Data Mining on NIJ data Sangjik Lee
Unstructured Data Mining Text Keyword Extraction Structured Data Base Data Mining Image Feature Extraction Structured Data Base Data Mining
Handwritten CEDAR Letter
Document Level Features 1. Entropy 2. Gray-level threshold 3. Number of black pixels 4. Stroke width 5. Number of interior contours 6. Number of exterior contours 7. Number of vertical slope components 8. Number of horizontal slope components 9. Number of negative slope components 10. Number of positive slope components 11. Slant 12. Height Measure of Pen Pressure Measure of Writing Movement Measure of Stroke Formation Slant Word Proportion
Character Level Features S y (i,j) tan S x (i,j)
Gradient : (192) Structure : (192) Concavity : (128) Character Level Features
Writer dataFeature data (normalized) Gen Age Han Edu Ethn Sch M F 85 L R H C H W B A O U F dark blob hole slant width skew ht int int int real int real int Writer and Feature Data
Instances of the Data (normalized) Feature document level data (12 features) Entropy dark pixel blob hole hslope nslope pslope vslope slant width ht real int int int int int int int int real int int
White maleWhite femaleBlack femaleBlack male Data Mining on sub-group
Data Mining on sub-group (Cont.) Subgroup analysis is useful information to be mined. 1-constraint subgroups {Male: Female}, {White : Black : Hispanic}, etc. 2-constraints subgroups {Male-white: Female-white}, etc. 3-constraints subgroups {Male-white-25~45: Female-white-25~45}, etc. Gen Age Han Edu Ethn Sch M F 85 L R H C H W B A O U F There are a combinatorially large number of subgroups.
Gender Age Handedness Ethnicity eDucation Schooling G W SDEHA If |W| < support, reject Constraints 1 GAGHAHAEADASHEHDHSEDESDSGSGDGE 2 GAEGADGAHGASGHEGHDGHSGEDGESGDSAHE 3 ……... GAHEDS... subgroups
Database ~ Normalized feature data Raw feature data Writer data Color Scale
Feature Database (White and Black) 12~24 25~44 45~64 >= 65 whiteblack Female whiteblack Male
What to do 1. Feature Selection Process that chooses an optimal subset of features according to a certain criterion (Feature Selection for knowledge discovery and data mining by Huan Liu and Hiroshi Motoda) Since there are limited number of writer in each sub-group, reduced subset of features is needed. To improve performance (speed of learning, predictive accuracy, or simplicity of rules) To visualize the data for model selection To reduce dimensionality and remove noise
Feature Selection Example of feature selection Feature 1-2 ~ 2-3Feature 6-10 ~ Feature 9-10 ~ Knowing that some features are highly correlated to some others can help removing redundant features
What to do 2. Visualization of trend (if any) of writer sub-groups Useful tool so that we can quickly obtain an overall structural view of the trend of sub-group Seeing is Believing !
Implementation of Subgroup Analysis on NIJ Data Writer Data Find a subgroup that has enouth support Data Preparation Subgroup Classifier Feature Data Task: Which writer subgroup is more distinguishable than others (if any)?
The Result of Subgroup Classification Results Procedure for writer subgroup analysis Find subgroup that has enough support Choose ‘the other’ (complement) group Make data sets(4) for Artificial Neural Network Train ANN and get the results from two test sets Limit 3 categoris are used (gender, ethnicity and age) up to 2 constraints are considered only Document-level features are used
1 Subgroup Classifier dark blob hole slant height Artificial neural network (11-6-1) This is a test. This is a sample writing for document 1 written by an author a. Feature space representation of Handwritten document is This is a test. This is a sample writing for document 1 written by an author a. of Handwritten document is Feature extraction Writer is Which group?
The Result of Subgroup Classification Results
They’re distinguishable, but why... Need to explain why they’re distinguishable ANN does a good job, but can’t explain clearly its output 12 features are too many to explain and visualize Only 2 (or 3) dimensions are visualizable Question : Does a reasonable two or three dimensional representation of the data exist that may be analyzed visually? Reference : Feature Selection for Knowledge Discovery and Data Mining - Huan Liu and Hiroshi Motoda
Feature Extraction Common characteristic of feature extraction methods is that they all produce new features y based on the original features x. After feature extraction, representation of data is changed so that many techniques such as visualization, decision tree building can be conveniently used. Feature extraction started, as early as in 60’s and 70’s, as a problem of finding the intrinsic dimensionality of a data set - the minimum number of independent features required to generate the instances
Visualization Perspective Data of high dimensions cannot be analyzed visually It is often necessary to reduce it’s dimensionality in order to visualize the data The most popular method of determining topological dimensionality is the Karhunen-Loeve (K-L) method (also called Principal Component Analysis) which is based on the eigenvalues of a covariance matrix(R) computed from the data
Visualization Perspective The M eigenvectors corresponding to the M largest eigenvalues of R define a linear transformation from the N- dimensional space to an M-dimensional space in which the features are uncorrelated. This property of uncorrelated features is derived from a theorem stating that if the eigenvalues of a matrix are distinct, then the associated eigenvectors are linearly independent For the purpose of visualization, one may take the M features corresponding to the M largest eigenvalues of R
Applied to the NIJ data 1. Normalize each feature’s values into a range [0,1] 2. Obtain the correlation matrix for the 12 original features 3. Find eigenvalues of the correlation matrix 4. Select the largest two eigenvalues should be chosen 5. Output the chosen eigenvectors associated with the chosen eigenvalues. Here we obtain a 12 * 2 transformation matrix M 6. Transform the normalized data D old into data D new of extracted features as follows: D new = D old M The resulting data is of 2-dimensional having the original class label attached to each instance
Applied to the NIJ data
Sample Iris data (the original is 4-dimensional)