Thesis (Bachelor of Science in Computer Science)

Browse

Subject: search.filters.subject.Machine Learning

Search Results

Now showing 1 - 10 of 48
  • Thumbnail Image
    Item
    Cardiovascular disease prediction model using Machine Learning Algorithm
    (BRAC University, 2022-09) Arefin Mirdha, MD Shamsul; Mostakim, Moin
    This research uses machine learning to anticipate and detect the symptoms of specific diseases after examining some of the important elements of these diseases in order to better understand them and develop new and better treatment techniques. This study uses machine learning and generated data sets to evaluate and categorize the signs and symptoms of heart diseases. We’d like to see whether we can improve individual disease prediction processes so that we can predict cardiovascular diseases and their modalities more accurately. Therefore, the aim of this study is also to develop a more diversified model from the existing ones. We are focusing on cardiovascular diseases, which is among the world’s top causes of death. Multiple machine learning (ML) algorithms are being used more frequently to predict cardiovascular disease. We want to evaluate and describe how well ML algorithms generally forecast cardiovascular illnesses. This research analyzes the classification of cardiovascular disease using machine learning methods including Random Forest (RF), Logistic Regression, Decision Tree, Na¨ıve Bayes, Linear Algorithm, Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and Neural Network. We anticipate finding effective and efficient results that will aid in better diagnosing these cardiovascular diseases and also will help us for developing better treatment procedures.
  • Thumbnail Image
    Item
    Predicting suicidal intent from social media text post using machine learning
    (BRAC University, 2022-05) Chowdhury, Md. Mubin Ul Islam; Hasan, Mehadi; Nayem, A.K.M Muhibullah; Meem, Humaira Tasnim; Arif, Hossain
    We are living in an age of modern science where cutting-edge technology has made the world so small. We can now easily connect with people worldwide via the internet and using social media. Social media has become a popular way to connect with people and to share our thoughts and feeling with the people we are connected. People are using social media as a tool where they share their feelings, daily life activities and so on. As a result, people are spending more times on those platforms to connect with people rather than in person. People who suffer from suicidal ideation are expressing their feelings and emotions on social platforms. As suicide is now an alarming problem in our society, we can use machine learning technology to determine suicidal ideation in the early stage based on social media data such as Twitter data and Reddit data. We have combined deep learning and an artificial neural network to make a model that we have named SIP (Suicidal Intent Prediction) which can detect suicidal ideation based on the text data of social media in the first place. In our proposed SIP model, we have used Functional, Word Embedding, Dense and GRU (Gated recurrent unit), Bi-directional LSTM, Bert to build our model. We have shown that our SIP model is able to determine the suicidal ideation with a higher training accuracy of 88%, a validation accuracy of 89% and training accuracy 98% and validation accuracy 99% from SIP (Sentiment) model.
  • Thumbnail Image
    Item
    Cyberbullying Detection using Machine Learning from Social Media comments in Bangla Language
    (BRAC University, 2022-05) Tuhin, Saikat Halder; Islam, MD Touhidul; Islam, MD. Tauhidul; Rahman, Mr. Tanvir; Bin Ahsraf, Mr.Faisal
    Cyberbullying which is defined as bullying perpetrated through the use of informa tion and communication technology is a serious problem nowadays. As a result of the invention of social networks friendships through different social media, relation ships, and social communications have all gone to a new level with new definitions. In fact, people become friends with someone whom he/she cannot even know face to face. With such a huge amount of users on the internet, cyberbullying has become a widespread global phenomenon. It not only makes a person mentally low but also has become one of the most important reasons for committing suicide. Being the seventh most speaking language in the world and increasing usage of the online platform, Bangla speaking people badly need an effective cyberbullying detection to handle this issue. In this thesis paper, we explore the spread of cyberbullying in fluence through the pairwise interactions between users. For cyberbullying through language, we will collect users’ unique comments from social media and check them with the help of psychological references. After that, those comments will be cat egorized using Word embedding, an evaluation tool to categorize text, so that the dataset will be shortened and ready for classification. Lastly, the dataset will be to a machine learning classifier named Random Forest in detecting the cyberbullying comments. The performance and accuracy of numerous frequently used machine learning approaches on Bangla text are investigated in this study. In addition, the influence of user-specific information, such as location, age, gender, number of likes, number of comments, and so on, is examined for the identification of Bangla cy berbullying. Random Forest is the top effective algorithm for Bangla cyberbullying identification when just posts or comments are used to identify, according to exper imental data, with 95.78% accuracy. Therefore, Random Forest is used for applying the approach on social media since it works better.
  • Thumbnail Image
    Item
    Analysis of financial data on the time series using data from the stock market
    (BRAC University, 2022-05) Shachcha, Ifad Bhuiyan; Siam, Muhammad Ziaus; Rasel, Annajiat Alim; Khan, Rubayat Ahmed
    Predicting financial data is really important for investors Often times investors do not have a proper tool to properly assess the market and forecast their predictions. Furthermore, not only investors in modern day civilians are also willing to invest as well and as there is an abundant amount of data available from the financial sector it is of utmost significance to find the optimal algorithm in a general case scenario. This project aims to show a comparison between the results found from some of the popular neural network algorithms. In this project we have employed the help of Dense Neural Network [DNN], Recurrent Neural Network [RNN], Long Short Term Memory unit [LSTM], Convolutional Neural Network [CNN] and a pipeline where we combined LSTM and CNN. We have kept some of the parameters similar and compared the results to determine an algorithm in a general case. This would help people take informed decisions while investing.
  • Thumbnail Image
    Item
    Occluded object detection for autonomous vehicles employing YOLOv5, YOLOX and Faster R-CNN
    (BRAC University, 2022-05) Mostafa, Tanzim; Chowdhury, Sartaj Jamal; Rhaman, Dr. Md. Khalilur; Rabiul Alam, Dr. Md. Golam
    Autonomous vehicles [AVs] are the future of transportation and they are likely to bring countless benefits compared to human-operated driving. However, there are still a lot of advances yet to be made before these vehicles can be considered com pletely safe and before they can reach full autonomy. Perceiving the environment with utmost accuracy and speed is a crucial task for autonomous vehicles, ergo mak ing this process more efficient and streamlined is of paramount importance. In order to perceive the environment, AVs need to classify and localize the different objects in the surrounding. For this research, we deal with the detection of occluded ob jects to help enhance the perception of AVs. We introduce a new dataset containing occluded instances of road scenes from the perspective of Bangladesh. We utilized transfer learning to train the YOLOv5, YOLOX and Faster R-CNN models, using their respective pre-trained weights on the COCO dataset. We then evaluate and compare the performance of the three object detection algorithms on our dataset. YOLOv5, YOLOX, and Faster R-CNN achieved mAP at 0.5 metric of 0.777, 0.849 and 0.688, and mAP at 0.5:0.95 of 0.546, 0.634, and 0.422 respectively in our test set. Therefore, we find YOLOX to be the best performing model on our dataset, and its high mAP scores demonstrate the effectiveness of the model as well as the dataset.
  • Thumbnail Image
    Item
    A deep learning approach to depression detection based on Convolutional Neural Networks and Transfer Learning
    (BRAC University, 2021-10) Sarmi, Kaniz Fatima; Rahman, Shaikh Mahmudur; Sultana, Nusrat Jahan; Anzoom Shanto, Khandaker MD. Asef; Parvez, Dr. Mohammad Zavid
    Depression and mental health issues (stress, nervousness, panic attacks, anxiety attacks etc.) are nowadays a major issue in the whole world. It is a common cause of mental illness that has been linked to an increased risk of dying young. Especially in our country, mental health is an issue which most of the families do not want to give as much attention as it is supposed to get and because of that so many people who are suffering from Major Depressive Disorder (MDD) are often helpless. Currently there are numerous ways to detect depression by various methods. For example: emotion recognition, social media records, analyzing daily routine with the help of machine learning and many more. This paper aims to detect depression by implementing various deep learning/ transfer learning models (for example: VGG16, Xception, ResNet152, MobileNetV2 etc.) using EEG brain signals to discover the model that provides the highest level of accuracy for our data type. In addition, we want to analyze why the particular model performs better and what might be the cases to make a model perform better to propose a model so that this method of modeling can be used in most cases for detecting depression and model improvement. Furthermore, we have made a custom model which gives the most accuracy (99.75%). We are successful at bringing the highest accuracy among the existing models which were implemented by us. For this reason, we are analyzing the EEG brain signal data of several healthy and MDD patients. We believe that this research will aid in the development of innovative strategies for building models and early identification of depression in our daily lives.
  • Thumbnail Image
    Item
    Detecting self-esteem level and depressive indication due to different parenting style using supervised learning techniques
    (BRAC University, 2022-05) Al Taawab, Abdullah; Rahman, Mahfuzzur; Islam, Zawadul; Mustari, Nafisa; Alam, Md. Golam Rabiul; Roy, Shaily
    Uprising a child is a psychological construct of parents, which is a combination of factors that evolves over time with the growth and development of the child. Parent ing style represents a set of strategies that have diverse influences on children. These approaches can create depressive symptoms in children’s minds, which can last even if they become adolescents. Moreover, these indications may affect their level of self confidence. In this research, supervised learning models are used to detect different parenting styles, depression indications of adolescents due to parenting and the level of their self-esteem. Due to the absence of publicly available data, we created our own data set of about 500 survey responses. Additionally, eleven psychological and nine linguistic attributes of Linguistic Inquiry and Word Count (LIWC) have been used to identify depression indications. Among all the supervised models, the Lo gistic Regression (LR), Gradient Boost Classifier (GBC) and Bi-Directional LSTM (Bi-LSTM) provide better results than other models. This research is capable of helping the parents to know their children’s psychology in a better way and make them have a more profound discussion on practical life.
  • Thumbnail Image
    Item
    Recall-Net: A CNN-based Model for Four-class Classification of Alzheimer’s Disease
    (BRAC University, 2021-09) Hasan, M. M. Kamrul; Angan, Farhan Faisal; Rashid, Tasmim; Bin Ashraf, Faisal; Parvez, Dr. Mohammad Zavid
    eep learning, a cutting-edge machine learning technique, has outperformed classical machine learning at detecting detailed structures in complex multi-dimensional data, particularly in the field of computer vision. As recent advances in neuroimaging techniques have created massive multimodal neuroimaging data, the application of deep learning to early diagnosis and automated categorization of AD has recently gotten a lot of interest. It also supports biomedical researchers in the identification of many diseases such as cancer, Alzheimer’s, Malaria, and blood cell detection, among others. Deep learning is a subclass of machine learning techniques for extracting features and applying them to classification, image processing, and other tasks. A thorough Google Scholar search was conducted before and during our research to find deep learning publications on AD published between January 2010 and July 2020. After reading and evaluating, these articles were categorized and summarized according to their used algorithms and neuroimaging techniques. In our research, we have used CNN also known as ConvNet which is one of the most efficient deep learning-based neural networks to classify Alzheimer’s patients from healthy individuals with the help of MRI data. We have collected our sMRI data from ADNI. Our dataset includes a total of 6400 patients who were categorized as non-demented or having mild to severe Alzheimer’s disease. Our deep learning approach automatically distinguishes different stages of AD subjects according to their severity. Our proposed model was designed to assist with accurate classifi cation of four classes e.g. NC, EMCI, MCI and AD. Though classification is very crucial for modeling a prediction model to find out the existence and intensity of the disease, it has always been quite difficult. Separating the distinctive features from the ROI is the most challenging part. Our proposed model has a classification accuracy of 79.77%. The performance of our work was compared to several other existing approaches for multi-class classification.
  • Thumbnail Image
    Item
    Critical retinal disease detection from optical coherence tomography images by deep convolutional neural network and explainable machine learning
    (BRAC University, 2021-01) Datta, Pranab; Islam, Saniul; Das, Retuparna; Zabir, Mihiran Uddin; Alam, Md. Golam Rabiul
    Retinal disease diagnosis by machine learning can be achieved using Deep Neural Network based predictors. Use of Explainable Artificial Intelligence (XAI) has the potential to explain the black box of those neural network models which are used in identifying critical retinal diseases. Due to lack of explanation in neural networks, Machine Learning based systems are not well trusted in medical field. People still have to rely on the doctor’s clarification to come into any conclusion with medical issues. In our proposed model, we have used several Convolutional Neural Net work (CNN) models leveraging transfer learning to identify some of the critical reti nal diseases such as Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), DRUSEN from Optical Coherence Tomography (OCT) images along with the explanation. For our classification task, we have used four different CNN models namely which are ResNet 50, inceptionv3, xception, VGG 16. At first, A dataset of several thousand OCT images consisting of four classes: CNV, DME, DRUSEN, and Normal were collected. Afterward, The dataset were pre-processed and applied to our proposed CNN models to classification. We achieved the accuracy gradually 95.20 %, 94.00 %, 96.30 % and 93.30 % by performing the four deep learning model respectively Inception V3, ResNet50, VGG16, and Xception. Eventually, in order to understand the results produced by the black box models, we applied a method of Explainable AI named Layer-wise Propagation (LRP) for a better understanding of retinal disease detection by the CNN models. To add with, the LRP have analysed the models with back propagation and focused on the area of the input image based on the model’s training parameters. To sum up, our proposed model has been able to perform critical retinal diseases detection as well as the explanation behind the identification.
  • Thumbnail Image
    Item
    Recognition of Bangladeshi sign language from 2D videos using openpose and LSTM based RNN
    (BRAC University, 2021-02) Dewanjee, Tanmoy; Nuder, Azibun; Malek, Md. Imtiaz; Nanjiba, Refah; Rahman, Atia Anjum; Alam, Md. Golam Rabiul
    Sign-language recognition is an essential part of computer vision to solve a communication obstacle between the deaf-mute and the common. Bangladeshi Sign Language (BdSL) is the medium of communication of the deaf and dumb community of Bangladesh. Where 2.4 million people cannot communicate without a sign language, developing countries like Bangladesh do not have sufficient facilities for these people [34]. Our research represents a sign-language recognizer in Bangladesh which is an approach to understanding Bangladeshi sign language so that it can become a bridge between the deaf-mute community and the normal world. Though many works have been done in this field for foreign languages, there are only a few remarkable works on the Bangladeshi Sign Language, among which they used techniques that are not accessible to all, and their accuracy was also not satisfactory. Moreover, there is a shortage of publicly available datasets of Bangladeshi Sign Language. Our objective is to deliver a compact and highly accurate system that will recognize Bangladeshi Sign Language. We propose an method based on estimation of the human keypoints. First of all, we develop a BdSL dataset containing 1151 videos with ten different words. Our algorithm uses OpenPose to extract human pose from 2D videos and feed the extracted features keeping their temporal nature to an LSTM based RNN classifier that accurately classifies the signs. Our proposed sign language model classifies the signs of Bangladeshi Sign Language with 96.54% accuracy.