maxhalford / xam

:dart: Personal data science and machine learning toolbox

xam

xam is my personal data science and machine learning toolbox. It is written in Python 3 and stands on the shoulders of giants (mainly pandas and scikit-learn). It loosely follows scikit-learn's fit/transform/predict convention.

Installation

• Install Anaconda for Python 3.x >= 3.5
• Run pip install git+https://github.com/MaxHalford/xam --upgrade in a terminal

⚠️ Because xam is a personal toolkit, the --upgrade flag will install the latest releases of each dependency (scipy, pandas etc.). I like to stay up-to-date with the latest library versions.

Table of contents

Usage example is available in the docs folder. Each example is tested with doctest.

• Ensembling
  • Groupby model
  • LightGBM with CV
  • Stacking
  • Stacking with bagged test predictions
• Exploratory data analysis (EDA)
  • Feature importance
• Feature extraction
  • Bayesian target encoding
  • Combining features
  • Count encoding
  • Cyclic features
• Feature selection
  • Forward-backward selection
• Linear models
  • AUC regressor
• Model selection
  • Ordered cross-validation
• Natural Language Processing (NLP)
  • NB-SVM
  • Norvig spelling corrector
  • Top-terms classifier
• Pipeline
  • Column selection
  • Series transformer
  • DataFrame transformer
  • Lambda transformer
• Plotting
  • Latex style figures
• Preprocessing
  • Binning
  • Groupby transformer
  • One-hot encoding
  • Resampling
• Time series analysis (TSA)
  • Exponentially weighted average
  • Exponential smoothing
  • Frequency average forecasting
• Various
  • Datetime range
  • Next day of the week
  • Subsequence lengths
  • DataFrame to Vowpal Wabbit
  • Normalized compression distance
  • Skyline querying
  • Fuzzy duplicates

Other Python data science and machine learning toolkits

• fastai/fastai
• Laurae2/Laurae
• rasbt/mlxtend
• reiinakano/scikit-plot
• scikit-learn-contrib
• zygmuntz/phraug2

License

The MIT License (MIT). Please see the license file for more information.

xam's Issues

BayesianTargetEncoder can not be pickled

I tried pickle.dump the BayesianTargetEncoder after fit(), but get this error
AttributeError: Can't pickle local object 'BayesianTargetEncoder.fit..'

It is for sure a problem in deployment, if you can not store away a pre-trained encoder. Is this a known issue? Can we get a fix for it? Thanks

Error while pip install xam

It does install the package but when the import has to be done in the code, it reports an error saying that "ModuleNotFoundError: No module named 'repos'
After doing pip install repos, again the error shows up as:
ModuleNotFoundError: No module named 'fabric.api'

Feature extraction methods in sklearn pipeline

Hey, I'm facing an issue with implementing BayesianTargetEncoder and CycleTransformer in a pipeline that I've made. This might be because of my lack of experience but any help would still be appreciated.

The pipeline I currently have is

model = RandomForestRegressor(n_estimators = 1000, min_samples_split = 2, max_features = 'sqrt',
max_depth = 20, random_state = 42, bootstrap = True, oob_score = True, n_jobs = 4)

ss = StandardScaler()
bte = BayesianTargetEncoder()
ct = CycleTransformer()

rf = Pipeline(
    [
    ("transformer", ColumnTransformer
        (
            [
                ("scaler", ss, num_cols),
                ("bayesian_target_encoder", bte, cat_cols),
                ("cycle_transformer"), ct, times
            ]
        )
    ),
    ("reg", model) 
    ]
)  

and the issue that I am facing is that it says TypeError: 'CycleTransformer' object is not iterable

If I remove CycleTransformer and proceed without it, then I get this issue:

C:\py\lib\site-packages\xam\feature_extraction\encoding\bayesian_target.py in transform(self, X, y)
     73             raise ValueError('X has to be a pandas.DataFrame')
     74 
---> 75         for cols in self.columns:
     76 
     77             if isinstance(cols, list):

TypeError: 'NoneType' object is not iterable

I understand it says that X has to be a pandas dataframe but i'm not sure what the work around to this would be.
Thank you

smooth target encoding

Hi,

thanks a lot for this handy toolkit! May i make a minor observation? It would be useful to cast the transformed column types into np.float32 explicitly here, while sometimes for unclear to me reason the type remains 'category', while in other cases it is casted into a float

cannot save count encoder to pickle

Hi,
I have problem to save count encoder into pickle. It throw the error below :
AttributeError: Can't pickle local object 'CountEncoder.fit..'
thank you

meaning of prior_weight

there is a prior_weight param used in Bayesian Target Encoding and a reference to a page to Additive Smoothing. However, on that page, I don't see any reference to the prior weight parameter. Is there some kind of explanation what that parameter means?

Bayesian target encoding for regression

Hi, I read a paper which says you can apply Bayesian target encoding for regression also. Does your lib work for that? any sample code?

AUC regressor range value

is it normal if we get prediction using AUC regressor with the range not between 0-1? does predict method here similar to predict_proba ? I am working on binary classification.