Leveraging NLP and graph-based machine learning for behavioral and transactional big data analytics in the Nigerian retail banking sector

The rapid digitalisation of Nigerian retail banking has generated unprecedented volumes of behavioural and transactional data, exposing the inadequacy of conventional rule-based and single-modality machine learning systems. Existing approaches—applied exclusively to structured tabular data—fail to exploit unstructured behavioural signals in customer communications or complex relational patterns in transaction networks. This paper proposes, implements, and empirically validates the NLP-GNN Integrated Analytics Framework for Nigerian Banking (NIAF-NB), the first integrated framework to combine Natural Language Processing (NLP) and Graph-Based Machine Learning (GBML) for multimodal analytics in Nigerian retail banking. The framework comprises three core components: (1) Nigerian FinBERT, a BERT-based language model adapted for Nigerian banking text, including Nigerian Pidgin English; (2) a Heterogeneous Graph Attention Network (HAN) modelling the transaction ecosystem as a multi-type relational graph; and (3) a Joint Cross-Attention (JCA) fusion mechanism that integrates NLP-derived semantic representations with GBML-derived topological representations. Evaluated on an anonymised dataset of approximately 1.53 billion transactions and 102 million customer interactions from three tier-1 Nigerian commercial banks over a 12-month period, NIAF-NB significantly outperforms all baselines across three tasks: customer churn prediction (AUC-ROC 0.967), credit risk assessment (AUC-ROC 0.893; 0.887 for thin-file customers), and Anti-Money Laundering (AML) detection (Precision 0.71, vs. 0.08 for existing rule-based systems). Ablation studies confirm the incremental contribution of each component. Fairness evaluation demonstrates equitable outcomes across demographic subgroups. The framework presents a deployable blueprint for Nigerian banks seeking to move from fragmented data silos to integrated, multimodal intelligence.