PCA Implementation in Java

Question

I need implementation of PCA in Java. I am interested in finding something that's well documented, practical and easy to use. Any recommendations?

Answer 1

There are now a number of Principal Component Analysis implementations for Java.

1. Apache Spark: https://spark.apache.org/docs/2.1.0/mllib-dimensionality-reduction.html#principal-component-analysis-pca

   SparkConf conf = new SparkConf().setAppName("PCAExample").setMaster("local");
   try (JavaSparkContext sc = new JavaSparkContext(conf)) {
       //Create points as Spark Vectors
       List<Vector> vectors = Arrays.asList(
               Vectors.dense( -1.0, -1.0 ),
               Vectors.dense( -1.0, 1.0 ),
               Vectors.dense( 1.0, 1.0 ));
   
       //Create Spark MLLib RDD
       JavaRDD<Vector> distData = sc.parallelize(vectors);
       RDD<Vector> vectorRDD = distData.rdd();
   
       //Execute PCA Projection to 2 dimensions
       PCA pca = new PCA(2); 
       PCAModel pcaModel = pca.fit(vectorRDD);
       Matrix matrix = pcaModel.pc();
   }
   

2. ND4J: https://javadoc.io/doc/org.nd4j/nd4j-api/latest/org/nd4j/linalg/dimensionalityreduction/PCA.html

   //Create points as NDArray instances
   List<INDArray> ndArrays = Arrays.asList(
           new NDArray(new float [] {-1.0F, -1.0F}),
           new NDArray(new float [] {-1.0F, 1.0F}),
           new NDArray(new float [] {1.0F, 1.0F}));
   
   //Create matrix of points (rows are observations; columns are features)
   INDArray matrix = new NDArray(ndArrays, new int [] {3,2});
   
   //Execute PCA - again to 2 dimensions
   INDArray factors = PCA.pca_factor(matrix, 2, false);
   

3. Apache Commons Math (single threaded; no framework)

   //create points in a double array
   double[][] pointsArray = new double[][] { 
       new double[] { -1.0, -1.0 }, 
       new double[] { -1.0, 1.0 },
       new double[] { 1.0, 1.0 } };
   
   //create real matrix
   RealMatrix realMatrix = MatrixUtils.createRealMatrix(pointsArray);
   
   //create covariance matrix of points, then find eigenvectors
   //see https://stats.stackexchange.com/questions/2691/making-sense-of-principal-component-analysis-eigenvectors-eigenvalues
   
   Covariance covariance = new Covariance(realMatrix);
   RealMatrix covarianceMatrix = covariance.getCovarianceMatrix();
   EigenDecomposition ed = new EigenDecomposition(covarianceMatrix);
   

Note, Singular Value Decomposition, which can also be used to find Principal Components, has equivalent implementations.

Answer 2

Here is one: PCA Class.

This class contains the methods necessary for a basic Principal Component Analysis with a varimax rotation. Options are available for an analysis using either the covariance or the correlation martix. A parallel analysis, using Monte Carlo simulations, is performed. Extraction criteria based on eigenvalues greater than unity, greater than a Monte Carlo eigenvalue percentile or greater than the Monte Carlo eigenvalue means are available.

Answer 3

check http://weka.sourceforge.net/doc.stable/weka/attributeSelection/PrincipalComponents.html weka in fact have many other algorithm that could be used with along with PCA and also weka is adding more algorithm from time to time. so i thing, if you are working on java then switch to weka api.

Answer 4

Smile is a full-fledged ML library for java. You give its PCA implementation a try. Please see: https://haifengl.github.io/smile/api/java/smile/projection/PCA.html

There is also PCA tutorial with Smile but the tutorial uses Scala.

Answer 5

You can see a few implementations of PCA in the DataMelt project:

https://jwork.org/dmelt/code/index.php?keyword=PCA

(they are rewritten in Jython). They include some graphical examples for dimensionality reduction. They show the usage of several Java packages, such as JSAT, DatumBox and others.