Graph mining for discovering infrastructure patterns in configuration management databases

Approximate graph mining with label costs

“…Since the search space is exponential, sampling methods have gained traction recently [4,9]. In our algorithm we employ the depth-first random edge extension strategy we proposed in [2], i.e., we employ random walks over the chains of the frequent subgraph partial order. Each random walk starts with an empty pattern and repeatedly adds a new edge to a new vertex, or connects two existing vertices in the pattern, to generate a new candidate.…”

“…Hence, sup(P ) is at least as large as MIS support. Other upper bounds for the MIS value have been proposed in gApprox [5] and CMDB-Miner [2] algorithms. The support function used in gApprox can be computed from the representative sets by enumerating the isomorphisms as described in the Sec.…”

“…Algorithms like those in [19], on the other hand, belong to the class of pattern growth algorithms in which a candidate pattern is generated by extending a frequent pattern with another edge. Sampling approaches like those proposed in [2,9,20] mine a representative set of maximal patterns from a database of graphs or a single graph; they are especially effective in large real-world graphs where complete graph mining is practically infeasible.…”

“…In [7], they proposed a definition of support based on overlapping ancestor isomorphisms. In [2], we proposed CMDBMiner to mine frequent patterns from a single large graph. Support of a pattern is defined as the maximum flow in an appropriately constructed flow network with capacities.…”

“…The search space is explored in a depth first order and the support of a pattern is computed by enumerating all its isomorphisms. This approach is not feasible for large graphs with label multiplicities as there are potentially an exponential number of isomorphisms [2]. In [13], they proposed APGM to mine approximate frequent subgraphs from a database of graphs.…”