BANK PRODUCT ADVISORY SYSTEM APPLYING CASE-BASED REASONING METHOD

Andrzej Kluza, Stanisław Jabłonowski
Department of Econometrics and Informatics, Warsaw Agricultural University, Poland

ABSTRACT

Case-Base Reasoning (CBR) method utilizes a base of implicit knowledge, included in stored cases. The analytical subset of CBR applications covers factual advisory systems, mentioned in the paper. A presented bank product advanced search engine build by authors is contained in this class also. Banking product extension is a source of widespread information and advertisement, but many Internet portals have deficient search engines comparing to complexity of bank product selection task. Advantages of described application are setting personal preferences and weights, which can be used to delimit personal preferences. Real world data of 142 bank accounts of 27 Polish banks are used as records in the case base. CBR method includes variety of result adaptation scenarios. Presented actual searches with adaptation of results show how to apply this phase of retrieval.

Key words: Case�Based Reasoning, Artificial Intelligence, bank products.

INTRODUCTION

Paper is a presentation of a support system, selecting a bank account, build by authors, a continuation of previous work [10]. The task, the system undertakes is distinguishing the bank account from market offer, most adjusted to preferences, presented by a potential client. This is the first system with flexible search in bank account selection, which utilizes the Case-Based Reasoning (CBR), known to the authors.

CBR method is seen as one of modern methods of advisory functions implementation in information system environment [4, 12]. This method belongs to a group of artificial intelligence methods. In distinguishing from Artificial Intelligence methods based on rules, in the CBR method, knowledge is stored in implicit form in a case base. System can inform of how to achieve a good solution on the basis of experience from the past. Following Kolodner [12], we can assign CBR application described in this paper to analytical CBR category. This class deals with of objects classifications, diagnoses and counselling; tasks of forecasting activities of time series, tasks of risks and cost estimation. We can indicate some author previous research of this field:

• real estate advisory services [11],

• detection of some dependencies in time series [7, 8, 9].

Numerous other examples of utilization of CBR method in counselling systems, applied by leading commercial and industrial enterprises are known [1, 6]. It is worth to mention some concerns using this type counselling technology: 3COM, PSA Citroen-Peugeot, MET Ericsson, Snecma, Ruhrgas, Cisco, NEC Computers, SNCF, Legrand, Daimler-Chrysler, and General Motors, to list the biggest only.

The Electronic Commerce (e-Commerce) is more and more frequent and one of most promising areas of employment of CBR method, compare [5, 14, 15]. Examples can be Phillips household product base search, and search in Analogy Devices base of electronic components. Wide Area Network application developer software, such WEBSELL [4], for creating intelligent applications of electronic commerce based on CBR technology are now available.

Financial counselling services are among specialized domains of investment banking services on developed markets. Large number of products that banks offer is the reason of an extensive information supply and intensive advertising. Internet network is a form of very convenient media for information of bank products. Polish portals like bankier.pl, money.pl, biznes.onet.pl, expander.pl, pb.pl (Business Pulse), www.wabank.pl advertise and inform of accounts, credits and other banking products, along with extensive support of future client. There are several types of information advisory systems available on these portals. Direct developers of bank products also offer detailed explanations of their investment products, along with comparison and calculators of profitability, e.g. cu.com.pl (Commercial Union), mbank.com.pl (mBank). Portal services propose criterion-driven search of banking products on the portal. This functionality goes along with current trends of greater utilization of Internet network media information for support of the whole cycle of sale and commerce in general [14]. Paper shows a counselling system that may outperform the referenced portal bank product selection solutions.

METHOD APPLICATION

CBR method uses a base of known cases. The indispensable employment conditions of this method consist of invariability of rules concerning case, phenomena repeating capability, and continuity of part of object features´ values. Every case in the case base represents knowledge related to a certain experience, and description of a procedure, which proved to bring success, solving problem connected with this experience [12].

Drawing 1 presents basic schema of CBR method. Past experiences data base, in the form of cards file, are compared in a query with single card placed on the left, which means current experience, otherwise called a problem. A similarity function is applied to decide which records of the pile are most similar to the problem. This function can be treated as inverse of certain metrics, see [4, 12, 13, 16]. The method admits minimization of distance function as well as maximization of similarity function. After finding a similar record from past, a solution of similar problem from the past S_SIM can be adapted for the present case as current problem solution S_o. This is the last step of the method, presented as dart directed to the left on Graph 1.

Having query features vector as such:

X_o = (x_o,1, x_o,2,.., x_o,n),

where:
n - number of all features of the product,
x_o,j - single desirable j-th feature value, j=1, 2, .., n.

The case base contains k data sets, describing bank account types. Each i-th bank account type is presented by n feature values x_i,j, i = 1, 2,... k, j = 1, 2,... n. The value of local similarity function specifies how similar is a certain j-th query feature and i-th product j-th feature x_i,j. The function is described on values [4, 12, 13]:

i = 1, 2,... k, j = 1, 2,... n,

where means a maximum difference between two values of certain j-th feature of i-th product X_i in the set of all k bank products.

Global similarity SIM(X_i) is meant as a sum of problem local similarities to i-th case X_i from case base. Both local and global similarity functions have values from set <0,1>.

We can count global similarity, concerning only some, chosen features of the product profile. A case completeness rule tells which features are designated for global similarity computations, see [3]. Let us define global similarity as weighted sum of local similarities:

,

where w_j - given feature weight, defining a degree of client preference, w_{j Î} (0, +¥). A factor m_i,j Î {0,1} applicates case completeness rule to the subsequent computations of i-th product. A j-th case feature of client concern has the value m_i,j = 1, and inputs j-th feature local similarity into computations. A client unconcerned feature or feature of no value for this i-th account type has the value m_i,j = 0. Global similarity function SIM(X_i) is an adaptation of city-block metrics see [13].

Both global and local similarity metrics can utilize non-linear functions. In the domain of bank account products this move can implement a complex expert knowledge showing how well a certain bank account characteristic is adjusted to given client preferences, producing a certain quantitative measure of supremacy of one account product over another out.

EMPIRICAL MATERIAL

Empirical data contains bank account profile records, extracted from Bankier.pl internet portal [2]. The material covers 142 bank account types, 106 of personal and 36 of company type. Data concerns 27 banks acting on the area of Poland.

The following Table 1 presents names of bank account types, their features and feature ranges. Subsequent associated local similarity function is noted. Local similarity functions were defined as follows [4, 12, 13]:

- for non-continuous variables,

- for continuous variables.

If client declares a monthly deposit level, features signed * can be exchanged by features signed **.

OPERATION OF COUNSELING SYSTEM

At the beginning of counselling process all of the bank account features with its value ranges from case base are specified to a potential bank customer for accustoming with market offer. Next, we have to interrogate the client, which of account features are important for functioning of prospective account and what values of these features are desired or at least accepted. After entering preferred values of account features the client should assign own subjective weights to each account feature. These weight values define a level of importance of each subsequent feature, and should be positive real numbers.

Inputting gathered personal preferences to a computer spreadsheet and running MS Excel procedure implementing earlier described mathematical tool gives us a list of bank account types, sorted from the most similar to the query first. The list is created according preference - filtering only company or personal account type.

The adaptation phase of the CBR method begins with presenting the obtained list to the client. At this moment the final choice can be done. But if the first account type on the list lacks any important features, needed for the client, the following positions on the list can fulfill client expectations. The client can also adapt the expressed preferences to the obtained solution, which can be otherwise very attractive. Another way for successful solution is changing client preferences or weights, rerunning computations, and searching an acceptable account type in the next set of results.

Table 2 shows data of three examples of bank product choice. For every query example the preferences and weight values are shown as well as the case which was most similar to the query.

Example 1. A person wants to open an account with preferred interest rate of 3% and 3000 PLN debit level. The less important preference is monthly deposit, declared as 1000 PLN. After querying the case base a better result is obtained: higher interest rate, smaller minimal deposit and desired debit limit.

Example 2. A prospective client searches an personal account. The most important are: an absence of monthly fee, zero transfer cost and an absence of bill payment charge. Only in the half important are monthly minimal deposit, declared as 800 PLN and interest rate of 12%. The first place on result list occupies an account type fulfilling most conditions, but monthly minimal deposit is set to 1500 PLN. If this recommendation would seem interesting, the client could agree, changing somewhat initial preferences.

Example 3. Company account preferences are set as follows: the least credit interest and zero transfer cost. The less important was debit limit of 100000 PLN. The least important were telephone and Internet access and trading, 3% interest rate and bank name known for quality service. Table 2 shows the most similar bank product to the query. Although we see that an absence of debit doesn´t suit the preferences. Here we can start negotiations - instead of the debit, client probably accepts drawing loans with talking terms. In other words we do not set a certain feature option value, but try to adapt other feature for desired functionality.

CONCLUSION

Case-Based Reasoning method as a method of artificial intelligence establishes a wide spread of real world applications. It shows many advantages not only in production and commerce areas, but financial counselling can utilize them as well [12]. Presented CBR application:

• can be applied without exhaustive understanding or description of laws that rule phenomena,

• unlike statistical tests, has no need to state any hypotheses in the course of computations,

• explicit knowledge of how to solve a certain problem is not necessary, distinguishing of artificial intelligence systems based on rules,

• suits to work on non-numeric type data and to recollect premises from which the result is drawn, distinguishing of artificial neural networks.

In spite of many advantages, search system shows the prerequisites that can be not unambiguous:

• it requires arbitrary setting of local as well as global similarity functions, option and weight values,

• a user have to understand what preferences are set against features of market offer,

• basic knowledge of the case base phenomena is necessary to form a similarity functions,

• a dependable case base is needed for reasoning.

Future work can comprise research on how expert knowledge can influence similarity functions, and how to set client preferences and similarity functions concerning different feature value ranges. Also some typical algorithms of query results adaptation can be prepared to help the advising process. While setting personal options of a prospective client a weights-adjusting method can be suitable to use, e.g. Analytical Hierarchy Process. The calculations could be incorporated in the system, acting as a beginning step in CBR query preparation.

REFERENCES

1. AI-KBG (Artificial Intelligence Knowledge-Based Group), 2003. Case-Based Reasoning homepage. University of Kaiserslautern, http://www.cbr-web.org/

2. BANKIER.PL, 2003. Listing of bank accounts of Polish banks from 2003-06-08 http://www.bankier.pl/fo/konta/narzedzia/zestawienie/

3. Bergman R., Wilke W., Vollrath I., Wess S., 1996. Integrating General Knowledge with Object-Oriented Case Representation and Reasoning. 4th German Workshop: Case-Based Reasoning - System Development and Evaluation.

4. Case-Based Reasoning Technology, From Foundations to Applications, 1998. ed. Mario Lenz, Brigitte Bartsch, Hans-Dieter Burkhard, Stefan Wess. Lecture Notes in Computer Science, Springer, Heidelberg.

5. Donner M., Roth-Berghoffer T., 1999. Architectures for Integrating CBR-Systems with Databases for E-Commerce. Proceedings of the 7th German Workshop on Case-Based Reasoning, Wurzburg, Germany.

6. Kaidara Software, 2003. web page http://www.acknosoft.com/

7. Kluza A., 2003. Odkrywanie trendów w przebiegach kursów akcji z zastosowaniem metody sztucznej inteligencji opartej o bazę przypadków. [Shares trends detecton using a case based Artificial Intelligence method.] Conference proceedings: Metody ilościowe w ekonomii, SGGW, Warszawa. [in Polish] in print.

8. Kluza A., 2003. Wnioskowanie o podobieństwie ciągu w szeregu czasowym do ekstremum lokalnego w oparciu o bazę przypadków. [Series similarities inference to local extremum in time series based on a case base.] Roczniki Informatyki Stosowanej, tom 4, Politechnika Szczecińska, Szczecin. pp. 391-399. [in Polish]

9. Kluza A., 2004. Odkrywanie ciągów samopodobnych w szeregu czasowym z zastosowaniem metody wnioskowania opartej o znane przypadki. [Self-similar trend discovery in time series using artificial intelligence case-based method.] Ekonometria - Zastosowania Metod Ilościowych, Zeszyty Naukowe Akademii Ekonomicznej, Wrocław. [in Polish] in print.

10. Kluza A., 2004. Zastosowanie sztucznej inteligencji do wyboru produktów bankowych z użyciem bazy przypadków [Case-based Artificial Intelligence in choosing banking products]. Ekonomika I Organizacja Gospodarki Żywnościowej. 51 (2003), Warszawa [in Polish] pp. 111-120.

11. Kluza A., Jabłonowski S., Kotlarska J., 2004. Zastosowanie sztucznej inteligencji do wyboru mieszkań na rynku wtórnym z użyciem bazy przypadków. [Adaptation of Case-based Artificial Intelligence method for real estetate advisory services] Conference proceedings: Modelowanie procesów ekonomicznych, WSH Kielce - SGGW Warszawa. [in Polish] in print.

12. Kolodner J., 1993. Case-Based Reasoning. Morgan Kaufman Publishers, Inc., San Francisco, CA.

13. Kukuła K., 2000. Metoda Unitaryzacji Zerowej. Biblioteka Ekonometryczna, Wydawnictwo Naukowe PWN, Warszawa.

14. Lenz M., 1999. Experiences from Deploying CBR Applications in Electronic Commerce. Proceedings of the 7th German Workshop on Case-Based Reasoning, Wurzburg, Germany.

15. Prasad B., 2003. Intelligent Techniques for E-Commerce. Journal of Electronic Commerce Research, VOL. 4. NO.2.

16. Richter M.M., 1995. On the Notion of Similarity in Case-Based Reasoning. Mathematical and Statistical Methods in Artificial Intelligence (eds. G. della Riccia, R. Kruse, R. Viertl). Springer Verlag. pp. 171-184.

Responses to this article, comments are invited and should be submitted within three months of the publication of the article. If accepted for publication, they will be published in the chapter headed 'Discussions' and hyperlinked to the article.