EP1629401A2

EP1629401A2 - Method and system, in addition to computer program comprising program coding elements and computer program product for analyzing user data organized according to a database structure

Info

Publication number: EP1629401A2
Application number: EP03794981A
Authority: EP
Inventors: Michael Haft; Reimar Hofmann
Original assignee: Siemens AG
Current assignee: Panoratio Database Images GmbH
Priority date: 2002-09-02
Filing date: 2003-09-02
Publication date: 2006-03-01
Also published as: WO2004025501A2; JP2005537585A; US20060173889A1; AU2003264251A1

Description

"Method and arrangement as well as computer program with program code means and computer program product for the analysis of user data structured according to a database structure".

description

Method and arrangement as well as computer program with program code means and computer program product for the analysis of useful data structured according to a database structure

The invention relates to an analysis of user data structured according to a database structure, such as customer or product data of a company. 0

Almost every process in a company, such as every contact of the company with a customer or every logistical process within a company, starting with an order for a product up to the delivery of the 5 finished product, is now carried out electronically supported, controlled and controlled.

Based on systematic data, such as customer data or product data, recorded and logged, the basis for 0 economic, business and / or strategic marketing analyzes, with which the data into usable eco _^ nomic, business and / or arktstrategische findings are implemented.

5 Because of their economic, economic and / or market-strategic importance, these company data represent a significant asset for the companies. As a result, the companies make great efforts to collect and analyze this data.

Various well-known systems are available for the collection of such company data, such as customer relationship management systems ( _C RM ₎ [1], supply chain management systems (SCM) [2] or data warehouses [3]. After collection, the data are usually stored in databases and stored in a structured manner. As a rule, data sets Di = (Ai, Bi, Ci, ...) are formed, the index i denoting the respective data set Di.

Each data record Di represents a specific object from a group of objects, for example a specific customer from all registered customers of a company or a specific product from a product line of a company.

Each data record comprises a predeterminable number of entries, Ai, Bi, Ci, ..., the individually recorded data, with categories or attributes A, B, C, ... These categories or attributes represent properties of an object group pe, such as age (A), income (B), product purchased (C), .... The entries Ai, Bi, Ci, ... for the respective categories A, B, C, ... can be numerical or semantic.

Statistical methods, so-called data mining methods [4], [10], [11], [12], are used to analyze such company data. Many of these data mining methods are based on a static framework, i.e. they are formulated in a statistical language.

A well-known and frequently used data mining process is a so-called decision tree [5].

Other known and used data mining methods are so-called clustering methods [6] or association rules [9].

A disadvantage of many of the known and mentioned analysis methods is that they cannot be used adequately when analyzing large amounts of data. As a rule, there is one or multiple access to the whole Analyzing data stock, which is stored, for example, in a database, is necessary.

With large amounts of data, this leads to long access times, long computing and response times and, as a result, poor performance. A high computing power or computing capacity is also required.

A determination of a common probability model P (A, B, C, ..., X) for a data structure (A, B, C, ...) based on a hidden variable X is known from [7].

From [8] a determination of a common probability model P (A, B, C, ...) for a data structure (A, B, C, .-.) Is known based on a structure learning.

The invention is based on the object of specifying an analysis method for the analysis of structured useful data which can also be used with large amounts of useful data and also has a high performance there.

This object is achieved by the method and the arrangement as well as by the computer program with program code means and the computer program product for analyzing user data structured according to a database structure with the features according to the respective independent patent claim.

In the method for the analysis of user data structured according to a database structure, a common statistical probability model is first determined for the user data structured according to the database structure. The user data structured according to the database structure is then analyzed using a statistical analysis method, the statistical analysis method used in the analysis being applied to the common statistical probability model, not, as is customary, directly to the output data. The arrangement for analyzing user data structured according to a database structure has:

a modeling unit with which a common statistical probability model for the user data structured according to the database structure can be determined, and

an analysis unit with which the user data structured according to the database structure can be analyzed using a statistical analysis method in such a way that the statistical analysis method used in the analysis is applied to the common statistical probability model.

The invention is clearly based on a two-stage procedure.

The first step is to assume user data that can be specified and structured according to a database structure. A database structure of this type should be understood to mean that the user data is based on a higher-level, fixed structure, for example, data sets with the same structure (Ai, Bi, Ci, ...) with the same entry categories A, B, C, .... Such structures are generally known.

A common, purpose-independent probability model, as described, for example, in [7], [8], is formed from these user data to be analyzed and structured according to a database structure.

This represents a general, complete and precise representation of a statistic of the data structure of the structured user data (“analytical database representation”). It is also a highly compressed form of knowledge about the user data. The general image can then subsequently be used as the basis for the analysis by the statistical methods. These then no longer access the entire user data inventory or the individual user data, but use the statistical image created, ie the common probability model, for the analysis.

This allows access, computing and response times to be reduced during analysis and thus increased performance.

The computer program according to the invention with program code means is set up to carry out all steps according to the analysis method according to the invention when the program is executed on a computer.

The computer program product with program code means stored on a machine-readable carrier is set up to carry out all steps according to the analysis method according to the invention when the program is executed on a computer.

The arrangement and the computer program with program code means, set up to carry out all steps according to the inventive analysis method when the program is executed on a computer, and the computer program product with program code means stored on a machine-readable medium, set up all steps according to the Carrying out analysis methods according to the invention when the program is executed on a computer are particularly suitable for carrying out the analysis method according to the invention or one of its further developments explained below.

Preferred developments of the invention result from the dependent claims. The further developments described below relate both to the method and to the arrangement.

The invention and the further developments described below can be implemented both in software and in hardware, for example using a special electrical circuit.

Furthermore, an implementation of the invention or a further development described in the following is possible by means of a computer-readable storage medium on which the computer program with program code means which carries out the invention or further development is stored.

Also, the invention or any described below

Further development can be realized by a computer program product which has a storage medium on which the computer program with program code means which carries out the invention or further development is stored.

In a further development, structured user data are used in user data records, for example user data records from a database. Each user data record represents a specific object from a group of objects. The user data associated with the respective user data record describe properties of the respective object.

Statistical methods based on a hidden variable [7] or methods based on structural learning [8] can be used to determine the common statistical probability model. A combination of both methods is also possible.

Furthermore, it is expedient that the statistical analysis method is applied to the common statistical probability model in such a way that a common probability is used as an input variable for the statistical analysis method. driving is used. The common probability results directly from the common probability model. This avoids unnecessary intermediate steps, costs computing time and extends response times.

A method based on a data mining method [4], [10], [11], [12] can be used as the statistical analysis method, for example a clustering method [5] or a decision tree [6] or association rules [9].

When analyzing using the statistical analysis method, it is possible to determine dependencies between the user data and / or their significance based on a statistical test. This can be due to the highly compressed form of the user data, i.e. the common probability model, interactively and very efficiently.

Furthermore, it makes sense to determine the common statistical probability model and to analyze the common statistical probability model using the statistical analysis method at different times and locations.

For example, the analytical database image, i.e. the common probability model is newly formed at predefinable time intervals, such as daily or weekly. Education can take place at night or on weekends. The complete analytical database image is then available as required to significantly speed up analyzes.

The user data can be obtained from various data sources. The easiest way is to obtain the user data from a database in which the user data are stored and from which they are read. Because of the performance it can achieve in the analysis of data, the invention is particularly suitable where large amounts of data have to be processed or analyzed, such as in the area of customer relationship management (CRM) [1] or supply chain management [2]. or a data warehouse (DW) [3].

In the area of CMR, further training can be used, for example, to analyze customer data. In this case, the object is a customer, which is described by at least two of the following properties, age, income, product purchased, date of purchase, frequency of purchases. This enables eminently important issues to be solved for marketing departments, such as the customer behavior of certain customer groups. Based on this, targeted target groups can be determined when customers are acquired, customer groups can be selected for specific products and marketing campaigns, and customers can be served with more foresight.

An embodiment of the invention is shown in figures and is explained below.

Show it

Figure 1 sketch that schematically shows how a

Analysis system for analyzing customer data according to an embodiment shows;

Figures 2a to g sketches showing analysis results of an analysis system for analyzing customer data according to an embodiment.

Execution example: Analysis system for analyzing customer behavior at a bank based on a customer relationship management system The exemplary embodiment relates to an analysis system for analyzing customer data from a bank.

It must be said in advance that the analysis system described below can be used not only at banks, but also at any company for the analysis of corresponding company data, such as for example in department stores or manufacturing companies.

Functioning of the analysis system (Fig.l)

1 shows schematically the functioning 100 of the analysis system for analyzing bank customer data 110.

Functionality 100 is divided into knowledge acquisition 101 and implementation of the knowledge into intelligent operation of bank customers 102.

Large and thus difficult-to-handle amounts of customer data 110 are first condensed 111 into a statistical model 112, a common probability model, of customer behavior.

The shared probability model 112 used is one based on a hidden variable. The basics are described in [7].

It should be noted that other types of common probability models can also be used, such as those based on structural learning [8].

Using the common probability model 112, properties of the customer and in particular their behavior over time can be explored much more efficiently and flexibly than on the basis of the initial data. For this purpose, statistical methods 120, in general data mining methods and here in this case a decision tree, are used, which are based on the statistical model.

It should be noted that other data mining methods can also be used, such as clustering methods or association rules.

Basics of data mining procedures are described in [4], [10], [11], [12], a decision tree in [6] and clustering procedures in [5].

The coupling is made possible by the fact that the data mining methods or decision tree 120 are based on a statistical framework, and thus use the same statistical terms or the same statistical language as the common probability model 112.

Important questions (cf. FIGS. 2) can be answered 140 interactively using the decision tree 120 using the common probability model 112.

This not only enables a quantitative (how many customers?) But also a qualitative view of the customer (which type of customer), for example:

- How many and what quality of customers come through which partnerships or campaigns? How efficient are my advertising measures? - Which customer classes with which preferences and needs are there? How and when can these needs best be met?

The results of the questions can be further implemented 121 in intelligent customer service 130. Customer data ((Fig. 1, 110) The customer data 110 in the analysis system is collected as part of a customer relationship management (CRM) 150.

The basics of CRM are described in [1].

With the CRM 150, large amounts of data 110 about the bank customers from all sales channels of the bank, such as direct contacts, web, call center, are recorded and stored.

The following are recorded and saved for the customers (so-called attributes A, B, C, ...):

- the acquired bank products A in the respective chronological order (AI, A2, A3, ...),

- a time interval B between the purchase times of the acquired bank products (Bl-2, B2-3, B3-4, ...),

- a date of birth (C),

- an income (D),

- an address (E),

- the last bank visit (F), - the last account movement (G).

The data is stored in a database in the form of customer-specific data records Di (AI, A2, ..., Bl-2, B2-3, ..., C, D, ...), the index i being the respective bank customer i features.

Common probability model (Fig. 1, 112)

The knowledge about the bank customers that is hidden in this data 110 is then condensed into a model, the common probability model 112.

The common probability model 112 used is one based on a hidden variable X. The foundations for this are described in [7]. The common probability model 112 is written based on the hidden variable X as P (A, B, C, ..., X) for all attributes (A, B, C, ...).

Such a statistical image of data represents a highly compressed form of knowledge about customers and can be used to efficiently and interactively explore dependencies 120, 140.

With the help of the common probability model 112 created here, the knowledge about the customers can now be quickly and efficiently tapped; in particular, customer behavior can be studied easily and flexibly, typical customer behavior patterns and development cycles can be analyzed efficiently and intuitively, typical customer segments and determine and recognize their preferences reliably and clearly 120, 140.

In addition to the described analysis function, the common probability model 112 also provides quickly retrievable forecasts of further expected behavior and current needs of a customer. The forecasts can also be used to serve customers proactively and in a targeted manner and to provide proactive, personal offers 130.

Attachment of a decision tree to the common probability model (Fig.l, 120)

If the common probability model 112 is used further, the decision tree [6] is placed 120 on the statistical model 112, the common probability model 112.

This allows arbitrary edge distributions, such as those for a first split of the decision tree, namely P (A, X),

Determine P (B, X), P (C, X), ..., and also for all other splits in the decision tree. Furthermore, all basic probability distributions or basic probabilities P (A), P (B), ... and any conditional probabilities or probability distributions P (B | A), P (C | A), P (C | B ), ... determine.

From the common distribution P (A, B, C, ..., X) based on the hidden (or latent) variable X, the common distribution P (A, B, C, ...) goes over all attributes of the customers by summation over the hidden variable X.

Structural learning immediately provides a common distribution P (A, B, C, ...).

Any one-dimensional marginal distributions (margins) P (A), P (B), ..., lower-dimensional distributions P (A, B), P (B, C), ... and derive any conditional probabilities (one-dimensional or one-dimensional) P (B | A), P (C | A), P (A, C | B), ...

This is done as part of an inference process, as described in [13].

According to [13], the structure of the models, for example those with a predefined hidden variable or those that were generated by structure learning, or a combination of the above, is used to efficiently calculate necessary sums over the common distribution.

Decision trees are usually built using a known CHAID or a known CART procedure.

In general, to build a decision tree with a target variable (or dependent variable) A for the so-called first split, all pairwise distributions P (A, B), P (B, C), P (A, D),. ... A selection of a variable from the set of variables B, C, D, ..., for the first split then takes place in almost all known methods based on a statistical criterion (a statistical test and significance criteria) based on the pairwise distributions P (A, B), P (B, C), P (A, D), ... and a known number of data.

If, for example, the variable D with the two values dl and d2 was selected for the first split, then for the second split conditional pairwise distributions of the form P (A, B | dl), P (A, B | d2) are required, P (A, C | dl), P (A, C | d2), ....

The necessary probabilities or distributions for the construction of the decision tree (or as the basis for the necessary statistical tests) can (as usual) be determined from the data or also from the most accurate probability model described in the above (inference process).

Interactive analyzes (Fig.l, 140, Fig.2a to 2g)

2a to 2g show examples of some of the possible interactive analyzes 140 that can be carried out with the decision tree 120 using the common probability model 112.

2a shows probability distributions P (A1), P (A2) P (A3) P (A4), P (A5), P (Bl-2), P (B2-3), P (B3-4) and P (C) and P (D). P (Al = "Giro / Salary Account) = 56, 125% is particularly marked.

Fig. 2b now shows conditional probability distributions under the condition Al = "Giro / ^Salary Account ^λλ , namely P (A2 | A1 =" Giro / ^Salary Account "), P (A3 | A1 =" Giro / Salary Account ") , P (A4 | A1 = "Giro / Salary Account"), P (A5 | A1 = "Giro / Salary Account"), P (B1-2 | A1 = "Giro / Salary Account ^λ ), P ( B2-3 | A1 = "Gi ro / Salary Account "), P (B3-4 | A1 =" Giro / Salary Account ") and P (C | A1 =" Giro / Salary Account ") and P (D | A1 =" Giro / Salary - Account "). P (A2 = "insurance product | Al =" current account / salary account) = 29% and P (A2 = "savings / investment | Al =" current account / salary account) = 50% are particularly marked.

2c now shows conditional probability distributions under the conditions Al = "checking / salary account" and A2 = "insurance product", namely P (A3 | A1 = "checking / salary account", A2 = "insurance product"), P ( A4 | A1 = "Giro / Salary Account", A2 = "Insurance Product"), P (A5 | A1 = "Giro / Salary Account", A2 = "Insurance Product"), .... P (Bl-2 = "purchase gap between first and second product greater than 3 years | A1 =" current account / salary account ", A2 =" insurance product) is particularly marked here = 85%.

2d shows further conditional probability distributions under the conditions Al = "checking / salary account" and A2 = "saving / investment", namely P (A3 | A1 = "checking / salary account", A2 = "saving / investment" ), P (A4 | A1 = "Giro / Salary Account", A2 = "Saving / Investment"), P (A5 | A1 = "Giro / Salary Account", A2 = "Saving / Investment"), .. . The probability distributions P (B1-2 | A1 = "current account / salary account", A2 = "savings / investment") are particularly marked here.

Fig. 2e shows the probability distributions P (A1), P (A2) P (A3) P (A4), P (A5), P (Bl-2), P (B2-3), P (B3-4) and P (C) and P (D). P (Al = "Giro / Salary Account) = 56, 125% is particularly marked. Fig. 2e also shows the probability distribution of the hidden variable X, referred to here as segments, namely P (segments). Particularly marked st P (segments = 4) = 34%, which shows that 34% of all recorded bank customers fall into segment 4.

FIGS. 2f and 2g again show the conditional probability distributions, once under the condition segments = 4 (Fig.2f) and the other time under the condition C = date of birth between 980 and 1990 (Fig.2g).

The following publications are cited in this document:

[1] Customer Relationship Management System, available on August 31, 2002 at: http://www.crm-expo.com/.

[2] Supply Chain Management System, available on June 31, 2002 at: http: // www. sap-ag.de/germany/solutions/scm/.

[3] Data Warehouse, available on August 31, 2002 at: http: // www. data warehouse systems. de /.

[4] Heckermann, D., "Bayesian Networks for Data Mining", Data Mining and Knowledge Discovery, pages 79 to 119, 1997.

[5] Kass, G., "An exploratory technique for investigating large quantities of categorical data", Applied Statistics, 29: 2, pages 119 to 117, 1980.

[6] Bezdek, J.C., Pal, S.K., "Fuzzy Models for Pattern Recognition", IEEE Press, 1992.

[7] Everitt, B.S., "An Introduction to Latent Variable Models", London, Chapman and Hall, 1984.

[8] Reimar Hofmann, "Learning the structure of nonlinear dependencies with graphic models", dissertation at the Technical University of Munich, publisher: dissertation.de, ISBN: 3-89825-131-4.

[9] Ashoka Savasere, Edward Omiecinski, Shamkant B. Navathe, "An Efficient Algorithm for Mining Association Rules in Large Databases", The VLDB Journal, pages 432 to 444, 1995. [10] Usama M. Fayyad, Gregory Piatetsky-Shapiro, Padhraic Smyth and Ramasamy Uthurusamy, "Advances in Knowledge Discovery and Data Mining", American Association for Artificial Intelligence, CA, 1996. [11] lan H. Witten, Eibe Frank, Morgan Kaufmann, Data Mining, 2000.

[12] T. Hastie, R. Tibshirani, J.H. Friedman, "The Elements of Statistical Learning: Data Mining, Inference, and Prediction," Springer Series in Statistics.

[13] Jensen, V.J., "An Introduction to Bayesian Networks", UCL Press, London, 1996.

Claims

claims

1. Method for the analysis of user data structured according to a database structure, - in which a common statistical probability model is determined for the user data structured according to the database structure, - in which the user data structured according to the database structure are analyzed using a statistical analysis method, the statistical analysis methods used in the analysis is applied to the common statistical probability model.

2. The method as claimed in claim 1, in which the user data structured according to the database structure are structured in user data records, which user data records each represent an object, the user data of a user data record describing properties of the respective object.

3. The method of claim 1 or 2, wherein the common statistical probability model is determined based on a hidden variable.

4. The method according to any one of claims 1 to 3, wherein the common statistical probability model is determined based on a structure learning.

5. The method as claimed in one of the preceding claims, in which the statistical analysis method is applied to the common statistical probability model in such a way that a common probability of the common probability model is used as an input variable for the statistical analysis method.

6. The method according to any one of the preceding claims, in which a method based on a data mining method is used as the statistical analysis method.

7. The method according to claim 6, in which a clustering method is used as the statistical analysis method.

8. The method according to claim 6

In which a method known as "association rules" is used as the statistical analysis method.

9. The method according to claim 6, in which a decision tree is used as the statistical analysis method.

10. The method as claimed in one of the preceding claims, in which, during the analysis using the statistical analysis method, dependencies between the useful data are determined and / or their significance is determined based on a statistical test.

11. The method as claimed in one of the preceding claims, in which the determination of the common statistical probability model and the analysis of the common statistical probability model are carried out at different times and locations by the statistical analysis method.

12. The method according to any one of the preceding claims, in which the user data are stored in a database.

13. The method according to any one of claims 2 to 12, wherein the object is a customer, which is described by at least two of the following properties, age, income, purchased product, date of acquisition, frequency of purchases.

14. The method according to any one of the preceding claims, used in the data warehouse, the user data describing the data warehouse.

15. The method according to any one of claims 1 to 13, used in a customer relationship management or supply chain management, the user data being customer data or product data.

16. Arrangement for the analysis of user data structured according to a database structure,

with a modeling unit with which a common statistical probability model for the user data structured according to the database structure can be determined,

with an analysis unit with which the user data structured according to the database structure can be analyzed using a statistical analysis method in such a way that the statistical analysis method used in the analysis is applied to the common statistical probability model.

17. Computer program product, which comprises a computer-readable storage medium, on which a program is stored, which, after it has been loaded into a memory of the computer, enables a computer to carry out the following steps for analyzing user data structured according to a database structure,

- A common statistical probability model is used for those structured according to the database structure

User data determined,

- The user data structured according to the database structure are analyzed using a statistical analysis method, the statistical analysis method used in the analysis being applied to the common statistical probability model.

18. Computer-readable storage medium on which a program is stored which, after it has been loaded into a memory of the computer, enables a computer to carry out the following steps for analyzing user data structured according to a database structure,

a common statistical probability model is determined for the user data structured according to the database structure,

19. Computer program with program code means to all

Perform steps according to claim 1 when the program is executed on a computer.

20. Computer program with program code means according to claim 18, which are stored on a computer-readable data carrier.

21. Computer program product with program code means stored on a machine-readable carrier, in order to carry out all the steps according to claim 1 when the program is executed on a computer.