Skip to content

Latest commit

 

History

History
197 lines (145 loc) · 7.74 KB

README.md

File metadata and controls

197 lines (145 loc) · 7.74 KB
Raw

Build Status Coverage Status pub package Gitter Chat

ml_preprocessing

Data preprocessing algorithms

What is data preprocessing?

Data preprocessing is a set of techniques for data preparation before one can use the data in Machine Learning algorithms.

Why is it needed?

Let's say, you have a dataset:

    ----------------------------------------------------------------------------------------
    | Gender | Country | Height (cm) | Weight (kg) | Diabetes (1 - Positive, 0 - Negative) |
    ----------------------------------------------------------------------------------------
    | Female | France  |     165     |     55      |                    1                  |
    ----------------------------------------------------------------------------------------
    | Female | Spain   |     155     |     50      |                    0                  |
    ----------------------------------------------------------------------------------------
    | Male   | Spain   |     175     |     75      |                    0                  |
    ----------------------------------------------------------------------------------------
    | Male   | Russia  |     173     |     77      |                   N/A                 |
    ----------------------------------------------------------------------------------------

Everything seems good for now. Say, you're about to train a classifier to predict if a person has diabetes. But there is an obstacle - how can it be possible to use the data in mathematical equations with string-value columns (Gender, Country)? And things are getting even worse because of an empty (N/A) value in the Diabetes column. There should be a way to convert this data to a valid numerical representation. Here data preprocessing techniques come to play. You should decide, how to convert string data (aka categorical data) to numbers and how to treat empty values. Of course, you can come up with your unique algorithms to do all of these operations, but there are a lot of well-known techniques for doing all the conversions.

The aim of the library is to give data scientists, who are interested in Dart programming language, these preprocessing techniques.

Prerequisites

The library depends on DataFrame class from the repo. It's necessary to use it as a dependency in your project, because you need to pack data into DataFrame before doing preprocessing. An example with a part of pubspec.yaml:

dependencies:
  ...
  ml_dataframe: ^1.0.0
  ...

Usage examples

Getting started

Let's download some data from Kaggle - let it be amazing black friday dataset. It's pretty interesting data with huge amount of observations (approx. 538000 rows) and a good number of categorical features.

First, import all necessary libraries:

import 'package:ml_dataframe/ml_dataframe.dart';
import 'package:ml_preprocessing/ml_preprocessing.dart';

Then, we should read the csv and create a data frame:

final dataFrame = await fromCsv('example/black_friday/black_friday.csv', 
  columns: [2, 3, 5, 6, 7, 11]);

Categorical data

After we get a dataframe, we may encode all the needed features. Let's analyze the dataset and decide, what features should be encoded. In our case these are:

final featureNames = ['Gender', 'Age', 'City_Category', 'Stay_In_Current_City_Years', 'Marital_Status'];

One-hot encoding

Let's fit the one-hot encoder.

Why should we fit it? Categorical data encoder fitting - a process, when all the unique category values are being searched for in order to create an encoded labels list. After the fitting is complete, one may use the fitted encoder for the new data of the same source.

In order to fit the encoder, it's needed to create the instance of the Encoder class and pass the fitting data as an argument to the constructor, along with the features to be encoded:

final encoder = Encoder.oneHot(
  dataFrame,
  columnNames: featureNames,
);

Let's encode the features:

final encoded = encoder.process(dataFrame);

We used the same dataframe here - it's absolutely normal since when we created the encoder, we just fit it with the dataframe, and now is the time to apply the dataframe to the fitted encoder.

It's time to take a look at our processed data. Let's read it:

final data = encoded.toMatrix();

print(data);

In the output we will see just numerical data, that's exactly what we wanted to reach.

Label encoding

Another well-known encoding method. The technique is the same - first, we should fit the encoder and after that, we may use this "trained" encoder in some applications:

// fit encoder
final encoder = Encoder.label(
  dataFrame,
  columnNames: featureNames,
);

// apply fitted encoder to data
final encoded = encoder.process(dataFrame);

Numerical data normalization

Sometimes we need to have our numerical features normalized, which means we need to treat every dataframe row as a vector and divide this vector element-wise by its norm (Euclidean, Manhattan, etc.). To do so the library exposes Normalizer class:

final normalizer = Normalizer(); // by default Euclidean norm will be used
final transformed = normalizer.process(dataFrame);

Please, notice, that if your data has raw categorical values, the normalization will fail as it requires only numerical values. In this case, you should encode data (e.g. using one-hot encoding) before normalization.

Data standardization

A lot of machine learning algorithms require normally distributed data as their input. Normally distributed data means that every column in the data has zero mean and unit variance. One may reach this requirement using the Standardizer class. During the creation of the class instance, all the columns' mean values and deviation values are being extracted from the passed data and stored as fields of the class, in order to apply them to standardize the other (or the same that was used for the creation of the Standardizer) data:

final dataFrame = DataFrame([
  [  1,   2,   3],
  [ 10,  20,  30],
  [100, 200, 300],
], headerExists: false);

// fit standardizer
final standardizer = Standardizer(dataFrame);

// apply fitted standardizer to data
final transformed = standardizer.process(dataFrame);

Pipeline

There is a convenient way to organize a sequence of data preprocessing operations - Pipeline:

final pipeline = Pipeline(dataFrame, [
  toOneHotLabels(columnNames: ['Gender', 'Age', 'City_Category']),
  toIntegerLabels(columnNames: ['Stay_In_Current_City_Years', 'Marital_Status']),
  normalize(),
  standardize(),
]);

Once you create (or rather fit) a pipeline, you may use it further in your application:

final processed = pipeline.process(dataFrame);

toOneHotLabels, toIntegerLabels, normalize and standardize are pipeable operator functions. The pipeable operator function is a factory that takes fitting data and creates a fitted pipeable entity (e.g., Normalizer instance)