The following node is available in the Open Source KNIME predictive analytics and data mining platform version 2.7.1. Discover over 1000 other nodes, as well as enterprise functionality at
http://knime.com.
PNN Learner (DDA)
Trains a Probabilistic Neural Network (PNN) based on the DDA
(Dynamic Decay Adjustment) method on labeled data using
Constructive Training of Probabilistic Neural Networks as the
underlying algorithm.
This algorithm generates rules based on numeric data. Each rule is
defined as high-dimensional Gaussian function that is adjusted by
two thresholds, theta minus and theta plus, to avoid conflicts with
rules of different classes. Each Gaussian function is defined by
a center vector (from the first covered instance) and a standard
deviation which is adjusted during training to cover only
non-conflicting instances. The selected numeric columns of the
input data are used as input data for training and additional
columns are used as classification target, either one column holding
the class information or a number of numeric columns with class
degrees between 0 and 1 can be selected. The data output contains
the rules after execution along with a number of of rule
measurements. The model output port contains the PNN model,
which can be used for prediction in the PNN Predictor node.
Dialog Options
- Missing Values
-
Select one method to handle missing values: "Incorp" may generate
rules with missing values, if no replacement value has been
found during the learning process. "Best Guess" computes the optimal
replacement value by projecting the rule (with missing value(s))
onto the missing dimension(s) of all other rules. "Mean", "Min", and "Max"
replaces the missing value with each column's statistical property.
"Zero" and "One" perform a constant replacement by inserting either zero
or one.
- Advanced
-
Shrink after commit If selected, a shrink to reduce conflicting
rules is executed immediately after a new rule is committed, i.e. the
new rule is reduced so that conflicts with all other rules of different
classes are avoided.
Use class with max coverage If selected, only the class with
maximum coverage degree of the target columns is used during training,
otherwise all class columns are considered for coverage.
- Maximum no. Epochs
-
If selected, the option defines the maximum number of epochs
the algorithm has to process the entire data set, otherwise it repeats
this process until this rule model is stable, i.e. no new rule has been
committed and/or no shrink is executed.
- Target Columns
-
Select the target(s) to be used for classification. If more
than one column (only numeric) is selected, the columns must contain
class membership values between 0 and 1 for each class given by the
column name.
- PNN
-
Theta Minus This defines the upper boundary of activation for
conflicting rules: default value is 0.2.
Theta Plus This defines the lower boundary of activation for
non-conflicting rules: default value is 0.4.
Ports
Input Ports
0 |
Numeric data as well as class information used for training.
|
Output Ports
0 |
Rules as Gaussian functions, classification columns,
and additional rule measures.
|
1 |
PNN model can be used for prediction.
|
Views
- Learner Statistics
-
Displays a summary of the learning process.
This node is contained in KNIME Base Nodes
provided by KNIME GmbH, Konstanz, Germany.