neighbors-classifier

Fast, simple, but well-performing classifier based on weighted distances between data points.

Prerequisites

• python >= 3.7.3
• numpy >= 1.17.0
• pandas >= 0.24.2
• sklearn >= 0.21.2
• matplotlib >= 3.0.3

Installing

I recommend downloading the newest anaconda3 version. Then install the following packages e.g. via

pip install numpy pandas sklearn matplotlib

Clone the repository e.g. with

git clone https://github.com/schneider-source/neighbors-classifier.git

Running example

An example of how to use the classifier and a performance comparison with sklearn.neighbors.RadiusNeighborsClassifier using the Iris data set is provided. First, cd into the neighbors-classifier directory. Then execute the script via

python3 example.py

from the command line or

%run example

from inside an IPython session. Follow the console output for performance scores and runtime comparison, which should give you something similar to the following.

Neighbors classifier example and comparison with 
sklearn.neighbors.RadiusNeighborsClassifier

PCA: EVR = [0.84136038 0.11751808], SUM = 0.9588784639918416

Test performance of classifiers for full dataset

NeighborsClassifier:
accuracy_score    0.9452

sklearn.neighbors.RadiusNeighborsClassifier:
accuracy_score    0.888

For the confusion matrices see figure images/full_dataset/cm.png.

My NeighborsClassifier should have slightly better performance
than sklearn.neighbors.RadiusNeighborsClassifier.

For feature space classification see figure images/full_dataset/classification.png.

Test performance of classifiers for asymmetric data
(using no setosa, all 50 versicolor, and only 20 versicolor samples)

NeighborsClassifier:
accuracy_score    0.943333

sklearn.neighbors.RadiusNeighborsClassifier:
accuracy_score    0.774167

For the confusion matrices see figure images/asymmetric/cm.png.

My NeighborsClassifier should have significantly better performance
than sklearn.neighbors.RadiusNeighborsClassifier.

For feature space classification see figure images/asymmetric/classification.png.

Compare execution time of classifiers

Runtime NeighborsClassifier = 1.047 sec
Runtime sklearn.neighbors.RadiusNeighborsClassifier = 10.793 sec
My classifier is 10.31 x faster than the sklearn classifier.

More importantly, look for the generated figures in directory 'images', which should look similar to the following.

Full dataset

Using all 50 samples from each of three species of Iris (Iris setosa, Iris virginica and Iris versicolor).

Asymmetric label frequency

Using no Iris setosa, all 50 Iris versicolor, and only the first 20 Iris versicolor samples.

License

See the LICENSE for details.