Multi Input Neural Embedding Architecture for Predicting Digital Resource Adoption in Higher Education

Abstract

Digital transformation in higher education often under delivers due to inconsistent adoption of e-libraries, analytics dashboards, and virtual learning environments across students, faculty, and administrators. To address this gap and offer predictive insights, we present MINEA—a Multi-Input Neural Embedding Architecture—that forecasts individual adoption levels using raw system interaction logs collected entirely through cloud-native infrastructure. The training pipeline processes a wide-format dataset comprising over 10,000 cloud-logged interaction sessions with 39 heterogeneous features, by: (i) extracting chronometric patterns from three timestamp fields, (ii) applying z-score normalization to fifteen cloud-monitored resource usage counters, (iii) one-hot encoding categorical roles and agent types, and (iv) learning dense latent representations for high-cardinality user_id values via embeddings. These features are fed into a dual-branch neural network architecture, where the numerical and one-hot-encoded inputs are concatenated with user embeddings and passed through a stack of ReLU-activated layers (128, 64, 32 units) with dropout regularization. Trained over 200 epochs with early stopping, MINEA achieves strong predictive performance (MAE = 0.0354, RMSE = 0.0722) with an R² indicating that the model explains the vast majority of adoption variability. Residual analysis confirms quasi-normal, homoscedastic errors, while five-fold cross-validation on cloud-hosted data confirms robustness (mean MAE ≈ 0.0374).