Discovering Redundancy-Aware Top-k Anomalies in High Dimensional Data