Troubleshooting¶

Common issues and how to resolve them.

`knee` is `None`¶

If kl.knee returns None, no knee point was detected. Common causes:

Wrong `curve` or `direction`¶

This is the most common issue. If curve or direction don't match your data, the algorithm won't find a knee.

from kneed import KneeLocator, DataGenerator

x, y = DataGenerator.figure2()

# Wrong direction — returns None
kl = KneeLocator(x, y, curve="concave", direction="decreasing")
print(kl.knee)  # None

# Correct direction
kl = KneeLocator(x, y, curve="concave", direction="increasing")
print(kl.knee)  # 0.222

Fix: Use find_shape() to auto-detect, or plot your data to determine the correct values.

Data is too flat or linear¶

If the data has no meaningful curvature (e.g., a straight line), there's no knee to detect.

import numpy as np

x = np.arange(10)
y = x  # Perfectly linear — no knee

kl = KneeLocator(x, y, curve="concave", direction="increasing")
print(kl.knee)  # None

Sensitivity (`S`) is too high¶

A high S value makes the algorithm more conservative. Try lowering it:

# Too conservative
kl = KneeLocator(x, y, curve="convex", direction="decreasing", S=200)
print(kl.knee)  # May be None or late

# More sensitive
kl = KneeLocator(x, y, curve="convex", direction="decreasing", S=1)
print(kl.knee)  # Earlier knee detection

See the S parameter guide for details.

Wrong Knee Detected¶

If a knee is detected but it's not where you expect:

Try online mode¶

Offline mode returns the first knee found, which may not be the most significant:

kl = KneeLocator(x, y, curve="convex", direction="decreasing", online=True)
print(kl.knee)  # Often a better result

Adjust sensitivity¶

The S parameter controls how early or late a knee is detected. Experiment with different values:

for s in [1, 3, 5, 10]:
    kl = KneeLocator(x, y, curve="convex", direction="decreasing", S=s)
    print(f"S={s}: knee={kl.knee}")

Try polynomial interpolation¶

If your data is noisy, polynomial interpolation can smooth out the curve:

kl = KneeLocator(
    x, y, curve="convex", direction="decreasing",
    interp_method="polynomial", polynomial_degree=7,
)

See the interp_method guide for details.

Debugging with Internal Data¶

KneeLocator exposes intermediate calculation data that can help you understand what's happening:

kl = KneeLocator(x, y, curve="concave", direction="increasing")

# Plot the normalized difference curve
import matplotlib.pyplot as plt

plt.plot(kl.x_difference, kl.y_difference, label="difference curve")
plt.plot(kl.x_difference_maxima, kl.y_difference_maxima, "ro", label="maxima")
plt.plot(kl.x_difference_minima, kl.y_difference_minima, "go", label="minima")
plt.legend()
plt.title("Difference Curve")
plt.show()

Key attributes for debugging:

Attribute	Description
`x_normalized`, `y_normalized`	Normalized input data
`x_difference`, `y_difference`	The difference curve
`maxima_indices`	Indices of local maxima on the difference curve
`minima_indices`	Indices of local minima on the difference curve
`Tmx`	Threshold values for each maximum

scipy `interp1d` Deprecation Warning¶

If you see a deprecation warning about scipy.interpolate.interp1d, you can suppress it or switch to polynomial interpolation:

# Option 1: Use polynomial interpolation instead
kl = KneeLocator(x, y, curve="concave", interp_method="polynomial")

# Option 2: Suppress the warning
import warnings
warnings.filterwarnings("ignore", category=DeprecationWarning, module="scipy")
kl = KneeLocator(x, y, curve="concave")

Plotting Errors¶

If you get a ModuleNotFoundError when calling plot_knee() or plot_knee_normalized(), install matplotlib:

pip install kneed[plot]

Matplotlib is an optional dependency and is only required for the plotting methods.