【Python】Floating-point precision of the split node

[XTmingyue]

I have encountered an unusual situation:

A model shows the split point for a particular node as follows:

4:[f21<166.666702] yes=9,no=10,missing=9

If I feed in an observation where f21 = 166.6667 the python model incorrectly sends this to the yes node. I've limited the number of floating-point numbers in Python to four decimal places.

[trivialfis]

How do you limit floating point precision in Python?

[Denisevi4]

166.6667 < 166.666702

Should be a yes node. Or am I missing something?

[trivialfis]

Floating point printing is a very tricky thing, it has round off:

In [1]: print('{:.17f}'.format(166.666702))                                                                                                                                                                 
166.66670199999998658

And the round off mode is dependent on implementation.

[trivialfis]

Also 1.0.2 is now out. Feel free to try the experimental JSON format.