티스토리 뷰

728x90

1. Rule Generation from frequent itemset

 1. Frenquent itemset이 {A,B,C,D}일 떄 Candidate rules은 다음과 같다.

 Frenquent itemset이 K개 이면 총 생성할 수 있는 Candidate rules은  2^k-2개이다.

 공집합과 전체 집합은 뺴기 때문에 -2이다.

 

 2. 신뢰도(confidence)는 anti-monotone 성질을 가지지 않는다.

 Apriori 특성 사용이 어려움
    - c(ABC->D) can be larger of smaller than c(AB->D)

 동일한 항목집합에서 생성된 규칙에 대해서는 anti-monotone 성질이 성립

    - c(ABC->D) >= c(AB->CD) >= c(A->BCD)


 3. Candidate rule is generated by merging two rules 

 두 개의 Candidate Rule이 새로운 Rule을 만들어 내지만 생성된 Rule의 신뢰도는 기존 2개의 신뢰도 보다 작다.

 

 

2. Maximal Frequent Itemset

 1. Itemset은 maximal frequent를 가진다.

 

 2. Itemset은 maximal frequent 찾는 방법

▪ 먼저 Infrequent와 frequent itemset 사이의 itemset을 찾는다.

- d, bc, ad, adc

▪ 위에서 찾은 itemset의 immediate superset을 찾는다.

- d의 superset으로 ad, bd, cd가 있는데 ad는 frequent라서 d는 immediate superset이 아니다.

- bc는 adb와 bcd를 superset으로 갖는데, adc가 frequent

- ad와 abc의 superset은 모두 infrequent. 따라서 ad와 abc는 maximal frequnt

 

 

3. Closed Itemset

 1. Closed Itemset : support값이 superset의 support 값보다 크면 close

 2. Closed frequent itemset 찾는 방법

먼저 모든 frequent itemset을 찾는다. 위 예시에서는 최소 support값이 2이다. 파랑 테두리를 가지고 있다.

▪ superset의 support 값이 작은 close itemset을 찾는다. 두꺼운 테두리를 가지고 있다.

▪ 파랑 테두리(frequent itemset)중 close인 것들은 close frequent itemset이라고 한다.

 

 

4. Maximal vs Closed Itemsets

 1. Example

 

 2. Result

▪ 회색 : Closed frequent itemset

▪ 파랑 띠 : Maximal frequent itemset

▪ Frequent > Closed > Maximal 관계를 가진다.

 

 

5. 연관 규칙 평가(Pattern Evaluation)

 1. 연관 규칙 생성 알고리즘은 너무 많은 연관 규칙을 생성한다. 

▪ 생성된 모든 규칙이 유용하지는 않음

▪ {A, B, C} -> {D}와 {A, B} -> {D}가 동일한 지지도/신뢰도를 갖는다면, 이들 두 규칙은 중복임

▪ Interestingness measures(유용성 척도)는 유도된 규칙을 제거하거나 순위를 매기는데(prune or rank) 사용됨

지지도와 신뢰도(support & confidence)도 유용성 척도에 속함

 

2. 분할표(contingency table)를 사용하여 다양한 유용성 척도를 계산할 수 있다

 

3. 신뢰도의 단점. 

▪ 위 신뢰도를 보면 차를 마시는 사람은 커피를 마는 경향이 있다고 추론할 수 있다.

▪ 그러나 원래 전체 100명 중 커피를 마시는 사람이 80퍼센트이다. 즉 차를 마신다는 정보에서 커피를 마시는 사람에 대한 정보를 아는 것은 큰 의미가 없다.

 

 

6. Statistical Independence

 1. Statistical independence

▪ P(SnB) = P(S) ´ P(B) => Statistical independence
▪ P(SnB) > P(S) ´ P(B) => Positively correlated
▪ P(SnB) < P(S) ´ P(B) => Negatively correlated

 

 2. Exampel

  Population of 1000 students

    - 600 students know how to swim (S)

    - 700 students know how to bike (B)

    - 420 students know how to swim and bike (S,B)

 

▪ P(SnB) = 420/1000 = 0.42

▪ P(S) ´ P(B) = 0.6 ´ 0.7 = 0.42

P(SnB) = P(S) ´ P(B) => Statistical independence

 

 3. Measures that take into account statistical dependence

 

 4. 연관 규칙 평가 : Lift

▪ If Lift > 1, then X and Y appear more often together than expected
▪ If Lift < 1 then, X and Y appear less often together than expected
▪ If Lift = 1, then X and Y are independent.

 

 5. Example

▪ Lift less than 1, so Negatively correlated.

댓글
공지사항
최근에 올라온 글
최근에 달린 댓글
Total
Today
Yesterday
링크
«   2024/05   »
1 2 3 4
5 6 7 8 9 10 11
12 13 14 15 16 17 18
19 20 21 22 23 24 25
26 27 28 29 30 31
글 보관함