Authors - Sarthak, Utkarsh Kumar Singh, Ankur Yadav, Aarushi Sharma, Samarth Saxena, Vaishnavi Kumari Singh, Anisha Kumari Abstract - Cervical cancer prediction using machine learning is often limited by class imbalance, dataset variability, and insufficient control of false positive rates. While many existing models report high accuracy, they frequently fail to maintain a clinically appropriate balance between sensitivity and specificity, particularly across datasets with different sizes and feature structures [1]. Models trained on large clinical risk-factor datasets may not generalize well to smaller behavioral datasets, and recall-oriented optimization can significantly increase false positives. This study proposes a false positive–optimized ensemble framework combining behavioral and clinical risk factors and analyzes its performance across two heterogeneous datasets. Threshold tuning and ensemble techniques, including soft voting and stacking, are employed to increase minority-class detection while retaining specificity. Results indicate that independent classifiers show dataset-dependent instability, with trade-offs between recall and false positive control. However, ensemble methods provide more consistent accuracy, precision, recall, and F1-score across datasets. The findings demonstrate that threshold optimization combined with ensemble learning improves cross-dataset robustness and supports more clinically reliable cervical cancer prediction.
Authors - Sowmini Devi Veeramachaneni, Yaswanth Gavini, Arun K Pujari Abstract - Combining Particle Swarm Optimization (PSO) with gradientbased local search enhances efficiency in solving complex optimization problems. Existing hybrids often use fixed switching rules, causing premature convergence orwastedcomputation.We present an adaptive PSO–gradient descent method where stagnation detection triggers local refinement only when needed. Adam is employed for local search without extra parameters. Tests on seven benchmark functions show the approach achieves strong or competitive results on challenging cases while ensuring robust convergence on simpler ones.
Authors - Amna Ali, Rida Hijab Basit Abstract - With the advent of agentic Artificial Intelligence, systems have demonstrated significant ability to understand data and respond to changing business environments without human assistance. Agentic AI is largely being used in supply chain management (SCM) systems for automating the supply chain tasks - demand forecasting and planning, logistics and transportation optimization, supplier management and risk reduction, and warehouse management. Use of agentic AI in SCM represents a drastic shift from traditional rule-based systems to automated goal-driven systems that operate without human intervention. Such systems are supported by Natural Language Processing and deep learning models which have made the supply chain processes much easier, efficient and less prone to error. The organizations that have incorporated agentic AI in their business processes have reported operational efficiency and cost effectiveness. However, such advancements in technology have raised concerns related to privacy ethics and data security. In this paper, we have conducted the systematic review of the existing research on the usage of Agentic AI in Supply Chain Management. The paper discusses characteristics of agents in SCM, different types of architectures and analyses the limitations and challenges related to the usage of AI agents in supply chain management.
Authors - Bikram Bikash Das, Chukhu Chunka, Pantha Kanti Nath, Nippu Kumar Abstract - Credit card transaction analysis is challenged by severe class imbalance with evolving spending behavior and large-scale financial data. Many existing fraud detection approaches rely on supervised learning and assume stable fraud labels, limiting robustness under changing fraud prevalence. This study presents a large-scale, multi-year credit card trans action dataset stored in partitioned Parquet format and conducts a systematic comparison of classical machine learning, supervised deep learning, and unsupervised deep learning models for customer spend ing behavior analysis. An exploratory behavioral analysis characterizes spending heterogeneity, temporal regularities, and channel and category variations. Supervised sequence models based on LSTM and CNN ar chitectures are evaluated alongside unsupervised sequence autoencoders and hybrid detection pipelines across fraud rates ranging from 2-12%. To ensure fair evaluation under extreme imbalance, models are assessed using ranking-based metrics under fixed alert budgets, including pre cision–recall area under the curve and recall-at-K. A hybrid of Autoen coder and LSTM architectures achieves the highest performance for large systems. An integrated XAI module is introduced to derive important features providing interpretable insights.
Authors - Dao Khanh Duy, Nguyen Hoang Hieu, Karn Nasritha, Khanista Namee Abstract - This research examines the effectiveness of four state-of-the art transformer-based models (LaBSE, mBERT, XLM-RoBERTa, and mT5) for cross-lingual sentiment analysis of railway passenger feedback. We focus on transferring knowledge from high-resource languages (En glish, French, Vietnamese, and Korean) to Thai, a low-resource language in this domain. To address data imbalance and scarcity, the study inves tigates transfer learning strategies ranging from zero-shot to "ultra-shot" (using only 60 labeled samples) and high-shot paradigms. Experimental results demonstrate that while generative models like mT5 perform well in zero-shot settings, the LaBSE model achieves a superior accuracy of 94.65% under high-shot fine-tuning. Notably, our proposed ultra-shot strategy enables LaBSE to reach 90.42% accuracy with minimal data, effectively bridging the performance gap without extensive annotation. These findings suggest a strategic approach for AI systems in railway op erations: rather than investing in large-scale datasets or computationally heavy models, operators can implement the ultra-shot strategy by fine tuning robust sentence-embedding models like LaBSE with a small set of gold-standard data to achieve optimal performance and cost-efficiency.
Authors - Nazar Melnyk, Oleksandr Korochkin Abstract - Reliable prediction of rare critical events is a key enabler for modern risk management, civil protection, and decision support sys tems, yet it remains challenging due to extreme class imbalance and strict requirements on false alarm rates. We present an ensemble learn ing framework that combines a deep feed-forward neural network with a Random Forest classifier, complemented by temporal feature engineering and precision-oriented optimization. The approach addresses three ob jectives: extracting informative temporal and regional patterns from raw event logs, learning calibrated probabilistic scores under severe imbalance using focal loss, and tuning per-region decision thresholds to achieve high precision while preserving acceptable recall. As a case study we apply the framework to air alert prediction over 25 administrative regions across 38 months, totalling 774,125 hourly observations. The system attains 96.13% accuracy, 75.1% precision, and 77.9% recall, demonstrating that high-precision early warning is feasible in strongly imbalanced settings. The framework is applicable to a wide range of safety-critical rare event prediction tasks.
Authors - Lavinia Chiara Tagliabue, Silvia Meschini, Viviana Vaccaro, Hira Ovais, Silvana Dalmazzone, Gianluca Torta, Ferruccio Damiani, Stefano Rinaldi Abstract - Named Entity Recognition (NER) is an essential task for sequence labelling and information extraction that plays a fundamental role in subsequent Natural Language Processing (NLP) applications, such as information retrieval, question answering, knowledge graph development, and machine translation. Although significant advancements have been made in NER for high resource languages, achieving effective entity recognition in Indian languages continues to be an unresolved research challenge because of linguistic diversity, complex morphology, typological differences, flexible word order, script differences, and prevalent codemixing. The scarce presence of annotated datasets and the lack of standardized evaluation metrics further limit supervised and transfer learning methods in these low resource environments. This document introduces a multilingual NER framework rooted in Sentence embeddings derived from Large Language Models (LLMs) and inference guided by prompts. The suggested method employs contextual; language independent embeddings obtained from pretrained multilingual LLMs to encode semantic representations of Indian and foreign languages within a common embedding space. Rather than using traditional token level classification, entity recognition and classification are achieved via structured prompting, allowing for zero-shot and few-shot generalization without the need for task specific finetuning. The system guarantees that entity identification and retrieval take place in the same language as the input text, maintaining linguistic accuracy and reducing error propagation caused by translation. To tackle domain variability and informal writing, constraints/guardrails for prompts and simple rule-based normalization are utilized to manage orthographic differences, script inconsistencies, and codemixed phrases often found in user generated content and social media. Experimental assessment across various Indian languages shows reliable enhancements in precision, recall, and F1score compared to traditional neural and transformer-based benchmarks, especially in low resource conditions. The findings suggest that embeddings powered by LLMs along with prompt-based reasoning provide a scalable and data efficient option for multilingual NER. This project advances the development of resilient, inclusive, and language adaptive systems for extracting information in linguistically varied settings.
Authors - Murat Aydın Abstract - Combining Particle Swarm Optimization (PSO) with gradientbased local search enhances efficiency in solving complex optimization problems. Existing hybrids often use fixed switching rules, causing premature convergence orwastedcomputation.We present an adaptive PSO–gradient descent method where stagnation detection triggers local refinement only when needed. Adam is employed for local search without extra parameters. Tests on seven benchmark functions show the approach achieves strong or competitive results on challenging cases while ensuring robust convergence on simpler ones.
Authors - Hiep. L. Thi Abstract - Flexible Job Shop Scheduling Problems (FJSP) involve large discrete decision spaces and strict feasibility constraints, making them challenging for deep reinforcement learning methods. In this work, we study how state represen tation and feature extraction architecture influence the performance of action masked Proximal Policy Optimization (PPO) in flexible scheduling. The scheduling task is formulated as a sequential assignment of operations to machines with a fixed discrete action space, where infeasible actions are removed using a feasibility mask. The environment state is represented using three heter ogeneous feature blocks describing resource availability, operation readiness, and time-related attributes of assignment alternatives. We compare a baseline single-branch encoder with a multi-branch feature extraction architecture that processes these blocks separately before aggregation. Experiments were conducted on the Brandimarte MK benchmark suite (MK01 MK10). Under identical training conditions, the multi-branch representation achieved lower makespan on 9 out of 10 instances, with relative improvements ranging from 2.4% to 27.8% compared to the single-branch baseline. The largest reductions were observed on MK06 (−27.8%) and MK10 (−25.2%), while per formance remained comparable on MK08. Training results indicate improved stability and more consistent convergence for structured representations. These results demonstrate that structured state design and feature extraction ar chitecture are critical factors in action-masked reinforcement learning for flexible job shop scheduling.
Authors - Karn Na Sritha, Khang Tran Chi Nguyen, Dao Khanh Duy, Khanista Namee Abstract - Multimodal affective computing system (MACS) aims to improve the affect prediction performance by fusing the complementary cues in visual and audio channels. While late fusion approaches are modular and can be flexibly deployed, they often rely on static modality weights which pre-assumes fixed reliability among modalities. In practical situation, visual stream can be corrupted by occlusion, variation of illumination and motion artifact while audio stream could be interfered by noise and reverberation or channel mismatch. Moreover, domain shifts between different datasets further contribute to the problem of in consistent calibration across modalities, which results in inaccurate fused predic tion. In this paper, a reliability-aware late fusion model is proposed to enhance ro bustness for multimodal emotion recognition. Based on the independently trained branches of FER and SER, we conduct an analytical process for theoretical var iance-covariance stability analysis of linear late fusion with respect to a modality imbalance condition. We further investigate entropy-driven reliability estima tion and calibration-aware weighting schemes. Experiment results from original test report are incorporated into the theoretical framework, it makes evidence that one modality’s dominance is more related to entropy stable and calibration char acteristics than raw unimodal accuracy. Our results also indicate that reliability aware weighting increases robustness under simulated degradation and missing modalities, without the need for retraining unimodal models.
