<![CDATA[Large-scale administrative reporting systems generate complex multivariate signals, yet institutional risk identification remains challenging when verified labels are scarce. This study develops a deep representation learning framework for anomaly detection and institutional risk ranking under label-scarce conditions. The framework integrates a Variational Autoencoder (VAE) for unsupervised anomaly detection with a Hybrid Autoencoder and risk prediction head using Positive-Unlabeled learning, while XGBoost serves as a supervised benchmark. The models were evaluated on multivariate institutional signals from the Indonesian Higher Education Database using stratified cross-validation and ranking-based metrics. The VAE consistently assigned greater reconstruction deviations to institutions with verified violations, demonstrating its ability to identify anomalous multivariate patterns without relying on complete labels. The Hybrid Autoencoder produced strong risk-ranking performance that closely approached the supervised benchmark while preserving uncertainty among unlabeled institutions rather than treating them as negative cases. These findings demonstrate that combining deep representation learning with Positive-Unlabeled learning provides a practical approach to anomaly detection and risk prioritization when verified labels are severely limited. The proposed framework offers an adaptable foundation for signal-based early warning systems in large-scale distributed administrative reporting environments characterized by multivariate data and label scarcity]]>
Copyrights © 2026
