bartekpog / modelcreator Goto Github PK

Add a progress bar showing the status of learning.

This task requires writing custom GridSearchCV function.

I found that this may be achieved this way. I also found this lbrary which seems to be an easy to use progress bar module.

We shall give the user a possibility to choose whether to make a quick best model comparison or a robust, time-consuming analysis. It should influence the number of parameters and models analyzed in a Grid Search. It should be modifiable as an initial machine state, or later as a machine object parameter.

Test automodel with next datasets. Propositions:

• Titanic (kaggle)
• Chosen datasets from sklearn sklearn datasets
• First we can start with the "Toy datasets", then move on to the more complicated "Real world datasets"

Problem

Dask[dataframe] is now an optional dependency in dask- so it does not install automatically. Now after installing the modelcreator package user has to install it by himself.

Solution

Add dask[dataframe] dependency with the same version as dask to the setup.py.

We have to handle class imbalance. One of the methods is multiplying the rows with rare classes to equalize the class frequency in train dataset.

Add the proper method and expand the interface to handle this feature.

Plan first version of the visualization/report that will be generated after running the program. Suggested format is HTML. Can be divided into two separated tasks if necessary - visualization and report.

We need more imbalanced datasets for testing purposes. Find those and send links to those here.

What are the various ways to handle the imbalanced datasets?

One of them is multiplying rare classes, but are there other ways?

Translate all current issues to English.
Issues content shall not differ from the repository language. Let's keep things clean.

Add other and more various models to learning set.

The crucial ones are SVM models - both for regression and classification with multiple kernels and BalancedRandomForestClassifier from imblearn library.

Currently, all classification models are compared by their prediction accuracy. It is acceptable in some cases, but in some imbalanced datasets it may lead to poor grid search decisions and not choosing the best model.

Give user a chance to override the default accuracy metrics by others - for example, those already supported by Grid Search in sklearn library.