Transcript Why supervised classification?

Classification
Supervised and unsupervised
Tormod Næs
Matforsk
and
University of Oslo
Classificaton
• Unsupervised (cluster analysis)
– Searching for groups in the data
• Suspicion or general exploration
– Hierarchical methods, partitioning methods
• Supervised (discriminant analysis)
– Groups determined by other information
• External or from a cluster analysis
– Understand differences between groups
– Allocate new objects to the groups
• Scoring, finding degree of membership
Group 1
What is the difference?
?
Where?
New object X
?
Group 2
Why supervised classification?
• Authenticity studies
– Adulteration, impurities, different origin,
species etc.
• Raw materials
• Consumer products according to specification
• When quality classes are more important
than chemical values
• raw materials acceptable or not
• raw materials for different products
Flow chart for discriminant analysis
Main problems
• Selectivity
– Multivariate methods are needed
• Collinearity
– Data compression is needed
• Complex group structures
– Ellipses, squares or ”bananas”?
X2
Adulterated
Authentic
The selectivity problem
X1
Solving the selectivity problem
• Using several measurements at the same
time
– The information is there!
• Multivariate methods. These methods combine
several instrumental NIR variables in order to
determine the property of interest
• Mathematical ”purification” instead of wet chemical
analysis
Multivariate methods
Too many variables can also sometimes create
problems
–
–
–
–
Interpretation
Computations, time and numerical stability
Simple and difficult regions (nonlinearity)
Overfitting is easier (dependentent on method used)
• Sometimes important to find good compromises
(variable selection)
Conflict between flexibility and stability
Estimation error
Model error
Some main classes of methods
• Classical Bayes classification
– LDA, QDA
• Variants, modifications used to solve the
collinearity problem
– RDA, DASCO, SIMCA
• Classification based on regression analysis
– DPLS, DPCR
• KNN methods, flexible with respect to shape of
the groups
Bayes classification
• Assume prior probabilities pj for the groups
– If unknown, fix them to be pj= 1/C or
– equal to the proportions in the dataset
• Assume known probability model within
each class (fj(x))
– Estimated from the data, usually covariance
matrices and means
Bayes classification
• +
•
•
•
•
•
well understood, much used, often good properties, easy to validate
easy to modify for collinear data
Easy to updated, covariances
Can be modified for cost
Outlier diagnostics (not directly, but can be done, M-distance)
• • Can not handle too complex group structures, designed for elliptic
structures
• not so easy to interpret directly
• often followed by a Fisher’s linear discriminant analysis. Directly
related to interpreting differences between groups
Bayes rule
Maximise porterior probability
Normal data, minimise
1
Li  ( xi   j )  j ( xi   j )  log  j  2 log  j
T
Estimate model parameters,
ˆL  ( x  ˆ )T ˆ 1 ( x  ˆ )  log ˆ  2 log 
i
i
j
j
i
j
j
j
Mahalanobis distance plus determinant minus prior probability
Different covariance
structures
Mahalanobis distance is constant on ellipsoids
Best known members
• Equal covariance matrix for each group
– LDA
• Unequal covariance matrices
– QDA
• Collinear data, unstable inverted covariance
matrix (see equation)
– Use principal components (or PLS components)
– RDA, DASCO estimate stable inverse
covariance matrices
Classification by regression
• 0,1 dummy variables for each group
• Run PLS-2 (or PCR) or any other method which solves the
collinearity
• Predict class membership.
– The class with the highest value gets the vote
• All regular interpretation tools are available, variable
selection, plotting outliers diagnostics etc.
• Linear borders between subgroups, not too complicated
groups.
• Related to LDA, not covered here
• If large data sets, we can use more flexible methods
Example, classification of mayonnaise based on different oils
Indahl et al (1999). Chemolab
, Feasibility study, authenticity
The oils were
•soybean
•sunflower
•canola
•olive
•corn
•grapeseed
16 samples in each group
Start out low
Classification properties of QDA, LDA and regression
Comparison
• LDA and QDA gave almost identical results
• It was substantially better to use LDA/QDA
based on PLS/PCA components instead of
using PLS directly
Fisher’s linear discriminant
analysis
• Closely related to LDA
• Focuses on interpretation
– Use “spectral loadings” or group averages
• Finds the directions in space which distinguish the most
between groups
– Uncorrelated
• Sensitive to overfitting, use PC’s first
Fisher’s method.
Næs, Isaksson, Fearn and Davies (2001). A user friendly guide to cal. and class.
Plot of PC1 vs PC2
Not possible to distinguish the groups from each other
Mayonnaise data, clear separation
Canonical variates based on PC’s
Italian wines from same region, but based on different cultivars,
27 chromatic and chemical variables
Barbera
Barolo
Grignolino
PCA
Fisher’s method
Forina et al(1986), Vitis
Error rates
Validated properly
• LDA
– Barolo 100%, Grignolino 97.7%, Barbera
100%
• QDA
– Barolo 100%, Grignolino 100%, Barbera100%
KNN methods
• No model assumptions
• Therefore: needs data from “everywhere” and many data
points
• Flexible, complex data structures
• Sensitive to overfitting, use PC’s
New sample
KNN, finds the N samples which are closest
In this case 3 samples
Cluster analysis
Unsupervised classification
• Identifying groups in the data
– Explorative
Examples of use
• Forina et al(1982). Olive oil from different regions (fatty
acid composition). Ann. Chim.
• Armanino et al(1989), Olive oils from different Tuscan
provinces (acids, sterols, alcohols). Chemolab.
Methods
• PCA (informal/graphical)
– Look for structures in scores plots
– Interpretation of subgroups using loadings plots
• Hierarchical methods (more formal)
– Based on distances between objects (Euclidean or
Mahalanobis)
– Join the two most similar
– Interpret dendrograms
120 olive oils from one region in Italy, 29 variables (fatty acids, sterols, etc.)
Armanino et al(1989), Chem.Int. lab. Systems.