Recent Submissions

  • Sentiment Analysis: Amazon Electronics Reviews Using BERT and Textblob

    ElKafrawy, Passent; Mahgoub, Abdulrahman; Atef, Hesham; Nasser, Abdulrahman; Yasser, Mohamed; Medhat, Walaa M.; Darweesh, M. Saeed; External Collaboration; Artificial Intelligence & Cyber Security Lab; Computer Science; et al. (IEEE, 2023-01)
    The market needs a deeper and more comprehensive grasp of its insight, where the analytics world and methodologies such as “Sentiment Analysis” come in. These methods can assist people especially “business owners” in gaining live insights into their businesses and determining wheatear customers are satisfied or not. This paper plans to provide indicators by gathering real world Amazon reviews from Egyptian customers. By applying both Bidirectional Encoder Representations from Transformers “Bert” and “Text Blob” sentiment analysis methods. The processes shall determine the overall satisfaction of Egyptian customers in the electronics department - in order to focus on a specific domain. The two methods will be compared for both the Arabic and English languages. The results show that people in are mostly satisfied with the percentage of 47%. For the performance, BERT outperformed Textblob indicating that word embedding model BERT is more superior than rule-based model Textblob with a difference of 15% - 25%.
  • Improving the Performance of Semantic Text Similarity Tasks on Short Text Pairs

    ElKafrawy, Passent; Gamal, Mohamed Taher; External Collaboration; Artificial Intelligence & Cyber Security Lab; Computer Science; Gamal, Mohamed Taher (IEEE, 2023-01)
    Training semantic similarity model to detect duplicate text pairs is a challenging task as almost all of datasets are imbalanced, by data nature positive samples are fewer than negative samples, this issue can easily lead to model bias. Using traditional pairwise loss functions like pairwise binary cross entropy or Contrastive loss on imbalanced data may lead to model bias, however triplet loss showed improved performance compared to other loss functions. In triplet loss-based models data is fed to the model as follow: anchor sentence, positive sentence and negative sentence. The original data is permutated to follow the input structure. The default structure of training samples data is 363,861 training samples (90% of the data) distributed as 134,336 positive samples and 229,524 negative samples. The triplet structured data helped to generate much larger amount of balanced training samples 456,219. The test results showed higher accuracy and f1 scores in testing. We fine-tunned RoBERTa pre trained model using Triplet loss approach, testing showed better results. The best model scored 89.51 F1 score, and 91.45 Accuracy compared to 86.74 F1 score and 87.45 Accuracy in the second-best Contrastive loss-based BERT model.
  • A Core Ontology to Support Agricultural Data Interoperability

    ElKafrawy, Passent; Abdelmageed, Aly; Hatem, Shahenda; ael, Tasneem; Medhat, Walaa; König-Ries, Birgitta; Ellakwa, Susan F.; Algergawy, Alsayed; External Collaboration; Artificial Intelligence & Cyber Security Lab; et al. (Gesellschaft für Informatik eV, 2023)
    The amount and variety of raw data generated in the agriculture sector from numeroussources, including soil sensors and local weather stations, are proliferating. However, these raw data in themselves are meaningless and isolated and, therefore, may offer little value to the farmer. Data usefulness is determined by its context and meaning and by how it is interoperable with data from other sources. Semantic web technology can provide context and meaning to data and its aggregation by providing standard data interchange formats and description languages. In this paper, we introduce the design and overall description of a core ontology that facilitates the process of data interoperability in the agricultural domain.
  • Dynamic Modeling and Identification of the COVID-19 Stochastic Dispersion

    ElKafrawy, Passent; Gamal, Mohamed Taher; Hedaya, Mohammed M.; Bakeer, Bahi; Zakaria, Mahmoud; External Collaboration; Artificial Intelligence & Cyber Security Lab; Computer Science; Gamal, Mohamed Taher (IEEE, 2023-01)
    In this work, the stochastic dispersion of novel coronavirus disease 2019 (COVID-19) at the borders between France and Italy has been considered using a multi-input multi-output stochastic model. The physical effects of wind, temperature and altitude have been investigated as these factors and physical relationships are stochastic in nature. Stochastic terms have also been included to take into account the turbulence effect, and the random nature of the above physical parameters considered. Then, a method is proposed to identify the developed model's order and parameters. The actual data has been used in the identification and prediction process as a reference. These data have been divided into two parts: the first part is used to calculate the stochastic parameters of the model which are used to predict the COVID-19 level, while the second part is used as a check data. The predicted results are in good agreement with the check data.
  • Using Knowledge Graph Embeddings in Embedding Based Recommender Systems

    ElKafrawy, Passent; Ragab, Ahmed Hussein; External Collaboration; Artificial Intelligence & Cyber Security Lab; Computer Science; Ragab, Ahmed Hussein (IEEE, 2023-01)
    This paper proposes using entity2rec [1] which utilizes knowledge graph-based embeddings (node2vec) instead of traditional embedding layers in embedding based recommender systems. This opens the door to increasing the accuracy of some of the most implemented recommender systems running in production in many companies by just replacing the traditional embedding layer with node2vec graph embedding without the risk of completely migrating to newer SOTA systems and risking unexpected performance issues. Also, Graph embeddings will be able to incorporate user and item features which can help in solving the well-known Cold start problem in recommender systems. Both embedding methods are compared on the movie-Lens 100-K dataset in an item-item collaborative filtering recommender and we show that the suggested replacement improves the representation learning of the embedding layer by adding a semantic layer that can increase the overall performance of the normal embedding based recommenders. First, normal Recommender systems are introduced, and a brief explanation of both traditional and graph-based embeddings is presented. Then, the proposed approach is presented along with related work. Finally, results are presented along with future work.
  • Recommender Diagnosis System with Fuzzy Logic in Cloud Environment

    ElKafrawy, Passent; Elnemr, Rasha; Aboghazalah, Maie; Elsayed, Nedaa; elsayed, Ayman; External Collaboration; Artificial Intelligence & Cyber Security Lab; Computer Science; Aboghazalah, Maie (IEEE, 2023-01)
    Recommendation systems are now used in a wide range in many fields. In the medical field, recommendation systems have a great stature to both doctors and patients for its accurate prediction. It can reduce the time and efforts spent by doctors and patients. The present work introduces a simple and effective methodology for medical recommendation system based on fuzzy logic. Fuzzy logic is an important method to be used based on fuzzy input data. The input data for each patient are not the same, on which recommendation can differ. This work aims to develop techniques for handling the patient data to urge accurate lifestyle recommendations to the patient. Fuzzy logic is utilized to form different recommendations for the patient like lifestyle recommendations, medicine recommendations, and sports recommendations based on different patient factors like age, gender and patient diseases. After evaluating the system its efficiency reached 94%. This Experiment is the final module in a four modules recommendation system. The first one is responsible for diagnosing chest diseases using ECG signals. The second one makes diagnosis using X-ray images. The third is utilizing the security of the whole system through encryption when sending user data over the cloud.
  • Early Fall Prediction Using Hybrid Recurrent Neural Network and Long Short-Term Memory

    Lytras, Miltiadis; Kwok Tai Chui; Ryan Wen Liu; Mingbo Zhao; Miguel Torres Ruiz; External Collaboration; Computer Science (Springer International Publishing, 2022-10-21)
    Falls are unintentionally events that may occur in all age groups, particularly for elderly. Negative impacts include severe injuries and deaths. Although numerous machine learning models were proposed for fall detection, the formulations of the models are limited to prevent the occurrence of falls. Recently, the emerging research area namely early fall prediction receives an increasing attention. The major challenges of fall prediction are the long period of unseen future data and the nature of uncertainty in the time of occurrence of fall events. To extend the predictability (from 0.5 to 5 s) of the early fall prediction model, we propose a particle swarm optimization-based recurrent neural network and long short-term memory (RNN-LSTM). Results and analysis show that the algorithm yields accuracies of 89.8–98.2%, 88.4–97.1%, and 89.3–97.6% in three benchmark datasets UP Fall dataset, MOBIFALL dataset, and UR Fall dataset, respectively.
  • Artificial Intelligence in Brain Computer Interface

    Subasi, Abdulhamit; No Collaboration; Artificial Intelligence & Cyber Security Lab; Computer Science; Subasi, Abdulhamit (IEEE, 2022-06-09)
    A brain-computer interface (BCI) is a connection path among brain and an external device. Motor imagery (MI) is proven to be a useful cognitive technique for enhancing motor skills as well as for movement disorder rehabilitation therapy. It is known that the efficiency of MI training can be enhanced by using BCI approach, which provides real-time feedback on the mental attempts of the subject. Artificial intelligence (AI) methods play a key role in detecting changes in brain signals and converting them into appropriate control signals. In this paper, we focus on brain signals that have been obtained from the scalp to control assistive devices. In addition, signal denoising, feature extraction, dimension reduction, and AI techniques utilized for EEG-based BCI are evaluated. Moreover, Bagging and Adaboost are utilized to classify MI task for BCI using EEG signals. Different classifiers are used to enhance the performance of detecting the signals from the brain and make it on the real time and controlling any lateness. MI related brain activities can be categorized efficiently via AI techniques. This paper utilizes wavelet packet decomposition feature extraction approach to improve MI recognition accuracy. The proposed approach classifies MI-related brain signals using ensemble techniques. The results show that the proposed framework surpasses the traditional machine learning approaches. Furthermore, the proposed Adaboost with k-NN ensemble approach also yields a greater performance for MI classification with 94.57% classification accuracy for subject independent case.