TOWARDS MKDA: A DATA MINING SEMANTIC WEB SERVICE

Vincenzo Cannella, Giuseppe Russo and Roberto Pirrone

Universita’ degli Studi Palermo, Dipartimento Ingegneria Informatica - DINFO

Viale delle Scienze ed. 6 p.3, 90128 Palermo, Italy

Keywords:

Data Mining, Knowledge Discovery in Databases, Semantic Web Service, Medical Knowledge Discovery

Assistant, Knowledge Discovery Process.

Abstract:

Nowadays a huge amount of raw medical data is generated. These data, analyzed with data mining techniques,

could be used to produce new knowledge. Unluckily such tasks need skilled data analysts, and not so many

researchers in medical ﬁeld are also data mining experts. In this paper we present a web based system for

knowledge discovery assistance in Medicine able to advice a medical researcher in this kind of tasks. The

experiment speciﬁcations are expressed in a formal language we have deﬁned. The system GUI helps the user

in the their composition. The system plans a Knowledge Discovery Process (KDP). The KDP is designed on

the basis of rules in a knowledge base. Finally the system executes the KDP and produces a model as result.

The system works through the co-operation of different web services specialized in different tasks. The choice

of web services is based on the semantic of their functionalities, according to a common OWL ontology. The

system is still under development.

1 INTRODUCTION

In recent years the availability of huge medical data

collections has sometimes dramatically brought to

light the (in)ability to analyze them. Medical centers

have huge databases containing therapies, diagnoses

and personal data of their patients. Moreover, the au-

tomatic devices of relevant data acquisition, such as

MRI and PET, extract more and more accurate medi-

cal images of patients. Medical images are produced

in such a number that they can only be analyzed with

the help of complex systems. All these raw data could

be usefully investigated by medical researchers to ﬁnd

new knowledge. In this view a very important appli-

cation ﬁeld is the Knowledge Discovery in Databases

(KDD) that, according to (Fayyad et al., 1996), is de-

ﬁned as the “non-trivial process of identifying valid

novel, potentially useful and ultimately understand-

able patterns in data” . This task is brain-intensive.

It is usually designed by a human expert. Unluckily

the KDD techniques needs a speciﬁc skill, and usu-

ally doctors are not data mining experts. In this work

we propose Medical Knowledge Discover Assistant

(MKDA). It is a new tool that we are still develop-

ing, to help knowledge discovery process in medical

ﬁeld for non expert users in data mining techniques,

as doctors usually are. The system receives the formal

speciﬁcation of the medical experiment research, in-

cluding goals and the inputs characteristics. The user

should say “what” she wants, and not “how” to get it.

The system must plan and execute a suitable Knowl-

edge Discovery Process (KDP), designed according

to the user’s needs and the application domain. Fi-

nally, it returns the results. The interaction between

the user and the system must be designed carefully. A

not expert user must be free from the too technical as-

pects of the process, and she must be guided through

hits and helps.

The rest of the paper is arranged as follows. The next

paragraph describes the state of the art for Knowl-

edge Discovery Assistants. The third paragraph intro-

duces a new language to describe the speciﬁcations

of a medical experimental research. The forth one

presents the knowledge base that helps in construc-

tion of experiments. Then a new knowledgediscovery

workﬂow model is presented in the ﬁfth paragraph.

The sixth one describes the functionalities of MKDA

system followed by the system architecture in the sev-

enth paragraph. Finally, conclusions and future works

are reported.

129

Cannella V., Russo G. and Pirrone R. (2008).

TOWARDS MKDA: A DATA MINING SEMANTIC WEB SERVICE.

In Proceedings of the Fourth International Conference on Web Information Systems and Technologies, pages 129-134

DOI: 10.5220/0001521801290134

 SciTePress

2 THE STATE OF THE ART

In recent years many researches have been carried

out in KDD, with the aim of developing a tool able

to perform an autonomous data analysis. The in-

volved ﬁeld is essentially a combination of some as-

pects of many research areas such as knowledge based

systems, machine learning and statistics. In Mlt-

Consultant (Sleeman et al., 1995) the selection of a

machine learning method is made with the support of

a knowledge-based system. Mlt-Consultant chooses

the learning methods on the basis of the syntactic

properties of their inputs and outputs according to a

set of rules. Another approach is the meta-learning

approach. NOEMON (Kalousis and Hilario, 2001)

relies on a mapping between dataset characteristics

and inducer performance to propose inducers for spe-

ciﬁc dataset steps. The most appropriate classiﬁer for

a dataset is suggested on the basis of the similarity of

the dataset with existing ones and on the performance

of the classiﬁer for the latter.

The DM Assistant System(Charest et al., 2006)

used the case-based reasoning. It has a collection of

pre-deﬁned cases. Every time, the system compares a

new case with this collection of cases, and establishes

the most similar one. This system has been inspired

to the CRISP-DM model (cri, 2000).

Another approach is related to the possibility to

build the entire process needed to achieve the goal.

As an example the IDEA System (Bernstein et al.,

2005) starts from characteristics of the data and of

the desired mining result. Then it uses an ontology

to search for and enumerate the data mining processes

that are valid for producing the desired result from the

given data. Each search operator corresponds to the

inclusion in the DM process of a different data min-

ing technique. In this ﬁeld some commercial tools as

IBM DB Intelligent have also been built. Miner (Han

et al., 1996) integrates a relational database system,

a Sybase SQL server, with a concept hierarchy mod-

ule, and a set of knowledge discovery modules. An-

other commercial tool is Clementine (Engels, 1996)

(Wirth et al., 1997). In this system the user-guidance

module uses a task/method decomposition to guide

the user through a stepwise reﬁnement of a high-level

data mining process. Important issues in this ﬁeld are

open source. Yale (Eliassi-Rad et al., 2006) is an en-

vironment for machine learning experiments and data

mining. It supports the paradigm of rapid prototyp-

ing. Yale provides a rich variety of methods which al-

lows rapid prototyping for new applications. Yale can

be used mainly by users skilled in KDD. Yale uses

Weka (Witten and Frank, 1999), a collection of Java

implementations of machine learning algorithms. The

preparation of data is supported in Yale by numerous

feature selection and construction operators. How-

ever, Yale is applied to a single input data table. The

Mining Mart software (Euler, 2005) can be used to

combine data from several tables, or to prepare large

data sets inside a relational database instead of main

memory as in Yale. Mining Mart also provides oper-

ators that ease the integration with Yale.

3 MEDICAL EXPERIMENTAL

RESEARCH SPECIFICATION

The medical researcher has to deﬁne the experiment

formally, setting its speciﬁcation. She has to list the

collection of features of her research, as data, goals,

metrics and models.

We have deﬁned an XML-compliant experiment

speciﬁcation language, a suitable formal language to

describe the inputs. Formally the deﬁned language

ESL (Experiment Speciﬁcation Language) is used to

represent: the set of data useful to problem deﬁnition,

the set of user goals, the set of metrics used to evalu-

ate the process results and the representation used for

results. Composing such a description could be too

hard for a not expert user. To solve this problem, we

have designed and implemented a very simple GUI

interface. The user can express her needs graphically,

and the GUI composes automatically the correspon-

dent description of the speciﬁcation. The deﬁnition

of possible goals drives the entire process of data dis-

covery. The goal deﬁnition makes possible the ﬁtting

of user choices with system capabilities. Some of the

most important goals in data mining are represented

in the following list.

• Association Analysis: it deﬁnes the process of

ﬁnding frequent and relevant patterns in terms of

composition rules;

• Correlation Analysis: it is used to deﬁne the de-

gree of relation in the association analysis. The

correlation analysis gives a measure of the cor-

rectness degree of the association;

• Classiﬁcation: given a certain number of at-

tributes useful to identify a class, the classiﬁcation

goal is used to ﬁnd a model describing the situa-

tion;

• Prediction: the goal is the same of the classiﬁca-

tion but inputs are continuous;

• Relevance Analysis: it is used to deﬁne which the

relevant patterns are to describe a certain model

useful to aggregate data

WEBIST 2008 - International Conference on Web Information Systems and Technologies

130

• Cluster Analysis: it is used to classify data not

previously classiﬁed into clusters;

• Outlier Analysis: it follows the cluster analysis

and is used to estimate which are the characteris-

tics of not included data in the clusterization pro-

cess;

• Evolution Analysis: deﬁnes the time or space

data evolution in terms of a model that represents

changes.

Due to complexity of processes many possible

metrics to evaluate the system have to be used. It’s

possible to distinguish them in terms of computation

load, usefulness of new founded patterns, novelty.

Measures are mostly related to particular data min-

ing algorithms or tasks. In fact, they are in direct re-

lation to goals that user wants to obtain. The same

considerations are also valid for the set of possible

task-dependent representations. The input data are

of different types: numerical data, categorical data,

complex symbolic descriptions, rules. A deeper dif-

ferentiation of data is in relation to data composition.

Three different input classes have been deﬁned. The

ﬁrst is the object class: data matrix, dissimilarity ma-

trix, single values, graphs are some possible exam-

ples. The second is the special input class, as, for in-

stance, numerics, dates, therapy, diagnoses, diseases,

patients, counts, IDs, binary data that are used for im-

ages or videos, texts and documents. The third class is

the variable type class like internal variables, symmet-

ric or asymmetric binary variables, discrete variables,

continue variables, scaled variables. The type of in-

put data is a ﬁrst factor to discriminate the possible

choices in the design of the workﬂow.

4 THE KNOWLEDGE BASE

The Knowledge Base of the system supports the gen-

eration of a complete experiment starting from user

requests expressed in a formal language. The knowl-

edge base is built in a modular way and is organized

in two main levels. In the highest level there is the

deﬁnition of concepts that are used for the experi-

ments. The concepts have been grouped in relation to

the roles they have in the KDP. The knowledge base

is a composition of different aspects. Some key terms

involved in the process have to be deﬁned. An ex-

periment is the composition of a workﬂow, a model, a

set of evaluations about the model ﬁtting with initial

problem and a representation of the obtained model to

obtain a possible goal. A model is a formal and well

deﬁnition speciﬁcation of the result obtained from an

experiment over some particular data. A model is ob-

tained through a sequence of steps in a workﬂow. The

workﬂow steps are essentially grouped in three aggre-

gates: data pre-processing, data mining process and

data post-processing. Also the composition rules have

been added inside the knowledge bases. Rules deﬁne

the workﬂow steps sequence and the choice method

to select operators for a particular step.

The design of the Knowledge Base has been in-

spired mainly to the frames. The operators are de-

ﬁned as frame. In general, also on the basis of the

paradigm of OWL-S, each operator has four princi-

pal characteristics. It can receive an input, produce

an output, be activated under certain constraints, and

change the general conditions of the process at the

end of its execution. The operators are organized ac-

cording to a taxonomy. If an operator belong to a

certain parental line, it inherits the characteristics of

its ancestors, but they have been redeﬁned. Opera-

tors in the Knowledge Base are classiﬁed according

to the step of he workﬂow in which they are involved,

too. In particular,there are: data generation operators,

I/O operators, pre-processing operators, mining oper-

ators, post-processing operators, validation operators,

and visualization operators. This classiﬁcation helps

the expert system during the planning phase, reduc-

ing the search space. Inputs and outputs have been

classiﬁed and organized according a taxonomy too.

There are many different input-output operators, ac-

cording to the type of the data source or to the ﬁle for-

mat of the data ﬁle. Many different types of outputs

are possible. Some operators manages data. They can

change the content or the structure of the data input.

Other operators are involved in the generation of the

model resulted from the data mining. There are many

possible models. They are classiﬁed into: Bayesian

ones, neural nets, numerical classiﬁer, numerical re-

gression and prevision models, rules, trees. Each of

these classes is divided into sub-classes. The expert

system can choose an operator on the basis of the

model it is able to produce. To deﬁne domain struc-

ture an OWL-Dl (owl, 2004) ontology has been built.

As previously seen, it is possible to split the ontology

in different sub-ontologies. The links between the el-

ements in the same sub-ontology are homogeneous

and deﬁnes structural properties. The links through

different sub-ontologies deﬁne the relations between

different types of elements.

5 OUR KDD WORKFLOW

MODEL

Knowledge discovery in database can be planned as a

process consisting of a set of steps. The sequence of

TOWARDS MKDA: A DATA MINING SEMANTIC WEB SERVICE

131

Figure 1: The workﬂow of a Knowledge Discovery Process.

Figure 2: The system architecture and the data ﬂow.

these steps is described with a workﬂow of the pro-

cess. We designed a general workﬂow model as syn-

thesis of different workﬂows described by literature.

Because most of them cover only partially the knowl-

edge discovery process, we tried to redesign a more

general workﬂow (see ﬁgure 1).

The workﬂow is the combination of some possi-

ble phases. Phases in the workﬂow must not be nec-

essarily followed linearly. The process may contain

loops. The workﬂow can be divided into main macro-

phases, in their turns divided into sub-phases: prob-

lem analysis and speciﬁcations deﬁnition, choice of

the tasks and their execution, results analysis. These

three phases are recognizable at a very high view of

the process. During the ﬁrst phase, the analyst inter-

acts with the user to understand her needs and expec-

tations to build consequently effective experiments.

The analyst must explicit objectives and knowledge

discovery goals clearly and correctly, or the entire

process design could be wrong. For this reason the

ﬁrst phase is particularly tricky. To encompass prob-

lems related with this phase, as just said, we devel-

oped a particular problem deﬁnition language. We

wanted to make the problem deﬁnition simpler espe-

cially for all not very skilled users. The user must

deﬁne evaluation criteria too. On the basis of these

criteria ﬁnal results of the workﬂow can be evaluated

to establish the satisfaction of user’s needs. The sec-

ond phase of the workﬂow is tightly related to the ﬁrst

one. It deals with the deﬁnition of relevant knowledge

on the application domain. This knowledge is used

by the analyst to extract some domain-driven process

choices. At the end this phase is documented, and a

possible loop is sometimes necessary. In data access

phase, the user deﬁnes the speciﬁc characteristics of

the dataset that must be mined. This phase often needs

to consult different sources and bring together data

in a common format with consistent deﬁnitions for

ﬁelds and keys. Collected data could contain either

too much, less or irrelevant information. These prob-

lems are solved during the preparation phase, before

the application of the modeling and discovery tech-

niques. Data transformation mainly is performed in

two ways: horizontally (changing the dimensionality

of the data) and vertically (changing the number of

data items). The preparation phase is usually the most

time and hardware resources consuming one. In re-

turn for this computation load, this phase makes pos-

WEBIST 2008 - International Conference on Web Information Systems and Technologies

132

sible saving resources in the next phases of the work-

ﬂow and getting better results. The preparation phase

can be split into four different sub-phases: data selec-

tion, data cleaning, missing values handling and data

reduction. None of these phases is mandatory, and the

execution of the preparation phase can contain many

loops.

During the data mining phase data are analyzed

through the chosen techniques. The application of

data mining techniques requires parameters calibra-

tion to optimal values. Therefore, another data prepa-

ration and transformation step is often needed. After

the model has been produced, it is converted in an au-

tonomous application able to implement the model.

The deployment phase deals with this task. Some-

times, the model can be emulated directly through the

development tool, which it has been developed with.

Other times, it is necessary to implement a new ap-

plication in a speciﬁc programming language. The

evaluation phase establishes how the model is suitable

for user’s needs according to the success and evalua-

tion criteria speciﬁed in the ﬁrst phase. If the system

satisﬁes user’s requests, the entire workﬂow can be

recorded into a repository. In this way, it can be em-

ployed again in similar tasks.

6 SYSTEM FUNCTIONALITIES

We are developing a Web intelligent system able to

analyze data in an experiment, and to design a knowl-

edge discovery process in those data to extract new

knowledge from them. The inputs of the system are

the data, the preferences of the user and the domain

knowledge regarding the problem treated with the ex-

periment. The system has two outputs: the model

describing the new knowledge mined from the data,

and the workﬂow applied to get the model. The lat-

ter one can be recorded in a repository and re-used in

new similar tasks. Initially the system must be able

to analyze data evaluating the presence of problems

like noise or missing data. The system must resolve

these problems to get data that can be used for the

construction of the model. It must reach a trade-off

between the accuracy and the time cost. The charac-

teristics of the produced model should be chosen by

the user. The choices of the system are not manda-

tory. The user can change some parts of the work-

ﬂow to get a different result. In other cases, the sys-

tem can propose different possible workﬂows and the

user chooses what she prefers. In these cases, the user

can choose to try many different workﬂows to ﬁnd

the best one. The user can interact with the system in

two different ways. As just said, we have developed

a problem deﬁnition language. The user can deﬁne

the experiment through this language. This function-

ality has been developed for expert users or repeti-

tive processes. On the other hand, the system has

a simple GUI which allows the user to deﬁne easily

the problem and that compose automatically the de-

scription. The interaction process has been inspired

to programmable interaction with users like in chat-

bot systems. Unlike such systems, our interaction is

graphical. In particular, we refer to ALICE chatbot

(ali, ), a system that owns a repository composed of

question-answer patterns which are called categories.

These categories are structured with the Artiﬁcial In-

telligence Markup Language an XML-compliant lan-

guage. The dialogue is based on algorithms for auto-

matic detection of patterns in the statements. Our in-

teraction process functionalities are developed in the

same manner: the couples question-answer are cat-

egorized and through this mechanism is possible to

have a tight interaction. This interaction allows the

system to collect information about both tasks and

needs of the users.

7 SYSTEM ARCHITECTURE

MKDA system has been designed according to the

well-known web service architecture in conjunction

with the client-server three-tier one (see ﬁgure 2).

The core application is executed on an HTTP

server. The client connects to this server remotely.

Through an intuitive GUI she composes the charac-

teristics of the experiment. The interface of the client

has been developed using the AJAX technologies,

which stands for Asynchronous JavaScript and XML.

The web pages developed through this technology are

more ﬂexible, and can easily and quickly be reconﬁg-

ured on the basis of the content that has to be shown

or on the basis of the user interaction. The interface

dynamically composes the description of the request

of the user in the language described previously. At

this point the request is sent to the server. After-

wards, the request is sent to the workﬂow generator

web service (WGWS), which analyzes this request,

and constructs the correspondent knowledge discov-

ery workﬂow. At this aim, it queries the knowledge

base module. At the same time, it consults the Ex-

periment Comparator Service, which, on the basis of

a case-based reasoning, lists all past experiments in

the repository matching the user’s requests. Then it

consults the data mining tasks catalogue managed by

the tasks catalogue web service (TCWS). TCWS ad-

vertises the list of tasks that the system can employ.

These tasks are the basic ones that make a workﬂow.

TOWARDS MKDA: A DATA MINING SEMANTIC WEB SERVICE

133

This list is very long. It includes all operators of li-

braries and systems as, for instance, Weka, Yale and

Ptolemy. These three systems have been embedded

into web services, which make possible to use re-

motely their functionalities. The tasks are advertised

in WSDL. The data retrieved in this way are inserted

into the description of each step of the ﬁnal work-

ﬂow. The workﬂow is generated automatically by the

expert system through a planning process. The sys-

tem must evaluates the possible actions, and plan a

sequence of actions able to produce the desired goal.

The choices in the planning phase are related to the

characteristics of the actions. The planner links ac-

tions together, matching the data ﬂowing in the work-

ﬂow. The system must bind these services with the

steps of the workﬂow. Actually, the binding is syntac-

tical, based on a shared ontology. The system matches

the functionalities of each step to the functionalitiesof

the services in the web, and produces a description of

how the workﬂow should be executed. During the ex-

ecution of the workﬂow, these data are used to know

where each task has to be executed. After the work-

ﬂow has been generated, it is sent to the workﬂow ex-

ecutor web service (WEWS). It manages and coordi-

nates the steps of the workﬂow. Each step is executed

resorting to the Yale, Weka and Ptolemy web services.

These services can access the database, and return the

result of the execution of the workﬂow as model. The

model is returned by the WEWS to the application on

the HTTP server and sent to the client as reply to its

initial request. The model together with the workﬂow

can be recorded into the repository. They form an ex-

periment. The collection of experiments can be con-

sulted. In this way the user can eventually re-employ

a past workﬂow when she must work with a similar

experiment.

8 CONCLUSION AND FUTURE

WORKS

In this work we have proposed a new web based sys-

tem to help knowledge discovery in medical ﬁeld for

non expert users. We have described system archi-

tecture and functionalities. The system is able in a

very simple manner to collect the characteristic of

the treated experiments. Then a knowledge discovery

worﬂow is generate according to the workﬂow model

we have designed. Finally the system is able to ex-

ecute the workﬂow and produces a model as result.

The user should concentrate on the speciﬁcation of

the problem, while most of the implementation should

be delegated to the system. The system is under devel-

opment yet. The web infrastructure and the workﬂow

model has been realized and future work is focused

on its whole development and test.

REFERENCES

Alice. http://www.alicebot.org/.

(2000). Crisp-dm. http://www.crisp-dm.org/.

(2004). Owl web ontology language use cases and require-

ments.

Bernstein, A., Provost, F. J., and Hill, S. (2005). Toward

intelligent assistance for a data mining process: An

ontology-based approach for cost-sensitive classiﬁca-

tion. IEEE Trans. Knowl. Data Eng., 17(4):503–518.

Charest, M., Delisle, S., Cervantes, O., and Shen, Y. (2006).

Invited paper: Intelligent data mining assistance via

cbr and ontologies. In DEXA ’06: Proceedings of the

17th International Conference on Database and Ex-

pert Systems Applications, pages 593–597, Washing-

ton, DC, USA. IEEE Computer Society.

Eliassi-Rad, T., Ungar, L. H., Craven, M., and Gunopu-

los, D., editors (2006). Proceedings of the Twelfth

ACM SIGKDD International Conference on Knowl-

edge Discovery and Data Mining, Philadelphia, PA,

USA, August 20-23, 2006. ACM.

Engels, R. (1996). Planning tasks for knowledge discovery

in databases; performing task-oriented user-guidance.

In KDD, pages 170–175.

Euler, T. (2005). Publishing operational models of data min-

ing case studies. In Proc. of the Workshop on Data

Mining Case Studies at the 5th IEEE International

Conference on Data Mining (ICDM), page 99.

Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P., and Uthu-

rusamy, R., editors (1996). Advances in Knowledge

Discovery and Data Mining. AAAI/MIT Press.

Han, J., Fu, Y., Wang, W., Chiang, J., Gong, W., Koperski,

K., Li, D., Lu, Y., Rajan, A., Stefanovic, N., Xia, B.,

and Za¨ıane, O. R. (1996). Dbminer: A system for

mining knowledge in large relational databases. In

KDD, pages 250–255.

Kalousis, A. and Hilario, M. (2001). Model selection via

meta-learning: A comparative study. International

Journal on Artiﬁcial Intelligence Tools, 10(4):525–

554.

Sleeman, D. H., Rissakis, M., Craw, S., Graner, N., and

Sharma, S. (1995). Consultant-2: pre- and post-

processing of machine learning applications. Int. J.

Hum.-Comput. Stud., 43(1):43–63.

Wirth, R., Shearer, C., Grimmer, U., Reinartz, T. P.,

Schl¨osser, J., Breitner, C., Engels, R., and Lindner,

G. (1997). Towards process-oriented tool support

for knowledge discovery in databases. In PKDD

’97: Proceedings of the First European Symposium on

Principles of Data Mining and Knowledge Discovery,

pages 243–253, London, UK. Springer-Verlag.

Witten, I. H. and Frank, E. (1999). Data Mining: Practi-

cal Machine Learning Tools and Techniques with Java

Implementations. Morgan Kaufmann.

WEBIST 2008 - International Conference on Web Information Systems and Technologies

134