|
import streamlit as st |
|
import numpy as np |
|
import matplotlib.pyplot as plt |
|
|
|
st.subheader("K nearest neighbor (KNN) classification") |
|
|
|
st_col = st.columns(1)[0] |
|
|
|
K = st.slider('Number of nearest neighbors (K)', min_value=1, max_value=10, value=5, step=1) |
|
|
|
from sklearn.neighbors import KNeighborsClassifier as KNN |
|
from sklearn.model_selection import cross_val_score |
|
|
|
from sklearn.datasets import make_blobs |
|
X, y = make_blobs(n_samples=1000, centers=3, n_features=2, cluster_std=6, random_state=42) |
|
ntrain = 100 |
|
|
|
x_train = X[:ntrain] |
|
y_train = y[:ntrain] |
|
x_test = X[ntrain:] |
|
y_test = y[ntrain:] |
|
|
|
knn = KNN(n_neighbors=K) |
|
knn.fit(x_train, y_train) |
|
plt.figure() |
|
|
|
x = np.linspace(np.min(x_test[:, 0]), np.max(x_test[:, 0]), 200) |
|
y = np.linspace(np.min(x_test[:, 1]), np.max(x_test[:, 1]), 200) |
|
xx, yy = np.meshgrid(x, y) |
|
xy = np.c_[xx.ravel(), yy.ravel()] |
|
|
|
y_predicted = knn.predict(xy) |
|
|
|
plt.pcolormesh(xx, yy, y_predicted.reshape(200, 200), cmap='jet', alpha=0.2) |
|
y_unique = np.unique(y_train) |
|
markers = '*x+' |
|
colors = 'bgr' |
|
for i in range(len(y_unique)): |
|
plt.scatter(x_train[y_train == y_unique[i], 0], |
|
x_train[y_train == y_unique[i], 1], |
|
marker=markers[i], |
|
c=colors[i]) |
|
|
|
|
|
with st_col: |
|
st.pyplot(plt) |
|
|
|
hide_streamlit_style = """ |
|
<style> |
|
#MainMenu {visibility: hidden;} |
|
footer {visibility: hidden;} |
|
subheader {alignment: center;} |
|
</style> |
|
""" |
|
st.markdown(hide_streamlit_style, unsafe_allow_html=True) |
|
|
|
st.markdown(""" |
|
There are several points to note on the effect of K on the quality of model fit: |
|
* Models with extremely small values of K learn the local patterns and do not generalize well thus they have a high variance or overfitting effect. |
|
* Models with extremely high values of K suffer from averaging effect over the entire space and thus do not do well even on the train points. This is known as a high bias or underfitting effect. |
|
""") |
