Tradeoff in Rule Induction for Semantic Query Optimization

Semantic query optimization (SQO) is a promising approach to the optimization of increasingly complex query plans in global information systems. The idea of SQO is to use semantic rules about data to reformulate a query into an equivalent but less expensive one. Since it is difficult to encode required semantic rules, a complete SQO system also includes a rule induction system and a rule maintainer. To maximize the net utility of learning, a rule induction system needs to learn those rules that are effective in reducing the query execution cost while robust against data changes to minimize the rule maintenance cost. This paper focuses on this tradeoff between effectiveness and robustness in the rule induction for SQO. The solution is to explicitly estimate the degree of the robustness of rules. The system can use the estimated robustness to make decisions to guide rule construction, guide rule repair, and control the size of a rule set. This paper also briefly reviews how robustness can be efficiently estimated and reports the initial experimental results. Learning for Semantic Query Optimization Semantic query optimization (SQO) (King 1981) promising query optimization technique for intelligent information mediators (Wiederhold 1992; Arens et al. 1993; Levy, Srivastava, & Kirk 1995) that integrate heterogeneous information sources because it can complement conventional query optimization techniques to overcome the heterogeneity and considerably reduce query execution cost. The essential idea of semantic query optimization is to use semantic rules about data to reformulate a query into a more efficient but semantically equivalent query. Example 1: Suppose we have a query that retrieves the ship classes and the maximal draft of the ships in those classes which satisfy the following conditions: the ships in the class are capable of carrying containers, and their draft is less than 50 feet. This query is specified as follows: query (?ship_class, ?draft) : 1 : ship_class (?ship_class, _, ?draft, _, ?ctnr), 2: ship (_, ?ship_class ,_, _, ?status), 3: ?ctnr ---"Y", 4: ?status ---"active", S: ?draft < BO. In addition to the query, suppose the system possesses a set of semantic rules. Among them two rules are applicable to this query: RI: IF ship(_,?class,?status,_,_) ship_class (?class, _, ?draft, _,_) & ?draft < 50 THEN ?status = "active" R2: IF ship_class(?class .......?ctnr) ?ctnr = "Y" THEN ship (_,?class ...... ) R1 states that If the mamimum draft of a ship is less than 50, then its status is active. R2 states that If a ship class has container capability, then there must emist some ships that belong to that ship class in the