University of Ibadan Journal of Science and Logics in ICT Research

Design and Implementation of a Bonus-Malus System for Vehicle Insurance in Nigeria

A. U. Rufai ; O. B. Okunoye , E. Imohi — Mon, 07 Sep 2026 00:00:00 +0000

Insurance companies often categorize risks based on observable factors when determining premium rates. However, many unobservable factors can also impact risk. In a competitive market, it's challenging to cross-subsidize different risk categories. To enhance profitability and growth, insurance companies must prioritize efficient pricing models. Experience rating, also known as No Claim Discount or Bonus-Malus Systems, adjusts premiums based on claims history. This study employed a risk-based adjustment model that incorporates decisions about costs fairly and equitably based on individual characteristics. Data on risk criteria, claims impact, and placement decisions for motor insurance liability portfolios in Nigeria were collected and analysed using a 5-by-5 matrix computational approach. The study concluded that motor insurance risks are influenced by individual risk criteria and that a risk-based adjustment pricing approach is essential for fair and equitable cost allocation among insured individuals.

Development of an Adaptive Fingerprint Multi-Filter Enhancement Technique for Improved Reliability in Biometric Systems

Rukayat Afolake Olaniran; Olushola David Adeniji , Toyin Oguntunde — Mon, 07 Sep 2026 00:00:00 +0000

Fingerprint identification system is among the most dependable biometric identifiers due to its uniqueness and accessibility, but distorted fingerprint images usually reduce recognition accuracy. Existing multi-filter enhancement techniques improve ridge quality through a fixed sequence approach which usually leads to overenhancement and computational complexity. This study proposed an optimized multi-filter technique based on Fingerprint Noise Detection and Adaptive Removal (FINDAR) approach which detects noise patterns and applies adaptive de-noising techniques with reduced runtime and memory usage. FINDAR was implemented in MATLAB using 6000 real fingerprint images from Sokoto Coventry Fingerprint (SOCOFing) repository. Contrast enhancement and segmentation technique were employed as pre-processing steps to increase the visibility of the fingerprint images and to separate the foreground from the background, respectively. The developed model was evaluated solely using multi-noise detection accuracy and against Coherence Diffusion+Log-Gabor filter, Finger U-Net, Radial Hilbert Transform Edge Enhancement (RHLT) and Gabor+HE+CLAHE Multistage Enhancement Technique using Structural Similarity Index (SSIM), Natural Image Quality Evaluator (NIQE), Entropy, Execution time and Memory Usage, as metrics. FINDAR achieved high multi-noise detection accuracy average of 91.7% and also outperformed the benchmarked models with the highest SSIM (0.93), lowest NIQE (3.11) and highest entropy (7.02). FINDAR also demonstrated competitive runtime efficiency (17.71ms) and moderate CPU usage (295MB) compared to others. It was also observed that Finger U-Net that used deep learning approach recorded fastest runtime (13.9ms) but consumed significantly higher memory (970MB). These outcomes showed how effectively FINDAR addressed the issue of poor-quality fingerprint photos with minimal processing resources for practical biometric applications.

Ensemble Machine Learning Approach to Multimodal Deception Detection

Adetoye Oluwatoyin ADEDOKUN; Adebola K. OJO — Mon, 07 Sep 2026 00:00:00 +0000

Demand for reliable deception detection systems has stimulated the application of machine learning techniques to behavioural analysis. Conventional approaches either rely on single or multiple modalities, including language, facial expression, or gestures. The single modality may not adequately capture the complex nature of deceptive behaviour, while the multimodal approaches also suffer from inaccuracies due the inability of a single algorithms to be able to effectively capture different aspects of the human behavioural indicators. To address thislimitation, an ensemble machine learning framework that integrates linguistic and facial cues for multimodal deception detection was developed. The framework combines psycholinguistic features extracted from real life dataset obtained from YouTube, using facial cues derived from the Facial Action Coding System (FACS) and linguistic cues gotten from the Linguistic Inquiry and Word Count (LIWC). An ensemble machine learning approach was employed, using the Random Forest classifier to learn facial deceptive traits, Support Vector Machine (SVM) to learn deceptive linguistic patterns, and fusing the predicted outputs by using the Extreme Gradient Boosting (XGBoost) asthe meta-classifier for producing the final classification decision. Experimental evaluation showed that the ensemble multimodal framework outperformed the individual classifiers across multiple performance metrics. The proposed system achieved an overall classification accuracy of 89.8%, an F1- score of 89.8%, and an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.93, indicating strong discriminative capability between deceptive and truthful instances. This study affirms the effectiveness of ensemble strategies for improving deception detection accuracy by using modality-specific classifiers in multimodal feature analysis

Internet of Things-Based Soil Nutrient Monitoring Systems for Crop Yield Prediction: A PRISMA Systematic Review of Methodologies, Machine Learning Models, and Research Gaps

Mon, 07 Sep 2026 00:00:00 +0000

Global food production must rise by at least 60% by 2050, yet the inputs needed to reach that target risk accelerating the soil degradation that limits yields in the first place. The Internet of Things (IoT) and Machine Learning (ML) are increasingly used to monitor soil health and forecast crop performance in response to this pressure. Despite growing research in this area, no comprehensive synthesis has yet mapped the common methodologies in use, and it remains unclear how consistently different ML models perform once deployed across varied farming environments — a gap this review addresses. Using a PRISMA-compliant systematic review of 87 peer-reviewed studies published between 2017 and 2026, this paper examines how IoT sensor networks, wireless communication protocols, cloud architectures, and ML models are being combined to support precision agriculture. This review covers sensor deployment techniques, wireless protocols (WiFi, LoRaWAN, and NB- IoT), and ten standard ML models for multiple crop types, with the random forest regression (RFR) showing the best performance with an R2 of 0.97 across the datasets considered. The review also covers how explainable AI techniques can help farmers to comprehend and have confidence in results from model. Three common gaps are identified: a lack of interdisciplinary validation, a lack of studies from Sub-Saharan Africa, and a lack of a standardized approach to IoT sensor calibration. These findings offer a reference point for researchers, policymakers, and developers of future smart-farming systems.

Development of a Smart Campus Navigation System with Real-Time Guidance and Accessibility Features

Sunday. J Agbolade — Mon, 07 Sep 2026 00:00:00 +0000

Navigating your way around a large campus can be difficult, especially for newcomers. This study introduces a
Smart Campus Navigation System, a web-based platform designed to make campus travel easier with real-time
navigation, QR code indoor guidance, virtual tours, and a user-friendly interface. The system was built using Agile
methods, which allowed for ongoing development, regular testing, and user feedback at every stage. ReactJS was
used for the frontend, and Django powered the backend. Navigation functionalities were implemented using
mapping technologies, including Google Maps API and Leaflet.js. Usability testing took place in real campus
settings to assess the system’s performance, accessibility, and user satisfaction. Findings show that the system
significantly enhances navigation efficiency and overall user experience. Users frequently highlighted its intuitive
interface, precise location searches, and smooth navigation via QR codes. The solution effectively addresses
typical challenges associated with finding one’s way around university campuses. Potential future improvements
could involve integrating indoor positioning, adding voice-assisted navigation, and scaling the system for use
across various educational institutions.

Development of an Artificial Intelligence Chatbot System for Depression

Sunday. J. Agbolade — Mon, 07 Sep 2026 00:00:00 +0000

The growing global burden of mental health disorders, particularly depression, has revealed significant gaps in the
accessibility, affordability, and scalability of traditional therapeutic services. In response, this study presents the
design and implementation of an AI-powered chatbot system aimed at providing mental health support to
individuals experiencing symptoms of depression. The chatbot leverages advancements in artificial intelligence,
natural language processing (NLP), and evidence-based therapeutic practices primarily Cognitive Behavioral
Therapy (CBT), to offer immediate, empathetic, and structured support in a conversational format. The system
was developed using a modular architecture, incorporating a ReactJS-based frontend, a Flask backend interactions,
Rasa for NLP and dialogue management, and MongoDB Atlas for secure data storage. The chatbot was trained
using a preprocessed Kaggle dataset containing anonymized mental health-related user, enabling it to accurately
detect user emotions, classify intents, and generate context-sensitive responses. Its conversational flow was
designed around key CBT principles, guiding users through mood tracking, thought identification, cognitive
restructuring, and the application of coping strategies.

Predicting Information Technology Career Paths Using Stacking Ensemble Learning Model

A. A. Udosen ; A. Okoro , T. Oduba , S. Awodele , S. Alimi , D. Nteziryayo — Mon, 07 Sep 2026 00:00:00 +0000

Selecting a profession route in the fast-changing Information Technology (IT) industry is now an even more complicated issue among students and job applicants, and this has resulted to job dissatisfaction and skill gaps. The conventional counselling practices are often based/ on subjectivity and are not capable of examining the large number of modern tech positions and the individual ability of a person. This paper outlines the creation of an IT Career Prediction System that was developed to deliver recommendations based on personal data, using technical skills, including Database Fundamentals and AI/ML, and psychological traits, including Openness and Conscientiousness. A stacking technique was employed to implement an ensemble machine learning approach in order to achieve high predictive reliability by combining four base classifiers, such as Support Vector Machine (SVM), Decision Tree (DT), Random Forest Classifier (RFC), and Naive Bayes (NB). The developed ensemble learning method combined the computational powers of the four based classifiers used with a performance accuracy of 99.1%, and was effective by applying interpretability and balancing the risk of overfitting when giving practical advice. Hence, the paper concludes that the automated system provides a scalable, objective, and accurate alternative to conventional career guidance, which ends up enhancing more effective career alignment and longterm professional success in the IT industry.

On the Construction of Knowledge Graphs from Unstructured Text using Neuro-Symbolic Networks

E. P. Fasina ; B. A. Sawyerr , A. Murainah — Mon, 07 Sep 2026 00:00:00 +0000

In the past few years, Knowledge Graphs (KGs) have emerged as crucial tools for organizing structured information and building complex queries that support applications such as Websites, search engines, recommendation systems, LLMs, and data integration platforms. For KGs to scale with these applications, domain experts must transition from the current practice of curating KGs manually to generating them automatically from the huge volumes of unstructured text generated daily by different sources on the Internet. However, the difficulty of understanding the nuances of natural language and the need to extract precise relationships between entities in unstructured text are some of the major challenges of automatic KG generation. In this work, this challenge is tackled by combining entity and relationship extraction with neural networks (NNs) and the coherent representation of knowledge using symbolic reasoning techniques. An NN combined with a symbolic reasoning module can then be used for the accurate generation of KGs from a variety of unstructured text sources. Our methodology involves designing a generic neuro-symbolic network (GNSN) for KG extraction from a given text and then implementing a variant using open-source libraries. The implemented GNSN variant is trained and evaluated on WebNLG+2020 Text-to-RDF benchmark with a measured precision of 82.2%, recall of 82.4%, and an F1 score of 82.3%, which is competitive with other neuro-symbolic networks reported in the literature. This work aims to advance the state of the art in automatic KG generation from unstructured text

Punitive-IDS: A Reputation-Based Punishment Framework for Enhancing Network Performance in Malicious Node Management for MANETs

Mon, 07 Sep 2026 00:00:00 +0000

Malicious node activity in Mobile Ad-hoc Networks (MANETs) degrades not only security integrity but also
fundamental network performance indicators, including throughput, packet delivery ratio (PDR), and end-to-end
delay. Existing intrusion detection systems predominantly optimise classification accuracy while neglecting the
network-level consequences of detection decisions, particularly the disruptive effects of false positives on routing
availability and connectivity. This paper presents Punitive-Intrusion Detection System (IDS), a reputation-based
punishment framework for an Intrusion Detection System (IDS) that integrates a Q-Learning-enhanced Artificial
Neural Network Intrusion Detection System (ANN-IDS) with a Random Forest classifier and a severity-aware
punitive decision engine. Detected malicious nodes receive graded reputation score reductions and are excluded
from routing decisions, with reinstatement permitted upon subsequent behavioural improvement. Evaluation
across four attack scenarios and eight node density configurations (25 to 200 nodes) using NS-3 simulation
demonstrates that Punitive-IDS achieves a throughput of 779.4 kbps, packet delivery ratio of 96.7%, and end-toend delay of 118.2 ms under active attack conditions, representing improvements of 27.3%, 40.3%, and 44.9%
respectively over an unprotected MANET baseline. Comparative analysis confirms that Punitive-IDS outperforms
Static ANN-IDS with Random Forest, SVM-based, and KNN-based intrusion detection frameworks across all
network-level metrics, while simultaneously achieving a superior classification accuracy of 99.21%.

Q-Learning-Driven Adaptive Intrusion Detection for Malicious Node Classification in Mobile Ad-Hoc Networks

Mon, 07 Sep 2026 00:00:00 +0000

Mobile Ad-hoc Networks (MANETs) remain highly exposed to routing-layer attacks such as Sybil, Wormhole, Blackhole, and Denial-of-Service (DoS) due to their decentralised structure and constantly shifting topology. Intrusion detection systems built on static machine learning models struggle to keep pace with these shifting attack patterns, leading to a growing number of missed detections as network conditions change. This work proposes an adaptive intrusion detection framework that embeds Q-Learning reinforcement within an Artificial Neural Network-based Intrusion Detection System (ANN-IDS), with classification finalised by a Random Forest ensemble. The Q-Learning agent refines its detection policy through a composite reward signal that jointly rewards true positive rate and packet delivery ratio while penalising false positives and end-to-end delay. Tested across four attack types (Sybil, Wormhole, Blackhole, and DoS) and eight node density settings (25 to 200 nodes) in NS-3, the framework attains an overall accuracy of 99.21%, precision of 98.51%, recall of 99.68%, and an F1-score of 99.09%. The agent's detection policy stabilised within roughly 140 training episodes and sustained this performance at the 200-node scale. Across every evaluated condition, the proposed approach surpasses Static ANN-IDS with Random Forest, Support Vector Machine, K-Nearest Neighbour, and an unprotected MANET baseline, supporting its use in large, dynamic MANET deployments.

Multi-agent Deep Reinforcement Learning for Energy-Efficient Cluster Head Selection in IoT-Based Wireless Sensor Networks

Ademola A. Omilabu — Mon, 07 Sep 2026 00:00:00 +0000

To monitor large physical spaces, modern industrial and urban Internet of Things (IoT) architectures depend on data collected by Wireless Sensor Networks (WSNs). Crucially, battery life remains the absolute limiting factor. Because these tiny sensor nodes run on limited, non-rechargeable power, keeping the network alive for long periods is incredibly difficult. To stretch these limited power reserves, current systems group nodes into clusters. These protocols group the nodes together so that a single coordinator—the Cluster Head (CH)—collects all local data and routes it to the main base station. However, static heuristics and centralized systems are simply too rigid. They fail completely if data traffic surges or if the physical network layout changes unexpectedly. To bridge this gap, we designed a decentralized Multi-Agent Deep Reinforcement Learning (MADRL) framework specifically for selecting CHs efficiently. Instead of relying on a central controller, we treat each sensor node as an independent agent. These agents work together to cut down overall energy loss while keeping network coverage wide. We built a Deep Q-Network (DQN) directly into each agent, allowing them to learn smart clustering habits on the fly. The nodes make these choices by tracking their own remaining power, how close they are to neighbors, and their distance to the base station. Our simulations show excellent results. This decentralized MADRL setup drastically lowers energy use, keeps individual nodes alive longer, and extends network lifespan far better than traditional LEACH or genetic algorithms.

Intelligent Traffic Congestion Control in MANETs: A Comparative Analysis of Neuro-Fuzzy and Metaheuristic-Enhanced Routing Strategies

A. O Adepoju ; M. O Oyediran , O. S. Ojo , O. E Makinde , F. W. Ipeayeda — Mon, 07 Sep 2026 00:00:00 +0000

Mobile Ad Hoc Networks (MANETs) are well known to suffer from congestion due to their decentralized architecture, variable topology, and limited resource availability. The conventional shortest-path routing protocols were typically found to be the least efficient for packet delivery, and caused a slower latency when a great deal of traffic is occurring due to the fact that they have basically just used hop count in deciding on routing without taking into account the current state of the network. Although Adaptive Neuro-Fuzzy Inference Systems (ANFIS) are used to predict congestion and improve routing strategies, it is still reliant on how membership function parameters are set. We compare three different routing methods in this study: the Shortest Path First (SPF) baseline model, the conventional ANFIS, and an advanced form of ANFIS with the Firefly Algorithm (Firefly-ANFIS) designed for congestion control in MANETs. We build a simulation environment in MATLAB with 25 mobile nodes, measured with a dataset of traffic records of 1,200 instances. Assessment metrics were congestion index, packet success rate, end-to-end latency and hop count as well as traffic overhead statistics and transmission time. The findings show that Firefly-ANFIS led to better results on all metrics, including the lowest average congestion index (0.043), minimal average end-to-end latency (0.1040 s), an average hop count as low as 2.45, and the lowest traffic overhead compared to SPF baseline and standard ANFIS models. Despite a significantly improved packet success rate of 92.76% with standard ANFIS, it was also faced with increased worst-case latencies associated with unnecessary route detours in bursty conditions. Firefly techniques provided significant mitigation through optimized tuning processes for membership function parameters, allowing for more standardized routing choices to be delivered to all member routing types. The conclusions of the present study show that combining metaheuristic optimization methods with neuro-fuzzy methods for predicting congestion can provide a better balance between reliability gains and latency reduction using dynamically influenced systems typical of MANETs.

Federated Explainable Artificial Intelligence for Early Diabetes Prediction Across Multi-Hospital Data

Mon, 07 Sep 2026 00:00:00 +0000

In healthcare system, detection of disease earlier helps in patient health management. The same is applicable to diabetes mellitus management and treatment. Centralized data is where conventional machine depends on, but the centralized data has issues related to privacy and security in the healthcare system. Federated Explainable Artificial Intelligence (FedXAI) framework in early diabetes prediction across multiple hospital settings is the focus of this work. This work, proposed using federated learning to collaboratively train model, with restrictions to sensitive patient data. The integration of SHapley Additive explanations (SHAP) to the system is for transparency enhancement and clinical trust. SHAP integration is insightful in the provision of interpretability in the model predictions. The Pima Indians Diabetes dataset helps stimulate distributed hospital settings for independent and non-independent (nonIID) data distribution. The outcome of the work indicated federated model is of higher competitiveness in comparison to the centralized approach. There is a decrease in accuracy with heterogeneous distributions. The glucose level, body mass index (BMI), and age remains the major predictor of diabetes as regards to the explainability analysis. Practically, the proposed system enhances privacy protection and preservation, solves interpretable issues in application for real- world healthcare deployments.

Design and Experimental Evaluation of an IoT-Integrated Smart Lighting System with Adaptive Control and Wireless Power Transfer

A.D. Omiyale; O. F. Odeyinka , M. A. Adewuyi , G. A. Omiyale — Mon, 07 Sep 2026 00:00:00 +0000

Energy usage connected with conventional lighting systems is a critical problem in residential and commercial
spaces. The purpose of this paper is to develop and experimentally evaluate an Internet of Things (IoT)-based
smart lighting system that incorporates adaptive illumination control, Bluetooth Low Energy (BLE)
communication, and resonant wireless power transfer in one system. For experimental testing, the developed
system was built using an ESP32 microcontroller, a TSL2561 digital light sensor, an NRF52832 BLE module, and
a power switch circuit that operates with the use of MOSFET. During experiments, energy efficiency, control
responsiveness, and wireless power transfer performance were estimated. The developed smart lighting system
provided 30% energy savings compared to a conventional relay-based lighting system. BLE commands' processing
was characterised by 0.5s of average response time and a 95% success rate within 10 m range. Additionally, the
wireless power transfer subsystem ensured an energy transfer efficiency of approximately 85% at a distance of 10
cm.

Lightweight CNN-RL Adaptive Feedback for Gamified Learning Platforms: A Review

A. A Udosen ; M. Eze , O. Ebiesuwa , O. Akande , S. O. Kuyoro , E. E. Onuiri — Mon, 07 Sep 2026 00:00:00 +0000

The rapid increase in online and blended learning has added pressure on the necessity to have adaptive, engaging,
and efficient educational technologies. This paper provides a review and syntheses of the current developments
in lightweight convolutional neural networks (CNNs) and reinforcement learning (RL) to support the
establishment of real-time, personalised feedback in a gamified learning setting. The paper presents the theoretical
premises, system designs, and experimental findings of the recent designs, and emphasises the potential and
difficulties of implementing these systems to resource-constrained platforms. As we have shown in the analysis,
the combination of lightweight CNNs with RL agents can allow adaptive learning to become scaled, responsive,
and efficient, and become the foundation of the intelligent education platform of the next generation.