Transcript Metody Inteligencji Obliczeniowej

Computational intelligence methods for
information understanding
and information management
Włodzisław Duch
Department of Informatics
Nicolaus Copernicus University, Torun, Poland
&
School of Computer Engineering,
Nanyang Technological University, Singapore
IMS2005, Kunming, China
Plan
What is this about ?
•
•
•
•
How to discover knowledge in data;
how to create comprehensible models of data;
how to evaluate new data;
how to understand what computational intelligence (CI)
methods really do.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
AI, CI & Data Mining
Forms of useful knowledge
Integration of different methods in GhostMiner
Exploration & Visualization
Rule-based data analysis
Neurofuzzy models
Neural models, understanding what they do
Similarity-based models, prototype rules
Case studies
From data to expert system
AI, CI & DM
Artificial Intelligence: symbolic models of knowledge.
• Higher-level cognition: reasoning, problem solving,
planning, heuristic search for solutions.
• Machine learning, inductive, rule-based methods.
• Technology: expert systems.
Computational Intelligence, Soft Computing:
methods inspired by many sources:
• biology – evolutionary, immune, neural computing
• statistics, patter recognition
• probability – Bayesian networks
• logic – fuzzy, rough …
Perception, object recognition.
Data Mining, Knowledge Discovery in Databases.
• discovery of interesting rules, knowledge => info understanding.
• building predictive data models => part of info management.
Forms of useful knowledge
AI/Machine Learning camp:
Neural nets are black boxes.
Unacceptable! Symbolic rules forever.
But ... knowledge accessible to humans is in:
• symbols and rules;
• similarity to prototypes, structures, known cases;
• images, visual representations.
What type of explanation is satisfactory?
Interesting question for cognitive scientists but ...
in different fields answers are different!
Forms of knowledge
• Humans remember examples of each category and
refer to such examples – as similarity-based, case
based or nearest-neighbors methods do.
• Humans create prototypes out of many examples –
as Gaussian classifiers, RBF networks, or neurofuzzy
systems modeling probability densities do.
• Logical rules are the highest form of summarization of
simple forms of knowledge;
• Bayesian networks present complex relationships.
3 types of explanation presented here:
• logic-based: symbols and rules;
• exemplar-based: prototypes and similarity of structures;
• visualization-based: maps, diagrams, relations ...
GhostMiner Philosophy
GhostMiner tools for data mining & knowledge discovery,
from our lab + Fujitsu: http://www.fqspl.com.pl/ghostminer/
• Separate the process of model building (hackers) and knowledge
discovery, from model use (lamers) =>
GhostMiner Developer & GhostMiner Analyzer (ver. 3.0 & newer)
• There is no free lunch – provide different type of tools for
knowledge discovery.
Decision tree, neural, neurofuzzy, similarity-based, SVM,
committees.
• Provide tools for visualization of data.
• Support the process of knowledge discovery/model building
and evaluating, organizing it into projects.
Wine data example
Chemical analysis of wine from grapes grown in the same region in
Italy, but derived from three different cultivars.
Task: recognize the source of wine sample.
13 quantities measured, all features are continuous:
•
•
•
•
•
•
alcohol content
ash content
magnesium content
flavanoids content
proanthocyanins phenols content
OD280/D315 of diluted wines
•
•
•
•
•
•
•
malic acid content
alkalinity of ash
total phenols content
nonanthocyanins phenols
content
color intensity
hue
proline.
Wine sample => 13 numerical quantities => feature space rep.
Complex structures: no feature space, only Similarity(A,B) known.
Exploration and visualization
General info about the data
Exploration: data
Inspect the data
Exploration: data statistics
Distribution of feature values
Proline has very large values, most methods will benefit from data
standardization before further processing.
Exploration: data standardized
Standardized data: unit standard deviation, about 2/3 of all data should
fall within [mean-std,mean+std]
Other options: normalize to [-1,+1],
or normalize rejecting p% of extreme values.
Exploration: 1D histograms
Distribution of feature values in classes
Some features are more useful than the others.
Exploration: 1D/3D histograms
Distribution of feature values in classes, 3D
Exploration: 2D projections
Projections on selected 2D
Projections on selected 2D
Visualize data
Hard to imagine relations in more than 3D.
Linear methods: PCA, FDA, PP ... use input combinations.
SOM mappings: popular for visualization, but rather inaccurate, there
is no measure of distortions.
Measure of topographical distortions:
map all Xi points from Rn to xi points in Rm, m < n, and ask:
how well are Rij = D(Xi, Xj) distances reproduced
by distances rij = d(xi,xj) ?
Use m = 2 for visualization,
use higher m for dimensionality reduction.
Visualize data: MDS
Multidimensional scaling: invented in psychometry by Torgerson
(1952), re-invented by Sammon (1969) and myself (1994) …
Minimize measure of topographical distortions moving the x
coordinates.
1
S1  x  
2
R
 ij
R
i j
ij
 rij  x  
2
MDS
i j
1  r  x  
1
S2  x  
 Rij

1
S3  x  
 Rij
 1  r  x 
i j
i j
i j
i j
2
ij
Sammon
Rij
ij
Rij 
2
MDS, more local
Visualize data: Wine
3 clusters are clearly distinguished, 2D is fine.
The green outlier can be identified easily.
Decision trees
Simplest things first:
use decision tree to find logical rules.
Test single attribute, find good point to split the data, separating
vectors from different classes.
DT advantages: fast, simple, easy to understand, easy to program,
many good algorithms.
Decision borders
Univariate trees:
test the value of a single attribute x < a.
Multivariate trees:
test on combinations of attributes.
Result:
feature space is divided into
large hyperrectangular areas with
decision borders perpendicular to
axes.
Splitting criteria
Most popular: information gain, used in C4.5 and other trees.
Which attribute is better?
Which should be at the top
of the tree?
Look at entropy reduction,
or information gain index.
E(S )  Pt lg2 Pt  Pf lg2 Pf
S
S
G( S , A)  E( S ) 
E ( S ) 
E ( S )
S
S
CART trees use Gini index
of node purity:
C
Gini  node   1   Pi 2
i 1
Non-Bayesian selection
Bayesian MAP selection: choose max a posteriori P(C|X)
P(C,A1)
P(C,A2)
A=0
0.0100
0.0900
0.0300
0.1300
A=1
0.4900
0.4100
0.4700
0.3700
P(C0)=0.5
P(C1)=0.5
P(C|X)=P(C,X)/P(X)
MAP is here equivalent to a majority classifier (MC): given A=x, choose
maxC P(C,A=x)
MC(A1)=0.58, S+=0.98, S-=0.18, AUC=0.58, MI= 0.058
MC(A2)=0.60, S+=0.94, S-=0.26, AUC=0.60, MI= 0.057
MC(A1)<MC(A2), AUC(A1)<AUC(A2), but MI(A1)>MI(A2) !
Problem: for binary features non-optimal decisions are taken!
But estimation of P(C,A) for non-binary features is not reliable.
SSV decision tree
Separability Split Value tree:
based on the separability criterion.
Define subsets of data D using a binary test f(X,s) to split the data
into left and right subset D = LS  RS.
LS  s, f , D   X  D : f ( X, s )  T
RS  s, f , D   D  LS  s, f , D 
SSV criterion: separate as many pairs of vectors from different classes
as possible; minimize the number of separated from the same class.
SSV ( s)  2 LS  s, f , D  Dc  RS  s, f , D 
cC

 D  Dc 
 min LS  s, f , D  Dc , RS  s, f , D  Dc
cC

SSV – complex tree
Trees may always learn to achieve 100% accuracy.
Very few vectors are left in the leaves – splits are not reliable and will
overfit the data!
SSV – simplest tree
Pruning finds the nodes that should be removed to increase
generalization – accuracy on unseen data.
Trees with 7 nodes left: 15 errors/178 vectors.
SSV – logical rules
Trees may be converted to logical rules.
Simplest tree leads to 4 logical rules:
1.
2.
3.
4.
if proline > 719 and flavanoids > 2.3 then class 1
if proline < 719 and OD280 > 2.115 then class 2
if proline > 719 and flavanoids < 2.3 then class 3
if proline < 719 and OD280 < 2.115 then class 3
How accurate are such rules?
Not 15/178 errors, or 91.5% accuracy!
Run 10-fold CV and average the results.
85±10%?
Run 10X and average
85±10%±2%? Run again ...
SSV – optimal trees/rules
Optimal: estimate how well rules will generalize.
Use stratified crossvalidation for training;
use beam search for better results.
1. if OD280/D315 > 2.505 and proline > 726.5 then class 1
2. if OD280/D315 < 2.505 and hue > 0.875 and malic-acid < 2.82
3.
4.
5.
then class 2
if OD280/D315 > 2.505 and proline < 726.5 then class 2
if OD280/D315 < 2.505 and hue > 0.875 and malic-acid > 2.82
then class 3
if OD280/D315 < 2.505 and hue < 0.875 then class 3
Note 6/178 errors, or 91.5% accuracy!
Run 10-fold CV: results are 85±10%? Run 10X!
Logical rules
Crisp logic rules: for continuous x use linguistic variables (predicate
functions).
sk(x)  True [Xk x  X'k], for example:
small(x) = True{x|x < 1}
medium(x) = True{x|x  [1,2]}
large(x)
= True{x|x > 2}
Linguistic variables are used in crisp (prepositional, Boolean) logic
rules:
IF small-height(X) AND has-hat(X) AND has-beard(X)
THEN (X is a Brownie)
ELSE IF ... ELSE ...
Crisp logic decisions
Crisp logic is based on rectangular membership
functions:
True/False values jump from 0 to 1.
Step functions are used for partitioning of the
feature space.
Very simple hyper-rectangular decision borders.
Expressive power of crisp logical rules is very
limited!
Similarity cannot be captured by rules.
Logical rules - advantages
Logical rules, if simple enough, are preferable.
• Rules may expose limitations of black box solutions.
• Only relevant features are used in rules.
• Rules may sometimes be more accurate than NN and other CI
methods.
• Overfitting is easy to control, rules usually have small number
of parameters.
• Rules forever !?
A logical rule about logical rules is:
IF the number of rules is relatively small
AND the accuracy is sufficiently high.
THEN rules may be an optimal choice.
Logical rules - limitations
Logical rules are preferred but ...
• Only one class is predicted p(Ci|X,M) = 0 or 1; such
black-and-white picture may be inappropriate in many applications.
• Discontinuous cost function allow only non-gradient optimization
methods, more expensive.
• Sets of rules are unstable: small change in the dataset leads to a
large change in structure of sets of rules.
• Reliable crisp rules may reject some cases as unclassified.
• Interpretation of crisp rules may be misleading.
• Fuzzy rules remove some limitations, but are not so
comprehensible.
Fuzzy inputs vs. fuzzy rules
Crisp rule Ra(x) = Q(xa) applied to uncertain input with uniform
input uncertainty U(x;Dx)=1 in [xDx, xDx] and zero outside is true
to the degree given by a semi-linear function S(x;Dx):
Input uncertainty and the probability
that Ra(x) rule is true.
For other input uncertainties similar
relations hold!
For example, triangular U(x):
leads to sigmoidal S(x) function.
For more input conditions rules are
true to the degree described by soft
trapezoidal functions, difference of
two sigmoidal functions.
Crisp rules + input uncertainty  fuzzy rules for crisp inputs = MLP !
From rules to probabilities
Data has been measured with unknown error.
Assume Gaussian distribution:
x  Gx  G( y; x, sx )
x – fuzzy number with Gaussian membership function.
A set of logical rules R is used for fuzzy input vectors:
Monte Carlo simulations for arbitrary system => p(Ci|X)
Analytical evaluation p(C|X) is based on cumulant function:
 a  x 
1
  a  x    G  y; x, sx  dy  1  erf 
      (a  x) 
2 

 sx 2  
a
  2.4 / 2sx
Error function is identical to
logistic f. < 0.02
Rules - choices
Simplicity vs. accuracy.
Confidence vs. rejection rate.
 p
p  true | predicted   
 p
p
p
p r  p

p r  p
p is a hit; p false alarm; p is a miss.
Accuracy (overall)
Error rate
Rejection rate
Sensitivity
Specificity
A(M) = p+ p
L(M) = p+ p
R(M)=p+r+pr= 1L(M)A(M)
S+(M)= p+|+ = p++ /p+
S(M)= p| = p /p
Rules – error functions
The overall accuracy is equal to a combination of sensitivity and
selectivity weighted by the a priori probabilities:
A(M) = pS(M)+pS(M)
Optimization of rules for the C+ class;
large g means no errors but high rejection rate.
E(M;g)= gL(M)A(M)= g (p+p)  (p+p)
minM E(M;g)  minM {(1+g)L(M)+R(M)}
Optimization with different costs of errors
minM E(M;a) = minM {p+ a p} =
minM {p1S(M))  pr(M) + a [p1S(M))  pr(M)]}
ROC curves
ROC curves display S+(1S) for different models (classifiers) or
different confidence thresholds:
Ideal classifier: below some
threshold S+ = 1 (all positive
cases recognized) for 1-S= 0
(no false alarms) .
Useless classifier (blue): same
number of true positives as
false alarms for any threshold.
Reasonable classifier (red):
no errors until some threshold that allows for recognition of 0.5 positive
cases, no errors if 1-S > 0.6; slowly rising errors in between.
Good measure of quality: high AUC, Area Under ROC Curve.
AUC = 0.5 is random guessing, AUC = 1 is perfect prediction.
Gaussian fuzzification of crisp rules
Very important case: Gaussian input uncertainty.
Rule Ra(x) = {xa} is fulfilled by Gx with probability:

p  Ra (Gx )  T    G  y; x, sx  dy     ( x  a) 
a
Error function is approximated by logistic function;
assuming error distribution (x)1 x)),
for s2=1.7 approximates Gauss < 3.5%
Rule Rab(x) = {b> x a} is fulfilled by Gx with
probability:
b
p  Rab (Gx )  T    G  y; x, sx  dy     ( x  a)      ( x  b) 
a
Soft trapezoids and NN
The difference between two sigmoids makes a soft trapezoidal
membership functions.
Conclusion:
fuzzy logic with soft trapezoidal membership functions
(x)  (x-b)  to a crisp logic + Gaussian uncertainty of inputs.
Optimization of rules
Fuzzy: large receptive fields, rough estimations.
Gx – uncertainty of inputs, small receptive fields.
Minimization of the number of errors – difficult, non-gradient, but
now Monte Carlo or analytical p(C|X;M).
2
1
E { X }; R, sx     p  Ci | X ; M     C ( X ), Ci  
2 X i
•
Gradient optimization works for large number of parameters.
•
Parameters sx are known for some features, use them as
optimization parameters for others!
Probabilities instead of 0/1 rule outcomes.
Vectors that were not classified by crisp rules have now non-zero
probabilities.
•
•
Mushrooms
The Mushroom Guide: no simple rule for mushrooms; no rule like:
‘leaflets three, let it be’ for Poisonous Oak and Ivy.
8124 cases, 51.8% are edible, the rest non-edible.
22 symbolic attributes, up to 12 values each, equivalent to 118 logical
features, or 2118=31035 possible input vectors.
Cap Shape: bell, conical, convex, flat, knobbed, sunken
Cap Surface: fibrous, grooves, scaly, smooth
Cap Color: brown, buff, cinnamon, gray, green, pink, red, white, yellow
Bruises: bruised, not bruised
Odor: almond, anise, creosote, fishy, foul, musty, none, pungent, spicy
Spore print color: black, brown, buff, chocolate, green ...yellow.
Gill Attachment: attached, descending, free, notched
Gill Spacing: close, crowded, distant
Population: abundant, clustered, numerous, scattered, several, solitary
Habitat: grasses, leaves, meadows, paths, urban, waste, woods ...
Mushrooms data
Mushroom-3 is: edible, bell, smooth, white, bruises, almond, free,
close, broad, white, enlarging, club, smooth, smooth, white, white,
partial, white, one, pendant, black, scattered, meadows
Mushroom-4 is: poisonous, convex, smooth, white, bruises, pungent,
free, close, narrow, white, enlarging, equal, smooth, smooth, white,
white, partial, white, one, pendant, black, scattered, urban
Mushroom-5 is: poisonous, convex, smooth, white, bruises, pungent,
free, close, narrow, pink, enlarging, equal, smooth, smooth, white,
white, partial, white, one, pendant, black, several, urban
Mushroom-8000 is: poisonous, convex, smooth, white, bruises,
pungent, free, close, narrow, pink, enlarging, equal, smooth, smooth,
white, white, partial, white, one, pendant, brown, scattered, urban.
What knowledge is hidden in this data?
Mushroom rule
To eat or not to eat, this is the question!
Well, not any more ...
Safe rule for edible mushrooms found by the SSV decision tree and
MLP2LN network:
IF
odor is none or almond or anise
AND spore_print_color is not green
THEN mushroom is edible
This rule makes only 48 errors, is 99.41% correct.
This is why animals have such a good sense of smell !
What does it tell us about odor receptors in animal noses ?
This rule has been quoted by > 50 encyclopedias so far!
Mushrooms rules
To eat or not to eat, this is the question! Well, not any more ...
A mushroom is poisonous if:
R1) odor =  (almond  anise  none); 120 errors, 98.52%
R2) spore-print-color = green
48 errors, 99.41%
R3) odor = none  stalk-surface-below-ring = scaly
 stalk-color-above-ring =  brown 8 errors, 99.90%
R4) habitat = leaves  cap-color = white
no errors!
R1 + R2 are quite stable, found even with 10% of data;
R3 and R4 may be replaced by other rules, ex:
R'3): gill-size=narrow  stalk-surface-above-ring=(silky  scaly)
R'4): gill-size=narrow  population=clustered
Only 5 of 22 attributes used! Simplest possible rules?
100% in CV tests - structure of this data is completely clear.
Recurrence of breast cancer
Institute of Oncology, University Medical Center, Ljubljana.
286 cases, 201 no (70.3%), 85 recurrence cases (29.7%)
9 symbolic features: age (9 bins), tumor-size (12 bins), nodes involved
(13 bins), degree-malignant (1,2,3), area, radiation, menopause, nodecaps.
no-recurrence,40-49,premeno,25-29,0-2,?,2, left, right_low, yes
Many systems were used on this data with 65-78% accuracy reported.
Best single rule:
IF (nodes-involved  [0,2]  degree-malignant = 3 THEN recurrence
ELSE no-recurrence
77% accuracy, only trivial knowledge in the data: highly malignant
cancer involving many nodes is likely to strike back.
Neurofuzzy system
Sometimes crisp logical rules may fail.
Fuzzy: mx0,1 (no/yes) replaced by a degree mx[0,1].
Triangular, trapezoidal, Gaussian or other membership f.
M.f-s in many
dimensions:
Feature Space Mapping (FSM) neurofuzzy system (1995).
Neural adaptation, estimation of probability density distribution (PDF)
using single hidden layer network (RBF-like), with nodes realizing
separable functions (like in Naive Bayes):
G  X ; P    Gi  X i ; Pi 
i 1
FSM
Initialize using clusterization or decision trees.
Triangular & Gaussian f. for fuzzy rules.
Rectangular functions for crisp rules.
Between 9-14 rules with triangular membership functions are
created; accuracy in 10xCV tests about 96±4.5%
Similar results obtained with Gaussian functions.
Rectangular functions: simple rules are created, many nearly
equivalent descriptions of this data exist.
If proline > 929.5 then class 1 (48 cases, 45 correct
+ 2 recovered by other rules).
If color < 3.79285 then class 2 (63 cases, 60 correct)
Interesting rules, but overall accuracy is only 88±9%
Prototype-based rules
C-rules (Crisp), are a special case of F-rules (fuzzy rules).
F-rules (fuzzy rules) are a special case of P-rules (Prototype).
P-rules have the form:
IF P = arg minR D(X,R) THAN Class(X)=Class(P)
D(X,R) is a dissimilarity (distance) function, determining decision
borders around prototype P.
P-rules are easy to interpret! F-rules may always be presented as
P-rules, so this is an alternative to neurofuzzy systems.
IF
X=You are most similar to the P=Superman
THAN You are in the Super-league.
IF
X=You are most similar to the P=Weakling
THAN You are in the Failed-league.
“Similar” may involve different features or D(X,P).
P-rules
Euclidean distance leads to a Gaussian fuzzy membership functions +
product as T-norm.
D  X, P    d  X i , Pi    Wi  X i  Pi 
i
mP  X  e
 D X,P 
2
i
e

 d  X i , Pi 
i
 e
Wi  X i  Pi 
i
2
  mi  X i , Pi 
i
Manhattan function => m(X;P)=exp{|X-P|}
Various distance functions lead to different MF.
Ex. data-dependent distance functions, for symbolic data:


DVDM  X, Y     p  C j | X i   p  C j | Yi  
i  j



DPDF  X, Y     p  X i | C j   p  C j | Yi  
i  j

Promoters
DNA strings, 57 aminoacids, 53 + and 53 - samples
tactagcaatacgcttgcgttcggtggttaagtatgtataatgcgcgggcttgtcgt
Euclidean distance, symbolic
s =a, c, t, g replaced by x=1, 2, 3, 4
PDF distance, symbolic
s=a, c, t, g replaced by p(s|+)
P-rules
New distance functions from info theory => interesting MF.
MF => new distance function, with local D(X,R) for each cluster.
Crisp logic rules: use Chebyshev distance (L norm):
DCh(X,P) = ||XP|| = maxi Wi |XiPi|
DCh(X,P) = const => rectangular contours.
Chebyshev distance with thresholds P
IF DCh(X,P)  P THEN C(X)=C(P)
is equivalent to a conjunctive crisp rule
IF
X1[P1P/W1,P1P/W1] …XN [PN P/WN, PNP/WN]
THEN C(X)=C(P)
Decision borders
D(P,X)=const and decision borders D(P,X)=D(Q,X).
Euclidean distance from 3
prototypes, one per class.
Minkovski a=20 distance from
3 prototypes.
P-rules for Wine
Manhattan distance:
6 prototypes kept,
4 errors, f2 removed
Chebyshev distance:
15 prototypes kept, 5 errors, f2,
f8, f10 removed
Euclidean distance:
11 prototypes kept,
7 errors
Many other solutions.
Scatterograms for hypothyroid
Shows images of training vectors mapped by
neural network; for more than 2 classes
either linear projections, or several 2D
scatterograms, or parallel coordinates.
Good for:
•
analysis of the learning process;
•
comparison of network solutions;
•
stability of the network;
•
analysis of the effects of regularization;
•
evaluation of confidence by perturbation of
the query vector.
...
Details: W. Duch, IJCNN 2003
Neural networks
• MLP – Multilayer Perceptrons, most popular NN models.
Use soft hyperplanes for discrimination.
Results are difficult to interpret, complex decision borders.
Prediction, approximation: infinite number of classes.
• RBF – Radial Basis Functions.
RBF with Gaussian functions are equivalent to fuzzy systems with
Gaussian membership functions, but …
No feature selection => complex rules.
Other radial functions => not separable!
Use separable functions, not radial => FSM.
• Many methods to convert MLP NN to logical rules.
What NN really do?
• Common opinion: NN are black boxes.
NN provide complex mappings that may involve various kinks
and discontinuities, but NN have the power!
•
Solution 1 (common): extract rules approximating NN mapings.
•
Solution 2 (new): visualize neural mapping.
RBF network for fuzzy XOR, using 4
Gaussian nodes:
rows for =1/7,1 and 7
left column: scatterogram of the hidden
node activity in 4D.
middle columns: parallel coordinate view
right column: output view (2D)
Wine example
• MLP with 2 hidden nodes, SCG training, regularization a=0.5
•
After 3 iterations: output, parallel, hidden.
After convergence + with noise var=0.05 added
Rules from MLPs
Why is it difficult?
Multi-layer perceptron (MLP) networks: stack many perceptron units,
performing threshold logic:
M-of-N rule: IF (M conditions of N are true) THEN ...
Problem: for N inputs number of subsets is 2N.
Exponentially growing number of possible conjunctive rules.
MLP2LN
Converts MLP neural networks into a network performing logical
operations (LN).
Input
layer
Output:
one node
per class.
Aggregation:
better features
Linguistic units:
windows, filters
Rule units:
threshold logic
MLP2LN training
Constructive algorithm: add as many nodes as needed.
Optimize cost function:
1
p
p 2
E ( W )   ( F ( Xi ; W)  ti )  minimize errors +
2 p i
1
2
2
W
 ij 
enforce zero connections +
i j
2
W

2
i j
2
ij
(Wij  1) (Wij  1)
2
2
leave only +1 and -1 weights
makes interpretation easy.
L-units
Create linguistic variables.
Numerical representation for R-nodes
Vsk=(1,1,1,...) for sk=low
Vsk=(1,1,1,...) for sk=normal
L( X )  S1 W1x  b  S2 W2 x  b '
Product of bi-central functions is
logical rule, used by IncNet NN.
L-units: 2 thresholds as adaptive
parameters;
logistic (x), or tanh(x)[1, 1]
Soft trapezoidal functions change
into rectangular filters (Parzen
windows).
4 types, depending on signs Si.
Iris example
Network after training:
iris setosa: q=1
(0,0,0;0,0,0;+1,0,0;+1,0,0)
iris versicolor: q=2
(0,0,0;0,0,0;0,+1,0;0,+1,0)
iris virginica: q=1
(0,0,0;0,0,0;0,0,+1;0,0,+1)
Rules:
If (x3=s  x4=s) setosa
If (x3=mx4=m) versicolor
If (x3=l  x4=l) virginica
3 errors only (98%).
Learning dynamics
Decision regions shown every 200 training epochs in x3, x4 coordinates;
borders are optimally placed with wide margins.
Thyroid screening
Clinical
findings
Garavan Institute, Sydney,
Australia
15 binary, 6 continuous
Training: 93+191+3488
Validate: 73+177+3178

Determine important
clinical factors

Calculate prob. of
each diagnosis.
Age
sex
…
…
Hidden
units
Final
diagnoses
Normal
Hypothyroid
TSH
T4U
T3
TT4
TBG
Hyperthyroid
Thyroid – some results.
Accuracy of diagnoses obtained with several systems – rules are
accurate.
Method
Rules/Features Training %
Test %
MLP2LN optimized
4/6
99.9
99.36
CART/SSV Decision Trees
3/5
99.8
99.33
Best Backprop MLP
-/21
100
98.5
Naïve Bayes
-/-
97.0
96.1
k-nearest neighbors
-/-
-
93.8
Thyroid – output visualization.
2D – plot scatterogram of the vectors transformed by the network.
3D – display it inside the cube, use perspective.
ND – make linear projection of network outputs on the polygon
vertices
Feature selection
Feature Extraction, Foundations and Applications. Eds. Guyon, I, Gunn
S, Nikravesh M, and Zadeh L, Springer Verlag, Heidelberg, 2005.
NIPS 2003 competition: databases with 10.000-100.000 features,
data obtained from text analysis and bioinformatics problems.
Without feature selection or extraction analysis is not possible.
Our InfoSel++ library (in C++) implements >20 methods for feature
ranking & selection based on: mutual information, information gain,
symmetrical uncertainty coefficient, asymmetric dependency
coefficients, Mantaras distance using transinformation matrix, distances
between probability distributions (Kolmogorov-Smirnov, KullbackLeibler), Markov blanket, Pearson’s correlations coefficient (CC),
Bayesian accuracy, etc.
Such indices depend very strongly on discretization procedures; care
has been taken to use unbiased probability and entropy estimators and
to find appropriate discretization of continuous feature values.
Breast cancer data
Walker AJ, Cross S.S & Harrison R.F, Lancet 354, 1518-1522, 1999
Fine-needle aspirates of breast lumps were performed.
Age plus 10 observations made by experienced pathologist were
collected for each breast cancer case; the final determination whether
the cancer was malignant or benign was confirmed by biopsy.
SVM with optimized Gaussian kernel gives in 10xCV tests very good
accuracy 95,49±0,29%, but gives no understanding of the important
clinical factors.
PCA feature extraction
Using first principal component allows for 93.7% accuracy, but second
dimension is useless and linear mixture of features makes
interpretation difficult.
BC visualization
MDS shows interesting clusters with similar clinical profile, differing
mostly by the age of patients; clusters correspond to different subtypes
of benign cancer types, while malignant types are more diverse (left).
SVM support vectors may be used for unsupervised clustering in the
PCA space (right).
BC logical rules
Feature Extraction, Foundations and Applications. Eds. Guyon, I, Gunn
S, Nikravesh M, and Zadeh L, Springer Verlag, Heidelberg, 2005.
SSV achieves 95.4,% accuracy, 90% sensitivity, and specificity of 98%.
The presence of necrotic epithelial cells is the most important factor. The
decision function can be presented in the form of the following logical
rules:
1. if F#9 > 0 and F#7 > 0 then class 0
2. if F#9 > 0 and F#7 < 0 and F#1 > 50.5 then class 0
3. if F#9 < 0 and F#1 > 56.5 and F#3 > 0 then class 0
4. else class 1
A forest of trees generated for the breast cancer data reveals another
interesting classification rules:
if F#8 > 0 and F#7 > 0 then class 0 else class 1.
Accuracy is lower, 90.6%, but the rule has 100% specificity.
BC classification results
4 prototypes give 96.4% accuracy, sensitivity 92.3%, specificity 98.5%. 2
prototype solutions have accuracy around 94%.
11 features may be reduced to 6 without decrease of accuracy.
Taking first principal component + 2 prototypes same accuracy is found.
Average accuracy
Standard deviation
kNN
95,87
0,27
SVM
95,49
0,29
NRBF
94,91
0,29
SSV Tree
94,41
0,46
FSM
92,74
1,59
Naive Bayes
88,24
0,45
Classifier
Summary of black box classification results.
Psychometry
Use CI to find knowledge, create Expert System.
MMPI (Minnesota Multiphasic Personality Inventory) psychometric
test.
Printed forms are scanned or computerized version of the test is
used.
•
Raw data: 550 questions, ex:
I am getting tired quickly: Yes - Don’t know - No
•
Results are combined into 10 clinical scales and 4 validity scales
using fixed coefficients.
•
Each scale measures tendencies towards hypochondria,
schizophrenia, psychopathic deviations, depression, hysteria,
paranoia etc.
Psychometry: goal
• There is no simple correlation between single values and final
diagnosis.
• Results are displayed in form of a histogram, called ‘a
psychogram’. Interpretation depends on the experience and skill
of an expert, takes into account correlations between peaks.
Goal: an expert system providing evaluation and interpretation of MMPI
tests at an expert level.
Problem: experts agree only about 70% of the time;
alternative diagnosis and personality changes over time are
important.
Psychometric data
1600 cases for woman, same number for men.
27 classes:
norm, psychopathic, schizophrenia, paranoia, neurosis, mania,
simulation, alcoholism, drug addiction, criminal tendencies,
abnormal behavior due to ...
Extraction of logical rules: 14 scales = features.
Define linguistic variables and use FSM, MLP2LN, SSV - giving
about 2-3 rules/class.
Psychometric results
Method
Data
N. rules
Accuracy
+Gx%
C 4.5
♀
55
93.0
93.7
♂
61
92.5
93.1
♀
69
95.4
97.6
♂
98
95.9
96.9
FSM
10-CV for FSM is 82-85%, for C4.5 is 79-84%.
Input uncertainty +Gx around 1.5% (best ROC) improves FSM results
to 90-92%.
Psychometric Expert
Probabilities for different classes.
For greater uncertainties more classes are
predicted.
Fitting the rules to the conditions:
typically 3-5 conditions per rule, Gaussian
distributions around measured values that fall
into the rule interval are shown in green.
Verbal interpretation of each case, rule and
scale dependent.
Visualization
Probability of classes versus input
uncertainty.
Detailed input probabilities around the
measured values vs. change in the single
scale; changes over time define ‘patients
trajectory’.
Interactive multidimensional scaling:
zooming on the new case to inspect its
similarity to other cases.
Summary
Computational intelligence methods: neural, decision trees, similaritybased & other, help to understand the data.
Understanding data: achieved by rules, prototypes, visualization.
Small is beautiful => simple is the best!
Simplest possible, but not simpler - regularization of models;
accurate but not too accurate - handling of uncertainty;
high confidence, but not paranoid - rejecting some cases.
• Challenges:
hierarchical systems – higher-order rules, missing information;
discovery of theories rather than data models;
reasoning in complex domains/objects;
integration with image/signal analysis;
applications in data mining, bioinformatics, signal processing, vision ...
The End
We are slowly addressing the challenges.
The methods used here (+ many more) are included in the
Ghostminer, data mining software developed by my group,
in collaboration with FQS, Fujitsu Kyushu Systems
http://www.fqspl.com.pl/ghostminer/
Completely new version of these tools is being developed.
Papers describing in details some of the ideas presented here
may be accessed through my home page:
Google: Duch
or
http://www.phys.uni.torun.pl/~duch