System information

ML.Net 1.5.2
.Net Framework 4.7.2

Issue

I have hundreds of projects, and they all have tree data structure like this:

A
  AA
     AAA
  BB
     BBB

Or like this:

A
  AA1
     AAA1
  BB2
     BBB2

Each project has its own tree structure which is modified from a standard tree structure. What I am trying to do is to map project's tree structure to the standard tree structure, like this:

A          <--- A
  AA       <---   AA1
     AAA   <---      AAA1
  BB       <---   BB2
     BBB   <---      BBB2

Or like this:

(The mapping really depends on the text instead of the node's level. )

Now I'm using multi class classification in ML.Net. First I map the existing projects' tree to the standard tree manually and save the results in the database, like this:

| Label      | Level1         | Level2         | Level3         |
| --------   | -------------- | -------------- | -------------- |
| A          | A              |      *         |       *        |
| A-AA       | A              |      AA1       |       *        |
| A-AA-AAA   | A              |      AA1       |      AAA1      |
| A-BB       | A              |      BB2       |       *        |
| A-BB-BBB   | A              |      BB2       |      BBB2      |
| A          | A              |      *         |       *        |
| A-AA-AAA   | A              |      AAA1      |       *        |
| A-BB       | A              |      BB2       |       *        |
| A-BB-BBB   | A              |      BB2       |      BBB2      |  

Because data in the column in ML.Net cannot be a missing value, so I replace them with *. And my tree has 15 levels (feature columns).

The multi class classification algorithm I choose is SdcaMaximumEntropy. Hopefully I can use the prediction to map the tree instead of doing this manually.

I successfully implemented the prediction. However, the prediction result is really poor.

So my question is:

1. Is the way I do this right?
2. If yes, should I remove the duplicate rows and should I replace the missing value with *?

Thanks in advance.