Professional Documents
Culture Documents
davis1992
davis1992
A. J. GAMMERMAN
Heriot-Watt University, Edinburgh
1. Introduction
Inductive systems are used to generalize concepts from a set of examples. The
examples constitute previous experience, and induction is one of the very basic ways
humans use for learning. As a result, the learning usually creates some knowledge.
The syntactic structure and semantic organization of knowledge is what distinguishes
it from information. Many knowledge representation formalisms have been used,
and one of them we are going to deal with is a decision tree. Induction that is
used to extract concepts from data should somehow be capable of distinguishing
among them. For example, in credit-card assessment, two possible concepts are
'creditworthy' (or 'good' applicants who paid back their debts) and 'noncredit-
worthy' (or 'bad' applicants). A large enough set of card-application examples can
be used in order to find a 'typical' set of attributes with the corresponding concept
or class (good or bad).
Our research aim is to apply different machine-learning and inference techniques
to the credit-card assessment problem. In particular, there are two major objectives:
(1) to describe a few different algorithms and systems for the classification of
applicants; (2) to compare the effectiveness of different models by comparing the
accuracy with the results of classification by bank managers. We estimate the
'accuracy' here by classifying a test set of examples, and then calculating the per-
centage of 'correct' classifications (the ones that correspond to the bank managers'
decisions).
We consider here two approaches. The first uses Bayesian inference without
assuming independence—the G&T model (Gammerman & Thatcher 1990)—and the
second is based on neural-network algorithms, or more precisely, on a connectionist
expert system technique (Gallant 1988). We shall describe briefly both models and
their design and implementation, and then compare their performance using a
banking example.
Training
procedure
Coding
Selection
Calculations
Final Fitting
decisions
(FD)
Evaluation Decision
making
Computer
diagnoses
Training set
of data
Pocket
algorithm
Knowledge
User Inference
interface mechanism
The job of the matching part is to fit the observed attributes of newly arrived
applicants into the combinations of relevant attributes, and also to find their
probabilities.
The G&T system allows a decision to be made for each new applicant by selecting
the class with the highest probability selected (we shall call it the computer
classification). The system also allows the evaluation of the accuracy of classification,
by comparing the computer classification with the applicant's final classification
recorded by bank managers.
Pocket algorithm
(1) set P to the 0 vector, where P is a perceptron weight vector
(2) let P be the current perceptron weights.
randomly pick a training example Ek (with corresponding classification C*)
(3) IF P correctly classifies Ek
THEN
IF (the current run of correct classifications with P is longer
than the run of correct classifications for the weight vector W^ in your
pocket)
THEN
replace the pocket weight Wm by P
4. Results
We now present results of the performance of the models. In each case the training
set was 609 applicant records, and 255 were used as the testing set.
TABLE 1
Selected combinations of attributes for class 1 ('good1)
Combinations D N P L U
255 applicants. That is, for each of the 255 clients, we calculate the probabilities for
both classes, and take as a 'computer classification' the classification with the highest
probability. As a result, 71.4% of the applicants were correctly classified.
TABLE 2
Accuracy of classification: G&T
Computer classification
Final classification
(Threshold 10.5) GOOD BAD Total
GOOD 18 62 80
BAD 11 164 175
Total 29 226 255
Table 3 shows the accuracy achieved by the 'simple Bayes' method on the same set
of data.
MACHINE-LEARNING ALGORITHMS FOR CREDIT-CARD APPLICATIONS 49
TABLE 3
Accuracy of classification: simple Bayes
Computer classification
GOOD 38 42 80
BAD 37 138 175
Total 75 180 255
TABLE 4
Accuracy of classification: pocket algorithm
Computer classification
GOOD 39 41 80
BAD 50 125 175
Total 89 166 255
5. Conclusions
Two computational models have been described and applied to a set of banking data.
The present results are limited by the relatively small number of training examples
and test set. Given this fact, both techniques performed well. Still there were various
differences among the techniques worth noting. First, G&T has been generally more
'selective' than the connectionist model. It selected only relevant attributes and
constructed a set of combinations of attributes (a decision tree) for each of the classes.
It would be even more obvious had we processed the data with more than two classes.
The accuracy achieved by G&T was on the level of (~71%). As can be seen the
'proper Bayes' model has achieved a slightly higher level of accuracy than the 'simple
Bayes' model (~69%). The connectionist model was much slower in the learning
part than G&T model, and achieved about 64% of accuracy.
50 R. H. DAVIS, D. B. EDELMAN, AND A. J. GAMMERMAN
Acknowledgements
We wish to thank Mr. Hakon Styri and Mr. Zhiyuan Luo for useful discussions and
some results with both systems. Data were used with permission of Bank of Scotland.
MACHINE-LEARNING ALGORITHMS FOR CREDIT-CARD APPLICATIONS 51
REFERENCES
GALLANT, S. 1988 Connectionist expert systems. Commun. ACM 31, 152-69.
GAMMERMAN A., & THATCHER, A. R. 1990 Bayesian inference in an expert system without
assuming independence. Advances in Artificial Intelligence (Ed. M. C. Columbic) pp.
182-95. Springer-Verlag.