Schema Evolution

1
Schema Evolution
2
Schema Evolution

• Ref Alex Borgida K. E. Williamson.
  Accommodating Exceptions in Databases, and
  Refining the Schema by Hearing From Them. VLDB,
  1985.
• SCHEMA
• User
• Describes the domain of discoveries of the
  database
• Check the correctness of the data entry
• DBMS
• Storage efficiency (store extra attributes and
  non-unique keys, attribute domains constrain
  violations)
• Faster retrieval
• Semantic query optimization
• Integrity constraints - used to enforce
  additional conditions not captured by the
  restrictive language of class definition
• All employees have a supervisor
• Supervisor wage gt Employee wage

3

• Exceptions In Database
• Some employees make more than their supervisor.
• Some employees may earn a foreign degree, which
  is not even close to the American ones.
• How To Live With Exceptions
• Storage and access efficiency.
• Semantic computation (employee is a consultant).
• Sharing (US dollars are now Swiss francs).
• Validating of future updates.
• Use object concept to represent exceptional
  information so that such information can be
  maintained
• Alert the user when exceptional information is
  being manipulated as well as when constraints are
  violated

4

• Learning From Exceptions
• Modifying integrity constraints to apply in more
  restricted circumstances.
• Adding new attributes to a class definition.
• Defining a new class and placing it in the
  current sub-hierarchy.
• Using Empirical Generalization For Schema
  Refinement
• DBA keeps track of violations and exceptions to
  constraints in the schema.
• When sufficient evidence is accumulated, it
  suggests one or more alternative changes to the
  schema.

5
The Generalization Algorithm

• Introduce/expand range
• Introduce/generalize class
• Supervisor Accountant -gt Supervisor Employee
• Drop attribute specification
• Research (age 30, Ph.D.) and Research (age 35,
  MS)
• -gt Research (age 30-35, MS-Ph.D.)

6
Refinement to the Generalization Process

• If certain attribute specification is dropped,
  then the resulting description becomes overly
  generalized.
• If no longer detects some of the errors it
  detects earlier.
• Need heuristic for determining when an attribute
  constraint is essential.
• Notion of RELEVANCE
• Keep the attribute in the generalization if more
  relevant
• Discard the attributes in the generalization if
  less relevant

7

• Integrity constraints must be refined to
  disregard the contradictions raised by exception
  value.
• Machine learning to find commonalities among the
  elements of a set of objects, such as exception
  to a rule encountered so far and characterize
  these in the form of a class description.
• Use such characterization to adjust schema.
• Changing restriction on property value range
• Modifying integrity constraints
• Define new classes
• Shortcomings
• Heuristic nature of the RELEVANCE function.
• Inability to handle negation and disjunction.

8
A Proposal For Automatic Schema Evolution

• Predefine Schema templates
• Monitoring the exception frequency of constraints
• Switch over to new schema when exception
  frequency reaches a certain pre-specified level
• New schema is selected based on the exception
  characteristics

9

• Ref Q. Li and D. McLeod. Object Flavor
  Evolution Through Learning in an Object-Oriented
  Database System. Proceedings of the 2nd
  International Conference on Expert Database
  Systems, 1989, pp. 469-495.
• PKM Personal Knowledge Model
• A simple object-oriented model
• Based on a small number of simple concepts
• Stepwise develop more complex and abstract
  database models
• Level 1 Atomic Object - Non-decomposable units of
  information at a certain stage (e.g., John
  Smith).
• Level 2 Open Atomic Set - A collection of atomic
  objects, members are not fixed (e.g., P
  names).

10

• Closed Atomic Set - A collection of atomic
  objects, members are fixed e.g., sex, age (gt 0,
  integer)
• Mapping - Mapping from one object to another
• - Domain range mapping (e.g.,
  has sender mapping from letter to person)
• Social - Concept describes relationship (e.g.,
  Jon has GPA 3.75)
• Procedure - Operation/methods (e.g., method for
  creating or destroying objects)

11

• Level 3 Open Social Set - classified a
  collection of objects (e.g., applicants)
• Closed Social Set (e.g., courses)
• Composed Mapping - virtual objects
  derived by composing other objects
• Composed Procedure - virtual object derived
  by grouping a collection of
  other procedure objects

12
Evolution System Architecture

• ICE Intelligent Concept Evolver
• How, what when to evolve
• PKM Personal Knowledge Manager
• LFI Learning From Instruction (passive learning)
• Initiated by user
• Acquire knowledge from user
• Dialogue-based
• Involved some inference
• LFE Learning From Exceptions (passive learning)
• Learning is triggered by
• Violations of constraints
• Definition of the object in the DB
• Involved more inference than LFI

13
Evolution System Architecture (contd)

• LFO Learning From Observation (active learning)
• Check if any evolution should be conducted
• Discovery of target evolution
• Creation of classification criteria for set
  object and similar tasks
• LFO is automatic, more inference than LFI and LFE

14
Examples

• 1) Find GPA of all students in the third world
  country. Since third world country is not
  defined in the schema, ICE conducts LFI (learning
  from instruction)
• Acquire the concept of third world country given
  the member of the country and ask the user to
  identify if they are a third world country.
• ICE learning all the third world countries in the
  DB introduces a new object countries-class with
  member third world, second world, and first
  world. Countries evolved to countries class.
• LFO (learning from observation) is triggered to
  find mapping cardinality constraint of has
  class, also determines if it is a many-to-one
  mapping.
• 2) CS 10 -gt CS 10a, 10b, 10c
• CS 10 violation! ICE calls LFE and LFI
• LFI learns there are 10a, 10b, 10c (same course
  name, unit number, instructor, and students)
• Students of each section should be proper subset
  of original student set.

15

• 3) Honor student has GPA gt 3.75
• Honor students can also be selected based on
  recommendation letters from Professors but GPA lt
  3.75
• ICEs LFO (learning from observation) learning
  that every instance satisfies a generalization
  condition GPA gt 3.70
• ICE changes the honor students condition from
  GPAgt3.75 to 3.70

16
(No Transcript)
17
Examples Of Distributed Intelligent Database
Systems
18
Planning Scheduling Initiative (PI)

• Sponsored by the Defense Advanced Research
  Project Agency (DARPA) Rome Laboratories (RL)
• Goal
• Promote automated planning technology
• Smooth transition of the planning technology to
  operational planning and scheduling problems
• Approach
• Tier 1 Independent research effects
• develops automated reasoning capabilities that
  are relevant to PI
• Tier 2 Integrated Feasibility Demos
• Merging efforts in Tier 1 into a single system
• and single operational problem
• Sizes Involved more than 20 contractors

19
Operational Problems Focus

• Transportation Planning
• Provide transportation planner with an ability to
  rapidly generate/evaluate/modify transportation
  plan
• Initial version used in DESERT SHIELD
• Features
• Interconnected with a large number of
  Heterogeneous DBs (thousands)
• Distributed, Multi-Agent Architecture
  (delegation, negotiation)
• Knowledge-based (representative, acquisition)
• Dynamic
• Common Prototype Environment
• Visionary Demo

20
Database Requirements

• Distributed
• Heterogeneous
• Intelligent (KB/DB)
• Temporal Evolutionary
• Multimedia
• Security