8333 research outputs found
Sort by
A comparative analysis of Deep Learning Models for Spatio-Temporal Crime Prediction in Chicago
No full textChicago’s complicated crime patterns provide a serious public challenge. While crime is a widespread issue, Chicago’s high rates have received special attention. The city’s varied neighbourhoods experience varying degrees of crime, emphasising the importance of localised crime prevention initiatives. This study tries to address this difficulty by studying the efficacy of deep learning models. ATTNBILSTM, LSTM, and CNN-LSTM for predicting crime patterns. Additionally, geospatial clustering approaches, such as HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) and PCA-aided clustering, are used to detect and visualise crime hotspots. Analysing crime datasets and using these models can reveal trends in crime patterns, identify high-crime areas, and influence data-driven crime prevention efforts
Leveraging ResNet Architectures for Enhanced Detection of Great Apes in Video Data
No full textDetecting great apes in their natural habitats is essential for conservation, but not solely for providing critical insights into their behavior and population dynamics. It is also to reduce the manual work done by researchers when trying to process thousands of videos, resulting is great use of labour and possible avenues for human error. This thesis investigates the performance and efficiency of a ResNet-101 model with integrated Spatial Convolutional Modules (SCM) and Temporal Convolutional Modules (TCM) for great ape detection and behavior recognition using the PanAf500 dataset. The study compares the implementation of this model on both GPU and TPU, evaluating metrics such as precision, recall, F1 score, mean Average Precision (mAP), training time, and resource usage. Contrary to common expectations, the initial findings indicate that the TPU implementation exhibited longer training times and higher validation loss compared to the GPU implementation, which benefited from a Cosine Annealing Learning Rate Scheduler. This discrepancy highlights the importance of workload optimization and batch size considerations for each platform. Additionally, the research encountered unusually high reported memory usage and model size metrics, suggesting potential measurement or reporting errors that require further validation. This research also explores the impact of model quantization on reducing computational resource requirements and improving generalization. This study contributes to the fields of machine learning and wildlife conservation, aligning with several UN Sustainable Development Goals (SDGs), including climate action, life on land, quality education, and industry innovation and infrastructure. It supports the development of optimized deep learning models for real-world conservation efforts, enhancing the ability to monitor and protect great apes in their natural environments and more importantly, reduce the labour of researchers in data processing for large volumes of video data
A Data-Driven Approach to Customer Segmentation and Customer Lifetime Value Prediction in Retail and E-Commerce
No full textThe rapid advancement of digital technologies has transformed the retail and e-commerce landscape, boosting market competition and reshaping traditional business practices, making it essential for businesses to understand customer needs and behaviors. This proposed research focuses on improving Customer Segmentation and Customer Lifetime Value (CLTV) prediction in the Retail and E-Commerce sectors by integrating RFM (Recency, Frequency, Monetary) analysis with advanced clustering and regression techniques. The primary objective was to assess the effectiveness of these combined methods in enhancing marketing strategies through more accurate customer segmentation and value prediction. The methodology followed the Cross-Industry Standard Process for Data Minin (CRISP-DM) framework, starting with data collection and preprocessing, followed by the implementation of RFM analysis Hierarchical Clustering, Gaussian Mixture Models for segmentation, and Linear Regression, Random Forest, and Support Vector Regression models for CLTV prediction. The findings showed that Linear Regression using Scikitlearn, particularly when applied to log-transformed data, outperformed other models, with a significant improvement in accuracy and error reduction. LazyPredict further confirmed these results, with ElasticNetCV and LassoCV models showing superior performance. Despite limitations related to data quality and industry-specific findings, the research successfully proved the potential of these techniques in optimizing customer relationship management and targeted marketing efforts, contributing valuable insights for the retail and e-commerce industries
Tailored Resume Generation for Job Applications using RAG with LLMs
No full textIn an era of digital job applications, it is very crucial that the resume should stand out from other applications. The generic traditional resume fails to effectively highlight the skills for diverse job specifications. The research introduces a novel method of tailoring the resume that aligns with job description that leverages Retrieval Augmented Generation (RAG) with Large Language Models (LLM). The resume parsing is the information retrieval step to extract each section such as Education, Experience, skills, etc. which has also been performed using LLMs such as LLAMA-3, Mixtral and Gemma model. For retrieval, the resume data has been stored in Pinecone (as chunked PDF data) and MongoDB (entire JSON resume) vector databases. The Re-ranking techniques such as BM25 and Cross Encoder has been used to enhance the performance of the retriever. In this study, the BM25 which is a keyword-based search outperforms the cross-encoder. The prompt engineering is used to instruct the model to generate the tailor resume with one-shot approach where one example has been given a model to better understand. The variants of LLAMA-3 were used with ChatGroq to mitigate the computational load and faster inference. To provide the memory to the LLM models, the conversational buffer memory of langchain has been utilised by storing the last five previous chat history, so that the user can keep on refining the tailor resume till they get satisfied. The efficiency of the methodology is highlighted through a series of evaluations such as RAGAs Metrics, BERTScore, and Custom metrics such as Job Alignment and Content Preservation. By comparing the cosine similarity for content preservation between LLM and RAG with LLM are 72% and 89% respectively. The RAG with LLM outperforms compared to using only LLM to generate the tailored resumes. The Chatbot interface is designed for a user to have a seamless and highly user-friendly experience
Predicting Maternal Mortality Risk: Sub Urban and Rural India
No full textMaternal mortality being the critical issue in the global healthcare, especially in developing countries like India, where there is limited access to maternal health care and complications during pregnancy are often fatal. This study addresses these challenges of predicting maternal mortality risk in the rural regions of India by integrating machine learning (ML) and deep learning (DL) models by using data collected through IoT. The study compares the performances of three models: Random Forest, XG Boost and FNN. A hybrid model is built to combine the strengths of these individual models. The dataset includes features such as age, blood sugar levels, temperature of body and heart rates. Data was extensively pre-processed and hyper-parameter tuning was performed to optimize each model. The evaluation of these models is calculated based on accuracy, recall, precision and f1-score. The results showed that the XG Boost model achieved the highest accuracy Manik et al. (2020), slightly outperforming Random Forest model. However, the developed hybrid model which takes the combined outputs of all models has input, demonstrated the improved stability in the predictions. This research shows the potential of machine and deep learning models to predict maternal mortality risk and aiming in reducing the rate and improving maternal health outcomes in resource-limited settings
Eight-state Protein Secondary Structure Prediction Using NLP and Deep Learning
No full textThe protein secondary structure prediction(PSSP) is a significant task in bioinformatics, as it determines the structural arrangement such as -helices, beta-sheets, and random coils, of amino acids. These structures are used to identify the 3D structure of protein, which in turn determines the function of each protein. This research mainly investigates the effects of Natural Language Processing (NLP) techniques in integration with deep learning models to predict the eight-state protein secondary structure prediction. NLP methods such as Word2Vec, GloVe, and ESM are used for retrieving embeddings from the amino acid sequences and the study compares their effectiveness in capturing contextual protein features. The LSTM and BiLSTM with attention mechanisms used for model training, improve prediction accuracy, while challenges such as class imbalance and the inability to identify all DSSP8 states remain. The findings highlight the potential of language models but emphasize the need for incorporating additional features like PSSM and resampling strategies to enhance class prediction. This study lays a foundation for future work in integrating contextual information for improved PSSP accuracy
Liver Tumour Segmentation Using Deep Learning
No full textThis study focuses on improving the liver tumour segmentation by using state-of-art deep learning architectures including U-Net and V-Net integrated with residual networks such as ResNet and Inception. The study addresses the challenge of in methods used in medical image analysis particularly the detection of tumours is computationally expensive and is inaccurate. Transfer learning and a hybrid of loss functions (focal loss, Dice loss) increase computational efficiency to improve segmentation time from minutes to milliseconds, the results of empirical testing indicate an improvement of accuracy and the decrease of computational time of the model, based on such criteria as precision and recall, as well as the F1-score. These attributes have important clinical advantages as they accelerate and improve the accuracy of the diagnosis and location of liver tumours
Predicting Employee Attrition Using Machine Learning in Tech Industry: A Methodological Approach
No full textEmployee attrition in tech industry is a critical issue which leads to increased recruitment costs and operational disruptions as well as a loss of valuable institutional knowledge. Traditional methods such as exit interviews and employee surveys which provide limited insights since they focus on post-departure feedback which is often too late for intervention. This research explores data-driven approaches for predicting employee attrition, allowing organizations to identify employees at risk before they decide to leave. Some significant ethical considerations regarding the use of employee data for this purpose are also discussed, keeping in view the main ideas of privacy and bias along with transparency.
Predictive models have still been relatively underutilized, while there is a huge gap in research around the ethical consequences of the usage of such predictive models. The present study tries to fill the current gaps by presenting a comprehensive framework in the studies for employee turnover prediction with an emphasis on embedding ethical safeguards relating to privacy, bias detection, and transparency. This will make sure that predictive tools prove to be effective without compromising standards for ethics. Logistic regression, decision trees, and random forests were applied to the IBM HR Analytics Attrition Dataset in order to find the most influential factors on employee turnover. A high degree of data pre-processing was performed to ensure accuracy and robustness of the results. Ethical considerations-such as anonymizing data and mitigating model bias-were integrated within the research process in an effort to protect privacy and fairness for employees.
The random forest model proved to be the most accurate in identifying key predictors of attrition, such as career development opportunities and work-life balance. Also, the study ensured that demographic factors such as age and gender did not have an undue impact on the model's predictions. These findings will therefore give HR professionals actionable insights by which to design effective retention strategies while maintaining ethical integrity.
This report uses the statistical models to predict the turnover among tech employees, thereby contributing to growing field of HR analytics. It also emphasizes need to incorporate ethical considerations like the protection of privacy and the minimization of bias, in data-driven methods so that predictive models are both fair and interpretable.
This research provides significant insights that HR professionals can use to prevent employee turnover by implementing targeted intervention programs. One also needs to ensure that the model adheres to ethical guidelines that prevent using the model for malicious intents and thereby becomes a part of the ethical technology movement. This enables organizations to strike a balance between business requirements and ethical responsibility.
Although the models described demonstrate impressive predictive power, additional studies are in order, examining the influence of predictive tools on workplace culture, including employee trust. There also remain questions about how much of a say employees should have in the use of their data, and whether they should be notified when predictive models are applied to their personally identifiable information. Further research is needed to find a balance between predictive performance vs. employee consent vs. long-term ethic
Privacy-Preserving Predictive Analytics in Healthcare Using Federated Learning and Deep Learning Models
No full textFederated Learning (FL) is one of the most used strategies to solve the issue of preserving privacy in deep learning applications in healthcare. Centralized models of machine learning are a challenge for data sharing because of the patient privacy and data protection laws which do not allow data sharing between institutions. FL offers a distributed learning solution in which the data is stored in different institutions while sharing the update of the model to ensure that privacy is not infringed while at the same time learning. In this study, a decentralized FL model performed slightly better than standard benchmark models in predicting patient length of stay. Even though centralized models exhibited similar performance like de-centralised deep learning model, the objective of this research was not to achieve the highest accuracy possible but to prove that competitive performance can be achieved at the same time as privacy preservation. In the context of healthcare applications, Federated Learning’s capability to train models without exchanging the data and yet obtain high accuracy. The motivation for this research is to establish an approach for achieving high accuracy in decentralized systems that also respects privacy and security. The results stress that Federated Learning could be an effective solution that provides both high predictive accuracy and solid protection of privacy-sensitive health data
Analysis of Scalable and Efficient Approaches for Liver disease detection
No full textThe liver disease detection using machine learning has seek significant attention due to its potential to improve diagnostic accuracy and patient outcomes. This study explores an unique approach by combining texture-based features that is Gray-Level Co-occurrence Matrix (GLCM), with high-dimensional pre-trained ResNet50 deep learning model. The research aims to address gaps identified in the literature, particularly the challenge of leveraging unlabeled medical imaging datasets.
The workflow encompassed machine learning models building steps like data preprocessing, feature extraction, clustering using unsupervised techniques (KMeans, Agglomerative Clustering), and supervised classification (Random Forest, SVM, XGBoost). The classification optimization followed 6 phased approach to assess the efficacy of feature sets and clustering technique. This phase include establishing baseline performance using extracted features from GLCM and ResNet50 features from the dataset that were tested both with Kmeans cluster and agglomerative cluster. ResNet50 features combined with Agglomerative cluster labels yielded near-perfect accuracy, achieving 99% with RF and SVM, and 100% with XGBoost outperforming ResNet50 features with KMeans labels. Among the finding, the GLCM features combined with KMeans cluster labels & agglomerative cluster labels yielded the near to perfect accuracy of 100% which indicate effective capture of the patterns in the feature but it is unusual in real-world scenario and indicated the sign of over-fitting. The study attempted to advanced the performance by Pseudo-labeling, generated from clustering, that facilitated the result of classifiers. The approach was followed by the integration of GLCM + ResNet50 features against Kmeans cluster and agglomerative clusters from which combine features of GLCM + ResNet50 with agglomerative target variable outperform kmeans target variable and enhanced classification performance, achieving up to 99% accuracy with the SVM classifier on combined features.
The results demonstrated the promising potential of combining texture and deep learning-based features in detecting liver diseases with high precision. This work provides a scalable framework for integrating diverse feature sets and contributes to the advancement of machine learning applications in medical imaging