The adaptive engagement framework: enhancing banking customer experience through AI-powered invisible marketing.

The proliferation of AI-driven personalization in digital banking has created new opportunities for delivering targeted financial recommendations without triggering customer resistance. This study proposes an Adaptive Engagement Framework that operationalizes invisible marketing by classifying customers according to interaction propensity and aligning engagement timing, channel, and content with empirically identified behavioral receptivity signals. A dataset of 45,211 customer records from Bank Mellat, comprising 5,308 High Interaction and 39,903 Low Interaction customers, was analyzed. Class imbalance was addressed through Random OverSampling applied within each cross-validation fold. Mutual Information was employed for feature importance ranking and dimensionality reduction. Four classification algorithms, namely Random Forest, Decision Tree, Support Vector Machine, and Deep Neural Network, were evaluated under 5-fold stratified cross-validation. A novel feature-to-image transformation pipeline was developed to encode tabular customer records as 64 × 64 grayscale images, enabling evaluation of three CNN architectures: Plain CNN, CNN-SE, and the proposed Residual Dual-Attention Depthwise-Separable CNN (RDAD-CNN). Inter-model differences were assessed using the Wilcoxon signed-rank test, Friedman test, and Cohen's d effect sizes, supported by permutation-based and Gini importance interpretability analyses. Call duration and account balance were identified as the two dominant predictors of customer interaction class, with normalized Mutual Information scores of 1.000 and 0.995 respectively. Among classical classifiers, Random Forest achieved the highest performance (accuracy = 0.9698 ± 0.0014; AUC = 0.9997 ± 0.0001; High Interaction recall = 0.9990 ± 0.0003). The proposed RDAD-CNN achieved superior performance across all nine evaluated metrics (accuracy = 0.9906 ± 0.0005; AUC = 0.9942 ± 0.0003; MCC = 0.9812 ± 0.0009), with statistically significant improvements confirmed by the Friedman test (χ