Fix SHAP prediction explanations bug

[eccabay]

Closes #3220

Proposed solution is to remove the usage of the logit link function in our explanations altogether. The only difference between the function is what space the final SHAP values are reported in. With the logit function they sit in log-odds space, but between -1 and 1 with the identity function. Due to the lack of significant difference between the two methods, I figure there’s no significant impact in making this change.

[codecov]

Codecov Report

Merging #3221 (0b8dc24) into main (119c8c7) will increase coverage by 0.1%.
The diff coverage is 100.0%.

@@           Coverage Diff           @@
##            main   #3221     +/-   ##
=======================================
+ Coverage   99.8%   99.8%   +0.1%     
=======================================
  Files        326     326             
  Lines      31388   31395      +7     
=======================================
+ Hits       31297   31304      +7     
  Misses        91      91             

Impacted Files	Coverage Δ
...derstanding/prediction_explanations/_algorithms.py	100.0% <ø> (ø)
...s/prediction_explanations_tests/test_algorithms.py	100.0% <100.0%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 119c8c7...0b8dc24. Read the comment docs.

[freddyaboulton]

Looks good to me @eccabay ! Thank you for the fix. Can you please run the model understanding perf tests to be sure we're not accidentally introducing a regression?

[angela97lin]

LGTM, thank you @eccabay! Agreed with Freddy's suggestion about running perf tests, and I'm almost tempted to say that perhaps the identity method could be easier to understand than the logit link function :)

[bchen1116]

Agreed with both above! Nice fix!

[chukarsten]

Thank you!

[eccabay]

@freddyaboulton great suggestion for perf testing! Results look pretty consistent to me, lmk what you think!
report.html.zip

[freddyaboulton]

@eccabay Thanks for running the tests! Send it!