Efficiently Identifying Exploratory Rules’ Significance - Data Mining: Theory, Methodology, Techniques, and Applications - page 73

Table 2. Improved OPUS IR Filter

Algorithm: OPUS IR Filter(Current, Available, parent rule, M )

1 SoFar := ∅ ;
2 FOR EACH P in Available
2.1 New := Current ∪ P
2.2 IF New satisfies all the prunable constraints in M except the nontrivial
constraint THEN
2.2.1 Derive the statistics of coverset ( Current ) −coverset ( New ) , according to lemma
1.
2.2.2 IF the mean of New

→ target is not significantly improved comparing to

coverset ( New ) THEN
go to step 2.2.4;
2.2.3 ELSE use the circular intersection to comparing the mean of New

coverset ( Current ) −

→ target with

the mean of its direct parents other than parent rule
2.2.3.1 IF the mean New

→ target is significantly improved comparing to all its

direct parents THEN
record New

target to rule list ;
OPUS IR Filter(New, SoFar, New → target );
SoFar := SoFar ∪ P;
2.2.3.2 END IF;
2.2.4 END IF;
2.3 END IF;
3 END FOR

→

The new algorithm for impact rule discovery with filters is shown in table 2. In
this table, the is the corresponding rule for the node whose children

we are currently exploring. The antecedent of

parent rule

parent rule is current .

6 Experimental Evaluations

In order to explain how the techniques introduced in this paper can practically
improve the eciency of rule discovery, we did our experiments by applying the
new algorithm to 10 databases chosen from the UCI Machine Learning reposi-
tory [6] and the UCI KDD archives [3]. The databases are described in table 3.
We applied 3-bin equal-frequency discretization to map all the quantitative at-
tributes, except the target attribute, into qualitative ones. The significance level

Table 3. Basic information of the databases

database records attributes conditions Target

Abalone 4117 9 24 Shuckedweight

Heart 270 13 40 Max heart rate

Housing 506 14 49 MEDV

German credit 1000 20 77 Credit amount

Ipums.la.97 70187 61 1693 Total income

Ipums.la.98 74954 61 1610 Total income

Ipums.la.99 88443 61 1889 Total income

Ticdata2000 5822 86 771 Ave. income

Census income 199523 42 522 Wage per hour

Covtype 581012 55 131 Elevation