Science
Permanent URI for this communityhttp://repository.kln.ac.lk/handle/123456789/1
Browse
10 results
Search Results
Item Generalizability of Music Emotion Classifiers: An Evaluation of the Applicability of miremotion in Emotion Classification of Sri Lankan Folk Melodies(4th International Research Symposium on Pure and Applied Sciences, Faculty of Science, University of Kelaniya, Sri Lanka, 2019) Charles, J.; Lekamge, S.Music conveys and evokes powerful emotions, owing to various musical characteristics such as rhythm, melody, and orchestration. This amazing ability has motivated the researchers worldwide to discover relationships between music and emotion. As a result, various data mining tasks are carried out where state-of-the-art machine learning techniques are utilized in music emotion classification. However, the literature reveals that these studies frequently employ western or western classical music. Since the emotional expression in music is carried out through various musical characteristics which are cultural-specific, generalizability of classifiers trained using different ground-truth music in new contexts warrants further research. Therefore, in our study miremotion which is an existing classification model defined in MATLAB MIRToolbox was investigated for its applicability in Sri Lankan folk melodies which is an abundant source of emotion expression. miremotion is a classification model trained using Western film music which is presented by previous scholars after extensive research. Our study comprised of a listening experiment, subjects being thirty university students (age 30-35 years; from non-music disciplines) in which subjective ratings for ‘happy’, ‘sad’, ‘tender’, ‘anger’, and ‘fear’ were obtained using a seven point Likert scale. Thirty music files (30s; 44100Hz; stereo; 32bit; .wav) were employed as the music stimuli. Objective predictions using miremotion were obtained for the same on the above-mentioned emotion categories. To identify the differences among the subjective ratings and the objective predictions, paired t-tests were performed. Between the two groups no significant difference was noted only for ‘anger’ (p > 0.05) where significant differences were noted for all other emotions: ‘happy’ (p < 0.001), ‘sad’ (p < 0.05), ‘tender’ (p < 0.05), and ‘fear’ (p < 0.05). The findings reveal that music-emotion classification models cannot always be generalized and that their applicability varies depending on the emotion. Findings are also in support of the need for developing classification models for cultural-specific melodies. The study further aims to apply machine learning techniques for emotion classification of Sri Lankan folk melodies followed by a comparison among different standard classification algorithms. Use of a limited number of thirty music files in this study was due to the need for considering a reasonable number of stimuli for human listeners in an experiment. The study was also limited by the inclusion of a homogeneous study population where further studies are needed which pay higher emphasis on the annotator profiles, as music-emotion perception is highly influenced by sociocultural and educational background.Item Language identification at word level in Sinhala-English code-mixed social media text(IEEE International Research Conference on Smart computing & Systems Engineering (SCSE) 2019, Department of Industrial Management, Faculty of Science, University of Kelaniya, Sri Lanka, 2019) Shanmugalingam, K.; Sumathipala, S.Automatic analyzing and extracting useful information from the noisy social media content are currently getting attention from the research community. It is common to find people easily mixing their native language along with the English language to express their thoughts in social media, using Unicode characters or the Unicode characters written in Roman Scripts. Thus these types of noisy code-mixed text are characterized by a high percentage of spelling mistakes with phonetic typing, wordplay, creative spelling, abbreviations, Meta tags, and so on. Identification of languages at word level become a necessary part for analyzing the noisy content in social media. It would be used as an intimidate language identifier for chatbot application by using the native languages. For this study we used Sinhala-English codemixed text from social media. Natural Language Processing (NLP) and Machine Learning (ML) technologies are used to identify the language tags at the word level. A novel approach proposed for this system implemented is machine learning classifier based on features such as Sinhala Unicode characters written in Roman scripts, dictionaries, and term frequency. Different machine learning classifiers such as Support Vector Machines (SVM), Naive Bayes, Logistic Regression, Random Forest and Decision Trees were used in the evaluation process. Among them, the highest accuracy of 90.5% was obtained when using Random Forest classifierItem Use of LIME for Human interpretability in Sinhala document classification(IEEE International Research Conference on Smart computing & Systems Engineering (SCSE) 2019, Department of Industrial Management, Faculty of Science, University of Kelaniya, Sri Lanka, 2019) Kumari, P. K. S.; Haddela, P.S.With advancement of technology in Sri Lanka, use of Sinhala text usage has grown rapidly over the time where automatic categorization is helpful for efficient content management. As a result, experts tend to use machine learning application to categorize this large volume of data in an efficient and accurate manner. Most of these learning models are operating in a black-box where there is no way to understand how the model has decided which category an instance is assigned. Understanding the reason behind why learning model makes these predictions is very important to trust such models and to provide reasonable justifications in real world application. Intention of this research is to present the work carried on related to document classification model prediction interpretation where a set of text classifiers has been studied with use of SinNG5, freely available Sinhala Document corpusItem Distribution cost optimization using Big Data Analytics, Machine Learning and Computer Simulation for FMCG Sector(IEEE International Research Conference on Smart computing & Systems Engineering (SCSE) 2019, Department of Industrial Management, Faculty of Science, University of Kelaniya, Sri Lanka, 2019) Adikari, A.M. C.; Amalan, T. P.Developments in information and communication technology has made significant impact on every sector. Unfortunately, limited research exists regarding information systems for the distribution networks in Supply Chain. This study made an effort to investigate the linkage between information systems and transportation cost optimization in FMCG (Fast Moving Consumer Goods) sector. Information systems should support the management at operational and strategic level. The study focused on the operational level implementation of information system with machine learning and big data analytics. Factors, variables and constraints affecting the cost of transportation were identified from industry experts and literature. Then a case study approach applied by analyzing the distribution network data of a Sri Lankan FMCG company. A quantitative model was developed to reflect the transport cost structure and a software model was developed considering the constraints and the cost structure, to reduce the cost of transportation by big data analytics, machine learning and computer simulation. Developed model has been compared with the existing model of transportation in the FMCG manufacturer to benchmark the optimization. In proposed model, the usage of vehicles are reduced, thereby minimizing the transportation cost by increasing the consolidation possibilities, route planning and stacking models.Item A study on classifying the store positioning from the transactional data(International Research Conference on Smart Computing and Systems Engineering - SCSE 2018, 2018) Takahashi, M.; Tanaka, Y.This paper describes a customer analysis for store positioning, using data gathered from supermarkets in Japan. Among the retail industry in Japan, there are many types of reward cards used for customer retention purposes. The rewards cards or “Point Card”, is originally aimed for customer analysis purposes, but at present the full benefits have not been extracted due to issues in data analytics. This reward card has only become a method of simply distributing “virtual money” to the customer. For the efficient use of gathering data, we propose a classification method of the customer based on the objectives of visiting stores. In this study, the customers were classified into their objectives.Item An assessment of machine learning-based training tools to assist Dyslexic patients(International Research Conference on Smart Computing and Systems Engineering - SCSE 2018, 2018) Sathsara, G.W.C.; Rupasinghe, T.D.; Sumanasena, S.P.Dyslexia is a language based disability, where the patients often have difficulties with reading, spelling, writing and pronouncing words. The reading speed of Dyslexics tend to be lower than their equivalents, because of slow letter and word processing. Inspite of this disorder, a dyslexic person can be trained to read in normal speed. There are manual methods and some technical improvements can be reported such as the live-scribe smart pen, Dragon Naturally Speaking, Word processors, and Video Games. This study provides an assessment about the Machine Learning (ML) based techniques used for Dyslexic patients via a systematic review of literature, and a proposed ML based algorithm that will lay foundation for future research in the areas of machine learning, augmented and healthcare training devices.Item Analysis and detection of potentially harmful Android applications using machine learning(International Research Conference on Smart Computing and Systems Engineering - SCSE 2018, 2018) Kavneth, G.A.S.; Jayalal, S.With the rapid advancement of technology today, smartphones have become more and more powerful and attract a huge number of users with new features provided by mobile device operating systems such as Android and iOS. Android extended its lead by capturing 86% of the total market in 2017 (Gartner, 2017) and became the most popular mobile operating system. However, this huge demand and freedom has made the hackers and cybercriminals more curious to generate malicious apps towards the Android operating system. Thus, research on effective and efficient mobile threat analysis becomes an emerging and important topic in cybersecurity research area. This paper proposes a static-dynamic hybrid malware detecting scheme for Android applications. While the static analysis could be fast, and less resource consuming technique and dynamic analysis can be used for high complexity and deep analysis. The suggested methods can automatically deliver an unknown application for both static and dynamic analysis and determine whether Android application is a malware or not. The experimental results show that the suggested scheme is effective as its detection accuracy can achieve to 93% ∼ 100%. The findings have been more accurate in identifying Android malwares rather than separating those two static and dynamic behaviors. Furthermore, this research compares the machine learning algorithms for static and dynamic analysis of the Android malwares and compare the accuracy by the data used to train the machine learning models. It reveals Deep Neural Networks and SVM can be used for and higher accuracy. In addition, era of the training and testing dataset highly effect the accuracy of the results regarding Android applications.Item Machine learning based model for Android malware analysis and detection.(International Research Symposium on Pure and Applied Sciences, 2017 Faculty of Science, University of Kelaniya, Sri Lanka., 2017) Kavneth, G. A. S.; Jayalal, S.Rapid advancement of technology has enabled smartphones to become extremely powerful. They are capable of attracting a considerable amount of users with new features provided by mobile device operating systems such as Android and iOS. Android extended its lead by capturing 86 percent of the total market in 2017, and became the most popular mobile operating system. The Android operating system, which is found on a wide range of devices is owned by Google and powered by the Linux kernel. It is an open source operating system that enables mobile application developers to access unlocked hardware and develop new apps as they wish. However, this huge demand and freedom has made the hackers and cybercriminals more curious to generate malicious apps towards the Android operating system. They constantly target the security vulnerabilities in the operating system to gain access within the system. Even though, Google provides a primary set of security services, there are possibilities for potentially harmful applications in the Google Play store and other third party application stores. Thus, research on effective and efficient mobile threat analysis becomes an emerging and important topic in cybersecurity research area. Many researchers proposed various security analysis and evaluation strategies such as static analysis and dynamic analysis. In this research, we propose a hybrid approach, which aggregates the static and dynamic analysis for detecting security threats and attacks by Android malware application. This approach has two phases. First phase is the static analysis for applications, which will analyze the mobile application without execution. This focuses on extracting app APK file and examining permission requests made by Android apps that have declared in AndroidManifest.xml, as a means for detecting malwares. Because, in most of cases extra permissions granted by applications will allow the attacker to exploit the device. As the next phase, we perform dynamic analysis for mobile application. This phase focuses on runtime data obtained from the applications such as CPU, scheduler information from every running application, network calls, sensor data and so forth. For both phases, we have used supervised, machine learning based algorithms to train models and recognize malwares. In the first phase, potentially harmful applications can be identified as well as in the proposed hybrid mechanism, which is a combination of both phases. Data that was collected by several cybersecurity research centers were used for the evaluation of the proposed hybrid approach and both real-life malware and benign app data demonstrated a good detection performance with high scalability. The initial findings have been more accurate in identifying Android malwares rather than separating those two static and dynamic behaviors.Item Detection of cyber bullying on social media networks(Faculty of Science, University of Kelaniya, Sri Lanka, 2016) Priyangika. S.; Jayalal, S.Social Media is becoming an integral part of people’s daily lives today. It is an effective way of sharing one’s life experiences, special occasions, achievements and other events with their friends and family. Although it is a fruitful way to communicate with groups, some people find themselves being insulted or offended by others who are involved in certain post or conversations. These insulations can be based on racism, using profanity or any other vulgar or lewd language. This cyber bullying needs to be monitored and controlled by the social media site owners since it will highly effect on the number and safety of the active site membership. Currently, there is no automated process of identifying offensive comments by the social network site itself. It can be only diagnosed by humans after reading the comments, flagging or reporting them to the owner of the site or blocking the offender. Considering the massive big data set generated in social media daily, automatically detection of offensive statements is required to reduce insulation effectively. For this purpose, text classification approach can be applied where a given text will be categorized as insulting or not, through learning from a pre-learned model. In order to develop the model, data was collected from the popular data repository site named www.kaggle.com. The dataset consists of comments posted on Facebook and Twitter. Firstly the dataset was divided into training data set and test data set. Then the collected data was preprocessed by removing the unwanted strings, correcting words and eliminating duplicate data fields. In the next step, features or keywords were extracted which are qualified to distinguish a statement as ‘insulting’ using N-grams model and counting methods. Feature selection is done using Chi- Squared test and finally apply classification algorithms for separating insulting comments and non-insulting comments from a dataset given. Machine learning algorithms such as Support Vector Machines (SVM), Naïve Bayes, Logistic Regression and Random Forest are used for this. Out of the classification algorithms, SVM is to be performed better than other algorithms since this is a two-class classification problem and a comment is to be classified only into two separate classes which are ‘insulting’ and ‘neutral’. With an exact separation of a given comment into ‘insulting’ and ‘neutral’ category, cyberbullying happening through offensive comments posted on social media sites can be detected.Item WatchDog: An Advanced Surveillance System(Faculty of Science, University of Kelaniya, Sri Lanka, 2016) Ganepola, G.A.U.E.; Wijayasiriwardhane, T.K.Surveillance systems have become an integral part of the business world today due to the intensive care given to ensure the security of properties with a considerable monetary value. As a result, Closed-Circuit Television (CCTV) cameras are widely used in organizations. However, these systems have added an additional complexity to the user’s day-to-day work due to considerations like footage review and storage. The most common solution to this problem is incorporation of intelligence and automation to these systems. Typically, image processing and machine learning concepts are employed to implement such surveillance systems. However, the currently available advanced surveillance systems are not affordable for small and medium enterprises. The most widely used freely available advanced surveillance systems only detect motion. On the other hand, the systems that can identify the presence of people and even recognize them cost a considerable amount that does not fit into the budget of most, small scale businesses. Further, the most of the available free surveillance systems have not been designed in a way to achieve both storage efficiency and giving feedback on footages. In fact, most of them do record the footage 24x7. To address all those issues, in this paper, we present “WatchDog”, an advanced surveillance system that is implemented as a 100% free and open source product with features like detection of human presence, storage efficiency mode where the footage is stored only when there is a human in the frame, feedback and reporting facilities and recognizing people in the footage. The system detects people, and only those frames are recorded in high quality while rest of the video is saved in low quality to achieve storage efficiency. Using facial feature recognition, the system can predict factors such as gender and age of people in the footage. At the end of each day, the system produces a report with detailed information. This report would be a great relief from the user’s point of view since it drastically reduces the time to review the footages when required. Viola Jones algorithm, Haar features, Integral image, Adaboost and Cascading concepts are used for Human detections and facial feature recognition in this system. Our aim of this research is to answer the 3 major problems in surveillance systems such as affordability, storage efficiency and intelligence all at once.