Pinebioml

description:
1. RF tuner search range change: cpp :[1e-3, 1e-1] -> [1e-5, 1e-1]. the larger dataset needs smaller cpp.
2. Pine do not retune the model while doing cv evaluation.
3. pca_plot standardization add elpsilon to prevent zero std.
4. c4.5 selection add add elpsilon to prevent zero std.
5. tuner name for print.
6. add argument: return=True to function fit in sklearn estimator wrapper.
7. modified the scorer sparser for roc_auc's multiple response method
8. wrapper's argument name was wrong.

description:
1. task support:
\ data_overview | selection | tuner | summary | Pine
binary | v v v v v
multi | v v v v v
regression | v v v v v

2. selection revised.
3. Pine scorer revised: polymorphism for regression and classification.
4. Pine do_stage revised: polymorphism for regression and classification
5. plot functions revised: sturctural adjustment.
6. tuner metric sparser reviesd: kargs supported now. Nearly full support sklearn scoerers

bug:
1. feature bagging:
- contradicts to log domain in selection method (the negative output of pca)
- the variance of feature will change during bagging (variance matters)
2. Lasso logistic regression in grid search need to to modified. In temporary using binary search.

fixed:
1. C++ 14.0 issue can be solved by: https://blog.csdn.net/Lc_001/article/details/129195335
2. IO didn't reset repeatued index.

On Going:
0. A reliable tutorial. (including mac) / Document reference
1. Pine monitor, progress bar and report. experiment setting visualization.

ToDo:
0. threshold tuner
1. report and diagnose, Pine pipeline graph.
2. using pretty, beautiful, good-looking, precise packages:
a. pca
b. The only OPLS da reliable(compare to others), alive, python implement
https://github.com/Omicometrics/pypls?tab=readme-ov-file
3. barutoSHAP
n. add parameter dict(json or yaml-like)
n. interactive interface or GUI (maybe nicegui/ plotly)

description:
1. Tutorial
2. task support:
\ data_overview | selection | tuner | summary | Pine
binary | v v v v v
multi | v v v v
regression | v v v

3. tuner metrics easy mode.


fixed:
1. RF too slow: In some reason(still unknown) the sklearn rf using "thread" mode as joblib.parallel_backend. That leads to single processing. Fixed by manually set joblib.parallel_backend to loky.
2. DT random_state
3. Lasso in essemble selector returns all zero score.

description:
1. add Pine experiment manager. it's a tool to select the best pipeline.
a. example_Pine.ipynb: the example of how to use Pine.
2. tuner's parameters change.
3. IO move from preprocessing to PineBioML
4. report.classification_scores: a tool to compute common classification scores
5. document update.

bug:
1. feature bagging:
- contradicts to log domain in selection method (the negative output of pca)
- the variance of feature will change during bagging (variance matters)
2. Lasso logistic regression in grid search need to to modified. In temporary using binary search.
3. Issue about installation of pandas lacking of Microsoft Visual C++ 14.0:
https://learn.microsoft.com/zh-tw/cpp/windows/latest-supported-vc-redist?view=msvc-170
4. RF tuner too slow.
5. elasticnet and DT have random behavier
On Going:
0. verbosity or or progress bar for Pine
1. A reliable tutorial. (including mac)
2. One click everything (almost)
ToDo:
1. report and diagnose, Pine pipeline graph.
2. Document reference
3. using pretty, beautiful, good-looking, precise packages:
a. pca
b. The only OPLS da reliable(compare to others), alive, python implement
https://github.com/Omicometrics/pypls?tab=readme-ov-file
n. add parameter dict(json or yaml-like)
n. interactive interface or GUI (maybe nicegui/ plotly)

description:
update readme.
bug:
1. bagging:
- contradicts to log domain in selection method (the negative output of pca)
- the variance of feature will change during bagging (variance matters)
to do:
1. add parameter dict(json or yaml-like)
2. auto ml (auto sklearn, mljar, H2O)
3. document (on going)
4. report
n. interactive interface or GUI

description:
1. adjusting missing value in volcano plot. In the past, we use imputed value to construct volcano plot, which may lead to misestimate the group size.

On going:
1. normalize issue.
2. tutorial for mac
3. tutorial for general using
4. auto ml (auto sklearn, mljar, H2O)

bug:
1. bagging:
- contradicts to log domain in selection method (the negative output of pca)
- the variance of feature will change during bagging (variance matters)
2. logistic regression + L1 grid search need to to modified. In temporary using binary search.
to do:
1. add parameter dict(json or yaml-like)
2. report
n. interactive interface or GUI (maybe nicegui/ plotly)

2024/6/8 for_present
description:
1. remove h2o futrue tabulate from requirment
2. set random states (need to check)
3. boosting mehtod sub sampling.