Presentation is loading. Please wait.

Presentation is loading. Please wait.

Performance Measures II

Similar presentations


Presentation on theme: "Performance Measures II"— Presentation transcript:

1 Performance Measures II
Prof. Kislaya Prasad BUDT 733 (Spring 2015)

2 The Lift Charts Lift charts analyze the “class of interest”
Fraudulent claims, responders to mailing, patients at risk of a heart attack, … , or customers that prefer regular beer! Builds on the probability of belonging to the class of interest Sorts all records in a descending order of probability of belonging to the class Cumulative number of cases is on the x-axis Cumulative number of true positives is on the y-axis College Park, Spring 2015

3 Sort by Probability of Success
Row Id. Predicted Class Actual Class Prob. for 1 (success) Log odds Gender Married Income Age Cumulative Actual Class 99 1 $70,981.00 45 84 $53,768.00 34 2 100 $48,259.00 32 3 59 $49,585.00 4 91 $43,593.00 27 5 74 $42,775.00 6 98 $47,495.00 35 7 60 $44,802.00 8 78 $30,841.00 24 9 79 $52,341.00 42 10 83 $55,281.00 48 11 57 $36,821.00 12 80 $41,779.00 39 13 70 $26,598.00 29 14 86 $63,555.00 62 15 72 $46,626.00 46 16 $38,176.00 40 61 $43,194.00 47 17 $28,513.00 63 $38,671.00 44 18 $23,234.00 31 $44,955.00 53 93 $46,273.00 56 19 28 $28,094.00 $37,263.00 49 $43,741.00 54 50 $26,186.00 38 22 $29,293.00 43 $42,051.00 58 $32,739.00 $33,027.00 51 $22,408.00 $27,078.00 $36,030.00 55 88 $52,875.00 71 20 33 $39,923.00 $24,302.00 $25,440.00 25 $40,582.00 66 $35,678.00 64 Construct cumulative of Actual Class Plot 20 of 40 cases in validation data are successes. In any M randomly selected cases we will have M/2 successes. This line is denoted in red Sort by Probability of Success College Park, Spring 2015

4 Lift Chart (Validation Data)
Lift Charts generated in R (See script for details) Reference line, represents selecting records (in our case customers) using the Naïve rule

5 ROC (Receiver Operating Characteristic) Curve
Developed in the 1950s for signal detection theory to analyze noisy signals Characterizes the trade-off between positive hits and false alarms Performance of each classifier represented as a point on the ROC curve Changing the threshold of an algorithm, the sample distribution or cost matrix changes the location of the point Decide where to set the cutoffs to find out false positives and false negatives College Park, Spring 2015

6 ROC (Receiver Operating Characteristic) Curve
The ROC curve plots the pairs (1-specificity, sensitivity) Interesting points on the curve: (0,0): declare everything to be 0 (1,1): declare everything to be 1 (1,0): perfect classification Diagonal line: Random guessing Below diagonal line: prediction is opposite of the true class – the model is worse than random guessing! Cutoff 0 Cutoff 1 College Park, Spring 2015

7 Creating a ROC curve with R
We have already calculated sensitivity and specificity in our data table example Add 1-specificity column Insert plot, with 1-specificity as the x values and sensitivity as the y values See script for details College Park, Spring 2015

8 Using ROC for Model Comparison
In this example neither model consistently outperforms the other M1 is better for small sensitivity values M2 is better for larger sensitivity values College Park, Spring 2015

9 Summary The performance measures differ depending on the use of the model In data-mining we are usually more interested in prediction accuracy than statistical fit When predicting numerical values we can use numerous error measures such as the mean absolute error or RMSE. When classifying we again have many options: accuracy, sensitivity, misclassification costs, the ROC curve etc. Whenever possible, we should use a validation sample to estimate the performance of the model If the sample size is too small to split up the model – there are methods to estimate the prediction accuracy College Park, Spring 2015

10 Next We will introduce two more classification methods!
Naive Bayes – for classification K-nearest neighbor – for classification and prediction Both identified as one of the top 10 algorithms by the IEEE International Conference on Data Mining (ICDM) in December 2006! (see here: College Park, Spring 2015


Download ppt "Performance Measures II"

Similar presentations


Ads by Google