Multi-RelationalDataMiningusingUMLforILP1资源-CSDN文库

数据挖掘

需积分: 3 119 浏览量 2008-04-16 21:41:29 上传评论收藏 141KB PDF 举报

资源推荐

资源详情

资源评论

Multi-Relational Data Mining, using UML for ILP

Arno J. Knobbe

, Arno Siebes

, Hendrik Blockeel

, Daniël van der Wallen

Kiminkii, Postbus 171, 3990 DD Houten, The Netherlands,

a.knobbe@kiminkii.com

CWI, P.O. Box 94079, 1090 GB Amsterdam, The Netherlands

arno@cwi.nl

K.U.Leuven, Dept. of Computer Science, Celestijnenlaan 200A, B-3001 Heverlee, Belgium

hendrik.blockeel@cs.kuleuven.ac.be

Dantec B.V., Pontanusstraat 21, 1093 RX Amsterdam, The Netherlands

daniel@dantec.nl

Abstract.

Although there is a growing need for multi-relational data

mining solutions in KDD, the use of obvious candidates from the field

of Inductive Logic Programming (ILP) has been limited. In our view

this is mainly due to the variation in ILP engines, especially with

respect to input specification, as well as the limited attention for

relational database issues. In this paper we describe an approach which

uses UML as the common specification language for a large range of

ILP engines. Having such a common language will enable a wide range

of users, including non-experts, to model problems and apply different

engines without any extra effort. The process involves transformation

of UML into a language called CDBL, that is then translated to a

variety of input formats for different engines.

1 Introduction

A central problem in the specification of a multi-relational data mining problem is the

definition of a model of the data. Such a model directly determines the type of

patterns that will be considered, and thus the direction of the search. Such

specifications are usually referred to as declarative or language bias in ILP [14]. Most

current systems use logic-based formalisms to specify the language bias (e.g., Progol,

S-CART, Claudien, ICL, Tilde, Warmr [13, 12, 6, 7, 3, 8]). Although most of these

formalisms are quite similar and make use of the same concepts (e.g., types and

modes), there are still differences between the formalisms that make the sharing of the

language bias specification between engines a non-trivial task. The main reasons for

this are:

•

the different formalisms each have their own syntax; the user needs to be familiar

with all of them

First published at PKDD 2000.

•

many formalisms contain certain constructs, the semantics of which, sometimes

in a subtle way, reflect behavioral characteristics of the inductive algorithm.

The use of different ILP-systems would be simplified significantly if a common

declarative bias language were available. Such a language should have the following

characteristics:

•

The common language should be usable for a large range of ILP systems, which

means that it should be easy to translate a bias specification from the common

language to the native language of the ILP system

•

It should be easy to learn. This means it should make use of concepts most users

are familiar with. In the ideal case, the whole language itself is a language that

the intended users are familiar with already

•

The bias should not just serve as a necessary prerequisite for running the

induction algorithm, but should also be usable as a shared piece of information or

documentation about a problem within a team of analysts with varying levels of

technical expertise

•

It should be easy to judge the complexity of a problem from a single glance at the

declarative bias. A graphical representation would be desirable.

In this paper we propose the use of the Unified Modeling Language (UML) [2, 15, 16,

17] as the language of choice for specifying declarative bias of such nature. Over the

past few years UML has proven itself as a versatile tool for modeling a large range of

applications in various domains. For ILP the Class Diagrams with their usefulness in

database modeling are specifically interesting. Our discussion will be based on these

diagrams.

Why do we wish to use UML to express bias? First of all, as UML is an intuitive

visual language, essentially consisting of annotated graphs, we can easily write down

the declarative bias for a particular domain or judge the complexity of a given data

model [9]. Another reason for using UML is its widespread use in database (as well as

object oriented) modelling. UML has effectively become a standard with thorough

support in many commercial tools. Some tools allow the reverse engineering of a data

model from a given relational database, directly using the table specifications and

foreign key relations. If we can come up with a process of using UML in ILP

algorithms, we would then have practically automated the analysis process of a

relational database. Finally, UML may serve as a common means of stating

declarative bias languages used in the different ILP engines.

Although it is clear that UML is a good candidate for specifying first order

declarative bias, it may not be directly clear how the different engines will actually be

making use of the UML declarations. Its use in our previously published Multi-

Relational Data Mining framework [10, 11] is straightforward, as this framework and

the related engine Fiji2 have been designed around the use of UML from the outset.

To translate UML bias declarations to logic-based bias languages, we use an

intermediate textual representation, called Common Declarative Bias Language

(CDBL). CDBL is essentially a set of Prolog predicates, which can be easily

processed by the different translation procedures. Translation procedures for the

popular engines Tilde, Warmr and Progol are currently available. The whole process

of connecting an ILP engine to a relational database now becomes a series of

translation steps as is illustrated by the diagram in figure 1. We have implemented and

embedded each of these steps into a single GUI.

The investigation of UML as a common declarative bias language for non-experts

was motivated by the efforts involved in the Esprit IV project Aladin. This project

aims at bringing ILP capabilities to a wider, commercial audience by embedding a

range of ILP algorithms into the commercial Data Mining tool, Clementine.

The outline of this paper is as follows. A section describing UML and its potential

as first order declarative bias follows this introduction. We then give a short overview

of the syntax of the Common Declarative Bias Language. In

Translating CDBL

give an algorithm for translating CDBL to ILP input. Next we analyze the usefulness

of UML as a declarative bias language compared to other approaches in

Comparing

UML to traditional languages

. This section is followed by a

Conclusion

2 UML

From the large set of modelling tools provided by UML we will focus on the richest

and most commonly used one: Class Diagrams [16]. These diagrams model exactly

the concepts relevant for ILP, namely tables and the relation between them. In fact

when we write UML in this paper we are referring specifically to Class Diagrams.

There are two specific concepts within the Class Diagrams that we will be focusing

on. The first is the concept of

class.

A class is a description of a set of objects that

share the same features, relationships, and semantics. In a Class Diagram, a class is

represented as a rectangle. Typically, a class represents some tangible entity in the

problem domain, and maps to a table in the database.

The second concept is that of

association.

An association is a structural

relationship that specifies that objects of one class are connected to objects of another.

An important aspect of an association is its

multiplicity.

This specifies how many

Figure 1 The complete process of using UML with existing engines.

剩余11页未读，继续阅读

评论收藏

内容反馈

morre

粉丝: 187
资源: 2337

Multi-Relational Data Mining using UML for ILP 1

最新资源

Multi-Relational Data Mining using UML for ILP 1

Speeding Up Multi-Relational Data Mining

Multi-Relational Data Mining in Medical Databases

Prospects and Challenges for Multi-Relational Data Mining

Numbers in Multi-Relational Data Mining

multi-relational data mining

Introduction to multi-relational data mining

Multi-relational data mining in Microsoft SQL Server 2005

A Relational Tucker Decomposition for Multi-Relational Link Prediction.pdf

Community Mining from Multi-relational

A-Relational-Model-of-Data-for-Large-Shared-Data-Banks.pdf

COMPOSITION-BASED MULTI-RELATIONAL GRAPH CONVOLUTIONAL NETWORKS

南开大学数据库原理课件lecture2-Relational-Data-Model.ppt

前端项目-backbone-relational.zip

A-Relational-Model-of-Data-for-Large-Shared-Data-Banks

pagerankmatlab代码-Link_Prediction_in_Multi-relational_Networks:Link_Pred

HIBERNATE - Relational Persistence for Idiomatic Java

HIBERNATE - Relational Persistence for Idiomatic Java.chm

Spark SQL- Relational Data Processing in Spark(Paper).rar

第一讲：python网络爬虫基础知识

Python基于机器学习实现的股票价格预测、股票预测源码+数据集，机器学习大作业

身份证前6位对应的省市区代码（超详细）

抖音用户浏览行为数据集

两阶段鲁棒优化/综合能源系统/需求响应/微电网/多目标优化/优化调度matlab-yalmip-cplex/gurobi文章复现

SPSS中介效应分析插件（Process和mediate插件）

基于在线教学平台的数据挖掘与学习行为分析超星集团数据集

regress函数实例代码

Fragstats V4.2 软件计算景观指数的参数文件示例

新闻数据集（对应新闻文本分类案例）

最新资源