[R] Fix parsing decision stump.

[trivialfis]

• Fix in dataframe construction.
• Fix in plotting.

Still need tests for more sophisticated cases like mixed tree types.

Close #7669

[trivialfis]

@hetong007 Could you please take a look into this fix? I'm a little bit confused by the regex parsing.

[hetong007]

Sure. What exactly is the confusing part?

[trivialfis]

@hetong007 For instance this comment in the code:

skip some indices with spurious capture groups from anynumber_regex

I'm not entirely sure what's it skipping. In general, I would have more confidence about the correctness of the code if we use the JSON dump along with jsonlite for parsing. But right now in the R API, the user can directly pass a text string input so I can't make the change otherwise it would be breaking.

[trivialfis]

Would be great if you can review the PR and see if the fix makes sense.

[hetong007]

I looked into the code with an example.

The to-parse text is "f0<0.5] yes=1,no=2,missing=1,gain=0.0366666317,cover=2", with branch_rx="f(\\d+)<([-+]?[0-9]*\\.?[0-9]+([eE][-+]?[0-9]+)?)\\] yes=(\\d+),no=(\\d+),missing=(\\d+),gain=([-+]?[0-9]*\\.?[0-9]+([eE][-+]?[0-9]+)?),cover=([-+]?[0-9]*\\.?[0-9]+([eE][-+]?[0-9]+)?)".

After the regmatches call, we have the results as

[[1]]
 [1] "f0<0.5] yes=1,no=2,missing=1,gain=0.0366666317,cover=2"
 [2] "0"
 [3] "0.5"
 [4] ""
 [5] "1"
 [6] "2"
 [7] "1"
 [8] "0.0366666317"
 [9] ""
[10] "2"
[11] ""

So basically the skipped indices contains either the input or an empty string, thus the list c(2, 3, 5, 6, 7, 8, 10) just slices those informative ones out.

Overall I think your fix here makes sense, with one inline comment.

[trivialfis]

@hetong007 Thank you for digging into this! Your explanation of the code is helpful.

On the issue of using constant instead of string value NA, I reverted the change as by the end of the function, there are procedures to convert each column to the correct type using as.numeric and as.integer. We need to use string value NA first, as the parsed values are strings and need to be consistent with the columns initialized by strings NA.

[hetong007]

@hetong007 Thank you for digging into this! Your explanation of the code is helpful.

On the issue of using constant instead of string value NA, I reverted the change as by the end of the function, there are procedures to convert each column to the correct type using as.numeric and as.integer. We need to use string value NA first, as the parsed values are strings and need to be consistent with the columns initialized by strings NA.

Cool! This makes sense.