Visualising a Classification in High Dimension, part 2

A few weeks ago, I published a post on Visualising a Classification in High Dimension, based on the use of a principal component analysis, to get a projection on the first two components. Following that post, I was wondering what could be done in the context of a classification on categorical covariates. A natural idea would be to consider a correspondance analysis, and to run a similar code.

Consider here the dataset used in a recent post,

> source("http://freakonometrics.free.fr/import_data_credit.R")

If we consider a correspondance analysis, we get

> library(FactoMineR)
> acm=MCA(train.db,quali.sup = 
+ which(names(train.db,)=="class"),ncp=10)

For the covariates (including also the variable we want to model, considered here as some supplementary variable), the visualisation – on the first two components – is

and for the individuals

The extension of the previous post (based on a principal component analysis) is easy : based on a point in the (projection) space above, we want to get the point in the original space, use our classifier to get a prediction, then visualise it on that projection. The problem is that when we get back to the original space, we might not end up in some categories: instead of having a gender “male” or “female“, our point might be 83% “male” and 17% “female“. We we can run our classifier in that fuzzy space, that’s fine. The good thing is that we can, in the context of a logistic regression (as discussed in a previous post, entitled Classification with Categorical Variables (the fuzzy side)).

I should also mentioned that when I wrote my initial code, I got some technical problem to get back to the original space (one dimension was missing)

> X=tab.disjonctif(train.db[,-which(
+ names(train.db)=="class")])
> ncol(X)
[1] 56
> res.mca = MCA(train.db,quali.sup =
+ which(names(train.db)=="class"),ncp=56)
> ncol(res.mca$var$coord)
[1] 55

So I asked Julie who kindly helped me by sharing some code she got.

> don <- train.db[,-which(names(train.db)=="class")]
> moy.p <- function(V, poids) {
+   res <- sum(V * poids, na.rm = TRUE)/sum(poids[!is.na(V)])
+ }
> row.w = rep(1/nrow(don), nrow(don))
> tab.disj.comp=tab.disjonctif(don)
> M = apply(tab.disj.comp, 2, moy.p, row.w)/ncol(don)
> Z = sweep(tab.disj.comp, 2, apply(tab.disj.comp, 2, moy.p, 
+     row.w), FUN = "/")
> Z = sweep(Z, 2, apply(Z, 2, moy.p, row.w), FUN = "-")
> Zscale = sweep(Z, 2, sqrt(M), FUN = "*")
> svd.Zscale = svd.triplet(Zscale, row.w = row.w)
> eig.shrunk = (svd.Zscale$vs[1:2])

Now, we can get a function that returns a point in the original space (in a fuzzy sense) given a location in the projected space

> reconstruct=function(x1,x2){
+ rec = tcrossprod(sweep(cbind(x1,x2), 2, 
+ eig.shrunk, FUN = "*"), svd.Zscale$V[, 1:2])
+ tab.disj.rec = sweep(rec, 2, sqrt(M), FUN = "/") + matrix(1, 
+ nrow(rec), ncol(rec))
+ tab.disj.rec = sweep(tab.disj.rec, 2, 
+ apply(tab.disj.comp, 2, moy.p, row.w), 
+ FUN = "*")
+ colnames(tab.disj.rec)=colnames(tab.disj.comp)
+ return(tab.disj.rec)
+ }

For instance, if we consider the origin in the projected space

> reconstruct(0,0)[1:10]
 [1] 0.25931677 0.06677019 0.27173913 0.40217391
 [5] 0.42546584 0.48447205 0.09006211 0.29503106
 [9] 0.09161491 0.61335404

This point is actually the avereage individual in our dataset

> (apply(tab.disj.comp,2,mean))[1:4]
       CA < 0 euros      CA > 200 euros 
         0.25931677          0.06677019 
CA in [0-200 euros[ No checking account 
         0.27173913          0.40217391

And indeed, this individual belongs to the fuzzy set of our covariates: this individual has 40% chance to get no account, 26% chance to get a negative balance, 27% to get a positive balance, below 200 euros, and 7% chance to get a positive balance above 200 euros. So it cannot be an actual observation from our dataset.

Our classifier should be in that fuzzy space, not on the space of categorical covariates. This was discussed in our previous post

> credit_disj=data.frame(class=train.db$class,
+ tab.disjonctif(don))
> reg=glm(class~.,data=credit_disj,
+ family=binomial)
> nd=as.data.frame(t(apply(tab.disj.comp,
+ 2,mean)))
> names(nd)=names(credit_disj)[-1]
> predict(reg,newdata=nd,type="response")
0.1934358

Here, we get exactly the same prediction using our reconstruction function,

> nd=data.frame(reconstruct(0,0))
> predict(reg,newdata=nd,type="response")
0.1934358

We can define a function to get a prediction

> prev=function(x1,x2){
+ newd=data.frame(retour(x1,x2))
+ return(predict(reg,newdata=newd,
+ type="response"))
+ }

It is then possible to get predictions on a grid, on the projective space

> xgrid=seq(-1,2,length=51)
> ygrid=seq(-1,1,length=51)
> xygrid=as.matrix(expand.grid(xgrid,ygrid))
> rec = tcrossprod(sweep(xygrid, 2,
+ eig.shrunk, FUN = "*"), svd.Zscale$V[, 1:2])
> tab.disj.rec = sweep(rec, 2, sqrt(M), 
+ FUN = "/") + matrix(1,
+ nrow(rec), ncol(rec))
> tab.disj.rec = sweep(tab.disj.rec, 2, apply(
+ tab.disj.comp, 2, moy.p, row.w), FUN = "*")
> colnames(tab.disj.rec)=colnames(tab.disj.comp)
> newd=data.frame(tab.disj.rec)
> z_vect=predict(reg,newdata=newd,
+ type="response")
> zgrid=matrix(z_vect,51,51)

We can then visualize our classifier, on that projected space

> plot(acm$ind$coord[,1:2],xlab="Dim 1",ylab="Dim 2",col="white")
> abline(h=0,col="grey")
> abline(v=0,col="grey")
> points(acm$ind$coord[train.db$class=="0",1],
+ acm$ind$coord[train.db$class=="0",2],
+ col="blue",pch=19,cex=.5)
> points(acm$ind$coord[train.db$class=="1",1],
+ acm$ind$coord[train.db$class=="1",2],
+ col="red",pch=19,cex=.5)
> contour(xgrid,ygrid,zgrid,add=TRUE,
+ levels=prev(0,0),lwd=2)
> contour(xgrid,ygrid,zgrid,add=TRUE,col="grey")

 


3 thoughts on “Visualising a Classification in High Dimension, part 2”

  1. Hi Arthur!
    How about trying out t-SNE instead of PCA?
    http://blog.kaggle.com/2012/11/02/t-distributed-stochastic-neighbor-embedding-wins-merck-viz-challenge/

    ..a friend of mine recently tried t_SNE to do scatterplot of ukrainian parliamentaries’ voting alignment and the method was able to catch ‘behind the curtain’ links between politics pretty good.

    http://sakva.net/wp-content/uploads/rada.html

    maybe you want to compare it vs PCA on your dataset.

    Regards,
    igor.

  2. Hi Arthur,
    In the function you have defined to get prediction (prev): by retour(x1,x2), you just mean return(x1,x2)?.

Leave a Reply

Your email address will not be published. Required fields are marked *