Informatica 26 (2002) 47-56 Construction and application of hierarchical socioeconomic decision models Marjan Krisper and Blaž Zupan Faculty of Computer and Information Science University of Ljubljana, Slovenia Phone: -h386 1 476 8388m Fax: -t-386 1 426 4647 E-mail: marjan.krisper@fri.uni-lj.si, blaz.zupan@fri.uni-lj.si Keywords: socioeconomic models, socioeconomic development, decision support, hierarchical decision models, what-if analysis, comparative analysis, data visualization Received: January 17,2000 The article presents a utiUty of multi-attributed hierarchical modeUing approach to represent, analyze and study socioeconomic processes. The models are based on criteria tree for which the expert specifies the utility functions. The specific advantages of the approach are structuring the problem domain, a relative ease to build the models and the existence of underlying tools for comparative and what-if type of data analysis. We use these tools to construct two socioeconomic models, one for assessment of country's knowledge infrastructure and the other one for assessment ofquality ofpolitical and economic system. We demonstrate the utility of these two models through experimental application in the analysis of real-world data from Word Competitiveness Yearbook. Introduction A determined orientation of the developed countries to fos­ter the growth of Information infrastructure that will allow their transition to Information society [7] shows that we are undergoing a period that will exert a decisive influence on their future development. This is aiso or even more true for the Central European countries like Slovenia, Czech Re­public, and Poland where the change of the political, eco­nomic, and legal system is the basis for their gradual tran­sition to a modem society and their prospective integration within European Union. In order to monitor and evaluate such transition, com­pare countries' successfulness, and investigate for the alter­native development scenarios, one may benefit from mod­els that assess the value of country's system given a se­lection of its observable criteria. A well-known example of such approach has been carried out by International In­stitute for Management Development (IMD), a non-profit foundation from Lausanne, Switzerland. IMD systemat­ically coUects different criteria from over 40 world-wide countries (roughly one half of them being OECD members and another half being newly industrialized and emerging market economies), resulting in a yearly report called The World Competitiveness Vearbook (see, for example, [6]; in this article we will refer to as Vearbook). Each Vearbook normally includes more than 200 crite­ria, of which about two thirds present measurable quanti­ties (e.g. GDP, unemployment, etc), while the other third is obtained from the Executive Opinion Survey. Different as­pects of world competitiveness are described by eight/ac­tors (like Domestic Economy, Government, Finance, etc.) which are derived from observable criteria. To organize the criteria further, each factor includes several criteria sub­groups — in this tree-like three-level structure (Figure 1), each criterion belongs to a single subgroup, and conse­quently to a single factor. Factor and factor subgroups thus represent an aggrega­tion of the observable criteria. The observable criteria are first scaled, and then weighted and summed to obtain the value of their corresponding factor subgroup (see [6] for details). Finally, factors are computed as the sum of their corresponding factor subgroups. The country's data is then analyzed by presenting country's rank when considering each of the criteria, subgroups or factors. The advantage of IMD's approach lies in the high number of quality crite­ria being gathered, and providing a simple two-level struc­ture in which these criteria are aggregated and studied. The disadvantage, however, is that the criteria aggregation by means of weighted sum may be over-simplistic as it does not take into consideration any potentially more complex criteria interaction. Furthermore, IMD's evaluation proce­dure that assigns ali measurable criteria an equal weight of 1 and aH the survey criteria an equal weight of about 0.9 may be too restrictive as it would be expected that differ­ent criteria are differently important (relevant). And finally, the Vearbook is in a sense static and calls for a computer-supported environment that vvould allow an Interactive use of underlying evaluation model, supporting the decision making in terms of what-if and comparative analysis. Crucial to the utility of such computer-supported models is their ability not only to reach a valid and (hopefully) ac­curate conclusions, but also to explain why such conclusion were reached [11, 10]. The modeling methodology should provide grounds for explorative analysis of alternatives be­ing evaluated, making the model and decision support envi­ Informatica 16 (2002) 47-56 M. Krisper et al. JO. JD. factor O ••• O O ••• o subgroups \ //\ //\ //\ :"S:oo-o o o-o o o-o o o-o Figure 1: General systematization schema for criteria used by IMD. ronment a valuable tool for decision expert. In these terms, classical numerical decision models that are based on cri­teria weighting [5] may be inadequate and pose problems where modeling a more complex interdependence of cri­teria is required [3]. This article builds on alternative ap­proach for multi-attribute decision making that hierarchi­cally organizes the criteria in the criteria tree and introduces new aggregate criteria. The aggregate criteria simplify util­ity function elicitation and play major role for explorative analysis. The approach was first proposed by Efstathiou and Rajkovič [8] and subsequently used in many applica­tions, including the evaluation of R&D projects [3], evalua­tion of applications for nursery schools [14], priority rank­ing of applications for housing loans [1], portfolio analy­sis [9] and strategic planning [16]. In this article, we refer to its implementation in an expert system shell for decision supportDEX[2]. Compared to IMD's three-level (criteria-subgroups­factors) criteria tree, we define models that have arbitrary number of layers, and refer to ali internal nodes of the trees as intermediate criteria and the root of the tree as target criteria. Intermediate and target criteria are also referred to as aggregated criteria, as their value is computed from other underlying criteria rather than provided as an input to the model. The leaves of the trees represent criteria se­lected from those defined in the Yearbook — we refer to them as basic criteria. Using this terminology, the IMD's criteria subgroups are intermediate, and factors are target criteria. We propose two different models, one for Knowl­edge infrastructure and one for System target criteria. The article is organized as follovvs. Section 2 intro­duces DEX paradigm for hierarchical multi-attribute de­cision models. DEX-based socioeconomic models for Knowledge infrastructure and the Quality of Political and Economic System are presented in Section 3. Section 4 il­lustrates the benefits of the DEX methodology through us­ing the two socioeconomic models for tasks such as what-if and comparative analysis for the countries and data from the Yearbook, as well as for Slovenia — a country at the tirne of the writing of this paper not (yet) enlisted in the Vearbook but interesting since being a country in transi­tion. Section 5 summarizes the results and concludes the article. 2 Hierarchical multi-attribute decision models Hierarchical multi-attribute decision models as used by DEX consist of criteria tree and utiUty functions. Fig­ure 2 shows a simple decision model — constructed only for illustrative purposes — to assess the quality of coun­try's knowledge infrastructure from the quality of educa­tion, telecommunication netvvork and computer deploy­ment. Knovvledge infrastructure (ki ) is the overall utility or a target criterion, located at the root of the tree, that is modeled and derived from a set of basic criteria which are found at leaves of the criteria tree and which include the level of general education (educ) , the quality of telecom­munication network (tel ) and the level of computer de­ployment (comp). The basic criteria are those that can be measured and/or obtained for specific country. The criteria tree also includes an internal node, which is an intermedi­ate criterion that assesses the quality of technical infras­tructure (in f r) . Both k i and in f r are also referred to as aggregated criteria, as their value is determined from the values of other criteria in the criteria tree {e.g., in f r from comp and tel , and k i from edu c and in f r) . The ag­gregated criteria are those that can not be directly observed or measured, but are besides the target criterion useful to be modeled. For the real-world problems, a criteria tree would include several aggregated criteria, depending on a complexity of the domain being modeled. educ infr ki lov? low med med high low high low low med med high high high med high low med high high med low high high fedud comp tel infr low low low low high med med low low med high high high low med high high high fcompj te l Figure 2: A simple decision model with three basic cri­teria (educ , comp, tel ) and one intermediate criterion (infr) . DEX ušes qualitative criteria, i.e., every criterion in the criteria tree is assigned a finite value domain. In our čase. CONSTRUCTION AND APPLICATION OF HIERARCHICAL... the value domain for ki , educ , and comp is {low, med, high } and the value domain for te l is {low, high} . Utility functions are used to compute the value of aggre­gated criteria. DEX utility'functions use soTcalled if-then: decision rules, where each rule includes a specific combi­nation of values for criteria entering the criteria function (the i/part) and associated utility (the then part of the rule). These rules can then be represented with utiUty table. For example, in Figure 2, the utility function for in f r spec­ifies that when comp is lo w and te l is low, the value of the technical infrastructure is lo w (the first line in the utility table). Differently, when comp is med and te l is high , the value of in f r is hig h (the fourth line in the utility table). Within DEX, the rules in utility tables are defined man­ually, most often in a setup where a domain expert collabo­rates with a knovvledge engineer. Once a sufficient number df the rules for some aggregated criteria have been entered, DEX assists in the elicitation of the new rules by propos­ing a viable set of values of the corresponding aggregated criteria. The complete process of defining the criteria struc­ture and utility tables typically takes from one to five days, where a definition of criteria tree is often a more demand­ing task. DEX models are evaluated from bottom up, starting at the aggregated criteria that depend solely on basic criteria to finally derive the overall utility. For our model from Fig­ure 2, aggregated criterion in f r is evaluated first based on criteria comp and tel , and then the overall utility k i is obtained from the values of edu c and in f r . For exam­ple, given the values of basic criteria for the two countries A and B from Table 1, the same Table shows the derived value of the intermediate criterion and the overall utility. 3 Socioeconomic models for knowIedge infrastructure and the quality of political and economic system Using DEX modeling paradigm, we have developed two different socioeconomic models, the first one modeling the level of the knovvledge infrastructure (Knovvledge infras­tructure model) and a second one the quality of political and economic system with respect to their support of the economy and business (System model). Each model ušes a separate set of basic criteria taken from the World Compet­itiveness Yearbook (Table 2). The Knovvledge infrastructure target criteria (KI) rep­resents the level of development of knovvledge infras­tructure to support business and economic development. The KI model employs the criteria hierarchy as given in Figure 3. The model incorporates the utilization (IT_USAGE) and level of development of Information technology (TEC_INFRA) and the quality of education (EDUCATION). The general education with regard to IT Informatica 26 (2002) 47-56 depends on computer literacy (C_LIT) and the overall quality of general education (GEN_EDUCATION). The development of technological infrastructure is estimated from diffusion of computers (C_INFRA) and the state of development of telecommunications (TELECOMM), vvhich in turn depends on the current level and de­velopment potential of telecommunication infrastructure (TEL_INFR_INV) and accessibility and diffusion of tele­phones (TELEPHONES). The quality of political and economic system in regard to their support of the economy and business is modeled as a target criteria SYSTEM. Its dependency on intermediate and basic criteria is outlined in a criteria tree shown in Fig­ure 4. The value of the S YSTEM depends on the quality of government and economic system (QUAL_GOV_ECON), vvhich aggregates the value of economic system and poli­cies (ECONOMIC) and quality of government with re­spect to the support economy (GOV_QUALITY) and on the qualityof politics and public trust (QUAL_POL). The later aggregates the values of quality of system and poli­cies (POLITICAL) and the value of public trust to the cur­rent political system (TRUST). In its quality of govern­ment subtree, the model includes also the aggregated crite­ria that estimate the impact of lobbying (LOBBYING) and the governmerit effectiveness and openness (EFFECT). The knovvledge infrastructure model ušes 12, vvhile the model for system ušes 15 basic criteria. Each basic crite­rion has a domain of four values labeled "1 " to "4", vvhere "I " denotes the "vvorst" value of the criterion, i.e., the one that has a negative influence to the value of the target crite­rion, and "4 " denotes the "best" value of the criteria, again with respect to the influence to the target criteria. In this sense, the criteria values are nominal and ordered. The same domain definition was used for aH aggregated criteria. Together, the tvvo models define 17 aggregated crite­ria. Presenting ali utility functions defined is beyond the scope of this article, and for illustrative reasons we pro­vide only an example. Consider thus one of the utility functions for knovvledge infrastructure model that aggre­gates the value of educational system (EDUC_SYS) and in-company training (TRAINING) to the value of aggre­gated criteria for general education (GEN_EDUCATION, see Table 3). The utility function defines ali 16 possible combinations of values for EDUC_SYS and TRAINING. For example, consider the rule number 7, vvhich states that the value of general education level is 3 if the quality of educational system is 3 and in-company training is 2. We found that this pointvvise definition of utility functions pro­vides means to straightforvvard elicitation of knowledge from experts, since the experts find relatively easy to an­svver concrete questions (such as, "what is GEN_EDUCAT if the level for EDUCAT_SYS is 3 and TRAINING is 2"). Pointvvise definition allovvs for defining non-linear func­tions. For example, in the function for GEN_EDUCAT the outcome never exceeds 2 if one of the input criteria (EDUC_SYS and TRAINING) has the value of 1. Non­linearity in the aggregate function for GEN_EDUCAT is Informatica 26 (2002) 47-56 M. Krisper et al. Basic criteria Aggregated criterion Overall utility Option name educa t comp te l infr a k i Country A med lo w lo w lo w lo w Country B hig h hig h lo w med hig h Table 1: Evaluation results for countries A and B Knowledge infrastructure MANAG_IT Management of Information teclinology: utilization of and familiarity witli information teclinoiogy by management IT Information technology: exploiting by companies C_LIT Computer literacy among employees EDUC_SYS The educational system: educational system meets the needs of competitive economy TRAINING In-company training: investing of companies in fraining of their empIoyees C_USE Computers in use: share of worldwide computers in use C_PC Computers per capita: number of computers per person INF_REQ Infrastructure requirements, Telecommunications: Extend to which infrastmcture meets business re­ quirements TEL_INFR Telecommunications infrastnicture INVEST State investments in telecommunications TEL_LINES Telephones: number of main lines in use per 1000 inhabitants TEL_COST International telephone costs System INTER F State interference: State interference does not hinder die development of business SUBSID Subsidies: Government subsidies are directed tovvards future vvinners CONTROL Control of enterprises: State control of enterprises does not distort fair competition in the country IMP„PRAC T Improper practices (such as bribing and corruption) EXTEN T Lobbying; Extent to which lobbying accelerates govemment decision making INT_GROUPS Lobbying by special interest groups RESPONS Government responsiveness: Ability to quickly adapt policies to new realities DECENTRAL Administrative decentralization: Decision-making independence of local/regional authorities from central government PUB_SEC Public sector contracts: openness to foreign bidders POLICIES Government economic policies: Extend to vvhich government adapts its policies to new realities effec­ tively ADAPTATION Political system: Extend to which political system is well adapted to today's economic challenges TRANSPAR Transparency of government towards citizens POL_RISK Political risk rating GOV_POL Government policies: Supporting by puhlic consensus SUPPORT Public consensus and support for economic policies Table 2: List of basic criteria used by Knowledge infrastructure and System model, respectively. Figure 3: Criteria hierarchy for Knowledge infrastructure model. Figure 4: Criteria hierarchy for System model. CONSTRUCTION AND APPLICATION OF HIERARCHICAL. Informatica 26 (2002) 47-56 51 rule # EDUC_SYS TRAINING GEN_EDUCAT 1. 1 1 1 2. 2 1 1 3. 3 1 2 4. 4 1 2 3. 1 2 1 6. 2 2 2 7. 3 2 3 8. 4 2 3 9. 1 3 2 10. 2 3 3 11. 3 3 3 12. 4 3 4 13. 1 4 2 14. 2 4 3 15. 3 4 3 16. 4 4 4 Table 3: An example of a utility function defined within the knowledge infrastructure model. also evident from a graphical presentation of decision rules (Figure 5). GEN_EDUCAT Figure 5: A graphical presentation of utility function from Table 3. The pointvvise definition of utility functions follovvs a case-based human way of thinking and as such implic­itly States the relevance of each of the criteria. In prac­tice, besides requiring linear relationships between input and aggregated criteria, eliciting explicit weights from the expert is usually a difficult task, as it forces the expert to think in more abstract way [2]. Note that not ali of these vvere manually defined by expert, since DEX incorporates a mechanism that, based on the currently entered rules, pro­vides suggestions for the rules not defined. In practice, we needed to define only about one half of the rules in util­ity functions for the two socioeconomic models — for the other half the expert most often accepted the suggestions provided by DEX. 4 Socioeconomic models in use To demonstrate the applicability of the models defined in the previous section, we have first prepared the data set to be used. The models vvere buiit such that their set of basic criteria was taken from the list of criteria included in the World Competitiveness Yearbook (WCY, [6]). ObviousIy, the data of the countries included in WCY constitutes our basic data set. For each of the criteria from WCY, the values were first ordered such that low values would potentially lower the modeFs outcome (final criterion) and that high values would increase it. Since DEX models require criteria to be qualitative {i.e., "1" , "2", "3" , and "4"), the criteria val­ues needed to be discretized. Discretization used quantiles, such that each resulting qualitative value would represent roughly the same number of countries for that criterion. Note that in this setup the qualitative values of criteria can also be interpreted as: "1 " as "low", "2 " as "below aver­age", "3 " as "above average", and "4 " as "high". The models developed can be used in a number of dif­ferent ways. First, the models and their utility functions may provide additional insight to the domains. Next, the models can be used to evaluate the countries' data and de­rive corresponding values for aggregated and final criteria. The differences between two or more countries can then be studied by means of graphical comparison of criteria val­ues. Finally, a specific country may be studied to see the effect of changing the values of basic criteria and studying its good and bad points. 4.1 Analysis of the model The decision model as such can be analyzed locally by in­specting each of the defined utility function or globally by observing the overall impact of basic criteria on the target criterion. For the first task, DEX provides several tools. First, the utility functions can be visualized by selecting two input criteria and observing the output of aggregated function (see Figure 5 for an example). Another interest­ing DEX's tool is construction of aggregated rules from the set of elementary rules. For instance, an example of utility function that represents the function from Table 3 but is ex­pressed by aggregated rules is given in Table 4. Note that instead of 16 there are just 9 rules required to define the aggregated criteria GEN_EDUCAT. Also, the utility func­tion is much easier to comprehend. For example, from the last ruie it is easy to see that GEN_EDUCAT can reach the highest value (4) only when EDUC_SYS is 4. Further­more, the first two rules indicate that GEN_EDUCAT is 1 whenever one of the input criteria is 1 and the other less than or equal to 2. We have further used both socioeconomic models to es­timate the relevance of basic criteria to the value of the tar­get criteria. For these, from each model a dataset was con­structed that consisted of only basic criteria values and cor­responding value of a target criterion. We have arbitrarily sampled each model with about 2000 such "data points", and then used the information measure (IM) score as de­fined in [13] to estimate the relevance. IM was originally used in recursive partition algorithms for decision tree in­duction to identify most appropriate {i.e., important) crite­ria for decision tree nodes [15]. The criterion importance Inforraatica 26 (2002) 47-56 M. Krisper et al. agg. rule # EDUC. .SYS TRAINING GEN_EDUCAT 1. I < 2 1 2. < 2 1 1 3. > 3 1 2 4. 2 2 2 5. 1 > 3 2 6. 3 > 2 3 7. > 3 2 3 8. 2:3 > 3 3 9. 4 > 3 4 Table 4: An example of aggregated rules for utility function from Table 3. is assessed in independence of the other basic criteria: only the relationship with the target criterion is observed. For the two socioeconomic models, the basic criteria are ranked according to their importance in Table 6. For knowledge infrastructure model, the three most important basic criteria are management of Information technology, computer literacy and the value of education system. The three most important criteria from the System model are the control of enterprises, the level of state interference, and government subsidies. These results in general meet experts' intuitive expectations. 4.2 Comparative data analysis The Knovvledge infrastructure and System models were used to derive the value of the corresponding target crite­ria (KI and SYSTEM, respectively). Although DEX can be used for this task, another system called Vredana [17] vvas employed instead. Besides graphical presentation, the unique feature of Vredana is that it can evaluate each coun­try not only to a single qualitative value of the target crite­rion, but can also estimate country's relative position within this range. For example, consider that the two countries having the values of EDUC_SYS and TRAINING 1 and 1 or 2 and 1, respectively, would both be classified to 1 for GEN_EDUCAT (see Table 3). In such čase, Vredana would — within the qualitative value of 1 — rate the first country a bit lovver than the second one by assigning a lower quantitative adjustment to the first country. In gen­eral, the gain of such rating is that Vredana allows further differentiation of the countries that were evaluated to the same qualitative rank. Since it is beyond the scope of this paper to further describe Vredana's evaluation algorithm, please see [4] for details. Before presenting the results of comparative analysis, we needed to consider that the World Competitiveness Year­book data we have used contains missing values. Both DEX and Vredana can properly handle these by deriving a range of values (probability distribution) for aggregated and final criteria. Although this is often a very desired fea­ture, the requirement for the analysis in this section vvas that we required to unambiguously rank the countries and thus we needed crisp evaluation outcomes. For this pur­ rank 1 2 3 4 5 6 7 8 9 10 11 12 rank 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 criterion MANAG_IT C Lrr EDUC_SYS TRAINING IT TEL_LINES C PC INF_REQ TEL_COST TEL INFR C USE INVEST cntenon CONTROL INTERF SUBSID 1MP_PRACT POLICIES EXTENT PUB_SEC RESPONS ADAPTATION DECENTRAL INT_GROUPS GOV_POL SUPPORT POL_RISK TRASPAR IM 0.2200 0.1297 0.0694 0.0618 0.0576 0.0201 0.0174 0.0164 0.0102 0.0097 0.0043 0.0033 IM 0.1526 0.1463 0.1404 0.0438 0.0351 0.0263 0.0202 0.0186 0.0125 0.0092 0.0059 0.0054 0.0052 0.0019 0.0012 Table 5: Ranking of basic criteria from Knowledge infras­tructure (above) and System (below) model. pose, the missing values were estimated as follows. If a country C vvas having a missing value for some criterion, we first found three other countries having most similar GDP/capita to the country C. Next, we replaced a miss­ing criterion value of C with the average for this criterion over the three countries found. Using the above introduced schema for handling miss­ing values, the results of evaluation for both models are given in Figure 6. Note that in terms of the knovvledge in­frastructure, Finland, Svveden, Singapore, Hong Kong, and Germany rank the highest. It vvould be expected that USA vvould rank very high here, but additional analysis shovvs that it is ranked in class "3 " because of the low value of general education. The specific comparison of Finland and USA that also highlights this deficiency is shovvn in Fig­ure 7. In terms of the quality of political and economic system in regard to their support of economy and business Figure 6 shovvs that Hong Kong, Singapore, and Malaysia — the Tiger countries — are the ones that rate the highest. We have further explored the relation betvveen country ratings of the tvvo models by means of correlation coef­ficients. Three other ratings were used as well based on the follovving measures: GDP/capita, average value of cri­teria computers per capita (C_PC) and Information tech­nology (IT) from Knovvledge infrastructure model (sel.KI), and average value of criteria government economic policies (POLICIES) and adaptation of political system (ADAP­TATION) from System model (sel.SVSTEM). The corre­lation coefficients are given in Table 6. Note that com­pletely correlated ranks would have a coefficient of 1, and CONSTRUCTION AND APPLICATION OFHIERARCHICAL.. Informatica 26 (2002) 47-56 Fll» i i , 4 ; GER ji.AU S i AUT i N" Ml ' 3 N^'- MAL : . THA i. IRE i 2 : KOR • • i » UNIPO' ; rm [ ^ IND 1 • ARE POR ; ( ^ ; 1 ' 1 ' 1 Figure 6: Vredana's graphical representation of the results of the evaluation for Knowledge infrastructure and System model (Figure on the right shows an enlargement of the quadrantKI=l and SYSTEM=1). Criteria USA Finland KI 3 4 IT^USAGE 4 4 MANAG IT 4 4 IT 4 4 EDUCATION 2 4 C LIT 3 4 GEN EDUCAT 2 4 EDUC SYS 2 4 TRAINING 2 4 TEC_INFRA 4 4 C_INFRA 4 4 C_USE 4 3 C_PC 4 4 TELECOMM 4 4 TEL INFR INV 3 4 INF_REQ 4 4 TEL_INFR 4 4 INVEST 1 2 TELEPHONES 4 4 TEL LINES 4 4 TEL_COST 3 4 Figure 7: Comparison of criteria for knovvledge infrastruc­ture for USA and Finland. The difference between the two countries (3 to 4) can be contributed to the differences of quality of the education (2 to 4), of which the subtree is printed in bold. SVSTEM sel.KI sel.SVS GDP/cap KI 0.452 0.822 0.201 0.658 SVSTEM 0.377 0.754 0.150 sel.KI 0.216 0.730 sel.SVSTEM 0.050 Table 6: Correlation coefficients for ranks obtained from the two socioeconomic models (KI and SVSTEM), se­lected basic criteria from each model (sel.KI and sel.SYS) and GDP/capita. uncorrelated a coefficient of 0. The ranks of the two mod­els are found weakly correlated (0.452). Not surprisingly, GDP/capita correlates better with Knowledge infrastruc­ture than with System (0.658 > 0.150). As expected, the outcome of the two models best correlate with the ranks derived from the two averaged selected criteria, i.e., sel.KI and sel.SYSTEM respectively. A more focused ways of comparing the countries in Vredana in shown in Figure 8. The user selected four coun­tries (Svveden, Austria, Poland, and Slovenia, and three cri­teria (political and economical system, and knovvledge in­frastructure) upon vvhich these countries are compared. For Slovenia, the values of the basic criteria vvere estimated by local experts. For the analysis in Figure 8 we did not replace the unknovvn values, so one can observe that for Poland and Slovenia the minimal and maximal value for specific criteria is shovvn (for example, for Slovenia, the value of political system lies vvithin 2 and 3). The radar charts shovv that there is a balance among knovvledge in­frastructure, political and economical system for the tvvo highly developed members of the EU, whe:reas for the tvvo associated countries in transition Poland and Slovenia it is evident that knovvledge infrastructure and economical sys­ Informatica 26 (2002) 47-56 Figure 8: A snapshot of Vredana showing a radar chart that compares four countries with respect to three selected criteria. tem do not follovv yet the positive changes in political sys­tem. 4.3 What-if analysis For an example of "what-if" analysis, we have studied Slovenia through the knovvledge infrastructure model. Ini­tially, Slovenia evaluates to "2-3 " which ranks it into mod­erately developed countries in this respect. The question posed to "what-if" analysis is whether its knovvledge in­frastructure will be improved provided that Slovenia pri­vatizes telecommunications. Namely, at present the Tele­com Slovenia is the only telecommunication provider in the country (until 2001), thus holding a complete monopoly. The privatization of telecommunications in the European countries (including those in transition) boosted the devel­opment increasing both the quality of infrastructure and services. We have simulated such čase and raised the values of basic criteria TELJNFR and TEL_COST to 4. The two evaluated criteria trees, /. e., the one for original data and the one with adjusted values due to privatization in telecommu­nications, are shown in Figure 9. According to the model, it is TEL_INFR whose improvement propagates through the intermediate criteria of telecommunication infrastruc­ture ali the way up to the target criteria, such that the new value of knovvledge infrastructure is 3. We have additionally attempted to change the value of management and information technology criteria (MANAG_IT). Increasing utilization and familiarity with IT by management is an already undergoing process, so we expect changes it this area in the near future. Ali other ba­sic criteria being equal, raising MANAGJT from "2-3" to "3-4" first results in increased IT_USAGE from 2-3 to 3-4, which finally results in improvement of knovvledge infras­tructure to 3 from previously 2-3. Overall, we found the "vvhat-if" analysis by DEX as ex­emplified above a very flexible and useful tool, especially as it provides the explanation through tracing of criteria tree M. Krisf>er et al. of why and to what degree did the changes influenced the final score. This feature of moders transparency further­more increases decision maker's confidence to the model and veracity of results. Criteria SLO SLO* KI 2:3 3 IT_USAGE 2:3 2:3 MANAGJ T 2:3 2:3 IT 3:4 3:4 EDUCATION 3 3 C_UT 3:4 3:4 GEN EDUCAT 3 3 EDUC_SYS 4 4 TRAINING 2 2 TECJNFRA 2 3 C_INFRA 2 2 C_USE 1 1 C_PC 3 3 TELECOMM 2 3 TEL_INFR_1NV 2 3 INF_REQ 2:3 2:3 TEL INFR 2 4 INVEST 2 2 TELEPHONES 3 3 TEL LINES 3 3 TEL_COST 2 4 Figure 9: Original (SLO) and modified (SLO*) evaluated criteria tree for Slovenia considering the pending changes in privatization of telecommunications. The differences are highlighted (criteria printed in bold). 4.4 Advantages and Disadvantages Another feature of DEX that can support socioeconomic data analysis is the display of advantages and disadvan­tages for some selected country. Advantageous criteria are considered to be those that have especially positive effect to the value of the target criteria. Criteria that potentially most lovver the final outcome are considered as disadvantages. An example of advantages/disadvantages analysis for Japan using knovvledge infrastructure model is shovvn in Figure 10. One can see that the major advantages of this country are in the area of usage of IT and in education, while the only disadvantage is the International telephone cost. Note that both disadvantages and advantages are shovvn as the criteria subtrees, so one can easily trace the propagation of positive (negative) effects through the crite­ria tree. For Japan, vve can see that the advantageous crite­ria propagated ali the way up in the IT usage and education subtrees, but these advantages vvere not strong enough to make the final outcome of maximal grade of the highest grade (the value of knovvledge infrastructure evaluates to 3). 5 Conclusion We have described the DEX paradigm to construction of hierarchical decision-support models and presented a čase study to show hovv it can enable the efficient construction and application of socioeconomic models. In particular, CONSTRUCTION AND APPLICATION OF HIERARCHICAL. Advantages: rr.USAG E MANAG_IT EDUCATION C_LIT GEN_EDUCAT EDUC_SYS TRAINING C_USE 4 INF_REQ 4 Disadvantages: TEL_COST 1 Figure 10: Advantages and disadvantages for Japan. - DEX enables an efficient model construction that con­sist of identification of hierarchical structure and con­struction of rules for aggregated criteria. Since the original problem (mapping of many basic criteria to final criteria) is decomposed by introduction of in­termediate criteria, the aggregation functions include only a few attributes and can be efficiently specified by means of pointwise rules elicited from the experts. - The use of intermediate criteria not only decomposes a problem of model construction to simpler subprob­lems, but also makes these intermediate criteria ob­servable — this is specifically useful in application of the model, since it can provide structured explanation and can ease the process of data analysis. - Once the models are built, DEX can provide further inspection to aggregated functions by means of visu­alization and of presenting rules in an aggregated way. Moreover, the models can be used to study the overall relevance of the basic criteria. - Data is provided to DEX models in terms of values for the basic criteria. DEX evaluates the data (derives the values of aggregated criteria) and can additionally be used to answer what-if questions, compare options (data for different countries), and structurally outline the advantages and disadvantages of each option. - In addition to DEX, a Vredana tool can be used to visualize the data and compare the options. When constructing and applying the socioeconomic models for knowledge infrastructure and value of political and economical system, we found ali of the above advan­tages of the DEX and Vredana approach very useful. Of specific help was Vredana tool, which we believe should be the tool of the choice for performing what-if analysis and comparative studies. The weakness of the proposed approach is the fact that DEX and Vredana are available only as a separate tools that communicate through common model and option definition data file. It is expected that on­going work on their integration will not only make data analysis more efficient, but will enable a deeper analysis of the model, such that, for example, the effects of changing Informatica 26 (2002) 47-56 55 the aggregation rules on the values of aggregated criteria for some set of options (countries) could be immediately observed through visualization. Another possible methodological improvement is a func­tion decomposition technique [18] to model development. Namely, in the čase where a dataset exists that gives the values of the target criterion for a number of combinations of basic criteria, the aggregated functions can be automati­cally induced from the dataset. This data mining approach can potentially shorten the model development tirne as well as maintain the integrity of the model with some preexist­ing classified data. A pilot study that used this framework for construction of knovvledge infrastructure model is de­scribedin [12]. We have proposed two different socioeconomic models, first modeling the value of country's knowledge infrastruc­ture and second modeling the quality of political and eco­nomic system in regard to their support of economy and business. There are of course many other interesting so­cioeconomic models that could be employed in drilling in the country's socioeconomic data, getting insight to its present state and constructing and evaluating its potential future development scenarios. In our further work, we plan to extend the existing and construct new socioeconomic de­cision support models and correspondingly extend the data base of basic criteria using different sources, including the Yearbook of International Institute for Management Devel­opment, World Bank, International Monetary Found, and Institute for Economic Research from Ljubljana, Slovenia. As proposed in this article, these models will be built, inte­grated in and applied to support decision making and data analysis within DEX-Vredana framework. References [1] M. Bohanec, B. Cestnik, and V. Rajkovič. A manage­ment decision support system for allocating housing loans. In R Humphreys, L. Bannon, A. McCosh, and P. Migliarese, editors, Implementing Systemfor Sup­porting Management Decisions, pages 34—43. Chap­man & Hali, London, 1996. [2] M. Bohanec and V. Rajkovič. DEX: An expert system shell for decision support. Sistemica, 1(1): 145-157, 1990. [3] M. Bohanec, V. Rajkovič, B. Semolič, and A. Pogačnik. Knowledge-based portfolio analysis for project evaluation. Information & Management, 28:293-302,1995. [4] M, Bohanec, B. Urh, and V. Rajkovič. Evaluating op­tions by combined qualitative and quantitative meth­ods. Acta Psychologica, 80:67-89, 1992. [5] V. Chankong and Y. Y. Haimes. Multiattribute de­cision making: Theory and methodology. North-Holland, 1983. Informatica 26 (2002) 47-56 1^6] C. Decosterd, S. Garelli, M. Hediger, M. Linard de Guertechin, and C. Travers. The World Compet­itiveness Yearbook. International Institute for Man­agement Development, Lausanne, Svvitzerland, 1996. [7] J. Deior. Growth, competitiveness, employment: The challenges and ways forward into 21st century. White Book, 1997. [8] J. Efstathiou and V. Rajkovič. Multiattribute decision-making using a fuzzy heuristic approach. IEEE Trans, on Systems, Man and Cybemetics, 9:326-333, 1979. [9] M. Krisper, V. Bukvič, V. Rajkovič, and T. Sagadin. Strategic planning with expert systems based portfo­lio analysis. In Proč. IITT International Conference EXPERSYS-91, pages 283-288, Pariš, France, 1991. [10] M. Krisper and L. Sočan. Heuristic modelling of socio-economic scenarios. In Z. Kaltnekar, editor, Proč. International Conference on Organisation and Information Systems, pages 64-72, Bled, Slovenia, 1992. [11] M. Krisper, L. Sočan, and T. Zrimec. Expert sys­tem for the measurement and forecasting of socio­economic development. In Proč. lASTED Interna­tional Conference, Geneve, Svvitzerland, 1987. [12] M. Krisper and B. Zupan. Synthesis of hierarchical decision support models from socioeconomic data. In C. Bavec and M. Gams, editors. Proč. Information So­ciety Conference, pages 60-63, Ljubljana, Slovenia, 1998. [13] J. Mingers. An empirical evaluation of selection mea­sures for decision-tree induction. Machine Learning, 3:319-342,1989. [14] M. Olave, V. Rajkovič, and M. Bohanec. An appii­cation for admission in pubiic school systems. In I. Th. M. Snellen, W. B. H. J. van de Donk, and J.-R Baquiast, editors, Expert Systems in Public Adminis­tration, pages 145-160. Elsevier Science Publishers (North Holland), 1989. [15] R. Quinlan. Induction of decision trees. Machine Learning, 1(1):81-106,1986. [16] V. Rajkovič, M. Krisper, T. Sagadin, and S. Grubar. An expert systems-based portfolio analysis. In Proč. IFIP Congress, Madrid, Spain, 1992. [17] A. Šet, M. Bohanec, and M. Krisper. Vredana: A pro­gram for evaluation and analysis of options in multi­attribute decision making (in slovene). In F. Solina and B. Zaje, editors. Prof Fourth Electrotechnical and Computer Conference ERK'95, pages 157-160, Portorož, Slovenia, 1995. M. Krisper et al. [18] B. Zupan, M. Bohanec, J. Demšar, and I. Bratko. Fea­ture transformation by function decomposition. IEEE Intelligent Systems & Their Applications, 13(2):38­43, March/April 1998.