1. Data Mining CSCI 307, Spring 2019 Lecture 3
Fielded Applications
Ethics

2. Fielded Applications
The result of learning -- or the learning method itself -- is deployed in practical applications
These are actual ML problems put into use versus research problems or the toy problems....that we'll look at this semester.
Processing loan applications
Screening images for oil slicks
Electricity supply forecasting
Diagnosis of machine faults
Marketing and sales (Big $$)
Web Mining (PageRank, query content mining, etc.)
Separating crude oil and natural gas (rules used in setting parameters and now take 10 minutes rather than a day)
Reducing banding in rotogravure printing (reduced artifacts from 500 down to 30)
Finding appropriate technicians for telephone faults (saved $10 million)
Scientific applications: biology, astronomy, chemistry (analyzing cell structure accelerates drug discovery, cataloging celestial objects, etc.)
Automatic selection of TV programs
Monitoring intensive care patients

3. Processing Loan Applications
(American Express)
Make a decision
Decisions involving judgment.
Given: Questionnaire with financial and
personal information
Question: Should money be lent?
Simple statistical method covers 90% of cases Borderline cases (the other 10%) referred to loan officers
But: 50% of accepted borderline cases defaulted!
Solution: Reject all borderline cases?
No! Borderline cases are most active customers....

4. Enter Machine Learning
1000 training examples of borderline cases
20 attributes:
age
years with current employer
years at current address
years with the bank
other credit cards possessed, …
Learned Rules: correct on 67% of cases
Human Experts: correct only 50% of cases
Rules could be used to explain decisions to customers

5. Screening Images Given: radar satellite images of coastal waters
Produce attributes
Given: radar satellite images of coastal waters
Problem: detect oil slicks in those images
Oil slicks appear as dark regions with changing size and shape
Not easy: Lookalike dark regions can be caused by weather conditions (e.g. high wind)
Expensive process requiring highly trained personnel
I am clearly not trained, I have no idea what these are.

6. Enter Machine Learning
Classification scheme is not what is deployed but the learning scheme itself.
Enter Machine Learning
Extract dark regions from normalized image
Attributes:
size of region
shape, area
intensity
sharpness and jaggedness of boundaries
proximity of other regions
info about background
Constraints/Problems:
Few training examples—oil slicks are rare!
Unbalanced data: most dark regions aren’t slicks
Regions from same image form a batch—Different batches have different backgrounds
Requirement: adjustable false-alarm rate
Input: images
Output: Smaller images with marked regions
Use image processing algorithms to normalize AND produce the attributes.
The learning scheme is applied to the attributes.

7. Fifteen years of data, plus "additive" factors.
Load Forecasting
Do it faster
Electricity supply companies need forecast of future demand for power
Forecasts of min/max load for each hour ==> significant savings
Given: Manually constructed load model that assumes “normal” climatic conditions
Problem: Adjust for weather conditions
Static model consists of:
base load for the year
load periodicity over the year
effect of holidays
Fifteen years of data, plus "additive" factors.

8. Enter Machine Learning
Prediction corrected using “most similar” days
Attributes:
temperature
humidity
wind speed
cloud cover readings
plus difference between actual load and predicted load
Average difference among eight “most similar” days added to static model
Linear regression coefficients form attribute weights in similarity function
The resulting system yielded the same performance as expert forecaster, but took seconds to compute rather than hours.

9. Diagnosis of Machine Faults
Let the expert help by tweaking the rules
Diagnosis of Machine Faults
Diagnosis: Classical domain of expert systems
Given: Fourier analysis of vibrations measured at various points of a device’s mounting
Question: Which fault is present?
Preventative maintenance of electromechanical motors and generators
Information very noisy
So far: diagnosis by expert/hand-crafted rules
Not IF there is a fault, but WHAT is the fault?

10. Enter Machine Learning
Available: 600 faults with expert’s diagnosis
~300 unsatisfactory, rest used for training
Expert was not satisfied with initial rules because they did not relate to his domain knowledge
Further background knowledge resulted in more complex rules that were satisfactory
Learned rules outperformed hand-crafted ones

11. Marketing and Sales I
Make money
Companies precisely record massive amounts of marketing and sales data
Applications:
Customer loyalty:
Identifying customers that are likely to defect by detecting changes in their behavior
(e.g. banks/phone companies)
Special offers:
Identifying profitable customers
(e.g. reliable owners of credit cards that need extra money during the holiday season)

12. Marketing and Sales II Market Basket Analysis
Association techniques find groups of items that tend to occur together in a transaction (used to analyze checkout data)
Historical analysis of purchasing patterns
Identifying prospective customers
Focusing promotional mailings (targeted campaigns are cheaper than mass-marketed ones)
Recall Beer/Diapers on Thursdays (just a few items) and Saturdays (full shopping)

13. The Data Mining Process

14. Machine Learning and Statistics
Historical difference (grossly oversimplified):
Statistics: testing hypotheses
Machine learning: finding the right hypothesis
But: huge overlap
Decision trees
Nearest-neighbor methods
Today: perspectives have converged
Most ML algorithms employ statistical techniques
Chapter 4

15. Generalization as Search
Data Mining as search through all possible rule sets:
Set of all possible rule sets is very large
It's difficult to search the entire set
Noise in the data could eliminate all rule sets
Biases in searches help address issues
Language Bias (How are rule sets described?)
Search bias (Finding a good rule rather than the best rule)
Overfitting avoidance bias (Simpler trees may be better at generalizing to new examples)

16. Data Mining and Ethics I
Ethical issues arise in practical applications
Anonymizing data is difficult
85% of Americans can be identified from just zip code, birth date and sex
Data mining often used to discriminate
e.g. loan applications: using some information (e.g. sex, religion, race) is unethical
Ethical situation depends on application
e.g. same information ok in medical application
Attributes may contain problematic information
e.g. area code may correlate with race
Anonymizing is hard.
A Massachusetts governor received his health records in the mail after a MA hospital released “anonymized” data of records!
Less damaging “problematic data” and the desire to get more data!
I got a phone call “qualifying” me to go back to school. The benign questioning turned to the caller saying they had all my information, but they just needed to “verify” it.

17. Data Mining and Ethics II
Important questions:
Who is permitted access to the data?
For what purpose was the data collected?
What kind of conclusions can be legitimately drawn from it?
Caveats must be attached to results
Purely statistical arguments are never sufficient!
People must take results of DM along with their own knowledge and make decisions of how and whether to apply them.
-- Fake LinkedIn accounts were created so that data from real accounts could be scraped and used.
Also, the fake profiles themselves create an issue of trust in the system itself.
++ To prevent fires before they start, NYC Fire department used predictive models (using building age, vacancy rate, presence of suppression systems, etc.) to help compile a list of potentially-at-risk addresses so the city building inspectors can prioritize their work. (Used to choose buildings at random.)
++ To reduce waits in physician’s office, predict cancellations with more accuracy….with smart scheduling the claim is doctor’s can see lots more patients each week and wait times are reduced.