Mining Approximate Top-K Subspace Anomalies in Multi-Dimensional Time-Series Data