Subscribe Now Subscribe Today
Research Article
 

A Decision Support System for Supervised Assignment in Banking Decisions



George Rigopoulos, John Psarras and Dimitrios Th. Askounis
 
Facebook Twitter Digg Reddit Linkedin StumbleUpon E-mail
ABSTRACT

This study presents a Decision Support System (DSS) which supports assignment of actions (e.g., numbers, projects, people etc.) into predefined categories according to their score on evaluation criteria. It implements a novel classification algorithm based on multicriteria analysis and fuzzy preference relations. More detailed, assignment to classes is based on the concept of category threshold, which defines at what degree an alternative can be included in a specific category. For each category a threshold is defined by the corresponding decision maker, which indicates its lower limit with respect to the evaluation criteria. Actions are then evaluated according to the criteria and fuzzy inclusion degrees are calculated for each category. Finally, an action is assigned to the category for which the inclusion degree is the maximum. The DSS implements the above classification algorithm, providing a user-friendly interface, which supports decision makers to formulate and solve similar problems. In addition to the DSS, we present a real world application at a classification problem within the environment of a Greek bank. Results derived from evaluation experiments in the business environment provide evidence that the proposed methodology and the DSS can effectively support decision makers in classification decisions. The methodology as well as the proposed DSS can be used to classification problems not only in financial domain but to a variety of domains such as production, environmental, or human resources.

Services
Related Articles in ASCI
Similar Articles in this Journal
Search in Google Scholar
View Citation
Report Citation

 
  How to cite this article:

George Rigopoulos, John Psarras and Dimitrios Th. Askounis, 2008. A Decision Support System for Supervised Assignment in Banking Decisions. Journal of Applied Sciences, 8: 443-452.

DOI: 10.3923/jas.2008.443.452

URL: https://scialert.net/abstract/?doi=jas.2008.443.452

INTRODUCTION

Classification has been active research topic during the past decades with wide range of applications. Assignment of actions (numbers, people, projects etc.) to categories is common objective in decision-making problems to a variety of fields (Doumpos and Zopounidis, 2001). Algorithms and methodologies to solve such problems come from operations research, neural networks, mathematics and machine learning, reflecting active research interest. Computational complexity of classification problems has been reduced due to developments in information technology, resulting to Decision Support Systems which implement resource intensive classification algorithms.

Classification can be divided in, (a) supervised, which requires decision maker’s contribution and refers to predefined categories which depending on whether they are ordered or not is referred as sorting or classification and (b) unsupervised, which does not require decision maker’s contribution, is executed automatically mostly based on appropriate algorithms, categories are not predefined and is referred as clustering. Multicriteria analysis offers a variety of methodologies and tools to solve sorting and classification problems (Figueira et al., 2005). However, existing multicriteria methodologies and decision support tools focus mostly on ordered categories, resulting in relative lack for classification in non-ordered categories.

Within this framework, we present NeXClass, a novel supervised classification algorithm as well as a decision support system based on multicriteria analysis, which solves classification problems to predefined non-ordered categories. The classification algorithm is based on the concept of category entrance threshold. The problem that we want to solve is to classify an action to a specific category with respect to action’s score to the evaluation criteria, considering a set of actions, a set of predefined non-ordered categories and a set of evaluation criteria. In general, for each predefined category, the decision maker defines an entrance threshold. This threshold represents the minimum requirements for an action in terms of performance on the evaluation criteria in order to be included in this category. Decision maker defines actions’ score on the criteria as well as all required parameters’ values. For each action, its performance on the criteria is compared with the entrance threshold of every category and finally the action is assigned to the category for which it has the maximum distance from the entrance threshold, in terms of a fuzzy inclusion degree.

In this study, we present the algorithm and the decision support system, which implements it, as well as a real world case study demonstrating its usage. Initially, we introduce the basic definition of NeXClass algorithm and the integrated classification methodology. Next, we present the architecture and major functionalities of NeXClass DSS. Finally, a real world case study is presented in order to demonstrate the methodology and DSS usage. Following the application of DSS, evaluation results from DSS usage provide evidence for DSS efficiency and applicability to classification problems. Although we focus on financial classification problems, the proposed DSS can be utilized to support decision makers at a variety of fields, such as human resources, production and environment where similar assignment problems can occur.

MATERIALS AND METHODS

NeXClass methodology: The proposed methodology NeXClass (Non Excluding Classification) for multicriteria classification is based on the concept of inclusion (non exclusion) to a category. Inclusion (non exclusion) of an action is determined by the value of actions’ fuzzy inclusion degree which is calculated following concordance/non-discordance concepts generalizing ELECTRE III method (Roy, 1991). In NeXClass categories are defined by an entrance threshold, which can be considered as a virtual action that marginally satisfies inclusion requirements. The objective of the methodology is to classify actions to categories considering the inclusion (non exclusion) concept.

In the following we present the definition of fuzzy inclusion degree and the methodology which utilizes it to classify actions to categories. The following notations will be used:

A = {a1, a2,..., an} is a set of actions for assignment to categories.
C = {C1, C2,..., Ch} is a set of categories, defined by their least typical representatives referred as entrance thresholds.
G = {g1, g2,..., gm} is a set of evaluation criteria.
W = {w(g1), w(g2),..., w(gm)} is the set of criteria importance weights.
B = {b1, b2, ..., bh} is the set of entrance thresholds per category.
Action’s ai score on the evaluation criteria gj is defined as gj (ai).
Entrance threshold’s bh score on the evaluation criteria gj is defined as gj (bh), where bh is the threshold of category Ck ε C.

The problem that we want to solve is defined as follows: Having a set of actions (e.g., projects, people, numbers etc.), a set of non-ordered categories which are defined by their least typical representative (entrance threshold) and a set of evaluation criteria, assign actions to categories with respect to their score on the evaluation criteria and the inclusion (non exclusion) concept.

Fuzzy inclusion degree: In order to utilize the inclusion (non exclusion) concept for the classification of actions an appropriate method is needed to quantify actions’ inclusion degree to a category. To support this, we introduce the fuzzy inclusion relation P(a, b) as a binary relation between an action ai and a category threshold bh.

The relation P(a, b) states that an action ai is preferred over a threshold bh (and can be thus included in the category Ch) if there is a majority of evaluation criteria supporting preference of action ai over threshold bh and there is no strong opposition to this. In order to evaluate the relation P(a, b) we utilize concordance/non-discordance principle, defining appropriate inclusion/non-inclusion indexes and calculate action’s fuzzy inclusion degree for a category.

Partial inclusion index: A criterion is said to be concordant if it expresses agreement about the assertion that action ai is preferred over threshold bh or equivalently gj (ai)>gj(bh) for the specific criterion. However, in order to overcome imprecision in definition of data, we define two discrimination thresholds q(gj) and p(gj) (indifference and preference thresholds respectively) for each criterion, resulting in three areas of inclusion (no inclusion, weak inclusion and strong inclusion). Considering these areas, we define the partial inclusion index Cj (a, bh) for the evaluation of concordance per criterion as follows:

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
(1)

Comprehensive inclusion index: For the evaluation of concordance degree of all criteria on the assertion that ai is preferred over threshold bh, we define the comprehensive inclusion index for action ai aggregating partial inclusion indexes as:

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
(2)

Where, wj is the importance weight of criterion gj.

Non-inclusion index: In some cases a criterion can express negative judgment about classification of action ai to a class Ch. More specifically, a criterion gj can express a significant opposition (or is discordant) to assertion that action ai is preferred over threshold bh. In order to measure the discordance degree we define the non-inclusion index Dj (ai, bh) for every criterion. To handle imprecision, we define a veto threshold v(gj) for each criterion as the minimum value which is incompatible with the assertion that the criterion is discordant, resulting in three areas of non-inclusion as follows:

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
(3)

Comprehensive fuzzy inclusion relation: In order to evaluate the assertion ’ai is preferred over a threshold bh (and can be thus included in the category Ch)’ appropriate consideration of all concordant and discordant criteria should be executed aggregating inclusion and non-inclusion indexes (Eq. 2, 3) into a comprehensive relation. Among other aggregation operators that have been proposed, we follow the ELECTRE III method and define the comprehensive fuzzy inclusion relation aggregating the inclusion relations (Eq. 2) weakened by non-inclusion (Eq. 3) as:

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
(4)

Fuzzy inclusion degree: Finally, in order to measure the credibility degree of the comprehensive fuzzy inclusion relation (Eq. 4) we define the fuzzy inclusion degree of each action aiε A for every category Ch ε C, as:

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
(5)

Classification methodology: The classification procedure using the fuzzy inclusion degree as defined above (Eq. 5) is comprised of the following phases.

Problem definition: Decision maker formulates the problem, setting all appropriate parameters. In details:

Categories: Decision Maker (DM) defines the set of categories C = {C1, C2, ..., Ch} for the classification of actions reflecting problem requirements.
Evaluation criteria: DM defines the set of evaluation criteria G = {g1, g2, ..., gm} according to problem requirements.
Criteria weights: DM defines importance weights W = {w(g1), w(g2), ..., w(gm)} of criteria.
Actions: DM defines the set of actions A = {a1, a2, ..., am} for classification and scores them on the evaluation criteria 1(a), g2(a), ..., gm(a)}.
Entrance thresholds: DM defines appropriate entrance thresholds bh for each category C = {C1, C2, ..., Ch} and scores them on the evaluation criteria gj(bh).
For each criterion DM defines preference p(gj), indifference q(gj) and veto v(gj) thresholds respectively.

Classification: After problem formulation, the following steps are followed for the classification:

For each action ai fuzzy partial inclusion relations (Eq. 1) are calculated over all thresholds bh and categories Ch.
For each action ai comprehensive inclusion relations (Eq. 2) and non-inclusion indexes (Eq. 3) are calculated over all thresholds bh and categories Ch.
For each action ai fuzzy inclusion degree is calculated for every category Ch (Eq. 4).
Action ai is assigned to the category for which fuzzy inclusion degree is maximum.

Results assessment: DM assesses the results and in case of major misclassifications, modifies the parameters accordingly and reruns the model.

NeXClass DSS: The above algorithm and methodology is implemented in NeXClass DSS, a Decision Support System which was developed in order to support decision makers to interactively solve classification problems (Fig. 1). The DSS was developed in C++ and is running under Windows OS.

DSS architecture: The DSS is oriented towards usage from single decision makers. Thus, it was developed as stand alone application, which can be installed and executed in Windows workstations. Since data necessary for classification can reside in several data sources, DSS supports XML, text files and RDBMS as data sources. For the development of the DSS we followed a N-tier architecture model, defining the following major layers: data layer, application layer and presentation layer. Every layer comprises from appropriate modules which implement the above presented methodology phases. Below we present the basic operations of major modules (Fig. 2).

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
Fig. 1: NeXClass DSS

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
Fig. 2: DSS modules

Presentation layer: This layer includes the user interface, which supports user operations. Design follows the methodology and it was based on the required functionalities. It provides a flexible guided interface for the formulation of classification problems reducing thus complexity for users. Layers’s modules are:

User management: Since it is possible that a classification model handles confidential data, (such as financial data including costs, budgets etc.) and classification results may be confidential, it is imperative to protect usage of DSS from unauthorized users. This module provides user authentication procedure and in the case of multiple users, restricts access only to user’s own models.
User interface: This module provides all the necessary screens to user in appropriate format according to the methodology.

Application layer: Within this layer, basic operations as well as the classification algorithm takes place. As described earlier, the entire methodology is separated in 3 main phases. Application layer is responsible for the implementation of these phases. Layers’s modules are:

Configuration: This module provides general configuration capabilities to user to customize the DSS interface, such as font selection, sizing, colour and other interface parameters.
Model import: In the case of large quantities of data, a user can import a model from a data source, instead of inserting all the values manually. In this case the module imports all the data from the external source and formats the classification model.
Model creation: In the general case, a user creates a new model from scratch. This module provides all the functionality to create a new model following the steps of the problem definition phase of the methodology.
Model reporting: After the model creation/import, this module provides overview of the model, allowing corrections to it.
Classification: This is the module which implements the classification algorithm, either on a training set or the entire set of the alternatives.
Results reporting: After the classification, this module presents the results in appropriate format. Results include not only the alternatives’ assignment to classes, but evaluations of inclusion degrees, concordance and discordance indexes as mentioned in the methodology.

Data layer: This layer includes the data model necessary for the application. All data are stored in appropriate format in database tables, within a relational database. Since, decision support problems may require relative large amount of data, we have included the capability to import data from XML files or text files with appropriate format. Layers’s basic module is:

Data management: This module is responsible for handling the data operations from and to data sources. This module is responsible to support data operations to data stored in text files, XML files and RDBMS in appropriate formats.

RESULTS AND DISCUSSION

Here, we present a real world application of the proposed methodology and the DSS in order to demonstrate its applicability in classification problems within business environment. The specific problem refers to a top Greek bank, which aims to reorganize its electronic payment network, consisted of retail companies that are equipped with terminals for online electronic payments. Electronic payments between retail companies and consumers are executed mostly through appropriate terminals (EFTPoS/Electronic Fund Transfer at Point of Sale), which form a network connecting banks, retailers and consumers in a triangular relationship. Banks or financial organizations are usually the owners of such networks, charging users of network services with transaction fees (Alexander et al., 1992). During the early development of payment networks, such terminals were offering basic payment services. However, today’s EFPoS terminals, which have evolved technologically and are capable for advanced payment services, can be utilized by banks as point of differentiation to competitors, offering a set of value added services to both retailers and individual consumers (Smith, 1987; Alexander et al., 1991). Although in several countries ownership of network is separated from service provision, in Greece, due to market particularities, several private EFTPoS networks owned by banks coexist, resulting in increased competition that suppresses banks’ revenues. Retailer evaluation is thus a critical issue for network viability.

Problem definition: The bank’s payment network has currently an extended installed base of more than 5.000 EFTPoS devices located at several retailer stores. However, analysis posed several inefficiencies, which in combination with high investment and operational costs as well as technical and support issues result in low profitability of the payment network. In addition, advanced capabilities and enhancements of EFTPoS devices are not efficiently deployed to support customers’ needs. Trying to minimize network inefficiencies, the bank decided to follow a renovated customer centric approach and consider its EFTPoS devices not as dummy terminals for payment execution, but as points to deliver added value payment services to its customers. Following this strategic approach, the bank’s objective is the assignment of EFTPoS retailers to appropriate categories reflecting differentiated development strategies. A retailer will be assigned to a specific category if he satisfies the minimum entrance requirements for inclusion to this, considering bank’s preferences as expressed by bank’s expert. Based on the above requirements, the proposed NeXClass methodology was applied in order to support the entire decision process. Below, we present the assignment procedure for a subset of 20 retailers following the steps of the methodology.

Table 1: Defined categories for retailer classification
Image for - A Decision Support System for Supervised Assignment in Banking Decisions

Table 2: Defined criteria for the evaluation of retailers
Image for - A Decision Support System for Supervised Assignment in Banking Decisions

Methodology and DSS application: Initially, the bank formed a two-dimensional conceptual evaluation framework comprised of retailer’s site potential and profitability indicators. The EFTPoS domain was thus segmented into four partitions. The classification categories were defined, relying on this segmentation. For the sake of simplicity, bank’s expert defined 4 categories depicting the relevant importance of retailer for the bank (Table 1). The categories are also linked to a marketing strategy that the bank will follow for the retailers classified to each one.

The next step was the definition of a set of appropriate evaluation criteria to represent both profitability and site’s potential factors, according to the segmentation. Criteria selection was based on relevant studies (Abdul-Muhmin and Alzamel, 2001; Ironfield and McGoldrick, 1988; McFayden, 1987) and availability of data for easy quantification. In total, thirteen criteria were defined satisfying bank’s requirements and Simos procedure was followed (Figueira and Roy, 2002) for the definition of importance weights, since it is straightforward and decision maker can easily understand the process. Scoring of retailers on each criterion was executed by bank’s expert, based on data collection from bank’s resources. Decision maker scored each retailer according to its performance on each criterion and values were converted to 1-100 scale using appropriate procedure for each criterion.

Criteria are presented in Table 2, along with the defined scales and importance weights. Criteria parameters were imported by DM to the DSS using the appropriate form as shown in Fig. 3.

Following the methodology, appropriate values were defined for the required parameters. Entrance thresholds for the categories were defined by bank’s expert DM setting values for each criterion in the scales defined previously. For simplicity, the expert defined low indifference and preference zones. In addition, veto thresholds were set in relative high values in order to restrict exclusions only for extreme cases. Threshold values were imported by DM to the DSS using the appropriate forms as shown in Fig. 4, Table 3.

Next, a subset of 20 retailers was selected from the existing customer base as alternatives for classification and expert evaluated their scores on criteria using bank’s data (Table 4). As mentioned earlier, expert initially scored retailers with respect to bank’s resources and values were normalized in 1-100 scale with appropriate functions for uniformity. Score values were imported by DM to the DSS in the appropriate forms as shown in Fig. 5.

Finally, the model was executed and classification results were derived. Calculated inclusion degrees and classification results are shown in Table 5, 6, respectively. We also include in Table 6 classification of this set from expert using existing procedure. As it can be seen from this set, the model is in accordance with experts’ opinion using existing procedure except some misclassifications in categories C3 and C4. In Fig. 6 and 7, the degrees as well as the classification results are depicted in screens of the DSS.

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
Fig. 3: DSS screen for criteria definition

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
Fig. 4: DSS screen for category threshold definition

Table 3: Categories’ entrance thresholds
Image for - A Decision Support System for Supervised Assignment in Banking Decisions

Table 4: Retailers’ performance on criteria
Image for - A Decision Support System for Supervised Assignment in Banking Decisions

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
Fig. 5: DSS screen for scores’ definition

Validation of NeXClass: In order to validate the NeXClass methodology, we executed a number of experiments, testing the validity of results on the one hand, as well as functionality from decision maker’s point of view. Validation testing, measured the correctness of the results and the accuracy of the algorithm. Since, supervised classification incorporates decision maker’s preference there is no absolute solution set to classification problems. For the specific bank setting, we considered a number of solution sets as defined by the expert following the traditional decision making procedure. These sets were used as benchmarks for methodology and DSS evaluation and a number of tests were executed in order to verify methodology’s classification accuracy. Below we present the results from a number of experiments regarding classification accuracy according to existing procedure (Table 7).

As it can be seen, at the initial run DSS presents a percentage of misclassifications. Existing procedure is heuristic based and definition of category thresholds is not a straightforward process. So we tuned the thresholds and after a second run misclassification percentage was reduced proving thus that these misclassifications were due to inappropriate parameter definition.

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
Fig. 6: DSS screen for inclusion degrees result

Image for - A Decision Support System for Supervised Assignment in Banking Decisions
Fig. 7: DSS screen for classification results

Our findings from DSS application and interaction with decision makers provide valid evidence that the proposed methodology and DSS can provide sufficient support for classification problems. It formulates the entire problem in a structured way, enhancing decision makers understanding, reducing thus misclassifications derived by existing heuristics. However, some limitations and issues for future study exist.

Table 5: Inclusion degrees per category
Image for - A Decision Support System for Supervised Assignment in Banking Decisions

Table 6: Classification results (Retailers per category)
Image for - A Decision Support System for Supervised Assignment in Banking Decisions

Table 7: Misclassification results from DSS validation experiments
Image for - A Decision Support System for Supervised Assignment in Banking Decisions

Deriving parameters is a time consuming and relative complex procedure, which requires decision maker to be familiar enough with the methodology. As a future enhancement, we plan to extract parameters from past decisions, in order to minimize decision maker’s effort. From the above experience, we believe that both methodology and DSS can become a valuable tool for decision makers, in similar classification problems in a variety of domains, including production, human resources and environment to mention a few.

REFERENCES
1:  Abdul-Muhmin, A.G. and I.A. Alzamel, 2001. Retailers' experiences with and attitudes toward the Saudi Arabian EFTPoS system. Int. J. Retail Distribution Manage., 29: 188-199.
Direct Link  |  

2:  Alexander, N., J. Hine and J. Howells, 1991. EFTPoS, before and after; retailer reaction. Int. J. Retail Distrib. Manage., 19: 9-15.
CrossRef  |  Direct Link  |  

3:  Alexander, N., J. Howells and J. Hine, 1992. EFTPoS: Impact on channel relationships. Int. J. Bank Market., 10: 38-44.
CrossRef  |  Direct Link  |  

4:  Doumpos, M. and C. Zopounidis, 2002. Multi-criteria classification methods in financial and banking decisions. Int. T. Operat. Res., 9: 567-581.
CrossRef  |  Direct Link  |  

5:  Figueira, J. and B. Roy, 2002. Determining the weights of criteria in the ELECTRE type methods with a revised Simos' procedure. E. J. Operat. Res., 139: 317-326.
Direct Link  |  

6:  Figueira, J., S. Greco and M. Ehrgott, 2005. Multiple Criteria Decision Analysis: State of the Art Surveys. Springer Verlag, Boston, Dordrecht, London.

7:  Ironfield, C.E. and P.J. McGoldrick, 1988. EFTPoS systems determinants of shoppers' awareness and usage. Int. J. Retail., 6: 60-61.

8:  McFadyen, E., 1987. Retailers' attitudes to EFTPoS. Retail Distrib. Manage., 15: 19-20.
CrossRef  |  Direct Link  |  

9:  Smith, C., 1987. Will EFTPoS really help the retailers improve service at the checkout desk? EFTPoS. Int. Bull., 5: 6-9.

10:  Roy, B., 1991. The outranking approach and the foundations of electre methods. Theory Decis., 31: 49-73.
CrossRef  |  Direct Link  |  

©  2021 Science Alert. All Rights Reserved