7. Decision Trees and Decision Rules

Transcription

1 立 林 National Yunlin University of Science and Technology 良 料 理 立 林

2 良 良 良 論

3 (self-organizing map, or SOM) [Kohonen 1974, 1990, 1996] 類神 路 度 料 留 料 拓 度 度 料

4

5 [Kohonen 1996] 識 [Visa 1990] 理流 [Kasslin 1992, Simula 1992] 識 [Mantysalo 1994] 識 [Vapola 1994] 年 料 理 [Lagus 1996] 療 [Chen 2000, Kramer 2000] 理 [Kasabov 2000, Deboeck 2000]

6 Area Dept. Amount Count Area Dept. Amount Count Taipei E.E North Taiwan Engineering College High 23 Taoyuan C.E Central Taiwan Management College Low 10 Hsinchu E.E South Taiwan Engineering College Median 8 Miaoli M.E Taichung MIS Kaohsiung E.E 料 料 Attribute-Oriented Induction (AOI): 量 料 精 料

7 理數 料 料都 料 留 度 數

8 良 理 類 料 料 料

9 練 練 料 2.1 (BMU) 2.2 神 3. 2 立 m i =[7, 5, 8] x1=[8, 5, 9] x2=[7, 4, 2] x1=[8, 5, 9] 練 料

10 m = [ m1, m2,..., x = [ x1, x2,..., m n x n ] R n ] R n // SOM unit // input data d( x, m ) = = ( i x mi ( x BMU = arg min{ d( x, m i k = 1, n k i )} m ik ) 2 ) 1/ 2 m i =[7, 5, 8] x1=[8, 5, 9] x2=[7, 4, 2] 練 料 x1=[8, 5, 9]

11 神 m i ( t + 1) = m ( t) + α( t) h, ( t)[ x( t) m ( t)] 0 < α( t) < ( ) h b, i t i b i 1 率 數 數 離 BMU i h b, i = rb ri exp( 2 2σ ( t) 2 ) mi=[7, 5, 8] x1=[8, 5, 9]

12 理數 料 神 神 都 利 里 離 d ( x, m ) = = ( i x mi ( x k = 1, n k m ik ) 2 ) 1/ 2 x1=[8, 5, 9] m i =[7, 5, 8] d(x1, m i )=[ (8-7) 2 +(5-5) 2 +(9-8) 2 ] 1/2 Name Favorite Drink Amount Jane Coke 70 Mary Pepsi 70 Tom Water 70

13 理 類 (1) 類 料 度 (2) 度 (3) 不 易 (4) 欄 Name Favorite Drink Amount Jane Coke 70 Mary Pesi 70 Tom Water 70 Name Coke Pesi Water Amount Jane Mary Tom d(jane, Mary)=d(Jane, Tom)= 2

14 樂 念 料 樂樂 料類 類 金類 料 糖 糖 茶Name Favorite Drink Amount Jane Coke 70 Mary Pesi 70 Tom Water 70

15 類 念 m.i AB ANY A B E F G m.i=(n, d) N 葉 d 離 m.i= (A, 0.2) A 0.2

16 念 兩 離 AB ANY X Y A B E F G X Y = d + d - 2d X Y LCP(X, Y) Where X: (A, d X ) Y: (B, d Y ) LCP(X, Y): least common point of X and Y

17 樂 類 m i =( 樂 1.7) 量 0.8 x1=( 樂 3) m i =( 樂 2. 5) 料類 類 金類 x1=(2) m i =( 樂 0.9) x1=( 茶 4) m i =( 茶 1.1) 料 樂樂 料 糖 糖 茶

18 良 vs. m(t+1)=[(coke, 1.3), 40.8] m(t+1)=[1.2, 1.0, 1.7, 40.8] 量 量 m(t)=[(coke, 0.5), 40] m(t)=[0.4, 0.2, 0.9, 40] x=[(coke, 0), 70] x=[1, 0, 0, 70]

19 數 min.~max. 類 念

20 良 料 類 料 料 料 UCI Adult Data Set

21 料 -- 類 料 GNo Department Program Product DataCount 1 EE 2-year Coke 20 2 CE 4 year Pepsi 10 3 MIS Ph.D. Latte 20 4 MBA MS Mocca 10 5 VC Day Orange_Juice 20 6 SD Night Apple_Juice 10 ENG EE CE

22 Kohonon SOM

23 良 SOM

24 良 SOM-- 類 料

25 G. No. Dept. Program Amount Count 1 MBA 2-year (400, 20) 60 2 MBA 4-year (60, 3) 30 3 MBA 4-year (200, 10) 30 4 MIS 2-year (600, 30) 60 5 MIS 4-year (100, 5) 30 6 MIS 4-year (40, 2) 30 7 EE 2-year (300, 15) 60 8 EE 4-year (700, 35) 30 9 EE 4-year (600, 30) CE 2-year (500, 25) CE 4-year (20, 1) CE 4-year (100, 5) 30

26 料 練 ~6 理 7~12 4, 8, 9 2, 6, 11, 12

27 類

28 Area Dept. Amount Count Area Dept. Amount Count Taipei E.E North Taiwan Engineering College High 23 Taoyuan C.E Central Taiwan Management College Low 10 Hsinchu E.E South Taiwan Engineering College Median 8 Miaoli M.E Taichung MIS Kaohsiung E.E 料 Attribute-Oriented Induction (AOI): 量 料 精 料 料

29 Area Dept. Amount Count Area Dept. Amount Count Taipei E.E North Taiwan Engineering College High 23 Taoyuan C.E Central Taiwan Management College Low 10 Hsinchu E.E South Taiwan Engineering College Median 8 Miaoli M.E Taichung MIS Kaohsiung E.E 度 略 數

30 留 Age Prod Cnt Drink 18 Cola Soda 35 =3 =0.8 Tea Soda Cola 19 Juice 7 Age Prod Cnt 15 Tea 5 18 Cola 45 =3 13 Milk 4 16 Soda 35 Age Prod Cnt 12 Milk_tea 4 Youth-{16, 18} Drink-{Cola, Soda} 20 Youth Drink 100 料 EAOI AOI

31 留 S i =< ( vi, 1, ci,1),( vi,2, ci,2 ),...,( vi, j, ci, j ) > J CS, = c, c, β where 1 α i α α k= 1 i k j= 1 i j J M i = v 1 k α, v in S } { i, k i, k i S prod = <(Cola, 45), (Soda, 35), (Juice, 7), (Tea, 5), > CS prod,2 = (45+35)/( )=0.8 M prod = {Cola, Soda} Drink - M prod = Drink - {Cola, Soda} 留

32 數 Median High 40K 79K 80K 90K {40K~79K} Median Income {80K~} High Income T T T1 Median T2 High

33 不 念値 料 Age Prod Cnt 18 Cola Soda 35 EAOI 19 Juice 7 Age Prod Cnt 15 Tea 5 (18, 0) Cola 45 AOI 13 Milk 4 (16, 0) Soda 35 Age Prod Cnt 12 Milk_tea 4 (15.4, 2.9) Drink-{Cola, Soda} 20 Youth Drink 100

34 料 Modified SOM Extended AOI Patterns of clusters Rawdata Clustered data Traditional AOI Patterns of clusters Extracted patterns 良

35 料 Sex (ratio) Status (ratio) BirthPlace (ratio) Age Amount Class Cnt Male (100%) Ph.D. (80%) M.S. (10%) M.A. (8%) Senior (2%) Taiwan (40%) Japan (30%) SKorea (15%) NKorea (15%) N(30,2) N(90,2) College of engineering 60 Female (100%) Senior (65%) Junior (25%) Sophomore (5%) Freshman (5%) Vietnam (70%) Myanmar (20%) Nepal (5%) Thailand (5%) N(20,2) N(60,2) College of design 60 Male (60%) Female (40%) Freshman (60%) Sophomore (40%) Turkey (55%) Jordan (45%) N(19,1) N(80,2) College of management 50

36 良 練

37 EAOI -- Sex (ratio) Status (ratio) BirthPlace (ratio) Age Amount Class Cnt Male (100%) Ph.D. (80%) M.S. (10%) M.A. (8%) Senior (2%) Taiwan (40%) Japan (30%) SKorea (15%) NKorea (15%) N(30,2) N(90,2) College of engineering 60 Par. =3, =0.65, =10 Sex Status BirthPlace Age Amount Support C1 (Engineering College ) M Ph.D Taiwan (30.2;2.3) (90.8;2.0) 0.3 M Ph.D Japan (29.2;2.1) (90.4;1.5) M Ph.D EAsia-{Taiwan, Japan} (30.2;1.5) (90.4;1.6) M Graduate-{Ph.D} Taiwan (30.0;1.7) (90.7;1.9) 0.1 M Graduate-{Ph.D} EAsia-{Taiwan, Japan} (30.5;1.7) (90.8;1.5) M Graduate-{Ph.D} Japan (30;0) (92.0;0) M Undergraduate Japan (29;0) (90.0;0) 0.017

38 AOI -- Sex (ratio) Status (ratio) BirthPlace (ratio) Age Amount Class Cnt Male (100%) Ph.D. (80%) M.S. (10%) M.A. (8%) Senior (2%) Taiwan (40%) Japan (30%) SKorea (15%) NKorea (15%) N(30,2) N(90,2) College of engineering 60 Par. Traditional AOI =3, =10 (also, EAOI =3, =10, =1) Sex Status BirthPlace Age Amount Support C1 (Engineering College) M Graduate EAsia MidAged Huge 0.45 M Graduate EAsia Young Huge M Graduate EAsia Young Large M Graduate EAsia MidAged Large M Undergraduate EAsia Young Huge 0.017

39 料 年 料 (adult) Machine Learning Data Sets, UCI (University of California, Irvine ) 48,842 料 15 欄 9 類 6 數 Education, Sex, Age, Work_Hours, 欄 50K 76% >50K:~24%

40 Training Results of Adult Data Set

41 >50K 50K 率 Group Count No. of <=50K No. of >50K Ratio of <=50K Ratio of >50K % 2.26% % 7.55% % 10.68% % 52.31% % 32.7% % 11.71% % 7.47% Total % 23.91%

42

43

44

45 錄行 N(>50K)*10 + N( 50K)*0 - [N(>50K)+N( 50K)]*2 練

46 N(>50K)*10 + N( 50K)*1 - [N(>50K)+N( 50K)]*2 練

47 N(>50K)*10 + N( 50K)*2 - [N(>50K)+N( 50K)]*2 練 利

48 論 良 理 料 留 度 數 良 料