Browsing by Author "Munasinghe, Lankeshwara"

Now showing 1 - 5 of 5

Anomaly detection in cloud network data
(Department of Industrial Management, Faculty of Science, University of Kelaniya, Sri Lanka, 2020) Yasarathna, Tharindu Lakshan; Munasinghe, Lankeshwara
Cloud computing is one of the most rapidly expanding computing concepts in the modern IT world. Cloud computing interconnects data and applications served from multiple geographic locations. A large number of transactions and the hidden infrastructure in cloud computing systems have presented a number of challenges to the research community. Among them, maintaining the cloud network security has become a key challenge. For example, detecting anomalous data has been a key research area in cloud computing. Anomaly detection (or outlier detection) is the identification of suspicious or uncommon data that significantly differs from the majority of the data. Recently, machine learning methods have shown their effectiveness in anomaly detection. However, identifying anomalies or outliers using supervised learning methods still a challenging task due to the class imbalance and the unpredictable nature and inconsistent properties or patterns of anomaly data. One-class classifiers are one feasible solution for this issue. In this paper, we mainly focused on analyzing cloud network data for identifying anomalies using one-class classification methods namely One Class Support Vector Machine(OCSVM) and Autoencoder. Here, we used a benchmark data set, YAHOO Synthetic cloud network data set. To the best of our knowledge, this is the first study that used YAHOO data for detecting anomalies. According to our analysis, Autoencoder achieves 96.02 percent accuracy in detecting outliers and OCSVM achieves 79.05 percent accuracy. In addition, we further investigated the effectiveness of a one class classification method using another benchmarked data set, UNSW-NB15. There we obtained 99.10 percent accuracy for Autoencoder and 60.89 percent accuracy for OCSVM. The above results show the neural network-based methods perform better than the kernel-based methods in anomaly detection in cloud network data.
Hate Words Detection Among Sri Lankan Social Media Text Messages
(Department of Industrial Management, Faculty of Science, University of Kelaniya Sri Lanka, 2022) Shalinda, J. A. D. U.; Munasinghe, Lankeshwara
The number of Sri Lankan social media users have been increased with the rapid growth of 23% between 2020 and 2021, reaching 7.9 million in 2021 January. Social media platforms became more popular when they started supporting native languages. The problems with social media also evolved as popularity grows. Social media platforms were banned for Sri Lankan users in 2019 to prevent the spreading of hate messages and incorrect information among citizens. The lack of automatically recognizing tools for hate messages in Sinhala and Romanized Sinhala was reported as the reason for the ban. It’s also a waste of time and money to manually identify them. Many studies have been conducted to identify hate messages in both English and Sinhala separately. Users in Sri Lanka tend to combine Sinhala, Romanized Sinhala, and English phrases while expressing their opinions.” Mama job ekakata apply kara,” for example. To train, an open-source data set which consists of 2500 comments, was used. And the comments were categorized as either hateful or non-hateful. To pre-process the data set, an Open-source stop word corpus and stem word corpus in Sinhala were utilized, and two corpus were manually converted into Romanized Sinhala stop word corpus and Romanized Sinhala stem word corpus to identify stop words and stem words in Romanized Sinhala. All English words were recognized using an open- source English word corpus, and a library was utilized to obtain stop word corpus and stem English words. As a result, doing research to identify hate speech in all of the languages indicated above will be more effective in reaching Sri Lankan users. The bag of words and term frequency-inverse document frequency were compared for feature engineering. Linear Support vector classifier, Random Forest Classification, SGD classifier, Logistic Regression, XGBoost classifier and multinomial Naive Bayes classifier are used as classification algorithms and evaluated. Using the SGD classification using TF-IDF with uni&bi-gram, the highest accuracy was determined to be 74.2%.
Review on Decision Support Systems used for Resource Allocation in Health Crises
(Faculty of Computing and Technology (FCT), University of Kelaniya, Sri Lanka, 2021) Diyasena, Deshanjali; Arambepola, Nimasha; Munasinghe, Lankeshwara
A disaster or crisis can be stated as a serious disruption occurring for a certain period of time, which could cause loss of human lives, properties, and disrupt the day-to-day life of people. Managing such situations is always a challenge due to various reasons. Especially, allocating and providing resources to manage disaster situations to restore the normal life of people is the main challenge in a disaster situation. Having a proper mechanism for resource allocation could save thousands of human lives as well as properties. Modern smart technologies play a vital role in designing and developing solutions for efficient and effective resource allocation mechanisms. For example, the COVID-19 pandemic has forced people to work from home using digital platforms. Those digital platforms have been able to support people to do their routine work while maintaining social distancing which minimizes the spread of Covid-19. On the other hand, those digital platforms provide an easy and fast way for healthcare officials to reach infected patients to provide necessary treatments and care. Present research critically reviews the past research on managing resources in health crises particularly falls under pandemics and epidemics.
Temporal preferential attachment: Predicting new links in temporal social networks
(Department of Industrial Management, Faculty of Science, University of Kelaniya Sri Lanka, 2021) Wickramarachchi, Panchani; Munasinghe, Lankeshwara
Social networks have shown an exponential growth in the recent past. It has estimated that nearly 4 billion people are currently using social networks. The growth of social networks can be explained using different models. Preferential Attachment (PA) is a widely used model, which is often used to link prediction in social networks. PA tells that the social network users prefer to get linked with popular users in the network. However, the popularity of a node depends not only on the node’s degree but also on the node's activeness which is reflected by the amount of active links the node has at present. Activeness of a link can be quantified using the timestamp of the link. The present work introduces a novel method called Temporal Preferential Attachment (TPA) which is defined on the activeness and strength of a node. Strength of a node is the sum of weights of links attached to the node. Here, the weights of the links are assigned according to their activeness. Thus, TPA captures the temporal behaviors of nodes, which is a vital factor for new link formation. The novel method uses min - max scaling to scale the time differences between current time and the timestamps of the links. Here, the min value is the earliest timestamp of the links in the given network and max value is the latest timestamp of the links. The scaled time difference of a link is considered as the temporal weight of the link, which reflects its activeness. TPA was evaluated in terms of its link prediction performance using well-known social network data sets. The results show that TPA performs well in link prediction compared to PA, and show a significant improvement in prediction accuracy.
What makes job satisfaction in the information technology industry?
(Department of Industrial Management, Faculty of Science, University of Kelaniya Sri Lanka, 2021) Arambepola, Nimasha; Munasinghe, Lankeshwara
Having a rich human resource is critical for an organization to move towards success. Especially, for business organizations such as technology companies, the human resource is the driving factor of the company's growth which depends on employees' motivation, skills and quality of work. Employees often change their jobs when they are not satisfied with it. Different factors may cause a change in the level of job satisfaction of an employee. For example, the dynamic nature of the Information Technology (IT) industry is an impactful factor that determines the job satisfaction of IT professionals. Foreseeing the employees' job satisfaction makes it easy for a company to take swift actions to improve the job satisfaction of its employees. In this research, we analyzed the effectiveness of machine learning (ML) methods for predicting job satisfaction using employee job profiles. There are job-specific factors in each job domain, and those factors may influence job satisfaction levels. Therefore, this research focused on the following fundamental questions: 1) How do existing ML models perform when predicting job satisfaction of software developers? 2) Can the job satisfaction prediction models be generalized to the other job roles in the IT industry? This study compared the performance of classification models: Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Neural Network (NN) in predicting the level of job satisfaction. Our experiments used two benchmark datasets: Stack Overflow developer survey and IBM HR analytics dataset. The experimental analysis shows that both employee-related factors and company-related factors contribute similarly to predicting job satisfaction. On average, the above ML models predict the job satisfaction of software developers with an accuracy of around 79%.