en-de
en-es
en-fr
en-pt
en-sl
en
en-zh
0.25
0.5
0.75
1.25
1.5
1.75
2
Learning in the real world
Published on Aug 02, 20115236 Views
Quite properly, the early days of computational and machine learning assumed the data were well-behaved when developing learning algorithms and strategies. Unfortunately, such an assumption is unwise
Related categories
Chapter list
Learning in the real world00:00
In science:02:36
The ‘details’ in learning theory include:04:02
The ‘details’ arise from the context06:35
Example07:20
A vehicle to illustrate the ideas08:46
Basic structure10:03
Problem-based criteria11:09
Classification accuracy criteria12:45
A case study: Comparing credit scorecards14:26
Example 1: Building a new (better) scorecard15:51
Extreme illustration17:17
To tackle use ‘reject inference’18:26
Example 2: Choosing a new scorecard20:19
‘Old’ score card and ‘New’ score card22:55
Joint population density function23:44
The crunch24:44
Class scores - 125:28
Class scores - 225:44
Class scores - 326:05
Class scores - 426:36
Class scores - 527:06
Class scores - 627:35
The effect28:26
Example 3: Fraud detection29:33
Predicted class - True class31:27
Problems 1 and 232:16
A superior measure33:52
Included timeliness in the count c34:42
Overall performance measure for given threshold35:12
Issue 2: Bias in evaluation - 136:50
Issue 2: Bias in evaluation - 237:21
Issue 2: Bias in evaluation - 338:08
Issue 2: Bias in evaluation - 439:04
Consider straightforward estimates39:39
Consider a single terminating account with c fraudulent transactions40:30
suppose c = 141:31
Example 4: Discrimination42:34
US Equal Credit Opportunity Act, 197443:45
Credit scorecards do not include sex as a predictor variable44:42
Solution45:05
Solution depends on the real aim46:09
Build separate models for men and women46:54
Build a single model using all the variables we can think of, including sex and any proxies47:19
Current situation is neither (A) nor (B)48:10
Relevance48:25
1978: female employees sued Los Angeles Department of Water and Power48:43
Issues49:08
Suppose cost of driving insurance is equalised at a weighted mean of men and women50:37
Conclusions51:21
Thank you51:36