使用決策樹的預(yù)測(cè)建模-閱讀頁
2025-07-24 13:08本頁面
【正文】 %等實(shí)際響應(yīng)者的百分?jǐn)?shù)。記得總體的響應(yīng)率大約為20%。因此提升圖描繪的是基于基礎(chǔ)線的相對(duì)改進(jìn). Cumulative %ResponseCumulative Lift Value我們知道在按預(yù)測(cè)概率排序的前10% %. % 除以20% (baseline rate) 得到略大于4的數(shù), 這表明你在這個(gè)組里會(huì)得到比相同人數(shù)的簡單隨機(jī)樣本里4 倍多的欺詐者.你可能會(huì)提出這樣的問題：每一個(gè)組里響應(yīng)者占全部響應(yīng)者的百分?jǐn)?shù)是多少？ 。要查看計(jì)算結(jié)果，選擇%Captured Response旁邊的圓鈕。你可以看到如果你拒絕申請(qǐng)者的百分?jǐn)?shù)為 40%,你可能挑出了大約80% 將會(huì)欺詐的人(提升比例大于2!). 關(guān)閉Lift Chart 和Assessment Tool窗口.In order to choose the appropriate threshold to classify observations positively or negatively, the cost of misclassification must be considered. In the home equity line of credit example, you are modeling the probability of a default, which is coded as a 1. Therefore, Enterprise Miner sets up the profit matrix as shown above.Assume that every two dollars loaned returns three dollars if the borrower does not default. Rejecting a good loan for two dollars forgoes the expected dollar profit. Accepting a bad loan for two dollars forgoes the twodollar loan itself (assuming that the default is early in the repayment period).The costs of misclassification are shown in the table.One way to determine the appropriate threshold is a theoretical approach. This approach uses the plug in Bayes rule. Using simple decision theory, the optimal threshold is given by q.Using the cost structure defined for the home equity example, the optimal threshold is 1/(1+(2/1)) = 1/3. That is, reject all applications whose predicted probability of default exceeds .You can obtain the same result using the Assessment node in Enterprise Miner by using the profit matrix to specify the profit associated with the level of the response being modeled (in this case, a loan default or a 1). As a bonus, you can estimate the fraction of loan applications you must reject when using the selected threshold.選擇決策閾值首先考慮從理論上決定的決策閾值 1. 返回到項(xiàng)目Project1 流程圖，打開 Default Tree 節(jié)點(diǎn),并選擇打分（Score）標(biāo)簽 2. 選擇 Training, Validation, and Test旁的選箱，這樣預(yù)測(cè)值被加到數(shù)據(jù)集合。4. 在Default Tree 節(jié)點(diǎn)后加入一個(gè)Insight節(jié)點(diǎn)。7. 從這個(gè)Default Tree節(jié)點(diǎn)選擇驗(yàn)證數(shù)據(jù)集合。10. 關(guān)閉節(jié)點(diǎn)，保存所做的改動(dòng)。該數(shù)據(jù)集合的新變量之一是P_BAD1, 它是預(yù)測(cè)的貸款人目標(biāo)變量＝1（貸款欺詐）的概率。 Y.14. 在Y欄涂黑 P_BAD1 并選擇Asc/Des 以排序15. 選擇 OK.16. 在數(shù)據(jù)表中移動(dòng)鼠標(biāo)，你可以看到380個(gè)觀測(cè)記錄其預(yù)測(cè)的欺詐概率大于1/3.因此根據(jù)理論方法，1967個(gè)申請(qǐng)人中380個(gè)將被拒絕，或者近似地19%的申請(qǐng)人要被拒絕。 Lift Chart.5. 在提升圖窗口的左下角，選擇Edit… 來定義一個(gè)目標(biāo)特征（target profile）.6. 在Editing Assessment Profile for BAD 窗口里, 右擊列有向量和矩陣（vectors and matrices）的區(qū)域的空白區(qū)并選擇Add.7. 涂黑這個(gè)新的Profit matrix, 并在矩陣?yán)镙斎霐?shù)值如下圖! For credit screening, a target value of 1 implies a default and, thus, a loss. A target value of 0 implies a paid repaid loan and, thus, a profit. The fixed cost of processing each loan application is insubstantial and taken to be zero.8. 右擊Profit matrix 并選擇Set to Use. 現(xiàn)在profit matrix 處于活動(dòng)狀態(tài)，因?yàn)樗懊嬗?號(hào)。前20%的部分顯示了拒絕申請(qǐng)人的利潤。! In Enterprise Miner, the NonCumulative profit chart never dips below zero. This is because a cutoff value is chosen and there is no cost below this level because there is no action. As a result, the cumulative profit chart can be mislea
環(huán)評(píng)公示相關(guān)推薦
鄂ICP備17016276號(hào)-1