Re: RuleGrowth/ERMiner
Date: March 13, 2017 04:16PM
Hello,
For these algorithms, we find the rules directly
The reason is that the rules found by ERMiner and RUleGrowth are of the form X --> Y where the items in X and Y are unoredered, but X must appear before Y. For example, a rule {a,b,c} --> {e} means that a,b,c appeared in any order but were followed by {e}.
Some other algorithms like RuleGen find the sequential patterns first and produce some rules X-->Y such that X and Y are sequential patterns. For example, they could find patterns such as a,b,c--> d,e,f which means that if a,b,c appears one after the other, then they will be followed by d,e,f in exactly that order. However, the problem with these rules are that they are too specific and not noise tolerant. For example, the rule a,b,c -> d,e,f would be seen as different from b,a,c -> d,e,f or a,c,b -> d,e,f or a,b,c-> d,f,e but actually they are probably the same rules, because in real life there is noise. In my experiments for web log prediction, I found that these rules are too specific and do not work well for prediction (see my TKDE paper about Rulegrowth or ADMA 2012 paper about sequence prediction for details about these experiments). That is why, in my work such as ERMiner and RUleGrowth, I use the rules X--> Y where the antecedent and consequent are unoredered.
So because the rules found by RulEgrowth/ERMiner have an unordered antecedent and consequent, these algorithms do not find sequential patterns first, and instead find the rules directly. Actually, the ERMiner algorithm is inspired by the Eclat algortihm for itemset mining. My goal was to apply a similar idea for rules. The RuleGrowth algorithms on the other hand is a pattern growth algorithm. The idea was inspired by PrefixSpan for sequential pattern mining.
Best regards,
Edited 3 time(s). Last edit at 03/13/2017 04:22PM by webmasterphilfv.