Some Confusion About ADWIN

[9sea]

Discussed in #1535

^{Originally posted by 9sea April 26, 2024}
The code is in line 244 of river.drift.adwin_c.pyx.
I found that the code at line 244 of the file adwin_c.pyx, which reads "for k in range(bucket.current_idx - 1)", differs from the realization in MOA. I believe it should be "for k in range(bucket.current_idx)". Of course, this is a minor difference, but it might also affect the conditional statement at line 272, causing that condition to always be false.
I would appreciate it if someone could clarify this for me. Thank you!

[smastelini]

Hi @9sea, thanks for reporting that.

Are you referring to this line, in MOA?

https://github.com/Waikato/moa/blob/5474e181c778f8a8cd8a549fa83c08975ff99dde/moa/src/main/java/moa/classifiers/core/driftdetection/ADWIN.java#L504

The loop in MOA uses the <= operator while River's version implicitly uses <. Although I did not do the original port from Java to Python and later Cython, I think you might have a point.

Have you tried to run some benchmarks to assess the implications of such a change?

[9sea]

Hi @smastelini , Yes, it is line 504 of the file. If the '<' is used, it will not iterate to the last bucket, which is different from the logic of the algorithm in the original paper. I have not run any benchmarks yet, and I suspect that there might have been a mistake when converting the code from Java to Python, possibly mistakenly writing '<' instead of '<='.

[smastelini]

It would be nice to run a few simple benchmarks to measure the impact of such a change and then create a PR with the needed changes. @9sea, are you willing to take care of that? I have a few things at the top of my priority list at the moment, so I can't check it right now.