Architecture for PatternBase Management Systems
Manolis Terrovitis
National Technical Univ. of Athens,
Dept. of Electrical and Computer Eng.,
Athens, Hellas
[email protected]
Panos Vassiliadis
University of Ioannina,
Dept. of Computer Science,
Ioannina, Hellas
[email protected]
Outline
• Introduction and Motivation
• “Manifesto”-User requirements
• PBMS architecture
2
Outline
• Introduction and Motivation
• “Manifesto”-User requirements
• PBMS architecture
3
The Problem
• Vast amounts of raw data.
• Data mining is not enough to extract
knowledge.
• The core of the problem:
– Keeping, Storing and Manipulating knowledge
from raw data.
4
Knowledge as Patterns
“Patterns are compact and rich in
semantics representations of raw data “
5
The need for a new PBMS
• Current DBMS’s cannot address the new
user requirements patterns impose.
• We need a system to provide powerful and
specialized manipulation abilities for
patterns.
• We need a Pattern Base Management
System.
6
Main topics
• A “manifesto”-like list of requirements for a
PBMS.
• A reference architecture for a PBMS.
• Discussion of the differences between
ORDBMS’s and PBMS.
7
Outline
• Introduction and Motivation
• “Manifesto”-User requirements
• PBMS architecture
8
Requirements for the PBMS
• User requirements for the Data Model.
• User requirements for the PBMS
architecture.
• User requirements for the query language
and processing.
9
Requirements for the Data Model
• Compact and Rich in semantics.
• Based upon a generic uniform model that
covers all kinds of patterns.
• Support different types of patterns in an
extensible fashion.
• Allow semantically similar patterns to be
identified.
• Support composite patterns, generated
from simpler ones.
10
Requirements for the Architecture
•
•
•
•
Mechanisms for representing and storing
its entries.
Cooperation with DBMS’s storing raw
data.
Ability to manage pattern extraction and
creation.
Access to intermediate results of pattern
creation algorithms.
11
Requirements for Query language
and Processing
• A PBMS processes different kinds of
queries, possibly even on raw data and
returns more intuitive results to users.
• A PBMS employs a query language which
can at least perform the following tasks:
– Pattern matching
– Logical inferences
– Meta queries
12
Outline
• Introduction and Motivation
• “Manifesto”-User requirements
• PBMS architecture
13
PBMS
Architecture
PBMS
Class Layer
PatternsOf
Experiment145
MyClustersOn
TableEMP
Ass.
Rules
Type
Member of
Type Layer
Dec.
Trees
Type
Cluster
Type
Primary
Key
Type
Cyclical
Cluster
Type
Instance of
• PBMS
– Pattern Layer
– Pattern Type Layer
– Class Layer
Patterns
Layer
DBSCAN
Cluster m
DBSCAN
PK #1
Ass. Rule
DBSCAN
Cluster 2
n
Cluster 1
Ass. Rule
Dec. Tree
Ass. Rule
1
#1
2
• Intermediate Results
Intermediate Mappings
• RDBMS
– Raw data
Pattern
Recognition
Algorithms
Data Mining
Algorithms
Raw
Data
Layer
DB1
DB2
Flat file
14
Pattern Type Layer
• Pattern types: describe the syntax of the
patterns.
– Structure
– Declarative specification
• Patterns are instantiations of Pattern
Types, which follow the structure of the
types.
• Pattern Type layer is extensible.
15
Class Layer
• Classes: collections of patterns which
share some semantic property.
• Patterns of a certain class must all be
instances of the same type.
• Classes are used to create patterns with
predefined semantics given by the
designer.
• Classes are defined over features.
16
Is it really different from an
OO/ORDBMS?
•
•
Semantically rich representation of patterns.
Novel querying requirements:
–
–
–
•
Ad hoc operations over the source and pattern
spaces and their mapping
Pattern matching
Reasoning facilities based on the declarative
specification of patterns
Alternative storage, indexing and query
optimization techniques.
17
Summary
• A “manifesto”-like list of requirements for a
PBMS
• A reference architecture for a PBMS
• Discussion of the differences between
ORDBMS’s and PBMS
18
Descargar

Architecture for Pattern-Base Management Systems