2026 IEEE International Conference on AI and Data Analytics
(ICAD 2026)
June 11 – 12, 2026
Boston, Massachusetts
Room 115 abstracts and final papers
189: Validation-Conditioned Dynamic Ensemble Regression with Applications in Biomedical Data presented by Brandon Warner (Verseon International Corp.)*; Edward Ratner (Verseon International Corp.); Elliot Farmer Garcia (Verseon International Corp.)
Ensemble learning enhances predictive accuracy through two complementary mechanisms: the construction of diverse base learners and the strategy by which those learners are combined at inference time. Although dynamic ensemble selection (DES) has been extensively studied in classification settings, its application to regression remains comparatively limited due to difficulties associated with continuous response variables, heightened sensitivity to noise, and the risk of validation leakage. In this paper, we introduce a simple and leakage-aware dynamic ensemble selection framework tailored for regression tasks that extends a globally trained ensemble with instance-specific adaptive weighting derived from validation data. For each query sample, a local region of competence is defined using KD-tree–based neighborhoods within a held-out validation set, and model competence is assessed using neighbor-conditioned regression error rather than distance-weighted loss functions. A validation-tuned gating mechanism subsequently balances local and global ensemble contributions based on relative distance and relative error cues. Empirical evaluation on standard regression benchmarks and a large-scale biomedical dataset demonstrates consistent performance gains over static ensembles and existing regression-focused DES methods. The proposed framework is modular, interpretable, and computationally efficient, offering a robust and transparent baseline for dynamic ensemble regression without dependence on meta-learning strategies or auxiliary predictive models.
194: Can Physics-Informed Neural Networks Learn Regulatory Compliance Bounds? Isolating Calibration Drift in Smart Heat Meters presented by Gideon Mbiydzenyuy (University of Borås)*; Fimon Yacob (Ekkono Solutions); Saleh Javadi (Blekinge Institute of Technology)
Thermal energy meters in district heating networks must comply with regulatory Maximum Permissible Error (MPE) thresholds. Detecting meters exceeding these limits due to calibration drift—rather than consumption pattern changes—is critical for compliance and billing accuracy. However, observed measurement deviations contain both calibration drift (physical sensor degradation) and behavioral drift (changes in consumption patterns). We investigate under what conditions a physics-informed neural network (PINN) can learn a latent calibration bias parameter β (serving as a risk indicator, not a direct calibration correction) that aligns with regulatory MPE thresholds. Using thermodynamic energy conservation as a physics constraint, we train an autoencoder on smart heat meters from three Swedish utilities and analyze the learned β. Our key findings are: (1) β decomposes into a fleet-level systematic offset (e.g., mean |β| ≈ 2.4–5.9% during a winter regime) and meaningful per-meter variation (e.g., σ_β = 0.5–2.3%); (2) a temporal stability metric η_stable = 1 – MAD(β)/mean|β| achieves 0.73–0.79 in winter, indicating stable within-meter calibration estimates; (3) across three utilities and 620 meters (winter), 16–26% exceed ±6% MPE bounds, providing actionable compliance candidates.
198: ULAF: Learnable Geometry Activation for Deep Neural Networks presented by Saikatesh Dash (TEKsystems Inc.)*; Sradhanjali Paty (Intone Networks Inc.); Elif Kongar (University of New Haven)
We evaluate ULAF under controlled experimental conditions on representative vision and language benchmarks, comparing it against strong contemporary activation baselines. The results demonstrate consistent performance improvements, stable convergence behavior, and favorable calibration characteristics across tasks. Learning activation geometry thus provides a principled extension to existing activation design, enabling more expressive and adaptable nonlinear transformations in modern deep neural networks.
201: Topological Data Analysis Integrated with Deep Sequence Models for Multivariate Time Series-based Solar Flare Prediction presented by Syed Abrab Mahmood (University of Massachusetts Lowell)*; Ruizhe Ma (University of Massachusetts Lowell)
Accurate solar flare prediction is critical for protecting astronauts, spaceborne equipment, and terrestrial infrastructure. We propose a hybrid deep learning framework that integrates Topological Data Analysis (TDA) with sequence-based models to enhance multivariate time-series (MVTS) forecasting. While traditional models focus on temporal dependencies, TDA captures the intrinsic geometric and topological structures of active region evolution. By transforming MVTS trajectories into persistence images, our method preserves both global and local topological features that effectively distinguish flaring from non-flaring regions. Utilizing a rigorous preprocessing pipeline, including imputation, normalization, and class-imbalance mitigation, we evaluate our model on the Space Weather Analytics for Solar Flares (SWAN-SF) dataset. Experimental results demonstrate that incorporating topological descriptors significantly improves model performance, confirming TDA as a powerful tool for solar flare prediction.
204: Automated Highlight Generation using Deep Learning in Soccer Matches presented by Jenish Kothari (Northeastern University)*; Yash Phalle (Northeastern University)
This paper presents a deep learning-based framework for automated highlight generation from full-length soccer matches. Given the sparsity and unpredictability of events such as goals, fouls, red-cards etc., manual highlight generation remains labor-intensive. We address this challenge by leveraging spatiotemporal architectures, including 3D Residual Network architecture (ResNet-3D) and hybrid Convolutional Neural Network models such as ResNet with Gated Recurrent Units (GRU) and ResNet with Long Short Term Memory Networks (LSTM), to classify 10-second video segments into fine-grained action categories. Our approach is trained and evaluated on the SoccerNet dataset using k-fold cross-validation and extensive hyperparameter tuning. This approach significantly reduces the time and efforts required for manual highlight generation and enables fans, editors and analysts to focus on important moments without watching the entire match.
227: Spectral Learning for Crack and Corrosion Risk Prediction presented by Insha Yaqoob Sheikh (Missouri University of Science and Technology)*; Sarangapani Jagannathan (Missouri University of Science and Technology)
This paper proposes a novel geometry aware spectral optimization framework that overcomes the limitations of existing spectral and singular value decomposition (SVD)-based regularization techniques in order to reduce gradient instability and decay through a reflector-based orthogonal parameterization, ensuring temporally coherent spectral evolution during training. Singular value perturbations are applied within a consistent spectral frame, and optimization is carried out directly in the factorized domain, enabling structure preserving and error driven learning. The proposed framework is applied to long short-term memory (LSTM) architectures for automated crack and corrosion risk forecasting. Empirical evaluations show improved predictive accuracy in long horizon sequence modeling, demonstrating the advantages of geometry consistent spectral learning for complex spatiotemporal inference tasks.
246: Exploring the Impact of Dataset Statistical Effect Size on Model Performance and Data Sample Size Sufficiency presented by Arya Hatamian (University of California, Riverside); Lionel Levine (UCLA); Haniyeh Ehsani Oskouie (UCLA)*; Majid Sarrafzadeh (UCLA)
Having a sufficient quantity of quality data is a critical enabler of training effective machine learning models. Being able to effectively determine the adequacy of a dataset prior to training and evaluating a model’s performance would be an essential tool for anyone engaged in experimental design or data collection. However, despite the need for it, the ability to prospectively assess data sufficiency remains an elusive capability. We report here on two experiments undertaken in an attempt to better ascertain whether or not basic descriptive statistical measures can be indicative of how effective a dataset will be at training a resulting model. Leveraging the effect size of our features, this work first explores whether or not a correlation exists between effect size, and resulting model performance (theorizing that the magnitude of the distinction between classes could correlate to a classifier’s resulting success). We then explore whether or not the magnitude of the effect size will impact the rate of convergence of the learning curve, (theorizing again that a greater effect size may indicate that the model will converge more rapidly, and with a smaller sample size needed). Our results appear to indicate that this is not an effective heuristic for determining adequate sample size or projecting model performance, and therefore that additional work is still needed to better prospectively assess adequacy of data.
249: Uncertainty-Aware Forecasting with Shifting Seasonal Matrix Factorization presented by Jacob Munson (Montana State University)*; Breschine Cummins (Montana State University)
Uncertainty estimation remains a difficult problem in machine learning. Where possible, principled uncertainty estimation often requires extensive model alterations, repeated model runs, or expensive sampling-based methods. In this paper, we extend the existing Shifting Seasonal Matrix Factorization (SSMF) model to include meaningful uncertainty estimates without major model modifications. We use a conformal prediction framework with multiple variants to estimate prediction intervals based on a user-specified quantile. Our results show that permitting time-dependent variance in model residuals accurately models coverage over tight prediction intervals.
202: Modular Deep Learning for Multivariate Time-Series: Decoupling Imputation and Downstream Tasks presented by Joseph Arul Raj (Kings College London)*; Zina Ibrahim (Kings College London); Linglong Qian (Kings College London)
Missing values are pervasive in large-scale time-series data, posing challenges for reliable analysis and decision-making. Many neural architectures have been designed to model and impute the complex and heterogeneous missingness patterns of such data. Most existing methods are end-to-end, rendering imputation tightly coupled with downstream predictive tasks and leading to limited reusability of the trained model, reduced interpretability, and challenges in assessing model quality. In this paper, we call for a modular approach that decouples imputation and downstream tasks, enabling independent optimisation and greater adaptability. Using the largest open-source Python library for deep learning-based time-series analysis, PyPOTS, we evaluate a modular pipeline across six state-of-the-art models that perform imputation and prediction on eight datasets spanning multiple domains. Our results show that a modular approach maintains high performance while prioritising flexibility and reusability – qualities that are crucial for real-world applications. Through this work, we aim to demonstrate how modularity can benefit multivariate time-series analysis, achieving a balance between performance and adaptability.
218: Generation of Synthetic Samples by Large Language Models in the Context of Cross-Country Surveys presented by Elizaveta Kuznechenkova (Higher School of Economics), Maria Nesterskaya (Higher School of Economics)*,Vera Semyonova (Higher School of Economics)
This paper examines the feasibility of using large language models (ChatGPT and YandexGPT) to generate synthetic survey data that emulate responses in cross-national sociological surveys. Using items from the World Values Survey, including ethical topics (e.g., sexual and reproductive practices, suicide and euthanasia, and the justification of violence), we analyze how models answer normative questions in general and whether they reproduce empirical response variance. We compare the distributions of real and synthetic responses, quantify cross-country deviations using Brunner-Munzel statistics, and assess the coherence of country rankings using Kendall’s rank correlation coefficient across related items and between synthetic and empirical data. We find that both models distort key distributional properties: they tend to compress variance (often via centering) and, for several legality/violence items, produce uniformly strong moral disapproval with near-zero variance, in contrast to real survey patterns. Both models also differentiate countries along a stereotypical “progressive”–“conservative” axis; this clustering is more systematic for ChatGPT and less consistent for YandexGPT, and it aligns with WVS evidence only partially. Country-level socioeconomic predictors (e.g., GDP and HDI) exhibit no ro- bust effects, suggesting that cross-national differentiation reflects learned stereotypes rather than explicit developmental indicators. Finally, using fixed-effects and country-level regressions, we show that ChatGPT is highly responsive to prompt-provided demographic attributes, whereas YandexGPT displays weaker and more selective sensitivity, with religious affiliation emerging as a recurrent predictor. A supplementary re-run in autumn 2025 indicates temporal instability in point responses but partial persistence of higher-level patterns, underscoring the need for further research on reproducibility, cross-national bias, and the ethical and methodological implications of LLM-based synthetic data in the social sciences.
ROOM 117 ABSTRACTS AND FINAL PAPERS
128: Privacy-Preserving Federated AI on Blockchain for Secure and Compliant Banking Data Management presented by Vineeth Lakkadi (University of the Cumberlands)*; Praneetha Reddy Donthi (University of the Cumberlands); Sravani Kandula (University of the Cumberlands); Munish Kumar (University of the Cumberlands)
The paper suggests a blockchain-based artificial intelligence (AI) system of managing customer data securely, auditably, and with data sovereignty in the banking industry by using federated and privacy-preserving learning algorithms. The TGNN-Transformer is a hybrid model combining behavioral, relational, and compliance-contextual data. Hybrid Temporal Graph Neural Network-Transformer Proposes to detect fraud, personalize the services, and handle the risk, but incorporates privacy-guaranteed via functional encryption and zero-knowledge proofs. These cryptographic units produce verifiable and encrypted model advances which cannot be traced back to individual clients but which can be merged with a permission blockchain-coordinated federation. The architecture is a consortium-based one and has the capability to achieve high throughput, low-latency execution that is suitable to be implemented across heterogeneous banking infrastructures. Tests on emulated multi-bank environments revealed a Fraud Detection Accuracy (FDA) of 96.8 percent and a Federated Convergence Rate (FCR) of 92 Percent, a throughput of 1,450 transactions per second, and a privacy assurance score of 0.98. The findings verify that federated AI, using blockchain, can increase security, compliance with regulatory requirements, and customer trust in the next-generation banking applications, and manage financial data ecosystems at scale and in a privacy aware manner.
145: HAGS: A Hand and Glove Segmentation Dataset for Collaborative Assembly presented by Shivansh Sharma (The University of Texas at Austin); Mathew Huang (The University of Texas at Austin); Sanat Nair (The University of Texas at Austin); Alan Wen (The University of Texas at Austin); Christina Petlowany (The University of Texas at Austin); Juston Moore (Los Alamos National Laboratory); Selma Wanna (Los Alamos National Laboratory)*; Mitch Pryor (The University of Texas at Austin)
Manufacturing in U.S. Department of Energy (DOE) research facilities differs from commercial production: it is low-volume, highly specialized, and often involves handling hazardous materials in confined spaces. Industry 4.0 concepts, such as collaborative robotics, could improve safety and throughput in DOE research and prototyping workflows. However, this automation demands reliable, real-time perception systems that can operate under strict safety criteria. One necessary capability is glove segmentation. Hand segmentation datasets from prior work do not generally capture visual complexity stemming from industrial workflows, e.g., reflective surfaces, robots, or protective equipment such as gloves; and the datasets which target industrial domains largely rely on synthetic data, which may not transfer to real-world tasks. Often both datasets neglect out-of-distribution (OOD) examples necessary for safety validation via uncertainty estimation. To address these gaps, we introduce HAGS: the Hand And Glove Segmentation dataset for glovebox environments which includes surrogate, small-batch cobotic manufacturing scenarios and OOD images generated via green-screen augmentation to support uncertainty analysis. Our \href{https://dataverse.tdl.org/dataset.xhtml?persistentId=doi:10.18738/T8/85R7KQ}{dataset} and \href{https://github.com/UTNuclearRoboticsPublic/assembly\_glovebox\_dataset}{baselines} are publicly available.
215: Integrating Graph Neural Networks and Large Language Models for Climate Policy Shock Contagion in Bank Lending Networks presented by Rohit Nimmala (Bank of America); Pavan Nutalapati (Independent Researcher); Jagrut Nimmala (Independent Researcher); Milan Parikh (Independent Researcher); SIVA RAMA KRISHNA varma Bayyavarapu (Independent researcher)*; Tejas Patel (Independent researcher)
Current regulatory climate stress tests model bank exposures independently, ignoring network amplification effects that propagate losses through interconnected financial systems. Empirical research shows that supply chain contagion amplifies climate-driven losses by a factor of 4 – 5x (Tabachov´a et al., 2024). Building on this foundation, we introduce ClimateNet, a hybrid architecture combining Graph Neural Networks and Large Lannguage Models for climate policy shock analysis in bank lending networks. Our system uses Relational Graph Convolutional Networks with auditable shock injection mechanics to model contagion dynamics across loan portfolios and employs LLMs to generate interpretable scenario narratives through GNN-RAG retrieval of contagion pathways. We evaluate ClimateNet on Hungarian central bank data, demonstrating network-amplified loss estimates consistent with prior empirical findings while adding learned representations and natural language explanations. We further present a validation protocol comprising no-graph ablation, edge intervention tests, and LLM faithfulness audits. ClimateNet extends existing supply chain contagion methods with interpretable outputs for financial regulators and institutional risk managers.
150: Forecast to Replenishment: Interpretable Machine Learning for Retail Inventory Optimization presented by Marsa Rayani (California state University DH)*; Samarasimha Donthireddy (California state university DH); Alireza Izaddoost (California state university DH); Benyamin Ahmadnia (California state university DH)
151: A Systems Engineering Driven Framework for Integrating Large Language Models into Clinical Documentation to Increase Efficiency presented by Anoushka Vijay (Notre Dame High School San Jose CA); Jyotirmay Gadewadikar (Mitre)*
Artificial Intelligence (AI) is rapidly changing the landscape of the healthcare industry with its promise to make the process significantly efficient. Large Language Models (LLMs) are becoming powerful tools to support clinicians in managing patient interactions and related documentation. Using LLMs in healthcare flows not only require technical capability, but also deeper integration of checks for safety and compliance. This paper reviews the real-world potential of LLMs to reduce administrative workload and improve the quality of clinical documentation. Specifically, we explore how LLMs can streamline patient data collection during clinical visits. We propose a structured framework to detect and mitigate hallucination risks through quantitative error analysis. Our prototype simulation validating the transcript-to-EHR extraction phase indicates a potential 37.7% overall workflow efficiency increase, reducing administrative time by an average of 42.44 minutes per session. By addressing both the potential to integrate LLMs and identifying the risks, this paper contributes to an actionable approach to the responsible use of AI in healthcare.
166: Adoption of Safe and Responsible AI for Streamlining Communication with Verbally Impaired Communities presented by Mahendra Samarawickrama (Meta61)*
Deaf and verbally impaired communities face significant communication barriers across social, educational, and clinical settings. Developments within artificial intelligence (AI) and computer vision create pathways towards inclusive communication support. This paper presents a comprehensive examination of responsible AI use in real-time hand-sign recognition, with a particular focus on the design, implementation, and evaluation of a lightweight portable prototype. The system outlined utilises hand-landmark extraction and a classification model to recognise American Sign Language (ASL) letters from webcam input, implemented in a browser environment. This prototype demonstrates accessible deployment with general-purpose resources. A structured literature review evaluates existing assistive technologies, legal and policy frameworks, and evolutions of AI use in communication support systems. Further analysis integrates risk and strategy frameworks, alongside psychological and moral theories, to assess the broad impact, sustainability concerns, and ethical implications of implementing AI-assisted hand-sign recognition. The findings indicated widespread potential applications and positive psychological benefits for the impacted communities. The importance of ethical safeguards and inclusive design is further underscored. Recommendations for future development and diverse sector-specific applications are provided, situating this paper’s work within global efforts to advance sustainability goals and promote equity in communication.
219: Diversity Under Domain Conditioning: Comparing SimSon and CONSMI for QAC Generation presented by Shahana Shultana (George Mason University ); Shiva Ghaemi (George Mason University ); Farzad Ahmed (George Mason University ); Amarda Shehu (George Mason University ); Daniel Barbara (George Mason University )*
These findings establish that contrastive objective choice is a primary control lever for the diversity-fidelity trade-off in domain-conditioned molecular generation, with practical implications for when to prioritize exploration (SimSon) versus semantic concentration (CONSMI) in controllable synthesis.
248: Benchmarking Hybrid Deep Learning Architectures for Predictive Maintenance in Industry 4.0 presented by Zhengyang Gu (Liveline Technologies)*; Joseph Hernandez (Liveline Technologies); Thomas Cook (Liveline Technologies); John Burtenshaw (Liveline Technologies); Sean Scott (Liveline Technologies); Chris Couch (Liveline Technologies)
We found that while Transformers excelled at tracking stable, slow-moving processes, they tend to overreact to chaotic data, mistakenly taking sensor noise for meaningful signals [6]. We also found that the hybrid method that combines a Long Short-Term Memory (LSTM) layer with a Transformer layer is more resilient to noisy data from factory shops [7]. Functioning as a noise filter, the LSTM smooths out data volatility, allowing the Transformer to focus on the bigger picture without being distracted [8]. The hybrid model did not just improve accuracy; it proved to be significantly more consistent than complex models, delivering reliable predictions regardless of how chaotic the underlying system became.
234: On The Limits of Predicting Sex-Offense Escalation from Criminal Histories presented by Suli Adeniye (Arizona State University)*; Dominique Roe-Sepowitz (Arizona State University); Arunabha Sen (Arizona State University)
Predicting whether an individual will escalate to a sex offense, given their prior offenses, remains a critical but unresolved challenge in criminal-justice analytics. Prior work in recidivism prediction, spanning traditional risk-assessment tools and modern machine-learning models, has reported mixed findings, often limited by short observation windows, sparse temporal information, or reliance on demographic factors with well-documented fairness concerns. In this study, we investigate the limits of predicting sex-offense escalation using a large, real-world dataset obtained from a major U.S. metropolitan police department, focusing on individuals with prior drug involvement. The dataset consists of image-based booking records that we digitize, verify, and reconstruct into long-form offense trajectories containing drug offenses, other non–sex offenses, and sex offenses. We engineer sequence-based features using sliding windows of varying lengths, incorporating offense types, temporal spacing, and booking-time attributes such as age, gender, tattoo indicators, multi-state arrest histories, and charge patterns.
ROOM 135 ABSTRACTS AND FINAL PAPERS
160: Shazam4Code: An explainable method for detecting derivative clones for ensuring code providence presented by Jamie Heller (Tufts University)*; Samuel Guyer (Tufts University, Veracode Inc.)
Copying and pasting code is common and has implications in software maintenance, security, legal, and plagiarism applications. This behavior may have a poison pill effect that propagates a bug or even violates a software license. Despite the availability of AI to search large dataset corpus, there are fundamental limitations that inhibit LLMs from scaling efficiently and providing explainable results for these code providence applications. One major issue is the number of one-to-one matches (O(N)) that AI must perform, which doesn’t scale and is too costly. LLMs also suffer from semantic hallucination and are sensitive to prompt quality. AI agentic code generators also create duplicate code blocks, increasing the risk of vulnerabilities and maintenance issues. Furthermore, there are no quantitative metrics of assessment that are independent of human evaluators to establish the ground truth for code clones. Utilizing hybrid approaches that use offline processing to generate embedded representations can help alleviate some of these LLM challenges. The goal of this work is to present an efficient, explainable search system to determine if a code snippet is a derivative of known code, resulting from copying and pasting with possible edits. This will help to quickly identify code provenance, reducing the risk of license violations and helping to trace bug propagation. The methods we introduce will help address these foundational limitations of AI approaches and establish a new quantitative metric called shazaM trEe Edit disTance pRoxy(MEETER) to help reduce the dependency on human evaluators when establishing code clone ground truth datasets. The new framework approach called Shazam4Code has three major stages, source code normalization through canonicalization, functional fingerprinting through Abstract Syntax Tree (AST) decomposition, and efficient fingerprint matching. The search cost of our approach grows sub-linearly with database size. We evaluate our framework for accuracy and performance using BigCloneBench and the iJaDataset.
212: Manual Elasticity Models to Scalable Pricing Intelligence: An Industry Case Study presented by Girish Vasudevan (Vanguard)*
Pricing differences between competing financial products play a significant role in influencing fund flows; however, many financial institutions rely on manual and legacy approaches to assess pricing sensitivity. Traditional elasticity analyzes are difficult to scale, require substantial analyst intervention, and often focus primarily on price while under representing broader market and behavioral drivers. This paper presents an industry case study describing the development and deployment of a scalable machine learning based pricing intelligence framework designed for enterprise use. This framework systematically evaluates pricing sensitivity across products by including price, non-price factors, including market conditions, product attributes, and customer behavior signals. Special importance was placed on interpretability, governance, and operational integration, rather than algorithmic complexity. The model outputs were embedded into existing business intelligence workflows to support recurring analysis and decision-making by the pricing and product teams. The resulting framework reduced manual analytical effort, improved consistency across pricing evaluations, and improved organizational confidence in pricing related decisions. This paper discusses practical implementation, governance requirements, considerations, and lessons learned relevant to financial services organizations seeking to modernize pricing analytics in a controlled and sustainable manner
163: Applied AI-Enhanced RF Interference Rejection presented by Rahul Jain (MIT Lincoln Laboratory)*; Pierre Trepagnier (MIT Lincoln Laboratory); Rick Gentile (MIT Lincoln Laboratory); Joey Botero (MIT Lincoln Laboratory); Alexia Schulz (MIT Lincoln Laboratory)
AI-enhanced interference rejection in radio frequency (RF) transmissions has recently attracted interest because deep learning approaches trained on both the signal of interest (SOI) and the signal mixture (SOI plus interference) can outperform traditional approaches which only consider the SOI. The goal is to detect, demodulate, and decode signals over a range of signal-to-interference-plus-noise (SINR) levels without having a detailed, design-level knowledge of the interfering signal or the propagation conditions. Our present AI interference suppression results are based on Autoregressive Transformer Decoder models which exhibit orders of magnitude faster throughput at inference time than WaveNet models developed in earlier work. As a specific example, we investigate an analog FM “Walkie Talkie” radio signal of interest in the presence of an Orthogonal Frequency-Division Multiplexing (OFDM) interferer. This type of interferer is near-ubiquitous in the current RF landscape. Our results clearly show the benefits of transformer-based interference mitigation in tactical settings. We show that unintelligible transmissions become intelligible via metrics such as Perceptual Evaluation of Speech Quality (PESQ), while overall latency is kept to a minimum using readily available lightweight GPUs such as a Jetson AGX Orin. We believe these same techniques can also be applied to a broader set of national security scenarios, as well as having commercial applications.
217: Interpretable Transfer Learning for Classifying Exoplanet Atmospheric Stability presented by Beneyaz Begum (University of Central Florida)*; Ramses Ramirez (University of Central Florida)
This paper presents an Interpretable Transfer Learning (I-TL) framework for exoplanet atmospheric stability classification using features derived from 1-D radiative-convective, cloud-free climate model. The model ingests synthetic atmospheric data comprising N2-CO2-H2O-H2 abundances and generates an estimate of net outgoing IR flux (FIR), net incoming stellar flux (FSOL), and planetary albedo, which helps classify exoplanets thermodynamic regimes. Building on Random Forest (RF), Neural Network (NN)-based variants (with a regularizer, PCA-, MI-, and autoencoder) and CNN classifiers, the I-TL framework facilitates generalization and knowledge transfer across three Sun-like star datasets, reducing computational costs and training time. This approach helps in both binary and multiclass classification. The models demonstrate comparable predictive performance, achieving F1 scores between 0.96 – 0.99 for the dominant “extreme cooling” and “extreme heating” classes while preserving the precision and recall. The feature attribution reveals FIR as most influential driver for thermal equilibrium and in determining the inner habitable zone.
238: Operationalization of Machine Learning with Serverless Architecture: An Industrial Implementation for Harmonized System Code Prediction presented by Sai Vineeth Kandappa Reddi gari (Schneider Electric Global)*; Santhoshkumar Jagadish (Schneider Electric Global)
Industrializing machine learning operations (MLOps) requires architectures that deliver scalability, automation, and operational reliability. This paper introduces a serverless MLOps framework that orchestrates the entire model life-cycle (data ingestion, training, deployment, monitoring, and retraining using event-driven pipelines and managed services). The architecture is model-agnostic, supporting diverse
inference patterns through standardized interfaces, enabling rapid adaptation without infrastructure over-head. To demonstrate its practical applicability, we present an industrial implementation for Harmonized System (HS) code prediction, a compliance-critical classification task in which short, unstructured product descriptions must be mapped to standardized codes used by customs authorities for global trade. Frequent updates and ambiguous descriptions make accurate classification challenging, and errors can lead to shipment delays and financial losses. Our solution employs a custom text embedding encoder and multiple deep learning architectures, with the best-performing model (Text CNN) achieving 98 percent accuracy on ground truth data. Beyond accuracy, the pipeline ensures reproducibility, auditability, and SLA adherence under variable loads, leveraging auto-scaling. A distinguishing feature of the framework is automated A/B testing, enabling dynamic model selection and safe promotion in production environments. One key consideration in model choice is cost-efficiency; while transformer-based models may achieve similar accuracy, their long-term operational cost for training and maintenance is significantly higher. While generative AI offers flexibility for exploratory tasks, deterministic classification with predictable latency and explainability is prioritized here over generalized models. Importantly, the architecture remains extensible to integrate transformer variants and LLM-based inference when appropriate. The first section of the paper introduces the deep learning architectures used, along with simulations and model comparison results. The second section discusses the industrialization of MLOps through a serverless architecture, demonstrating
the feasibility of automated retraining, prediction, and validation of HS codes. This work provides a replicable blueprint for the operationalization of machine learning using serverless architecture, enabling enterprises to scale confidently while optimizing both performance and economics.
206: Coordinate-Driven Random Forests – A Transferable Approach for Graph Data presented by Hansi Kalpana Yasodara Paththini Hetti Arachchige (Colorado State University)*;https://ieee-icad.org/wp-content/uploads/2026/06/206-1.pdf Anura Jayasumana (Colorado State University)
A framework combining Topology Coordinate (TC) embeddings with Random Forest classifiers is proposed as a lightweight and interpretable alternative to deep graph neural networks for large-scale graph learning. Structural information is encoded using topology coordinate embeddings computed via anchor-based distance measurements, while prediction is performed using a classical tree-based classifier. Extensive experiments on vision, molecular, and social graph benchmarks show that the framework achieves competitive accuracy while significantly reducing computational cost, memory footprint, and energy consumption. Cross-dataset evaluation from Pascal VOC-SP to COCO-SP demonstrates effective transfer learning in a zero-shot setting, where a Random Forest model trained once on the source dataset can be directly reused on the target dataset without any retraining. Energy and latency profiling further confirm that this reuse enables efficient deployment in resource-constrained and edge environments. The results indicate that explicit structural representations combined with non-neural learning models support efficient, reusable, and sustainable graph learning across datasets.
211: Diagnosing and Mitigating Privacy Risks in Natural Language Interfaces for Sensitive Databases presented by Suli Adeniye (Arizona State University)*; Dominique Roe-Sepowitz (Arizona State University); Arunabha Sen (Arizona State University)
Natural language interfaces offer an intuitive way to query complex graph databases, but in sensitive domains such as criminal justice, user queries must be sanitized before being sent to cloud-based LLMs. Building on our prior NL2Cypher interface for law-enforcement data, this paper reveals that upstream entity masking, intended to protect personally identifiable information, fails under irregular or adversarial user inputs. Using a formal threat model and a collection of 200 real-world queries obtained in coordination with a major U.S. metropolitan police department (100 standard queries for diagnostic analysis and 100 adversarially-perturbed variants), we identify systematic vulnerabilities in both regex- and NER-based masking, including missed cue-less ages, boundary fragmentation, and obfuscation induced failures. To address these weaknesses, we develop an adversarial evaluation framework aligned with the operational constraints of law-enforcement environments. We then propose a hybrid regex-first and NER-repair pipeline with span expansion and obfuscation normalization, substantially reducing leakage across multiple masking failure modes under both standard and adversarial conditions. Integrated into the NL2Cypher workflow, our approach provides stronger front-end privacy guarantees without modifying downstream LLM translation.
These findings show that secure natural language interfaces for sensitive databases require adversarially robust masking at the natural language layer and demonstrate how such robustness can be achieved even under the constraints of real-world law-enforcement settings.
233: The Efficacy of ChatGPT Model GPT-3.5 in Rewriting Bias out of Text while Retaining User Engagement presented by Aarav Daftary (Cambridge Center for International Research)*
This study evaluates the capability of ChatGPT as an evaluator of media bias while also examining to serve as an automated debiasing assistant for news content. We propose a three-phase pipeline – Identify, Rewrite, Evaluate – applied to a stratified dataset drawn from news outlets across the political spectrum, selected using Ad Fontes Media bias ratings, using a novel scoring framework to quantify bias across three dimensions: Framing, Emotional language, and Divisive language (F.E.D. scores). Following an explicit validation of the bias evaluation framework against established bias ratings and inter-model consistency, automated analysis of high-bias sentences (F.E.D. ≥ 6) shows significant reductions in perceived bias: 79%, 69%, and 78% across the three dimensions, respectively. Human evaluation confirms that rewrites are perceived as less biased (original texts selected as more biased 356 times vs. 90 for rewrites). However, a key trade-off emerges: participants preferred the original, more biased content for engagement (90 selections) over ChatGPT rewrites (48). This indicates that naive debiasing capabilities may inadvertently strip away engaging journalistic elements. We conclude that while large language models are effective for bias detection and suggestion, their optimal role is assistive, flagging content for human editors rather than operating autonomously. Future work must address the core challenge of reducing bias without incurring a significant engagement penalty.
205: Bytes to Threats: Byte-Sequence Transformer For Malware Detection presented by Abdelrahman Eljamal (University of Rhode Island)*; Abdeltawab Hendawi (University of Rhode Island)
Existing malware detection approaches heavily rely on handcrafted features derived from static or dynamic analysis of the executable files. These methods require specialized domain expertise, dedicated analysis tools, and time-consuming analysis, which limits their scalability and robustness in the face of the increasing complexity and diversity of modern malware. Consequently, there is a growing interest in approaches that operate directly on raw bytes, which contain rich semantic and structural information. However, capturing long-range dependencies within executable files remains challenging for conventional static and CNN-based methods. This paper presents MalwareFormer, a transformer-based, byte-level malware detection framework that operates directly on raw bytes, without requiring domain knowledge or feature extraction. MalwareFormer achieves 98.13% accuracy and 99.17% AUC, which substantially outperforms CNN-based baselines, demonstrating that transformers provide a powerful and scalable approach for byte-level malware detection.
197: Nibras: A Non-Invasive EEG Brain–Computer Interface for Assistive Communication in Non-Verbal Individuals presented by Ahmed Ibrahim (Prince Sultan University)*; Tareq Ghazi (Prince Sultan University); Anas Houri (Prince Sultan University); ElMoatez Billah Nagoudi (Prince Sultan University)
Many severe neurological issues hinder patients’ ability to express their thoughts through speech or movement. Nibras is a proof of concept that aims to establish a feasibility study for a communication pathway for individuals with severe motor or speech impairments. To enable this project, a dataset was constructed by collecting electroencephalogram (EEG) recordings from a set of healthy volunteers as they performed motor imagery (MI) tasks. These signals were later used to train a set of supervised classifiers that predict different motor imagery actions. To map those MI actions to letters, a predefined MI-to-letter mapping was created. To ensure efficiency and meet realistic deployment constraints, Nibras relies on consumer-grade EEG recordings, and it avoids visual or motion-dependent stimuli to enable users to generate letters with an accuracy of 92.8 percent and an inference time of 0.29ms.
224: Structured Skill Taxonomies vs. Large Language Models: A Comparative Study of Curriculum-Level Skill Inference presented by Yash Pankhania (Northeastern University)*; Nicholas Brown (Northeastern University)
Accurately inferring workforce-relevant skills from academic curricula is critical for aligning higher education with labor market demands, yet existing approaches rely on fundamentally different AI paradigms. Structured skill taxonomies emphasize standardized, ontology-driven representations, while large language models (LLMs) enable flexible, generative skill inference without fixed taxonomic constraints. This paper presents a comparative study of curriculum-level skill inference using these two approaches. We analyze 476 graduate-level course syllabi from the College of Professional Studies at Northeastern University, applying a taxonomy-based skill extraction system derived from Lightcast and LLM-based inference using multiple state-of-the-art models from OpenAI and Anthropic. All methods are evaluated using identical syllabus inputs and a common output format to ensure comparability across models and runs. We assess performance in terms of alignment with structured taxonomies, coverage and novelty of inferred skills, and consistency across repeated inferences. The results reveal consistently low alignment between LLM-generated skills and the structured taxonomy, alongside very high generative novelty across all models and runs. While taxonomy-based extraction provides stable and interpretable skill representations, LLMs infer substantially broader and more context-sensitive skill sets that extend beyond predefined ontologies and exhibit measurable variability across repeated runs. These findings highlight fundamental trade-offs between standardization and generative flexibility and motivate hybrid skill inference frameworks that combine taxonomy-driven validation with LLM-based skill discovery for workforce-aligned curriculum analytics.
226: Evaluating Large Language Models as Symbolic Music Theory Assessors: A Multi-Metric Approach presented by Chauncey Barnes (North Carolina A&T State University); Kiana Katouzian (North Carolina A&T State University); Ahmad Patooghy (North Carolina A&T State University)*
Large language models (LLMs) and Artificial Intelligence (AI) have found their way into education, but their performance in specialized fields such as music theory is not well studied. This paper provides a systematic comparison of four LLM families—DeepSeek-v3.1, GPT-4o, Gemini-2.0-Flash, and LLaMA-4-Maverick—on 200 curriculum-based symbolic music theory tasks including harmony, rhythm, intervals, and meter analysis. We evaluate accuracy as well as pedagogical transparency, response efficiency, and confidence calibration with the help of a novel multi-metric framework. Findings indicate that DeepSeek-v3.1 performs best in terms of accuracy (65\%), whereas Gemini-2.0-Flash is the most efficient with 1.5-second average response time. Nonetheless, the level of overconfidence is high in all models (Expected Calibration Errors of 0.279 to 0.403), and the models use few pedagogically necessary features: Roman numeral notation in 33\% of responses, concrete examples in less than 10%, and meter clarity in approximately 19%. Our assessments show a fundamental conflict between correctness and pedagogical transparency as the models often find the correct answers without being able to reason why. We contribute (1) a structured benchmark reflecting undergraduate and advanced secondary curricula, (2) an automated evaluation pipeline integrating accuracy, explanation quality, latency, and calibration metrics, and (3) actionable insights that educators can use to understand the current capabilities and limitations of LLMs as music theory assessors.
ROOM 145 ABSTRACTS AND FINAL PAPERS
177: AI Companions for E-commerce: Proactive Customer Question Resolution presented by Lokanatha Reddy Gandikota (Verizon)*; Venkata Raghavendra Swamy Gudipati (Remington University); Sudheer Kumar Aluvala (HCL Tech); Sateeshkumar Ponugoti (Publicis Groupe); Satya Tulasi Ram Konda (Verizon)
Traditional e-commerce customer support systems operate reactively, requiring customers to initiate contact. This paper introduces AI Companion, a proactive support framework that anticipates customer questions before they arise. The framework integrates four core components: Data Ingestion Engine, Contextual User Profiler, Predictive Analytics Module, and Natural Language Processing Pipeline. We present complete architectural specifications, mathematical formulations for key metrics, and rigorous experimental validation. Pilot deployments across three retail domains (electronics, apparel, grocery) involving 5,000 customers over six weeks demonstrate significant improvements: 22.7% reduction in support tickets (p<0.001), 11.7% decrease in cart abandonment (p<0.001), and 18.3% improvement in customer satisfaction (p<0.001). We introduce the Customer Experience Score (CX), a composite metric achieving 0.77, up from 0.54 in the baseline. Ablation studies confirm predictive analytics as the critical component. Results establish AI Companion as a practical, reproducible framework for deploying proactive intelligent assistants in real-world e-commerce environments.
179: Lightweight and Maintainable Approach for Table Detection presented by varsha venkataraman (newcastle university)*; lei shi (Newcastle University)
190: RAG-Enhanced Explainable AI for Regulatory-Compliant Credit Risk Assessment presented by Aman Goyal (CMU); Jothsna Praveena Pendyala (Clark University)*; Jyoti Sondager (Independent Researcher); Shaozhi Jiang (AI/ML Independent Researcher); Prerak Manish Shah (Northeastern University)
Credit risk assessment machine learning models are coming under increasing regulatory scrutiny demanding transparent auditable explanations. Although the mathematical interpretability of explainable AI methods such as SHAP and counterfactual reasoning can be assessed, they cannot meet the regulatory compliance and stakeholder understanding standards. The current paper presents a framework of Retrieval-Augmented Explanation Generation that is able to fill the gap between formal model interpretability and adherence to regulatory requirements. Our method combines credit risk models, SHAP-based feature attribution, counterfactual reasoning, and retrieval-augmented
generation and bases the narratives generated by the LLM on bank policies, regulatory documents, model cards, and historical precedents. Assessment of the German Credit Dataset (N=50) indicates that RAEG eliminates citation hallucinations (38.7) to 6.4, regulatory compliance scores (61) to 83, and has 0.891 correlation with the ground-truth SHAP values, and is rated 4.2/5 by compliance officers. Index Terms—Explainable AI, Credit Risk Assessment, Retrieval-Augmented Generation, Regulatory Compliance, Counterfactual Explanations, Financial AI
199: Credit Card Fraud Detection Using Machine Learning presented by Fahd AlHaidari (Salem State university)*; Dania Alkhulaifi (Imam Abdulrahman Bin Faisal University); Wafa Hantom (Imam Abdulrahman Bin Faisal University); Ali Bazarah (Stonehill College); Jamal Alhiyafi (Kettering University)
Credit card fraud occurs when payment card information or related electronic payment methods are misused without authorization to gain financial or material benefits. The widespread adoption of online shopping and e-commerce platforms has introduced new challenges for fraud detection, particularly as users with varying levels of technical awareness engage in digital transactions. Sensitive information may be exposed through unsecured platforms or through identity theft, making reliable fraud-detection mechanisms essential for protecting users and ensuring secure financial transactions. In this paper, a real-world European credit card transaction dataset is used to evaluate supervised machine learning models for fraud detection. Several classifiers, including Random Forest and Support Vector Machine (SVM), are implemented and compared. The experimental results demonstrate that Random Forest based models outperform SVM, achieving an accuracy of 99% with higher precision and recall. In particular, models incorporating class imbalance handling techniques significantly improve detection performance, achieving near-perfect recall and F1-score. This highlights the need for metrics beyond accuracy and confirms the effectiveness of machine learning for fraud detection.
222: SearchAny: Agentic Federated Natural Language Analytics over Heterogeneous Enterprise Data Lakes via Self-Driven Exploration and Reflexive Orchestration presented by Jagadish Krishna Pilla (Independent Researcher)*
223: Algorithmic evidence and fair trial rights in EU Criminal Justice presented by Eleni Papargyri (European University of Cyprus)*; Lamprini Papargyri (Stanford University)
European criminal justice systems are increasingly relying on algorithmic outputs generated or shaped by machine-learning systems, including biometric match scores, risk assessments, link-analysis results, automated triage of seized devices, and other AI-mediated inferences. While such outputs are often introduced as investigative aids, they can acquire decisive evidential weight and function as algorithmic proof. This paper examines the legal status, admissibility, and contestability of algorithmic evidence under EU law, focusing on whether existing safeguards adequately protect fair trial rights. It analyses the interaction between national procedural autonomy and supranational constraints arising from Article 6 ECHR and Articles 47–48 of the EU Charter, alongside the governance regimes of the GDPR, the Law Enforcement Directive, and the EU AI Act. The paper argues that these instruments provide meaningful upstream protections through legality, accountability, impact assessment, logging, transparency, and human oversight obligations, and can support integrity across an extended chain of custody. However, they remain primarily ex ante and regulatory, and do not reliably translate into enforceable procedural entitlements enabling effective defence challenge. Persistent gaps—opacity, proprietary and secrecy barriers, technical complexity, resource asymmetries, and limited access to validation and error information—risk undermining equality of arms and judicial assessment of probative value. The paper therefore proposes a procedural “translation layer” of minimum EU rules on algorithmic evidence, including an algorithmic chain-of-custody record for the inferential pipeline, disclosure standards tailored to effective challenge, structured judicial reasoning on reliability and bias, and guaranteed access to independent expertise and testing. Where these conditions cannot be met and the output is likely to be decisive, opacity should be treated as a fairness defect capable of necessitating exclusion rather than merely affecting weight.
240: Keystroke-Free Reformulation: Design and Evaluation of Micro-Rewrite Widgets for Mobile Search presented by VENKATA Raghavendra Swamy Gudipati (Remington College)*; Lokanatha Reddy Gandikota (Verizon Wireless)
Abstract— Mobile search often falters when users must reformulate a query on small touchscreens. We present micro-rewrite widgets—thumb-scale chips that apply parameterized query edits (e.g., entity disambiguation, scope broaden/narrow, time window) and preview expected result changes before committing via live deltas (estimated hit count and source mix). The goal is keystroke-free, intent-level reformulation that reduces effort while preserving control and transparency. We specify a design mapping from common reformulation intents to safe rewrite templates with latency budgets for responsive previews, and we compare this approach with standard suggestion lists and faceted filtering across fact-finding and comparison tasks on phones in a within-subjects study (N = 36, 24 tasks per participant). Measurements include reformulation effort (keystrokes, edit count), task success and time, over-/under-narrowing errors, perceived workload (NASA-TLX), and usability (UMUX-Lite), analyzed with mixed-effects models and multiple-comparison control. Micro-rewrite widgets reduced keystrokes by 42% (median 14→8), and task time by 18% (median 78→64 s) relative to suggestions, while maintaining or improving task success by +6.3 pp (81.7%→88.0%); they lowered over-narrowing errors by 31% compared to facets when editing targeted query intent rather than result attributes. Facets remained preferable for strictly attribute-bounded filtering (e.g., brand-only constraints). Participants reported a lower workload (NASA-TLX: −8.7, 0–100) and higher perceived control (+0.6 on a 1–7 scale) with live previews. We conclude with design guidelines for integrating keystroke-free reformulation into mobile information retrieval systems, emphasizing transparent previews, safe rewrite templates, and error avoidance. This paper introduces micro-rewrite widgets, a keystroke-free, intent-level query reformulation technique designed for smartphones. Micro-rewrite widgets enable users to apply small, interpretable query edits using thumb-scale touch controls while providing preview-before-commit feedback that transparently communicates the expected impact of each rewrite. We implemented a mobile search prototype supporting five common reformulation intents and evaluated it in a within-subjects study with 36 participants across 24 search tasks. Compared to query suggestions and faceted navigation, micro-rewrite widgets significantly reduced typing effort and task completion time while maintaining or improving task success and reducing over-narrowing errors. These findings demonstrate the potential of keystroke-free reformulation interfaces to improve mobile search efficiency and user control.
Created and maintained by Ballos Associates