---

classificator : a convenience wrapper for classification methods in Python's scikit-learn

Latest Release
License
Build Status

About

The classificator is a wrapper around scikit-learn’s classification suite, whose primary function is to normalize and automate the process by which common machine learning classification models are trained and validated. This is specifically in the context of text feature spaces, which require NLP techniques for preprocessing and vectorization. There is an additional focus on group cross validation, where labeled groups exist as data features, and variance is known to be smaller within groups.

Companion API

The classificator package has a companion repository containing scripts to build and host an API and associated browser based UI called the classificator-server.

Features

Load a dataset

• User points to a dataset locally or in Amazon S3
• User defines simple data specs
• User defines feature columns, label column and group columns

Load a configuration

• User specifies preprocessing options, model choices, and hyperparameter grids per model choice
• User specifies training session meta parameters

Train and validate various model choices

• Builds pre-processing pipeline to vectorize or encode text and standardize and impute numeric feautres
• Uses common cross validation and grid search techniques, and chooses the best model based on a user supplied scoring option
• User defines output location for training and validation reports and pickled model object

Installation

Source Code: https://github.com/denver1117/classificator
Package Index: https://pypi.python.org/pypi/classificator

Binary installers for the latest released version are available at the Python package index:

# PyPI
pip install classificator

Quick Start

• See scripts/test.py for a simple example with a stock data set and configuration
• Here is a sample configuration (config.json):

{
    "data_specs": {
        "loc": "doc/data/tickers.tsv", 
        "label_column": "CapBucket_Label", 
        "feature_columns": ["Industry", "Sector"], 
        "feature_methods": ["vectorize", "encode"], 
        "group_columns": [], "sep": "\t",
        "out_loc": "tmp",
        "model_name": "model1.obj",
        "run_name": "my_model_run"
    },
    "meta_specs": {
        "n_jobs": 1, "bypass": 0, "split_ratio": 0.9, 
        "random_seed": 15, "k": 5, "score": "Accuracy", 
        "train_final": 1
    },
    "vectorizer": {
        "model": "TF-IDF Vectorizer", 
        "args": {
            "ngram_range": [1,2], "min_df": 0.001, 
            "decode_error": "ignore", 
            "analyzer": "word"
        }
    },
    "selector": {
        "model": "Chi-Squared", 
        "grid": {"alpha": [0.05, 0.10]}
    },
    "pre_processors": {"methods": [[]]},
    "classifiers": {
        "Decision Tree": {"min_samples_split": [2,0.01]},  
        "Logistic Regression": {"C": [1, 10]}
    }
}

• Running the following:

    from classificator.classify import Classificator

    clf = Classificator(config_loc="config.json")
    clf.choose_model()

will build out a text pre-processing pipeline using a TF-IDF Vectorizer (with the supplied parameters) for the column Industry and a OneHotEncoder for the column Sector in order to predict classes CapBucket_Label, with data located here: doc/data/tickers.tsv. It will send logging, validation metrics, and a pickled model object here: tmp. It will select features from the vectorized data using Chi-Squared feature selection, attemtping multiple alpha values in grid search. It will run separate grid search processes for a Decision Tree and for a Logistic Regression using the same pre-processing pipeline. The best model will be chosen by the best Accuracy score among parameter maps across classifier choices.

Authors

Evan Harris - Initial work began at Return Path

License

This project is licensed under the MIT License - see the LICENSE file for details