KNIME - Seven Techs For Dimensionality Reduction

Seven Techs for Dimensionality
Reduction
Subtitle
Contents
Introduction
Rules
Missing Values
Low Variance Filter
High Correlation Filter
PCA
Random Forests
Backward Feature Elimination
Forward Feature Construction
Comparison
Combining Dimensionality Reduction Techniques
Introduction
KDD Challenge: Customer relationship prediction.
Churn represents a contract severance by a customer;
Appetency the propensity to buy a service or a product
Upselling the possibility of selling additional side products to the main one
The data set came from the CRM system of a big French
telecommunications company
Huge data set with 50K rows and 15K columns
The problem is not the size of the data set, but rather the number of
input columns
Dimensionality reduction is based on supervised classification
algorithms
Used a cascade of the most promising techniques, as detected in the
first phase of the project on the smaller data set.
Evaluate data columns reduction based on:
High number of missing values
Low variance
High correlation with other data columns
Principal Component Analysis (PCA)
First cuts in random forest trees
Backward feature elimination
Forward feature construction
For the methods that use a threshold, an optimal threshold can be defined
through an optimization loop maximizing the classification accuracy on a
validation set for the best out of three classification algorithms:
MLP
decision tree
Nave Bayes.
Missing Values
Removing Data Columns with Too Many Missing Values
Example: if a data column has only 5-10% of the possible
values, it will likely not be useful for the classification of most
records.
Remove those data columns with too many missing values, i.e.
with more missing values in percent than a given threshold.
To count the number of missing values in all data columns, we
can either use a Statistics or a GroupBy node.
Ratio of missing values = no of missing values / total number
of rows
The bottle neck of this whole operation, in terms of execution

time, occurs when the missing values are counted for all input
data columns, either using the Statistics node or the GroupBy
node. This is due to the underlying data sorting
Threshold value 0.4 led to a dimensionality reduction of 71%
and a corresponding accuracy of 76% by the MLP.
Low Variance Filter
Measure the column variance
Remove those columns with a variance value below a given
threshold.
Numerical columns only.
Need to normalize data column ranges (to [0, 1] for instance).
Best threshold value : 0.03.
Classification accuracy: 82%.
Dimensionality reduction :73%.
Higher threshold values actually produce worse accuracy
values
High Correlation Filter
Correlated features: They depend on one another and carry
similar information
One of the two correlated columns can be removed without
decreasing the amount of information available for future
tasks dramatically
Correlation coefficient
Pearsons Product Moment Coefficient (numerical data)
Pearson's chi square value (nominal or categorical data)
No correlation coefficient is defined between a numerical and a nominal
data column.
Normalization for uniform data ranges
PCA
PCA transformation is sensitive to the relative scaling of the
original variables. Data column ranges need to be normalized
before applying PCA.
The new coordinates (PCs) are not real system-produced
variables anymore
Applying PCA to your data set loses its interpretability.
If interpretability of the results is important for the analysis,
PCA is not recommended.
PCA only affects the numerical columns and does not act on
the nominal columns
PCA skips the missing values. On a data set with many missing
values, PCA will be less effective.
With loss of 4% of the original data set information, PCA
PCA
62% reduction rate and an accuracy of 74%
Threshold at 96% to keep accuracy over 70%.
Random Forests
Useful for feature selection in addition to being effective classifiers.
One approach to dimensionality reduction is to generate a large and
carefully constructed set of trees against a target attribute and then
use each attributes usage statistics to find the most informative
subset of features.
Generate a set of very shallow trees, with each tree being trained on
a small fraction of the total number of attributes.
Example: 2000 trees were trained. Each tree was trained independently on 2
levels and 3 attributes:
If an attribute is often selected as best split, it is most likely an
informative feature to retain.
Calculate a score for each attribute, by counting how many times it
has been selected for a split and at which rank (level) among all
available attributes (candidates) in the trees of the ensemble
Score = #splits(lev.0)/#candidates(lev.0) +
#splits(lev.1)/#candidates(lev.1)
Backward Feature Elimination
The Backward Feature Elimination loop performs
dimensionality reduction against a particular machine learning
algorithm.
At each iteration, the selected classification algorithm is
trained on n input features.
Then we remove one input feature at a time and train the
same model on n-1 input features n times.
Finally, the input feature whose removal has produced the
smallest increase in the error rate is removed, leaving us with
n-1 input features.
The classification is then repeated using n-2 features n-1 times
and so on.
The algorithm starts with all available N input features and
Each iteration k produces a model trained on n-k features and an error
rate e(k)
Selecting the maximum tolerable error rate: define the smallest number
of features necessary to reach the desired classification performance
with the selected machine learning algorithm
The default classification algorithm is Nave Bayes (can be other
classification algorithms)
The main drawback of this technique is the high number of iterations
for very high dimensional data sets, possibly leading to very long
computation times.
It is better to apply it only after some other dimensionality reduction
had taken place.
Only 77 features -> circa 10 hours on a quad-core laptop with 8GB RAM.
Forward Feature Construction
Builds a number of pre-selected classifiers using an
incremental number of input features.
Loop starts from 1 feature and adds one more feature at a
time in the subsequent iterations.
Drawback: Long computational time for high dimensional data
sets
Using the Nave Bayes as classification algorithm and 20 final
input features (91% reduction rate): 83% accuracy on the
validation set
Comparison
The machine learning algorithm with the highest accuracy on the
validation set is the neural network
The missing value ratio seems a safe starting point. it obtains a
good reduction rate without compromising performances and
applies to both numerical and nominal columns.
The random forest based method achieves the highest reduction
rate
Backward elimination and forward construction techniques produce
great figures in terms of accuracy. However, their reduction rate
proved to be the highest. ?
PCA seems to perform the worst in the compromise between
reduction rate and model performance
Combining Dimensionality
Reduction Techniques
First: use missing values with threshold 90%
reduced the number of data columns from 150000 to 14578 (3%
reduction)
Second: Low Variance Filter on a subset of the data rows and
with a 0.003 threshold value
Number of input columns was further reduced from 14578 to 3511
(76% reduction)
Finally: PCA with allowing a loss of 1% of the information
data set dimensionality could be reduced from 3511 to 2136 data
columns (39% reduction).
Total dimensionality reduction : 86% (15000 original data
columns to 2136 reduced ones).

KNIME - Seven Techs For Dimensionality Reduction

Uploaded by

Copyright:

Available Formats

KNIME - Seven Techs For Dimensionality Reduction

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

KNIME - Seven Techs For Dimensionality Reduction

Uploaded by

Copyright:

Available Formats

Seven Techs for Dimensionality

The bottle neck of this whole operation, in terms of execution

You might also like