Graph Classification SEG 5010 Week 3
A Summary of Graph Features Fingerprint Maccs keys Tree and cyclic patterns Frequent subgraphs Graph fragments
A Boosting Approach to Graph Classification (NIPS04) Apply boosting to graph classification Weak learner: decision stump Definition of the gain function Learning the best weak learner mining the optimal subgraph An upper bound of the gain function and branch-and-bound search
Leap Search (SIGMOD08) The first study to mine the optimal subgraph given “general” user-specified objective functions Vertical pruning: branch-and-bound An objective function may not be anti-monotone, but its upper bound could be anti-monotone Horizontal pruning: structural proximity If two sibling branches are similar in structure, they may be similar in objective function scores There is a lot of redundancy in the graph pattern search tree
COM (CIKM09) Pattern co-occurrences: for effectiveness Joint discriminative power of multiple graph patterns Individual subgraphs are not discriminative, but their co-occurrences become discriminative A different pattern exploration order: for efficiency Complementary discriminative patterns are examined first Generate patterns with higher scores before those with lower scores Rule-based classifiers: a greedy generation process