A Pragmatic Approach to Summarize Association Rules in Business Analytics Projects - Technologies and Applications of Artificial Intelligence - page 328

Another interesting aspect is that, when the antecedent support threshold is
reduced, the maximum support value of each rule summary remains unchanged. For
example, the rule summary for consequent 'I' remains at 34.21%, despite the fact that
different antecedent support thresholds were used. This validates Property 3.
When examining a rule summary, the range of its interestingness metric indicates
the appropriateness of the minimum interestingness metric threshold. If a rule
summary has a narrow interestingness metric range, it suggests that the threshold may
be too high and should be adjusted down. If the interestingness metric range of a rule
summary is too wide , it suggests that the threshold may be too low and should be
adjusted up. For example, when the minimum antecedent support threshold is set at
5% in Table 3, the support ranges become quite wide and one might consider
increasing the support threshold.
Notice that Table 3 also illustrates the effect of adjusting the interestingness
thresholds. The effect of increasing the interestingness threshold will reduce
consequent frequency, but narrow the range of interestingness metrics. On the other
hand, decreasing the interestingness threshold will increase consequent frequency, but
widen the range of interestingness metrics.
For the sake of completeness, we have also obtained summaries of rules generated
from the Online Purchase dataset, with the antecedent support threshold fixed at 20%
and confidence threshold reduced gradually from 60% to 55%, then 50% and 45%.
We notice that the rule summaries tend to exhibit the same properties as those
observed in Table 3. Due to the page limit for wiring this paper, we do not show the
rule summaries here, but shall provide the details in a future publication.

4.4 A Case of Using CARS in a Business Analytics Project

Here, we illustrate how the proposed method can be used in conjunction with a
heuristic procedure to help find interesting association rules in the Sales Transaction
dataset. The procedure is as follows:

Step 1: Rule Generation). Given a dataset, define each attribute as both an input and
output (i.e., an attribute can be antecedent in Rule i , and it can be consequent in Rule
j , such that i ≠ j ). Use reasonably low interestingness thresholds to generate as many
rules as possible for summarization.

Step 2: Rule Summary Evaluation). Summarize the rules using CARS. Sort the rule
summaries based on the maximum confidence. Select rule summaries with
consequents that are of pragmatic importance in the domain of application. Evaluate
the impact of rule summaries using rule confidence and antecedent support ranges.
Validate each interesting rule summary by examining the credibility of frequent
antecedents in light of knowledge about the business domain.

Step 3: Result Refinement). Let the selected important consequents remain as both
input and output, and set the rest of the attributes as inputs. If required, adjust (and