions into valid Scalable Vector Graphics (SVG) code for precise diagrams [3]. The integrated approach allows for automated, scalable, and customizable diagram generation, offering an engaging and accessible learning experience for different problems. Ultimately, the smart system combines problem generation and visualization into a unified web and mobile application, providing diverse learners with powerful tools to engage with math.
Math Visualization, Accessibility in Education, Dual-LLM Architecture, AI-Powered Learning Tools.
Apurva Sinha and Ekta Gujral, Walmart Global Tech Sunnyvale, USA
Product attribute extraction is a growing field in e-commerce business, with several applications including product ranking, product recommendation, future assortment planning and improving online shopping customer experiences. Understanding the customer needs is critical part of online business, specifically fashion products. Retailers use assortment planning to determine the mix of products to offer in each store and channel, stay responsive to market dynamics and to manage inventory and catalogs. The goal is to offer the right styles, in the right sizes and colors, through the right channels to fostering customer loyalty. In this paper we present PAE, a product attribute extraction algorithm for future trend reports consisting text and images in PDF format. Most existing methods focus on attribute extraction from titles or product descriptions or utilize visual information from existing product images. Compared to the prior works, our work focuses on attribute extraction from PDF files where upcoming fashion trends are explained. Our contributions are three-fold: (a) We develop PAE, an efficient framework to extract attributes from unstructured data (text and images); (b) We provide catalog matching methodology based on BERT representations to discover the existing attributes using upcoming attribute values; (c) We conduct extensive experiments with several baselines and show that PAE is an effective, flexible and on par or superior (avg 92.5% F1-Score) framework to existing state-of-the-art for attribute value extraction task.
Attribute Extraction, PDF files, Large Language Model (LLM), Text and Images, BERT embeddings
Venkata Duvvuri, Chetan Kulkarni, Sritha Gogineni
COVID-19 pandemic has created a major impact around the world. Governments and businesses small or big around the world are facing unprecedent decisions to either close up or reopen or drive other policies based on the sentiment of people. While, understanding this sentiment and accompanying emotions has been researched especially in social media channels like Twitter, we propose a novel way to capture sentiment and emotions using intelligent chatbots (EmoBot) that reduces the participants biases inherent in prior analysis. We devise Emotion Extraction Layers (EEL) based on latest deep learning techniques like BERT (Bidirectional Encoder Representations from Transformers) and compare these models with traditional machine learning models. We show for a variety of emotions that the new deep learning models predict 1-5% (Sad, Fearful & Angry) better than traditional machine learning techniques. Further, we showcase that leveraging retail sentiment data using transfer learning techniques can help cross the cold start chasm of having no chatbot data initially, and this technique achieves -8% closer in performance when compared to having enough COVID sentiment data.
COVID, sentiment analysis, chatbots, BERT, deep learning, transfer learning.
Shafaq Khan, Anirudh Grack, Saksham Thukral, Kausar Fatema, Aryan Batra, Jaykumar M Kadiwala, School of Computer Science, University of Windsor, Windsor, ON
Predicting energy consumption in buildings based on factors such as size, design, usage patterns, and weather presents significant challenges. Comprehensive analysis of this research focuses on developing data-driven models for early-phase energy forecasting in buildings. This initiative is pivotal for energy- efficient building design and contributes significantly to energy planning, management, and conservation efforts. Utilizing data from 1500 buildings across various categories, the study integrates data preprocessing, mining, and analysis techniques along with machine learning algorithms such as Linear Regression, Decision Trees (DT) , Extreme Gradient Boosting (XGBoost), and Light Gradient-Boosting Machine (LightGBM). Unique to the approach is the explicit incorporation of advanced feature engineering techniques and weather data into the modeling process, filling critical gaps in existing methodologies. This study demonstrates that LightGBM outperforms alternative models and ensures the designed model is adaptable across various meter types, enhancing its universal application. Aligned with the United Nations (UN) 2030 The research agenda is dedicated to advancing global environmental sustainability by significantly reducing CO2 emissions and fostering substantial investments in building energy efficiency.
Forecasting, Energy consumption, Light-GBM, XGBoost, Decision Trees, Linear Regression, Load prediction.
Anurag Sing1and Sheenu Rizvi, Department Of_Computer_Science & Engineering, Amity University,India
Artificial Intelligence (AI) hasemerged as a transformativeforce in medical, providing innovative ideas that enhance diagnostic accuracy and better patient outcomes. The incorporation of A.I. technologies like machinelearning(ML), deep learning(DL), and natural languageprocessing(NLP) into diagnostic processes holds the potential to revolutionize medical practice. By rapidly analyzinglarge datasets, recognizing complex patterns, and facilitating data-driven decisions, AI is aiding healthcare professionals in reducing diagnostic errors and delivering personalized treatments. This paper provides an in-depth analysis of AIs impact on diagnostic processes in healthcare, real-world applications, benefits, challenges, ethical considerations, and future directions for AI technologies. The paper also explores how AI-enabled systems are reshaping healthcare workflows and empowering healthcare providers to make more informed decisions, ultimately leading to better patient outcomes.
Artificial Intelligence, Healthcare, Diagnostic Accuracy, Machine Learning, Deep Learning, Personalized Medicine, Patient Outcomes, Ethical Considerations.
Yuting Gao1, Ang Li2, 1Crean Lutheran High School, 12500 Sand Canyon Ave, Irvine, CA 92618, 2Computer Science Department, California State University, Long Beach, CA 90840
MuseComposer is a user-friendly app for generating music and lyrics using AI [1]. It simplifies music creation by offering intuitive interfaces and personalized outputs based on user prompts. Experiments revealed high satisfaction with ease of use and melody generation but identified improvements needed in lyrics accuracy [2]. The apps integration of AI, a robust database, and dynamic playback ensures accessibility and creativity. With future refinements, MuseComposer can redefine music composition for users of all skill levels [9].
AI-powered music generation, User-friendly interface, Personalized melodies, Creative lyric generation, Music composition accessibility.