D MCSHERRY The template used by Top Case to explain why it is recommend ing a particular case depends on whether that case exactly matches the user's query and whether preferred values have been elicited for all case attributes. In the example template below, Case X is the rec- nded case. attributes-I are attributes in which it differs from the user's query, and attributes-2 are attributes for which preferred values have not been elicited because they cannot affect the recommendation Case X differs from your query only in attributes-I and is the best case no matter what attributes-2 you prefer As in First Case(McSherry, 2003b), the aim of highlighting any compromises associated with the recommended case is to help the user decide whether or not to accept the system's recommendation. Also informing the user that her preferences with respect to any remaining attributes cannot affect the solution may help to increase her confi- dence in the recommendation. How effective the explanations provided by Top Case are in achieving these objectives is an important issue te be addressed by further research 4.2. Example case TheTravelcaselibrary(www.ai-cbr.orgthatweusetoillustrateour approach to explanation in Top Case is a standard benchmark con- taining the descriptions of over 1,000 holidays. Attributes in the case library and importance weights assigned to them in Top Case are price(8), month(7), location(6), persons (5), duration(4), type(3), accommodation(2), and transport(1). Attributes price, persons, and duration are treated in Top Case as nearer-is-better(NIB) attributes. The remaining attributes are treated as nominal attributes. As usual in practice, we define the similarity of two given values x and y of a NIB attribute a to be lx sima(x, y)=l max(a)-min(a) where, for example, max(a) is the maximum value of a in the case library. Our similarity measure for each of the nominal attributes sim ply assigns a score of l to a given case if its value for the attribute is the same as the preferred value and 0 if it is not the same Figure 3 shows an example dialogue based on the Travel case library. Minor editing in the interest of clarity includes the format of the user's initial query. The current similarity of each retrieved case,190 D. MCSHERRY The template used by Top Case to explain why it is recommend￾ing a particular case depends on whether that case exactly matches the user’s query and whether preferred values have been elicited for all case attributes. In the example template below, Case X is the rec￾ommended case, attributes-1 are attributes in which it differs from the user’s query, and attributes-2 are attributes for which preferred values have not been elicited because they cannot affect the recommendation. Case X differs from your query only in attributes-1 and is the best case no matter what attributes-2 you prefer As in First Case (McSherry, 2003b), the aim of highlighting any compromises associated with the recommended case is to help the user decide whether or not to accept the system’s recommendation. Also informing the user that her preferences with respect to any remaining attributes cannot affect the solution may help to increase her confi- dence in the recommendation. How effective the explanations provided by Top Case are in achieving these objectives is an important issue to be addressed by further research. 4.2. Example case library The Travel case library (www.ai-cbr.org) that we use to illustrate our approach to explanation in Top Case is a standard benchmark con￾taining the descriptions of over 1,000 holidays. Attributes in the case library and importance weights assigned to them in Top Case are price (8), month (7), location (6), persons (5), duration (4), type (3), accommodation (2), and transport (1). Attributes price, persons, and duration are treated in Top Case as nearer-is-better (NIB) attributes. The remaining attributes are treated as nominal attributes. As usual in practice, we define the similarity of two given values x and y of a NIB attribute a to be: sima(x, y)=1− |x −y| max(a)− min(a), where, for example, max(a) is the maximum value of a in the case library. Our similarity measure for each of the nominal attributes sim￾ply assigns a score of 1 to a given case if its value for the attribute is the same as the preferred value and 0 if it is not the same. Figure 3 shows an example dialogue based on the Travel case library. Minor editing in the interest of clarity includes the format of the user’s initial query. The current similarity of each retrieved case