Advanced Data Mining with Weka

Class 4 – Lesson 1
What is distributed Weka?

Mark Hall

Pentaho

weka.waikato.ac.nz
Lesson 4.1: What is distributed Weka?

Class 1 Time series forecasting

Lesson 4.1 What is distributed Weka?

Class 2 Data stream mining

in Weka and MOA Lesson 4.2 Installing with Apache Spark

Class 3 Interfacing to R and other data Lesson 4.3 Using Naive Bayes and JRip
mining packages

Lesson 4.4 Map tasks and Reduce tasks

Class 4 Distributed processing with
Apache Spark
Lesson 4.5 Miscellaneous capabilities

Class 5 Scripting Weka in Python

Lesson 4.6 Application:
Image classification
Lesson 4.1: What is distributed Weka?

 A plugin that allows Weka algorithms to run on a cluster of machines

 Use when a dataset is too large to load into RAM on your desktop, OR
 Processing would take too long on a single machine
Lesson 4.1: What is distributed Weka?

 Class 2 covered data stream mining

– sequential online algorithms for handling large datasets
 Distributed Weka works with distributed processing frameworks that use
map-reduce
– Suited to large offline batch-based processing
 Divide (the data) and conquer over multiple processing machines
 More on map-reduce shortly…
Lesson 4.1: What is distributed Weka?

 Two packages are needed:

 distributedWekaBase
– General map-reduce tasks for machine learning that are not tied to any particular map-
reduce framework implementation
– Tasks for training classifiers and clusterers, and computing summary statistics and
correlations
 distributedWekaSpark
– A wrapper for the base tasks that works on the Spark platform
– There is also a package (several actually) that works with Hadoop
Lesson 4.1: What is distributed Weka?
Map-reduce programs involve a “map” and “reduce” phase

Dataset Map tasks Reduce task(s)

• Summarize:
• Processing:
• E.g. sorting, <key, result> • E.g. counting,
Data split adding,
filtering, computing
averaging
partial results

• Summarize:
• Processing:
• E.g. counting,
• E.g. sorting,
Data split adding,
filtering, computing
averaging
partial results <key, result>

Map-reduce frameworks provide orchestration, redundancy and fault-tolerance

Lesson 4.1: What is distributed Weka?

 Goals of distributed Weka

– Provide a similar experience to that of using desktop Weka
– Use any classification or regression learner
– Generate output (including evaluation) that looks just like that produced by desktop
Weka
– Produce models that are normal Weka models (some caveats apply)
 Not a goal (initially at least)
– Providing distributed implementations of every learning algorithm in Weka
• One exception: k-means clustering
– We’ll see how distributed Weka handles building models later…
Lesson 4.1: What is distributed Weka?

 What distributed Weka is

 When you would want to use it
 What map-reduce is
 Basic goals in the design of distributed Weka
Advanced Data Mining with Weka
Class 4 – Lesson 2
Installing with Apache Spark

Mark Hall

Pentaho

weka.waikato.ac.nz
Lesson 4.2: Installing with Apache Spark

Class 1 Time series forecasting

Lesson 4.1 What is distributed Weka?

Class 2 Data stream mining

in Weka and MOA Lesson 4.2 Installing with Apache Spark

Class 3 Interfacing to R and other data Lesson 4.3 Using Naive Bayes and JRip
mining packages

Lesson 4.4 Map tasks and Reduce tasks

Class 4 Distributed processing with
Apache Spark
Lesson 4.5 Miscellaneous capabilities

Class 5 Scripting Weka in Python

Lesson 4.6 Application:
Image classification
Lesson 4.2: Installing with Apache Spark

 Install distributedWekaSpark via the package manager

– This automatically installs the general framework-independent
distributedWekaBase package as well
 Restart Weka
 Check that the package has installed and loaded properly by
starting the Knowledge Flow UI
Lesson 4.2: Installing with Apache Spark

The hypothyroid data

 A benchmark dataset from the UCI machine learning repository

 Predict the type of thyroid disease a patient has
– Input attributes: demographic and medical information
 3772 instances with 30 attributes
 A version of this data, in CSV format without a header row, can be found in
${user.home}\wekafiles\packages\distributedWekaSpark\sample_data
Lesson 4.2: Installing with Apache Spark

Why CSV without a header rather than ARFF?

 Hadoop and Spark split data files up into blocks

– Distributed storage
– Data local processing
 There are “readers” for text files and various structured binary files
– Maintain the integrity of individual records
 ARFF would require a special reader, due to the ARFF header only being
present in one block of the data
Lesson 4.2: Installing with Apache Spark

 Getting distributed Weka installed

 Our test dataset: the hypothyroid data
 Data format processed by distributed Weka
 Distributed Weka job to generate summary statistics
Advanced Data Mining with Weka
Class 4 – Lesson 3
Using Naive Bayes and JRip

Mark Hall

Pentaho

weka.waikato.ac.nz
Lesson 4.3: Using Naive Bayes and JRip

Class 1 Time series forecasting

Lesson 4.1 What is distributed Weka?

Class 2 Data stream mining

in Weka and MOA Lesson 4.2 Installing with Apache Spark

Class 3 Interfacing to R and other data Lesson 4.3 Using Naive Bayes and JRip
mining packages

Lesson 4.4 Map tasks and Reduce tasks

Class 4 Distributed processing with
Apache Spark
Lesson 4.5 Miscellaneous capabilities

Class 5 Scripting Weka in Python

Lesson 4.6 Application:
Image classification
No slides for Lesson 4.3
Advanced Data Mining with Weka
Class 4 – Lesson 4
Map tasks and Reduce tasks

Mark Hall

Pentaho

weka.waikato.ac.nz
Lesson 4.4: Map tasks and Reduce tasks

Class 1 Time series forecasting

Lesson 4.1 What is distributed Weka?

Class 2 Data stream mining

in Weka and MOA Lesson 4.2 Installing with Apache Spark

Class 3 Interfacing to R and other data Lesson 4.3 Using Naive Bayes and JRip
mining packages

Lesson 4.4 Map tasks and Reduce tasks

Class 4 Distributed processing with
Apache Spark
Lesson 4.5 Miscellaneous capabilities

Class 5 Scripting Weka in Python

Lesson 4.6 Application:
Image classification
Lesson 4.4: Map tasks and Reduce tasks
How is a classifier learned in Spark?

Dataset Map tasks Reduce task

Either:
Data split Learn a model 1. Aggregate models
to form one final
model of the same Results
type OR
2. Make an ensemble
classifier using all
Data split Learn a model the individual
models
Lesson 4.4: Map tasks and Reduce tasks

Cross validation in Spark

 Implemented with two phases (passes over the data):

1. Phase one: model construction

2. Phase two: model evaluation
Lesson 4.4: Map tasks and Reduce tasks
Cross-validation in Spark phase 1: model construction

Reduce tasks:
Map tasks: build partial Aggregate the partial
Dataset models on parts of folds models for each fold
Fold 1 M1: fold 2 + 3 M1: fold 2 + 3 Results
M1
Fold 2 M2: fold 1 + 3 M1: fold 2 + 3
Fold 3 M3: fold 1 + 2
M2: fold 1 + 3
M2
M2: fold 1 + 3
Fold 1
M1: fold 2 + 3
Fold 2 M3: fold 1 + 2
M2: fold 1 + 3
Fold 3
M3
M3: fold 1 + 2 M3: fold 1 + 2
Lesson 4.4: Map tasks and Reduce tasks
Cross-validation in Spark phase 2: model evaluation

Map tasks: evaluate Reduce task

Dataset fold models
Fold 1 M1: fold 1 Results
Fold 2 M2: fold 2
Fold 3 M3: fold 3
Aggregate all partial
evaluation results

Fold 1 M1: fold 1

Fold 2 M2: fold 2
Fold 3 M3: fold 3
Lesson 4.3 & 4.4: Exploring the Knowledge Flow templates

 Creating ARFF metadata and summary statistics for a dataset

 How distributed Weka builds models
 Distributed cross-validation
Advanced Data Mining with Weka
Class 4 – Lesson 5
Miscellaneous capabilities

Mark Hall

Pentaho

weka.waikato.ac.nz
Lesson 4.5: Miscellaneous capabilities

Class 1 Time series forecasting

Lesson 4.1 What is distributed Weka?

Class 2 Data stream mining

in Weka and MOA Lesson 4.2 Installing for Apache Spark

Class 3 Interfacing to R and other data Lesson 4.3 Using Naive Bayes and JRip
mining packages

Lesson 4.4 Map tasks and Reduce tasks

Class 4 Distributed processing with
Apache Spark
Lesson 4.5 Miscellaneous capabilities

Class 5 Scripting Weka in Python

Lesson 4.6 Application:
Image classification
Lesson 4.5: Miscellaneous capabilities

 Computing a correlation matrix in Spark and using it as input to PCA

 Running k-means clustering in Spark
 Where to go for information on setting up Spark clusters
Lesson 4.5: Miscellaneous capabilities

Further reading

 Distributed Weka for Spark

– http://markahall.blogspot.co.nz/2015/03/weka-and-spark.html
 Distributed Weka for Hadoop
– http://markahall.blogspot.co.nz/2013/10/weka-and-hadoop-part-1.html
 K-means|| clustering in distributed Weka
– http://markahall.blogspot.co.nz/2014/09/k-means-in-distributed-weka-for-hadoop.html
 Apache Spark documentation
– http://spark.apache.org/docs/latest/
 Setting up a simple stand-alone cluster
– http://blog.knoldus.com/2015/04/14/setup-a-apache-spark-cluster-in-your-single-
standalone-machine/
Advanced Data Mining with Weka
Class 4 – Lesson 6
Application: Image classification

Michael Mayo
Department of Computer Science
University of Waikato
New Zealand

weka.waikato.ac.nz
Lesson 4.6: Application: Image classification

Class 1 Time series forecasting

Lesson 4.1 What is distributed Weka?

Class 2 Data stream mining

in Weka and MOA Lesson 4.2 Installing for Apache Spark

Class 3 Interfacing to R and other data Lesson 4.3 Using Naive Bayes and JRip
mining packages

Lesson 4.4 Map tasks and Reduce tasks

Class 4 Distributed processing with
Apache Spark
Lesson 4.5 Miscellaneous capabilities

Class 5 Scripting Weka in Python

Lesson 4.6 Application:
Image classification
Lesson 4.6: Application: Image classification

 Image features are a way of describing an image using numbers

 For example:
f1 50%
– How bright is the image (f1)?
– How much yellow is in the image (f2)? f2 2%
– How much green is in the image (f3)? f3 65%
– How symmetrical is the image (f4)?
f4 50%

f1 50%

f2 50%

f3 10%

f4 100%
 Lesson 4.6: Application: Image classification
@relation butterfly_vs_owl

 Image filters extract the same @attribute filename string

@attribute f1 numeric
features for a set of images @attribute f2 numeric
@attribute f3 numeric
@attribute class {BUTTERFLY,OWL}
@relation butterfly_vs_owl
@attribute filename string @data
@attribute class {BUTTERFLY,OWL} mno001.jpg,3,7,0,BUTTERFLY
@data mno002.jpg,1,2,0,BUTTERFLY
mno001.jpg,BUTTERFLY mno003.jpg,3,4,0,BUTTERFLY
mno002.jpg,BUTTERFLY mno004.jpg,6,3,0,BUTTERFLY
mno003.jpg,BUTTERFLY owl001.jpg,3,5,0,OWL
mno004.jpg,BUTTERFLY owl002.jpg,7,3,0,OWL
owl001.jpg,OWL owl003.jpg,3,5,0,OWL
owl002.jpg,OWL owl004.jpg,7,5,1,OWL
owl003.jpg,OWL
owl004.jpg,OWL
Lesson 4.6: Application: Image classification

1. Install imageFilters package using the Package Manager

2. Create your own ARFF file or use the example at
%HOMEPATH%/wekafiles/packages/imageFilters/data
3. Open the ARFF file in the WEKA Explorer
4. Select an image filter from
filters/unsupervised/instance/imagefilter
5. Set the filter’s imageDirectory option to the correct directory
6. Click the Apply button
7. Repeat 5-7 if you wish to apply more than one filter
8. (Optional) Remove the first filename attribute
9. Select a classifier and perform some experiments
Lesson 4.6: Application: Image classification

 Summary
– Image features are mathematical properties of images
– Image filters can be applied to calculate image features for an
entire dataset of images
– Different features measure different properties of the image
– Experimenting with WEKA can help you identify the best
combination of image feature and classifier for your data
Lesson 4.6: Application: Image classification

 References
– LIRE: Mathias L. & Chatzichristofis S.A. (2008) Lire: Lucene Image Retrieval – An
Extensible Java CBIR Library. Proceedings of the 16th ACM International Conference on
Multimedia, 1085-1088.
– MPEG7 Features: Manjunath B., Ohm J.R., Vasudevan V.V. & Yamada A. (2001) Color and
texture descriptors. IEEE Trans. on Circuits and Systems for Video Technology, 11, 703–
715.
– Bird images: Lazebnik S., Schmid C. & Ponce J. (2005) Maximum Entropy Framework for
Part-Based Texture and Object Recognition. Proceedings of the IEEE International
Conference on Computer Vision, vol. 1, 832-838.
– Butterfly images: Lazebnik S., Schmid C. & Ponce J. (2004) Semi-Local Affine Parts for
Object Recognition. Proceedings of the British Machine Vision Conference, vol. 2, 959-
968.
Advanced Data Mining with Weka
Department of Computer Science
University of Waikato
New Zealand

Creative Commons Attribution 3.0 Unported License

creativecommons.org/licenses/by/3.0/

weka.waikato.ac.nz