A Hierarchical Fuzzy Classification of Online Customers - Diuf-Unifr

A Hierarchical Fuzzy Classification of Online Customers Nicolas Werro University of Fribourg Department of Informatics Bd. de Pérolles 90 1700 Fribourg, Switzerland [email protected]

Henrik Stormer University of Fribourg Department of Informatics Bd. de Pérolles 90 1700 Fribourg, Switzerland [email protected]

Andreas Meier University of Fribourg Department of Informatics Bd. de Pérolles 90 1700 Fribourg, Switzerland [email protected]

Abstract A key challenge for companies in the e-business era is to manage customer relationships as an asset. In today’s global economy this task is getting, at the same time, more difficult and more important. In order to retain the potentially good customers and to improve their buying attitude this paper proposes a hierarchical fuzzy classification of online customers. A fuzzy classification, which is a combination of relational databases and fuzzy logic, allows customers to be classified in several classes at the same time and can therefore precisely determine the customers’ value for an enterprise. This approach allows companies to improve the customer equity, to launch loyalty programs, to automate mass customization and to refine marketing campaigns in order to maximize the customers’ value and, this way, the companies’ profit.

1. Introduction The growing importance of the e-business in today’s economy forces the enterprises to adapt their behaviour towards the different actors of the market. This is particularly true for the customer relationship management (CRM) as the traditional means based on the human relationship are no more available. In this area, the customer retention and the cross/add-on selling are special issues because the global economy enabled by the Internet allows, on the one hand, the companies to offer their products or services worldwide and, on the other hand, also allows the customers to easily compare the different products/services and their prices. This paper proposes a new approach for managing customer relationships based on the customer equity principle by the mean of a fuzzy classification [9].

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

Unlike a sharp classification, a fuzzy classification allows elements to be classified in several classes at the same time. The notion of partial membership in the different classes provides more information by integrating the potential and the possible weaknesses of the classified elements. This approach can therefore precisely determine the customers’ value according to an enterprise. In order to better retain the potentially good customers and to improve their buying attitude, the fuzzy classification approach can improve the customer equity, launch loyalty programs, automate mass customization and refine marketing campaigns. The fuzzy classification approach has two other important advantages. First, it allows the users to work on a semantic level, i.e. with linguistic expressions. Secondly, it offers the possibility of decomposing complex classifications into a hierarchy of classifications. The decomposition mechanism allows classifications to keep a small number of classes with a proper semantic even if many attributes are taken into account. More important, it allows marketers to precisely analyse the potential and the weaknesses of the classified customers in the different levels of the hierarchy. The remainder of the present paper has the following structure: Section 2 introduces the fuzziness with relational databases as well as the fuzzy classification concept, query language and architecture. Based on the fuzzy classification approach, the concepts of customer equity, mass customization, marketing campaign and decomposition are explained in Section 3. The definition of the hierarchical classification space based on available attributes of online customers is then described in Section 4. Finally, Section 5 gives a conclusion and an outlook.

2. Fuzzy Classification 2.1. Database & Fuzziness In practice, information systems are often based on very large data collections, mostly stored in relational databases. Due to an information overload, it is becoming increasingly difficult to analyze these collections and to generate marketing decisions. In this context, a toolkit for the analysis of customer relationships which combines relational databases and fuzzy logic is proposed. Fuzzy logic, unlike statistical data mining techniques such as cluster or regression analysis, enables the use of non-numerical values and introduces the notion of linguistic variables. Using linguistic terms and variables results in a more human oriented querying process. The proposed toolkit reduces the complexity of customer data and extracts valuable hidden information through a fuzzy classification. The main advantage of a fuzzy classification compared to a classical one is that an element is not limited to a single class but can be assigned to several classes. Furthermore, each element has one or more membership degrees which illustrate to what extend this element belongs to the classes it has been assigned to. The notion of membership gives a much better description of the classified elements and also helps to find out the potential or the possible weaknesses of the considered elements. The fuzzy classification is achieved by extending the relational database schema with a context model. A fuzzy Classification Query Language (fCQL) can directly operate on the underlying database so that no migration of the raw data is needed.

2.2. Fuzzy Classification with Linguistic Variables In order to define classes in the relational database schema, we extend the relational model by a context model proposed by Chen [4]. This means that to every attribute Aj defined by a domain D(Aj), we add a context K(Aj). A context K(Aj) is a partition of D(Aj) into equivalence classes. A relational database schema with contexts R(A,K) is then the set A=(A1,…,An) of attributes with associated contexts K=(K1(A1),…,Kn(An)) [14]. Throughout Sections 2 and 3, a simple example of relationship management is used as a concrete fuzzy classification is exposed in Section 4. In this example, customers will be evaluated based on only two attributes, turnover and payment behaviour.

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

• Turnover in dollars per month: The attribute domain is defined by [0, 1000] and is divided into the equivalence classes [0, 499] for low turnover and [500, 1000] for high turnover. • Payment behaviour: The domain of the attribute {in advance, on time, behind time, too late} has the two equivalence classes {in advance, on time} for an attractive payment behaviour and {behind time, too late} for an unattractive one. To derive fuzzy classes from sharp contexts, the qualifying attributes are considered as linguistic variables, and verbal terms are assigned to each equivalence class [19]. With the help of linguistic variables, the equivalence classes of the attributes can be described more intuitively (see Figure 1). In addition, every term of a linguistic variable represents a fuzzy set. Each fuzzy set is determined by a membership function μ over the domain of the corresponding attribute (see Figure 2).

Figure 1. Concept of linguistic variable

The definition of the equivalence classes of the attributes turnover and payment behaviour determines a two-dimensional classification space shown in Figure 2. The four resulting classes C1 to C4 could be characterized by marketing strategies such as ‘Commit Customer’ (C1), ‘Improve Loyalty’ (C2), ‘Augment Turnover’ (C3), and ‘Don’t Invest’ (C4).

Figure 2. Fuzzy classification space defined by turnover and payment behaviour

With the context model, the usage of linguistic variables and membership functions, the classification space becomes fuzzy. This fuzzy partition has an important outcome, it implies the disappearance of the classes’ sharp orders, i.e. there are continuous transitions between the different classes. This means that a customer can belong to more than one class at the same time and that his membership degrees in the different classes can be calculated.

Then the classification results are displayed to the user or returned to the application. Before querying the fCQL toolkit, the data architect has to define the fuzzy classification (see case 3).

2.3. Fuzzy Classification Query Language The classification language fCQL is designed in the spirit of SQL [12]. Instead of specifying the attribute list in the select-clause, the name of the object column to be classified is given in the classify-clause. The from-clause specifies the considered relation, just as in SQL. Finally, the where-clause is changed into a withclause which specifies a classification predicate. An example in customer relationship management could be given as follows: classify Customer from CustomerRelation with Turnover is high and PaymentBehaviour is attractive

This classification query would return the class C1 (Commit Customer) defined as the aggregation of the terms ‘high turnover’ and ‘attractive payment behaviour’. The aggregation operator is the γ-operator which was suggested as compensatory and was empirically tested by Zimmermann and Zysno [18]. In this simple example, specifying linguistic variables in the with-clause is straightforward. However if customers are classified on three or more attributes, the capability of fCQL for a multidimensional classification space is increased. This can be seen as an extension of the classical slicing and dicing operators on a multidimensional data cube.

2.4. Architecture of the fCQL Toolkit The architecture of the fCQL toolkit shown in Figure 3 illustrates the interactions between the user, the fCQL toolkit and the relation database management system (RDBMS). The fCQL toolkit is an additional layer above the relational database system; this particularity makes fCQL independent of underlying database systems and thus enables fCQL to operate with every RDBMS. It also implies that the user can always query the database with standard SQL commands (see case 1). The user (resp. an application) can also formulate unsharp queries to the fCQL toolkit (see case 2). Those queries are analyzed and translated into corresponding SQL statements for the RDBMS.

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

Figure 3. Overview of the fCQL toolkit

2.5. Other Fuzzy Approaches A number of fuzzy query languages have been proposed in the literature. A well-known enhancement of the relational domain calculus is the Fuzzy Query Language proposed by Takahashi [15]. It provides a theoretical basis for the development of a humanoriented interface with relational databases. Another query language is the fuzzy SQL published by Bosc and Pivert [3]. This language allows gradual predicates interpreted in the framework of the fuzzy set theory. Finally, there is FQUERY from Kacprzyk and Zadrozny [7] which extends Microsoft’s Access in order to answer imprecisely specified questions. In practice, marketers do not use fuzzy classification approaches but clustering methods in order to derive a market segmentation. In this area, many contributions can be found showing the advantages of overlapping and fuzzy clustering methods compared to traditional means. For instance, Arabie et al. [1] showed the advantages of the overlapping approach for the product positioning and the market segmentation. An empirical comparison between classical and fuzzy clustering methods has shown that the latter has a greater internal validity [6]. Nguyen et al. [10] treated the opposition between the market size and the homogeneity of the segments and used the fuzzy clustering approach to adequately find a compromise between the two aspects. Setnes and Kaymak [13] also used fuzzy clustering for the target selection in direct marketing and could reach a substantial target improvement.

3. Fuzzy Customer Classes 3.1. Customer Equity Managing customers as an asset requires measuring and treating them according to their real value [2]. With sharp classes, i.e. traditional customer segments, this is not possible because all the elements within a class are treated the same way and there usually exists a strong discrepancy between the classes. In Figure 4 for instance, customers Brown and Ford have similar turnover as well as similar willingness to pay. However, Brown and Ford are treated differently: Brown belongs to the winner class C1 (Commit Customer) and Ford to the loser class C4 (Don’t Invest). In addition, a traditional customer segment strategy treats the top rating customer Smith the same way as Brown, who is close to the loser Ford.

their value for the enterprise [17]. For that purpose, a discount rate can be associated with each fuzzy class: for instance C1 (Commit Customer) gets a discount rate of 10%, C2 (Improve Loyalty) a discount of 5%, C3 (Augment Turnover) 3%, and C4 (Don’t Invest) 0%. The individual discount of a customer can then be calculated by the aggregation of the discount of the classes he belongs to in proportion to his membership degrees. With a fuzzy classification, the customers of Figure 4 get the following discounts: • Smith (C1:1.00, C2:0.00, C3:0.00, C4:0.00): 1.00*10% + 0.00*5% + 0.00*3% + 0.00*0% = 10% • Brown (C1:0.28, C2:0.25, C3:0.25, C4:0.22): 0.28*10% + 0.25*5% + 0.25*3% + 0.22*0% = 4.8% • Ford (C1:0.22, C2:0.25, C3:0.25, C4:0.28): 0.22*10% + 0.25*5% + 0.25*3% + 0.28*0% = 4.2% • Miller (C1:0.00, C2:0.00, C3:0.00, C4:1.00): 0.00*10% + 0.00*5% + 0.00*3% + 1.00*0% = 0% Using a fuzzy classification leads to a transparent and fair judgment: Smith gets the maximum discount and a better discount than Brown who belongs to the same customer class. Brown and Ford get nearly the same discount rate; they have comparable customer values although they belong to opposite classes. Miller, who sits in the same class as Ford, does not benefit from a discount.

3.3. Marketing Campaigns Figure 4. Customer equity example based on turnover and payment attitude

Those dilemmas can be adequately solved by the use of a fuzzy classification where the customers can belong to several classes. The notion of partial membership brings the disappearance of the sharp borders between the customer segments. Fuzzy customer classes better reflect the reality and allow companies to treat customers according to their real value. By driving the customer equity, a company can significantly improve the retention rate as well as the buying behaviour of potentially good to top customers.

3.2. Mass Customization According to the customer equity principle, the fuzzy classification approach allows marketers to automate mass customization. Indeed the membership degrees of the customers in the different classes can precisely determine the privileges the customers deserve, for example a personalized discount reflecting

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

Launching a marketing campaign can be very expensive. It is therefore crucial to select the most appropriate customers and then to be able to verify the impact of the marketing campaign in order to modify the target group or to improve the strategy. The fuzzy classification approach offers marketers convenient means for selecting customer subgroups and for measuring the efficiency of the marketing campaign. One strategy example, shown in Figure 5, could be to select customers without payment problems and with a low turnover in order to propose them new or premium products. Using membership functions, a subset of customers in class C1 has been chosen. The application of membership functions allows marketers to dynamically modify the size of the target group in order to respect the available campaign budget. Modifying the size of the target group is also a valuable mean to increase or decrease the homogeneity between the targeted customers depending if the proposed products are very specific or intended for a large public [10]. If the marketing campaign or a testing process has been launched, the fuzzy customer

classes can be analyzed again. It is then possible to verify the impact of the campaign by measuring the new value of the targeted customers, i.e. if the target group has moved in the direction of class C1.

Figure 5. Evolution of the target group due to the marketing campaign

Based on the same principle, a fuzzy classification allows marketers to monitor the evolution of single customers through the classes. By comparing the value of a customer over time, it is possible to detect if this customer is increasing, maintaining or decreasing his value. Based on those observations, marketing actions can be taken on the customer level. The most pertinent application of the customers monitoring is the detection of churning customers. With the fuzzy classification approach, an automated trigger mechanism that supervises the customers’ evolution based on several criteria can be easily implemented. If a good customer is suspected of having a churning behaviour, an alert is sent to the marketing department which can take the appropriate action in order to retain this customer.

3.4. Decomposition Principle Customers can be classified based on many attributes. In the classification example shown in Figure 2, customers were only evaluated regarding the turnover and the willingness to pay. This simple example cannot be effective in real case applications. On the other hand a fuzzy classification with many linguistic variables and terms leads to a multidimensional classification space with a large number of classes. By combining all the available attributes, it may not be possible to extract a clear semantic for each resulting class. This problem, also true for sharp classifications, is partially resolved by the use of the fuzzy classes. By having a continuous transition between the classes, only few equivalence

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

classes (linguistic terms) are required, reducing this way the number of final classes. The problem remains however for the number of dimensions (linguistic variables) which exponentially increases the number of classes. In order to keep a small number of classes with a proper semantic it is possible to decompose a complex fuzzy classification into a hierarchy of fuzzy classifications. By regrouping attributes of a given context in sub-classifications it is on the one hand possible to derive a precise definition of the classes and, on the other hand, to derive a new concept expressing a higher semantic than the basic attributes taken separately. The decomposition principle also reduces the complexity of the initial problem allowing a better definition and optimization of the different fuzzy sub-classifications. A good example of decomposition and a central perspective for a company is the customer loyalty. Many loyalty concepts have been proposed in the marketing literature. Harrison, for instance, proposes to express the customer loyalty based on two dimensions, the attachment and the buying behaviour of the customers as shown in Figure 6 [5].

Figure 6. Fuzzy classification for the concept of customer loyalty

The combination of the attributes attachment and buying behaviour leads to the concept of customer loyalty which is a very important indicator for an enterprise. By assigning grades of loyalty to the fuzzy classes it is then possible to derive the grade of loyalty of the customers towards the company. The results of this fuzzy classification can either be queried normally, either be used in a higher-order fuzzy classification in order to build a hierarchy. In other words, the decomposition principle allows marketers to merge given attributes in order to build more consistent and valuable concepts. Those concepts are then integrated as new dimensions in other fuzzy classifications leading to a hierarchy of fuzzy classifications.

4. Hierarchical Fuzzy Classification of Online Customers 4.1. Analysis of Online Customers The analysis of online customers compared to traditional ones has the advantage that a lot of information about the customers’ behaviour is automatically logged in the system. If an online shop system is considered, explicit and implicit information can be used for the analysis of the customer relationships. Explicit information is the data provided directly by the customer, like the product ratings, the forum entries and the orders, whereas implicit information is retrieved from his interaction with the system, e.g. the clickstream information. • Orders: the orders can determine the turnover and the margin of a customer as well as his buying frequency. Indirectly the payment delay and the return rate can also be identified. • Clickstream: the clickstream information can establish the visiting frequency. • Product ratings & forum entries: the ratings of products as well as forum entries can reveal the involvement frequency. The use of the order information is straightforward as most of the companies are using the turnover, margin, payment delay and return rate information in order to analyse and reward their customers. The clickstream data is a new source of information coming from the interaction of the users with the online shop. All the actions performed by the users are logged with a time stamp into the clickstream data allowing the system to determine the visiting frequency for each user [8]. Many online shops provide their customers a way to communicate and to share knowledge by means of product ratings and forum entries. This social involvement can lead to a virtual community which increases the trust and the attachment towards the company [11].

hierarchy expresses a concept defined by a fuzzy classification; therefore it is possible to derive for every customer the appropriate marketing actions in order to augment their value. For instance, the marketing department can increase the profitability of the customers by analysing them in the profitability sub-classification. At this level it can be observed if the weakness of a customer comes from a too low financial input, from excessive service costs or from a combination of the two aspects. By going further down, the origin of the problem can be precisely identified.

Figure 7. Hierarchical fuzzy classification

The top fuzzy classification shown in Figure 8 expresses the customer value and is based on the concepts of customer profitability and loyalty defined in the next sections. The profitability concept, which covers the financial context, is a main aspect as it expresses the real financial contribution of the customers towards the company. The loyalty concept, which expresses the strength of the relationship between the customers and the enterprise, is a strong indicator of the customers’ future buying behaviour. The resulting classification space is made up of four classes which have been assigned a customer value ranging from 0 to 100 in order to derive the individual value of the customers.

4.2. Hierarchy of Fuzzy Classifications Based on the above mentioned attributes and the decomposition principle explained in Section 3, it is possible to build a hierarchical fuzzy classification of online customers (see Figure 7). The use of a hierarchical fuzzy classification for the analysis of customers allows marketers, on the one hand, to extract the precise value of the online customers and, on the other hand, to analyse the potential and the weaknesses of the classified customers in the different levels of the hierarchy. Indeed, each level of the

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

Figure 8. Top fuzzy classification

4.3. Decomposition of the Profitability Concept The profitability of a customer is the difference between the gains he brings and the costs he generates. The sub-classification profitability is then defined by the concepts gain and service costs (see Figure 9).

Figure 11. Service costs sub-classification

4.4. Decomposition of the Loyalty Concept

Figure 9. Profitability sub-classification

The gain concept represents the profits of a customer’s purchases. This concept can be easily calculated based on the attributes turnover and margin derived from the order information (see Figure 10). Depending on the goods sold, the turnover can be either the total turnover of a given customer or a partial turnover realized during a given period.

The loyalty sub-classification, shown in Figure 12, is defined by the dimensions buying behaviour and attachment [5]. In contrast to the attachment, the buying frequency or repurchases attribute can be directly derived from the order information.

Figure 12. Loyalty sub-classification

Figure 10. Gain sub-classification

The service costs concept reflects the indirect costs that the customers can generate by returning products and by not paying on time (see Figure 11). The attribute return rate is the percentage of returned products; its domain has been limited to [0, 20] as 20 percents is already a non-acceptable return rate. The payment delay attribute represents the average delay in days of a customer after the product has been shipped. Note that the payment delay can be replaced by the qualitative attribute payment behaviour of Figure 2.

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

The attachment concept can be realized by the attributes visiting frequency and involvement frequency as shown in Figure 13. The involvement frequency has a deeper meaning than the visiting frequency as sharing experiences denotes a greater attachment as just browsing an online shop. The definition of a hierarchical fuzzy classification is highly context dependent. The choice of the attributes and their domain, the specification of the linguistic terms and the associated membership functions, the definition of the semantic and the value of the classes can only be wisely determined for a given context. The proposed hierarchical fuzzy classification is a generic example for the analysis of online customers in the context of an electronic shop.

Figure 13. Attachment sub-classification

5. Conclusion and Outlook The fuzzy classification approach is an effective mean for managing customer relationships. By providing an accurate information about the classified elements, it allows companies to drive the customer equity, to better retain top and potentially good customers by the mean of individualized privileges, to launch and verify marketing campaigns and to avoid the churn of good customers. With the help of linguistic variables and terms, the fuzzy classification approach also enables an intuitive querying process based on the terminology of the marketing department. Furthermore, the decomposition mechanism allows marketers to define hierarchies of fuzzy classifications in order to better model the reality and to apply appropriate marketing actions. All those tools help companies to maximize the value of their customers and, this way, their profits. The proposed fuzzy classifications hierarchy of online customers will be implemented in the open source webshop eSarine [16] in order to provide small and medium sized enterprises an integrated e-business tool. Further researches on the use of a fuzzy classification in the marketing mix theory and other application fields are still in progress.

6. References [1] P. Arabie, J.D. Carroll, W.S. DeSarbo and J. Wind, “Overlapping clustering: A new method for product positioning”, in Journal of Marketing Research, 18, 1981, pp. 310-317. [2] R.C. Blattberg, G. Getz and J.S. Thomas, “Customer equity - Building and managing relationships as valuable assets”, Harvard Business School Press, Boston, 2001. [3] P. Bosc and O. Pivert, “SQLf query functionality on top of a regular relational database management system”, in

IEEE International Conference on e-Business Engineering (ICEBE'06) 0-7695-2645-4/06 $20.00 © 2006

Knowledge Management in Fuzzy Databases, O. Pons, M. A. Vila, J. Kacprzyk, Physica Publisher, Heidelberg, 2000, pp. 171-190. [4] G. Chen, “Fuzzy logic in data modeling – Semantics, constraints and database design”, Kluwer Academic Publishers, London, 1998. [5] T. Harrison, “Financial Services Marketing”, Pearson Education, Essex, 2000. [6] H. Hruschka, “Market definition and segmentation using fuzzy clustering methods”, in International Journal of Research in Marketing, 3 (2), 1986, pp. 117-34. [7] J. Kacprzyk and S. Zadrozny, “Data mining via fuzzy querying over the internet”, in Knowledge Management in Fuzzy Databases, O. Pons, M. A. Vila, J. Kacprzyk, Physica Publisher, Heidelberg, 2000, pp. 211-233. [8] J. Lee, M. Podlaseck, E. Schonberg and R. Hoch, “Visualization and Analysis of Clickstream Data of Online Stores for Understanding Web Merchandising”, in Data Mining and Knowledge Discovery, 5, 2001, pp. 59-84. [9] A. Meier, N. Werro, M. Albrecht and M. Sarakinos, “Using a fuzzy classification query language for customer relationship management”, in Proc. of the 31st VLDB Conference, Trondheim, Norway, 2005. [10] P.T. Nguyen, G. Cliquet, A. Borges and F. Leray, “Opposition between size of the market and degree of homogeneity of the segments: a fuzzy clustering approach” (in French), in Décisions Marketing, 32, 2003, pp. 55-69. [11] H. Rheingold, “The Virtual Community - Homesteading on the Electronic Frontier”, Addison Wesley, New York, 1993. [12] G. Schindler, “Fuzzy data analysis through contextbased database queries” (in German), Deutscher Universitäts-Verlag, Wiesbaden, 1998. [13] M. Setnes and U. Kaymak, “Fuzzy modeling of client preference in data-rich marketing environments", Research Paper ERS; ERS-2000-49-LIS, Erasmus Research Institute of Management (ERIM), RSM Erasmus University, 2000. [14] S. Shenoi, “Fuzzy sets, information clouding and database security”, in Fuzziness in Database Management Systems, P. Bosc, and J. Kacprzyk, Physica Publisher, Heidelberg, 1995, pp. 207-228. [15] Y. Takahashi, “A fuzzy query language for relational databases”, in Fuzziness in Database Management Systems, P. Bosc, and J. Kacprzyk, Physica Publisher, Heidelberg, 1995, pp. 365-384. [16] N. Werro, H. Stormer, D. Frauchiger and A. Meier, “eSarine - A struts-based webshop for small and mediumsized enterprises”, in Proc. of the EMISA Conference Information Systems in E-Business and E-Government, Luxembourg, 2004. [17] N. Werro, H. Stormer and A. Meier, “Personalized discount - A fuzzy logic approach”, in Proc. of the 5th IFIP International Conference on eBusiness, eCommerce and eGovernment, Poznan, Poland, 2005. [18] H.-J. Zimmerman and P. Zysno, “Latent connectives in human decision making”, in FSS, 4, 1980, pp. 37-51. [19] H.-J. Zimmermann, “Fuzzy set theory and its applications”, 2nd ed., Kluwer Academic Publishers, London, 1992.