Project Marketplace
Project Topics & Materials
Search curated materials across every major department, faculty and institution.
Showing 13 materials
Computer SciencePHISHING WEBSITE DETECTION USING URL AND CONTENT-BASED FEATURE ANALYSIS
Admin
About This Research Topic Phishing remains among the most prevalent and financially damaging cyber-attack categories, exploiting deceptive websites impersonating legitimate services to harvest credentials. According to Anti-Phishing Working Group (APWG) trend reports , phishing consistently ranks among top reported attack vectors, with attackers increasingly leveraging short-lived, rapidly rotating domains that evade reactive defenses. Traditional blocklist-based browser warnings check visited URLs against databases of known malicious sites. While widely deployed, they are inherently reactive, unable to protect against newly registered zero-day phishing sites not yet catalogued. Machine-learning-based detection addresses this by learning characteristic patterns in URL structure, domain properties, and page content to enable proactive classification at point of access. This article presents a complete pipeline combining URL-lexical, domain/host-based and content-based features with explicit ablation quantification, packaged as low-latency detection service. For additional cybersecurity research materials, see ScholarNestHub cybersecurity collection and recent studies on hybrid feature-based phishing detection . Main Abstract Phishing remains one of the most prevalent and financially damaging cyber-attack categories exploiting deceptive websites impersonating legitimate services to harvest credentials and financial information, with industry reports ranking it among most reported attack categories. Blocklist-based defenses flagging known malicious sites remain widely deployed but inherently reactive, unable to protect against newly registered sites not yet catalogued. This study designs, implements and evaluates machine-learning-based phishing website detection system combining URL-lexical, domain/host-based and page-content features enabling proactive classification at point of access. Study adopted Design Science Research methodology combined with CRISP-DM for data-driven components. Combined dataset of 11,430 labelled websites constructed from PhiUSIIL phishing URL dataset and Kaggle-sourced phishing websites dataset incorporating 30 engineered features spanning three categories: URL-lexical (URL length, IP address presence, URL-shortening services, suspicious character counts), domain/host-based (domain age, WHOIS registration length, DGA-like pattern), and content-based (login form presence, ratio of external to internal links, favicon origin, mismatch between visible link text and target). Data cleaned and used to train and compare four models: Logistic Regression, Random Forest, XGBoost, and Multi-Layer Perceptron with feature-category ablation experiments isolating incremental contribution of URL-only, content-only and combined sets. XGBoost trained on combined feature set achieved strongest performance with accuracy 97.8%, precision 97.2%, recall 97.6%, F1-score 97.4%, outperforming URL-only subset (94.1% accuracy) and content-only subset (93.6% accuracy), demonstrating complementary rather than redundant discriminative signal. Trained model packaged as lightweight browser-extension-style detection service exposed via Flask backend evaluating visited page URL and rendered content in real time displaying risk indicator, achieving average end-to-end classification latency 140 milliseconds comfortably within range required for non-intrusive browsing. Study concludes combining URL-lexical and content-based features within gradient-boosted model provides materially more robust phishing detection capability than either category alone, and recommends periodic retraining and integration with live blocklist feeds as complementary safeguards. Keywords: phishing detection, machine learning, URL analysis, content-based features, cybersecurity, XGBoost, ablation study
Computer ScienceExplainable AI Framework for Medical Diagnosis Decision Support
Elijah T
About This Research Topic Artificial intelligence has moved from a research curiosity into a quiet, everyday presence inside clinics and hospitals, supporting decisions on everything from cardiovascular risk to diabetes screening. Yet the models that tend to predict best, gradient-boosted ensembles and deep neural networks in particular, are often the hardest to interpret. A clinician handed a risk score with no supporting rationale is effectively being asked to trust a black box with a patient's welfare, and that is a proposition regulators, professional bodies, and clinicians themselves are increasingly unwilling to accept without question. This tension between predictive accuracy and interpretability sits at the centre of explainable AI (XAI) research in healthcare, and it is the problem this article works through in depth. Drawing on a completed applied research project, the discussion below walks through the design, implementation, and evaluation of an XAI framework for medical diagnosis decision support. Rather than treating explanation as an afterthought bolted onto a finished model, the underlying study places three complementary explanation techniques, SHAP, LIME, and rule extraction, alongside high-performing diagnostic classifiers built for heart disease and diabetes prediction, and then tests those explanations directly with practising clinicians and final-year medical students. That last step is what distinguishes this work from a great deal of published research in the field. Many studies apply an explanation technique to a medical model and treat the mere technical presence of that explanation as evidence of transparency, without ever asking the people who will actually use it whether the explanation makes clinical sense. The sections that follow set out exactly how this study answered that question, what it found, and what it means for anyone building, evaluating, or procuring a clinical decision-support tool. Main Abstract Machine learning models now deliver strong predictive performance across many medical diagnosis tasks, but the models that perform best, particularly ensemble tree-based methods and deep learning architectures, are often opaque, giving little indication of the reasoning behind any single prediction. That opacity is a genuine barrier to clinical adoption. Clinicians carry professional and ethical responsibility for diagnostic decisions, and a growing body of regulatory guidance calls for some form of explanation to accompany automated decision support in healthcare. This study responds to that gap by designing, building, and evaluating an explainable AI framework that pairs high-performing diagnostic classifiers with several complementary post-hoc explanation techniques, and by testing the resulting explanations empirically with clinical volunteers rather than relying on model accuracy alone as a stand-in for clinical usefulness. The project followed Design Science Research (DSR) methodology alongside the Cross-Industry Standard Process for Data Mining (CRISP-DM) for its data-driven components, drawing on two established public clinical datasets, the UCI Heart Disease dataset and the Pima Indians Diabetes dataset, together covering 1,536 patient records after cleaning. Each condition was modelled separately given the differing feature schemas. Logistic Regression, Random Forest, and XGBoost were trained and compared for each condition, and the best-performing model for each was paired with SHAP for global and instance-level feature attribution, LIME for local surrogate-model explanations, and a rule-extraction technique that produced human-readable if-then rules for clearly separable cases. XGBoost delivered the strongest diagnostic performance across both conditions, reaching 88.9% accuracy and an 87.2% F1-score for heart disease prediction, and 84.6% accuracy with a 78.3% F1-score for diabetes prediction, outperforming both Logistic Regression and Random Forest. Twelve clinical volunteers, comprising final-year medical students and practising clinicians, reviewed SHAP, LIME, and rule-based explanations for a shared set of de-identified sample cases, rating each technique on clarity, clinical plausibility, and trustworthiness. SHAP explanations scored highest on clarity and trustworthiness (4.3 and 4.1 out of 5), closely followed by rule-based explanations (4.0 and 4.2), while LIME scored lowest on stability, with several evaluators noting that repeated LIME runs on similar cases sometimes surfaced different top features, a known consequence of its local sampling procedure. The trained models and explanation modules were integrated into a Flask-based clinical decision-support dashboard combining a diagnostic risk score, a SHAP-based feature-contribution chart, and, where applicable, a corresponding rule, with an average combined prediction-and-explanation response time of 0.31 seconds. The study concludes that combining several complementary explanation techniques, empirically validated with clinical end users rather than assumed to work in advance, offers a more clinically grounded route to explainable medical AI than relying on a single technique in isolation.
Computer ScienceMachine Learning Intrusion Detection System: Building a Smarter Network Defence
Elijah T
About This Research Topic Every network defence team eventually runs into the same uncomfortable truth: attackers do not always repeat themselves. A machine learning intrusion detection system is built to deal with exactly that problem. Rather than waiting for a security analyst to write a new rule for every fresh attack pattern, it learns the underlying shape of normal and malicious traffic directly from data, so that it can flag suspicious behaviour even when the exact attack has never been logged before. This article walks through a complete undergraduate research project that puts that idea to the test. It compares classical machine learning, ensemble methods, and a stacked ensemble model across two well-known network security datasets, tests three feature-selection strategies, and wraps the strongest model in a working alert dashboard. If you are exploring similar territory for your own final-year research, our Computer Science project topics library has further reference projects that show how this kind of methodology chapter, results chapter, and system implementation typically come together. Main Abstract Signature-based intrusion detection systems remain useful, but they share one structural weakness: they can only catch what they already recognise. Once an attacker varies their technique even slightly, a purely signature-driven system has no way of raising an alarm, because no matching entry exists in its database. This weakness is what has pushed so much recent research toward machine-learning-based detection, which builds a model of what normal and malicious traffic look like from historical data rather than from a fixed rulebook. This study designs, builds, and tests a machine-learning-based network intrusion detection pipeline, following the Design Science Research approach alongside the CRISP-DM process for the data-driven stages of the work. Two public benchmark datasets anchor the evaluation: NSL-KDD, a cleaned-up successor to the long-standing KDD Cup 1999 dataset, and CICIDS2017, a newer and more realistic dataset built by the Canadian Institute for Cybersecurity that captures a wider spread of modern attack behaviour, including brute-force attempts, denial-of-service traffic, web-based attacks, infiltration, and botnet activity. After cleaning and encoding the data, three feature-selection techniques were tested side by side — Recursive Feature Elimination, Mutual Information, and Lasso-based selection — ahead of training four models: Naive Bayes, Random Forest, XGBoost, and a stacked ensemble that blends Random Forest, XGBoost, and Extra-Trees under a Logistic Regression meta-model. Both a binary classification task (normal versus attack) and a multi-class task (identifying the specific attack type) were evaluated. The stacked ensemble, paired with Recursive Feature Elimination, came out on top across both datasets, reaching 99.6% accuracy and a 99.4% F1-score on CICIDS2017, and 99.1% accuracy with a 98.7% F1-score on NSL-KDD. It consistently outperformed the standalone Random Forest, XGBoost, and Naive Bayes models. The multi-class results told a more nuanced story: overall performance stayed strong, but recall dropped noticeably for rare attack categories such as infiltration and certain web-attack subtypes, a pattern directly tied to how few training examples those classes have relative to normal traffic. To close the loop between research and practice, the trained binary model was deployed behind a Flask-based monitoring dashboard that reads a simulated live traffic feed, scores each flow for intrusion risk, and raises a security alert when something looks malicious — with an average classification latency of just 8 milliseconds per flow. Taken together, the findings support stacked ensembles combined with careful feature selection as a strong, computationally realistic foundation for machine-learning-based intrusion detection, while cautioning that the near-perfect accuracy figures often reported in this field need to be read alongside dataset-specific class imbalance and cross-dataset generalisation, not treated as a promise of identical real-world performance.
Computer ScienceAI-Based Traffic Congestion Prediction System
Elijah T
About This Research Topic Most traffic apps tell you what is happening right now, which is a bit like checking the weather after you are already soaked. This project set out to build something genuinely predictive instead, a system that forecasts congestion up to several hours ahead of time, tested four different modelling approaches against each other, and packaged the winner behind a live dashboard. This article walks through how that system was built, from feature engineering on historical traffic sensor data to a head-to-head comparison of a statistical baseline, Random Forest, LSTM, and a hybrid CNN-LSTM architecture. Readers exploring related technical projects can browse the Computer Science project collection on ScholarNest for comparable studies in machine learning and systems design. What follows covers the background to AI-based traffic forecasting, the specific problem this project addresses, its objectives and research questions, the key technical terms used throughout, and closes with frequently asked questions for students and developers working on similar predictive systems. Main Abstract Traffic congestion imposes substantial economic, environmental, and quality-of-life costs on urban populations, through lost productivity, increased fuel consumption and emissions, and extended commute times, motivating sustained interest in predictive systems capable of forecasting congestion ahead of time to support proactive traffic management and route planning. This study addresses this problem by designing, implementing, and evaluating a machine learning system for short-term traffic congestion prediction that combines historical traffic sensor data with a simulated real-time data feed, moving beyond purely reactive, current-state traffic reporting toward a genuinely predictive capability. The study adopted the Design Science Research methodology combined with the Cross-Industry Standard Process for Data Mining for the data-driven components of the work. It used the publicly available Metro Interstate Traffic Volume dataset, comprising hourly traffic volume readings from a Minneapolis-St Paul interstate corridor spanning several years, combined with corresponding weather and holiday indicator features, from which a congestion-level target variable (Low, Moderate, High) was derived using volume and historical speed-relationship thresholds. Data were cleaned, engineered with cyclical time-of-day and day-of-week features and lagged historical volume features, and used to train and compare four models: a Seasonal ARIMA statistical baseline, Random Forest, a Long Short-Term Memory network, and a hybrid CNN-LSTM architecture combining convolutional feature extraction with recurrent temporal modelling. Models were evaluated on both a regression formulation, predicting continuous traffic volume using RMSE and MAE, and a complementary classification formulation, predicting discrete congestion level using accuracy, precision, recall, and F1-score. The hybrid CNN-LSTM model achieved the best performance on both formulations, with a regression RMSE of 312 vehicles per hour and a classification accuracy of 89.4% (macro F1-score of 87.6%) for one-hour-ahead prediction, outperforming the standalone LSTM (RMSE 356, accuracy 85.1%), Random Forest (RMSE 428, accuracy 79.8%), and SARIMA (RMSE 612, accuracy 68.3%). Prediction accuracy degraded gracefully with increasing forecast horizon, remaining above 80% classification accuracy up to a three-hour-ahead horizon before declining more sharply. The trained model was deployed behind a Flask-based dashboard that ingests a simulated real-time traffic feed, replayed historical data standing in for a live sensor connection, and displays current and predicted congestion levels for the monitored corridor on an interactive map-style interface, with an average end-to-end prediction latency of 45 milliseconds. The study concludes that hybrid CNN-LSTM architectures, informed by both historical patterns and current real-time conditions, offer a practical basis for short-term traffic congestion prediction capable of supporting proactive traffic management, and recommends extension to a road-network-wide, spatially-aware modelling approach as a direction for future work.
Computer ScienceA Facial Recognition Attendance System with Anti-Spoofing Measures
Elijah T
About This Research Topic A facial recognition attendance system sounds foolproof until someone simply holds up a photo to the camera. That's the gap most systems leave open, and it's exactly what this study set out to close — building a system that checks not just who is in front of the camera, but whether they're actually there, live, in the flesh. This piece walks through how a dedicated anti-spoofing stage was built, benchmarked, and combined with face recognition into a working attendance system. Readers interested in how applied machine learning performs on other real-world classification problems may also want to look at our project on predicting hospital readmission rates with machine learning , which covers a different domain but a similarly structured evaluation approach. What follows carries the full research structure — background, problem statement, aim and objectives, research questions, significance, scope, and definitions — rebuilt for a wider readership while preserving the original study's technical focus and reported results. Main Abstract Manual and card- or fingerprint-based attendance systems remain widely used in academic and workplace settings despite well-documented weaknesses: manual roll-call is time-consuming and susceptible to proxy attendance, while card- and fingerprint-based systems, though automated, still permit proxy attendance through credential sharing and raise hygiene concerns in shared-device settings. Facial recognition offers a contactless, difficult-to-share biometric alternative, but a system that only performs identity matching remains vulnerable to presentation attacks, in which an impostor presents a printed photograph, a video replay, or a mask of an enrolled individual in place of their own live face. This study addresses that problem by designing, implementing, and evaluating a facial recognition-based attendance system that incorporates an explicit anti-spoofing (liveness detection) stage, ensuring attendance is recorded only for a live, physically present individual rather than a static or replayed representation. Following the Design Science Research methodology combined with the Cross-Industry Standard Process for Data Mining, a face-recognition enrolment dataset of 40 volunteer individuals (roughly 25 images per individual, captured under varied lighting and pose) was combined with the publicly available CelebA-Spoof dataset for anti-spoofing model training, comprising live and spoof (print, replay, and cut-photo) face images. Face detection used a Multi-task Cascaded Convolutional Network, face recognition embeddings were generated using a pretrained FaceNet model, and identity matching was performed via cosine-similarity comparison against enrolled embeddings. For anti-spoofing, a MobileNetV2-based binary CNN classifier was benchmarked against a classical Local Binary Pattern texture baseline with an SVM classifier, and an eye-blink-based liveness heuristic using the Eye Aspect Ratio. The MobileNetV2 anti-spoofing model achieved the best performance — 97.8% accuracy, 97.2% precision, 98.1% recall, and a 97.6% F1-score on a held-out test set spanning print, replay, and cut-photo attacks — outperforming the LBP+SVM baseline (89.4% accuracy) and the EAR-based blink heuristic (81.7% accuracy, and specifically vulnerable to video replay attacks that include natural blinking). The face-recognition component achieved a rank-1 identification accuracy of 98.5% and a false acceptance rate of 0.6% on the 40-person enrolment set. The combined system, implemented as a desktop/web-hybrid application using OpenCV for camera capture, achieved an average end-to-end attendance-marking time of 1.1 seconds per individual. The study concludes that combining a dedicated CNN-based anti-spoofing stage with embedding-based face recognition substantially improves resistance to common presentation attacks relative to either face recognition alone or simple heuristic liveness checks, and recommends periodic model updates and expansion to 3D-mask attack resistance as directions for future work.
Computer ScienceMachine Learning for Credit Risk Scoring in Microfinance and Fintech Lending
Elijah T
About This Research Topic Most credit scoring was built for people who already have a credit history — which is exactly the problem for the millions of microfinance and fintech borrowers who don't. This piece walks through a machine learning approach built specifically for that gap: blending conventional loan data with alternative signals like mobile-money activity, benchmarking several modelling approaches against each other, and — just as importantly — checking whether the resulting model treats different borrower groups fairly. Readers curious about how machine learning performs on related prediction tasks may also want to look at our project on machine learning algorithms for credit risk prediction , which covers a closely related modelling problem in more depth. What follows carries the full research structure — background, problem statement, aim and objectives, research questions, significance, scope, and definitions — rebuilt for a wider readership while preserving the original study's technical focus and reported results. Main Abstract Access to credit remains a critical enabler of small business growth and household resilience in developing economies, yet microfinance institutions and fintech lenders serving these markets frequently lack the extensive, formal credit history data that traditional credit scoring relies upon in established banking systems — a condition commonly termed the thin-file problem. This study addresses that problem by designing, implementing, and evaluating a machine learning model for credit risk scoring that combines conventional loan-application and repayment-history features with alternative, non-traditional behavioural indicators, while explicitly evaluating model performance and fairness properties relevant to responsible lending. Following the Design Science Research methodology combined with the Cross-Industry Standard Process for Data Mining, the study used two complementary datasets — the Statlog (German Credit) benchmark and a Kaggle-sourced microfinance/small-business loan dataset incorporating mobile-money transaction regularity, utility-payment history, and business-registration status — comprising 9,578 loan records after cleaning and combination. Four models were trained and compared for binary default-risk classification: Logistic Regression, Random Forest, XGBoost, and a feed-forward Artificial Neural Network, evaluated on accuracy, precision, recall, F1-score, and ROC-AUC, with particular attention to recall on the default (minority) class given the asymmetric cost of misclassifying a genuinely high-risk borrower as low-risk. Model interpretability was addressed using SHAP values, and a fairness audit compared false-positive and false-negative rates across gender and business-sector subgroups. XGBoost achieved the best overall performance — 88.7% accuracy, 79.4% default-class recall, 74.1% precision, an F1-score of 76.7%, and a ROC-AUC of 0.91 — outperforming Logistic Regression (61.3% recall), Random Forest (74.8% recall), and the ANN (76.2% recall). SHAP analysis identified prior repayment delinquency, debt-to-income ratio, and mobile-money transaction regularity as the three most influential predictors of default risk, with the alternative mobile-money feature contributing meaningfully alongside conventional financial attributes. The fairness audit found a modest but non-negligible 6.1 percentage-point disparity in false-positive rate between gender subgroups, flagged as a finding requiring further mitigation rather than a settled result. The trained model was deployed behind a Flask-based loan-officer decision-support dashboard returning a risk score, a recommended decision band, and the top SHAP-derived contributing factors per application, with an average scoring response time of 0.09 seconds. The study concludes that gradient-boosted models incorporating alternative behavioural features can meaningfully improve default-risk identification relative to conventional logistic-regression-based scorecards common in microfinance practice, while underscoring that fairness auditing and human-in-the-loop review remain essential complements to model deployment in a lending context with direct financial consequences for applicants.
Computer ScienceAn AI-Powered Chatbot for Student Academic Advising and Course Registration Support
Admin
About This Research Topic Registration week has a familiar rhythm on most university campuses: long queues outside the adviser's office, the same handful of questions repeated dozens of times a day, and students with genuinely complicated situations stuck waiting behind ones who just want to confirm a prerequisite. It is a capacity problem more than a knowledge problem, and it is exactly the kind of bottleneck conversational AI is well suited to relieve. Students exploring a similarly applied AI project can browse ScholarNest's computer science project topics for related ideas in natural language processing and intelligent systems. This article walks through a complete undergraduate research project built around that exact problem: an AI-powered chatbot that handles routine academic advising and course registration queries through natural conversation, and hands off anything genuinely complex to a human adviser. Rather than settling for a single intent classifier, the study benchmarks three approaches — a classical TF-IDF/SVM baseline, a BiLSTM, and a fine-tuned DistilBERT transformer — then wraps the strongest model inside a full dialogue system grounded in a structured course and policy knowledge base, and tests it end-to-end with real users. What follows breaks down the study's background, problem statement, objectives, and scope, for students, researchers, and anyone curious about how conversational AI is being applied to student support. Main Abstract Academic advising and course registration support are essential but resource-intensive services in tertiary institutions, typically requiring students to queue for limited adviser appointments to resolve routine questions about course prerequisites, registration deadlines, credit-load limits, and graduation requirements. This demand is heavily concentrated around the start of each semester, and it frequently overwhelms available advising capacity, leaving many routine student queries unresolved in a timely manner. This study designs, implements, and evaluates an AI-powered chatbot capable of handling common academic advising and course registration queries through natural language conversation, escalating only genuinely complex or policy-ambiguous cases to a human adviser. The work follows a Design Science Research methodology paired with an Agile development approach for the conversational system itself. A corpus of 3,600 utterances, collected through a structured student survey soliciting example questions and synthetically generated paraphrases, was manually labelled across fourteen intent categories — course prerequisite inquiry, registration deadline inquiry, credit-load inquiry, GPA calculation, and adviser escalation, among others — and used to train and compare three intent classification approaches: a TF-IDF plus Support Vector Machine baseline, a Bidirectional LSTM classifier, and a fine-tuned DistilBERT transformer classifier. The chatbot's dialogue manager combines the intent classifier with a slot-filling component for extracting entities such as course codes and semesters, and a rule-based decision engine that queries a structured knowledge base of courses, prerequisites, and registration policies to generate a response. The fine-tuned DistilBERT model achieved the best intent classification performance, with an accuracy of 93.8% and a macro-averaged F1-score of 92.6%, outperforming the BiLSTM (89.1% accuracy) and the SVM baseline (83.4% accuracy). In an end-to-end task-completion evaluation involving 15 test users completing 5 representative advising tasks each, the chatbot achieved a task-completion rate of 86.7%, with most incomplete tasks attributable to queries falling outside the chatbot's trained intent set and correctly escalated to a human adviser. System testing showed an average response time of 0.6 seconds per turn. A System Usability Scale evaluation returned a mean score of 78.4, corresponding to a 'good' usability rating. The study concludes that an intent-classification-driven chatbot, grounded in a structured institutional knowledge base, can meaningfully reduce the routine advising burden on human academic advisers while reliably escalating queries beyond its competence, and recommends integration with a live student information system and periodic retraining on real deployment queries as future work.
Computer SciencePredictive Maintenance Model for Industrial Equipment Using Sensor Data
Admin
About This Research Topic A machine that fails without warning does more than stop production. It forces a scramble for spare parts, idles an entire line, and often costs far more to fix under pressure than it would have under a planned schedule. For decades, manufacturers have managed that risk with two blunt tools: run equipment until it breaks, or service it on a fixed calendar regardless of its actual condition. Predictive maintenance offers a third option, using sensor data and machine learning to flag a failing component before it fails, so intervention happens only when it is genuinely needed. Students exploring a similarly applied machine learning topic can browse ScholarNest's computer science project topics for related ideas in sensor data, classification, and industrial AI. This article walks through a complete undergraduate research project built around that exact problem, tackled from two complementary angles: predicting whether a piece of equipment is about to fail, and estimating how much useful operating life a degrading component has left. The study benchmarks four models for the failure-classification task and two for the remaining-useful-life estimation task, using two independently sourced datasets, then deploys the strongest classifier behind a live monitoring dashboard. What follows breaks down the study's background, problem statement, objectives, and scope, for students, researchers, and anyone curious about how machine learning is being applied to industrial reliability. Main Abstract Unplanned downtime caused by unexpected industrial equipment failure remains one of the most significant sources of lost productivity and maintenance cost in manufacturing environments. That reality has driven a sustained shift away from purely reactive, run-to-failure maintenance and fixed-interval preventive maintenance, toward predictive maintenance, where sensor-derived condition data is used to anticipate impending failure and schedule intervention only when it is genuinely warranted. This study designs, implements, and evaluates a machine learning model for predicting industrial equipment failure from multivariate sensor data, and extends that capability to a remaining-useful-life (RUL) estimation task for a degrading component. The work follows a Design Science Research methodology paired with the CRISP-DM process for its data-driven components. Two complementary datasets were used: the AI4I 2020 Predictive Maintenance dataset, a synthetically generated but operationally realistic set of 10,000 milling-machine operating records with binary failure labels and failure-mode annotations, used for the failure-classification task; and the NASA C-MAPSS turbofan degradation dataset, used for the remaining-useful-life regression task. Data were cleaned and engineered with rolling-window statistical features (mean, standard deviation, and rate of change over sliding sensor-reading windows), then used to train and compare four models for failure classification — Logistic Regression, Random Forest, Gradient Boosting via XGBoost, and a one-dimensional Convolutional Neural Network applied to short sensor-reading sequences — and two models for RUL regression: a Random Forest Regressor and a Long Short-Term Memory (LSTM) network. XGBoost delivered the strongest failure-classification performance, reaching 98.4% accuracy, 91.7% precision, 88.3% recall, and an 89.9% F1-score on the minority failure class, ahead of Logistic Regression (71.2% F1-score), Random Forest (86.1% F1-score), and the 1D-CNN (87.4% F1-score). For RUL estimation, the LSTM model achieved a Root Mean Squared Error of 18.9 cycles, outperforming the Random Forest Regressor's 24.6 cycles. The best-performing failure-classification model was deployed behind a Flask-based monitoring dashboard that ingests simulated streaming sensor readings, displays a live equipment health status and failure-risk score for each monitored unit, and generates a maintenance alert once the risk score crosses a configurable threshold. System testing showed an average per-reading inference latency of 12 milliseconds, comfortably supporting near-real-time monitoring. The study concludes that gradient-boosted tree models offer a strong, computationally efficient basis for sensor-based failure classification on tabular condition-monitoring data, while recurrent architectures hold a meaningful advantage for sequence-dependent remaining-useful-life estimation, and recommends integration with real industrial IoT sensor streams and cost-sensitive threshold tuning as directions for future deployment.
Computer ScienceSentiment Analysis of Nigerian Social Media Discourse Using NLP
Admin
About This Research Topic Scroll through X (formerly Twitter) during any major Nigerian news cycle — an election, a subsidy announcement, a currency policy shift — and you will find opinion coming in fast, in large volume, and in a mix of English, Nigerian Pidgin, and phrases borrowed from Hausa, Igbo, and Yoruba, often within the same sentence. That mix is exactly what makes Nigerian social media discourse so hard for standard sentiment analysis tools, most of which are trained on tidy, monolingual English text and stumble the moment code-switching, slang, and informal spelling enter the picture. For students exploring a similarly grounded NLP topic, ScholarNest's computer science project topics page is a good place to compare related project ideas in applied machine learning and language processing. This article breaks down a complete undergraduate research project built around that exact problem: an NLP pipeline that reads Nigerian social media posts and classifies their sentiment as positive, negative, or neutral. Rather than testing one model on convenient data, the study collects and manually annotates its own topic-focused dataset, then benchmarks three model families — classical machine learning, a BiLSTM, and a fine-tuned multilingual BERT (mBERT) — before deploying the strongest performer inside a topic-monitoring web dashboard. What follows walks through the study's background, problem statement, objectives, and scope, for students, researchers, and anyone curious about how NLP is adapting to Nigeria's linguistically layered social media conversation. Main Abstract Social media platforms, particularly X, have become a dominant space for Nigerians to weigh in on political, economic, and social issues, generating a volume of unstructured text that is simply impractical to read and interpret by hand. Understanding the sentiment carried in this discourse matters to policymakers, businesses, and researchers alike, but automated sentiment analysis of Nigerian social media text is complicated by widespread code-switching between English, Nigerian Pidgin, and indigenous languages such as Hausa, Igbo, and Yoruba, plus informal spelling, slang, and heavy use of hashtags and emojis. This study designs, implements, and evaluates an NLP pipeline for classifying the sentiment of Nigerian social media posts as positive, negative, or neutral. The work follows a Design Science Research methodology paired with the CRISP-DM process for its data-driven components. A dataset of 12,500 tweets tied to prominent Nigerian discourse topics — the 2023 general election, the fuel subsidy removal, and the Naira redesign policy — was collected via the X API and a Python scraping pipeline, manually annotated by three independent annotators using a majority-vote labelling scheme, and cleaned through a preprocessing pipeline that handled hashtags, mentions, emojis, and Nigerian Pidgin-aware tokenisation. Three model classes were trained and compared: classical baselines (Naive Bayes and Support Vector Machine) using TF-IDF features, a Bidirectional LSTM network with trainable word embeddings, and a fine-tuned multilingual BERT (mBERT) transformer. The fine-tuned transformer model came out on top, reaching 84.7% accuracy and an 83.1% macro-averaged F1-score, ahead of the BiLSTM (79.4% accuracy) and both classical baselines (SVM at 74.2%, Naive Bayes at 68.9%). Error analysis showed that most misclassifications fell into the neutral class or involved tweets with heavy Nigerian Pidgin or code-mixed content, which lines up with the broader low-resource status of these language varieties. The trained model was deployed behind a Flask web dashboard that lets a user submit a topic or hashtag and view an aggregated sentiment breakdown and trend chart drawn from recently collected tweets. Testing showed an average inference time of 0.18 seconds per tweet on a standard CPU-based server. The study concludes that transformer-based multilingual models currently offer the most practical route to reasonably accurate sentiment analysis of Nigerian social media discourse, and recommends further work on expanding annotated data for Nigerian Pidgin and indigenous languages to close the remaining performance gap on code-mixed content.
Computer ScienceFake News Detection Using Deep Learning: Building an Explainable BERT-Based Classifier
Admin
About This Research Topic Every day, social media feeds and messaging platforms mix verified reporting with skilfully disguised falsehoods, and most readers have no easy way to separate the two at a glance. This blurring of fact and fabrication is what researchers call fake news: content engineered to resemble legitimate journalism while carrying misleading or entirely invented claims. For final-year Computer Science and allied ICT students, teaching a machine to spot that difference sits squarely at the intersection of natural language processing, deep learning, and everyday digital literacy, which makes it one of the richer project areas currently available. Students hunting for an equally practical, technically demanding topic can browse ScholarNest's computer science project topics for related ideas in applied machine learning and NLP. This article walks through a complete undergraduate research project built around that exact challenge: a system that reads the text of a news article and classifies it as fake or real using deep learning. Rather than reporting a single accuracy figure in isolation, the study benchmarks four distinct modelling families side by side — classical machine learning, a convolutional neural network, a bidirectional LSTM, and a fine-tuned BERT transformer — on a combined dataset of almost 45,000 labelled articles, then deploys the strongest performer inside a usable, explainable web application. What follows is a structured breakdown of the study's background, problem statement, objectives, and scope, written for students, researchers, and anyone curious about how modern language models are being applied to the misinformation problem. Main Abstract The ease and low cost of publishing on websites and social media has been matched by an equally rapid rise in fabricated and misleading content passed off as news. This kind of content has repeatedly been shown to shape public opinion, disrupt electoral processes, and, in the worst cases, contribute to real-world harm. Manual fact-checking remains accurate but is fundamentally too slow to match the volume of content produced online every day, which is what makes automated, learning-based detection worth pursuing. This study designs, builds, and evaluates a deep-learning system that classifies news articles as fake or real using only their textual content. The work follows a Design Science Research approach paired with the CRISP-DM process for the data-driven components. A combined corpus of 44,898 labelled articles was assembled from the ISOT Fake News dataset and a Kaggle-sourced Fake and Real News dataset, spanning political and general news, and processed through a cleaning pipeline that handled HTML residue, punctuation, stop words, and tokenisation. Four model families were trained and compared under identical conditions: classical baselines (Logistic Regression and Multinomial Naive Bayes) using TF-IDF features, a Convolutional Neural Network with trainable word embeddings, a Bidirectional LSTM network, and a fine-tuned BERT (bert-base-uncased) transformer. The fine-tuned BERT model produced the strongest results, reaching 98.6% accuracy, 98.4% precision, 98.5% recall, and a 98.4% F1-score, ahead of the BiLSTM (95.8% accuracy), the CNN (94.1% accuracy), and both classical baselines (Logistic Regression at 92.3%, Naive Bayes at 89.7%). A closer look at the errors that did occur showed they clustered around short articles with limited context and satire-adjacent writing whose style overlaps heavily with genuine opinion pieces. The trained BERT model was then deployed behind a Flask web application that accepts pasted article text or a URL and returns a predicted label, a confidence score, and the specific words that most influenced the prediction, generated through a model-agnostic explanation technique. Testing showed an average inference time of 0.35 seconds per article on a standard CPU-based server. The study concludes that fine-tuned transformer models currently offer the most practical route to accurate, automated fake news detection from article text alone, while stressing that such tools work best as decision-support for human fact-checkers rather than as a replacement for them, particularly given the persistent difficulty of classifying short or satirical content.
Computer ScienceA Machine Learning Model for Early Crop Disease Detection Using Leaf Image Classification
Admin
About This Research Topic A smallholder farmer who spots a strange leaf pattern rarely has an agronomist a phone call away. By the time an extension officer makes it out to the farm, or the farmer guesses wrong and applies the wrong treatment, the disease has often already spread across the plot. This project tackles that gap directly: a machine learning model that reads a photograph of a leaf and returns a disease diagnosis, a confidence score, and a suggested next step, in under half a second. It's a strong example of the kind of applied computer science and machine learning project work that goes beyond a benchmark score and actually gets built into something a non-technical user could open on a phone. This article walks through how the system was trained, from the PlantVillage image dataset through the transfer-learning model that ended up outperforming every alternative tested, to the web application that puts the diagnosis in a farmer's hands. It closes with what the results mean for smallholder agriculture and where the approach still needs work. Main Abstract Plant diseases remain one of the most significant threats to global food security, with smallholder farmers in developing regions particularly vulnerable because of limited access to agricultural extension officers and diagnostic laboratories. Conventional disease diagnosis, which relies on manual visual inspection by agronomists, is slow, subjective, labour-intensive, and difficult to scale across large farmlands. This study addresses that problem by designing, implementing, and evaluating a machine learning model capable of detecting and classifying crop diseases at an early stage from images of plant leaves. The study adopted the Design Science Research (DSR) methodology, combined with the Cross-Industry Standard Process for Data Mining (CRISP-DM) for the data-driven components of the work. A dataset of leaf images spanning healthy and diseased classes of tomato, maize, and cassava was sourced from the PlantVillage repository and augmented with locally simulated field images to improve generalisation to real farm conditions. The images were preprocessed through resizing, normalisation, and augmentation (rotation, flipping, brightness adjustment) before being used to train a Convolutional Neural Network (CNN) built on a transfer-learning backbone (MobileNetV2), which was benchmarked against a custom shallow CNN and classical machine learning baselines (Support Vector Machine and Random Forest trained on handcrafted colour and texture features). The proposed system was implemented as a web-based application with a Flask backend and a lightweight interface that allows a farmer or extension worker to upload a leaf photograph and immediately receive a predicted disease class, a confidence score, and a suggested remedial action. The transfer-learning model achieved the best performance, with an overall test accuracy of 96.4%, precision of 95.8%, recall of 96.1%, and an F1-score of 95.9%, outperforming the custom CNN (91.2% accuracy) and the classical baselines (SVM: 84.7%; Random Forest: 81.3%). System-level testing further showed an average inference response time of 0.42 seconds per image on a standard CPU-based server, indicating suitability for near real-time deployment. The study concludes that transfer-learning-based CNN models offer a practical, low-cost, and reasonably accurate route to early crop disease detection and recommends further work on expanding the dataset to more crop species, incorporating disease-severity estimation, and deploying the model on offline-capable mobile devices for use in low-connectivity rural areas.
Computer ScienceAn AI-Powered Academic Performance Prediction System for Undergraduate Students
Admin
About This Research Topic By the time a university releases semester results, the window for helping a struggling student has already closed for that term. Continuous assessment scores, attendance records, and study habits all carry warning signs weeks earlier, but most institutions never systematically look at them until it's too late to act. This project takes a different approach: building an AI system that reads those signals early and hands academic advisers a ranked, explainable list of students who need attention now, not next semester. It's a strong example of the kind of applied computer science project work that combines real machine learning technique with a genuinely deployable tool, rather than stopping at an offline accuracy score. This article walks through how the system was built, from the 220-student survey and institutional records behind it, through the four competing machine learning models tested, to the web dashboard that turns predictions into something an adviser can actually act on. It closes with what the results mean for academic advising and where a system like this can go next. Main Abstract Early identification of undergraduate students at risk of poor academic performance is critical for enabling timely academic intervention, yet most tertiary institutions still rely on end-of-semester results to identify struggling students, by which point corrective action is often too late to be effective. This study addresses that problem by designing, implementing, and evaluating an artificial-intelligence-powered system that predicts undergraduate academic performance from a combination of academic history, continuous assessment scores, attendance records, and self-reported study-habit and engagement factors. The study adopted the Design Science Research (DSR) methodology together with the Cross-Industry Standard Process for Data Mining (CRISP-DM) for the data-driven components of the work. Data were collected through a structured questionnaire administered to 220 undergraduate Computer Science students, covering demographic information, study habits, class attendance, and engagement indicators, combined with matching institutional academic records (CGPA and course-level grades) obtained with appropriate consent. The combined dataset was cleaned, encoded, and used to train and compare four predictive models: Logistic Regression and Random Forest for at-risk classification (pass/at-risk), and a Random Forest Regressor and a feed-forward Artificial Neural Network (ANN) for continuous CGPA-band prediction. Models were evaluated using accuracy, precision, recall, F1-score, and confusion-matrix analysis for the classification task, and R-squared, RMSE, and MAE for the regression task. The Random Forest classifier achieved the best classification performance, with an accuracy of 89.6%, precision of 88.7%, recall of 89.1%, and an F1-score of 88.9%, outperforming Logistic Regression (81.4% accuracy). For continuous performance prediction, the ANN achieved an R-squared of 0.81 and an RMSE of 0.34 grade points, marginally outperforming the Random Forest Regressor (R-squared of 0.77). Feature-importance analysis identified continuous assessment score, class attendance rate, and weekly self-study hours as the strongest predictors of academic outcome. The trained Random Forest classifier was embedded in a web-based dashboard that allows academic advisers to view a ranked list of at-risk students each semester, together with the key factors driving each prediction, while a complementary student-facing view allows individual students to see their own risk indicator and general improvement suggestions. System testing showed an average dashboard response time of 0.6 seconds, and a usability evaluation with academic advisers and students returned generally favourable ratings for clarity and perceived usefulness. The study concludes that ensemble machine learning models, informed by both institutional records and self-reported engagement data, can meaningfully support early identification of at-risk undergraduates, and recommends integration with existing student-information systems and periodic model retraining as the student population and curriculum evolve.
Computer ScienceDESIGN AND IMPLEMENTATION OF EDUQUEST: A GAMIFIED LEARNING APPLICATION FOR IMPROVING STEM ENGAGEMENT AMONG SECONDARY SCHOOL STUDENTS
Admin
About This Research Topic Walk into almost any Nigerian secondary school classroom during a Mathematics or Basic Science period, and a familiar scene plays out: a teacher writes formulas on a chalkboard while rows of students copy silently, waiting for the bell. This is the environment that continues to shape how millions of JSS and SS students experience Science, Technology, Engineering, and Mathematics, and it is a large part of why so many of them grow up believing STEM subjects are difficult, dry, and disconnected from real life. This article rewrites and expands a university research project titled “Design and Implementation of EduQuest: A Gamified Learning Application for Improving STEM Engagement Among Secondary School Students,” a study that treats classroom disengagement not as a fixed reality but as a design problem. The researcher built EduQuest, a web-based application that borrows the mechanics of games — points, badges, levels, and leaderboards — and applies them to everyday STEM lessons. For students and supervisors browsing Computer Science project ideas, the underlying approach also connects naturally to Scholarnest's growing library of computer science project topics , where similar systems-design and software-evaluation studies are catalogued. What follows is a complete rewrite of the project's abstract and first chapter, reorganised for readability and search visibility while preserving every objective, research question, and scope boundary of the original study. Nothing has been invented: every statistic, hypothesis, and definition below reflects what the original researcher reported. Main Abstract Across much of the developing world, and Nigeria in particular, Science, Technology, Engineering, and Mathematics (STEM) education is treated as the engine of national innovation and long-term economic growth, yet secondary school students routinely show falling interest and weak performance in STEM subjects. Researchers point to a familiar set of culprits: instruction that leans almost entirely on lectures, laboratories that are too poorly equipped to support hands-on experimentation, and a habit of presenting scientific and mathematical ideas in the abstract rather than connecting them to anything a student can see or touch. This study responds to that gap by designing and implementing EduQuest, a gamified, web-based learning platform built specifically to lift STEM engagement among secondary school students through recognisable game mechanics — points, badges, leaderboards, progressive levels, and interactive challenges. The research combined a survey of 120 students and teachers with an iterative build process guided by the Agile Scrum framework, all situated within a Design Science Research (DSR) methodology. Before writing a single line of code, the researcher mapped out exactly where existing learning tools fall short: static content, no interactivity, and no feedback loop. The resulting system design, expressed through UML diagrams, an entity relationship diagram, and data flow diagrams, layers in adaptive quizzes, real-time scoring, achievement badges, and a progress-analytics dashboard on top of that foundation. EduQuest itself was built on a conventional and dependable stack — HTML5, CSS3, and JavaScript on the front end, PHP via the Laravel framework on the back end, and MySQL for data storage — arranged in a three-tier client-server architecture. To test whether any of this actually moved the needle, the researcher ran a usability and acceptance study using a five-point Likert-scale questionnaire with the same 120 participants, then analysed the results with descriptive statistics and Pearson correlation. The numbers were encouraging: a statistically significant positive relationship emerged between gamification elements and student engagement (r = 0.71, p < 0.05), and 84.2% of respondents said the application made them more motivated to study STEM subjects. Under simulated concurrent use, the platform also held up technically, returning an average page response time of 1.3 seconds and a 94.6% task completion success rate. Taken together, the findings support a straightforward conclusion: gamified learning tools, when their game mechanics are grounded in sound pedagogy rather than novelty for its own sake, can meaningfully improve STEM engagement, motivation, and outcomes among secondary school students. The study recommends that education policymakers and school administrators treat gamified digital platforms as a complement to — not a replacement for — traditional STEM teaching in the secondary school curriculum.
Can't find your topic? Request a custom material →