auto-modelling

Auto-modelling is a convenient library to train and tune machine models automatically.

Its main features include the following:

1. preprocessing columns in all datatypes. (numeric, categorical, text)
2. train machine models and tune parameters automatically.
3. return top n best models with optimized parameters.
4. Apply stacking technique to combine the n best models returned by the repo or self-determined fitted models together to get an even better result.

The machine learning models include the following:

• Classification:
  • ExtraTreesClassifier
  • RandomForestClassifier
  • KNeighborsClassifier
  • LogisticRegression
  • XGBClassifier
• Regression:
  • ExtraTreesRegressor
  • GradientBoostingRegressor
  • AdaBoostRegressor
  • DecisionTreeRegressor
  • RandomForestRegressor
  • XGBRegressor
• Stack:
  • for classify: LogisticRegression
  • for regression: LinearRegression

reference: https://github.com/EpistasisLab/tpot/blob

Installation

pip install auto-modelling

Usage Example

from auto_modelling.classification import GoClassify
from auto_modelling.regression import GoRegress
from auto_modelling.preprocess import DataManager
from auto_modelling.stack import Stack

# preprocessing data
dm = DataManager(directory = 'preprocess_tools')
train, test = dm.drop_sparse_columns(x_train, x_test)
train, test = dm.process_data(x_train, x_test)
# the encoders are stored in the directory called data_process_tools.

# use the same processing tools to process new data
predict_data = dm.process_predict_data(predict_x)
# predict_x should have the same format as x_train/x_test

# classification
clf = GoClassify(n_best=1)
best = clf.train(x_train, y_train)
y_pred = best.predict(x_test)

# regression
reg = GoRegress(n_best=1)
best = reg.train(x_train, y_train)
y_pred = best.predict(x_test)

# get top 3 best models
clf = GoClassify(n_best=3)
bests = clf.train(x_train, y_train)
y_preds = [m.predict(x_test) for m in bests]

# Stack top 3 best models
stack = Stack(n_models = 3)
level_0_models, level_1_model = stack.train(x_train, y_train, x_test, y_test)

There are examples test.py and sample.py in the root directory of this package. run python test.py/python sample.py.

Development Guide

• Clone the repo

• Create the virtual environment

mkvirtualenv auto
workon auto
pip install requirements.txt

if you have issues in installing xgboost reference: https://xgboost.readthedocs.io/en/latest/build.html# https://www.ibm.com/developerworks/community/blogs/jfp/entry/Installing_XGBoost_on_Mac_OSX?lang=en

Note

• TO DO: Feature selection, evaluation metricss

Thoughts

• Ideally, any dataframe being throw into this repo, it should be processed.

1. pre-processing

   • drop column that have too many null(Done)
   • fill na for both numeric and non-numeric values(Done)
   • encoded for non-numeric values(Done)
   • scale values if needed
   • balance the dataset if needed

2. model-training

   • mode = classification, regression, auto(Done)
   • split data-set
   • tuning parameters and model selection (Done)
   • feature selection
   • return a model with parameters, columns and a script to process x_test(Done)
   • stacking with customized fitted models (Done)

3. model-evaluation

Other reference

Packaging your project