Download presentation
Presentation is loading. Please wait.
Published byAustin Tucker Modified over 9 years ago
1
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Wednesday, February 7, 2001 William H. Hsu Department of Computing and Information Sciences, KSU http://www.cis.ksu.edu/~bhsu Readings: “Using Inductive Learning to Generate Rules for Semantic Query Optimization”, Hsu and Knoblock KDD Presentation (3 of 3): Rule Induction Lecture 10
2
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Presentation Outline Paper –“Using Inductive Learning to Generate Rules for Semantic Query Optimization” –Authors: C.-N. Hsu and C. A. Knoblock –In Advances in Knowledge Discovery in Databases (Fayyad, Piatetsky-Shapiro, Smyth, Uthurusamy, eds.) Overview –Learning semantic knowledge Rule induction Purpose: semantic query optimization (SQO) –Analogue: inductive logic programming (ILP) Knowledge representation: Horn clauses Idea: use reformulation of queries to learn (induce) rules Application of Machine Learning to KDD: Issues –Rules: Good hypothesis language for performance element (SQO)? –How are goals of database query speedup achieved? –Key strengths: straightforward induction method; can use domain theory
3
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Induction as Inverted Deduction: Design Principles Recall: Definition of Induction –Induction: finding h such that D. (B D x i ) | f(x i ) A | B means A logically entails B x i ith target instance f(x i ) is the target function value for example x i (data set D = { }) Background knowledge B (e.g., inductive bias in inductive learning) Idea –Design inductive algorithm by inverting operators for automated deduction –Same deductive operators as used in theorem proving Theorem Prover Deductive System for Inductive Learning Training Examples New Instance Assertion { c H } Inductive bias made explicit Classification of New Instance (or “Don’t Know”)
4
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Induction as Inverted Deduction: Example Deductive Query –“Pairs of people such that u is a child of v” –Relations (predicates) Child (target predicate) Father, Mother, Parent, Male, Female Learning Problem –Formulation Concept learning: target function f is Boolean-valued i.e., target predicate –Components Target function f(x i ): Child (Bob, Sharon) x i : Male (Bob), Female (Sharon), Father (Sharon, Bob) B: {Parent (x, y) Father (x, y). Parent (x, y) Mother (x, y).} –What satisfies D. (B D x i ) | f(x i )? h 1 : Child (u, v) Father (v, u).- doesn’t use B h 2 : Child (u, v) Parent (v, u).- uses B
5
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Induction as Inverted Deduction: Advantages and Disadvantages Advantages (Pros) –Subsumes earlier idea of finding h that “fits” training data –Domain theory B helps define meaning of “fitting” the data: B D x i | f(x i ) –Suggests algorithms that search H guided by B Theory-guided constructive induction [Donoho and Rendell, 1995] aka Knowledge-guided constructive induction [Donoho, 1996] Disadvantages (Cons) –Doesn’t allow for noisy data Q: Why not? A: Consider what D. (B D x i ) | f(x i ) stipulates –First-order logic gives a huge hypothesis space H Overfitting… Intractability of calculating all acceptable h’s
6
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Inverting Resolution: Example C: Pass-Exam Study C 2 : Know-Material Study C 1 : Pass-Exam Know-Material Resolution C: Pass-Exam Study C 2 : Know-Material Study C 1 : Pass-Exam Know-Material Inverse Resolution
7
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Goals (Section 17.1) –Use semantic rules to find “shortcuts” to queries Example: all CIS 864 students have studied basic probability Query: “Find all CIS 864 students who have had courses in probability and stochastic processes” Can drop condition –Learn rules from data Observe when query can be simplified Generalize over these “training cases” Background (Section 17.2) –Queries: Datalog select-from-where subset of Structured Query Language (SQL) –Semantic rules: Horn clauses (cf. Prolog) Learning Framework (Section 17.3) –Concept: SatisfyInputQuery (+ iff instance, i.e., tuple, satistifes query) –Algorithm for dropping constraints (generalization): greedy min-set-cover –Heuristic (preference bias): gain/cost ratio Semantic Query Optimization (SQO) Methodology
8
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Learning Framework and Algorithm Given: Few Example Queries, Data Set D (Many Tuples) Methodology (Sections 17.3-4) –Step 1 (Optimizer): optimize queries by dropping constraints if possible Use Greedy-Min-Set-Cover algorithm Call learning module to add rules to rule base –Step 2 (Find Alternative Queries): 2a (Construct Candidate Constraints): use gain/cost ratio (number of – cases excluded / syntactic length of constraint) Rationale: Occam’s Razor bias, min-set-cover (ratio-bounded approximation) 2b (Search for Constraints): build on newly-introduced relations –Step 3 (Update Rule Bank): apply newly discovered rules Put newly-induced rules into rule base Use inference engine (Prolog) to generate facts that will shorten query search
9
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Design Rationale Problem (Sections 17.1-4) –How to generalize well over reformulable queries? –Want to make sure inducer does not overfit observed pattern of training examples Solution Approach (Section 17.3-4) –Idea: Occam’s Razor bias Prefer shorter hypotheses, all other things being equal Why does this work? Types of Bias –Preference bias Captured (“encoded”) in learning algorithm Compare: search heuristic –Language bias Captured (“encoded”) in knowledge (hypothesis) representation Compare: restriction of search space aka restriction bias
10
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Experimental Method Experimental Results (Section 17.5) –Improvement using SQO by rule induction (Table 17.4) Reformulation using induced rules improves short and long queries (about uniformly) Speedup Breakdown of savings by NIL queries vs. overall Claims (Section 17.5) –SQO is scalable: can use rule induction on large DBs –SQO is general: can apply other search techniques, heuristics
11
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Summary: Content Critique Key Contribution –Simple, direct integration of inductive rule learning with SQO –Significance to KDD: good way to apply ILP-like learning in DB optimization –Applications Inference Decision support systems (DSS) Strengths –Somewhat generalizable approach Significant for KDD Applies to other learning-for-optimization inducers –Formal analysis of SQO complexity –Experiments: measure Speedup learning % time saved How wasted time is saved (NIL queries, short vs. long queries) cf. performance profiling Weaknesses, Tradeoffs, and Questionable Issues –Insufficient comparison of alternative heuristics (MDL, etc.) –Empirical performance of exhaustive search?
12
Kansas State University Department of Computing and Information Sciences CIS 830: Advanced Topics in Artificial Intelligence Summary: Presentation Critique Audience: Researchers and Practitioners of –AI (machine learning, intelligent database optimization) –Database management systems –Applied logic Positive and Exemplary Points –Good, abstract examples illustrating role of SQO and ILP –Real DB optimization example (3 Oracle DBs) Negative Points and Possible Improvements –Insufficient description of analytical hypothesis representations –Semantics: not clear how to apply other algorithms of rule induction Decision tree First-order ILP (e.g., FOIL)
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.