Περίληψη | In this paper we review the applicability of representative inductive machine learning
approaches in multicriteria decision making. We limit our review to four systems. We use
SICLA and KBG as representative conceptual clustering systems and ID3 and CN2 as
representative learning from examples systems. We demonstrate our results by way of two
real world decision making exemplars. The first exemplar concerns the evaluation of retail
outlets [15]. The second exemplar concerns venture capital assessment [16]. We discuss the
conditions under which inductive learning methodologies can be effectively implemented to
support decision making.
Inductive machine learning was pioneered by Michalski [9]. It aims at the derivation
of knowledge from a set of observations, or facts. In cases where facts are known to belong
to a certain class we speak of concept acquisition or learning from examples. In such
an instance we target our inquiry towards the derivation of concept identification rules.
Rules may be either discriminant or characteristic. When concept classes underlying fact
membership are not known we speak of learning from observations, or conceptual clustering.
Accordingly, we look forward toward the partitioning of facts into a meaningful and disjoint
set of clusters. A cluster represents a “coming together in space and time so that the density
of whatever is clustered contrasts with the density around” [6, p.33]. Generalization and
specialization are essential processes when making inductive inferences. The basic premise
characterizing any inductive inference is falsity preservation. The derivation of a hypothesis
H from facts E is falsity preserving in the sense that “if some facts falsify E, then they
must also falsify H” [9, p.89].
Although inductive machine learning is a rather new field there are several and
successful ‘fielded’ applications [7, 8]. Carter and Catlett [2] propose a methodology for
credit card assessment using inductive learning techniques. Also, Shaw and Gentry [14]
present an approach for company risk assessment that is based on inductive learning. Both
applications are exploratory; they, however, stress the potential of inductive learning in
decision making support. We maintain that learning is a trait of decision making: “quite
often the decision maker is interested in finding out what his weights are or what they
should be under different decision circumstances. In this sense, the weights of importance
could be considered as desirable outputs rather than independent inputs of an analysis.
Weights must be revealed or learned through a careful interactive process”, [17, p.22] -
emphasis is ours.
In this paper we discuss the methodological issues underlying the application of
inductive learning techniques in business decision making. We limit our endeavor to four
representative and well-known inductive learning systems, ID3 [12, 11, 13], CN2 [5, 4], KBG
[1] and SICLA [3]. These systems are part of the Machine Learning Toolbox [7, 18]. We
explore inductive system suitability by way of three decision making exemplars. We draw
our exemplars from retail outlet evaluation and venture capital assessment. We target our
inquiry toward the evaluation of pros and cons, concerning the application of the selected
inductive learning systems, in real world business decision making. Specifically, our research
focuses around the following lines:
1. Grouping of alternatives into disjoint cluster groups. We use a Lexicographic Evaluation
Functional, LEF, criterion to optimize clustering [10].
2. Identification of the most significant criteria for either alternative discrimination or
alternative characterization. Suppose that we have two alternative courses of action,
a1 and a2. We are interested in differences between a1 and a2, or in what a1 and a2
are all about. Furthermore, we present a methodology for inducing criteria weights.
3. Identification of relevant and accurate discrimination and recognition rules. We associate
this line with the previous one.
4. Identification of the most representative alternative for each decision class. We steer
our venture in the direction of deriving a conceptual indexing scheme for alternative
courses of action.
5. Identification of bias and error resulting from contextual factors. We define context
to represent the decision making environment.
Furthermore, we explore the implications of our research in decision making. We
place emphasis upon the expert critiquing and case based reasoning paradigms. | en |