Introduction
Here are some problems:
- What eggs to select for IVF?
(p2)
- What will software cost to develop?
- What diseases does a patient have
p25
- Which loan applications to fund (p22)?
- What houses will have the best resale value?
- Which parts of the program need more inspection?
- What products are best to sell to what markets?
(p26)
- What cows to keep and which to send to the abattoir ?
(p2)
- How to teach a satellite to distinguish between cloud shadows and oil spills
(p23)?
- How much electricity will b needed in two hours (i.e. what cola-powered generators to fire up)?
(p24)?
In the modern era, these problems have data mining solutions.
- Lots of data: world's databases doubling in size every 20 months;
- Internet, Radio Frequency Identification (RFID) tracking, on-line shopping (patterns of sales tracked at Amazon)
But, like Spider Man says, with great power comes great responsibility; e.g.
- Can consumers audit how their loans applications are analyzed?
- Is it right that our learners tell us what embryos are not implanted ?
- How can a Hindu protest at letting vital ethical decisions (like deciding what cows are
Aghanya--that which must not be slaughtered)
being made by a machine?
But before we can discuss the ethical implications of data mining, we must first understand both the power and limitations of the technology.
So we'll get to the ethics after discussing the technology.
The Technology
Data (Arff format)
@relation weather.symbolic
@attribute outlook {sunny, overcast, rainy}
@attribute temperature {hot, mild, cool}
@attribute humidity {high, normal}
@attribute windy {TRUE, FALSE}
@attribute play {yes, no}
@data
sunny,hot,high,FALSE,no
sunny,hot,high,TRUE,no
overcast,hot,high,FALSE,yes
rainy,mild,high,FALSE,yes
...
Summarized by J48 (decision tree learner)
- Find the attribute value that most divides up the data
- Split the data on that value
- Recurse on each subset
- Each split is a node in a decision tree.
outlook = sunny
| humidity = high: no
| humidity = normal: yes
outlook = overcast: yes
outlook = rainy
| windy = TRUE: no
| windy = FALSE: yes
How good is this theory?
- 10 times,
- divide data into 90% train, 10% test,
- learn on train,
- apply on test.
- Report average
results across the 10-way
a b <-- classified as
5 4 | a = yes
3 2 | b = no
To get some summary statistics out of this:
TP Rate FP Rate Precision Recall F-Measure Class
0.556 0.6 0.625 0.556 0.588 yes
0.4 0.444 0.333 0.4 0.364 no
Summarized by Naive Baues
A Naive Bayes classifier collects internal statistics, does not generate an explicit explanation:
Class yes: Prior probability = 0.63
outlook: Discrete Estimator. Counts = 3 5 4 (Total = 12)
temperature: Discrete Estimator. Counts = 3 5 4 (Total = 12)
humidity: Discrete Estimator. Counts = 4 7 (Total = 11)
windy: Discrete Estimator. Counts = 4 7 (Total = 11)
Class no: Prior probability = 0.38
outlook: Discrete Estimator. Counts = 4 1 3 (Total = 8)
temperature: Discrete Estimator. Counts = 3 3 2 (Total = 8)
humidity: Discrete Estimator. Counts = 5 2 (Total = 7)
windy: Discrete Estimator. Counts = 4 3 (Total = 7)
Results of the Naive Bayes classifier:
- 10 times,
- divide data into 90% train, 10% test,
- learn on train,
- apply on test.
- Report average
results across the 10-way
a b <-- classified as
7 2 | a = yes
4 1 | b = no
TP Rate FP Rate Precision Recall F-Measure Class
0.778 0.8 0.636 0.778 0.7 yes
0.2 0.222 0.333 0.2 0.25 no
Definitions
In general, these are hard questions:
- What does learning mean?
- What does intelligence mean?
- Does a slipper learn the shape of your foot?
But a simpler kind of "learning" is much easier to understand
PERFORMANCE systems
(e.g. Naive Bayes)
.------------->-----------.
| |
| v
TRAINING | PREDICTIONS
data --> LEARNING (on test data)
| ^
| |
.----> GENERALIZATION -->-.
EXPLANATION systems
(e.g. J48)
(This is the kind of learning explored in this class.)
Explanation systems build some human-readable intermediary,
then use that intermediary to make predictions.
Performance systems don't bother with explaining themselves, they just make predictions.
In the above, decision trees are an explanation system and Naive Bayes was a performance system.
The Over Fitting problem
Fixation on irrelevant details ⇒ over fitting.
Cause: noise; i.e. spurious signals not connected
to the output class
Symptoms:
- Overly complex theory
- Poorer performance on future examples
Solutions:
- Testing: assess a learned theory via its performance on data not seen during training (as done above).
- Pruning: after a theory is built, try throwing bits away; e.g. prune a decision tree back from the leaves, see how that changes performance.
confidence limit for pruning = 0.1 (very selective)
c0.1
Horsepower <= 82: _0 (9.0)
Horsepower > 82
| Horsepower <= 190: _20 (70.0/4.0)
| Horsepower > 190: _40 (14.0/5.0)
a b c d <-- classified as
9 1 0 0 | a = _0
0 65 5 0 | b = _20
0 7 5 0 | c = _40
0 0 1 0 | d = _60
TP Rate FP Rate Precision Recall F-Measure Class
0.9 0 1 0.9 0.947 _0
0.929 0.348 0.89 0.929 0.909 _20
0.417 0.074 0.455 0.417 0.435 _40
0 0 0 0 0 _60
confidence limit for pruning = 0.25 (default, less selective)
c0.25
Horsepower <= 82: _0 (9.0)
Horsepower > 82
| Horsepower <= 190: _20 (70.0/4.0)
| Horsepower > 190
| | Drive_train_type = 1
| | | Highway_MPG <= 26: _40 (4.0/1.0)
| | | Highway_MPG > 26: _20 (2.0)
| | Drive_train_type = 0: _40 (7.0/1.0)
| | Drive_train_type = 2: _20 (1.0)
a b c d <-- classified as
9 1 0 0 | a = _0
0 65 5 0 | b = _20
0 7 5 0 | c = _40
0 0 1 0 | d = _60
TP Rate FP Rate Precision Recall F-Measure Class
0.9 0 1 0.9 0.947 _0
0.929 0.348 0.89 0.929 0.909 _20
0.417 0.074 0.455 0.417 0.435 _40
0 0 0 0 0 _60
Generalization as Search
Give a set of possible concept descriptors
- e.g. age > 6; wealthy
...and a set of combination methods
- e.g. and, or, not, if-then-else, it-unless, etc
How to search the space of possible descriptors*combinations?
Problem: the space of descriptors*combinations is usually impossibly large for a complete search
Solution: heuristic search (cut corners)
- e.g. 1R: only ever learn decision trees of depth one
- i.e. always try combinations on a single attribute
- i.e. never try combinations of attributes
- BTW, works ok, but usually beaten by other methods
Different kinds of bias
Search bias
When growing a theory, what concepts*descriptors do you look at first?
- Depending on your goals your search will find different conclusions.
- e.g. are you seeking the theory with highest performance on new test data, or the
smallest theory, or a combination of both, see (minimum description length [p179][wt]).
- e.g. decision trees: what goes into the root (thereby effecting all sub-trees)?
- e.g. greedy search (only consider the next best thing) vs ε-greedy search (consider anything within ε of the current best thing).
Over fitting Avoidance Bias
When shrinking a theory, how do you control pruning?
- What trade-offs do you allow for size vs performance?
- What do you try pruning first, then second, then...
- e.g. if throwing away a decision tree sub-branch reduces performance by 5% but decision tree size by 50%, is that a goods prune?
Sample Bias
Learn from available examples, not the space of all possible examples.
- e.g. this data only comes from the east coast and they do things differently over there.
Language Bias
What is space of legal combinations and legal descriptors?
- e.g. many learners can't understand numbers
- then it can never learn, day > Thursday and must report a clumsier theory, day=Friday OR day=Saturday OR day=Sunday
- e.g. Classification learners find connections from many independent attributes to
a single class Dependent Attribute
- While association rule learners find connections between sets of attributes.
- So an classification learners can't predict for sets of attributes.
Evaluation bias
In the above examples, we assessed our theories by mean performance
and size of the learned theory and the explain-ability of the learned
theory.
Different measures yielded different conclusions about the best
learner.
Even widely-used measures are surprisingly problematic.
- e.g. 10 hobos with no jobs enter Bill Gate's office.
- Q: What is their mean income?. A: ((10*0)+$24 billion)/11
=2.2 billion.
- Note that this number is uninformative regarding the social status of both the hobos and Bill Gates
Many ways to evaluate the performance
of a learner, but we'll need some theory first.
In the meantime:
- Well keep using means (simplest);
- But we'll know we can better
- And we'll be interested in how changing the
evaluation criteria changes our opinion about the the learner.
Many, many biases
From "Separate-and-conquer rule learning"
Ethics
Issue: different learners use different heuristics
- Consequently, they learn different theories.
So the same data can generate different theories, depending on the search bias (the heuristic search method) used to explore the space of descriptors*combinations
- i.e. there is no one best theory.
More generally, if you using induction on historical data to predict the future, then:
- There is no such thing as an "unbiased opinion".
- All theories are bias (but only some admit it).
- They must be biased else there is no way to decide what bits are most important and which bits can be ignored.
- So bias blinds us and, paradoxically, lets us see (predict) the future.
Anyway, you have an ethical dilemma
- You can generate different theories: which do you report?
- Who gets hurt or helped by the different theories?
Do the users of your theories fully appreciate the bias and limitations of the learning methods used to generate this theory>
- Can your users audit the bias?
- Are you use an explanation system where the learned theory can be browsed or just a performance system that only knows how to make conclusions and not how to explain them?
Also, should the users audit the bias?
- Should you be able to access and audit the theory that assigns your credit rating?
- Should we tell spammers what kinds of emails will get rejected?
- Should we tell bombers how we will screen airline passengers for explosives?
How to handle bias
Data miners: you have responsibilities to your users;
- Explore the space of possible theories: any common conclusions?
- Ensure reproducibility, documented the biases;
- And always remember : above all, do no harm.
Users of theories from data miners
- DO NOT be an passive consumer of other people's conclusions;
- ALWAYS be an active reviewer of ideas.
- Require:
- Access to the data used for training;
- Access to the software used for learning;
And if our users demand the right theory, we can't give it them (the Bias problem).
- Knowledge relativism? All ideas are valid?
- No!
- Sure, one data set supports many theories.
- But there are many many more theories that are unsupported.
So while no idea is right ...
- ... some things are useful (perform well on test data) ...
- ... and many many many more ideas are useless
]]>
