University of Sindh Journal of Information and Communication Technology

Evaluating the Robustness of RNN Models on Diverse Sequential Datasets

2025-06-16T10:43:39+00:00

Recurrent Neural Networks are deep neutral networks that handle sequence data due to their ability to capture and reserve temporal dependencies in the data. This characteristic makes them ideal for applications such as natural language processing, speech recognition, and time series prediction. In this research paper RNN methods are applied and analyzed using the UCI Student Performance dataset and the UCI Human Activity Recognition (HAR) dataset. It studies the impact of hyperparameter optimization and data set balancing on critical performance system of measurement such as model accuracy, memory usage, and time taken for prediction. By altering critical features, the number of RNNs, the dropout rate, and the batch size, the effect of these parameters on model performance is analyzed. The results show that varying the number of RNN units and batch size optimization improves model performance, as well as the effect of computational efficiency. The RNN test produces 99.8% accuracy in the UCI Student Performance data set and 95.11% in the HAR data set. In adding, balanced datasets help to improve generalization by reducing bias and overfitting. A train-test split of 70%-30% is testified to provide the best overall precision compared to computational resource costs. This research tells us how hyperparameter optimization and data handling modify and improve the performance of RNN models.

Analyzing User Information Seeking Strategies in Interactive Information Retrieval Systems A Mixed-Methods Evaluation Framework

2025-09-17T20:21:47+00:00

Abstract

Understanding how users adapt their ISS across task types is critical for evaluating and designing effective IIR systems. While traditional evaluations focus on performance metrics like relevance or precision, they often fail to capture user behavior's cognitive and strategic dimensions. This study proposes a mixed-methods evaluation framework to analyze ISS across factual, exploratory, and comparative tasks. A user study involving 30 participants used a custom academic search interface. Data were collected through interaction logs, satisfaction surveys, and post-task reflections.

The findings show that strategy use is highly task-dependent. Factual tasks encouraged direct lookup behaviors, whereas exploratory and comparative tasks involved iterative query refinement, multi-document synthesis, and opportunistic exploration. Satisfaction correlated positively with strategy diversity; expert users demonstrated greater adaptability in complex tasks. Five dominant strategy types were identified and linked to task complexity and user expertise using quantitative log data and qualitative user narratives. This study advances IIR evaluation by offering a replicable, strategy-sensitive framework beyond traditional click-based metrics. The framework provides actionable insights for designing adaptive, user-centered search systems that support a range of strategic behaviors across varying information needs.

Keywords: Interactive Information Retrieval, Information Seeking Strategies, Mixed-Methods Evaluation, User Behavior, Task-Based Search

Optimizing Tracking System: The Impact Of Facial Recognition On Security And Efficiency

2025-09-17T20:40:39+00:00

The demand for effective and safe attendance monitoring systems has increased in the modern world, affecting several industries such as public institutions, corporate settings, and education. Traditional methods, such as filling out attendance sections or card-based systems, are error-prone, time-consuming, and vulnerable to fraudulent activity. To solve these problems, this paper introduces an improved convolutional neural network (CNN)-based facial recognition system that increases the efficiency and accuracy of attendance tracking by adjusting the learning rate through adaptive learning and optimizing the features. The proposed system combines off-the-shelf deep architectures with an optimistic lightweight to reduce inference time and enhance their robustness to variable lighting and mask conditions. The optimization explicitly targets both accuracy and speed, improving inference efficiency by tuning learning rates and pruning redundant convolutional layers. The model is optimized experimentally, with the use of the AT&T, LFW and CASIA-WebFace datasets, showing that the optimized model is better than the standard CNN baselines. Such a framework can prevent identity fraud and so-called buddy-punching, as well as enhance data security by encrypting and storing it in privacy-protected storage. But then, the confidentiality problem that arises while using recognized face technology is also very important to focus on. The first step to maintaining the confidentiality of biometric information is to implement clear data protection protocols and to put in place reliable security measures. Lastly, facility biometrics devices could be used to retain attendance reports in place of customary methods that are exhausted by various challenges. The future outbreak of such systems may perform as a necessary tool for presence management process reformation at various industries. The combination of effectiveness, accuracy, and security will lead to effectiveness and operational excellence. The originality of this paper is the hybrid optimization of CNN parameters and introduction of a privacy-conscious design in the modern absence sense.

Deep Learning for Sugarcane Disease Detection: A Field-Validated EfficientNet-B4 Approach

2025-09-18T05:09:41+00:00

Abstract

Sugarcane (Saccharum officinarum) serves as a critical economic mainstay in Pakistan, where it ranks among the top agricultural commodities. Devastating fungal and bacterial infections – particularly red rot (Colletotrichum falcatum), smut (Sporisorium scitamineum), and leaf scald (Xanthomonas albilineans) – inflict annual yield losses exceeding 20-50%, translating to crippling financial blows surpassing $350 million. Conventional disease identification techniques remain impractical for most farmers: they demand specialized expertise, involve time-consuming laboratory processes (typically 3-5 days), and prove cost-prohibitive across rural regions where sugarcane cultivation predominates.

This research pioneers an intelligent vision-based diagnostic system leveraging deep convolutional neural networks (CNNs) to automate sugarcane disease recognition directly from field imagery. To overcome dataset limitations endemic to agricultural AI applications, we compiled a comprehensive repository of 15,000 high-resolution field images capturing healthy and diseased specimens across diverse growth stages and environmental conditions. Strategic data augmentation through geometric transformations (multi-axis rotation, flipping, scaling) and photometric adjustments (variable brightness/contrast) expanded this corpus to 26,500 training samples. Three state-of-the-art architectures were rigorously evaluated: custom 8-layer CNN, ResNet50, and EfficientNet-B4, with the latter fine-tuned using transfer learning principles initialized on ImageNet weights.

The optimized EfficientNet-B4 model demonstrated exceptional proficiency, achieving 94% classification accuracy, 92% recall, and a 91% F1-score during real-world field validation – significantly outperforming traditional methods. Deployment occurs through an intuitive cross-platform application (mobile/web) enabling farmers to capture leaf/stem images and receive instant diagnoses (<2 seconds) with management recommendations, even offline. This work delivers three pivotal contributions: 1) A novel disease detection framework validated under operational farm conditions, 2) Public release of the largest curated sugarcane pathology image dataset to date, and 3) A farmer-centric tool advancing UN Sustainable Development Goals (SDGs) – specifically SDG 2 (Zero Hunger) through yield protection and SDG 9 (Industry, Innovation) by democratizing cutting-edge AI for precision agriculture.

Keywords: Deep Learning, Sugarcane Diseases, CNN, ResNet, EfficientNet, Image Classification, Precision Agriculture, SDGs.

Smart Agriculture: AI-based System for Classification of Fresh, Rotten, and Formalin Fruits

2025-12-22T07:47:21+00:00

Ensuring the freshness and safety of fruits and vegetables is critical for protecting public health, minimizing food waste, and enhancing supply chain efficiency. Traditionally, the assessment of produce quality relies on manual visual inspection, which is time-consuming, labor-intensive, and prone to errors. Recent advances in artificial intelligence have opened opportunities for automating this process, offering faster, more accurate, and scalable solutions. In this work, we propose a lightweight Convolutional Neural Network (CNN) model designed to classify fruits into three categories based on freshness: fresh, rotten, and formalin-treated. The model architecture integrates four convolutional layers with regularization techniques, enabling efficient feature extraction while maintaining low computational cost. We conducted experiments on the peer-reviewed Fruit Vision dataset, consisting of 10,154 images across 15 fruit classes. The proposed CNN achieved a test accuracy of 92.19%, with training and validation accuracies of 94.10% and 92.84%, respectively, and an F1-score of 0.93, outperforming several pretrained baseline models such as VGG16, InceptionV3, and MobileNetV2. These results demonstrate the feasibility of applying deep learning for automated fruit quality inspection. Our approach offers a reliable and efficient framework that can contribute to enhanced food safety, reduced waste, and more intelligent supply chain management.

A ranking-based approach for effort-aware software defect prediction.

2025-12-22T07:51:40+00:00

Predicting software defects is a critical aspect of software quality assurance, as early identification of potential faults enables better resource allocation and ensures high-quality software production. Effort-aware defect prediction enhances testing and maintenance by prioritizing software modules that maximize defect detection while minimizing inspection effort. Traditional methods typically predict defect probability or defect density and rank modules accordingly, often optimizing metrics such as the Proportion of Found Bugs at 20% LOC (PofB@20%) using linear regression models. In this study, we proposed a novel ranking approach that directly predicts the total number of defects per software module using a Random Forest Regressor. Each module is assigned a custom score that balances two objectives: selecting modules with more predicted defects and minimizing effort, measured as lines of code. Particle Swarm Optimization (PSO) is employed to optimize the scoring function with respect to effort-aware metrics PofB@20%, Initial False Alarms (IFA), and Popt, ensuring early defect detection and near-ideal module ranking. After the initial ranking, a defect-aware re-ranking strategy adjusts the top 20% of LOC modules by replacing them with better candidates from the remaining modules, provided that doing so improves defect coverage without exceeding the LOC budget. Experimental results demonstrate that the proposed approach outperforms baseline methods, achieving a higher PofB@20% (0.402), a lower IFA (5.1), and a better Popt (0.756). The findings indicate that ranking modules based on predicted defects and inspection effort effectively helps testers detect more faults with reduced effort, confirming the superiority of the PSO-optimized ranking methodology over traditional approaches.