Design for Linear Discriminant Analysis (LDA)

[Sourish]

I am thinking about how we should do the Linear Discriminant Analysis (LDA) in CRRao. I am thinking out loud. Please correct me if I am saying something wrong. The design that I am thinking of is as follows:

container = @fitmodel(formula, data, modelClass,ClassificationType,CovarianceType)

Example: For binary classification:

container = @fitmodel(Specied~PetalLength+SepalLength,data,LinearDiscriminantAnalysis(),Binary)
container = @fitmodel(Specied~PetalLength+SepalLength,data,LinearDiscriminantAnalysis(),Binary,ShrinkageCov)
container = @fitmodel(Specied~PetalLength+SepalLength,data,LinearDiscriminantAnalysis(),Binary,PythonCov)

Example: For multi-class classification:

container = @fitmodel(Specied~PetalLength+SepalLength,data,LinearDiscriminantAnalysis(),Multi)
container = @fitmodel(Specied~PetalLength+SepalLength,data,LinearDiscriminantAnalysis(),Multi,ShrinkageCov)
container = @fitmodel(Specied~PetalLength+SepalLength,data,LinearDiscriminantAnalysis(),Multi,PythonCov)

The default covariance type would be sample covariance.

[Ajay Shah]

What's PythonCov?

What's the state of Julia packages for robust covariance matrices? Should we let the user pass a function as an argument?

[Sourish]

• By PythonCov, I meant Python's covariance estimation process.
• To my understanding, Julia is using simple covariance matrices.
• Regarding your question: Should we let the user pass a function as an argument? - I am giving CovarianceType as one of the options. There may be a default setup.

[Ajay Shah]

I was thinking that there will be a general technology for covariance matrix estimation: simple, multiple robust methods, maybe something that works for spare matrices in big data, etc. So there should be a default (simple) but the caller should be able to supply a function that computes the covariance matrix. Or alternatively maybe there will be a function compute.cov(X, method), the caller to LDA should be able to supply the method.

[Sourish]

Humm -- I like both ideas.

Idea 1) covariance matrix estimation: simple, multiple robust methods, shrinkage estimation methods etc.

Idea 2) a function compute.cov(X, method), the caller to LDA should be able to supply the method.

I like the second idea with a default robust method of R.

[Sourish]

@ajaynshah @ayushpatnaikgit @codetalker7 @ShouvikGhosh2048

Struggling to decide - should we use MultivariateStat.jl for LDA. AND/OR should we use Aman's Julia code from scratch for LDA

Ayush's point if we rely on too many packages - then some people will never able to use CRRao because some package will be broken

On the other hand - why bother we are going to rely on lazy load in any way...

Requesting your comment -- now I want to move to LDA development for CRRao

[Sourish]

For now, I am thinking about developing the LDA with Aman's code which is faster than R and Python sklearn but slower than MultivariateStat.jl

Once MultivariateStat.jl becomes stable - we can later adapt the LDA of MultivariateStat.jl as a fast option.

[Sourish]

@ajaynshah @ayushpatnaikgit

We will raise an issue with MultivariateStat.jl that predict is not working. If they provide a solution then we would go ahead and take it in CRRao.jl

Otherwise, we will contribute in MultivariateStat.jl.

[Ajay Shah]

Yes, great, let's put all our knowledge on LDA to work to make MultivariateStat.jl stronger. And then in CRRao we will just call that LDA. Let's do the usual hard work:

• Textual narration of what's wrong with the existing code
• test cases which demo that
• PR that solves this.

so that it gets rapidly accepted into the main package.

[Sourish]

I have created this issue with MultivariateStats.jl

JuliaStats/MultivariateStats.jl#204

Basically I said the predict for MulticlassLDA is not working.