Browsing by Author "Deshpande, Gouri"
Now showing 1 - 3 of 3
Results Per Page
Sort Options
- ItemOpen AccessAn AI-Based Human-Centered Approach to Support Multidisciplinary Requirements Engineering(2023-01-30) Salmani, Ali; Moshirpour, Mohammad; Duffett-Leger, Linda; Far, Behrouz; Deshpande, GouriMultidisciplinary teams are often a necessity for software projects as they provide the required expertise to effectively solve complex problems. However, efficient collaboration between teams with different disciplines is challenging due to several factors including gaps of knowledge areas, establishing a process, and different requirements from various groups of stakeholders. Agile methodologies such as scrum provide a powerful approach to effectively manage software projects through tools and approaches to properly address change which is often more common in multidisciplinary teams. In this study, we will leverage process evaluation tools and techniques to analyze the efficiency of our software development process. We have evaluated this approach based on the project data recorded in Jira and GitHub. This approach is applied to a case study of a virtual healthcare intervention system to measure the team's productivity. Several deficiencies have been identified and discussed based on the results. We conclude that the enumerated deficiencies are related to the requirements engineering (RE) process. To improve the RE process, a set of solutions have been analyzed to determine their feasibility. Automating the requirements engineering process can be an efficient approach to address the aforementioned issues. The main objectives of this thesis is to devise an automated approach to 1) identify the system requirements including the new features and bugs from the users' speech and break them down into tasks, 2) find similar Jira tickets that are already implemented, and 3) estimate the amount of effort needed for the new task. By providing smart and automated support for requirements analysis and elicitation, this solution seamlessly integrates with scrum and is expected to considerably improve the efficiency of the software development process for the virtual intervention system that is used as the case study of this thesis. As part of this thesis, we aim to implement a model to determine whether tasks are similar and a model to estimate the effort required to complete each new task, which is the second and third objectives. For finding the similarities between tasks that relate to objectives 2 and 3 of the thesis, S-BERT, one of the most powerful transformer-based machine learning techniques, was utilized and trained with a dataset that was collected, pre-processed, and normalized. For estimating the required effort of the tasks, we have used an approach that converts original commit instances into a high-dimensional feature space using Kernel-based Principal Component Analysis (KPCA) along with Adversarial Learning (AL). Based on the results, the trained model has improved its ability for topic segmentation and finding similarities between requirements. As well, our model has an accuracy of 86\% when it comes to estimating the required effort.
- ItemOpen AccessAnalysis of Deep Domain Adaptation Methods for Brain Magnetic Resonance Image Segmentation(2022-12-16) Saat, Parisa; Hemmati, Hadi; Souza, Roberto; Gavrilova, Marina; Deshpande, GouriAccurate brain segmentation is critical for magnetic resonance imaging (MRI) analysis pipelines. Machine-learning-based brain MR image segmentation methods are among the state-of-the-art techniques for this task. Nevertheless, the segmentations produced by machine learning models often degrade in the presence of expected domain shifts between the test and train sets data distributions. These domain shifts are expected due to several factors, such as scanner hardware and software differences, technology updates, MRI acquisition parameters, and differences across the scanned populations. Domain adaptation (DA) methods can make machine learning models more resilient to these domain shifts. In this thesis, I investigated supervised and unsupervised deep domain adaptation methods for brain MRI segmentation. Two scenarios are investigated. In the first scenario, data shifts occur due to hardware and software differences across different MRI scanner vendors (General Electric, Philips, and Siemens). In the second scenario, data shifts occur due to differences in the scanned populations. The source brain MRI data comes from adults, while the target data corresponds to pediatric patients, whose brains are still developing. The main findings of this thesis are that there is not a single DA technique that consistently outperforms others, and hyperparameter tuning and computational times for these methods still need to be addressed before adopting these methods in clinical practice. Another important finding is that the DA techniques worked better for data shifts resulting from hardware and software differences across different MR scanner vendors than data shifts from population differences. The labeled data and source code used in this thesis were made publicly available and serve as a benchmark for evaluating DA methods for brain MRI segmentation.
- ItemOpen AccessRequirements Dependency Extraction: Advanced Machine Learning Approaches and their ROI Analysis(2022-02-02) Deshpande, Gouri; Ruhe, Guenther; Rokne, Jon; Nayebi, Maleknaz; Ferrari, Alessio; Bento, MarianaDependencies among requirements significantly impact the design, development, and testing of evolving software products. Requirements Dependencies Extraction (RDE) is a cognitively complex task due to rich semantics in natural language-based requirements, which impose challenges in automating the extraction and analysis of dependencies. The challenges intensify further when dependency types are considered. RDE is a part of the extensive decision support system to make effective software release planning, development, and testing decisions. Recently, Machine Learning and Natural Language Processing techniques have successfully automated tasks in Requirements Engineering to a large extent. Despite this success, there are some challenges to the automation of RDE - 1) Due to the nature of the problem, it is cognitively difficult to identify all the dependencies among requirements; hence generating or procuring high-quality annotations for automation through Machine Learning is an arduous task. 2) In the real-world, unlabelled data is abundant and supervised ML techniques need a training set. Lack of data for training is one of the challenges when using ML for RDE. 3) Textual requirements lack structure due to natural language, and feature extraction (transformation of the raw text into suitable internal numerical representations i.e.feature vector) techniques of NLP lead to ML techniques’ success. However, feature extraction method identification and application are cost and effort-intensive. 4) While there is a broad spectrum of Machine Learning techniques to choose from for RDE automation, not all techniques are economically viable in all the scenarios considering data size and effort investment. Hence, there is a need to evaluate the ML techniques beyond just performance measures for effective decision making. This thesis addresses these challenges and provides solutions. The results described in this thesis are derived from a series of empirical studies on industry and open-source software (OSS) datasets. The main contributions are as follows: • Performed a comprehensive assessment of Weakly Supervised Learning and Active Learning (AL) to address the data acquisition challenges using public and OSS datasets. Additionally, we compared Active Learning with Ontology-based retrieval (OBR) and further developed a hybrid solution that showed a 50% reduction in the labeling (human) effort for the two industry dataset evaluations from: Siemens Austria and Blackline safety. • Evaluated and compared a conventional ML-based Transfer Learning and state-of-the-art Deep Learning (DL) method (Fine-tuned Bidirectional Encoder Representations from Transformers (BERT)) for 6 Mozilla products (OSS) to address lack of training data challenge. We showed that the DL method outperformed the within project’s conventional ML models by 27% to 50% (on F1-score measure). ii • Demonstrated that the state-of-the-art DL method (fine-tuned BERT) could successfully overcome the feature extraction challenge of RDE as fine-tuned BERT outperformed conventional ML methods by 13% to 27% on the F1-score for the Firefox, Redmine and Typo3 product’s datasets. Also, we showed that fine-tuned BERT successfully predicted the direction of dependency. • Utilized a nine-stage ML process model and proposed a novel ROI of ML classification modeling approach. ROI of ML classification showed scenarios when it is viable to utilize complex methods over conventional methods considering the cost and benefits of data accumulation. Utilizing OSS datasets for evaluations and practitioner inputs for cost factors, we showed accuracy and ROI trade-offs in ML approach selection for RDE. Thus, we have demonstrated empirical evidence of ROI as an additional criterion for ML performance evaluation