Presentation is loading. Please wait.

Presentation is loading. Please wait.

Cluster Analysis (군집 분석)

Similar presentations


Presentation on theme: "Cluster Analysis (군집 분석)"— Presentation transcript:

1 Cluster Analysis (군집 분석)

2 Cluster Analysis란 (1) Primary Objective
: 사전에 고려된 변수들에 기초를 두고, 다양한 특성을 지닌 대상들을 상대적으로 동질적인 집단으로 분류하는 것 (2) Basic Principle : High internal (Within-cluster) homogeneity and high external (between-cluster) heterogeneity 군집내의 소비자들은 서로 유사하고 한 군집의 소비자는 다른 군집 의 소비자와 서로 다르게 군집을 선택한다.

3 (3) Application ⅰ) Market Segmentation /Benefit Segmentation ⅱ) 구매행동 이해 : 동질구매집단 분류를 통한 특성 파악 ⅲ) 신제품 기회요인 도출 : brand와 Product를 clustering ⅳ) Test market 선정 ⅴ) Data 축소 (4) Cluster Vs. Factor Analysis cluster : 대상 분류 Factor : 변수(variable) 분류 (5) Cluster Vs. Discriminant Analysis - Object Classification Cluster : Cluster나 Group에 대한 사전 정보(분류기준)가 없는 경우 (독립 관계 분석) Discriminant : Cluster나 Group에 대한 사전 정보가 있는 경우 (종속 관계 분석)

4 Cluster Analysis 방법 Formulating the problem
Selecting a Distance Measure Selecting a Clustering Procedure Deciding on the Number of Clusters Interpreting and Profiling Clusters Assessing the Validity of Clustering

5 ▣ Basic Concept ● An Ideal Clustering Situation
● A Practical Clustering Situation Variable 1 Variable 1 Variable 2 Variable 2

6 (1) Formulating the Problem
: clustering의 기초가 되는 변수 선정 ⅰ) 군집되는 대상의 특성 분류 ⅱ) Cluster Analysis의 목적과 연결 (2) Similarity Measure : Distance Measure가 주로 이용됨 (주어진 질문에 대해 대답 간 차이의 제곱의 합으로 계산) ① Euclidean distance r dijE = ∑ (Xik - Xjk)2 (k=1,.....r) k=1 Xik : k차원에서 대상 i의 좌표 Xjk : k차원에서 대상 j의 좌표

7 Normalized distance function
: Raw data를 Normalization (Mean=0, Variance=1) 하여 scale상의 차이로 발생된 bias를 해결한 Euclidean distance ② Squared Euclidean distance Dij = ∑(Xik - Xjk)2 i=1 An example of Euclidean distance between two objects measured on two variables – X and Y. Y (X2-Y2) (Y2-Y1) Object 1 (X1-Y1) (X1-Y1) X Distance = (X2-X1) + (Y2-Y1) 2 2

8 ③ City-block distance (Manhattan distance)
r dijc = ∑ Xik - Xjk i=1 [문제점] ⅰ) 변수간에 correlation이 없다는 가정 ⅱ) Characteristic을 측정하는 단위(Scales)이 상이성이 가능 Object Purchase Commercial Distance Citi-block Probability(%) Viewing Time(min) (min) (second) A AB B AC C BC

9 ④ Mahalanobis distance
ⅰ) Standard Deviation으로 scaling해서 data 표준화 ⅱ) intercorrelation을 조정하기 위해서 within-group variance-covariance 합산하는 접근 방식 ⅲ) 변수간에 서로 correlated 되었을 때 가장 적합 ⑤ Minkowski distance dijM = [∑(Xik - Xjk)p]1/r

10 (3) Clustering Algorithms
Clustering Procedures Hierarchical Nonhierarchical Hierarchical Divisive Sequential Threshold Parallel Threshold Optimizing Partitioning Linkage Methods Variance Methods Centroid Methods Ward’s Method Single Linkage Complete Linkage Average Linkage

11 1) 계층적 군집방법 (Hierarchical Cluster Procedure)
① Agglomerative Procedure : 한 개의 대상에서 출발하여, 주위의 대상이나 cluster를 군집화하여 최종적으로 1개의 cluster로 만드는 방법 ⅰ) Single Linkage : minimum distance rule 군집이나 대상간의 최소거리로 군집화 ⅱ) Complete Linkage : maximum distance rule ⅲ) Average Linkage ⅳ) Ward's Method : W ● Within-cluster variance minimization rule ● Within-cluster distance의 전체 sum of square의 증가가 최소가 되게 cluster ⅴ) Centroid Method ● 대상이나 cluster의 Centroid(mean)간의 거리 최소화 ● 단점 : Metric data에만 적용 가능

12 ② Decisive Method : 큰 한 개의 cluster로 부터 분리시켜 가는 방법
Dendrogram illustrating hierarchical clustering. 01 02 03 04 05 06 07 08 Observation number

13 [Single Linkage : 단일기준 결합 방식]
1.5 A D 1.2 1.55 D 1.4 B B C C 1.3 [Complete Linkage : 완전기준 결합방식] A A 1.5 1.55 D D B B C C

14 [Average Linkage : 평균기준 결합방식]
1.45 D D 1.425 B B C C

15 [Ward Method] [Centroid Method] ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

16 2) 비계층적 군집방법 (Nonhierarchical Clustering Procedures)
= k-means clustering ⅰ) Sequential threshold procedure ① 하나의 cluster center를 선택하고 미리 산정된 거리 내에 있는 모든 대상을 그 cluster안에 포함시킨다. ② 두 번째 cluster center를 선택하고 미리 산정된 거리 내에 있는 ⅱ) parallel threshold Procedure ① 초기에 여러 개의 cluster center를 선정하여 가장 가까운 center 로 대상을 포함시킨다 ② threshold 거리는 조절될 수 있다 ⅲ) Optimizing Partitioning Method : 전체적인 optimizing criterion (e.g.,within-cluster distance의 평균)에 따라 나중에 대상을 cluster별로 재편입 시킬 수 있다

17 ▣ Nonhierarchical Clustering의 단점
② Cluster Center 선정이 임의적이다 ③ 결과가 data의 순서에 의존적이다 ▣ Nonhierarchical Clustering의 장점 ① center 선정에 있어서 nonrandorn ② Clustering 속도가 빠르다

18 3) 군집방법 선택 : Hierarchical Vs. Nonhierarchical
ⅰ) Hierarchical + Ward's Method + average linkage ⇒ 처음에 잘못 clustering되면 지속적으로 영향을 미친다 ⅱ) Hierarchical + Nonhierarchical ① Hierarchical procedure을 사용하여 최초 clustering 결과도출 (Ward Method + average linkage) ② 얻어진 cluster 숫자와 cluster centroid를 optimizing partitioning method의 input으로 사용

19 (4) Cluster 숫자 결정 ⅰ) 이론적, 개념적, 실제적 목적 고려 ⅱ) cluster간의 거리로 판단 ⅲ) Nonhierarchical clustering에서 Within Group Variance 을 도식화시켜 b/w Group Variance 꺾이는 부분을 찾아내어 cluster 숫자로 사용 ⅳ) cluster내에 case의 숫자로 판단 (one case를 가진 cluster는 바람직하지 않음)

20 (5) Cluster 해석 ⅰ) 보통 cluster centroid로 해석 ⅱ) Discriminant analysis 이용 (6) Validation ⅰ) 여러 가지 distance measure를 사용한 결과 비교 ⅱ) 여러 가지 Algorithm을 사용한 결과 비교 ⅲ) data를 임의로 둘로 나누어 각각의 cluster centroids 비교 ⅳ) 일부 data를 임의로 빼고 나머지에 대한 결과를 비교 ⅴ) Nonhierarchical Clustering은 자료의 순서에 의존적이므로 자료의 순서를 바꾸어 여러번 clustering하고 가장 안정적인 결과선택

21 Examples (1) Example 1 ■ 목적 : 신형 자동차를 출시하기 위해서 기존 시장의 15차종에 대한 특성 파악
■ 자동차 분류기준 (사전조사결과) : 외형크기와 배기량 ■ 외형 크기와 배기량은 표준화 자 동 차 종류 표준화된 승용차 속성의 평가 점수 외형적 크기 엔진 배기량 A B C D E F G H I J K L M N O 2.50 2.25 3.00 0.25 0.50 -0.25 -2.00 -1.50 -2.50 2.00 1.75 1.00 -0.50 -1.75 -2.25

22 ■ 승용차 특성을 2차원 도식화 [그림 18-4] X2 (엔진배기량) 승용차 특성의 2차원 도표 X1 (외향적 크기) A B
K L M N O A B D C J I H E F G

23 → Classification cluster center를 계산하여
■ SPSS의 Quick Cluster → Classification cluster center를 계산하여 각 cluster의 평균을 계산하여 다시 입력자료로 사용하는 방법 [그림 18-5] 단일결합방식 에 의한 결과 A D B C E G J F H I K N M O L 1 3 2 5 4 6 7 8 9 11 10 12 13 14

24 → Classification cluster center를 계산하여
■ SPSS의 Quick Cluster → Classification cluster center를 계산하여 각 cluster의 평균을 계산하여 다시 입력자료로 사용하는 방법 [그림 18-6] 완전결합방식 에 의한 결과 A D B C E G J F H I K N M O L 1 3 2 4 5 6 6 7 9 10 11 12 13 14

25 (2) Example 2 ■ 목적 : 회사 특성의 중요성 평가에 따른 고객 분류 (Stage 1) Partitioning Step 1 : Hierarchical cluster Analysis 1) Similarity measure : Squared Euclidean distances 2) Algorithm : Ward's method ⇒ within-cluster difference를 최소화 3) cluster 수 결정 : Two cluster가 최선안으로 결정

26 Agglomeration Coefficient to Next Level
TABLE 7.2 Analysis of Agglomeration Coefficient for Hierarchical Cluster Analysis Percentage Change in Agglomeration Coefficient to Next Level Number of] Clusters 10 9 8 7 6 5 4 3 2 1 8.9 8.5 9.2 9.3 12.1 17.0 17.6 61.9 -

27 Step 2 : Nonhierarchical Cluster Analysis
→ hierarchical procedure 결과를 Fine-tune ⇒ Hierarchical procedure의 결과 확인 Results of Nonhierarchical Cluster Analysis with Initial Seed Points from Hierarchical Results Mean Values* Cluster X X X X X X X Cluster Size Classification cluster centers 1 2 4.40 2.43 1.39 3.22 8.70 6.74 5.09 5.69 2.94 2.87 2.65 2.87 5.91 8.10 Final cluster centers 1 2 4.38 2.57 1.58 3.21 8.90 6.80 4.92 5.60 2.96 2.87 2.52 2.82 5.90 8.13 52 48

28 Variables Cluster M.S. Df Error M.S df F Ratio Probability
Significance Testing of Differences Between Cluster Centers X1 X2 X3 X4 X5 X6 X7 Delivery speed Price level Price flexibility Manufacturer’s image Overall service Sales force’s image Product quality .1883 2.1233 1 .9298 .7661 .8233 1.1778 .5682 .5786 1.2797 98.0 9.5959 .3314 3.6697 .000 .003 .566 .058 * X1 = Delivery speed : X2 = Price level : X3 = Price flexibility : X4 = Manufacturer’s image : X5 = Overall service : X6 = Sales force’s image : X7 = Product quality.

29 Stage Two : Interpretation
Group Means and Significance Level for Two-Group Nonhierarchical Cluster Solution Cluster Variables F Ratio Significance Stage Two : Interpretation X1 X2 X3 X4 X5 X6 X7 Delivery speed Price level Price flexibility Manufacturer’s image Overall service Sales force’s image Product quality 4.460 1.576 8.900 4.926 2.992 2.510 5.904 2.570 3.152 6.888 5.570 2.840 2.820 8.038 105.00 76.61 111.30 8.73 1.02 4.17 82.68 .0000 .0039 .3141 .0438 Stage Three : Profiling Other variables of interest X9 Usage level X10 Satisfaction level 49.88 5.16 42.32 4.38 21.312 26.545 .0000

30 Stage Two : Interpretation
- Table 7.4 참조 - X5는 두 그룹 사이에 차이가 없는 것으로 평가됨 - Cluster 1 focuses ⅰ) delivery speed ⅱ) price flexibility Cluster 2 focuses ⅰ) price ⅱ) manufacturer's image ⅲ) sales force image ⅳ) product quality Stage Three : Validation - Table 7.5 참조 (결과의 consistency 확인) ⇒ 무작위로 선택한 subset으로 clustering하여 비교

31 Cluster X1 X2 X3 X4 X5 X6 X7 Cluster Size
TABLE 7.5 Results of Nonhierarchical Cluster Analysis with Randomly Selected Initial Seed Points Mean Values* Cluster X X X X X X X Cluster Size Classification cluster centers 1 2 4.95 1.76 1.14 2.70 9.03 6.87 6.55 5.50 3.21 1.97 3.79 2.70 5.09 8.45 Final cluster centers 1 2 4.47 2.63 1.57 3.10 8.93 6.94 4.99 5.49 2.99 2.84 2.57 2.75 5.78 8.07 48 52 Significance Testing of Differences Between Cluster Centers Variables Cluster M.S Df Error M.S df F Value Probability X1 X2 X3 X4 X5 X6 X7 Delivery speed Price level Price flexibility Manufacturer’s image Overall service Sales force’s image Product quality 6.2640 .5883 .7477 1 .9016 .8454 .9367 1.2292 .5641 .5927 1.2007 98.0 5.0958 1.0428 1.2616 .000 .026 .310 .264 * X1 = Delivery speed : X2 = Price level : X3 = Price flexibility : X4 = Manufacturer’s image : X5 = Overall service : X6 = Sales force’s image : X7 = Product quality.


Download ppt "Cluster Analysis (군집 분석)"

Similar presentations


Ads by Google