Learn the Fundamentals of XCUITest Framework_ A Beginner's Guide.pdf
ML whitepaper v0.2
1. Use loops to automate reoccurring processes
exploreread process split model asses deploy
csv reader –
brows to the file
you wish to read
and run the node
xls or xlsx reader
– brows to the
file you wish to
read, select the
desired sheet
and run the node
Parquet files are
great for fast
loading of data –
brows to the file
you wish to read,
and run the node
The traffic light below nodes
signals their state :
Red – not configured needs
additional configuration to run
Yellow – configured and ready
for execution
Green - executed
Use 2D-3D
scatter plot to
visually explore
sparsity, correlation,
feature-interaction
and more
Use the line plot
node in order to
explore trends
and regression
results
Use the data
explorer node to
quickly examine
various stats
about the data
Use math
formula nodes to
manipulate data
and create new
features
Use Pivoting /
Unpivoting /
GroupBy nodes to
create aggregations
of the data per one
of the features
Remove outliers to
create a less skewed
dataset (use wisely as
you might also remove
some of the legitimate
variability of the data)
Use Missing data node to
deal with missing data
completion of missing
values can be by a statistical
value interpolation or other
Type conversions:
Use integer to string to
define a classification task
with numeric targets
Partition the data to
train and test
We can further
partition to train /
validation / test using
two such nodes
Partition the data
multiple times
using a loop
Use xgboost learner to
model classification tasks
(predicting categories from
data) by boosted trees
Parameters:
Objective – binary for 2 classes /
multiclass for more
Eta – the learning rate the lower it is
the more boosting round we will
need – use eta=0.05 as default with
boosting rounds=500-1000
Subsampling rate=0.8
Column sample rate by tree=0.7
Increase regularization by:
• Reducing maximum depth
• Increasing minimum child weight
Use xgboost learner
(regression) to
model regression
tasks (prediction of
sequential targets)
And the same with
regression
ports are the way for each node to receive & transfer
information with other nodes in the workflow
Use the scorer node to
evaluate prediction accuracy
and confusion matrix
(classification models)
Roc curve
You can automatically send
mails from a process using
the send mail node
use the csv writer to
write output data
to a csv file
Meta nodes
Use loop end node
to collect results of
all runs
Use the appropriate
predictor to generate
predictions from the
trained model
Use string to datetime in
order to extract datetime
related features from
datetime saved as strings
Again, use the
appropriate predictor
to generate predictions
from the trained model
Use random forest
learner to model
classification tasks
(predicting categories
from data) by bagging
many decision trees
Again, use the
appropriate predictor
to generate predictions
from the trained model
Data table
Flow variables
Model
Tree ensemble
model
Black triangle ports
represent Input / output
Contains table with data –
either strings / numerical /
integer values are allowed
Red circle ports represent
Input/output consisted of
one or more flow variables
that may be used to replace
parameters in future nodes
Blue ports represent a
trained model
grey ports represent a
trained tree ensemble model
Flow variable ports are
usually hidden for most of
the nodes to display them
right click the node and select
“show flow variable ports”
use the csv writer to
write output data to
either xls or xlsx file
Example for using feature selection loop
Use the numeric scorer node
to evaluate prediction MAE /
(regression models)
Its easy to create encapsulate several nodes to a meta-node
Just select all of the nodes you want to encapsulate, right click
one of them, and select “create meta-node”
Time series
Create lagged
columns to imitate
prediction mode
Remember to split the data by
time to make sure that validation
is consistent with the type of data
you are expected to get on
deployment
Read / write
trained network
learner
Time series
lagged feature creation
Deep learning LSTM training and prediction flow
Machine learning and data science with KNIME – Nathaniel Shimoni
Hyper-parameters
optimization loop