Basic Approaches Of Integration Between Data
Warehouse And Data Mining
Ina Naydenova, Kalinka Kaloyanova
“St. Kliment Ohridski” University of Sofia,
Faculty of Mathematics and Informatics
Sofia 1164, Bulgaria
ina@fmi.uni-sofia.bg,
kkaloyanova@fmi.uni-sofia.bg
Abstract. Data Warehouse and OLAP are essential elements of decision support, which has increasingly become a focus of the database industry. On the
other hand, Data Mining and Knowledge Discovery is one of the fast growing
computer science fields. Traditional query and report tools have been used to
describe and extract what is in a database. The user forms a hypothesis about a
relationship and verifies it with a series of queries. Data mining can be used to
generate an hypothesis. But there is some other possibilities for collaboration
between Data Warehouse and Data Mining technologies. We discuss several
approaches of such collaboration in different architecture levels.
1
Introduction
The role of information in creating competitive advantage for businesses and other
enterprises is now a business axiom: whomever controls critical information can
leverage that knowledge for profitability. The difficulties associated with dealing with
the mountains of data produced in businesses brought about the concept of
information architecture which has spawned projects such Data Warehousing
(DW)[1]. The goal of a data warehouse system is to provide the analyst with an
integrated and consistent view on all enterprise data which are relevant for the
analysis tasks. On the other hand, Data Mining and Knowledge Discovery is one of
the fast growing computer science fields. Its popularity is caused by an increased
demand for tools that help with the analysis and understanding of huge amounts of
data [3]. The question arises: ”Where is the intersection of these powerful technologies of the day?” The paper provides an overview of three basic approaches of integration and interaction between them:
1. Integration on the front end level combining On-Line Analytical Processing
(OLAP) and data mining tools into a homogeneous Graphic User Interface;
2. Integration on the database level adding of data mining components directly in DBMS;
3. Interaction on back end level – the usage of data mining techniques during
the data warehouse design process.
2
Ina Naydenova, Kalinka Kaloyanova
Initially, we present briefly the fundamentals of data warehouse, OLAP and data mining. Then we consider: (1) the main characteristics of On-Line Analytical Mining,
witch integrates OLAP with data mining knowledge; (2) the trends of inclusion of data mining engines in DBMS by manufacturers of popular relational DBMS as Oracle,
IBM and Microsoft; (3) the application of existing data mining techniques in support
of data warehouse building process.
2
Data Warehouse, OLAP and Data Mining
DW and OLAP are essential elements of decision support. They are complementary
technologies - a DW stores and manages data while OLAP transforms the data into
possibly strategic information. Decision support usually requires consolidating data
from many heterogeneous sources: these might include external sources in addition to
several operational databases. The different sources might contain data of varying
quality, or use inconsistent representations, codes and formats, which have to be reconciled. Since data warehouses contain consolidated data, perhaps from several operational databases, over potentially long periods of time, they tend to be orders of
magnitude larger than operational databases. In data warehouses historical and summarized data is more important than detailed. Many organizations want to implement
an integrated enterprise warehouse that collects information about all subjects spanning the whole organization. Some organizations are settling for data marts instead,
which are departmental subsets focused on selected subjects (e.g., a marketing data
mart, personnel data mart).
A popular conceptual model that influences the front-end tools, database design, and
the query engines for OLAP is the multidimensional view of data in the warehouse. In
a multidimensional data model, there is a set of numeric measures that are the objects
of analysis. Examples of such measures are sales, revenue, and profit. Each of the
numeric measures depends on a set of dimensions, which provide the context for the
measure. For example, the dimensions associated with a sale amount can be the store,
product, and the date when the sale was made. The dimensions together are assumed
to uniquely determine the measure. Thus, the multidimensional data views a measure
as a value in the multidimensional space of dimensions. Often, dimensions are hierarchical; time of sale may be organized as a day-month-quarter-year hierarchy, product
as a product-category-industry hierarchy [4]. Typical OLAP operations, also called
cubing operation, include rollup (increasing the level of aggregation) and drill-down
(decreasing the level of aggregation or increasing detail) along one or more dimension
hierarchies, slicing (the extraction from a data cube of summarized data for a given
dimension-value), dicing (the extraction of a “sub-cube”, or intersection of several
slices), and pivot (rotating of the axes of a data cube so that one may examine a cube
from different angles).
Other technology, that can be used for querying the warehouse is Data Mining.
Knowledge Discovery (KD) is a nontrivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns from large collections of data [6].
One of the KD steps is Data Mining (DM). DM is the step that is concerned with the
actual extraction of knowledge from data, in contrast to the KD process that is con-
Basic Approaches Of Integration Between Data Warehouse And Data Mining
3
cerned with many other things like understanding and preparation of the data, verification of the mining results etc. In practice, however, people use terms DM and KD
as synonymous.
ALL
Sofia
Toronto
STORE
ALL
PRODUCT
ALL
LAA
(ALL, ALL, ALL)
food
(ALL, books, Sofia)
books
sports
(ALL, clothing, ALL)
clothing
2001 2002 2003 2004
ALL TIME
Fig. 1. Cube presentation of data in multidimensional model
The knowledge discovery goals are defined by the intended use of the system. We can
distinguish two types of goals: verification and discovery. With verification, the system is limited to verifying the user’s hypothesis. That kind of functionality is mainly
supported by OLAP technology. With discovery, the system autonomously finds new
patterns. The two high-level primary discovery goals of data mining in practice tend
to be prediction and description. Prediction involves using some variables or fields in
the database to predict unknown or future values of other variables of interest, and description focuses on finding human-interpretable patterns describing the data. Data
mining involves fitting models to, or determining patterns from, observed data. Most
data-mining methods are based on tried and tested techniques from machine learning,
pattern recognition, and statistics: classification, clustering, regression, and so on
(there is an array of different algorithms under each of these headings)[5]. We introduce in brief only some of them having relation with the rest of the paper:
Classification is learning a function that maps (classifies) a data item into one of
several predefined classes. Classification-type problems are generally those where
one attempts to predict values of a categorical dependent variable (class, group
membership, etc.) from one or more continuous and/or categorical predictor variables.
Regression is learning a function that maps a data item to a real-valued prediction
variable. Regression-type problems are generally those where one attempts to predict the values of a continuous variable from one or more continuous and/or categorical predictor variables. There are a large number of methods that an analyst can
choose from when analyzing classification or regression problems. Tree classification and regression techniques, produce predictions based on few logical if-then
conditions is frequently used approach [16].
Clustering is a common descriptive task where one seeks to identify a finite set of
categories or clusters to describe the data. The categories can be mutually exclusive and exhaustive or consist of a richer representation, such as hierarchical or
overlapping categories. Closely related to clustering is the task of probability den-
4
Ina Naydenova, Kalinka Kaloyanova
sity estimation, which consists of techniques for estimating from data the joint multivariate probability density function of all the variables.
Dependency modeling consists of finding a model that describes significant dependencies between variables. Dependency models exist at two levels: the structural level of the model specifies (often in graphic form) which variables are locally dependent on each other and the quantitative level of the model specifies the
strengths of the dependencies using some numeric scale (e.g. probabilistic dependency networks).
Change and deviation detection focuses on discovering the most significant changes in the data from previously measured or normative values.
Summarization involves methods for finding a compact description for a subset of
data e.g. the derivation of summary or association rules and the use of multivariate
visualization techniques [6]. An association rule is an expression of the form A →
B, where A and B are sets of items. Normally, the result of data mining algorithms
for finding association rules consists of all rules that exceed pre-defined minimum
support (coverage) and have high confidence (accuracy). Association rule discovery techniques are generally applied to databases of transactions where each transaction consists of a set of items. In such a framework the problem is to discover all
associations and correlations among data items where the presence of one set of
items in a transaction implies (with a certain degree of confidence) the presence of
other items. The problem of discovering sequential patterns is to find intertransaction patterns such that the presence of a set of items is followed by another
item in the time-stamp ordered transaction set.
Monitoring & Administration
3
External sources
Operational
data bases
Meta Repository
2
1
OLAP
Servers
Analysis
Data Warehouse
Extract
Transform
Load
Refresh
Serve
Query/Reporting
Data Mining
Data Marts
Fig. 2. Typical data warehousing architecture and points of integration
3
Integration of OLAP with Data Mining
Most data mining tools need to work on integrated, consistent, and cleaned data,
which requires costly data cleaning, data transformation, and data integration as pre-
Basic Approaches Of Integration Between Data Warehouse And Data Mining
5
processing steps. A data warehouse constructed by such preprocessing serves as a
valuable source of high quality of data for OLAP as well as for data mining. Effective
data mining needs exploratory data analysis. A user will often want to traverse
through a database, select portions of relevant data, analyze them at different granularities, and present knowledge/results in different forms [8]. By integration of OLAP
and data mining, OLAP mining (also called On-Line Analytical Mining) facilitates
flexible mining of interesting knowledge in data cubes because data mining can be
performed at multi-dimensional and multi-level abstraction space in a data cube. Cubing and mining functions can be interleaved and integrated to make data mining a
highly interactive and interesting process. The desired OLAP mining functions are:
Cubing then mining: With the availability of data cubes and cubing operations,
mining can be performed on any layers and any portions of a data cube. This
means that one can first perform cubing operations to select the portion of the data
and set the abstraction layer (granularity level) before a data mining process starts.
For example, one may first tailor a cube to a particular subset, such as “year =
2004”, and to a desired level, such as at the “city level” for the dimension “store”,
and then execute a prediction mining module.
Mining then cubing: This means that data mining can be first performed on a data
cube, and then particular mining results can be analyzed further by cubing operations. For example, one may first perform classification on a “market” data cube
according to a particular dimension or measure, such as profit made. Then for each
obtained class, such as the high profit class, cubing operations can be performed,
e.g., drill-down to detailed levels and examine its characteristics.
Cubing while mining: A flexible way to integrate mining and cubing operation is to
perform similar mining operations at multiple granularities by initiating cubing operations during mining. By doing so, the same data mining operations can be performed on different portions of a cube or at different abstraction levels. For example, for mining association rules in a “market” data cube, one can drill down along
a dimension, such as “time”, to find new association rules at a lower level of abstraction, such as from “year to month”.
Backtracking: To facilitate interactive mining, one should allow a mining process
to backtrack one or a few steps or backtrack to a preset marker and then explore alternative mining paths. For example, one may classify market data according to
profit made and then drill-down along some dimension(s), such as store to see its
characteristics. Alternatively, one may like to classify the data according to another
measure, cost of product, and then do the same (characterization). This requires the
miner to jump back a few steps or backtrack to some previously marked point, and
redo the classification. Such flexible traversal along the cube at mining is a highly
desired feature for users.
Comparative mining: A flexible data miner should allow comparative data mining,
that is, the comparison of alternative data mining processes. For example, a data
miner may contain several cluster analysis algorithms. One may like to compare
side by side the clustering quality of different algorithms, even examine them when
performing cubing operations, such as when drilling down to detailed abstraction
layers.
It is possible to have other combinations in OLAP mining. For example one can perform “mining then mining”, such as first perform classification on a set of data and
6
Ina Naydenova, Kalinka Kaloyanova
then find association patterns for each class. In a large warehouse containing a huge
amount of data, it is crucial to provide flexibilities in data mining so that a user may
traverse a data cube, select mining space and the desired levels of abstraction, and test
different mining modules and alternative mining algorithms [7] .
4
An Integration of RDBMS with Data Mining
Traditionally, data mining tasks are performed assuming that the data is in memory.
In the last decade, there has been an increased focus on scalability but research work
in the community has focused on scaling analysis over data that is stored in the file
system. This tradition of performing data mining outside the DBMS has led to a situation where starting a data mining project requires its own separate environment. Data
is dumped or sampled out of the database, and then a series of special purpose programs are used for data preparation. This typically results in the familiar large trail of
droppings in the file system, creating an entire new data management problem outside
the database. Once the data is adequately prepared, the mining algorithms are used to
generate data mining models. However, past work in data mining has barely addressed the problem of how to deploy models, e.g., once a model is generated, how to
store, maintain, and refresh it as data in the warehouse is updated, how to programmatically use the model to do predictions on other data sets, and how to browse models. Over the life cycle of an enterprise application, such deployment and management
of models remains one of the most important tasks [9]. The objective of the integration on the database level is to alleviate problems in deployment of models and to
ease the data preparation by working directly on relational data. The business community already tries to integrate the KD process. The biggest DBMS vendors like
IBM, Microsoft and Oracle integrated some of the DM tools into their commercial
systems. Their goal is to make the use of DM methods easier, in particular for users
that use their DBMS products [3].
The IBM Solution
Data mining is a key component of an IBM's Information Warehouse framework. The
core of IBM Data Mining technology is the mining kernel. It’s contains modules that
perform data mining functions like classification, clustering, association, sequential
patterns [14].
IBM’s DM tool, called Intelligent Miner is a feature of the DB2 Universal Database
Data Warehouse Editions, which consists of three components: Intelligent Miner
Modeling, Intelligent Miner Visualization, Intelligent Miner Scoring. Intelligent Miner Modeling is a extension of the DB2 database that enables the development of customized mining applications. This modeling tool supports statistical functions and algorithms, including clustering, tree classification, regression, and associations
algorithm. When new relationships are discovered, Intelligent Miner Scoring allows
you to apply the new relationships in your data to new data in real-time. By isolating
both the mining model and the scoring logic (scoring is the process of predicting outcomes) from the application, you can provide continuous model improvement as
Basic Approaches Of Integration Between Data Warehouse And Data Mining
7
trends change or additional information becomes available - without disruption to the
application. Data-mining model analysis is available via Intelligent Miner Visualizer,
a Java-based results browser.
Another product - DB2 Intelligent Miner for Data provides a suite of tools to obtain a
single framework for database mining. This framework supports the iterative process,
offering data processing, statistical analysis, and results visualization to complement a
variety of mining methods [15].
The Microsoft Solution
SQL Server 2000 includes Analysis Services with data mining technology, which examines data, in relational data warehouse, as well as OLAP cubes to uncover areas of
interest to business decision makers and other analysts. Data mining models are exposed through a new OLE DB for Data Mining provider to calling applications and
reporting tools. OLE DB DM is an extension of the SQL query language that allows
users to train and test DM models. It has only two notions beyond traditional OLE DB
definition: cases and models. Input data is in the form of a set of cases (caseset). Data
mining model is treated as if it were a special type of “table:” we associate a caseset
with a model and additional meta-information while defining a model; when data is
inserted into the data mining model, a mining algorithm processes it and the resulting
abstraction is saved instead of the data itself. Once a model is populated, it can be
used for prediction, or its content can be browsed for reporting. Actually, OLE DB
DM is independent of any particular provider or software and is meant to establish a
uniform API. Microsoft SQL Server database management system exposes OLE DB
DM. The “core” relational engine exposes the traditional OLE DB interface. However, the analysis server component exposes an OLE DB DM interface (along with OLE
DB for OLAP) so that data mining solutions may be built on top of the analysis server
[9]. OLAP Data Mining allows architects to reuse the results from data mining and
incorporate the information into an OLAP Cube dimension for further analysis. In addition, the Decision Support Objects (DSO) library has been extended in order to accommodate direct programmatic access to the data mining functionality present within OLAP services [13]. Data mining algorithms are the foundation from which mining
models are created. Microsoft’s SQL Server 2000 incorporates two algorithms: classification trees and clustering. The algorithms available with new SQL Server 2005
(Beta) are much more: classification (decision) trees, clustering, Naïve Bayes (classification algorithm), time series (sequential patterns algorithm), association, neural
network (creates classification and regression mining models by constructing a multilayer network) [12].
The Oracle Solution
Oracle has integrated OLAP and data mining capabilities directly into the database
server. All the data mining functionality is embedded in Oracle9i Database, so the data, data preparation, data mining, and scoring, all occur within the database. Oracle9i
Data Mining (ODM) has two main components: Data Mining API and Data Mining
Server. Application developers access Oracle data mining's functionality through a
Java-based API. Programmatic control of all data mining functions enables automation of data preparation, model-building, and model-scoring operations [11]. The Data
Mining Server is the server-side, in-database component that performs the data mining operations within the 9i database, and thus benefits from its availability and scalability. It also provides a metadata repository consisting of mining input objects and
8
Ina Naydenova, Kalinka Kaloyanova
result objects [10]. ODM supports tree classification, clustering, associations and attribute importance (in fact only Naïve Bayes). Associated analysis for preparatory data exploration and model evaluation is extended by Oracle's statistical functions and
OLAP capabilities. Because these also operate within the database, they can all be incorporated into a seamless application that shares database objects. Models are stored
in the database and directly accessible for evaluation, reporting, and further analysis
by tools and application functions.
Oracle has a DM tool called Oracle Darwin, which is a part of the Oracle Data Mining
Suite. Oracle Data Mining Suite has a Windows GUI and is targeted towards data
analysts doing "ad hoc" data mining. Oracle9i Data Mining targets Java application
developers who want to automate the extraction of business intelligence and its integration into other business applications.
5 The Usage of Data Mining Techniques during the Data
Warehouse Design Process
The design phase is the most costly and critical part of a data warehouse implementation. In different steps of this process arises problems that can be solved by usage of
data mining techniques. For better understanding of where data mining can support
design phase, in [2] the process is divided to follow steps: Data Source Analysis,
Structural Integration of Data Sources, Data Cleansing, Multidimensional Data Modeling and Physical DW Design.
The first step of the warehouse design cycle is the analysis of these sources. They are
often not adequately documented. Data mining algorithms can be used to discover the
implicit information about the semantics of the data structures.
Data Source Analysis
Often, the exact meaning of an attribute cannot be deduced from its name and data
type. The task of reconstructing the meaning of attributes would be optimally supported by dependency modeling using data mining techniques and mapping this model against expert knowledge, e.g., business models. Association rules are suited for
this purpose. Other data mining techniques, e.g., classification tree and rule induction,
and statistical methods, e.g., multivariate regression, probabilistic networks, can also
produce useful hypotheses in this context.
Many attribute values are (numerically) encoded. Identifying inter-field dependencies
helps to build hypotheses about encoding schemes when the semantics of some fields
are known. Also encoding schemes change over. Data mining algorithms are useful to
identify changes in encoding schemes, the time when they took place, and the part of
the code that is effected. Methods which use data sets to train a “normal” behavior
can be adapted to the task. The model learned can be used to evaluate significant
changes. A further approach would be to partition the data set, to build models on
these partitions applying the same data mining algorithms, and to compare the differences between these models.
Data mining and statistical methods can be used to induce integrity constraint candidates from the data. These include, for example, visualization methods to identify distributions for finding domains of attributes or methods for dependency modeling.
Basic Approaches Of Integration Between Data Warehouse And Data Mining
9
Other data mining methods can find intervals of attribute values which are rather
compact and cover a high percentage of the existing values.
Once each single data source is understood, content and structural integration follows.
This step involves resolving different kinds of structural and semantic conflicts. To a
certain degree, data mining methods can be used to identify and resolve these conflicts.
Structural Integration of Data Sources
Data mining methods can discover functional relationships between different databases when they are not too complex. A linear regression method would discover the
corresponding conversion factors. If the type of functional dependency (linear, quadratic, exponential etc.) is a priori not known, model search instead of parameter search
has to be applied.
Data Cleansing
Data cleansing is a non-trivial task in data warehouse environments. The main focus
is the identification of missing or incorrect data (noise) and conflicts between data of
different sources and the correction of these problems. Typically, missing values are
indicated by either blank fields or special attribute values. A way to handle these records is to replace them by the mean or most frequent value or the value which is most
common to similar objects. Advanced data mining methods for completion could be
similarity based methods or methods for dependency modeling to get hypotheses for
missing values.
On the other hand, data entries might be wrong or noisy. For such kind of problems is
proper to be used tree or rule induction methods. If we have a model that describe dependencies between data, every transaction that deviate from the model is a hint for
an error or noise.
Different sources can contain ambiguous or inconsistent data, e.g., different values for
the address of a customer or the price of the same article. With clustering methods
you might find records which describe the same real-world entity but differ in some
attributes which have to be cleaned. Record linkage techniques are used to link together records which relate to the same entity(e.g. patient or customer) in one or more
data sets where a unique identifier is not available.
Multidimensional Data Modeling
Sometimes it is not sensible to model all the fields as dimensions of the cube as some
fields are functionally dependent and other fields do not strongly influence the
measures. Data mining methods can help to rank the variables according to their importance in the domain. Non-correlated sets of attributes can be found with correlation
analysis. Furthermore, using data mining methods to drive this cube design seems
promising as they can help to identify (possibly weak) functional dependencies,
which indicate non-orthogonal dimension attributes.
Sparse regions should be avoided during modeling. Using special cluster methods
(e.g. probability density estimation) data points can be identified as the center of
dense regions.
The multidimensional paradigm demands that dimensions are of discrete data type.
Therefore, attributes with a continuous domain have to be mapped to discrete values
if this attribute is to be modeled as a dimension. Algorithms, that find meaningful intervals in numeric attributes help to get discrete values.
Physical DW Design
10
Ina Naydenova, Kalinka Kaloyanova
Using OLAP tools, the business user can interactively formulate queries. Data mining
algorithms can be used to identify tasks and to find query and interaction patterns that
are typical for certain tasks or users. Thus, the input to the data mining process is a set
of sessions (each consisting of a sequence of multidimensional queries) that can be
extracted from a log file of the data warehouse system. The first step during the data
preparation phase of the knowledge discovery process is to find a formal representation of an individual query that contains the interesting properties of this query. This
involves mapping different representations of conceptually equal queries to the same
“query prototype”. A database administrator might be interested in patterns that concern the structural composition of queries (e.g., that certain dimensions are often queried together in certain phases of the analysis process). Mathematically, this first step
corresponds to the definition of equivalency classes on the space of all queries. The
actual design of the representation and abstraction function is mainly influenced by
the type of patterns that should be discovered (Metric Space, Hypertext, Sequence of
Events Approach -[2] ). The second step is to represent the transformed log file information as a structure that can serve as an input for an existing data mining method.
Summing up, the adequate representation enables the use of temporal and spatial data
mining methods in order to find groups of users that are similar concerning the time
of their queries (e.g., between 8 am and 10 am), the contents of their queries, their
navigational behavior, the granularity of their analysis, or the complexity of their
analysis [2].
6
Conclusions
In the paper we observe basic approaches for integration of data warehouse, OLAP
and knowledge discovery. We present the main characteristics of On-Line Analytical
Mining. OLAM integrates on-line analytical processing with data mining, which substantially enhances the power and flexibility of data mining and makes mining an interesting exploratory process. Important task for future study is the extension of
OLAP mining techniques towards advanced and special purpose database systems,
including extended-relational, object-oriented, text, spatial, multimedia and Internet
information systems. With multiple data mining functions available, one question,
which naturally arises is how to determine which data mining function is the most appropriate one for a specific application. To select an appropriate data mining function,
one needs to be familiar with the application problem, data characteristics, and the
roles of different data mining functions. Sometimes one needs to perform interactive
exploratory analysis to observe which function discloses the most interesting features
in the database. OLAP mining provides an exploratory analysis tool, however, further
study should be performed on the automatic selection of data mining functions for
particular applications [7].
We also examine the efforts of RDBMS manufacturers to integrate their warehousing,
OLAP and mining solutions in common framework. Such integration is a promising
direction for easier and more efficient mining of data warehouses. Regardless observed efforts at present the knowledge discovery industry is fragmented. Knowledge
discovery community generates new solutions that are not widely accessible to a
Basic Approaches Of Integration Between Data Warehouse And Data Mining
11
broader audience. Because of the complexity and high cost of the knowledge discovery process, the knowledge discovery projects are deployed in situations where there
is an urgent need for them, while many other businesses reject it because of the high
costs involved [3]. The observed integration contributes for better accessibility of data
mining solutions and reduction of project’s cost.
Finally we look at how data mining methods can be used to support the most important and costly tasks of data warehouse design. Hundreds of tools are available to
automate portions of the tasks associated with auditing, cleansing, extracting, and
loading data into data warehouses. But tools that use data mining techniques still are
rare. Building of such kind of tools is a promising direction for more efficient implementation of backend steps.
References
1. Trillium Software System: Achieving Enterprise Wide Data Quality , White Paper, 2000
2. Sapia C.; Höfling G.; Müller M.; Hausdorf C.; Stoyan H.; Grimmer U.:On Supporting the
Data WarehouseDesign by Data Mining Techniques.Proc.GI-Workshop DM and DW, 1999.
3. Cios K. J.; Kurgan L.: Trends in Data Mining and Knowledge Discovery, Knowledge Discovery in Advanced Information Systems, Springer, 2002
4. Chaudhuri S., Dayal U.: An Overview of Data Warehousing and OLAP Technology,
SIGMOD Record 26(1): 65-74 ,1997
5. Fayyad U. , Piatetsky-Shapiro G., Smyth P.: From data mining to knowledge discovery in
databases. AI Magazine, 17(3):37-54, 1996.
6. Fayyad U., Piatetsky-Shapiro G., Smyth P.: Knowledge discovery and data mining - Towards
a unifying framework. Proc. 2nd Int. Conf. on KDD'96, Portland, Oregon, pp. 82-88, 1996.
7. J. Han : OLAP Mining: An integration of OLAP with data mining. Proc. of IFIP ICDS, 1997.
8. Han J., Chee S., Chiang J. Y. : Issues for On-Line Analytical Mining of Data Warehouses.
Proc. of 1998, SIGMOD'96 Workshop on Research Issues on DMKD'98 , 1998, pp. 2:1-2:5
9. Netz, A.; Chaudhuri, S.; Fayyad, U.; Bernhardt, J.: Integrating data mining with SQL databases: OLE DB for data mining , ICDE 2001, Page(s): 379 -387
10. Oracle Corporation:Oracle9i Data Mining Concepts Release 2 (9.2),P.No.A95961-01, 2002
11. Oracle Corporation: Oracle9i Data Mining,
www.oracle.com/technology/products/oracle9i/pdf/o9idm_bwp.pdf, 2002
12. Paul S., MacLennan J., Tang Z., Oveson S.: Microsoft SQL Server 2005 Data Mining Tutorial. Microsoft Corporation, 2004
13. Charran E., Introduction to Data Mining with SQL Server (Part 2), http://www.sql-serverperformance.com/ec_data_mining2.asp, 2002
14. IBM: Data mining: Extending the information warehouse framework. White Paper,
http://www.almaden.ibm.com/cs/quest /paper/whitepaper.html, 2002
15. IBM: IBM Software - DB2 Intelligent Miner - Family Overview, http://www306.ibm.com/software/data/iminer/, 2005
16. StatSoft: Classification and Regression Trees (C&RT),
http://www.statsoft.com/textbook/stcart.html, 2003