A Pragmatic Approach to Summarize Association Rules in Business Analytics Projects - Technologies and Applications of Artificial Intelligence - page 330

profitable product compared to PMP-C1, then the last rule summary may outweigh
the first rule summary in practical context.
For simplicity of our subsequent analysis, we shall assume that all consequents are of
equal importance. Under this assumption, we use 5A80C to generate 324 rules of higher
confidence. These rules are summarized as four summaries shown in Table 6. These
summaries suggest that there are five product categories that are of common interest to
consumers, namely PMP-A1, PMP-B1, PMP-C1, PMP-C2, and PMP-D3. From the
product database, we find that there are 46 PMPs under these five commonly purchased
product categories. This information is important. For store layout, we know that there
are about 200 different models of PMPs sold at the store, these 46 commonly purchased
PMPs should be placed in the same area within the store. For product bundling, we
should focus on formulating bundles based on these five PMP categories. Bundles of
individual products may not yield the desired outcome because our rule mining results
suggest that products are only commonly purchased at product category level, and not at
individual products level.

Table 6. Summaries of 324 rules generated using 5A80C.

Support Range
(%)

Confidence Range
(%)

Consequent Top Antecedent

PMP-C1 PMP-B1, PMP-A1, PMP-C2 5.0 - 8.5 84.8 - 92.1

PMP-C2 PMP-B1, PMP-C1, PMP-A2, PMP-D3 5.0 - 7.9 80.0 - 85.8

PMP-B1 PMP-C1, PMP-A1, PMP-C2 5.0 - 7.8 80.0 - 85.0

PMP-D3 PMP-C1, PMP-A1, PMP-C2 5.1 - 5.5 80.1 - 80.2

5 Concluding Remarks

This paper has presented a consequent-based approach to summarize association
rules. The approach relaxes the notion of redundancy and combines rules with
different antecedents and interestingness values into one summary with common
consequent, and provides useful information such as antecedent ranking and range of
interestingness metrics to help users in finding important rules. Using the case study
on a real-world sales transaction dataset, we have also demonstrated how rule
summaries can be used to find rules that lead to actionable insights.
The idea of focusing more on consequent importance and less on objective
interestingness has two implications on the future research in association rule mining.
Firstly, our study indicates that consequent importance should be factored into the
design of practical association rule mining systems. We demonstrated that a
consequent-based approach distills a set of 4222 rules into a final set of rule
summaries that can help in product bundling or store layout improvement.
Secondly, this work suggests that while steady development of scientifically
motivated rule mining methods is important, this process should be complemented by
the development simple and practically useful rule mining systems to serve the
impending need of the user community. The proposed method is one such example; it