# What is the interpretation of sklearn's linear perceptron coefficients?

I'm stumped as to why this example doesn't do a better job fitting the data, I suspect it has to do with my interpretation of the perceptron object's coefficients. Note that I'm interested in the sklearn object's implementation, not necessarily the math behind a perceptron.

It may help to know the example below yields:

perc.coef_ = array([[ -34. , -681.9,  222. ]])


import seaborn as sns
from sklearn.linear_model import Perceptron
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.backends.backend_pdf import PdfPages

# clean data
df.dropna(inplace=True)
df_no_torg = df.loc[df['island'] != 'Torgersen', :]

# extract data
x = df_no_torg.loc[:, ('bill_length_mm', 'flipper_length_mm')].to_numpy()
y = df_no_torg.loc[:, 'island'].to_numpy() == 'Biscoe'
# manual bias term (one less thing to consider)
x = np.hstack((np.ones((x.shape[0], 1)), x))

# init / fit perceptron
perc = Perceptron(fit_intercept=False)
perc.fit(x, y)

def get_boundary(x):
w = perc.coef_[0]
m = - w[1] / w[2]
b = - w[0] / w[2]
return m * x + b

# plot results
x = np.linspace(30, 60, 100)
b = get_boundary(x)
with PdfPages('penguin_scatter_bound.pdf') as pdf:
plt.figure(figsize=(10, 5))
plt.plot(x, b, label='boundary')
sns.scatterplot(data=df_no_torg, x='bill_length_mm', y='flipper_length_mm', hue='island')
pdf.savefig()


The problem here lies in the hyperparameters. perc.n_iter_ is 6 after running your code; according to the defaults in the api, the default value for n_iter_no_change is 5, so it looks like there was no improvement for the first five iterations and training simply terminated (despite the documentation also saying that n_iter_no_change only matters when early_stopping is True).

By playing around with the number of iterations and step size, I was able to get a much better fit (perc.fit(x, y, tol=None, max_iter=20000)).