High Order Graphlets for Pattern Classification

Graph-based methods are known to be successful for pattern description and comparison. Their general principle consists in using graphs to model local features as well as their structural relationships and achieving pattern comparison with graph matching. Among these methods, sub-graph isomorphism is particularly effective but intractable for general and unconstrained graph structures. In this paper, we introduce an efficient and effective method for graph-based pattern comparison. The main contribution includes a new stochastic search procedure that allows us to efficiently extract, hash and measure the distribution of increasing order subgraphs (a.k.a graphlets) in large graph collections. We consider both low and high order graphlets in order to model local features as well as their complex interactions. These graphlets are partitioned into sets of isomorphic and non-isomorphic subgraphs using well designed hash functions with a low probability of collision; resulting into accurate graph descriptions. When combined with support vector machines, these high order graphlet-based descriptions have positive impact on the performance of pattern comparison and classification as corroborated through experiments on different standard databases.