Manchester Journal of Artificial Intelligence and Applied Sciences

Hybridizing a Conceptual Hydrological Model with Neural Networks to Enhance Runoff Prediction

2022-05-22T15:39:05+01:00

In this paper, a hybrid rainfall-runoff model is developed by embedding an artificial neural network (ANN) into a monthly lumped conceptual model termed the dynamic water balance model (dynwbm). Based on this approach, groundwater storage parameters are eliminated, and direct runoff and deep percolation outputs derived with three conceptual parameters are served as input to the ANN part. In this serial hybridization, the dynwbm parameters and ANN weights are calibrated simultaneously. The proposed hybrid model, dynNN, equipped with an automatic calibration algorithm, has been applied at a key location in the Gediz River Basin of western Turkey. From the performance measures assessed, it has been proven that a robust model has taken the strengths of conceptual and artificial neural networks while it has disregarded their drawbacks.

Radiological medical reports classification

2023-09-04T22:02:26+01:00

The data gathered from the information system of Neuroradiology, Department of Neurosciences in “Mother Teresa” University Hospital Center in Tirana, Albania is composed of text and scan images. The textual information is very important as it not only represents the data about the patient but also the radiological interpretation of the images can help the physicians in medical diagnosis and patient treatment. Text classification would help further in analyzing the data and automatic classification into classes of interest. A dataset of 1313 radiology reports classified by radiologists in 6 categories is used in this study to analyze and compare two types of classifiers: the Naïve Bayes and character level n-gram classifier trained using sequences of three, four, five and six characters. Experimental results show that the most successful classifier is the Naïve Bayes classifier with best efficiency and accuracy reaching 98%.

Intelligent Machine Learning Customer Segmentations Algorithm

2023-09-06T23:29:23+01:00

The rapid growth of online shopping and the direct reliance of most people on social media resulted in the accumulation of huge amounts of business data. The analysis of business data will provide management with the information that will help them in promoting their business and improve company revenues. This is achieved by utilising Machine Learning (ML) technologies to classify their customers based on purchasing preferences and similarity analysis. This will provide the companies with information regarding their own customers plus the ability to split their customers into categories which can lead to targeting the most loyal customers with priority in every transaction, plus the chance to improve the loyalty of the less loyal group of customers, by motivating them to purchase more by directing more of the highly preferred product to them according to their preferences. This project provides a proposed machine learning solution to customer segmentation and prediction using K-mean clustering, Decision Tree and Random Forest classification algorithms. It shows improvements compared to other classification models. The pre-processing of the dataset using data engineering methods significantly improves customer prediction accuracy after segmenting them into three clusters. The proposed method shows that the learning representations using K-mean for segmentation increases performance on the customer groups. This proposed method uses a simple K-mean clustering algorithm to evaluate the model on the E-Commerce Data Actual transactions of the UK retailer dataset. The aim is to make important visualisation, segmentation, and prediction of customers. From the results obtained, a more sophisticated model and further training and analysis will further improve the results obtained in this study

Towards a Privacy preserving Machine Learning based Access Control for the Internet of Things

2023-09-04T21:52:29+01:00

The Internet of Things (IoT) is pushing us towards adopting original security approaches to manage the huge amount of data they produce while respecting its computing and storage constraints. For that matter, this paper endorses getting the most out of machine learning (ML) algorithms as a means to build smart and dynamic Access Control (AC) solutions for the IoT. Unfortunately, ML requires abundant amount of data for training, whereas a large part of the data exchanged by IoT devices is private, hence this paper is offering a way out to reconcile these two concepts through inter-organizational differential privacy. The proposed solution is backed up by an implementation and shows promising results compared to popular AC models and standards.

Controlling Driver Behaviour in ADAS with Emotions Recognition System

2023-09-04T22:28:10+01:00

The work focuses on the development of a non-intrusive driver behaviour tracking system. This system can be used in Advanced Driver Assistance Systems (ADAS). The principle of the developed algorithm is to apply the ViolaJones method to determine human emotions. By capturing an image and performing its integral representation, the method performs video stream processing based on Haar-like features designed to find an identifiable object. By applying AdaBoost machine learning method, which sifts out low-informative features and creates strong classifiers from a set of weak ones, the algorithm recognises the driver's face, estimates his age and emotions in real-time and sends a warning signal to the user based on the received data. The paper demonstrates the effectiveness of the ViolaJones algorithm to develop a computer vision system for identifying and tracking a person's state and emotions. Combining the results, an ADAS-like system for tracking a driver while driving has been developed. Through simulations, the high accuracy and speed of the algorithm for determining the emotion and age of the driver has been demonstrated. The synthesised algorithm can be used to analyse driver behaviour while driving.

Skin Lesions Segmentation and Classification for Medical Diagnosis

2023-09-04T18:01:51+01:00

Classification and segmentation of various skin lesions play a very important role in the field of dermoscopy. Using computer-aided applications to detect cancerous cells and predict the lesion as benign and malignant can yield better results. Automatic estimation of skin disease from skin lesion images help practitioners to perform rapid diagnosis, provide early treatment and quick decision making. In this paper, Convolution Neural Network (CNN) is used to identify cancerprone skin lesions from dermoscopy images. Experiments were performed on ISIC 2016 data set with two lesion classes (Malignant and Benign). The training was carried out with the Multiple Residual Neural Network (ResNet) architecture, where the data is pre-processed with different methods. Finally, the comparative analysis with other methods was also performed. The results indicated that the performance of our proposed method is also in line with state of the art methods.

Predict the Match Outcome in Cricket Matches Using Machine Learning

2023-09-04T22:18:35+01:00

In sports, cricket is a very famous game that being watched everywhere in the world. It has tremendous spectators everywhere in the world. One-day International (ODI) cricket matches are popular globally. The impact of ODI is increasing financial and sponsorship every year. Betters and bookmakers are interested in match outcomes. Our focus is to predict the ODI Matches. ODI matches have several features like pregame, in-game, and post-game that affect the match outcomes: the venue, pitch condition, toss winner, toss decision, ground, host country, and dynamic strategies. We use 2000 to 2019 ODI matches dataset. The dataset introduced some new features like Pitch Condition, Venue Familiarity, Winner Inning, Inning First, Inning Second. This study investigates machine learning classifiers to predict the ODI match outcomes. We applied machine learning algorithms such as Naïve Bayes, SVM, KNN, Random Forest, and Decision tree to generate predictive models. The experiment demonstrates that the model performs well in both instances, Result, and Winner. In the decision tree and Random Forest result accuracy up to 98 %, SVM, Naïve Bayes, and Logistic Regression accuracy up to 87%,65%, and 67% respectively. It shows that the Random Forest and decision tree perform well in accuracy, precision, and recall compared to other Statistical Models. This predictive model will help the Gambling Industry, Sports Analysts, Reporting Media, and Cricketing boards to analyze the team performance.

Remedial Directed Topic Map on Personalized Scaffolding Adaptive Learning Management System

2023-09-04T22:07:04+01:00

The problem arises behind the mastery learning model, that is some students fail to achieve the minimum standard of completeness. The existing solutions have been ineffective in helping the study plan. The existing remedial method does not consider the previous formative test answers, but all material that has been mastered must also be tested. This remedial method did not give students more time to learn misconceptions or the material they had not mastered yet. We propose a framework for building a scaffold on a remedial learning path that is personal and adaptive to the material needs that students must learn. The results of exploration in this study are possible to implement the proposed framework by utilizing machine learning using mastery module to process the input of students' answers, domain expert's topic maps, and the relationship among answers to build a computer-based scaffold. It could be an effective remedial path for the failed student to learn just according to topics have not mastered.

Improving Sentiment Analysis Based on Gated Recurrent Unit Model by Using Feature Selection

2023-09-04T21:57:13+01:00

In recent years, sentiment analysis has received a great deal of research due to the dramatic growth in online social media use. For learning-based methods, many technically advanced strategies have been used to improve the performance of the forms. Feature selection is one of these strategies and has been studied by many researchers. However, one unresolved problem is choosing the right number of features to get the best sentiment analysis performance from learning-based methods.
Therefore, we examine the relationship between the number of features selected and the sentiment analysis performance of learning-based methods, aiming to reduce the complexity time and increase accuracy in the experimental data. In this paper, we worked on a dataset consisting of tweets for six major US Airlines and analyses the impact in sentiment analysis of combined methods of Feature Selection based on chisquare correlation with the GRU network. Finally, GRU with feature selection method scored the highest performance in all metrics and achieved an accuracy of 96.38%.

Developing a Protective – Preventive and Machine Learning Based Model on Child Abuse

2023-09-04T22:35:01+01:00

Online grooming is an ever-increasing problem in societies and the time spent online is recently started to rise drastically. People can become anonymous whilst posting, sharing his/her own opinion, and being a part of online chatting. Option to be anonymous also brings together the chance for hiding personal identity when making an attempt on illegal activities. Online grooming is one of the significant areas of aforementioned actions and sexual predators can easily use online chatting platforms to quickly build a friendly relationship with children or teenagers to gain their trust and make them share their obscene media files. These sexual predators mostly try to convince their victims to meet and it may lead to having sexual intercourse with a minor. In order to draw attention to the huge challenge that most societies face, this study mainly aims to identify predators in the early stage of online communication. The objective is to do an investigation to detect child grooming through online chat records by using Machine Learning techniques.
In the first part of the study, it has been achieved to make a multi-label classification on a Wikipedia dataset with more than 97 percent accuracy, where a given text gets classified based on the toxicity types. The outcome of this work is also used in the second stage and herein PAN12 dataset has been used to train and test our model.
We have ended up with more than 92 percent accuracy, where suspicious conversation messages from the chat records get identified and sexual predators can be recognized.

Evaluating the Performance of Agents for Deceptive Levels

2023-09-04T18:19:32+01:00

Deceptive levels are levels where agents are trapped in the game and the policy leads them away from the global optimum policy. In this paper, we examine the problem of evaluating deceptive levels. In our previous work, we generated deceptive levels and evaluated those using different agents. The interesting insight from that study was that there was no single agent being able to solve all types of deceptions. In order to further extend our work, we investigate the effect of different metrics on the evaluation of deceptive levels. In our current work, we use different metrics including reward penalization, discounted reward, noisy world, broken world, and broken forward model. Initially, we execute these agents by utilizing deceptive as well as non-deceptive levels to perform a comparison on the basis of total points and win-rate. After comparison, ranks are assigned to the agents. Contrary to our previous evaluation, these metrics give different results and rankings. The experiment implies that each agent has a different set of characteristics and weaknesses.

Deep-Immune-Network Model for Vulnerable Clone Detection

2023-09-04T22:22:40+01:00

Code-clone is primarily utilized for the implementation of the software reusability concept. The behavior of code reuse generally leads to the propagation of vulnerabilities in the case of the reproduction of a piece of a vulnerable code. Such a vulnerability is called a code-clone vulnerability. This paper is the first work that proposes a novel methodology, which is capable of discovering vulnerable codeclones in open-source software programs efficiently and accurately. We use the immense potential of immune network models for understanding robust code-clone detection from vulnerable software metrics. Therefore, we propose the first deepimmune-network model that learns which software metrics are useful to discover vulnerable code-clones. We use the immune network model with two deep neural network models to discover vulnerable code-clones. To the best of our knowledge, we apply the immune-network model to the domain of vulnerable code-clone discovery for the first time. We introduce an empirical assessment of our approach and examine its usefulness in practice in five different open-source software projects.
We have concluded that the suggested approach is both reasonable and valuable for the detection of the software vulnerability of code-clones.

Self HARMing Malware - An Adaptive IoT Honeypot for Automated, Repetitive Malware

2023-09-04T22:12:12+01:00

Internet-of-Things (IoT) contains physically constrained devices which impacts on the security of IoT deployments. This can make them vulnerable to dedicated IoT attack software such as Mirai and increases the attack surface available to IoT malware developers. This article describes the functionality, deployment and assessment of an adaptive IoT Honeypot for Automated, Repetitive Malware (HARM). It presents findings from the live deployment of HARM that uses reinforcement learning to learn from attack interactions. It also demonstrates that dedicated IoT honeypot datasets contain attack information that is repetitive and automated. Standard honeypot deployments have scripted responses which can terminate attack interactions. This article enhances standard honeypots by exploiting malware characteristics of automation and repetition. It identifies that the learning evolution of HARM correlates with the discovery of the entire attack sequence. It also uses the captured dataset to assess reinforcement learning policies so as to optimise further HARM deployments.

Design & Development Of Mapping Tool For The Blind Or Partially Sighted

2022-05-22T15:47:03+01:00

People who have visual impairment face many problems doing their simple daily activities without needing the help of other people or specially trained dogs. This study intends to design and develop an optimal solution. The solution will be based on artificial intelligence technologies such as computer vision, intelligent text to voice conversion, detecting objects and/or obstacles, and conveying the information verbally. This will offer modern assistance techniques such as smart glasses and canes, helping the needy navigate indoor and outdoor. The solution will also provide a book reading capability. In this study, a system offering indoor and outdoor object detection, obstacle detection and real-time text recognition and conversion to audio will provide users with information to help them overcome navigation and detection problems. A comprehensive test will be carried out to check the ability of the system to achieve its designated task effectively, environmental test conditions, including tool performance under normal daylight, high or bright light, and total darkness. The text to voice ability was tested, in which case the system detected video subtitles and text from images equally well. The results indicate that the system performs equally well in all cases.

Development of a weighted ensemble approach for prediction of blood glucose levels

2022-05-22T15:57:23+01:00

Type 1 diabetes (T1D) is a serious global problem. Several differential equation models have been proposed to model the blood glucose levels (BGLs) in the human body. However, a comparison of prediction of BGLs via non-recurrent models (multilayer perceptron or MLP), recurrent models (long short-term memory or LSTM), and statistical models (vector auto-regression or VAR) has received relatively little attention. This research's primary objective is to address this literature gap and propose BGL predictions via non-recurrent, recurrent, and statistical models.
A T1D dataset was compiled from a differential equation-based mathematical diabetes simulator, Automated Insulin Dosage Advisor (AIDA). A non-recurrent model (MLP), a recurrent model (LSTM), a statistical model (VAR), and a weighted ensemble model of the individual models were developed. In all models, the timeseries BGLs were divided into 80% training and 20% test. Results revealed that the recurrent LSTM model outperformed the non-recurrent MLP and statistical VAR models during training and test to account for the BGLs from the AIDA model. Furthermore, the weighted ensemble model performed the best to account for the BGLs from the AIDA model among all models. Overall, machine learning models may provide an alternative to differential equation models for simulating healthcare variables. We illustrate the significance of using ensemble models for blood glucose predictions.