Repository logo
 

Faculty of Accounting and Informatics

Permanent URI for this communityhttp://ir-dev.dut.ac.za/handle/10321/1

Browse

Search Results

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Credit score prediction using genetic algorithm-LSTM technique
    (IEEE, 2022-03) Adisa, Juliana; Ojo, Samuel; Owolawi, Pius; Pretorius, Agnieta; Ojo, Sunday O.
    In data mining, the goal of prediction is to develop a more effective model that can provide accurate results. Prior literature has studied different classification techniques and found that combining multiple classifiers into ensembles outperformed most single classifier approaches. The performance of an ensemble classifier can be affected by some factors. How to determine the best classification technique' Which combination method to employ' This paper applies Long Short-Term Memory (LSTM), one of the most advanced deep learning algorithms which are inherently appropriate for the financial domain but rarely applied to credit scoring prediction. The research presents an optimization approach to determine the optimal parameters for a deep learning algorithm. The LSTM parameters are determined using an optimization algorithm. The LSTM parameters include epochs, batch size, number of neurons, learning rate and dropout. The results show that the optimized LSTM model outperforms both single classifiers and ensemble models.
  • Thumbnail Image
    Item
    Estimating effect of total specific atmospheric attenuation on performance of FSO communication link in South Africa
    (Engineering and Technology Publishing, 2022) Maswikaneng, Solly P.; Adebusola, Samuel O.; Owolawi, Pius A.; Ojo, Sunday O.
    In comparison with Radio Frequency (RF), the Free Space Optical Communication (FSOC) provides higher bandwidth, free license operation, and less initial expenditure. However, its susceptibility to changes in atmospheric weather conditions. In this paper, the impact of irradiance fluctuation on FSO systems was estimated using Rytov theory for major cities in South Africa. The extent to which the refractive index structure parameter, propagation distance and link margin affect the optical signal power at the receiver is discussed and the different methods used in evaluating the atmospheric turbulence effect are investigated. In order to achieve the stated aim, meteorological data, altitude, visibility, and wind speed were obtained from the archive of South Africa Weather Services for a period of 3years (2016-2018) over seven locations which include Cape Town, Pretoria, Upington, Bloemfontein, Emalahleni, Polokwane, East London. Results show that Emalahleni was found to possess the poor visibility of 4.4 km because of foggy conditions due to the activities of miners and other environmental factors, followed by East London with average visibility of 4.8 km. From the analysis of link margin, it was shown that FSO link attenuation reduces at higher wavelengths and long link distances due to the effect of geometric and atmospheric losses. The results show that the rate of decrease in link margin is much higher in the inland regions as compared to the coastal regions; therefore, FSO systems are prone to outage during high rainfall and longer range of connections.
  • Thumbnail Image
    Item
    A bisociated research paper recommendation model using BiSOLinkers
    (Insight Society, 2022-01-01) Maake, Benard M.; Ojo, Sunday O.; Zuva, Keneilwe; Mzee, Fredrick A.
    In the current days of information overload, it is nearly impossible to obtain a form of relevant knowledge from massive information repositories without using information retrieval and filtering tools. The academic field daily receives lots of research articles, thus making it virtually impossible for researchers to trace and retrieve important articles for their research work. Unfortunately, the tools used to search, retrieve and recommend relevant research papers suggest similar articles based on the user profile characteristic, resulting in the overspecialization problem whereby recommendations are boring, similar, and uninteresting. We attempt to address this problem by recommending research papers from domains considered unrelated and unconnected. This is achieved through identifying bridging concepts that can bridge these two unrelated domains through their outlying concepts – BiSOLinkers. We modeled a bisociation framework using graph theory and text mining technologies. Machine learning algorithms were utilized to identify outliers within the dataset, and the accuracy achieved by most algorithms was between 96.30% and 99.49%, suggesting that the classifiers accurately classified and identified the outliers. We additionally utilized the Latent Dirichlet Allocation (LDA) algorithm to identify the topics bridging the two unrelated domains at their point of intersection. BisoNets were finally generated, conceptually demonstrating how the two unrelated domains were linked, necessitating cross-domain recommendations. Hence, it is established that recommender systems' overspecialization can be addressed by combining bisociation, topic modeling, and text mining approaches.