@Proceedings{IndabaXNG2026, title = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, editor = {Sakinat Folorunso and Roseline Ogundokun and Francisca Oladipo}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, volume = 319 } @InProceedings{pmlr-v319-folorunso26a, title = {Preface: Proceedings of {IndabaX} {Nigeria} 2026}, author = {Folorunso, Sakinat Oluwabukonla and Ogundokun, Roseline and Oladipo, Francisca}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {i--vi}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/folorunso26a/folorunso26a.pdf}, url = {https://proceedings.mlr.press/v319/folorunso26a.html}, abstract = {Preface to the Proceedings of IndabaX Nigeria 2026, Volume 319 of the Proceedings of Machine Learning Research. IndabaX Nigeria 2026 received 176 submissions, of which 32 papers were accepted for publication, representing an acceptance rate of 18.2%.} } @InProceedings{pmlr-v319-isichei26a, title = {Characterising the Nigerian Stock Exchange for Machine Learning-Based Portfolio Research: An Empirical Analysis of Return Distributions, Volatility Dynamics, Liquidity, and Correlation Structure}, author = {Isichei, Nnamdi A. and Oduwole, Kehinde}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {1--15}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/isichei26a/isichei26a.pdf}, url = {https://proceedings.mlr.press/v319/isichei26a.html}, abstract = {Machine learning portfolio systems encode the statistical assumptions of the markets they were trained on. Deploying such systems on the Nigerian Stock Exchange (NGX) without first characterising that market is methodologically indefensible, yet the NGX is almost entirely absent from the empirical ML-finance literature. This paper fills that gap. Using daily OHLCV data on 22 NGX stocks from 2015–2025, we document five structural features of the market. All 22 stocks reject normality under the Jarque–Bera test; median excess kurtosis is 4.79 and 19 of 22 stocks are positively skewed. Mean annualised volatility is 43.3%, with GARCH(1,1) persistence ($\alpha + \beta$) exceeding 0.90 in 12 stocks and a cross-sectional median of 0.912. The average pairwise return correlation is 0.109, far below typical developed-market levels, with a dense banking-sector cluster as the only dominant structure. Liquidity risk operates through volume episodicity rather than zero-volume illiquidity: only 7 zero-volume days occur across the entire dataset, yet the mean volume coefficient of variation is 1.85. At the index level, the NGX composite returns a Sharpe ratio of 1.40 over the sample period, nearly double that of the S&P 500 (0.71) at comparable volatility, though nominal Naira figures are materially affected by the 2023–2024 devaluation episode. Each finding is translated into a concrete design requirement for the two-stage Graph Neural Network and Reinforcement Learning portfolio framework that constitutes the broader research programme.} } @InProceedings{pmlr-v319-abraham26a, title = {Bridging the Domain Gap: Transfer Learning and Aggressive Fine-Tuning for Robust Plant Disease Detection in Low-Resource African Agriculture}, author = {Abraham, Sunday Aspita and Adeniyi, Abidemi and Mughal, Qurrat Ul Ain and Olanloye, Odunayo}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {16--25}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/abraham26a/abraham26a.pdf}, url = {https://proceedings.mlr.press/v319/abraham26a.html}, abstract = {Farmers across Sub-Saharan Africa lose harvests to diseases they cannot name. By the time leaf spots are obvious, the damage is done. Deep learning models such as CNN excel on benchmark datasets, yet when brought to a real farm in southwestern Nigeria during harmattan season—when exposed to factors such as dust, glare, and soil clutter—they fall apart. Field data were gathered in Ibadan and surrounding districts of Oyo and Osun States, totalling 1,742 photographs. Starting from ImageNet-pretrained MobileNetV2, naive fine-tuning helps a little, but not nearly enough. The real lift comes from a two-step training protocol with an auto-fallback safeguard. Tomato accuracy climbs from 34.69% to 97.96%; cassava goes from 52.42% to 98.39%. In low-resource agriculture, the bottleneck is not architecture, but training discipline.} } @InProceedings{pmlr-v319-iram26a, title = {Defending Against Text-Based Social Engineering Attacks Using Federated Adversarial Learning}, author = {Iram, Emdadul Haque and Nahiyan, Navid and Jahan, Musfika and Hasan, Md. Nazmul}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {26--36}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/iram26a/iram26a.pdf}, url = {https://proceedings.mlr.press/v319/iram26a.html}, abstract = {Text-based social engineering attacks such as phishing emails and scam messages have remained quite successful because language and obfuscation patterns are constantly adapted by adversaries, and centralised detection approaches raise serious privacy concerns as well as insufficient real-world deployment. This paper proposes a privacy-preserving and robust detection system combining Joint Embedding Predictive Architecture (JEPA) representation learning with federated learning, enabling multiple clients to collaboratively train a global model without exchanging raw user data. To overcome modelling capacity challenges in heterogeneous (non-IID) client distributions, a Mixture-of-Experts (MoE) design and a Kolmogorov–Arnold Network (KAN)-based prediction head are adopted. The global model is trained via iterative local optimisation and server-side aggregation, with robustness induced by a federated adversarial learning stage. Experimental results demonstrate that the JEPA-Federated-MoE/KAN pipeline consistently exhibits excellent detection performance with privacy preservation and enhanced flexibility against adversarial changes.} } @InProceedings{pmlr-v319-olurinola26a, title = {{EDAIL-EduAI-NG}: Benchmarking Educator {AI} Readiness for Scalable Deployment in Low-Resource Classrooms}, author = {Olurinola, Oluwakemi D. and Folorunso, Sakinat and Owor, Patrick}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {37--50}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/olurinola26a/olurinola26a.pdf}, url = {https://proceedings.mlr.press/v319/olurinola26a.html}, abstract = {Artificial Intelligence (AI) is increasingly accepted as a change agent in education, but scale-up in resource-scarce contexts is hindered by lack of educator readiness and the absence of longitudinal FAIR-aligned benchmarks. This paper presents EduAI-NG, an open-source, FAIR-aligned longitudinal dataset from the EDAIL (Educators’ AI Literacy) programme in Nigeria. The dataset consists of 2,239 pre-training and 1,068 post-training records, along with a matched dataset of 770 educators. A composite Teacher AI Deployment Readiness Index (TADRI-lite) is constructed with lightweight machine learning baselines to validate predictive utility. Results show substantial improvements following the intervention: mean AI understanding increased from 2.80 to 3.93 (Cohen’s $d = 0.852$, $p < 0.001$), while the composite readiness index demonstrated a very large effect ($d = 1.29$) with excellent internal consistency (Cronbach’s $\alpha = 0.93$). Context-aware professional development has significant potential to affect educator AI readiness in the Global South.} } @InProceedings{pmlr-v319-oladunjoye26a, title = {Efficient Forecasting of Economic Indicators Using Lightweight Time Series Models in Resource-Constrained Environments}, author = {Oladunjoye, Ahmad and Akindotuni, Precious}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {51--61}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/oladunjoye26a/oladunjoye26a.pdf}, url = {https://proceedings.mlr.press/v319/oladunjoye26a.html}, abstract = {The accurate prediction of economic indicators is essential for decision-making in organisations operating with restricted computational capabilities. This study investigates lightweight time series forecasting models—Naı̈ve Forecast, Exponential Smoothing, ARIMA, and Prophet—applied to predicting exchange rate, GDP growth, inflation rate, and interest rate. Results show that simpler models, particularly Naı̈ve and Exponential Smoothing, achieve competitive accuracy across most indicators while maintaining significantly lower computational cost. This study provides practical insights for deploying efficient forecasting solutions in low-resource settings.} } @InProceedings{pmlr-v319-adedokun26a, title = {Development of an Enhanced Machine Learning Model for Deception Detection Leveraging Facial Action Units and Linguistic Features}, author = {Adedokun, Adetoye Oluwatoyin and Ojo, Adebola K.}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {62--73}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/adedokun26a/adedokun26a.pdf}, url = {https://proceedings.mlr.press/v319/adedokun26a.html}, abstract = {This study proposes a bimodal machine learning approach for multiclass deception detection by integrating facial Action Unit (AU) features and linguistic features extracted from a real-life dataset of video interviews of suspected criminal perpetrators, persons of interest, and convicted criminals. Facial features were obtained using the Facial Action Coding System (FACS), while linguistic cues were derived using Linguistic Inquiry Word Count (LIWC) scores. A mid-level data fusion strategy combines the extracted features into a unified representation. A Random Forest classifier applied to 3,720 real-life samples with 80:20 split and 10-fold cross-validation achieved an overall classification accuracy of 88%. Results confirm that combining facial and linguistic cues from real-life datasets provides a richer representation of deceptive behaviour.} } @InProceedings{pmlr-v319-dang26a, title = {{FedFairGNN}: A Privacy-Preserving and Fairness-Aware Federated Graph Neural Network for Fraud Detection}, author = {Dang, Quang-Vinh and Nguyen, Ngoc-Son-An}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {74--86}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/dang26a/dang26a.pdf}, url = {https://proceedings.mlr.press/v319/dang26a.html}, abstract = {Graph Neural Networks (GNNs) have emerged as a powerful tool for fraud detection. However, existing approaches often rely on centralised training, which raises privacy concerns when data is distributed across institutions. Federated Learning (FL) addresses this but faces unique challenges on graph data: graph heterogeneity across clients and propagation of algorithmic bias. We propose FedFairGNN, a novel framework that simultaneously ensures privacy, fairness, and utility via three components: Fairness-Sensitive Edge Reweighting (FSER), Fairness-Task Gradient Decomposition (FTGD) with Differential Privacy, and Bi-Objective Frank-Wolfe Aggregation (BFWA). Experiments on YelpChi, Amazon, and Elliptic datasets with $K = 3$ clients demonstrate that FedFairGNN achieves a highly competitive performance-fairness trade-off while significantly reducing demographic disparity.} } @InProceedings{pmlr-v319-tenga26a, title = {{LinguaTriage}: Cross-Lingual Transfer and African Language Pretraining for Low-Resource Medical Triage in {Lingala}}, author = {Tenga, Patrick S. and Kholief, Mohamed A.}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {87--100}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/tenga26a/tenga26a.pdf}, url = {https://proceedings.mlr.press/v319/tenga26a.html}, abstract = {We introduce LinguaTriage, the first medical triage classification system for Lingala, a Bantu language of Central Africa spoken by over 45 million people with no prior supervised NLP benchmarks. Working from a 616-sample dataset of annotated symptom descriptions across three urgency levels, we develop a targeted augmentation pipeline and evaluate three architectures: fine-tuned XLM-RoBERTa (XLM-RFT), a two-stage cross-lingual transfer system (XLM-RCL), and fine-tuned AfriBERTa-Large (AfriBERTaFT). AfriBERTaFT achieves macro-F1 of 0.974 and perfect Emergency recall (1.00) on the internal test set. Mixing just 100 in-domain examples into training improves external accuracy from near-chance to 79%, demonstrating that minimal target-domain exposure far outweighs architectural choices for generalisation.} } @InProceedings{pmlr-v319-adeyemo26a, title = {Towards Trustworthy Email Phishing Detection: Integrating Multi-Modal Deep Learning, Federated Learning, and Explainable {AI}}, author = {Adeyemo, Adetoye A. and Emuoyibofarhe, Ozichi and Abiodun, Adeyinka and Adegboye, James and Ajagbe, Sunday}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {101--117}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/adeyemo26a/adeyemo26a.pdf}, url = {https://proceedings.mlr.press/v319/adeyemo26a.html}, abstract = {This research presents a robust phishing detection system integrating multi-modal deep learning with federated learning and explainable AI. The model combines email text analysis with URL structural features using embedding, LSTM, and feature fusion layers. In the centralised setup, the model achieved accuracy of 99.6%, precision of 99.8%, recall of 99.5%, F1-score of 99.72%, and AUC of 99.85%. The federated model maintained competitive performance while protecting user privacy. LIME and perturbation analysis reveal which word-level features drive phishing classification decisions. Federated learning offers a strong privacy-preserving alternative despite marginally lower metrics than centralised training.} } @InProceedings{pmlr-v319-elesemoyo26a, title = {{YOLOv11}-Based Deep Learning System for Accurate and Real-Time Tomato Disease Classification}, author = {Elesemoyo, Isaac Oluwafemi and Adeniyi, Emmanuel and Oyebade, Adedoyin and Faruk, Yusuf}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {118--130}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/elesemoyo26a/elesemoyo26a.pdf}, url = {https://proceedings.mlr.press/v319/elesemoyo26a.html}, abstract = {This study proposes a system for detecting and classifying tomato diseases using the YOLOv11 model in real time. The proposed system utilised the PlantVillage dataset of 18,160 images across ten diseased and one healthy tomato class. With data processing, transfer learning, and hyperparameter optimisation, the trained YOLOv11 model achieved an accuracy of 99.2% and mean average precision (mAP@0.5) of 0.93. The system was deployed on a lightweight web application built with React.js and FastAPI, enabling users to upload images and receive instant predictions. The system has potential for reducing dependency on expert physical inspection and minimising yield losses in Nigerian agriculture.} } @InProceedings{pmlr-v319-folorunso26b, title = {{ORIN-Lyrics}: A Multilingual Nigerian Song Lyrics Dataset and Baseline for Efficient Language Detection}, author = {Folorunso, Sakinat O. and Odunsi, Oluwagbenga and David, Ayodele and Olaleye, Daniel and Salami, Fatimah and Giwa, Oluwakemi}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {131--143}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/folorunso26b/folorunso26b.pdf}, url = {https://proceedings.mlr.press/v319/folorunso26b.html}, abstract = {This study presents ORIN-Lyrics, a multilingual dataset of Nigerian song lyrics following FAIR data principles for multilingual NLP and cultural AI research. The dataset includes 853 songs representing 22 musical genres and 18 language categories, featuring Yoruba, English, Nigerian Pidgin, and code-switched compositions. The corpus contains 124,801 tokens and 12,098 unique words. An embedding-based visualisation method displays distinct semantic groupings between language categories. A baseline genre classification experiment using TF-IDF features and multi-class logistic regression achieves an accuracy of 0.54, substantially exceeding the random baseline of approximately 4.5%, enabling the creation of African language technologies that match local needs.} } @InProceedings{pmlr-v319-alhassan26a, title = {Detecting Phishing Emails in {Nigerian} {Pidgin} {English} Using a Dialect-Aware and Behavioural {NLP} Model}, author = {Alhassan, Zubaida Muhtar}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {144--154}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/alhassan26a/alhassan26a.pdf}, url = {https://proceedings.mlr.press/v319/alhassan26a.html}, abstract = {This study proposes a dialect-aware and behaviourally informed NLP model for detecting phishing emails in Nigerian Pidgin, spoken by over 100 million people in Nigeria. A balanced dataset of 870 emails was created using a hybrid translation and generation process, validated by native speakers. The model combines TF-IDF-based linguistic features with seven behavioural indicators derived from persuasion theory, optimised via a Genetic Algorithm-tuned Random Forest classifier. The system achieved 93.89% accuracy, 100.00% precision, and 87.69% recall, demonstrating the importance of integrating behavioural and linguistic analysis for cybersecurity in low-resource language contexts.} } @InProceedings{pmlr-v319-olukayode26a, title = {Machine Learning Prediction of Groundwater Contamination Vulnerability Using Hydrogeophysical Indicators in {Ibadan}}, author = {Olukayode, Oluwakemi Omolara and Folorunso, Sakinat and Bayewu, Olateju and Ojo, Odunayo and Olukayode, David and Omotola, Olubunmi and Sokan-Adeaga, Adewale and Olaseeni, Olayiwola}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {155--166}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/olukayode26a/olukayode26a.pdf}, url = {https://proceedings.mlr.press/v319/olukayode26a.html}, abstract = {This study employs machine learning to estimate groundwater contamination vulnerability in Ibadan, southwestern Nigeria, using hydrogeophysical indicators from 353 Vertical Electrical Sounding (VES) surveys. A Random Forest classifier trained on GOD/GODT vulnerability labels achieved accuracy = 0.94, precision = 0.94, recall = 0.93, F1-score = 0.93, and AUC = 0.95. Overburden thickness and longitudinal conductance were the most significant predictors. The model identified fifteen high-vulnerability zones versus nine from conventional GODT, demonstrating its ability to capture nonlinear interactions that conventional methods miss.} } @InProceedings{pmlr-v319-abdulrahman26a, title = {{INT8} Quantisation for Cassava Leaf Disease Classification on {Raspberry Pi}}, author = {AbdulRahman, Tosho Abdulahi and Ogundokun, Roseline Oluwaseun and Owolawi, Pius Adewale and Bello, Rotimi-Williams and Mathonsi, Topside Ehleketani}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {167--178}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/abdulrahman26a/abdulrahman26a.pdf}, url = {https://proceedings.mlr.press/v319/abdulrahman26a.html}, abstract = {We investigate the impact of post-training INT8 quantisation on a CNN trained for the Kaggle Cassava Leaf Disease Classification task (5 classes). Comparing FP32 baseline versus statically quantised INT8 model, the FP32 model achieved 81.5% accuracy (F1 = 0.676, 6.23 MB, 3.04 ms CPU latency), whereas the INT8 model shrank to 1.87 MB but accuracy dropped to 11.9% (58 ms on Raspberry Pi 4). These results reveal a drastic size-accuracy trade-off, highlighting pitfalls and practical lessons for deploying CNNs on edge devices in resource-constrained settings.} } @InProceedings{pmlr-v319-umeadi26a, title = {{MRI}-Based Brain Tumor Detection for the {African} Context}, author = {Umeadi, Jideofor J.}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {179--190}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/umeadi26a/umeadi26a.pdf}, url = {https://proceedings.mlr.press/v319/umeadi26a.html}, abstract = {This study evaluates lightweight transfer learning for binary brain tumor detection under resource constraints in sub-Saharan Africa, with a focus on generalisation across domain shift from Western-heavy training data to actual Nigerian clinical scans. A frozen MobileNetV2 backbone trained on a skull-stripped Kaggle MRI dataset and evaluated on the BraTS-Africa cohort—scans from six Nigerian diagnostic centres—showed that skull-stripping preprocessing improved mean external sensitivity from $52.67% \pm 24.04%$ to $85.67% \pm 13.58%$ across three random seeds. These results demonstrate that targeted domain alignment through simple preprocessing is a viable approach to closing the generalisation gap in low-resource settings.} } @InProceedings{pmlr-v319-gogo26a, title = {Reframing {AI} Design Through {African} Women’s Livelihood Intelligence: A Review and Conceptual Framework for {SME} Contexts}, author = {Gogo, Jacqueline Akelo and Muturi, Emma and Oguntosin, Victoria}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {191--204}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/gogo26a/gogo26a.pdf}, url = {https://proceedings.mlr.press/v319/gogo26a.html}, abstract = {This paper develops a conceptual and methodological foundation for designing inclusive, context-aware AI systems grounded in women’s livelihood practices within small and medium enterprises (SMEs). Drawing on feminist economics, sustainable livelihoods, value-sensitive design, and feminist HCI, the paper synthesises knowledge on how women navigate constraints related to care, safety, informality, and resource access. A multi-layered conceptual framework connects livelihood practices with sociotechnical systems and AI development. The proposed research design combines qualitative inquiry and participatory methods to translate women’s lived experiences into AI design principles and evaluation metrics, advancing a novel research agenda for equitable AI in women-led African SMEs.} } @InProceedings{pmlr-v319-rahman26a, title = {Multi-Crop Reproductive Structure Detection and Counting Using a Lightweight {YOLOv8n} with Spatial–Channel Attention and Re-parameterizable Convolutions}, author = {Rahman, Md. Mushibur and Rahim, Umme Fawzia}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {205--216}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/rahman26a/rahman26a.pdf}, url = {https://proceedings.mlr.press/v319/rahman26a.html}, abstract = {This study presents a lightweight YOLOv8n-based framework for multi-crop reproductive structure detection and counting. A Spatial–Channel Attention Convolution module (C2f-SCA) highlights informative spatial regions and feature channels, while Reparameterizable Depthwise Convolution (RepDWConv) strengthens multi-scale feature representation. Evaluated on Cauliflower, Tomato Flower, and Maize Tassel datasets, the model achieves mAP@0.5 of 0.981, 0.969, and 0.965 with F1-scores of 0.954, 0.921, and 0.936, respectively. These results are achieved with approximately 12% fewer parameters than the YOLOv8n baseline, demonstrating a compact and generalizable solution suitable for precision agriculture in resource-constrained environments.} } @InProceedings{pmlr-v319-habib26a, title = {High-Resolution Micro-Patching: A Zero-Leakage Microaneurysm Baseline}, author = {Habib, Md. Fahim and Alamgir, Anika}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {217--231}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/habib26a/habib26a.pdf}, url = {https://proceedings.mlr.press/v319/habib26a.html}, abstract = {We propose a patient-isolated microaneurysm segmentation scheme that maintains lesion geometry via high-resolution micro-patching. Rather than scaling full fundus images, the technique takes native-resolution spatial crops, preserving small lesion structure that resizing erases. A hybrid encoder-decoder with spatial and channel attention mechanisms suppresses background and emphasises rare vascular abnormalities. Zero-leakage training is applied, and cross-domain performance is evaluated on datasets with varying demographics and resolutions. Results show that native resolution is more robust to distribution shift and consistent in diagnostic sensitivity across unseen domains.} } @InProceedings{pmlr-v319-awokoya26a, title = {Advanced Machine Learning Models for Network Traffic Prediction and Management}, author = {Awokoya, Ayodele E. and Olumurewa, Kayode and Adeduro, Oladapo and Oladapo, Kayode and Omoyen, Rejoice}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {232--243}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/awokoya26a/awokoya26a.pdf}, url = {https://proceedings.mlr.press/v319/awokoya26a.html}, abstract = {This study employs Random Forest, LSTM, and Support Vector Regression (SVR) to forecast network traffic volume at a university. Wireshark was used to capture a full month of campus network traffic. Three models were trained on real-world traffic data and evaluated across short-term (hourly), medium-term (daily), and long-term (weekly) prediction horizons. Across all horizons, LSTM consistently achieved lower RMSE and MAE than Random Forest and SVR, demonstrating its suitability for capturing temporal dependencies in network traffic prediction.} } @InProceedings{pmlr-v319-akande26a, title = {Graph–Neurosymbolic Neural Networks for Trustworthy Clinical Decision Support}, author = {Akande, Oluwatobi Noah and Misra, Amil and Adeniyi, Abidemi and Mughal, Qurrat Ul Ain and Orifama, Dagogo Godwin}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {244--263}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/akande26a/akande26a.pdf}, url = {https://proceedings.mlr.press/v319/akande26a.html}, abstract = {We propose a Graph–Neurosymbolic Framework for Trustworthy Clinical Decision Support that integrates graph-based neural learning with formal symbolic medical reasoning. Patients and clinical entities are represented as a heterogeneous clinical graph; domain knowledge encoded as first-order logic and Horn-clause rules explicitly constrains graph construction, neural message passing, and inference. The framework is evaluated on MIMIC-IV and eICU-CRD, demonstrating competitive predictive performance while substantially reducing clinical rule violations, producing high-fidelity rule-consistent explanations, and exhibiting improved robustness under distribution shift across institutions.} } @InProceedings{pmlr-v319-mahtab26a, title = {{BanglaNLI}: A Benchmark Dataset for {Bangla} Natural Language Inference}, author = {Mahtab, MD Ajmain and Ronan, Atif and Rahman, Sheikh Ayatur and Sajid, Saleh Mohammad and Tasnim, Sanjida and Sadeque, Farig}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {264--277}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/mahtab26a/mahtab26a.pdf}, url = {https://proceedings.mlr.press/v319/mahtab26a.html}, abstract = {We present BanglaNLI, a high-quality Bangla Natural Language Inference dataset with expert annotations. The dataset was constructed from 4,200 image captions from the BanglaLekha-ImageCaption dataset written by native Bangla speakers, with three hypotheses (entailment, contradiction, neutral) generated per premise, yielding 12,600 carefully annotated sentence pairs. Annotation artefacts were minimised by instructing annotators to avoid simple heuristics. Inter-annotator agreement measured by Cohen’s Kappa reached $\geq 0.88$, confirming high-quality annotations for this under-resourced language with over 200 million native speakers.} } @InProceedings{pmlr-v319-kuzayet26a, title = {Benchmarking Classification Performance for Binary-Class Fault Detection Under Real-World Imbalanced Data Conditions}, author = {Kuzayet, Bagai Glory and Honour, Eje Obed and Bala, Jibril Abdullahi}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {278--292}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/kuzayet26a/kuzayet26a.pdf}, url = {https://proceedings.mlr.press/v319/kuzayet26a.html}, abstract = {This study addresses predictive maintenance in Nigeria’s manufacturing sector by benchmarking Logistic Regression, Random Forest, SVM, and XGBoost on 7,672 real-time sensor measurements from rotating electromechanical machinery. Beyond standard benchmarking, we propose a stacking ensemble combining SVM, Random Forest, and XGBoost as heterogeneous base learners under a Logistic Regression meta-learner. The stacked ensemble achieves an accuracy of 98.37% and an F1-score of 91.35%, establishing an empirical foundation for intelligent, data-driven operations aligned with Industry 4.0 principles in Nigerian industrial settings.} } @InProceedings{pmlr-v319-oluwagbenga26a, title = {Parameter-Efficient Fine-Tuning with Culturally-Aligned Adapters for Cross-Lingual Transfer in {Nigerian} Low-Resource Languages}, author = {Oluwagbenga, Taiwo Timothy and Itanyi, Mamudu Francis}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {293--305}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/oluwagbenga26a/oluwagbenga26a.pdf}, url = {https://proceedings.mlr.press/v319/oluwagbenga26a.html}, abstract = {This paper introduces CulturalAdapt, a Parameter-Efficient Fine-Tuning (PEFT) framework adding Low-Rank Adaptation (LoRA) modules to language adapters grounded in Nigerian cultural and linguistic context. CulturalAdapt separates language-specific adaptation (tonal patterns, diacritics, code-switching, morphological structure) from task-specific fine-tuning. Evaluated on NaijaSenti, MasakhaNER 2.0, and AfriSenti, CulturalAdapt achieves state-of-the-art macro-F1 of 77.3 on NER, 79.0 on sentiment analysis, and 84.1 on cross-lingual sentiment transfer, using only 2.1% of trainable parameters and reducing peak GPU memory by $3.4\times$ relative to full fine-tuning.} } @InProceedings{pmlr-v319-sambo-magaji26a, title = {The Agentic Artificial Intelligence Venture Co-Founder ({AIVC}): An {AI} Operating System for Lean Experimentation, Strategic Decisioning, and Responsible Scaling in Technology Startups}, author = {Sambo-Magaji, Amina and Adewale, Muyideen Dele}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {306--319}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/sambo-magaji26a/sambo-magaji26a.pdf}, url = {https://proceedings.mlr.press/v319/sambo-magaji26a.html}, abstract = {This paper introduces the Agentic Artificial Intelligence Venture Co-Founder (AIVC), a conceptual multi-agent system that senses markets and regulatory signals, designs and executes lean experiments, maintains causal traction models, allocates resources, and enforces governance-by-design. The framework defines five capability bundles: sensing, experimentation, decisioning, governance, and venture memory. A conceptual architecture specifies model classes, APIs, and latency budgets across three deployment tiers. AIVC is distinguished from the build-measure-predict-learn model and is supported by four testable propositions, advancing entrepreneurship and ML systems research by shifting focus to the quality of the founder–AI decision loop under Knightian uncertainty.} } @InProceedings{pmlr-v319-ogundokun26a, title = {{HistoEffiCrossFormer}: Cross-Attention {CNN}–{Transformer} Fusion with Multi-Scale Tokens for Ovarian Cancer Histopathology Classification}, author = {Ogundokun, Roseline Oluwaseun and Bello, Rotimi-Williams and Owolawi, Pius Adewale and Tu, Chunling and Agbolade, Sunday and AbdulRahman, Tosho Abdulahi}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {320--330}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/ogundokun26a/ogundokun26a.pdf}, url = {https://proceedings.mlr.press/v319/ogundokun26a.html}, abstract = {We propose HistoEffiCrossFormer, a lightweight CNN–Transformer hybrid for five-class classification of ovarian cancer histopathology images. The pipeline uses Macenko-inspired stain normalisation, EfficientNet-B0 feature extraction, squeeze-and-excitation channel attention, multi-scale tokenisation, a 2-layer transformer encoder, and cross-attention fusion. Benchmarked against SqueezeNet, ShuffleNetV2, AlexNet, and Xception, HistoEffiCrossFormer achieves 0.8267 test accuracy with AUC 0.974, outperforming lightweight CNN baselines (0.7067–0.7867 accuracy) and closely matching Xception’s AUC (0.975), motivating further validation in external cohorts and whole-slide pipelines.} } @InProceedings{pmlr-v319-yusuf26a, title = {From Black-Box to Glass-Box: A Review of Explainable Neuro-Symbolic {AI} for Climate-Induced Food Insecurity Prediction}, author = {Yusuf, Adam Omeiza and Kolajo, Taiwo and Ogbuju, Emeka and Oladipo, Francisca}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {331--345}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/yusuf26a/yusuf26a.pdf}, url = {https://proceedings.mlr.press/v319/yusuf26a.html}, abstract = {We propose a Glass-Box Neuro-Symbolic framework that balances explainability and accuracy for food insecurity prediction in Nigeria. Predictions are grounded in a Knowledge Graph of a harmonised Data Lakehouse, yielding human-readable reasoning paths that relate specific climate anomalies to predicted agricultural outcomes. A user study with agricultural extension officers and policy analysts demonstrates that trust scores for Neuro-Symbolic explanations are significantly higher than SHAP visualisations, supporting the establishment of trusted AI systems for data-driven decisions under climate uncertainty.} } @InProceedings{pmlr-v319-christopher26a, title = {Benchmarking Multimodal Semantic Alignment Between Speech and Text Representations in the {Igbo} Language}, author = {Christopher, Chidiebere}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {346--354}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/christopher26a/christopher26a.pdf}, url = {https://proceedings.mlr.press/v319/christopher26a.html}, abstract = {We present the first systematic benchmark for cross-modal semantic alignment between speech and text in Igbo, a tonal Niger-Congo language with approximately 45 million speakers and virtually no prior multimodal NLP representation. Using 699 stratified utterance pairs from the WAXAL corpus, zero-shot cross-modal cosine similarity is $-0.0009$, statistically indistinguishable from random. A lightweight linear projection (147,840 parameters) trained with symmetric InfoNCE achieves Speech-to-Text Recall@1 of 0.0658 ($5.1\times$ over zero-shot), Recall@10 of 0.3362 ($3.0\times$), and MRR of 0.1557 ($2.7\times$). Alignment is statistically significant ($t = 15.95$, $p = 2.81 \times 10^{-54}$). All embeddings, evaluation code, and benchmark protocols are released.} } @InProceedings{pmlr-v319-olawoye26a, title = {Speech-Based {Parkinson’s} Disease Screening: A Deep Learning Approach Using Acoustic Biomarkers and Invertible Neural Networks}, author = {Olawoye, Prisca O. and Asani, Emmanuel O. and Adebiyi, Marion O.}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {355--367}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/olawoye26a/olawoye26a.pdf}, url = {https://proceedings.mlr.press/v319/olawoye26a.html}, abstract = {We propose a deep learning architecture combining deep residual encoders and squeeze-and-excitation attention mechanisms with invertible normalising flows for speech-based Parkinson’s disease (PD) screening. On the Telemonitoring dataset, the model achieves AUC 0.979, 100% specificity, 72.4% sensitivity, and 79.5% accuracy. On the MSR dataset, it achieves 71.1% accuracy, AUC 0.806, with symmetric sensitivity (71.0%) and specificity (71.2%). Invertible normalising flows enable exact density estimation and principled uncertainty quantification, supporting telemedicine applications for smartphone-based remote screening of Parkinson’s disease.} } @InProceedings{pmlr-v319-rohan26a, title = {Anchor-Guided Repair: A Defense Mechanism for Enhancing Stability of Compromised Pretrained Language Models Against Low-Precision and Weight Noise Attacks}, author = {Rohan, Abrar Mahir and Khan, Nafiz and Tanni, Tahsin Tajwar and Fardin, Fuad and Bushra, Anika}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {368--381}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/rohan26a/rohan26a.pdf}, url = {https://proceedings.mlr.press/v319/rohan26a.html}, abstract = {We propose Anchor-Guided Repair, a defense mechanism for stabilising large language models (LLMs) compromised by weight noise injection and low-precision quantisation attacks. The method retrains the attacked model on clean text with an anchor regularisation loss that penalises large parameter deviations from a clean reference model. The combined objective balances language modelling loss and anchoring regularisation. Tested across various quantisation levels and weighted Gaussian noise attack scenarios, Anchor-Guided Repair consistently improves stability and performance relative to attacked models, demonstrating that anchoring can recover reliability even without proprietary training data.} } @InProceedings{pmlr-v319-ogbonnia26a, title = {A Systematic Review of Causal Machine Learning Approaches in Road Crash Analysis}, author = {Ogbonnia, Emmanuel and Ojerinde, Oluwaseun Adeniyi and Aminu, Enesi Femi and Alabi, Isiaq and Adepoju, Solomon A. and Dorcas, Mitong}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {382--405}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/ogbonnia26a/ogbonnia26a.pdf}, url = {https://proceedings.mlr.press/v319/ogbonnia26a.html}, abstract = {Although machine learning (ML) has significantly increased the predictive accuracy of road crash severity and frequency models, traditional predictive classifiers and consequent interpretability tools have often failed to distinguish between correlation and causation. Such approaches do not have the counterfactual reasoning needed in sound policy development. To explore the paradigm shift of formal causal inference in traffic safety, this review adhered to PRISMA 2020 and synthesised 35 peer-reviewed articles published between 2021 and 2025. The synthesis classifies the literature into a three-level taxonomy of Predictive ML, Interpretable ML, and Causal ML, showing that most existing research remains rooted in purely predictive ensembles or explainability tools such as SHAP. A small but tightly developed body of work executes true causal ML methods: Doubly Robust Learning, Uplift Modelling, and Causal Graph Discovery are effective in determining heterogeneous treatment effects (HTE) and mitigating confounding bias in observational crash data. Critical methodological gaps persist, including the continued conflation of predictive feature importance with causal effect, sensitivity to unobserved heterogeneity, and the absence of standardised causal benchmarks. Comprehensive sensitivity analyses and integration of structural causal models are identified as prerequisites for maturing Intelligent Transportation Systems (ITS) toward proactive and evidence-based safety interventions.} } @InProceedings{pmlr-v319-nazeer26a, title = {{ADINT}: Machine Learning-Powered Advertisement Intelligence for Proactive Threat Detection in {Nigeria’s} Digital Ecosystem}, author = {Nazeer, Muhammad and Mohammed, Habib and Abdulkadir, Nafisat and Odion, Philip Oshiokhaimhele and Irhebhude, Martins Ekata and Shitu, Saifullahi Sadi}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {406--420}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/nazeer26a/nazeer26a.pdf}, url = {https://proceedings.mlr.press/v319/nazeer26a.html}, abstract = {Digital advertising platforms have become unexpected threat vectors in Nigeria: terrorist groups exploit social media ads for recruitment, cybercriminals launder funds through advertising-related fraud, and human traffickers lure victims via deceptive job postings. Yet Nigerian security agencies lack systematic capabilities to monitor advertising content as an intelligence source. This paper presents ADINT, the first machine-learning-based advertisement intelligence framework designed for Nigeria’s threat landscape. A domain-informed synthetic dataset of 3,000 advertisements across four categories—benign (54.93%), fraud (22.90%), terrorism (11.70%), and trafficking (10.47%)—incorporates realistic class imbalance, graduated ambiguity, and lexical noise to simulate operational conditions. A six-phase experimental pipeline evaluates four architectures: BERT achieves the highest accuracy (91.33%) with perfect recall on terrorism and trafficking; Random Forest (90.33%) offers a compelling efficiency-accuracy trade-off for resource-constrained deployment. A proposed two-stage cascade—Random Forest pre-filter followed by BERT refinement—is analytically projected to reduce analyst workload by 75–78% while maintaining zero false negatives on critical threat classes within the synthetic evaluation environment.} } @InProceedings{pmlr-v319-dorcas26a, title = {Optimized Transfer Learning Pipeline Using {ResNet-50} and {Bayesian} Optimization for Oil Spillage Detection}, author = {Dorcas, Mitong and Abdullahi, Muhammad Bashir and Ogbonnia, Emmanuel}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {421--434}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/dorcas26a/dorcas26a.pdf}, url = {https://proceedings.mlr.press/v319/dorcas26a.html}, abstract = {This paper presents an efficient transfer learning framework combining a pretrained ResNet-50 backbone with Bayesian hyperparameter optimisation for oil spill detection in Sentinel-1 SAR imagery. The architecture uses a frozen convolutional feature extractor followed by a compact classification head (Dense-256, Dropout, Sigmoid), with SGD optimiser hyperparameters automatically tuned: learning rate (0.001578), weight decay (0.012348), dropout (0.0129), and momentum (0.5872). On an imbalanced cohort with a large proportion of oceanic look-alike dark spots, the refined model achieves accuracy of 85.36%, weighted F1-score of 85.32, and balanced class-wise performance (Non-Oil F1 = 0.89; Oil F1 = 0.79). Training and validation curves demonstrate consistent convergence without overfitting, confirming the effectiveness of Bayesian optimisation in navigating the complex hyperparameter space of SAR-based oil spill classification. The framework provides a computationally efficient solution for operational monitoring with robust discrimination between genuine spills and natural look-alikes.} }