View Full Version : Lift in Association Mining

06-02-2009, 11:54 AM
I have been using WEKA for some time in classification (C4.5/J48). I am now trying association mining (apriori). Rather than using confidence as a metric I am trying lift.

Here is my question:

At the end of each rule are the following (an example)

conf :0.91) < lift :15.56)> lev :0.05) [78] conv :9.62)

conf: is the confidence - 0.91
lift: is the lift - 15.56

What are lev: and conv:?
What is the value in [ ]?
How are they calculated?

06-05-2009, 12:48 AM

Leverage and conviction are two further metrics for the goodness of association rules. Leverage is defined as:

p(premise & consequence) - (p(premise) * p(consequence))

I think it is fairly obvious what this is measuring. The value in brackets equates to the above * number of instances in the data.

Conviction is a bit more obscure and I'd have to chase down the paper(s) that discuss it to refresh myself as to what exactly it is measuring :-) It is defined as:

p(premise) * p(!consequence) / p(premise & !consequence)


06-11-2009, 05:43 AM
i juste begin to study the datamining.

so lift and leverage are big, it's better and bond between premise & consequence is strong,
for conviction alos it's big, it's better?