[Next] [Previous] [Top] [Contents]

Artificially Intelligent Specification and Analysis of Context-Dependent Attribute Preferences

1 Introduction


Applications of utility theory represent each consumer as a preference function which is independent of the reason for which any purchase is to be made. Moss and Edmonds = [23], however, have reported an intelligent marketing integrated system (IMIS) incorporating a view from the marketing profession that, in some markets at least, preferences are usefully related to the context of consumption. They demonstrated a specification of context-dependent attribute preference (CDAP) functions that tracked the relationship between prices and market shares of branded fast-moving consumer goods (FMCGs) better than corresponding OLS models when estimation was based on the short sample periods. That is, the CDAP model incorporated the qualitative judgements of domain experts but used less statistical data to produce more accurate market share simulations than did the OLS models. It also provided substantial qualitative information that was useful in determining marketing strategies. Such qualitative information cannot, of course, be provided by statistical models of any kind.

Moss and Edmonds did not argue that CDAP models are in some sense better than statistical models. They have different purposes and, in any case, integrating statistical analysis with CDAP analysis leads to a better grounded and more fully rounded picture of competitive relations in some markets than does either approach by itself.

CDAP models incorporate of the qualitative representations of demand factors in a way which enables us to test their consistency with the available statistical record. We follow the formal modelling literature in which "qualitative" relations are functional numerical relations which are invariant under a set of topological transformations. See, for example, [15, 9, 5, 26, 11]. In particular, the qualitative relations used here are mappings from verbal expressions into real number intervals.

Clearly, this approach is very different from the concerns of the literature on consumer behaviour typified by [1] or [8]. It is not concerned directly with how individuals formulate their preferences or their purchasing and consumption decisions. Instead, preferences are represented as a distribution of the attributes required by consumers who are purchasing branded goods for specific purposes such as alcoholic beverages for parties or for consumption individually within the home.

For each purchasing context, the distribution of attribute ideals reflects differences in individual tastes and also the tolerance of consumers to differences between the ideal intensity of an attribute and the actual value which, in their perception, is embodied in actual brands. Moreover, particular attribute values will be more or less critical to consumers purchasing for different purposes. Consequently, for each purchasing context identified by the marketing professionals, we identify the attributes which are identified with the various brands as well as the ordering of the intensity of such attributes. Thus, Chivas Regal whisky has more of the attribute "specialness" than does Harry's Old Horsegut Bourbon but Harry's Bourbon has more of the attribute "special ingredients" than does Chivas Regal. When celebrating a successful event, consumers might give a high priority to buying a drink which has a high degree of "specialness" but when contemplating the rich variety of life they might want something more distinctive. In the one case specialness will be important and in the other case it might not. The contemplatives, taken as a group, might have widely differing ideal levels of distinctiveness in their drink or they might be tolerant of substantial deviations from their ideal of the perceived uniqueness of a drink.

A detailed justification for this approach is given in section 2. The preference distribution functions are motivated and described in detail in section 3, the algorithms for extracting credible CDAP models from EPOS data is described in section 4 and then, in section 5, applied to the case of a market for alcoholic beverages.

The results reported in section 5 demonstrate that the formal incorporation of the verbal, qualitative judgements of domain experts into a model of a specific market supports the extraction of detailed qualitative information which is demonstrably consistent with reliable, numerical, time-series data. This result is consistent with earlier findings such as the demonstrations by [4] and by [22] that adding rulebased descriptions of expert knowledge into econometric forecasting models can substantially improve their accuracy. Indeed, there is a substantial and long-standing literature on the incorporation of qualitative, expert knowledge in formal models. See, for example, [21, 18]. There is also the well established literature on the reliability measures for qualitative judgements made independently by several domain experts. [24] review and extend this literature.


Artificially Intelligent Specification and Analysis of Context-Dependent Attribute Preferences - 03 NOV 97
[Next] [Previous] [Top] [Contents]

Generated with CERN WebMaker