AI-Driven Cybercrime Prevention in Online Retailing: A Socio-Technological Perspective

Kamran Razzaq and Mahmood Shah

Cybercrime poses an increasing challenge to the stability of online retailing, where both technological weaknesses and human behaviours contribute towards organisational vulnerabilities. Although traditional studies often treat cybersecurity through a purely technical lens, current research adopts a socio-technical approach to investigate the dynamic interactions among technological systems, human engagement, and organisational infrastructure. Through qualitative, semi-structured interviews with the Chief Information Security Officers, Information Technology Managers, Cybersecurity Practitioners, and Online Retail Executives, the study explored four significant areas of cybercrime prevention: technological systems integration, human awareness and education, governance, and inter-organisational collaboration. The findings indicate that the effectiveness of cybercrime prevention lies in both technical and socio-cultural conditions implemented within the organisation. A Socio-Technological Model of Cybercrime Prevention in Online Retailing is proposed, which shows the interplay among technology, people, and organisational structures. The study advances Information Systems research by reconceptualising cybersecurity as a socio-technical phenomenon rather than a conventional technical construct. The findings further indicate that sustainable cybercrime prevention emerges from the holistic integration of technology, human and organisational subsystems.

Advancing Energy-Efficient AI: A VLSI Exploration of Neuromorphic Architectures with Spiking and Continuous-Domain Neurons

Bhuvanesh Kayarabettu, Bhuvanesh Kayarabettu and Bhuvanesh Kayarabettu

This paper examines VLSI design trade-offs for next-generation AI accelerators, comparing two neuron paradigms: event-driven spiking neurons (SNNs) and continuous-domain classical neurons (ANNs). Hardware implementations are analyzed from individual silicon neuron cores through full system-on-chip configurations, with particular attention to convolutional workloads. At the synaptic operation level, analog spiking circuits show a measurable energy advantage; however, this advantage largely disappears at system scale, where spike-rate encoding overhead and the cost of data transfer and memory access dominate the power budget. Emerging resistive memory technologies can push per-synapse energy into the fJ range for both paradigms, but doing so simply shifts the bottleneck to interconnect and control logic. The central argument is that practical SNN deployment requires reducing spike activity at the application level—through sparse, event-driven sensor inputs-rather
than optimizing the neuron circuit in isolation. Without that change in approach, SNN hardware cannot close the efficiency gap with mature ANN accelerators.

FORTRESS (Framework for Organized RedTeam and Threat Response Evaluation for Security and Safeguards)

Bradley Ammerman

As physical security threats increase in sophistication and frequency, organizations face a critical gap in their physical security, the lack of a structured framework for physical security assessments and tests. FORTRESS was designed to fill this need by providing a categorized model of physical Tactics, Techniques, and Procedures (TTPs), each aligned to industry-standard compliance controls such as NIST, HIPAA, FedRAMP, ISO 27001, and PCI DSS. Designed for both red and blue teams, FORTRESS enables coordinated, repeatable physical security testing by consolidating adversarial behaviors into mapped TTPs. Each TTP is supported by detailed guidance for the execution, validation, and reporting of physical security gaps making assessments more consistent across multiple teams, business units, and future engagements. This paper presents the overall breakdown of FORTRESS, highlighting its application, structure, and flow. I believe this framework fills that critical void in operational and physical security and can be viewed as the foundation for enhancing and maturing physical security risk programs with real threat models as the base.

Predicting Occupant Visual Comfort in Educational Buildings Using Machine Learning (ML)

Jobayer Hossain, G. M. Rahad, Md. Raisul Islam Riyad and Md Shoriful Islam Shuvo

Ensuring visual comfort is essential in the design of sustainable and productive
educational environments, particularly in dense urban areas like Dhaka, Bangladesh, where natural and artificial lighting conditions vary across building zones. This study aims to predict occupant visual comfort within an educational building using machine learning (ML) techniques, providing insights to enhance indoor learning environments. A total of 552 data samples were collected through field surveys and environmental monitoring, focusing on key visual comfort indicators such as number of windows, classroom orientation and occupant distance from windows. The study employed three ML classifiers Random Forest (RF), Decision Tree (DT), and XGBoost (XGB) to model visual comfort. These models were assessed using key performance metrics, including Accuracy, Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). SHapley Additive exPlanations (SHAP) analysis was used to interpret feature importance and increase model transparency. The results showed that the DT achieved
the highest prediction accuracy of 95%, followed by XGBoost with 94% and RF with 90%. SHAP analysis revealed Predicted Mean Vote (PMV), floor number, and classroom orientation as the most influential features affecting visual comfort. The findings underscore the potential of interpretable ML models in diagnosing and improving visual comfort in educational buildings. Future work should expand the dataset to include seasonal variations and additional architectural features, enabling more generalized and adaptive design strategies for occupant-centric indoor environments.

Assessment of Residents Thermal Comfort in Mud Houses Using Machine Learning Classifiers

Md. Estiak Ahmed, Md. Tafhimul Hasan, Al Rafayet Rahman, Mohammad Nyme Uddin, Bishwajit Chandra Mazumdar, Md Sojib Sheikh and Md. Hasan

Mud houses remain a crucial example of sustainable rural living; however, their thermal efficiency is frequently overlooked in relation to climate adaptation and human comfort. This research introduces machine learning models to predict human thermal comfort in naturally ventilated mud houses by utilizing environmental monitoring and field survey information. A dataset containing 515 samples, including 24 parameters of 4 categories such as demographic (age, gender, income, etc.), behavioral (occupant thermal comfort), environmental (temperature, humidity, light intensity, etc.) & physical (opening area,
ceiling material, etc.) was gathered from the Gazipur district in Bangladesh, where mud houses are easily available. Three classification models, Random Forest (RF), Decision Tree (DT), and XGBoost, were predict utilizing confusion matrices and classification metrics. By using hyperparameter optimization and cross-validation on the tree-based models (DT, RF, XGBoost), overfitting was reduced, and generalizability was enhanced, resulting in predicted accuracy that were remarkable. XGBoost achieved the best performance with an accuracy of 81%, with Random Forest next at 77%, and Decision Tree lagging behind at 74%. The SHAP (SHapley Additive exPlanations) study increased explainability beyond prediction accuracy by identifying significant influencing elements such as age, outdoor temperature, ceiling surface temperature, and indoor CO2
concentrations. This is a unique study to apply explainable ML (RF, DT, XGBoost with SHAP) to analyze the thermal comfort in rural mud houses around Bangladesh. Additionally, it provides a scalable predictive model that facilitates context-aware thermal design for climate-resilient houses in tropical, resource-limited areas.

Machine Learning-Driven Thermal Comfort Prediction for Construction Workers in Dynamic Work Environments

Md. Jakaria, G. M. Rahad, Devesis Mondal Dipta, Songkalan Das and H.M. Nazmul Kabir

The construction work environment is very challenging with respect to worker safety and comfort, especially in terms of thermal comfort. Due to the dynamic nature of environmental factors, it has become a challenge to predict and manage thermal comfort among workers who are often exposed to varying environmental conditions. To fill this gap, this paper aims to develop a model that predicts the thermal comfort of construction workers and investigates the main factors that significantly influence it, using classification-based machine learning (ML) models. In this winter period (November 2025-January 2026),
a total of 308 data samples were gathered, including 15 qualitative and quantitative variables. Qualitative data were obtained from surveys on thermal perception, health condition, physical effort, work experience,
etc., along with demographic information (age, gender, weight, BMI, etc.). Smart meters were used to measure quantitative features like temperature, humidity, CO2 concentration, TVOC, and more. Three ML models were constructed for the analysis: Decision Tree (DT), Random Forest (RF), and eXtreme Gradient
Boosting (XGBoost). Their prediction ability was measured through accuracy, precision, recall, and F1-score. Hyperparameter tuning was performed using cross-validation, and SHapley Additive Explanations (SHAP) were used to illustrate the importance of features. All models demonstrated accuracies exceeding 80%. Among them, RF had the highest accuracy with 93%, highlighting its strong ability to predict accurately. SHAP analysis showed that CO2 level, temperature, height and weight, and Lighting level (LUX) were the most influential factors. The application also provides the insight that Such models could serve as a real-time tool for cool-season thermal comfort evaluations and building environments with construction site environments. In the future, this index should be tested on data from more than one season and other pollutants, such as particulate matter (PM2.5) and VOCs. Using IoT-based real-time
monitoring systems can make predictions more accurate and models more reliable.

Feature Ranking for Predicting Thermal Comfort in a School Building Using Machine Learning

G. M. Rahad, Jobayer Hossain, Mohammad Nyme Uddin, Md. Raisul Islam Riyad and Md Shoriful Islam Shuvo

Ensuring thermal comfort is essential in the design of sustainable and productive educational environments, particularly in dense urban areas like Dhaka, Bangladesh. The aim of this study is to predict occupant thermal comfort and rank the most influential features within a school building using machine learning (ML) techniques, providing insights to enhance indoor learning environments. A total of 393 data points were collected through field surveys and using environmental monitoring devices in the summer season. School students were the participants in this study. The study employed three ML classifiers, Random Forest (RF), Decision Tree (DT), and XGBoost. These models were assessed using key performance metrics, including Accuracy, Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). Mean Decrease Impurity (MDI) analysis was employed for feature ranking and to enhance model transparency. All three models performed significantly well, achieving an accuracy of more than 90%. The most dominant features influencing thermal comfort prediction are temperature at the window, humidity, opening orientation, study level, and period. The findings underscore the potential of interpretable ML models in diagnosing and improving thermal comfort in educational buildings. Future work should expand the dataset to include seasonal variations and additional architectural and psychological features, enabling more generalized and adaptive design strategies for occupant-centric indoor environments.

Assessing the Impact of Water Pollution on Community Quality of Life Using Artificial Intelligence (AI)

Md.Hasibul Islam, Md. Hasan, Md Imtiaz Hossain Shohag, Alip Kumar Paul, Azizul Hakim, Mohammad Nyme Uddin, Bishwajit Chandra Mazumdar and Md. Sojib Sheikh

Urban water pollution is becoming a bigger problem, especially in Bangladesh along the Turag River, where people who live by the river must deal with declining environmental conditions that affect their day-to-day existence. This study’s main goals are to evaluate how the quality of life has changed for those who live close to the Turag River and to pinpoint the major causes of water pollution, such as family habits, sociodemographic traits, and river usage patterns, by using machine learning (ML). Structured questionnaires were used to gather information from 400 respondents on a variety of topics, including household cleanliness, waterborne disease incidence, exposure to contaminated water, water usage patterns, and perceptions of river pollution. Three machine learning models, such as Decision Tree (DT), Random Forest (RF), and XGBoost, were applied and evaluated using Accuracy, Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). Feature importances analysis was employed to enhance model transparency and identify key features. According to the findings, every model attained an accuracy of more than 85%. The result shows that water source quality and its proximity to sanitation facilities are the most influential factors affecting the outcome. In order to promote adaptive river management measures and generalize findings, future studies should include seasonal fluctuations and longitudinal monitoring.

Forecasting the U.S. Unemployment Rate using Artificial Intelligence

Reagan Hennen, Nin Tran, Shuoshuo Huo and Vijay Srinivas Tida

Accurate unemployment forecasting is essential for economic planning and timely policy decisions. Traditional forecasting methods often rely on lagging macroeconomic indicators that fail to capture sudden changes in labor market conditions. Real-time digital behavioral data, such as job-related Google search activity, provides immediate signals of job-seeking behavior and labor market sentiment. This study develops a hybrid deep learning model that combines Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to forecast the United States unemployment rate. The model integrates the monthly unemployment rate, GDP growth, CPI growth, and six job-related Google Trends queries. Data are scaled using Min-Max normalization and evaluated with an 80/20 train–test split. Results show that the hybrid LSTM-GRU achieves lower prediction errors than the standalone LSTM and GRU models, with RMSE of 0.1958, MAE of 0.1490, and MSE of 0.0383. Although the models function as black-box neural networks without direct interpretability, the improvements in forecasting accuracy demonstrate the practical value of combining traditional economic indicators with real-time search data for short-term unemployment prediction.

What’s Slowing Us Down? A Passenger Vehicle Perspective on Free Flow Speed Using Machine Learning Models

Md. Iktekar Alam Imran and Soleman Rakib

Free Flow Speed (FFS) is a critical indicator of roadway operational performance,
particularly in heterogeneous traffic environments of developing economies. While prior studies have primarily relied on traditional statistical techniques to analyze FFS, limited attention has been given to prediction-oriented machine learning approaches that also incorporate driver-perceived operational and environmental factors. This study aims to investigate what is slowing us down by predicting FFS from a passenger vehicle perspective using advanced machine learning models. The study analyzed a dataset of 339 passenger-vehicle drivers, covering multiple vehicles with 10 variables. Decision Tree, Random Forest, and XGBoost classifiers were developed using GridSearchCV-based cross-validation to ensure robust and consistent model evaluation. Comparative results indicate that XGBoost outperformed other models, achieving an accuracy of 0.69 and an F1-score of 0.70, followed by Random Forest with an accuracy of 0.63. To enhance interpretability, SHapley Additive exPlanations (SHAP) reveal that perceived humidity, incident occurrence, adverse weather conditions, heterogeneous traffic conditions, and perceived temperature are the most influential factors leading FFS variation from a passenger vehicle perspective. The findings highlight the importance of integrating human-perceived environmental factors alongside conventional traffic variables in FFS assessment. From a policy perspective, the study emphasizes the need for improved incident management, weather-responsive traffic operations, and enhanced infrastructural efficiency. Future research may extend this framework using continuous FFS prediction with broader and multi-city datasets to support resilient and data-driven
transport planning.

A Java Radio Frequency Identification (RFID) Scan Prediction Algorithm Using Forward-looking Evaluation Metrics

Olivia Marshall and Michael Hart

In this study, a radio frequency identification (RFID) scan prediction algorithm is written in the Java programming language. The algorithm is engineered to optimize Internet of Things (IoT) resource usage while accurately predicting future RFID scan timestamps. To evaluate its efficacy, the algorithm’s predictive error rate is compared to the classical machine learning models of Decision Trees, KNN, and Random Forest. It was discovered that the proposed algorithm predicted scans are comparable to the Random Forest model, with minimal differences between KNN and decision trees. In addition to model evaluation, the Java prototype classifies the number of machine learning scans using time series forecasting. Results demonstrate that the proposed algorithm has the potential to improve the prediction of imminent RFID scan times. In future studies, the authors plan to improve the algorithm’s predictive formula, expand to additional data sets, and test other machine learning models, such as small language models.

A Python data classifier algorithm that detects fraudulent data in surreptitious database activity using machine learning

Joshua McNeal and Michael Hart

In this study, researchers developed a data classifier algorithm that detects fraudulent data activity in dirty database records using machine learning. Written in the Python programming language, the algorithm is engineered to accept input and output from varying types of machine learning methods to improve detection rates in surreptitious database activity. The algorithm evaluates fraudulent dirty records unfounded or inaccurately classified by existing detection methods. Its novel contribution includes the ability to feed models positively identifiable clandestine records to increase detection accuracy. It uses an advanced suspicious penalty formula that is capable of adapting to different types of fraudulent records. The authors compared the results of the new algorithm’s ability to detect fraudulent records with decision trees, k-nearest neighbors, and neural networks. During testing, the new algorithm has a comparative rate of succession of 92% compared to the optimal neural network machine learning algorithm. Overall, results demonstrate that the new algorithm accurately identifies more dirty records than certain predecessors but fails to surpass modern machine learning algorithms when deceitful records have easily identifiable characteristics. Despite this progress, fraudulent records still pose significant problems. Future research can expand the proposed algorithm by advancing the suspicious penalty formula, utilizing alternative machine learning algorithms, and using alternative data samples.

Dynamic Instrumentation Analysis of Cryptographic Implementation Vulnerabilities in IoT Mobile Applications

Mansoor Ahmad

Internet of Things (IoT) devices increasingly rely on companion mobile applications for device control, user management, and cloud connectivity. These applications often handle sensitive user data, yet their security practices remain understudied, particularly regarding application-layer encryption implementations. This research investigates whether cryptographic protections in intimate IoT mobile applications provide meaningful security or merely create an illusion of protection. Two Android companion applications for Chinese-manufactured IoT devices were analyzed using a combination of dynamic instrumentation via Frida and static analysis through APK decompilation. The analysis revealed significant cryptographic implementation failures in both applications, including hardcoded AES encryption keys and, in one case, the use of identical values for both the encryption key and initialization vector, a configuration that undermines the security properties of cipher block chaining mode. These weaknesses enabled complete decryption of application traffic, exposing a user’s personal information and device control commands. To operationalize these findings, a Burp Suite extension called CrypticBurp was developed to automate decryption, plaintext editing, and re-encryption of intercepted traffic, restoring full security testing capabilities against applications that would otherwise resist analysis. This research contributes a reproducible methodology for extracting cryptographic material through iterative OpenSSL function hooking, documents a pattern of security theater where multiple protection layers (certificate pinning and application-layer encryption) collapse under systematic analysis, and provides practitioners with reusable tooling applicable to any application employing similar encryption schemes. The findings underscore the need for IoT developers to adopt proper key management practices, including per-session key derivation and secure storage mechanisms, rather than relying on obfuscation through embedded secrets.

Multi-task Temporal Convolutional Network BiLSTM Attention with Bayesian Optimization for air pollutants prediction in South Africa

Israel Edem Agbehadji and Christiana Ibidun

The impact of air pollution on human beings is growing at an alarming rate and despite several technological attempts to resolve its effect on human health, the particulate matter (PM2.5 and PM10) continues to threaten human health due to overpopulation, heavy traffic and rapid industrialization. Conducting a temporal analysis of air pollutants and meteorological variable, help to address the challenging and accurate future time predictions due to the non-stationary and non-linear characteristics inherent in air pollutant information. While several deep learning methods have been utilized to address the aforementioned challenge, forecast accuracy is an open research topic leading to model hybridization. By utilizing data from the South African context, this research provides a multi-task temporal convolutional network Bi-directional long short-term memory (BiLSTM)-based Attention with Bayesian Optimization model for the prediction of both PM2.5 and PM10. The proposed model was evaluated with R2-score and root mean square error (RMSE), where R2-scores of 0.51 and 0.45 are recorded for both PM2.5 and PM10, as compared to the comparative models, which were lower.

Beyond Points: Event Associations with Overall Placing in NCAA Division I Conference Decathlons Using Ranks and Standardized Marks

Davis Reinhard and Rajeev Bukralia

This study evaluates which decathlon events most consistently align with final overall placing without using scoring-table points as inputs. This approach avoids relying on scoring-table point conversions, which may influence perceived event “importance” and can obscure competition-relevant alignment between event outcomes and final placing. Publicly available NCAA Division I conference championship results from the Big Ten, SEC, ACC and Big 12 were compiled (60 meets), restricting the analysis to top eight finishers who completed all ten events. Three questions guided the analysis: (RQ1) which events align most strongly with final placing, (RQ2) whether event-placing alignment patterns are consistent when using event placings versus within-meet standardized marks (z-scores), and (RQ3) how alignment varies across conferences and meets. Event-outcome alignment was assessed using two methods: Spearman rank correlations between each event placing and final overall placing, and Pearson correlations between
within-meet z-scores and final overall placing. For pooled event-level correlations, 95% confidence intervals were estimated using a clustered bootstrap by meet. Across pooled results, the discus throw and shot put showed the strongest alignment with final placing under both approaches, while the 1500m and high jump showed weaker alignment than others. These findings provide a practical, competition-relevant perspective on event emphasis that does not rely on scoring-table point transformations.

Comparing YOLOv11 and YOLOv12 for Bounding-Box Detection of Apple Ripeness in Digital Images

Busrat Jahan and Dr. Rajeev Bukralia

Accurate object detection is essential for maintaining quality standards in post-harvest sorting and automated evaluation within modern agriculture. Currently, YOLO (You Only Look Once) is widely recognized for its balance of speed, accuracy, and simplicity, performing detection in a single forward pass for real-time applications. A total of 4,000 images were collected from Roboflow and, after random augmentation, expanded to 9,594 images and labeled as Fresh or Rotten. This study investigates how architectural modifications, particularly in backbone, neck, and attention modules, influence both accuracy and computational efficiency. A quantitative experimental approach was used. The existing YOLOv11 and YOLOv12 architectures remained highly efficient, each with lightweight parameter counts of approximately 2.5M and low computational costs of 6.4–6.5 GFLOPs. The existing YOLOv12 model achieved the highest accuracy of 99% and an mAP50 of 0.994 mAP50-95 of 0.908. Notably, the customized YOLOv11 model showed exceptional training efficiency, reducing training time to 0.38 hours, a substantial improvement over the existing YOLOv11’s 2.35 hours while maintaining high accuracy (97%) and only slightly increasing computational cost to 6.8 GFLOPs. In contrast, the customized YOLOv12 became significantly larger, reaching 14.5M parameters and 110.7 GFLOPs, but achieved lower accuracy (95%) for this dataset. Overall, the findings highlight that the existing YOLOv12 model is the best performer in terms of speed, accuracy, and efficiency. Meanwhile, the customized YOLOv11 outperformed the customized YOLOv12, demonstrating a well-balanced, lightweight, and precise architecture.

A Comparative Analysis of SLMs for Machine Translation

Wyatt Clausen and Rajeev Bukralia

While large language models (LLMs) have advanced machine translation, their high computational costs and environmental impact have raised concerns. Small language models (SLMs) can address these concerns and be further adapted to specific tasks through parameter-efficient fine-tuning (PEFT) methods, including Quantized Low-Rank Adaptation (QLoRA). This study compares the Spanish-to-English translation quality of two 7-billion-parameter SLMs, Mistral and Llama, using BLEU, chrF, and COMET as evaluation metrics, and investigates whether QLoRA fine-tuning on the Europarl EN-ES dataset improves their performance. Results indicate that both models exhibited low lexical accuracy (BLEU, chrF) but high semantic accuracy (COMET). Notably, QLoRA fine-tuning reduced semantic and lexical accuracy across both models, and the fine-tuned Mistral variant produced significant hallucinations.

Characterizing Seasonal Deal Seekers: A Machine Learning Approach to Behavioral Segmentation

Kundai Chirimumimba

The rise of e-commerce has enabled consumers to strategically compare prices and time their purchases, resulting in two behaviorally distinct customer types. Loyalists are customers who consistently engage with brands, while deal seekers are those who wait for promotional events. Current segmentation models often rely on self-reported data prone to response bias, leaving retailers with a gap in analyzing raw transactional records. This study addresses this by proposing a two-stage machine learning pipeline to identify consumer types using the UCI Online Retail II dataset. Following RFM (Recency, Frequency, Monetary) feature engineering, K-means clustering is applied to uncover natural behavioral segments, while Random Forest analysis identifies the most discriminative features. The results, visualized through a Python Plotly Dash dashboard, reveal three distinct segments, namely, High-Value Loyalists, Occasional Browsers, and Seasonal Deal Seekers. This research provides a scalable and validated approach for retailers to characterize customer behavior using standard invoice-level data, offering both theoretical insights into digital consumerism and practical applications for targeted marketing efforts and implications.

Content-Based Vocal Repertoire Ranking Framework Using Duration-Weighted Pitch Distributions

Madeline Johnson, Flint Million and Rajeev Bukralia

Choosing vocal repertoire that fits a singer’s voice is important for vocal health, yet to the author’s knowledge, current tools based on tessituragrams help assess a piece only after it has been selected—they do not recommend new pieces likely to suit the singer. A key factor in vocal fit is tessitura: the pitches on which the voice spends the most time in a given piece, weighted by how long each pitch is sustained. This paper presents a proof-of-concept content-based ranking framework that uses duration-weighted tessituragrams—pitch-by-duration profiles extracted from machine-readable musical scores—to rank songs by how well they match a singer’s stated preferences. The singer specifies a comfortable vocal range, favorite pitches, and pitches to avoid; the system filters songs by range, constructs an ideal pitch profile from these preferences, and scores each candidate using a similarity measure (cosine similarity) that captures how closely the song’s pitch profile aligns with the ideal, minus a penalty for time spent on pitches the singer wants to avoid. We evaluate offline with synthetic self-retrieval: a profile is built from a target vocal line (the range of the vocal line, four pitches with the most total duration as favorites, and the two with the least as avoids), and we ask how often that same line ranks highly among other filtered lines. We evaluate two protocols on the OpenScore Lieder corpus: a compact library of 101 vocal lines, one per composition, and an expanded flattened library of 1,655 vocal lines drawn from 1,419 compositions (some works contribute more than one line). We compare the full model to two different baselines: null (range filter and random ranking) and cosine-only (without an avoid penalty). The full model hit rate at 1 (HR@1) is 76% vs 6% null (compact design) and 55% vs 2% (expanded); HR@5 is 86% vs 7% null (expanded). Rankings are stable under small preference edits (mean Kendall’s τ ≈ 0.84-0.85). The avoid penalty does not clearly improve HR@1 or mean reciprocal rank over cosine-only on these offline draws. Results support internal consistency under these offline protocols. Claims are limited to synthetic profiles and symbolic scores; human evaluation and other genres are future work.

How Transportable Are Tabular Heart-Disease Classifiers? A Multi-Site Analysis with Shift Diagnostics and Calibration

Yisihaq Yemiru

Machine learning models for heart disease prediction are typically validated within a single cohort, limiting understanding of cross-site transportability. This study assesses whether such models remain dependable when applied to new populations – and, when they fail, what drives failure. Transportability was tested across five datasets: four UCI Heart Disease datasets (Cleveland, Hungarian, Swiss, and VA Long Beach) and one Kaggle Cardiovascular Disease dataset, using logistic regression, random forest, XGBoost, and LightGBM. Under internal 80/20 validation (RQ1), AUC-ROC values range from 0.65 to 0.97. In external validation (RQ2), it was found that transportability between families with a common feature set (CFS) resulted in a loss of up to 0.39 in AUC scores, whereas pairwise transportability within the UCI Heart Disease datasets resulted in AUC scores being reduced to 0.26. Shift diagnostics (RQ3) identify maximum heart rate (PSI up to 11.50) and fasting blood sugar (PSI up to 9.20) as the most unstable
characteristics. Post-hoc recalibration (RQ4) reduces mean ECE from 0.25 to 0.14 (Wilcoxon test, p < 0.001). For severely mis-calibrated pairs, Platt scaling lowers ECE from 0.49 to 0.02, representing a 95.00% reduction. However, calibration may degrade in small test sets. Using a three variable CFS (RQ5) incurs a mean AUC penalty of 0.17, while the magnitude of cross-site degradation remains unchanged.

Auralis: A frequency- and Emotion-Aware Hybrid Music Recommendation Framework

Samuel Nono

Music recommendation systems often prioritize large-scale behavioral data while overlooking the acoustic properties that shape how emotionally experienced music is. This creates a gap between personalization and interpretability, particularly in cold-start scenarios where user data is limited. This study presents Auralis, a frequency- and emotion-aware hybrid recommendation framework that integrates acoustic feature modeling with adaptive user preference learning. The system extracts Mel-Frequency Cepstral Coefficients (MFCC) to represent timbral and spectral characteristics of audio signals, aggregates these features into fixed-length vectors, and maps them to interpretable valence-arousal aligned emotion scores. Cosine similarity is then used to retrieve emotionally aligned tracks from an indexed collection. A preliminary implementation includes a modular backend, a song indexing pipeline, and an interactive interface designed to support future conversational interaction. Auralis aims to bridge signal processing, emotional modeling, and personalization in a transparent and extensible recommendation architecture.

Quantum-enhanced Adaptive Decision Support Systems for Supply Chain Risk Management

Prakrati Maheshwari, Vishal Shah, Ayush Jaiswal and Yogesh Pandya

Global supply chains are exposed to risk of economic instability, natural disasters, and other operational problems, where dynamic approaches are necessary to manage risk. Traditional DSS cannot respond to adaptive demands in complex supply chains. We point out the development of a quantum-enhanced adaptive decision support system for improving supply chain resilience through integrating quantum computing into risk management. It discusses the architecture, key algorithms, and performance metrics of Quantum-Enhanced Adaptive Decision Support System (QADSS), together with the comparative analysis against classical DSS models. Simulation results indicate QADSS could bring about important reductions in decision latency and improvements in adaptive capacity.

Predicting Residential Natural Gas Consumption at Municipal Level using Machine Learning

Marten Weijer

The Netherlands is phasing out residential natural gas – officially by 2050, though the direction has been clear since 2018. More than 90% of Dutch homes still run on gas for heating, and the gap between that reality and the policy target is large enough that planners need detailed, regional-level forecasts just to know where to start. Most energy forecasting research is done at the national level or focuses on individual buildings. Municipal-level gas prediction using machine learning is mostly unexplored. This study addresses that gap. The analysis uses CBS open data covering 300+ Dutch municipalities from 2015 to 2021, roughly 3,175 observations combining residential energy use, housing stock characteristics, household income, demographics and weather data. Four models are compared: Ridge regression, Lasso regression, Random Forest, and XGBoost. The models are chosen on purpose: linear models are easier to interpret, ensemble methods often predict better. Testing both on the same data shows whether the added complexity earns its place. Performance is evaluated using Root Mean Squared Error (RMSE), R-squared (R2), and Mean Absolute Error (MAE). Random Forest achieves the highest accuracy (R2=0.777, RMSE=114 m3) on the 2020-2021 test period. Feature importance analysis revealed housing characteristics dominate predictions, with dwelling surface area and housing age being most influential.

A Tutoring AI Chatbot for Teaching CS 1

Syed Kamran Khatai and Cheng Thao

Many universities are teaching heavy where professors have heavy teaching load. In addition, accessing to professors during office hours or tutoring may not be practical for non-traditional students who work during the day and have long commute. In this research, we propose an AI tutor to bridge the gap of students getting help. We implemented a prototype to help students in an introductory computer science in a minority serving university. The AI tutor chatbot is available 24/7 to help students with their assignments and course materials. The AI tutor uses Gemni API, and backend database for course materials and assignments.

Implementing Ekub on Blockchain

Girma Degfe and Cheng Thao

This research investigates the feasibility of integrating the traditional Ethiopian Ekub system, a Rotating Savings and Credit Association (ROSCA) with blockchain technology to enhance financial inclusion for Micro, Small, and Medium Enterprises (MSMEs). While Ekub serves as a critical source of finance for 50% of MSMEs in Ethiopia providing as a main source of access to credit, issues of transparency, scalability, and susceptibility to fraud hinders its potential for being a reliable and scalable solution. This study proposes a decentralized framework leveraging Ethereum smart contracts to automate contributions and payouts. Ekub heavily depends on social capital for creating trust amongst its participants, which, while it works, limits its scalability. The research introduces a novel reputation token system to bridge traditional social capital with digital trust mechanisms. The proposed solution aims to reduce fraud, enable global accessibility, and provide a scalable, secure financial tool for underserved communities.

Evaluating the Robustness of Autonomous Traffic Sign Recognition Systems under Heavy Snow and Fog Conditions: Identifying Safety Thresholds

Sadman Sakib Abdullah

Deep learning models perform exceedingly well for traffic sign recognition (TSR) under clear weather conditions. However, their performance remains unquantified under adverse weather conditions. This study evaluates the performance of YOLOv8 and ResNet-50 on the German Traffic Sign Recognition Benchmark (GTSRB) dataset under four synthetically generated snow and fog levels using the Albumentations library. The performance of each model will be measured using top-1 accuracy and per-class F1 metrics. This study aims to identify the specific fog and snow density at which each model crosses the defined safety threshold – a drop of 15% or more below clean-weather baseline accuracy, producing a controlled baseline for weather robustness evaluation in autonomous TSR Research.

Same Skills, Different Words: A Comparative Quasi-Experimental Analysis of Lexical Variation Effects in Automated Resume Screening

Mashfika Jahan

Automated resume screening (ATS) increasingly relies on natural language Processing (NLP) and similarity scoring to rank candidates based on the job description. Prior work done by other researchers shows that wording choices such as communal vs. agentic language, keyword phrasing, and patterns related to demographics can influence both algorithmic and human hiring decisions. This study investigates whether small lexical resume variation (e.g., “python” vs “Python programming”) significantly changes the automated screening outcomes. We use a controlled quasi-experimental methodology to keep the substance of the resumes the same and change the wording in a systematic way across four types of variations: phrasing (substituting synonyms), abbreviation (expanding or shortening), word order (reordering tokens), and placement (restructuring sections). We use TF-IDF cosine similarity and BM25 to assess resume variants for four roles: AI Researcher, Data Scientist, Cybersecurity Analyst, and Software Engineer, using a synthetic public Kaggle dataset of 1,000 resumes. We examine whether lexical variation alone can affect screening scores, rankings, and shortlist inclusion.