Browsing by Author "Haile, Bekele"

Now showing 1 - 2 of 2
  • No Thumbnail Available
    Item
    Construction of a Computer Based Amharic Thesaurus in Water Technology for Use with Information Storage and Retrieval Systems
    (Addis Ababa University, 1992-05) Haile, Bekele; Tadesse, Taye (PhD)
    A thesaurus is an organized list of vocabulary in one or- more disciplines structured in such a way that synonymous, broade, narrower and related terms are presented in alphabetical sequence. It is used to guide the user as to which descriptor to select f r om the indexing language for indexing or getting· access into a database to retrieve relevant information. The objective of this work is to construct a computer- based Amharic Thesaurus in water technology. Considering the responsibilities entrusted upon and the amount of information generated by the Water Resources Commission of Ethiopia, this thesaurus is of immense importance in collecting, analyzing and indexing water related documents written in Amharic, which are hitherto not properly processed, catalogued and indexed for easy r etrieval . After extensive discussions with experts in the field and project site visits, an English-Amharic dictionary holding some 2000 terms was created for use in defining terms and identifying interrelationships among the terms. Then, relationships among these terms entered into the thesaurus was established according to the principles of thesaurus construction. Finally, a sample database of the indexinglanguage is formed using CDS / ISIS, THES. PAS Program translated into Amharic. The work is not fully complete. Further effort is needed to automate the thesaurus entirely and to include local terms that are not yet fully covered. The task of facilitating the development of a multi-lingual system is also an area to be seriously considered in the future.
  • No Thumbnail Available
    Item
    Interconnect Bypass Fraud Detection Model Using Data Mining Technique
    (Addis Ababa University, 2019-08-08) Haile, Bekele; Midekso, Dida (PhD)
    Interconnect bypass fraud is a process by which official interconnect termination routes are being bypassed by using VoIP to route international call traffics into a SIM-Box device where calls are terminated and subsequently regenerated as local calls. According to communication fraud control associate (CFCA, 2017), it is categorized under a type of damage fraud along with subscription fraud. Telecom industry has been expanded dynamically as a result of the development of affordable technologies and an increasing demand of communications. However, the expansion in telecommunication industries in parallel motivated fraudsters to commit telecom fraud using different methods and techniques resulting in the decreasing of the revenue and quality of service in telecommunication providers. This thesis work focuses on predicting interconnect bypass fraud using different classfication techniques such as multilayer perceptron (MLP), support vector machine (SVM), random decision forest (RF), and J48 algorithms. To achieve our objective, call detail records (CDR) are collected from ethio telcom billing system for two months, from 41 millions active mobile subscribers. We applied cross-industrial standard process for data mining (CRISP-DM) model to the collected raw data; extracted important features from customers CDRs, and derived additional new features so as to characterize the behavior of interconnect bypass fraud. In addition, we preprocessed, aggregated and formatted the datasets convenient for the selected ML algorithms. Each algorithm was trained with five different aggregated datasets such as 4 hours, 8 hours, 12 hours, daily and weekly using two training modes (10-fold cross validation and percent split). The performance of the models were compared using confusion matrix and we proposed the best models for interconnect bypass fraud prediction. From our experiments, we found that J48 and RF models gave us the highest accuracy as compared to MLP and SVM by giving the classification accuracy of 99.99%, 99.99%, 99.84% and 95.61% respectively on 8 hours aggregated dataset.