Machine Learning & Artificial Intelligence

Machine Learning for Social Internet of Things

Machine learning has become widely used in a vast array in everyday applications. In this research, we utilize Machine Learning techniques to design predictive analytics’ techniques for Social Internet of Things (SIoT) networks and make well-informed decisions accordingly. For example, data analytics can help process, understand, and enhance the data generated by the devices. Internet-of-things (IoT) becomes essential in a variety of civil, public, and military applications, which makes their complexity and size perpetually increasing. The growing number of connected devices requires advanced forms of collaboration to exploit their heterogeneity and improve their services effectively. The Social Internet-of-things (SIoT) concept has been emerged by allowing IoT devices to establish their own social networks. The paradigm aims to aid the smart objects to establish and maintain relations with their peers. The relationships in the network are not exclusive to machine-to-machine but can be extended between the users of the SIoT system, such as machine-to-human or even human-to-human relations. The social relations help assure trustworthiness between devices as the basis to share resources or collaborate on different services such as the share of computational needs. In fact, the relations between IoT devices may reflect their ownership, location, or past collaboration. However, understanding the structure of such complex and ubiquitous networks composed of diverse communicating nodes remains a challenging task. Novel data and graph analysis techniques can constitute an appealing solution to discern the SIoT network patterns and correlations among IoT devices.  It can also help understand the structure of IoT systems using unsupervised machine learning approaches such as classification and clustering methods to group IoT 1) infrastructures, e.g., by clustering devices to reduce the complexity of the vast IoT network or 2) services, e.g., by assigning IoT devices to tasks/services.

In this research, we develop a novel clustering approach based on Graph Neural Network (GNN), a deep learning algorithm, to discern SIoT structure. We aim to embed the features of devices as well as their connections from a real-world dataset using GNN and then apply an unsupervised learning algorithm to determine clusters of IoT devices sharing strong social relations. Results of GNN-based clustering are compared to the deterministic community detection approach. In the Figure, we illustrate a general SIoT data analytic framework where graph analysis and machine learning techniques are used to perceive the structure of SIoT system and the relations among its nodes. The IoT devices connect through a cloud gateway to exchange necessary data such as the location and specification of the devices. From this information and other IoT devices’ features, graphs modeling the different social relations between the devices can be established. GNN and unsupervised learning techniques are then employed to determine the clusters of devices sharing strong social relations, which can help better understand the structure of the network and use this extra level of knowledge for more effective service discovery or mobile crowdsourcing tasks. The developed clustering algorithm can outperform other community detection methods for certain metrics such as modularity and coverage. These promising results are encouraging to further examine similar and more advanced machine learners.

Machine Learning for IOT Service Discovery

Efficiently utilizing the vast network of devices in the Internet-of-Things (IoT) system to leverage computational resources can be beneficial for edge computing services. The edge computing affirms to bring the available distributed, but yet closer resources, to the devices requesting external computational and/or storage capabilities. In many cases, the resource sharing and computing capabilities are indispensable in the IoT system, considering many of terminal devices such as sensors, mobile phones, actuators, and personal computers that may lack these resources to accomplish specific tasks. Devices, in the neighborhood, can share their available computational resources to process tasks for the profit of their peers and help relieve the load of cloud and edge servers. Service discovery in IoT platforms can be used to seek for edge computing providers. Despite that, searching in the vast network to find suitable devices remains one of the major challenges in ubiquitous IoT networks, especially when devices are heterogeneous and require a variety of services with different levels of storage and computational capabilities at irregular time instants. To achieve a productive search of mobile edge computers in the large-scale IoT, it is required to find available, trustworthy, and reliable devices that can potentially handle the targeted computational tasks. The emerging concept of Social IoT (SIoT) can help achieve these goals using the social relations built among the devices, aka} social objects. These social relations transform the IoT system into a social network of devices or “friends” having common characteristics and criteria, which can raise the level of security and trustworthiness in such diverse networks. In this project, we develop a generic framework aiming at selecting appropriate edge computers that can handle the computational task of its peer in a trustworthy and rapid manner. To this end, we proceed with the two following phases: implement a community detection algorithm to divide the IoT network into multiple groups of IoT devices sharing strong social relations, then,  run a machine learner, trained on previous resource sharing activities as well as the computational and non-computational features of the devices and the task to be executed itself, to predict the time to process the task. Starting from a request of an IoT device, the proposed framework outputs an edge computer socially connected to the requester and capable of rapidly executing the task. The proposed solution reduces the complexity of the service discovery task by shrinking the search space and applying machine learning techniques that do not require recurrent training, which can be very beneficial for large-scale IoT systems.

Artificial Intelligence for Insurance Industry

Artificial Intelligence (AI) technology has shown remarkable improvement in helping making accurate decisions in many fields such as robotics, computer vision, and medical science. For instance, many AI techniques were designed to help in solving a variety of issues in the insurance industry such as analyzing and processing data, detecting frauds, minimizing risks, and automating the claim process. Such a technology can be also utilized to automate visual inspection and validation in order to help cope with claims leakage. Indeed, insurance companies are deprived of a tremendous amount of financial gain due to claims leakage. Fraudulent claims present a huge and a costly problem for insurance companies, potentially leading to billions of dollars of unnecessary expenses for the industry yearly. Insurance fraudsters will often exaggerate or fabricate situations to provide the basis for fraudulent claims. Insurers have historically relied on mathematicians to measure risk and formulate premium rates for policy underwriting that would ensure rational levels of payouts without endangering the company’s financial health. Traditional insurance fraud detection methods are complex and time-consuming. They mainly depend on expert scrutiny, adjusters, and special investigation services. Added to that, manual detection results in additional costs and inaccurate results. Moreover, late decisions might lead to extra losses for the insurance companies. Therefore, there is a pressing need to devise fast and efficient solutions to build fraud detection, risk measurement, and secure data management solutions that maintain a perfect balance between client personal data preservation, loss prevention savings, and investment of false alert detection. From that perspective, we propose to develop an effective framework for insurance companies to help confront such challenges.

In this research, we employ two machine learning methods to build fraud detection and risk measurement modules to detect, classify fraudulent auto insurance claims, predict suspected customers, and estimate their next claim amount based on their risk levels. The first method is based on a batch learning strategy where the algorithm trains the whole dataset at once. The second method is based on an online learning strategy which dynamically trains, updates, and upgrades the learning weights as new data enters the system, without the need to retrain the whole model from scratch as new information arrives. The obtained results reveal that, when applied to an auto insurance dataset, our proposed machine learner achieves high performance gains compared to other existing learning algorithms. For instance, it reaches 7% higher accuracy compared to decision tree models when detecting fraudulent claims. Moreover, we present a novel framework to detect, locate, and identify damage severity on vehicles using deep learning, convolutional neural networks, and transfer learning techniques. Unlike previous studies, our approach not only detects damages but also identifies their severity levels, localizes, and visualizes them on the vehicle’s images. Numerical results reveal that our transfer learning proposed solution, based on Inception-ResnetV2 pre-trained model followed by a fully connected neural network, achieves higher performances (e.g., a 15% higher precision) in features extraction and damage detection/localization than a state-of-the-art pre-trained model, i.e., VGG16.