11-02-2014 07:11 AM
I am in process of forming a methodology for identification of infrequent aka suspicious association rules through SAS Eminer. The rules should be exclude frequent or strong rules as well as rare rules. Rare rules can occur in case of one off purchases say between a dealer dealing in manufacturing textiles and purchasing some relevant machinery.
Few tweaks that I have done to default settings are: 1) set the lower min support and 2) export close to 0.1 million rules as a part of the output.
1) use a filter for extracting low confidence through user discretion,
2) use a metric which uses confidence of rule divided by product of support of LHS and RHS,
3) sort the low confidence rules on ascending order of the above metric and select say top 30 to 40% of these rules.
Would like to ask the following questions:
1) is this methodology useful to find infrequent rules,
2) will an additional filter for low support prior to selecting low confidence be of any use or is redundant?
Note that we can give user univariate and bivariate statistics on support, confidence and both taken together. Also a cascading feature can be implemented to enable users know what unique values of key metrics can be possible, basis the filters mentioned above.
Thanks in advance for your suggestions and look foward to hear from you.
a week ago
I would suggest you filter the rules you get from a low support level by using a metric called "interest" defined as below.
According to probability theory, and are independent if . So the rule is not interesting if , which means that a rule is not interesting if its antecedent and consequent are approximately independent. Wu et al. introduces the function If , where interest is a predefined threshold, then itemset is referred to as a potentially interesting itemset.
Hope it helps,