ranking-kNN-imputation

An imputation method based on ranking and nearest neighbor (kNN) as described in:

Beretta L, Santaniello A. "Nearest neighbor imputation algorithms: a critical evaluation". BMC Med Inform Decis Mak. 2016 Jul 25;16 Suppl 3:74. https://doi.org/10.1186/s12911-016-0318-z

The method imputes missing data combining a non linear feature selection and a nearest neighbor 
    imputation algorithm.

The algoroithm discards features that may be irrelevant or noisy for distance determination and hence for 
    NN determination.

The variable with missing instances is used as a class feature and ranking algorithms are used to select 
    the variables that best correlate with the "class". The search is performed consdiering complete cases 
    and works on a normalized (by span) dataset.

Compared to the original algorithm described by Beretta et al., the code allows the automated selection of 
    relevant features and optimizes the number k neighbors used for imputation.

The elbow method is used to automatically select the parameters. Alternatively, the number of features and 
    k neighobrs can be manually set.

Non-linear feature ranking algorithm are recommended to select relevant features. By default the algorithm works using ReliefF-based methods. The sklearn package Rebate (https://github.com/EpistasisLab/scikit-rebate) needs to be installed.

Installation

pip install rkNN_imputer

Usage

The algorithm is designed to be integrated directly into scikit-learn machine learning workflows. ReliefF algorithms are used as feature selection step to optimize the search of closest neighbors.

import numpy as np
from rkNN_imputer.rkNN import rkNN

dummy = np.array ((["0.2", "0.4", "5", "0.4", "6"],
                ["?", "0.4", "8", "0.5", "8"],
                ["0.1", "0.8", "3", "0.2", "6"],
                ["0.3", "0.2", "?", "0.1", "7"],
                ["0.2", "0.9", "3", "0.2", "4"]))

imp = rkNN ()
imp.fit (dummy)
imputed_dummy = imp.transform (dummy)

# show orignal and imputed dataset
print (dummy)
print ("")
print (imputed_dummy)

Usage details

class rkNN_imputer.rkNN rkNN (estimator=MultiSURF (), k=None)

_Parameters:

estimator: object, optional (default=MultiSURF)

(The base estimator that ranks features for the selection of the most informative attributes for nearest neighbour determination. Any ranking algorithm is suitable. Preferred choices, ReliefF-based algorithms (as in the Rebate package)).

k: integer, optional (default=None)

(The number of k-neighbours to consider to impute the missing value. If = None the number is autoamtically optimized. For values of k > number of instances, the method is equal to mean imputation).

_Attributes:

fit (X, Y=None)

Fit the imputer on X.

_Parameters:

X : {array-like, sparse matrix}, shape (n_samples, n_features)

(Input data, where n_samples is the number of samples and n_features is the number of features)

_Returns:

self : Imputer

transform(X)

Impute all missing values in X.

_Parameters:

X : {array-like, sparse matrix}, shape = (n_samples, n_features)

(The input data to complete, where missing data are coded as ?).

Returns: X: {array-like, sparse matrix}, shape = (n_samples, n_features)

(The transformed (imputed) dataset).