8333 research outputs found
Sort by
From LDA to BERTopic: Evaluating Topic Modelling Methods for Aviation Safety Reports in Brazilian Portuguese
No full textThis study applies five different topic models to aviation safety reports in Brazilian Portuguese. The techniques explored are Latent Dirichlet Allocation (LDA), LDA with stemming, a cross-language model which translates the texts to English and then perform LDA, word2vec with k-means and BERTopic. The research aims to explore the dataset that was not previously used in published research and evaluate how effective the approaches applied are in identifying topics withing the corpus of reports. BERTopic outperformed the other models achieving a coherence score of 0.4819. A composite score was calculated based on the coherence and perplexity scores and used to evaluate the LDA models. LDA with stemming demonstrated the best composite score. Furthermore, Word2Vec with k-means might be a better approach for more generalised classifications
An Advanced Personalized Tweet Recommendation and Friend Suggestion System Using ChatGPT-3.5 Large Language Model, K-means Clustering, and Dynamic User Profiling
No full textAs has been exhibited from this research paper, there is an imperative need to improve the tweet recommendations and friend suggestions in social media sites such as the ‘X’. The system uses some of the following advanced technologies such GPT-3. 5 for interest scoring, K-means clustering for organizing the content and cosine similarity for friend suggestions. It also uses dynamic user interest profiling which captures and evolves as the amount of interest changes and a new algorithm for recommendations. The paper also goes deep into the system specifications involving the technology architecture and the major algorithms. Performance findings reveal that the system successfully recommends the correct suggestions and friends. Thus, the existing challenges that are relevant to real-time processing of big data and the ethical issues remain unsolved, the study offers basic insights for further development of the recommendation systems of social media
Indian Sign Language Detection and Translation using Deep Learning and Text-to-Speech
No full textThis paper explores the integration of YOLOv10, a cutting-edge object detection model, for real-time static sign recognition and translation of Indian Sign Language (ISL) into regional Hindi text and speech. Motivated by the demand for more effective communication tools for the deaf community in India, particularly in non-English speaking regions, this research compares the performance of YOLOv10 against the established YOLOv5 model. The study focused on key metrics such as accuracy, Mean Average Precision (mAP), precision, and inference time to assess the efficacy of YOLOv10 in ISL detection. The results demonstrate that YOLOv10 significantly improves upon the mAP@50:95 accuracy of YOLOv5, which was 89%, achieving 99% mAP@50:95 accuracy with 25 epochs for trained ISL words, along with superior precision and faster inference times. However, the false positive cases suggest potential overfitting, indicating the need for future work to refine the model. The findings further suggest that YOLOv10 offers enhanced real-time performance, making it a viable solution for improving accessibility and communication in both rural and urban areas of India. However, the research also identifies limitations, particularly related to the availability of diverse, high-quality data and the time-intensive nature of manual annotation. Future work will address these challenges by expanding the dataset to include a wider range of words and dynamic gestures, and by exploring the integration of Long Short-Term Memory (LSTM) networks to better capture complex sign language elements. This study not only advances the field of sign language recognition but also holds significant potential for commercial applications, particularly in developing assistive communication tools tailored to the Indian context
Reinforcement Learning Modelling for Autonomous Vehicle Navigation
No full textThe research aimed at providing a detailed investigation of the application deep learning approaches for self-driving car navigation with particular emphasis on learning of steer angle from images. The project incorporated the Udacity self-driving car simulator, which is a robust method of image data gathering and performing of the models’ validation. Three CNN architectures were crafted and trained to improve the prediction of the steering angle. The performances of the models were assessed with metrics like Mean Squared Error (MSE) as well as the R² score, where the enhanced models evidenced great enhancements with regard to the variation in driving conditions. Three CNN models for the autonomous vehicle navigation were developed and their performance assessed. The Extended Neural Network resulted in Mean Squared Error (MSE) 0. 053 with, an R² score of -0. 12, the Deep Neural Network model poses mean squared error equals to 0. 050 and the R² of -0. 16 as compared to the other developed CNNs showing the least performance with an MSE of 0. D=071,and, R² of -0. 50
Detection of AI-Generated Images using Multimodal Approach
No full textGenerative Adversarial Networks have come with great challenges in image forensics, making it increasingly hard to distinguish an AI-generated image from an authentic one. In this work, therefore, a multimodal approach using Histogram of Oriented Gradients, Local Binary Patterns, Convolutional Neural Network with Support Vector Machines, and Logistic Regression is proposed for improving classification accuracy. The methodology combines various techniques of feature extraction, which are applied in a unique way to address the deficiencies of single-feature models in detection. On rigorous experimentation, while the SVM model delivered an accuracy of 81.12%, Logistic Regression went a notch higher, with an accuracy of 83.52%, thus outperforming several other existing models. The results were driven by an emphasis on the effectiveness of feature integration in capturing wide arrays of image artifacts for improving accuracy in detection. It points out the requirement of more diverse datasets and sophisticated feature extraction methodologies to further make these detection systems even more robust. Even though this research was focused on images produced by StyleGAN, future work shall be organized with datasets from several GAN architectures in order to increase generalizability and adaptiveness for detection models. Future studies should also aim at increasing the breadth of the dataset used and the adoption of hybrid methodologies so that more adaptability and applicability of the models to the real world would be very possible
Solar Sight Forecast: Deep Learning Approaches for Solar PV Power Prediction at Bui Solar Power Station Ghana
No full textThe major crisis faced by the BUI Power Authority was to cope with the consistent distribution and generation of solar energy, which is influenced by various climatic conditions like humidity, wind, ambient temperature, global irradiation, etc. This research aims to enhance the understanding of solar power generation and enable reliable energy distribution to the organisation. The previous research that deployed machine learning models like gradient boosting and random forest achieved an accuracy of 90% and a normalised mean absolute error of 1.18%. Based on these findings, the approach to understanding how a deep learning model like LSTM can be used to increase the accuracy and overall outcome was carried out in this study. These findings can result in the potential of deep learning techniques, which can help in assisting the BUI Power Authority in utilising energy appropriately. The implication of this research is to enhance the reliability of solar energy supply, resulting in the broader goal of sustainability of renewable resources
Dynamic Time Warping Enhanced CNN-LSTM for Robust Seizure Prediction in EEG Data
No full textEpilepsy is known to be one of the most frequent neurological disorders whose manifestations cause severe burden to affected patients and their families because of the strictly chronic and often unpredictable character of the disease. It is essential to detect epileptic seizures in real time to enhance patients’ safety and quality of life; nevertheless, there is a fundamental problem concerning the high non-linearity of actual EEG signals. This paper presents a new approach integrating DTW with CNN-LSTM and applies it to the problem of improved seizure prediction. DTW helps to extract temporal patterns, and the CNN-LSTM helps in feature representation and learning of sequences. To eradicate interferences in the EEG signal, the model employs a Butterworth bandpass filter, which does not eliminate all the frequencies deemed necessary. On a publicly available EEG dataset, though the proposed hybrid model check marked superior accuracy, sensitivity, and specificity over conventional approaches for mental disorder diagnosis yielding test accuracy of 95. 62%. This makes TMS capable of real-time clinical application, according to this study’s finding. Thus, the findings of the study provide theoretical framework for seizure prediction that will help in improving the potential of better management of epilepsy in future. More refinement of the model’s parameters will be done in future work and integration of patient-specific data to increase accuracy for future seizure watch and patient management
Enhancing Customer Churn Prediction in the Telecom Sector Using Advance Machine Learning Techniques and Explainable AI
No full textIn telecommunications industry, customer churn is one of biggest issues that is faced by the operators, it is phenomenon where the customers stop using their services or switch to other operators. Telecommunications companies profitability can be heavily impacted by high churn rates, as acquiring new customers is often easy when compared to keeping existing customers. The study proposes a robust and explainable machine learning model to predict and control the customer churn. The proposed framework integrates heterogenous multi-stacking of ensemble technique, combining base models like random Forest, XGBoost, k-Nearest Neighbors (KNN) with logistic regression as meta model. To select the significant features we have applied Recursive Feature Elimination (RFE), and Synthetic Minority Oversampling Technique (SMOTE) was implemented to handle the imbalance of the class. Stratified k-fold was applied to cross validate the performance of the models. The multi-stacked model outperformed all the base models with an accuracy of 81%, while maintaining balance between recall and precision. The evaluation metrics like Accuracy, Precision, Recall, F1-score, ROCAUC score and confusion matrix was used to validate the efficiency of the model. To address the “black box” nature of the ensemble model, the Explainable AI technique called as SHapley Additive exPlanations (SHAP) was used to improve the interpretability of the models, the technique provided insights for both global and local important features. SHAP helped to identify the significant features influencing the churn like contract type, tenure, and monthly charges. These insights help to gain the trust of the stakeholders and design targeted retention strategies
Comparative Analysis of Machine Learning Algorithms For XAU/USD Prediction: Integrating Economic Indicators And Sentiment Analysis
No full textXAU/USD which is an highly traded forex pair in the financial market, is known for its volatility. Its dependence on diverse factors like economic indicators and sentiment-driven market dynamics, makes it a difficult pair to predict. While existing studies focus on either quantitative data or sentiment data. This study sought to fill this gap by integrating both historical indicators and sentiment scores from news articles to predict XAU/USD hourly rate. The prominent machine learning models were implemented and evaluated using metrics like MAE, MSE and R². Random Forest was found to be the most efficient in terms of accuracy and interoperability with Mean Absolute Error of 7.2035. Deep learning models even though they are designed for sequential data were outperformed by ensemble models. While sentiment score contributed to the predictive capability of models, their influence was limited. This research can help the traders and financial analyst to effectively predict the XAU/USD trends by providing a reliable framework. Nevertheless, some additional investigation of sentiment-driven features and real-time analysis tools is necessary to enhance model generalizability and precision
Enhancing Surveillance Security Through Violence Detection Using Advanced Deep Learning Algorithms
No full textDetection of violent activities is a prime importance in terms of public safety, security monitoring, and law enforcement support. The growing dependence on extensive surveillance systems in the public and private domains made it necessary to ascertain these violent acts in real-time, which is quite a challenge. Violence detection in surveillance videos becomes a critical task with public safety, law enforcement, and security monitoring applications. Though quite challenging, real-time detection of violent activities remains difficult to accomplish due to the dynamic nature of video data, constraints on computational efficiency, and the need to be accurate across diverse situations. Existing solutions mostly rely on traditional and special techniques of computer vision or single deep learning models, which can get bogged down while performing both tasks of higher computation efficiency and accuracy in a complex environment. This paper presents a comprehensive framework that harnesses advanced deep learning algorithms: Dense Neural Networks, Long Short Term Memory, Gated Recurrent Units, and a hybrid LSTM+GRU model, for the task. Our methodology combines spatial and sequential feature extraction from video frames, preprocessing, data augmentation, and model training. Evaluation of these models is performed using accuracy, precision, recall, F1-score, AUC, and loss to identify the best model. The GRU model outperformed all, achieving slightly better accuracy and generalization, making it the best possible solution for any real-life application. As a practical application, we have developed a Flask-based web application so that users can upload videos, which could lead to detecting violent activities