Extension of Canonical Adjacency Matrices for Frequent Approximate Subgraph Mining on Multi-Graph Collections