[SPARK-18385][ML] Make the transformer's natively in ml framework to avoid extra conversion

[techaddict]

What changes were proposed in this pull request?

Follow Up of SPARK-14615
Transformer's added in ml framework to avoid extra conversion for:
ChiSqSelector
IDF
StandardScaler
PCA

How was this patch tested?

Existing Tests

[techaddict]

cc: @dbtsai@mengxr

[SparkQA]

Test build #68410 has finished for PR 15831 at commit a9483ef.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #68411 has finished for PR 15831 at commit 89e6858.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[sethah]

I see this patch was created as a result of the PR that separated the ml/mllib linalg packages, to avoid some inefficiencies in conversion. However, it also is a partial step toward feature parity. Typically, we would port full algorithms all at once, instead of just porting the transformer functionality as is done here, but I understand that this is not just about parity. I would suggest one of the following:

1. Port over full feature functionality. This increases the scope and therefore the algos should probably separated out individually into PRs.
2. Keep the scope the same, but avoid copying code.

For an example of option 2, for ChiSqSelector, we can implement new static methods in the mllib.ChiSqSelectorModel:

private[spark] defcompressDense(
selectedFeatures: Array[Int],
values: Array[Double]):Array[Double] = {
selectedFeatures.map(i => values(i))
}
private[spark] defcompressSparse(
compressedSize: Int,
selectedFeatures: Array[Int],
indices: Array[Int],
values: Array[Double]): (Array[Int], Array[Double]) = {
...
}

then in the actual model classes we can just do something like:

privatedefcompress(features: Vector):Vector= {
features match {
caseSparseVector(_, indices, values) =>valnewSize= selectedFeatures.length
val (newIndices, newValues) =ChiSqSelectorModel.compressSparse(newSize, selectedFeatures, indices, values)
Vectors.sparse(newSize, newIndices, newValues)
caseDenseVector(values) =>Vectors.dense(ChiSqSelectorModel.compressDense(selectedFeatures, values))
}
}

This approach would allow us to avoid copying a lot of code until we do full feature ports. What are others opinions? I lean towards the second option since it keeps the scope reasonable.

cc @dbtsai@yanboliang

[sethah]

btw there is no reason to have this case since Vector is a sealed trait

[techaddict]

@sethah I agree, 2nd approach is much more reasonable.

[yanboliang]

@techaddict@sethah I'm more prefer option 1, since we would like to remove spark.mllib package in a future release(may be 3.0) and we wouldn't like to make any change to it except bug fix. Could you make this improvement separately for relevant algorithms? Thanks.

[techaddict]

@sethah@yanboliang I've started with migrating IDF, can you review the WIP and if i'm going in the right direction https://github.com/techaddict/spark/pull/2/files
there is some code duplication were we can make mllib code actually depend on the ml one

[MLnick]

I'm also generally supportive of (1) - porting the code to ml and having the mllib code wrap the ml version - this is the approach for other models that have been done. Of course only once allmllib code has been ported over fully can we ultimately deprecate mllib.

I guess we can start doing this for some transformers like these - but ideally we should focus on porting stuff that's still missing in ml first.

I'd prefer that we create a top-level JIRA to track all the components that need to be done, and link everything appropriately. We also need to decide on priority - we may realistically be working on it over a 1-1.5 year time frame (of course hopefully it will take a lot shorter).

[techaddict]

@MLnick I will create a umbrella jira and start adding jira's for things I'm aware of of and you can start prioritising 👍 sounds like a plan ?

[zhengruifeng]

the same TODO also appear in HashingTF, what about include it in this PR?

[sethah]

I think we decided to go a different direction than what is proposed here? Actually, I still think there's merit in fixing the problem without having to do full feature ports. Either way, I'm not sure anyone is still taking on this task, so @zhengruifeng or @techaddict it would be great if you wanted to either revive this PR/help review, or start working on the larger umbrella JIRA and sub tasks...

[techaddict]

@sethah I will revive this pr thanks 👍

[zhengruifeng]

@techaddict@sethah I have some time to work on the porting, but I dont find the umbrella JIRA

[HyukjinKwon]

Hi @@techaddict, how is this PR going?

[techaddict]

@HyukjinKwon was busy, will restart this week.

[SparkQA]

Test build #77448 has finished for PR 15831 at commit 89e6858.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.