Microsoft PowerPoint - Microsoft10.pptx

Size: px

Start display at page:

Download "Microsoft PowerPoint - Microsoft10.pptx"

节镰郁
5 years ago
Views:

1 資料探勘在電信公司交叉銷售之應用李御璽銘傳大學資訊工程學系

2 專案目的某電信公司針對手機用戶推廣無線寬頻網卡找出潛在具有申裝無線寬頻網卡的客戶名單資料集中共有 1,114 筆資料已申裝無線寬頻網卡者 206 人 (18.49%) 未申裝無線寬頻網卡者 908 人 (81.51%)

3 Broadband 資料集欄位名稱資料型別意義值分布 CUST_ID int- 數值型客戶代碼 [1,1114] GENDER char(1)- 二元型性別 1/0 AGE int- 數值型年齡 [18,76] TENURE int- 數值型申辦門號月份數 [1,72] CHANNEL char(1)- 類別型申辦通路 1,2,3,4 AUTOPAY char(1)- 二元型自動轉帳扣繳 1/0 ARPB_3M float- 數值型最近 3 個月平均電話費 [68,2049] CALL_PARTY_CNT float- 數值型撥出號碼數 [0,51] DAY_MOU float- 數值型白天通話分鐘數 [0.0,1162.4] AFTERNOON_MOU float- 數值型下午通話分鐘數 [0.0,4915.0] NIGHT_MOU float- 數值型夜間通話分鐘數 [0.0,1502.1] AVG_CALL_LENGTH float- 數值型平均通話長度 [2.2,6.83] BROADBAND char(1)- 二元型無線寬頻註記 1/0

4 Training & Test Datasets 訓練資料 : 891 筆已申裝無線寬頻網卡者 165 人 (18.52%) 未申裝無線寬頻網卡者 726 人 (81.48%) 測試資料 : 223 筆已申裝無線寬頻網卡者 41 人 (18.39%) 未申裝無線寬頻網卡者 182 人 (81.61%)

5 Detailed Attribute Analysis

6 Detailed Attribute Analysis 用四分位數法偵測離群值 / 極端值 7 個數值型屬性有離群值 / 極端值

7 Detailed Attribute Analysis- GENDER ( ) 差距不大

8 Detailed Attribute Analysis- CHANNEL ( ) 不同申辦通路有影響

9 Detailed Attribute Analysis- AUTOPAY ( ) 不自動轉帳扣繳者較有興趣

10 Detailed Attribute Analysis- AGE ( ) 年紀越大者越感興趣

11 Detailed Attribute Analysis- AGE ( ) 70 歲以上直接判定為有興趣測試資料有 2 筆符合此規則且均有興趣

12 Detailed Attribute Analysis- TENURE ( ) 申辦門號月份越長者越感興趣

13 Detailed Attribute Analysis- TENURE ( ) 往來 70 個月以上直接判定為有興趣測試資料有 6 筆符合此規則且均有興趣

14 Detailed Attribute Analysis- ARPB_3M ( ) 有嚴重的離群值 / 極端值

15 Detailed Attribute Analysis- ARPB_3M ( ) 最近 3 個月平均電話費 >=700 直接判定為有興趣測試資料有 4 筆符合此規則且均有興趣

16 Detailed Attribute Analysis- CALL_PARTY_CNT ( ) 有離群值 / 極端值

17 Detailed Attribute Analysis- CALL_PARTY_CNT ( ) 撥出號碼數 >=45 直接判定為有興趣測試資料有 0 筆符合此規則

18 Detailed Attribute Analysis- DAY_MOU ( ) 有離群值 / 極端值

19 Detailed Attribute Analysis- DAY_MOU ( ) 白天通話分鐘數 >=800 直接判定為有興趣測試資料有 2 筆符合此規則且均有興趣

20 Detailed Attribute Analysis- AFTERNOON_MOU ( ) 有嚴重的離群值 / 極端值

21 Detailed Attribute Analysis- NIGHT_MOU ( ) 有嚴重的離群值 / 極端值

22 Detailed Attribute Analysis- AVG_CALL_LENGTH ( ) 有離群值 / 極端值

23 Detailed Attribute Analysis- AVG_CALL_LENGTH ( ) 平均通話長度 >=6.1 分鐘直接判定為有興趣測試資料有 1 筆符合此規則且均有興趣

24 Summary 訓練資料被直接判定為有興趣的有 31 筆, 且全部正確測試資料被直接判定為有興趣的有 8 筆, 且全部正確只要符合 1 個條件即可條件是以 0R 的型式來應用刪除了一些離群值及極端值, 且立即鎖定固定的目標不正常資料 ( 離群值及極端值 ) 常常隱含著某種關係比模型產生出的機率較高者還理想

25 New Training & Test Datasets 訓練資料 : 860 筆已申裝無線寬頻網卡者 134 人 (15.58%) 未申裝無線寬頻網卡者 726 人 (84.42%) 測試資料 : 215 筆已申裝無線寬頻網卡者 33 人 (15.35%) 未申裝無線寬頻網卡者 182 人 (84.65%)

26 Results in Training Dataset CART Model Accuracy=84.42% Precision=N/A Recall=0% F-Measure=N/A C5.0 Model Accuracy=93.37% Precision=97.53% Recall=58.96% F-Measure=73.49% Neural Network Model Accuracy=87.91% Precision=67.44% Recall=43.28% F-Measure=52.73%

27 Results in Test Dataset CART Model Accuracy=84.65% Precision=N/A Recall=0% F-Measure=N/A C5.0 Model Accuracy=85.58% Precision=57.14% Recall=24.24% F-Measure=34.04% Neural Network Model Accuracy=81.86% Precision=37.50% Recall=27.27% F-Measure=31.58%

28 Decision Tree

29 Decision Rule See File Two Interesting Rules Rule 12 for 1 (35; 0.971) if ARPB_3M > 309 and AUTOPAY = 0 and GENDER = 1 then 1 Rule 16 for 1 (19; 0.947) if ARPB_3M > 309 and AUTOPAY = 0 and GENDER = 0 and TENURE > 63 then 1

0 Model Accuracy=99.19% Precision=100% Recall=94.

30 Results in Training Dataset (Boosting) CART Model Accuracy=95% Precision=97.89% Recall=69.40% F-Measure=81.22% C5.0 Model Accuracy=99.19% Precision=100% Recall=94.78% F-Measure=97.32% Neural Network Model Accuracy=100% Precision=100% Recall=100% F-Measure=100%

31 Results in Test Dataset (Boosting) CART Model Accuracy=88.37% Precision=70.00% Recall=42.42% F-Measure=52.83% C5.0 Model Accuracy=90.70% Precision=93.33% Recall=42.42% F-Measure=58.33% Neural Network Model Accuracy=91.63% Precision=100% Recall=45.45% F-Measure=62.50%

32 Results in Training Dataset (Bagging) CART Model Accuracy=88.49% Precision=90.70% Recall=29.10% F-Measure=44.07% Neural Network Model Accuracy=97.44% Precision=99.12% Recall=84.33% F-Measure=91.13%

33 Results in Test Dataset (Bagging) CART Model Accuracy=86.51% Precision=75.00% Recall=18.18% F-Measure=29.27% Neural Network Model Accuracy=88.84% Precision=84.62% Recall=33.33% F-Measure=47.83%

34 Summary for Experiments Training Dataset CART Accuracy Precision Recall F-Measure Original 84.42% N/A 0% N/A Boosting 95.00% 97.89% 69.40% 81.22% Bagging 88.49% 90.70% 29.10% 44.07% Test Dataset CART Accuracy Precision Recall F-Measure Original 84.65% N/A 0% N/A Boosting 88.37% 70.00% 42.42% 52.83% Bagging 86.51% 75.00% 18.18% 29.27%

35 Summary for Experiments Training Dataset C5.0 Accuracy Precision Recall F-Measure Original 93.37% 97.53% 58.96% 73.49% Boosting 99.19% 100% 94.78% 97.32% Bagging N/A N/A N/A N/A Test Dataset C5.0 Accuracy Precision Recall F-Measure Original 85.58% 57.14% 24.24% 34.04% Boosting 90.70% 93.33% 42.42% 58.33% Bagging N/A N/A N/A N/A

Summary for Experiments Training Dataset NN Accuracy

73% Boosting 100% 100% 100% 100% Bagging 97.44% 99.

13% Test Dataset NN Accuracy Precision Recall F-Measure

36 Summary for Experiments Training Dataset NN Accuracy Precision Recall F-Measure Original 87.91% 67.44% 43% 52.73% Boosting 100% 100% 100% 100% Bagging 97.44% 99.12% 84.33% 91.13% Test Dataset NN Accuracy Precision Recall F-Measure Original 81.86% 37.50% 27.27% 31.58% Boosting 91.63% 100% 45.45% 62.50% Bagging 88.84% 84.62% 33.33% 47.83%

37 Summary for Training Dataset Boosting, F- Measure CART, 81.22% Original, F- Measure C5.0, 73.49% Boosting, F- Measure C5.0, 97.32% Boosting, F- Measure NN, 100% Bagging, F-Measure NN, 91.13% Bagging, F-Measure CART, 44.07% Original, F- Measure NN, 52.73% Original Bagging Boosting Original, F- Measure CART, 0% Bagging, F-Measure C5.0, 0%

38 Summary for Test Dataset Boosting, F- Measure CART, 52.83% Boosting, F- Measure C5.0, 58.33% Boosting, F- Measure NN, 62.50% Bagging, F-Measure NN, 47.83% Bagging, F-Measure CART, 29.27% Original, F- Measure C5.0, 34.04% Original, F- Measure NN, 31.58% Original Bagging Boosting Original, F- Measure CART, 0% Bagging, F-Measure C5.0, 0%

PK IBM Warren McCulloch Walter Pits MP 1949 Hebb Hebb Hebb 145

PK IBM Warren McCulloch Walter Pits MP 1949 Hebb Hebb Hebb 145 6 01 06 2006 12 5 PK 20 4 4 800 1000 IBM 1997 1943 Warren McCulloch Walter Pits MP 1949 Hebb Hebb Hebb 145 01 SQL Server 2008 Data Mining 1969 Seymour Papert Marvin Minsky XOR 1982 John Hopfield HNN 1986