y_transform doesn't work with unevenly spaced x values about kneed HOT 4 OPEN

Hi @MitchellAcoustics, thanks for checking out the repo. The x and y values are interpolated and normalized before they are transformed, so it should handle the issue of unevenly spaced x value input. Does this example help answer any questions?

import numpy as np
import matplotlib.pyplot as plt

from kneed import KneeLocator

n = 50
x = np.geomspace(1, 256, num=n)
y = np.linspace(0, 1, num=n)


plt.scatter(x, y)

Kneed seems to find a reasonable knee point.

kneed = KneeLocator(x, y, curve='concave', direction='increasing')

print(kneed.knee)
46.883891195871875

kneed.plot_knee()

from kneed.

Hi Thanks for this great algorithm. However I am struggling to get it work when x becomes log(x). For example for your same above example, if I did below, the found knee point is at end point which is not correct.

import numpy as np
import matplotlib.pyplot as plt

from kneed import KneeLocator

n = 50
x = np.linspace(1, 256, num=n)
x1=np.log(x)
y = np.linspace(0, 1, num=n)

kneed = KneeLocator(x1, y, curve='convex', direction='increasing')

from kneed.

Hi

I keep playing with interpolation method, polynomial degree and sensitivity parameters and I found the results change a lot by changing these parameters. and the best result I got is still not optimal. Now my question is how do we select the parameters automatically? I am using this algorithm to detect elbow point automatically for any curve. Could you please help ? Thank you so much.

import numpy as np
import matplotlib.pyplot as plt

from kneed import KneeLocator

n = 50

x = np.linspace(1, 256, num=n)
x1=np.log(x)
y = np.linspace(0, 1, num=n)
kneed = KneeLocator(x1, y, S=1,curve='convex', direction='increasing',online=True,interp_method="polynomial",polynomial_degree=3)

print(kneed.knee)
kneed.plot_knee_normalized()
kneed.plot_knee()

from kneed.

If I may chime in, I think the issue is that only the y value is transformed.
In knee_locator.py, there is a transform_y function that handles the entire transform. I think this could be fixed by adding a transform_x function that handles an x transformation.

    def transform_x(x: Iterable[float], direction: str, curve: str) -> float:
        """transform x to concave, increasing based on given direction and curve"""
        # convert elbows to knees
        if direction == "decreasing" and curve == "concave":
            x = x.max() - x
        elif direction == "increasing" and curve == "convex":
            x = x.max() - x

       return x

    def transform_y(y: Iterable[float], direction: str, curve: str) -> float:
        """transform y to concave, increasing based on given direction and curve"""
        # convert elbows to knees
        if curve == "convex":
            y = y.max() - y

       return y

Then changing this part of the code

 # Step 2: normalize values
        self.x_normalized = self.__normalize(self.x)
        self.y_normalized = self.__normalize(self.Ds_y)

        # Step 3: Calculate the Difference curve
        self.y_normalized = self.transform_y(
            self.y_normalized, self.direction, self.curve
        )

        self.x_normalized = self.transform_x(
            self.x_normalized, self.direction, self.curve
        )
        # normalized difference curve

from kneed.