Graph Indexing: Tree + Δ ≥ Graph

1. Peixiang Zhao Jeffrey Xu Yu Philip S. Yu CS@UIUC SEEM@CUHK IBM T. J. Watson Research Center September 12th, 2007 Graph Indexing: Tree + Δ ≥ Graph VLDB’07 Vienna, Austria

2. Synopsis • Introduction • Graph Containment Query • Algorithmic Framework • Related Work • Tree + Δ • Indexability of frequent Trees • Discriminative graph feature selection: Δ • Experimental Study • Conclusion VLDB’07 Vienna, Austria

3. Introduction • Graph is a mathematical construct and a general data structure representing relations among entities • The emergence and the dominance of graphs asks for effective graph data management and mining tools so that users can organize, access, and analyze graph data efficiently • Structural Pattern Mining: Given a graph database, what are the potentially interesting structural patterns and how can we find them? • Graph Indexing and Search:How can we index graphs and perform searching, either exactly or approximately, in large graph databases? VLDB’07 Vienna, Austria

4. Introduction • Graph Containment Query • Given a graph database G = {g1, g2, …, gN} and a query graph q, find the set • NP, since subgraph-isomorphism checking is NP-Complete • Infeasible to check subgraph isomorphism sequentially for every gi in G, especially challenging when graphs in G are large, or G is large and diverse • Graph indexing! VLDB’07 Vienna, Austria

5. Graph Indexing: Algorithmic Framework • Index construction generates the index feature set F from the graph database G. For each feature f, sup(f) is maintained • Query processingis performed in a filtering-verification fashion: • The filtering phase uses indexing features contained in q to compute the candidate answer set Every graph in Cqcontains all q's indexed features. Therefore, the query answer set, sup(q), is a subset of Cq • The verification phase checks subgraph isomorphism for every graph in Cq. False positives are pruned and the true answer set sup(q) is returned VLDB’07 Vienna, Austria

6. Query Cost Model • The cost of processing a graph containment query q upon G, denoted C, can be modeled as below • Cf: the filtering cost • Cv: the verification cost (NP-Complete) • Analysis • The key issue to improve query performance is to minimize |Cq| • The indexing feature set F is quite relevant to Cf and |Cq| • Index construction performance: the feature selection cost Cfs to construct F from among G VLDB’07 Vienna, Austria

7. Related Work • Path-based Indexing approach • All existing paths up to a certain length lp are enumerated as indexing features • Index can be constructed efficiently • Index size is quite large when lp is not small • Limited pruning power, mainly because the structural information exhibited in graphs is lost when breaking graphs into paths • GraphGrep (PODS’02) • Graph-based Indexing approach • Subgraphs of G with different characteristics are selected as indexing features • A costly index construction process • Compact index structure • Great pruning power, since structural information of graph is well-preserved • gIndex (SIGMOD’04, PODS’05), C-Tree (ICDE’06), GString (ICDE’07), GDIndex (ICDE’07), FG-Index (SIGMOD’07) VLDB’07 Vienna, Austria

8. An alternative approach: (Tree + Δ) • Tree-based Graph Indexing • Tree: Better indexability in comparison with path and graph • The majority of frequent graph-features of G are usually tree-features indeed • Frequent tree-features and graph-features share similar distributions and frequent tree-features have similar pruning power like graph-features • tree mining can be done much more efficiently than graph mining on G • Δ: On-demand select a small number of discriminative graph-features without conducting costly graph mining beforehand • Orders of magnitude smaller in index size, but performs much better than existing approaches in indexing construction and query processing VLDB’07 Vienna, Austria

9. Indexability of Path, Tree and Graph • Frequent features (paths, trees, graphs) expose intrinsic characteristics of a graph database, G. They are representatives to discriminate between different groups of graphs in a graph database • Which one should we index? Path, Tree or Graph? • The frequent feature set size: | F | • The feature selection cost: Cfs • the candidate answer set size: |Cq| VLDB’07 Vienna, Austria

10. The Frequent Feature Set Size: | F | • Evidences: • Among all frequent graph-features of G, a majority of them are trees indeed • All subtrees of a frequent graph are frequent • There is little chance that subtrees of frequent graph g coincide with those of frequent graph g’, due to the structural diversity and label variety • Frequent pathsshare a very small portion, because a path-feature has a simple linear structure, which has little variety in structural complexity • In terms of feature distributions, tree-features and graph-features share a very similar distribution w.r.t. feature size VLDB’07 Vienna, Austria

11. Experiments on Two Datasets w.r.t.| F | The Real Dataset The Synthetic Dataset VLDB’07 Vienna, Austria

12. The feature selection cost: Cfs • Given a graph database, G, and a minimum support threshold, σ, to discover the frequent feature set F (FP/ FT/ FG) from G • Tree • A good compromise between • the more expressive, but computationally harder general graph • the faster but less expressive path • Specialization of general graph avoiding undesirable theoretical properties and algorithmic complexity incurred by graph VLDB’07 Vienna, Austria

13. The Candidate Answer Set Size: |Cq| • We define the pruning power power(f) of a frequent feature f as • The pruning power of a frequent feature set S = {f1, f2 , …, fn} • Theorem 1: Given a frequent graph-feature g, and let its frequent sub-tree set be T (g) = {t1, t2 , …, tn}. Then, power(g) ≥ power(T (g)) • Theorem 2: Given a frequent tree-feature t, and let its frequent sub-path set be P (t) = {p1, p2 , …, pm}. Then, power(t) ≥ power(P (t)) VLDB’07 Vienna, Austria

14. Pruning Power • The pruning power of all frequent subtree features, T (g), of a frequent graph-feature g can be similar to the pruning power of g • There is a big gap between the pruning power of a graph-feature g and that of all its frequent sub-path features, P(g) The Real Dataset VLDB’07 Vienna, Austria

15. Indexability of Path, Tree and Graph • It is feasible and effective to select FT, instead of FG, as indexing features for the graph containment query problem • The frequent tree-feature set, FT, dominates FG • Discovering frequent tree-features from G can be done much more efficiently than mining frequent general graph-features • FTcan contribute similar pruning power like that provided by FG VLDB’07 Vienna, Austria

16. Δ Discriminative Graph Features • Consider a query graph q which contains a subgraph g • If power(T (g)) ≈power(g), there is no need to index the graph-feature g, because its subtrees jointly have the similar pruning power • if power(g) >>power(T (g)), it will be necessary to select g as an index feature because g is more discriminative than T (g), in terms of pruning • Discriminative graph-features (w.r.t. its subtree-features, controlled by ε0) are selected from queries on-demand, without mining the whole set of frequent graph-features from G beforehand • Discriminative graph-features are used as additional indexing features, denoted Δ, which can also be reused further to answer subsequent queries VLDB’07 Vienna, Austria

17. Discriminative Graph Selection • Given two graphs g, g’ q , where g g’ • If the gap between power(g’) and power(g) is large enough, g’ will be reclaimed from G; • Otherwise, g is discriminative enough for pruning purpose, and there is no need to reclaim g’ in the presence of g • Approximate the discriminative computation between g’ and g, in the presence of our knowledge on frequent tree-features discovered VLDB’07 Vienna, Austria

18. Discriminative Graph Selection • The occurrence probability of g in the graph database, G • the conditional occurrence probability of g’, w.r.t. g, models the probability to select g’ from G in the presence of g • The upper and lower bound of Pr(g’|g) • The conditional occurrence probability of Pr(g’|g), is solely upper-bounded by T (g’) VLDB’07 Vienna, Austria

19. Experimental Studies • The Real Dataset • The AIDS antiviral screen dataset from Developmental Theroapeutics Program in NCI/NIH • 42390 compounds retrieved from DTP's Drug Information System • 63 kinds of atoms in this dataset, most of which are C, H, O, S, etc. • Three kinds of bonds are popular in these compounds: single-bond, double-bond and aromatic-bond • On average, compounds in the dataset has 43 vertices and 45 edges. • The graph of maximum size has 221 vertices and 234 edges VLDB’07 Vienna, Austria

20. Experimental Studies • The real dataset: index construction VLDB’07 Vienna, Austria

21. Experimental Studies • The real dataset: false positive ratio (|Cq|/|sup(q)|) w.r.t. the database size (= 1,000; 2,000; 4,000; 8,000; 10,000) VLDB’07 Vienna, Austria

22. Experimental Studies • The Synthetic Dataset • Generated by a widely-used graph generator, which is controlled by the following parameters: VLDB’07 Vienna, Austria

23. Experimental Studies • The synthetic dataset: false positive ratio VLDB’07 Vienna, Austria

24. Conclusion • Graph indexing plays a critical role in graph containment query processing on large graph databases • Path-based and graph-based indexing approaches suffer from overly large index size, substantial index construction overhead and expensive query processing cost • (Tree+Δ) is an effective and efficient graph indexing feature to answer graph containment queries • (Tree+Δ) holds a compact index structure, achieves good performance in index construction and most importantly, provides satisfactory query performance for answering graph containment queries over large graph databases VLDB’07 Vienna, Austria

25. Thank you VLDB’07 Vienna, Austria