3
$\begingroup$

I am trying to use Support Vector Regression on a (neurophysiological) dataset where the position of points on a circular manifold in N dimensions is correlated with a circular variable (phase of an oscillation of a separate system).

Before starting, I am trying to solve this on a dummy dataset. I have dots from a circle $X$, plus some noise, and each is associated with a circular variable $y$ which is the angle of the dots, plus some noise (this recapitulates reasonably well my dataset, which has just more dimensions and a less regular circular manifold - just stating this in case there might be better approaches than what I am describing below).

from matplotlib import pyplot as plt
import numpy as np
from sklearn import svm
from sklearn.model_selection import train_test_split

n_points = 1000
rad_sd, phase_sd = 0.2, 0.5

angles = np.random.rand(n_points) * 2*np.pi
X = np.array([np.cos(angles)+np.random.randn(n_points)*rad_sd, 
              np.sin(angles)+np.random.randn(n_points)*rad_sd]).T
y = angles + np.random.randn(n_points) * phase_sd

f, ax = plt.subplots(figsize=(3, 2.5))
im = plt.scatter(X[:, 0], X[:, 1], c=y)
ax.set(xlabel="$x_1$", ylabel="$x_2$")
plt.colorbar(im, label="phase $y$")
plt.tight_layout()

enter image description here

I am then fitting the vanilla SVR to the data:

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.5, random_state=42)

regr = svm.SVR()
regr.fit(X_train, y_train)
predicted = regr.predict(X_test)

f, ax = plt.subplots(figsize=(4, 3))
ax.scatter(y_test, predicted, lw=0)
ax.set(ylabel="Predicted $y$", xlabel="Test $y$")

enter image description here

Which looks very good, apart from having a problematic transition from $0$ to $2\pi$.

I guess that what I would need it to somehow make the optimiser know that this is a circular variable, and that distances for the fit should be calculated accordingly. It is not clear to me however how could I add a step to the pipeline/modify the loss function to get that distance calculated. I get the feeling that any kind of optimisation I try in scikit-learn would need to deal with this issue

Any clue? Thank you very much in advance for any pointer that could help me with this problem!

$\endgroup$

1 Answer 1

2
$\begingroup$

Found a solution thanks to other posts here. It is possible to fit to more than one $y$ the model; so I regressed $\sin(y)$ and $\cos(y)$ in a MultiOutputRegressor model and took the $\arctan2$ of the solutions:

from sklearn.multioutput import MultiOutputRegressor

y_train_multi = np.array([np.sin(y_train), np.cos(y_train)]).T
y_test_multi = np.array([np.sin(y_test), np.cos(y_test)]).T

regressor = MultiOutputRegressor(SVR(kernel='rbf', C=1e3, gamma=0.1))

regressor.fit(X_train, y_train_multi)
prediction = regressor.predict(X_test)

angle_predicted = np.arctan2(prediction[:, 0], prediction[:, 1])

f, ax = plt.subplots(figsize=(4, 3))
ax.scatter(y_test, angle_predicted, lw=0)
ax.set(ylabel="Predicted $y$", xlabel="Test $y$")

enter image description here

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.