Computer Science

Permanent URI for this collectionhttps://etd.hu.edu.et/handle/123456789/76

Browse

Now showing 1 - 20 of 64

A MODEL TOWARDS PRICE PREDICTION FOR COMMODITIES USING DEEP LEARNING: CASE OF ETHIOPIAN COMMODITY EXCHANGE
(Hawassa University, 2022-10-03) SOLEN GOBENA
The development of information technology makes it possible to collect and store large amounts of data every second. Market Enterprises are generating large amounts of data, and it is difficult to use traditional data analysis methods to analyze and predict their future market price. Price predictions are an integral component of trade and policy analysis. The prices of agricultural commodities directly influence the real income of farmers and it also affects the national foreign currency. Haricot bean is produced in many areas of Ethiopia and it is rich in starch, protein, and dietary fiber, and is an excellent source of minerals and vitamins. Haricot bean is also the main agricultural commodity traded on the Ethiopian commodity exchange (ECX) market for the past 10 years. Though there are price prediction works for various crops in Ethiopia and abroad using machine learning and deep learning approaches, price prediction for Haricot bean has not been studied using machine learning as to the best of our knowledge,. The main objective of this study is to develop a price prediction model that can predict future prices of Haricot Bean traded at the ECX market based on time series data. Past 10 years, data has been obtained from the Ethiopian commodity exchange (ECX) with sample dataset size of 12272. Simple linear regression (SLR), multiple linear regression (MLR), and long short term memory (LSTM) were evaluated as predictive models. The results showed that LSTM outperformed other predictive models in all measures of model performance for predicting the Haricot Bean prices by achieving a coefficient of determination (R2 ) of 0.97, mean absolute percentage error (MAPE) of 0.015, and mean absolute error (MAE) of 0.032.
A URL-based Phishing Attack Detection and Data Protection Model
(Hawassa University, 2021-09-10) Yonathan Bukure Racho
Internet users are increasing rapidly in an uninterrupted way that is influencing the way of living. Every day billions of websites are accessed over the globe to facilitate different usage to people. This positive reinforcement is also resulting in internet abusing by hackers for their benefits. Most of the time internet abusing has experimented with over mobile phones or emails. The users are victimized by those abuse even without knowing that they are misused by hackers. Social engineering has become the tool for the hacker to manipulate users psychologically to reveal secret information. Phishing is a kind of social engineering attack with the potential to do harm to the individual or overall organization. Cybercriminal called Phisher comes up constantly in contact with individuals with creative ways to compromise the secret assets. Phishers uses the malicious URLs that are embedded over the webpage with severe threat and appears legitimate. When user clicks these links, redirects to malicious webpage where attackers ask some secrete information by misguiding user. Such kinds of attacks must be properly addressed. This thesis is focused on URL based phishing detection and data protection against such kind of attacks. Thus, the contribution of this thesis is divided into two phases that are: (1) URL based phishing attack detection, and (2) Protection of individual/organization assets. For the first phase, this thesis explored and implemented four machine learning algorithms like Decision tree, Random Forest, Naive Bayes, and Logistic Regression. Further performances of these algorithms are evaluated and compared against training and testing dataset. Based on performance result obtained, the best algorithm is recommended. For the second phase, thesis proposed a data protection model using a hybrid encryption method that combined AES and RSA algorithms. This model ensures the confidentiality of information assets as well as protect them against various kind of attacks. Overall proposed work is implemented in the Python programming language. The phishing detection phase concluded that Random forest outperforms and gave the highest accuracy of detection after important feature selection as compared to other algorithms. Results analysis conclude 96.89% and 99.06% detection accuracy over testing and training dataset respectively in Random forest. Similarly, the data protection phase encrypted and decrypted the data files very fast i.e., within few milliseconds and ensured the confidentiality of data in transit
Academic Performance Prediction Model for Teacher's Training Colleges Using Machine learning Approach
(Hawassa University, 2020-08-19) Firehiwot Getachew
Data mining is the process of extracting novel or previously unknown information from a large amount of data. The purpose of this study is to develop an academic performance prediction model and identifying the factors that affect academic performance of college student using data mining techniques. The data used for this study are 1023 active students from HCTE in 2018/19 academic year. For the consumption of this research, both primary and secondary data was used. Primary data such as age, gender, previous high school, department, library usage, study hours, sport interest, mother education, father education, time spent in social media, family support and economic status of family is collected by means of questionnaire. Secondary data was obtained from the HCTE registrar office. The prediction model was developed using multilayer perceptron (MLP) classification algorithm, Naive Bayes and J48 and correlation based feature selection (CFS) is applied to identify the predictive attributes of academic performance. Finally, Multilayer perceptron, Naive Bayes and J48 is compared using the same dataset. According to the result of the experiments, Multilayer perceptron using all attributes with test method of 10-fold cross validation and accuracy 60.6% gives better result compared to Naive Bayes, J48 and MLP after applying attribute selection. The study findings also showed that sex of the student, total courses credit hours taken by the students, study hours, assignment performance and library usage of the students are identified as a significant factor affecting academic performance. WEKA 3.8.1 tool was used for data mining process.
AMHARIC EXTRACTIVE TEXT SUMMARIZATION USING AmRoBERTa –BiLSTM MODEL
(Hawassa University, 2024-05) EDEN AHMED
Extractive text summarization is a crucial task in natural language processing, allowing users to quickly grasp the main ideas of lengthy documents. The manual summarization process is often labor-intensive and time-consuming. As the volume of information in the Amharic language continues to grow, the need for effective summarization systems has become essential. While various summarization techniques have been developed for multiple languages, research specifically focused on Amharic remains limited. Most existing studies rely on traditional methods that often lack of contextual embeddings, which are crucial for understanding the meaning within the text. Additionally, current approaches often struggle to capture long-range dependencies among sentences and none of the existing studies have utilized hybrid deep models, which have demonstrated state of-the-art performance in summarization tasks for other languages. This study addresses the challenge of extractive text summarization for Amharic news articles by proposing a hybrid deep learning model that combines the contextual understanding of AmRoBERTa with the sequential processing capabilities of Bidirectional Long Short-Term Memory. A dataset of 1,200 Amharic news articles, covering a variety of topics, was collected. Each article was segmented into sentences and labeled by experts to indicate their relevance for summarization. Preprocessing was conducted, including normalization and tokenization using AmRoBERTa, to prepare the data for modeling. The proposed model was trained using various hyperparameter configurations and optimization techniques. Its effectiveness was evaluated using ROUGE metrics. The results demonstrate that our model achieved significant performance, with a ROUGE-1 score of 44.48, a ROUGE-2 score of 34.73, and a ROUGE-L score of 44.47.
AMHARIC EXTRACTIVE TEXT SUMMARIZATION USING AmRoBERTa –BiLSTM MODEL
(Hawassa University, 2024-04-14) EDEN AHMED
Extractive text summarization is a crucial task in natural language processing, allowing users to quickly grasp the main ideas of lengthy documents. The manual summarization process is often labor-intensive and time-consuming. As the volume of information in the Amharic language continues to grow, the need for effective summarization systems has become essential. While various summarization techniques have been developed for multiple languages, research specifically focused on Amharic remains limited. Most existing studies rely on traditional methods that often lack of contextual embeddings, which are crucial for understanding the meaning within the text. Additionally, current approaches often struggle to capture long-range dependencies among sentences and none of the existing studies have utilized hybrid deep models, which have demonstrated state of-the-art performance in summarization tasks for other languages. This study addresses the challenge of extractive text summarization for Amharic news articles by proposing a hybrid deep learning model that combines the contextual understanding of AmRoBERTa with the sequential processing capabilities of Bidirectional Long Short-Term Memory. A dataset of 1,200 Amharic news articles, covering a variety of topics, was collected. Each article was segmented into sentences and labeled by experts to indicate their relevance for summarization. Preprocessing was conducted, including normalization and tokenization using AmRoBERTa, to prepare the data for modeling. The proposed model was trained using various hyperparameter configurations and optimization techniques. Its effectiveness was evaluated using ROUGE metrics. The results demonstrate that our model achieved significant performance, with a ROUGE-1 score of 44.48, a ROUGE-2 score of 34.73, and a ROUGE-L score of 44.47
AMHARIC MULTI-HOP QUESTION ANSWERING IN HISTORICAL TEXTS: A DEEP LEARNING APPROACH
(Hawassa University, 2024-11) BEREKET ENDALE
In our daily lives, questioning is the most effective way to gain knowledge. However, manual extraction of answers is time-consuming and requires expertise in the field. As a result, implementing fully question answering could accelerate extraction times and reduce the requirement for human labour. Numerous studies have been done on question answering in full resource languages like English, and others using various recent techniques. However, unlike previous research, which concentrated exclusively on single hop question answering, this thesis proposes the concept of multi-hop question answering in Amharic. Until yet, no studies have investigated multi-hop question answering in the context of the Amharic language, which includes reasoning over numerous pieces of evidence or documents to generate an answer. Furthermore, there is no existing question answering data set to address these issues; therefore, this study used deep learning for the Amharic multi-hop question answering problem, a neural network method. To do this, we preprocess our dataset using tokenization, normalization, stop word removal, and, padding before feeding it to a deep learning model such as CNN, LSTM, and Bi-LSTM to create question type classification based on the given input. Because there is no multi-hop Question answering training dataset in Amharic, training data must be created manually, which is time-consuming and tedious. It is around 1500 questions and contexts associated with five classes. The class depicts as ((0) for factoid_date, (1) for factoid_person, (2) for factoid_location, and (3) for factoid_organization. Accuracy, precision, the F-measure, and the confusion matrix are performance metrics used to evaluate the model's overall efficiency when applied to the provided dataset. According to performance measurements, the maximum achievable accuracy rates for this study's LSTM, CNN, and Bi-LSTM were 96%, 96.38%, and 97.04%, respectively. The findings indicated that the suggested Bi LSTM outperformed the other two models in terms of Amharic multi-hop questions type classification.
AMHARIC MULTI-HOP QUESTION ANSWERING IN HISTORICAL TEXTS: A DEEP LEARNING APPROACH
(Hawassa University, 2024-07-03) BEREKET ENDALE
In our daily lives, questioning is the most effective way to gain knowledge. However, manual extraction of answers is time-consuming and requires expertise in the field. As a result, implementing fully question answering could accelerate extraction times and reduce the requirement for human labour. Numerous studies have been done on question answering in full resource languages like English, and others using various recent techniques. However, unlike previous research, which concentrated exclusively on single hop question answering, this thesis proposes the concept of multi-hop question answering in Amharic. Until yet, no studies have investigated multi-hop question answering in the context of the Amharic language, which includes reasoning over numerous pieces of evidence or documents to generate an answer. Furthermore, there is no existing question answering data set to address these issues; therefore, this study used deep learning for the Amharic multi-hop question answering problem, a neural network method. To do this, we preprocess our dataset using tokenization, normalization, stop word removal, and, padding before feeding it to a deep learning model such as CNN, LSTM, and Bi-LSTM to create question type classification based on the given input. Because there is no multi-hop Question answering training dataset in Amharic, training data must be created manually, which is time-consuming and tedious. It is around 1500 questions and contexts associated with five classes. The class depicts as ((0) for factoid_date, (1) for factoid_person, (2) for factoid_location, and (3) for factoid_organization. Accuracy, precision, the F-measure, and the confusion matrix are performance metrics used to evaluate the model's overall efficiency when applied to the provided dataset. According to performance measurements, the maximum achievable accuracy rates for this study's LSTM, CNN, and Bi-LSTM were 96%, 96.38%, and 97.04%, respectively. The findings indicated that the suggested Bi LSTM outperformed the other two models in terms of Amharic multi-hop questions type classification.
ANALYSIS AND MODELING OF 5G NETWORK PERFORMANCE BASED ON RESPONSE TIME REDUCTION
(Hawassa University, 2023-11) MEKASHA MEKURIA
Despite the fact that 5G technology has the benefits of meeting all of the key requirements for a 5G system and understanding the secrets for attaining a reduced response time, which was the most dominating component in 5G, the globe had adequate bandwidth in earlier generations for daily usage. However, response time was not a major concern, but for today's applications such as VANET and ongoing online gaming, as well as for vertical industries accessibilities such as SDN (software-defined network), NFV (network function virtualization), URLLC (ultra-reliable low latency communication), backhaul connection, and control or location update information, response time is more crucial than output. To address the aforementioned challenges and gaps, the study have analyzed the numerologies to 5G NR(radio network) recognizing KPI for cellular system analysis based on human demands and technological efforts to fulfil purpose, and address the aforementioned challenges by using the 5G toolbox for techniques of simulating hidden 5G numerologies. The simulation results show that our proposed approach outperforms state-of-the-art techniques because it yields the highest probability in regarding the requirements from the access network in response time reduction. As a practical implication of the study, the researcher have realized that the adaptable subframe structure leads to a very low symbol duration, which enables low response time, as time critical applications increased, and that wider subcarrier spacing could be used for users to provide them with very low response time symbol duration. In the future work, the study planned to incorporate the channel modeling of the mmwave band was relatively complex; which does not have any perfect channel model, high capacity backhaul connectivity, for its challenging for the exponentially growing data demands of 5G and would be required more additional exploration in depth and spectrum and interference management due to the scarcity of the spectrum resources and interference issues, thus needs efficiently manage the 5G spectrum, hence should be to conduct comparative performance analysis.
ASPECT BASED SENTIMENT ANALYSIS FOR AFAAN OROMOO TEXT USING BERT
(Hawassa University, 2024-08-14) FETIYA FURI
Aspect-based sentiment analysis (ABSA) is a more important and advanced task of sentiment analysis which determine both the sentiments and the aspects within the text. It is an essential research field within natural language processing, especially for languages that lack extensive resources. This study focuses on developing an ABSA model for Afaan Oromoo language, one of the widely spoken languages in Ethiopia. Despite the rich linguistic diversity of Afaan Oromoo, there is a scarcity of computational tools and datasets for sentiment analysis in this language. Our research addresses this gap by creating a comprehensive dataset annotated with BIO annotation scheme for aspect terms and integrates CNN and BiLSTM for aspect extraction, and BERT for aspect sentiment classification. We fine-tuned pre-trained BERT model on our annotated Afaan Oromoo dataset to perform aspect based sentiment analysis. The total of 2550 review text collected from FBC Afaan Oromoo Facebook page, BBC Afaan Oromoo and other relevant social media are used for this study. After data collection, two annotators’ annotated data manually into three classes (i.e., positive, negative and neutral). The aspect terms used for study are extracted from three domain, coffee, gold and flower. Basically ten aspect terms namely (qulqullinna bunaa, oomisha bunaa, foolii, dandhama, worqee baasuu, galii, gatii, diinagdee, agarsiisa worqee and al-ergii) are used for the study. CNN-BiLSTM is used for aspect extraction and performed 92.8% of accuracy. BERT model performed accuracy of 87% for aspect sentiment classification. This work not only contributes to the development of sentiment analysis for Afaan Oromoo but also provides a framework for applying advanced NLP techniques to other low-resource languages
AUTOMATIC FISH SPECIES IDENTIFICATION USING DEEP LEARNING TECHNIQUE
(Hawassa University, 2023-03-17) HABTAMUA ZERIHUN
In recent years, the growing global population has led to an increased demand for animal protein, including fish and other aquatic products. Aquaculture has emerged as a primary method for meeting this demand. There is a need for reliable and accurate methods to identify fish species. However, the accurate identification of fish species remains a challenge as there are various fish species endemic to different regions. This research focuses on addressing this challenge by developing a system for automatic fish species identification using deep learning technique, with a specific emphasis on convolutional neural network (CNN). To accomplish the objective of the research, fish species images were collected from Lake Hawassa. The collected dataset was certified by domain experts from the Centre for Aquaculture Research and Education (CARE) at Hawassa University. A custom dataset was prepared, consisting of a total of 6000 images of six fish species: Oreochromis niloticus, Clarias garipienus, LabeoBarbus intermedius, Barbus paludinosis, Garra quadrimaculata, and Aplocheilichthys. The proposed system for fish species identification implements a preprocessing module that involves image resizing and pixel value normalization to ensure uniformity and enhance training performance. Data augmentation techniques were utilized to generate diverse training examples. For classification, convolutional neural network (CNN) is employed, either trained using Convolutional neural network (CNN) architectures or utilizing pre-trained models such as Inceptionv3, VGG16, and ResNet50. Evaluation metrics were employed with two different dataset ratios: 70/30 and 80/20 and also three pre-trained models were used for comparison. The results demonstrate that our proposed model 70/30 ratio outperforms the pre-trained models in terms of training, testing accuracy, as well as loss. Our model achieved a training accuracy of 100%, validation accuracy of 99.7% and a testing accuracy of 99.5%, indicating better learning and classification capabilities. Additionally, the model achieved a recall, precision and f1 score of 100%. This research contributes to the field of fish species identification. By leveraging deep learning techniques, Particularly CNN, our model achieves better accuracy in automatic fish species identification. It reduces reliance on expert skills, addresses unresolved problems, and contributes to the progress of accurate fish species identification
BI-DIRECTIONAL NEURAL MACHINE TRANSLATION FOR
(Hawassa University, 2023-07-12) ANDU ALEMAYEHU
Machine translation is one of the Natural Language Processing applications, which enables the translation of text from one natural language to another. This study aimed to design and develop a bidirectional English-Sidaamu Afoo neural machine translation system, as the need system has become increasingly important due to the growing number of language users, it needs to increase its presence on the web, For effective communication and information sharing, translation of various official documents, news articles, and other written texts in both languages is necessary and last to need integrating the other high-level NLP tools, but no prior solution in this area. Recently, Neural Machine Translation has emerged as a promising approach to machine translation, delivering state-of-the-art translation quality. Unlike traditional machine translation methods, NMT uses a single neural network that can be continuously fine-tuned to improve translation performance. This study aimed to develop a bidirectional Sidaamu Afoo-English machine translation system using deep learning techniques, specifically LSTM and Transformer models. In an attempt to do this study, due to un availability of parallel data for machine translation, we opted to collect parallel data from a religious domain, specifically from Bible and Sidaamu Afoo conversation. After gathering the data, experiments were conducted using 15,000 parallel sentences from different domains. To determine the optimal model, the efficiency in terms of training time, memory usage, and BLEU score was evaluated. The results showed that the Transformer model yielded the best results, with a BLEU score of 0.413 for Sidaamu Afoo to English translation and 0.465 for English to Sidaamu Afoo translation. Future work to enhance the performance of the system could include further research and the addition of more clean data and larger corpus sizes
BI-DIRECTIONAL NEURAL MACHINE TRANSLATION FOR
(Hawassa Unversity, 2023-07-07) ANDU ALEMAYEHU
Machine translation is one of the Natural Language Processing applications, which enables the translation of text from one natural language to another. This study aimed to design and develop a bidirectional English-Sidaamu Afoo neural machine translation system, as the need system has become increasingly important due to the growing number of language users, it needs to increase its presence on the web, For effective communication and information sharing, translation of various official documents, news articles, and other written texts in both languages is necessary and last to need integrating the other high-level NLP tools, but no prior solution in this area. Recently, Neural Machine Translation has emerged as a promising approach to machine translation, delivering state-of-the-art translation quality. Unlike traditional machine translation methods, NMT uses a single neural network that can be continuously fine-tuned to improve translation performance. This study aimed to develop a bidirectional Sidaamu Afoo-English machine translation system using deep learning techniques, specifically LSTM and Transformer models. In an attempt to do this study, due to un availability of parallel data for machine translation, we opted to collect parallel data from a religious domain, specifically from Bible and Sidaamu Afoo conversation. After gathering the data, experiments were conducted using 15,000 parallel sentences from different domains. To determine the optimal model, the efficiency in terms of training time, memory usage, and BLEU score was evaluated. The results showed that the Transformer model yielded the best results, with a BLEU score of 0.413 for Sidaamu Afoo to English translation and 0.465 for English to Sidaamu Afoo translation. Future work to enhance the performance of the system could include further research and the addition of more clean data and larger corpus sizes.
Bi-Directional Sidaamu Afoo - Amharic Statistical Machine Translation
(Hawassa University, 2023-04-06) Kebebush Kamiso
Machine translation (MT) is the area of Natural Language Processing (NLP) that focuses on obtaining a target language text from a source language text using automatic techniques. It is a multidisciplinary field and the challenge has been approached from various points of view including linguistics and statistics. MT usually involves one or more approaches. Our preference for this study is to develop the bi directional Sidaamu Afoo - Amharic machine translation system, make use of a statistical machine translation (SMT) approach. To conduct the experiment, a parallel corpus was collected from all possible available sources. These include mostly the Old and New Testaments of the Holy Bible for both languages. We used the monolingual Contemporary Amharic Corpus and the Sidama Afoo corpus compiled by a research team in the Informatics Faculty of Hawassa University. Different preprocessing tasks such as tokenization, cleaning, and normalization have been done to make the corpus suitable for the system. To accomplish the objective of this thesis work, we conducted four experiments using word and morpheme-based translation units with SMT for Sidaamu Afoo - Amharic language pairs. The first two experiments focus on word-based SMT and the next two on morpheme-based translation using unsupervised morphological segmentation tool; Morfessor. For each experiment, we used 30,100 parallel sentences. Out of the total parallel sentences, we used 80% (24,100) of randomly selected parallel sentences for training, 10% (3,000) for tuning and another 10% (3,000) for testing. The basic tools used for accomplishing the machine translation are Moses for the translation process which is MGIZA ++ for word and morpheme alignment and KenLM for language modeling; Morfessor for morphological segmentation. For evaluation SacreBLEU package which are BLEU, ChrF and TER metrics. According to the experimental findings, the differences between Amharic to Sidaamu Afoo and Sidaamu Afoo to Amharic in the Word-based alignment translation were 6.2, 16, and 1.9 for BLUE, ChrF2, and TER, respectively. In the Morpheme-based alignment, the differences between Amharic to Sidaamu Afoo and Sidaamu Afoo to Amharic translation were 7.5, 20.4, and 5.1, for BLUE, ChrF2, and TER respectively. In conclusion, the results show that morpheme-based alignment performance is better than word based alignment, for Amharic to Sidaamu Afoo than Sidaamu Afoo to Amharic
CABBAGE DISEASE CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORK
(Hawassa University, 2024-04-07) Samrawit Feleke
Cabbage is a widely cultivated vegetable crop susceptible to various diseases, which can significantly reduce crop yield and quality. To effectively monitor these diseases, accurate and timely classification is essential. This study proposes cabbage plant disease classification using convolutional neural networks (CNN). The proposed method involves preprocessing input images, extracting relevant features using a developed CNN model, and accurately classifying diseases affecting cabbage plants. The cabbage leaf image dataset was collected from Sidama region Hawassa zuriya deneba and east shewa zone of the Oromia region dugda woreda meki town cabbage production areas in Ethiopia. The performance of the proposed method is evaluated on 5700 datasets of cabbage plant images with five types of diseases, and one healthy cabbage class such as Alternaria leaf spot, anthracnose, black rot, cabbageworm, downy mildew, and healthy cabbage. We used the data augmentation method to expand the training dataset. 70% of the dataset was used for training, and the remaining 30% was used for testing and validation. We used existing research to compare the proposed model and obtained better results. The proposed model is evaluated using certain metrics such as precision, recall, F1-score, and accuracy. 96.80% of precision, 96.65% of recall, and 96.67% of F1-score. The proposed model achieved 99.53% training accuracy and 98.51% test accuracy for classified experiment results. The results show that the proposed approach can provide accurate and efficient disease diagnosis for cabbage plants, enabling timely and targeted intervention to prevent crop loss
CLASSIFICATION OF INJERA QUALITY USING CONVOLUTIONAL NEURAL NETWORK
(Hawassa University, 2023-09-08) MEBRHIT G/GEWERGS TEKLU
INJERA is one of the most well-known Ethiopian foods. It is also a popular food all over the world because of its whole grain product and gluten-free nature. It is prepared from different grains like teff, maize, barley, etc. but teff is the most preferred grain and it contains many nutrients to have the advantage to our health. It can cause celiac disease and controls blood sugar levels. Nowadays, many enterprises are occupied with selling INJERA to hotels and individuals and also exporting to foreign countries. But, some of the enterprises sell INJERA by adulterating with foreign particles without the knowledge of consumers to gain their lone profit. Due to this, INJERA adulteration has become a serious problem now & this might be a health risk in the near future. This Adulterated INJERA is dangerous because it can be toxic and may affect one’s health. It could deprive nutrients crucial for proper growth and development. Classification of quality INJERA using the naked eyes and also through smelling, observing the appearance, and tasting is difficult due to their visual similarities. To address this problem, we proposed a deep learning algorithm for the classification of INJERA quality based on the INJERA images. To do so, the design science research methodology was followed. To conduct this study, a total of 2230 images were collected including 1115 pure Teff INJERA samples and 1115 mixed INJERA samples. After collecting the required images, we applied image pre processing such as image resizing, and image normalization on the image datasets before adding them to the model. In this study three different CNN models with different design options like the number of layers, stride size, kernel size, and padding and with and without dropout namely the CNN3L, CNN4L, and CNN5L, were trained to determine the quality of INJERA as pure or adulterated. All of the models were able to find a successful detection result after many experiments. The CNN3L model has 99.94 percent training and 99.11 percent testing accuracy, the CNN4L pre-trained model has 99.78 percent training and 99.55 percent testing accuracy, and the CNN5L has 99.89 percent training and 99.55 percent testing accuracy. The CNN3L model has a training loss of 0.72% and a testing loss of 2.80%, the CNN4L model has a training loss of 3.74% and a testing loss of 10.1%, and the CNN5L model has a training loss of 0.86 percent and a testing loss of 1.94 percent. The experimental result demonstrates that the CNN5L model is effective for the accurate recognition of INJERA images with higher accuracy (99.55 %) and less loss value (1.94 %) in this study
CLASSIFYING EFFECT OF E-BANKING SERVICE ON DEPOSIT MOBILIZATION USING MACHINE-LEARNING TECHNIQUES
(Hawassa University, 2024-10-03) BALCHA BEKELE
Identifying services that are more likely potential to E-banking product offering is an important issue. Cooperative Bank of Oromia S.C., being one of the former private banks in Ethiopia is offering E-Banking products. The main objective of this study is to apply machine learning algorithms for developing Deposit mobilization Performance prediction Model that forecast potential of E-banking channel service in Cooperative Bank of Oromia. This research follows experimental research. For modelling purpose, data was gathered from the institution head office. Since irrelevant features result in bad model performance, data pre processing was performed in order to determine the inputs to the model. This thesis investigates the creation and assessment of six machine learning algorithms to forecast deposit behavior from customers: CART, SVM, KNN, Naïve Bayes, Logistic Regression and Random Forest. Cross tables were used to show the results of precision calculations and confusion matrices used to evaluate the performance of these models. With an emphasis on the relevance of various attributes in predicting customer deposits, the suitability of various classification algorithms, the relative effectiveness of ensemble versus base learning models, and forecasting based on influential attributes, the study tackled three main research questions. Experimental results exhibit that, the ensemble learning model achieved 98.496% accuracy in categorizing deposits, outperforming individual algorithms like KNN (98.491%) and SVM (98.401%), emphasizing the superiority of ensemble methods for deposit mobilization prediction. Random Forest Classifier identified "other_debit," "gender," and "mobile banking" as the most significant predictors of deposit mobilization, with relevance scores of 20%, 18%, and 13% respectively. Moderately important features included "mobile_credit", "mobile_debit", "card_debit", and "marital_status", while "atm_card" and "other_credit" were negligible. Finally, this thesis shows the effectiveness of machine learning in financial prediction by offering a thorough comparison of six popular categorization methods. The result offer valuable insights for enhancing customer deposit strategies at CBO and potentially other banking institutions
COLLABORATIVE APPROACH OF AGILE AND DEVOPS FOR CONTINUOUS DELIVERY OF QUALITY SOFTWARE
(Hawassa University, 2023-08) DESSALEGN MENGESHA
We are in the era of high demand for quality software in many organizations in order to achieve their organizational goals. Many organizations around the globe have shown great interest in the automation of their business processes. This in turn causes emerging and improvement of different software development methodologies and the way of service provision dramatically. Among those methodologies, Agile Software Development Methodologies and DevOps culture/tool have become more popular due to their capability on supporting rapid software development, continuous integration, and continuous delivery. Even though the two methodologies are complementary and have their own significant role in the software development lifecycle, using the two approaches independently will not bring development process improvement to the optimum level. Contextualizing the software development process enables the practitioners to improve their development process and for better productivity. The objective of this thesis work is to integrate the two approaches together with minor modifications to the DevOps team structure by extending the role of the DevOps team to the development environment. The research is conducted as experimental research and the evaluation was done by using two working projects, one using classical Agile as a control group and the other by integrated approach of Agile and DevOps as an experimental group. The number of changes accepted and developed and the number of deliveries in a specific period of time are used as measurement parameters. The experiment was done using students who joined Hawassa University Application Development Team for practical attachments. The findings of the experiment demonstrate that the experimental group project, which utilized agile methodologies in conjunction with DevOps practices, achieved superior outcomes compared to the control group project, which relied on the department's standard Agile/Scrum approach. This improvement was evident in metrics such as accepted changes and committed deliveries. Furthermore, the guideline applied to the experimental group project was refined and is included in this paper to serve as a valuable resource for future researchers and developers.
COMPUTATIONAL MODEL FOR WOLAYITTA LANGUAGE SPELLING CHECKER
(Hawassa Inversity, 2020-08-10) RAHEL SHUME
Spelling checker systems are built, and researches are conducted worldwide to meet the needs of different languages. In Ethiopia, spelling checker researches were carried out for only Amharic language. These works have paved the way for researches on other languages like Wolayitta. A computational spelling checker is proposed and adopted for the Wolayitta language in this research. Word level spelling checker were built based on an edit distance algorithm and language model based open source tools and platforms. A dictionary database was constructed from 13,313 words from Wolayitta lexicon dictionary and 20,000 words from Wolayitta Bible. The database was then clustered into training and testing. The statistical model accept input from the user then check whether it is available in dictionary or not if it is available it will do nothing if not it gives suggestion. The testing was done in two phases where the first is error detection rate and the second evaluation is error correction. Tests were carried out on the input data from the user, and accuracy of 94.57% was achieved for spelling error detection, while accuracy of around 90.98 % was retrieved for the spelling suggestion model. The results were promising, and further researches can be entertained as per the recommendations made by the researcher
CONTEXT-BASED SPELL CHECKER FOR SIDAAMU AFOO
(Hawassa University, 2022-03-04) MEKONEN MOKE WAGARA
A spell checker is one of the applications of natural language processing that is used to detect and correct spelling errors in written text. Spelling errors that occur in the written text can be non-word errors or real-word errors. A non-word error is a misspelled word that is not found in the language and has no meaning whereas a real-word error, that is, the word is a valid word in the language but it does not fit contextually in the sentence. We designed and implemented a spell checker for Sidaamu Afoo that can detect and correct both non-word and real-word errors. Sidaamu Afoo is one of the languages spoken in the Sidaama region in the south-central part of Ethiopia. It is an official working language and is used as a medium of instruction in primary schools of the Sidaama national regional state in Ethiopia. To address the issue of spelling errors in the Sidaamu Afoo text, a spell checker is required. In this study, the dictionary look-up approach with a hashing algorithm is used to detect non-word errors, and the character-based encoder-decoder model is used to correct the non-word errors. The LSTM model with attention mechanism and edit distance is used to detect and correct the context based spelling error. To conduct the experiment, 55440 sentences were used, of which 90% were for training (i.e., 49,896) and 10% were for testing (i.e., 5544). According to the experimental results, for an isolated spell checker, dictionary lookup with hashing achieved an accuracy of 93.05%, a recall of correct words of 91.51%, and a precision of incorrect words of 72.37% for detection. The encoder decoder model achieved a recall of 91.76% for corrections. For a context-sensitive spell checker, the LSTM model with attention and edit distance achieved an accuracy of 88.8%, recall of the correct word of 86.18%, and precision of the incorrect word of 62.84% for detection. It achieved a recall of 74.28% for the correction. The results of the experiment show that the model used to detect and correct both non-word and real-word spelling errors in Sidaamu Afoo’s written text performed well. Finally, to improve the performance of the model, we recommend using additional data set and a state-of-the-art transformer model.
CONTEXT-BASED SPELL CHECKER FOR SIDAAMU AFOO USING HYBRID APPROACH
(Hawassa University, 2024-04-08) BEZABIH BEYENE
Spellcheck involves identifying and suggesting corrections for incorrectly spelled words within the text. Its integration spans various applications such as digitally correcting handwritten text, aiding user word corrections during retrieval, and more. This thesis outlines the creation, implementation, and assessment of a model intended to rectify both non-word and real-word errors. The central objective of this research is to devise a context-based spellchecker for Sidaamu afoo. This system relies on the language's error patterns, deduced from word sequences within input sentences. The chosen technique for this spellchecking entails an unsupervised statistical method, which is particularly beneficial for languages like Sidaamu afoo by enabling analysis without the need for extensive tagged datasets. The process of rectifying spelling unfolds through distinct phases: identifying errors, proposing potential corrections, and arranging these suggestions by priority. Error identification hinges on a combination of dictionary lookup and bigram analysis. Data for the dictionary and Bigram model, essential for error detection and correction, were collected from diverse sources by the researcher. Addressing non-word errors involves computing the similarity between the misspelled word and tokens in the dictionary, measured using the Levenshtein distance, resulting in ranking and correction suggestions. In cases of real-word errors, bigram frequency aids in error detection, while bigram probability informs the correction process for misspelled words. The experimental phase encompassed the utilization of 52,093 tokens and 5,788 tokens for model learning and testing, respectively. The outcome revealed a spellchecker recall score of 92.4% and an accuracy rate of 92.5% for both non-word and real-word errors. These findings, aligned with the gated result accuracy of 92.5%, underscore the system's capability to rectify Sidaamu afoo misspellings. Future enhancements could explore advanced neural architectures to improve model quality further