Data Mining - Motivation - Knowledge Engineering Group

Document technical information

Format pdf
Size 744.6 kB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Organizations

Places

Transcript

Data Mining - Motivation
"Computers have promised us a fountain of wisdom but
delivered a flood of data."
"It has been estimated that the amount of information in
the world doubles every 20 months."
(Frawley, Piatetsky-Shapiro, Matheus, 1992)
1
© J. Fürnkranz
Knowledge Discovery in Databases
(KDD)
Mining for nuggets of knowledge in mountains of Data.
2
© J. Fürnkranz
Definition
●
Data Mining is a non-trivial
process of identifying
●
●
●
●
valid
novel
potentially useful
ultimately
understandable
It employs techniques from
●
machine learning
●
statistics
●
databases
patterns in data.
(Fayyad et al. 1996)
Or maybe:
● Data Mining is torturing your database until it confesses.
(Mannila (?))
3
© J. Fürnkranz
Knowledge Discovery in Databases:
Key Steps
Key steps in the Knowledge Discovery cycle:
1. Data Cleaning: remove noise and incosistent data
2. Data Integration: combine multiple data sources
3. Data Selection: select the part of the data that are relevant for
the problem
4. Data Transformation: transform the data into a suitable format
(e.g., a single table, by summary or aggregation operations)
5. Data Mining: apply machine learning and machine discovery
techniques
6. Pattern Evaluation: evaluate whether the found patterns meet
the requirements (e.g., interestingness)
7. Knowledge Presentation: present the mined knowledge to the
user (e.g., visualization)
4
© J. Fürnkranz
Data Mining is a Process !
The steps are not followed linearly, but in an iterative process.
Source: http://alg.ncsa.uiuc.edu/tools/docs/d2k/manual/dataMining.html, after Fayyad, Piatetsky-Shapiro, Smyth, 1996
5
© J. Fürnkranz
Another Process Model
Source: http://www.crisp-dm.org/
6
© J. Fürnkranz
Pre-Processing
●
Databases are typically not made to support analysis with a
data mining algorithm

●
pre-processing of data is necessary
Pre-processing techniques:


Data Cleaning: remove inconsistencies from the data
Feature Engineering: find the right features/attribute set
●
●
●

Feature Subset Selection: select appropriate feature subsets
Feature Transformation: bring attributes into a suitable form
(e.g., discretization)
Feature Construction: construct derived features
Sampling:
●
select appropriate subsets of the data
7
© J. Fürnkranz
Unsupervised vs. Supervised
Pre-processing
●
Unsupervised

do not use information about the learning task
●
●
●
Supervised

use information about the learning task
●
●
only prior information (from knowledge about the data)
and information about the distribution of the training data
e.g.: look at relation of an attribute to class attribute
WARNING:
●
pre-processing may only use information from training data!



●
compute pre-processing model from training data
apply the model to training and test data
otherwise information from test data may be captured in the preprocessing step → biased evaluation
in particular: apply pre-processing to every fold in cross-validation
8
© J. Fürnkranz
Feature Subset Selection
●
●
Databases are typically not collected with data mining in
mind
Many features may be



●
Removing them can



●
irrelevant
uninteresting
redundant
increase efficiency
improve accuracy
prevent overfitting
Feature Subsect Selection techniques try to determine
appropriate features automatically
9
© J. Fürnkranz
Unsupervised FSS
●
Using domain knowledge

●
some features may be known to be irrelevant, uninteresting or
redundant
Random Sampling


select a random sample of the feature
may be appropriate in the case of many weakly relevant
features and/or in connection with ensemble methods
10
© J. Fürnkranz
Supervised FSS
●
Filter approaches:



compute some measure for estimating the ability to
discriminate between classes
typically measure feature weight and select the best n
features
problems
●
●
●
redundant features (correlated features will all have similar
weights)
dependent features (some features may only be important in
combination (e.g., XOR/parity problems).
Wrapper approaches


search through the space of all possible feature subsets
each search subset is tried with the learning algorithm
11
© J. Fürnkranz
Supervised FSS: Filters
●
foreach attribute A

W[A] = feature weight according to some measure of
discrimination
●
●
e.g., decision tree splitting criteria (entropy/information gain, giniindex, ...), attribute weighting criteria (Relief, ...), etc.
select the n features with highest W[A]
Basic idea:
● a good attribute should discriminate between the different
classes
● use a measure of discrimination to determine which attributes to
select
Disadvantage:
● quality of each attribute is measured in isolation
● some attributes may only be useful in combination with others
12
© J. Fürnkranz
RELIEF
(Kira & Rendell, ICML-92)
Basic idea:
● in a local neighborhood around an example R a good
attribute A should

allow to discriminate R from all examples of different classes
(the set of misses)
●

therefore the probability that the attribute has a different value for
R and a miss M should be high
have the same value for all examples of the same class as R
(the set of hits)
●
therefore the probability that the attribute has a different value for
R and a hit H should be low
→ try to estimate and maximize W [ A]=P a R ≠a M −P a R ≠a H 
where aX is the value of attribute A in example X
13
© J. Fürnkranz
RELIEF
(Kira & Rendell, ICML-92)
●
●
set all attribute weights W[A] = 0.0
for i = 1 to m (← user-settable parameter)
 select a random example R
 find
●
●

H: nearest neighbor of same class (near hit)
M: nearest neigbor of different class (near miss)
for each attribute A
●
d  A , H , R d  A , M , R
W [ A]  W [ A]−

m
m
where d(A,X,Y) is the distance in attribute A between
examples X and Y (normalized to [0,1]-range).
14
© J. Fürnkranz
FSS: Wrapper Approach
(John, Kohavi, Pfleger, ICML-94)
●
Wrapper Approach:



try a feature subset with the learner
improve it by modifying the feature sets based on the result
repeat
Figure by Kohavi & John
15
© J. Fürnkranz
FSS: Wrapper Approach
●
Forward selection:
1. start with empty feature set F
2. for each attribute A
a) F = F ∪ {A}
b) Estimate Accuracy of Learning algorithm on F
c) F = F \ {A}
3. F = F ∪ {attribute with highest estimated accuracy}
4. goto 2. unless estimated accuracy decreases significantly
●
Backward elimination:


●
start with full feature set F
try to remove attributes
Bi-directional search is also possible
16
© J. Fürnkranz
Example: Forward Search
Attrs: current set of attributes
Est: accuracy estimated by wrapper
Real: „real“ accuracy
Figure by John, Kohavi & Pfleger
17
© J. Fürnkranz
Properties
●
Advantage:



●
find feature set that is tailored to learning algorithm
considers combinations of features, not only individual feature
weights
can eliminate redundant features
(picks only as many as the algorithm needs)
Disadvantage:

very inefficient: many learning cycles necessary
18
© J. Fürnkranz
Comparison Wrapper / Relief
Note: RelieveD is a version of Relief that uses all examples instead of a random sample
●
on these datasets:

●
forward selection reduces attributes w/o error increase
in general, it may also reduce error
Figure by John, Kohavi & Pfleger
19
© J. Fürnkranz
Feature Transformation
●
bring features into a usable form
●
numerization


some algorithms can only use numeric data
nominal → binary
●

binary → numeric
●
●
a nominal attribute with n values is converted into n binary attributes
binary features may be viewed as special cases of numeric
attributes with two values
discretization

some algorithms can only use categorical data
●
transform numeric attributes into a number of (ordered) categorical
values
20
© J. Fürnkranz
Discretization
●
Supervised vs. Unsupervised:
 Unsupervised:
● only look at the distribution of values of the attribute
 Supervised:
● also consider the relation of attribute values to class values
●
Merging vs. Splitting:
 Merging (bottom-up discretization):
● Start with a set of intervals (e.g., each point is an interval)
and successively combine neighboring intervals
 Splitting (top-down discretization):
● Start with a single interval and successively split the interval
into sub-intervals
21
© J. Fürnkranz
Unsupervised Discretization
●
domain-dependent:
●
●
●
equal-width:
●
●
●
suitable discretizations are often known
age (0-18) →
baby (0-3), child (3-6), school child (6-10), teenager (11-18)
divide value range into a number of intervals with equal width
age (0,18) → (0-3, 4-7, 8-11, 12-15, 16-18)
equal-frequency:
●
●
●
divide value range into a number of intervals so that (approximately)
the same number of datapoints are in each interval
e.g., N = 5: each interval will contain 20% of the training data
good for non-uniform distributions (e.g., salary)
22
© J. Fürnkranz
Supervised Discretization:
Chi-Merge (Kerber, AAAI-92)
Basic Idea: merge neighboring intervals if the class information is
independent of the interval an example belongs to
●
initialization:


●
sort examples according to feature value
construct one interval for each value
interval merging:

compute 2 value for each pair of adjacent intervals
n
c
2
2
c
C
 A −E 
N i=∑ A ij C j = ∑ Aij
E ij = N i j
2=∑ ∑ ij ij
N
E ij
j =1
i=1
i =1 j=1
intervals
Aij = number of examples in i-th interval that are of class j
Eij = expected number of examples in i-th interval that are of class j
= examples in i-th interval Ni × fraction Cj/N of (all) examples of class j
●

merge those with lowest 2 value

when the 2 values of all pairs exceed a significance threshold
stop
23
© J. Fürnkranz
Supervised Discretization:
Entropy-Split (Fayyad & Irani, IJCAI-93)
Basic Idea: grow a decision tree using a single numeric attribute and
use the value ranges in the leaves as ordinal values
● initialization:



initialize intervals with a single interval covering all examples S
sort all examples according to the attribute value
initialize the set of possible split points


●
simple: all values
more efficient: only between class changes in sorted list
interval splitting:

select split point with the minimum weighted entropy


∣S AT∣
∣S A≥T∣
T max =arg min
Entropy S A T 
Entropy  S A≥T 
∣S∣
∣S∣
T

●
recursively apply Entropy-Split to S AT
stop



max
and S A≥T
max
when a given number of splits is achieved
or when splitting would yield too small intervals
or MDL-based stopping criterion (Fayyad & Irani, 1993)
24
© J. Fürnkranz
Example
Temperature
64
65
68
70
71
72
72
75
80
81
83
85
Play
Yes
No
Yes Yes Yes
No
No
Yes Yes Yes
No
Yes Yes
No
Slide taken from Witten & Frank
69
25
75
© J. Fürnkranz
Unsupervised Feature Construction
●
based on domain knowledge

●
weight kg 
BMI =
height  m2
Example: Body Mass Index
automatic

Examples:
●
kernel functions


●
principal components analysis

●
may be viewed as feature construction modules
→ support vector machines
transforms an n-dimensional space into a lower-dimensional subspace
w/o losing much information
GLEM:


uses an Apriori -like algorithms to compute all conjunctive combinations
of basic features that occur at least n times
application to constructing evaluation functions for game Othello
26
© J. Fürnkranz
Supervised Feature Construction
●
use the class information to construct features that help to
solve the classification problem
●
Examples:

Wrapper approach
●
●
●

use rule or decision tree learning algorithm
observe frequently co-occurring features or feature values
encode them as separate features
Neural Network
●
nodes in hidden layers may be interpreted as constructed features
27
© J. Fürnkranz
Scalability
●
databases are often too big for machine learning algorithms


●
ML algorithms require frequent counting operations and multidimensional access to data
only feasible for data that can be held in main memory
two strategies to make DM algorithms scalable


design algorithms that are explicitly targetted towards
minimizing the number of database operations (e.g., Apriori)
use sampling to work on subsets of the data
28
© J. Fürnkranz
Sampling
●
Random Sampling

●
Select a random subset of the data
Progressive Sampling


start with a small sample
increase sample size
●
●

●
arithmetic: increase sample size by fixed number of examples
geometric: multiply size with a fixed number (e.g., double size)
stop when convergence is detected
Sequential sampling

rule out solution candidates based on significant differences
29
© J. Fürnkranz
Windowing
●
Idea:

●
focus the learner on the parts of the search space that are not
yet correctly covered
Algorithm:
1. Initialize the window with a random subsample of the
available data
2. Learn a theory from the current window
3. If the learned theory correctly classifies all examples
(including those outside of the window), return the theory
4. Add some mis-classified examples to the window and goto 2.
●
Properties:


may learn a good theory from a subset of the data
problems with noisy data
30
© J. Fürnkranz
Weiterführende Lehrveranstaltungen
●
Computer Poker Challenge

besteht aus:
●
●


●
●
●
●
●
●
●
Seminar KE und ML in Spielen
Praktikum aus Künstliche Intelligenz
ACHTUNG: Beginn in der 2. März-Woche!
Teilnahme an einem Internationalen Computer Poker-Wettbewerb
Maschinelles Lernen: Statistische Verfahren 1 + 2 (Roth/Schiele)
Neural Networks (Stibor)
Einführung in die Künstliche Intelligenz (Fürnkranz, 3+1)
Web Mining (erst wieder SS09)
Data und Knowledge Engineering (A. Buchmann, Fürnkranz)
Seminare (wechselnde Themen, z.B. Mining in Graphs).
Hiwis gesucht!
31
© J. Fürnkranz
×

Report this document