Faculty of Engineering, Computer Science & Engineering

Permanent URI for this communityhttp://192.248.9.226/handle/123/47

Browse

Browsing Faculty of Engineering, Computer Science & Engineering by Author "Ahangama S"

Filter results by typing the first few letters
Now showing 1 - 4 of 4
  • Thumbnail Image
    item: Thesis-Abstract
    Cross - domain recommendation system for improving accuracy by focusing on diversity
    (2022) Herath AAHWS; Ahangama S
    With the rapidly developing technology world, recommender systems also improving day by day since customer expectations also vary from new angles making new business trends. As a result of this kind of situation, enterprise-level recommender systems require more modifications with new improvements to achieve a high user satisfaction level. in that case it seems currently most commercial recommender systems are struggling with low recommender quality by decreasing user trust and expectations. On the other hand, it senses only the recommender accuracy is not sufficient to measure recommender quality. Under the major domain recommender system, the cross-domain recommender system is one of the not much-explored areas and it needs more research works focused on diversity like subjective metrics rather than accuracy. With the purpose of improving accuracy by focusing diversity on CDRS here, I have built a matrix factorization-based collaborative filtering crossdomain recommender system using explicit user feedback with movilens100k research data set. When it comes to cross-domain recommender systems, the most frequent approach is to measure and evaluate their relevancy using standard predicted accuracy metrics such as root mean squared error (RMSE), mean absolute error (MAE), and so on. Since the more need than accuracy to maintain high-quality recommendations, we need to pay attention to a few specific areas beyond accuracy like diversity and novelty. We have measured our CDRS model’s performance via RMSE, MSE, MAE, FCP, hit ratio, and Precision@k and in all cases, CDRS has achieved good performance than the general CF model. Moreover, we measured the CDRS model’s diversity and novelty and could see both are increasing when top-N increasing. These findings would be pretty much worthy when we are implementing enterprise-level cross-domain recommender systems in the future to achieve success in each modern business use case with enhancing user satisfaction.
  • Thumbnail Image
    item: Thesis-Abstract
    Hybrid CNN-LSTM model for minute - wise stock market price prediction
    (2022) Vijithasena MRG; Ahangama S
    Stock market prediction is considered as a challenging problem because of the non-linear and dynamic price changes in stock markets. And need to deal with high volume and high frequency data. Despite the fact that a variety of machine learning and deep learning approaches can be applied to construct prediction algorithms, stock value prediction is difficult due to the high frequency data. Economic factors such as change in corporate policy, economic shifts, expectations of investors, other stock markets’ movements and government change influence the stock market movements. When developing a prediction model, these influenced factors need to be considered to get highly accurate results. The successful stock market prediction results in better decisions and high profits. Minute-wise stock market prices provide better understanding about stock price behavior within a particular day. Since it is very important to thoroughly analyze stock price behavior to make trading decisions, analyzing and predicting trading trends within a day is very crucial. Rather than predicting daily close price, open price and highest price, if we can predict the next upcoming couple of minutes or hours stock price with highest accuracy, then it is a great improvement in stock market prediction. Stakeholders including buyers and sellers can get good predictions and they can make proficient decisions on time. This paper considers implementing a hybrid CNN-LSTM model to predict minute wise stock market prices by using minute-wise stock market data which provides a best performance. Stock market data of different companies including Apple, Google and Amazon were collected from Yahoo Finance API. As for the evaluation, several benchmark models were created and compared their performance with the proposed model. Furthermore, proposed model was evaluated using various datasets and timeframes. The next 5 minutes forecasted stock prices were compared with the actual prices and measured the performance of model. In this research, as for the evaluation metrics, Mean Absolute Percentage Error and Root Mean Square Error were used and the best model was selected considering the validation results. Models were fine-tuned using different time windows, model parameters and selected the best parameters for the forecasting model. Finally, the proposed model outperformed the state-of-art models for predicting short-term stock market values.
  • Thumbnail Image
    item: Thesis-Abstract
    Neural collaborative filtering based recommendation system for purchased product recommendation
    (2022) Widanagamage D; Ahangama S
    In order to validate that the problem exists, I followed the procedure as explained below. First I grouped the data set with user id and the product. Then, for each user and item, I have derived the number of views, transactions and add to cart events. Then, I have created 10 new data sets. For the first five data sets, I have assigned different weights based on the event type (i.e. view, purchase or transaction). As for the second five data sets, they were created with different volumes of view, transaction and purchased events. Then I have verified that, with the presence of outliers (view events), the purchased products are not recommended to the user. To verify this behaviour I have used Bayesian Personalized Ranking, Neural Collaborative Filtering, Generalized Matrix Factorization, Most Pop, Item KNN adjusted and Multi-Layer Perceptron models. Thereafter, I have removed view data from the data set and grouped data records based on the product and user. Next I have used a weighting scheme combined with binning to derive a rating score. Next, I have used four models to verify my solution. These includes, Bayesian Personalized Ranking, Neural Collaborative Filtering, Item KNN adjusted, Generalized Matrix Factorization and Multi-Layer Perceptron. I have used fivefold cross validation to train the models and used a separate data set for validation. The results were promising. I received a Hit ratio 0.275 for HR@10. This was a major improvement as, before this the Hit ratio was near to 0.
  • Thumbnail Image
    item: Thesis-Abstract
    Rugby event detection in broadcast videos based on visual features using deep learning
    (2022) Jayasuriya DP; Ahangama S
    A sports play event is an athletic activity that is performed by multiple players during a sporting event. Sports Event Detection is a challenging task in the domain of sports video analytics. Numerous attempts were made to detect events occurring in sports such as soccer, basketball, and cricket. Our primary objective in this research is to detect events in a Rugby sports video. In comparison to other sports, this one is more difficult due to the sport’s chaotic nature. As a result, very little research is conducted on the Rugby sport. The Rugby Events Dataset is presented in this paper as a benchmark dataset for event detection in rugby. It contains videos with temporal annotations for events as well as images with bounding box annotations for the same. Nevertheless, using deep learning and computer vision techniques, this research was able to successfully train on this dataset and detect rugby events as well as temporally localize those events in broadcasted videos. A simple classification model is used to distinguish between sports fields and other scenes in these videos, while an object detection model is used to identify sporting events. Whereas current object detection models are used to detect objects, this research demonstrates that these models can be extended to detect sports events and still produce satisfactory results. Combining tracking with object detection models increased our accuracy of localizing events in the temporal domain even further. This project has released a Sports Event Detection Framework which can be deployed in any machine. The RugbyEvents dataset is publicly available in