IAES International Journal of Artificial Intelligence (IJ-AI)
Not a member yet
1769 research outputs found
Sort by
AI-driven hyper-personalization and transfer learning for precision recruitment
Get PDFThe research study demonstrates how artificial intelligence (AI)-powered models can transform the hiring process by maximizing the match between candidates and jobs, leading to better hiring options and increased worker productivity. Our research develops highly personalized AI-powered recruitment applications. By using hyper-personalization to tailor job recommendations based on job compatibility and big five personality traits, this study leverages AI to improve job matching. Unlike traditional recruitment models that depend only on complex skill matching, hyper-personalization combines soft skills and personality dimensions to achieve a more precise candidate-job alignment. Transformer-based models, including bidirectional encoder representations from transformers (BERT), RoBERTa, and cross-lingual language model (XLM)-RoBERTa, have shown exceptional performance in natural language processing (NLP) and classification tasks; thus, we apply them. Transfer learning helps us to fine-tune these models to improve the accuracy of personality classification. Compared to conventional models, experimental data achieves up to 80% accuracy in binary classification and 72% in multi-class classification. By demonstrating job-candidate compatibility, this study emphasizes the potential of AI-driven models to transform recruitment, leading to better hiring decisions and workforce productivity. Our outcomes play a crucial role in advancing hyper-personalized AI applications in talent
Grid graph convolutional network-cyclical learning rate EfficientNet for liver tumor segmentation classification
Get PDFLiver tumors are identified in computed tomography (CT) images, which are crucial for accurate disease diagnosis and treatment planning as they enable clear delineation of tumors. Hence, it is vital in the field of medical radiology to segment and classify CT images of liver tumors effectively. However, liver tumor locations are not captured accurately at the boundaries in terms of size and depth within the liver due to downsampled images, leading to reduced segmentation and classification results. This research proposes a grid-graph convolutional network-based cyclical learning rate EfficientNet (GGCN-CLREN) to accurately segment and classify liver tumors. GGCN addresses inaccurate liver tumor segmentation due to downsampled images, which capture spatial relationships effectively and preserve tumor boundaries as well as depth information. For classification, CLREN optimizes classification by adjusting the learning rate, which enhances convergence and accuracy. Therefore, GGCN-CLREN ensures enhanced segmentation and classification by addressing size and depth inaccuracies. Golden sine gray wolf optimization (GSGWO) selects the most appropriate features effectively. The GGCN-CLREN achieves commendable accuracies of 99.80% and 99.96%, respectively, for the LiTS17 and CHAOS datasets when compared to the existing techniques: enhanced swim transformer network with adversarial propagation (APESTNet) and adding inception module-UNet (AIM-UNet)
Impact of batch size on stability in novel re-identification model
Get PDFThis research introduces ConvReID-Net, a custom convolutional neural network (CNN) developed for person re-identification (Re-ID) focusing on the batch size dynamics and their effect on training stability. The model architecture consists of three convolutional layers, each followed by batch normalization, dropout, and max-pooling layers for regularization and feature extraction. The final layers include flattened and dense layers, optimizing the extracted features for classification. Evaluated over 50 epochs using early stopping, the network was trained on augmented image data to enhance robustness. The study specifically examines the influence of batch size on model performance, with batch size 64 yielding the best balance between validation accuracy (96.68%) and loss (0.1962). Smaller (batch size 32)and larger (batch size 128) configurations resulted in less stable performance, underscoring the importance of selecting an optimal batch size. These findings demonstrate ConvReID-Net’s potential for real-world Re-ID applications, especially in video surveillance systems. Future work will focus on further hyperparameter tuning and model improvements to enhance training efficiency and stability
Performance analysis and comparison of machine learning algorithms for predicting heart disease
Get PDFHeart disease (HD) is a serious medical condition that has an enormous effect on people's quality of life. Early as well as accurate identification is crucial for preventing and treating HD. Traditional methods of diagnosis may not always be reliable. Non-intrusive methods like machine learning (ML) are proficient in distinguishing between patients with HD and those in good health. The prime objective of this study is to find a robust ML technique that can accurately detect the presence of HD. For this purpose, several ML algorithms were chosen based on the relevant literature studied. For this investigation, two different heart datasets the Cleveland and Statlog datasets were downloaded from Kaggle. The analysis was carried out utilizing the Waikato environment for knowledge analysis (WEKA) 3.9.6 software. To assess how well various algorithms predicted HD, the study employed a variety of performance evaluation metrics and error rates. The findings showed that for both the datasets radio frequency is a better option for predicting HD with an accuracy and receiver operating characteristic (ROC) values of 94% and 0.984 for the Cleveland dataset and 90% and 0.975 for the Statlog dataset. This work may aid researchers in creating early HD detection models and assist medical practitioners in identifying HD
A hybrid steganography scheme with reduced difference expansion and pixel-value ordering
Get PDFSteganography embeds secret messages into public media while ensuring the stego content remains visually indistinguishable from the original. The primary challenge lies in maximizing embedding capacity and image quality without introducing noticeable distortions. This research proposes a novel reversible data hiding (RDH) scheme that integrates reduced difference expansion (RDE) with four directional pixel-value ordering (PVO) schemes, horizontal, vertical, diagonal-right, and diagonal-left, to enhance embedding efficiency and visual fidelity. Unlike existing RDH methods that apply RDE with fixed or limited PVO directions, the proposed scheme dynamically selects the optimal PVO orientation based on pixel pair characteristics, effectively improving local prediction accuracy and reducing embedding-induced distortion. Previous studies have largely overlooked this relationship between pixel pair selection and embedding performance. Experimental evaluation on medical images with secret data sizes ranging from 5 kb to 100 kb demonstrates significant gains over recent PVO-based methods. The proposed method increases the average embedding capacity from 0.8315 to 0.9781 bit per pixel (bpp) (a 17.6% improvement) and raises the average peak signal-to-noise ratio (PSNR) from 49.44 to 53.40 dB, reducing distortion by approximately 3.96 dB
Evaluating the influence of feature selection-based dimensionality reduction on sentiment analysis
Get PDFAs social media has become an integral part of digital medium, the usage of the same has increased multi-fold in recent years. With increase in usage, the sentiment analysis of such data has emerged as one of the most sought research domains. At the same time, social media texts are known to pose variety of challenges during the analysis, thus making pre-processing one of the important steps. The aim of this work is to perform sentiment analysis on social media text, while handling the noise effectively in the data. This study is performed on a multi-class twitter sentiment dataset. Firstly, we apply several text cleaning techniques in order to eliminate noise and redundancy in the data. In addition, we examine the influence of regularized locality preserving indexing (RLPI) technique combined with the well-known word weighting methods. The findings obtained from experiment indicate that, RLPI outperforms other algorithms in feature selection and when paired with long short-term memory (LSTM), the combination outperforms other classification models that are discussed
Optimized ensemble framework for predicting hydroponic stock and sales using machine learning
Get PDFThe increasing global demand for food necessitates the adoption of sustainable agricultural practices. Hydroponic farming, while efficient in resource utilization, faces challenges in accurately predicting stock levels and sales due to dynamic, ever-changing factors. This research presents an optimized ensemble framework for forecasting hydroponic stock levels and sales by integrating linear regression (LR), random forest (RF), and XGBoost, further enhanced through an evolutionary algorithm (EA). The proposed framework is evaluated using root mean square error (RMSE) and mean absolute error (MAE), demonstrating significant accuracy improvements over individual models. The ensemble model achieves an RMSE reduction of 43.82% for stock prediction and 55.3% for sales forecasting compared to the best-performing individual model. Additionally, local interpretable model-agnostic explanations (LIME) are employed to offer stakeholders clear insights into decision-making processes, such as identifying "number of harvested crops" and "sales data" as key drivers of prediction outcomes. This framework supports sustainable development goals (SDGs) 9.3, 12.3, and 12.C by promoting resource efficiency, reducing food waste, and improving small-scale farmer market access. Future research will explore real-time data integration for dynamic adaptation and further model enhancements
Object detection for indoor mobile robot: deep learning approaches review
Get PDFEfficient object detection is crucial for enabling autonomous indoor robot navigation. This paper reviews current methodologies and challenges in the field, with a focus on deep learning-based techniques. Methods like you only look once (YOLO), region-based convolutional neural networks (R-CNN), and Faster R-CNN are explored for their suitability in real-time detection in dynamic indoor environments. Deep learning models are emphasized for their ability to improve detection accuracy and adaptability to varying conditions. Key performance metrics such as accuracy, speed, and scalability across different object types and environmental scenarios are discussed. Additionally, the integration of object detection with navigation systems is examined, highlighting the importance of accurate perception for safe and effective robot movement. This study provides insights into future research directions aimed at advancing the capabilities of indoor robot navigation through enhanced deep learning-based object detection techniques
Ensemble reverse knowledge distillation: training robust model using weak models
Get PDFTo ensure that artificial intelligence (AI) can be aligned with humans, AI models need to be developed and supervised by humans. Unfortunately, it is possible for an AI to exceed human capabilities, which is commonly referred to as superalignment models. Thus, it raised the question of whether humans can still supervise a superalignment model, which is encapsulated in a concept called weak-to-strong generalization. To address this issue, we introduce ensemble reverse knowledge distillation (ERKD), which leverages two weaker models to supervise a more robust model. This technique is a potential solution for humans to manage a super-alignment of models. ERKD enables a more robust model to achieve optimal performance with the assistance of two weaker models. We tried to train a more robust EfficientNet model with weaker convolutional neural network (CNN) models in a supervised fashion. With this method, the EfficientNet model performed better than the model trained with the standard transfer learning (STL) method. It also performed better than a model that was supervised by a single weaker model. Finally, ERKD-trained EfficientNet models can perform better than EfficientNet models that are one or even two levels stronger
Optimizing citrus disease detection: a transferrable convolutional neural network model enhanced with the fruitfly optimization algorithm
Get PDFFungal, bacterial, and viral diseases significantly threaten citrus production and quality worldwide, prompting producers to explore technological solutions to mitigate the financial impact of these diseases. Image analysis techniques have emerged as powerful tools for detecting citrus diseases by differentiating between healthy and diseased specimens through the extraction of discriminative features from input images. This paper introduces a valuable dataset comprising 953 color images of orange leaves from the species Citrus sinensis (L.) Osbeck, which serves to train, evaluate, and compare various algorithms aimed at identifying abnormalities in citrus fruits. The development of automated detection systems is crucial for reducing economic losses in citrus production, with this research focusing on twelve specific diseases and nutrient deficiencies. We propose a novel approach to citrus plant disease detection utilizing a hyper-parameter tuned transferrable convolutional neural network (TCNN) model, referred to as the enhanced fruitfly optimization algorithm (EFOA)-TCNN model. This model optimizes the parameters of TCNN using the EFOA and enhances architectural design by incorporating three convolutional layers alongside an energy layer instead of a traditional pooling layer. Experimental results demonstrate that the proposed EFOA-TCNN model outperforms existing state-of-the-art methods, achieving a sensitivity of 0.975 and an accuracy of 0.995