The MBANALYSIS Procedure

The MBANALYSIS procedure performs association rule mining on transaction data in conjunction with an item that is specified in one or more hierarchies (also called groups or categories). Association rule mining is popularly known as market basket analysis. Market basket analysis is useful in retail marketing scenarios that involve tens of thousands of distinct items that are grouped in a hierarchy. PROC MBANALYSIS uses the transaction and hierarchical data and generates rules at multiple levels. These rules are called generalized association rules; they solve some of the problems, which are described below, that arise in simple market basket analysis.

Simple market basket analysis that is performed on transaction or point-of-sale data might miss potentially useful and significant associations. For example, consider a supermarket that sells different types of breads and a wide selection of wines, and assume that a large number of customers buy some type of bread with some type of wine. Simple market basket analysis might fail to spot a rule that links bread and wine at the transaction level. It computes the support for specific types of breads to specific types of wines, but it is possible that none of these support levels is large enough to generate a rule. To overcome this problem, the MBANALYSIS procedure combines hierarchies with the transaction data and computes the support for any type of bread to any type of wine.

In addition, simple association mining tends to create a large number of obvious and uninteresting rules along with the useful rules. A solution to the preceding problem that reduces the support would only exacerbate that problem. In practice, this tendency is one of the major drawbacks of association rule mining. If the support is set high, fewer rules are generated but more of them are obvious and thus useless. If the support is set low, too many rules are generated and the domain experts must evaluate the rules to determine which rules are useful. One solution to this drawback is generalized association rule mining.

The MBANALYSIS procedure enables you to generate an interestingness measure, which is called support lift. This measure is based on the deviation of a rule’s support from its estimated support, which is computed based on the support of the ancestors of the rule’s items. Objectively, greater deviation is more likely to be interesting. If no hierarchy of ancestors is specified, then PROC MBANALYSIS generates simple association rules.

For example, the following rule for supermarket means that if customers buy butter and bread, they also buy milk:

Support (coverage) is an indication of how frequently the items appear in the database. Support of a rule for A and B is expressed as

where X is the number of transactions that contain both A and B, and Y is the total number of transactions.

Confidence (accuracy) indicates the number of times the conditional statement of rule have been found to be true. Confidence of a rule for A and B is expressed as

where X is the number of transactions that contain both A and B, and Z is the number of transactions that contain A.