DataVec | ETL Library for Machine Learning

You have a sparse matrix, but you're representing it in Eigen as a dense matrix. The matrix file that you have is also dense, it would be more convenient to use if it was stored in sparse form, the Market format for example.

If I change the matrix to a sparse one, and use

You need to build an executable JAR.

See e.g. How can I create an executable JAR with dependencies using Maven? for information how to do this with Maven.

You can use the maven assembly or maven shade plugin.

You must use imports in your Java source code. Maven dependencies do not replace imports.

They make imports possible, though. Without the dependency, the import will fail.

Iterate through your DataSetIterator and for each DataSet entry, create two new DataSets, each for train and test.

The key is to use the splitTestAndTrain method, which accepts a double fractionTrain that will specify the amount of data to be trained (the rest to be tested). There are different overloads of the method, so you can choose the one that fits your needs best. If you wish to add all train and test datasets to a common iterator, you could store them in two different Lists, and get their corresponding iterator later. Something like:

There was a colon in the path and removing it solved the problem.

Next question: Why it didn't make any problem to build C++ apps?

The Java API of OpenCV found in the org.opencv package doesn't come with a loader, so the libraries need to be loaded by something else externally. In the case of the JavaCPP Presets for OpenCV, the libraries and wrappers are all bundled in JAR files and we can call Loader.load(opencv_java.class) to load everything as documented here:
https://github.com/bytedeco/javacpp-presets/tree/master/opencv#documentation

JavaCV, Deeplearning4j, and DataVec do not use that Java API of OpenCV, they use the API found in the org.bytedeco.opencv package, which loads everything automatically, so they do not need to call anything.

The important part of the exception is this:

I think you are forcing a newer version of netty than Spark supports.

By running mvn dependency:tree you can see what version Spark wants here, and use that instead of the one you've defined.

If you don't care about Spark, but want to just use DataVec to transform your data, take a look at https://www.dubs.tech/guides/quickstart-with-dl4j/. It is a little bit outdated concerning the dependencies, but the datavec part shows how to use it without spark.