A package for dealing with crowdsourced big data.
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
.circleci
docker
docs
examples
img
spark-crowd Changes version of methods with likelihood Dec 13, 2017
.gitignore Modifies examples to making them more interesting Oct 30, 2017
LICENSE
README.md

README.md

CircleCI branch Codecov Maven Central license

spark-crowd

A package for dealing with crowdsourced big data.

Crowdsourced data introduces new problems that need to be addressed by Machine learning algorithms. This illustration exemplifies the main issues of using this kind of data.

Crowdsourcing illustration

Installation

The package uses sbt for building the project, so we recommend installing this tool if you do not yet have it installed.

The simplest way to use the package is adding the next dependency directly into the build.sbt file of your project:

libraryDependencies += "com.enriquegrodrigo" %% "spark-crowd" % "0.1.5"

If this is not a possibility, you can compile the project and create a .jar file or, you can publish the project to a local repository, as explained below.

Creating a .jar file and adding it to a new project

In the spark-crowd folder, one should execute the command

> sbt package 

to create a .jar file. Usually, it is located in target/scala-2.11/spark-crowd_2.11-0.1.jar.

This .jar can be added to new projects using this library. In sbt one can add .jar files to the lib folder.

Publishing to a local repository

In the spark-crowd folder, one should execute the command

> sbt publish-local 

to publish the library to a local Ivy repository. The library can be added to the build.sbt file of a new project with the following line:

    libraryDependencies += "com.enriquegrodrigo" %% "spark-crowd" % "0.1.5"

Usage

Running the examples

Docker Automated buil Docker Build Statu

For running the examples of this package one can use our docker image with the latest version of spark-crowd. Let's see how to run the DawidSkeneExample.scala file:

docker run --rm -it -v $(pwd)/:/home/work/project enriquegrodrigo/spark-crowd DawidSkeneExample.scala

For running a spark-shell with the library pre-loaded, one can use:

docker run --rm -it -v $(pwd)/:/home/work/project enriquegrodrigo/spark-crowd 

One can also generate a .jar file as seen previously and use it with spark-shell or spark-submit. For example, with spark-shell: spark-shell --jars spark-crowd_2.11-0.1.5.jar -i DawidSkeneExample.scala

Types

This package makes extensive use of Spark DataFrame and Dataset APIs. The latter takes advantage of typed rows which is beneficial for debugging purposes, among other things. As the annotations data sets usually have a fixed structure the package includes types three annotations data sets (binary, multiclass and real annotations), all of them with the following structure:

example annotator value
1 1 0
1 2 1
2 2 0
... ... ...

So the user needs to provide the annotations using this typed data sets to apply the learning methods. This is usually simple if the user has all the information above in a Spark DataFrame:

  • The example variable should be in the range [0..number of Examples]
  • The annotator variable should be in the range [0..number of Annotators]
  • The value variable should be in the range [0..number of Classes]
import com.enriquegrodrigo.spark.crowd.types.BinaryAnnotation

val df = annotationDataFrame

val converted = fixed.map(x => BinaryAnnotation(x.getLong(0), x.getLong(1), x.getInt(2)))
                     .as[BinaryAnnotation]

The process is similar for the other types of annotation data. The converted Spark Dataset is ready to be use with the methods commented in the Methods subsection.

In the case of the feature dataset, the requisites are that:

  • Appart from the features, the data must have an example and a class columns.
  • The example must be of type Long.
  • Class must be of type Integer or Double, depending on the type of class (discrete or continuous).
  • All features must be of type Double. For discrete features one can use indicator variables.

Methods

The methods implemented as well as the type of annotations that they support are summarised in the following table:

Method Binary Multiclass Real Reference
MajorityVoting
DawidSkene JRSS
GLAD NIPS
Raykar JMLR

The algorithm name links to the documentation of the implemented method in our application. The Reference column contains a link to where the algorithm was published. As an example, the following code shows how to use the DawidSkene method:

import com.enriquegrodrigo.spark.crowd.methods.DawidSkene

//Dataset of annotations
val df = annotationDataset.as[MulticlassAnnotation]

//Parameters for the method
val eMIters = 10
val eMThreshold = 0.01 

//Algorithm execution
result = DawidSkene(df,eMIters, eMThreshold) 
annotatorReliability = result.params.pi
groundTruth = result.dataset

The information obtained from each algorithm as well as about the parameters needed by them can be found in the documentation.

Credits

Author:

  • Enrique G. Rodrigo

Contributors:

  • Juan A. Aledo
  • Jose A. Gamez

License

MIT License

Copyright (c) 2017 Enrique González Rodrigo

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.