Tomek Links in Machine Learning

Tomek Links: what they remove and when to try them

Tomek links are pairs of examples from different classes that are each other’s nearest neighbors. Removing selected members of these pairs is a data-cleaning approach, not a guarantee of better predictions: boundary examples can be valid observations.

In binary classification, imbalanced-learn’s default removes the majority-class member. Using sampling_strategy="all" removes both members. See the TomekLinks documentation.

Before trying it: split off evaluation data, fit scaling on training data, and apply resampling only to training rows. During cross-validation, resample inside each training fold using an imbalanced-learn pipeline. Compare against an unchanged baseline using precision and recall.

Learn the evaluation workflow first. Read the free Starter Kit excerpt on test sets to preview the €19 handbook. No email required.

Why Use Tomek Links?

• Improve class separation
• Reduce noise near boundaries
• Test whether minority-class performance improves
• Compare generalization against an unchanged baseline

They are widely used in machine learning, especially when dealing with imbalanced datasets.

How Tomek Links Work

1. Find nearest neighbors
2. Identify cross-class mutual neighbors
3. Mark Tomek Links
4. Remove majority samples

Python Implementation

from collections import Counter
from sklearn.datasets import make_classification
from imblearn.under_sampling import TomekLinks
from sklearn.model_selection import train_test_split

X, y = make_classification(
    n_samples=1000,
    weights=[0.9, 0.1],
    n_features=2,
    n_informative=2,
    n_redundant=0,
    random_state=42
)

X_train, X_test, y_train, y_test = train_test_split(
    X, y, stratify=y, test_size=0.2, random_state=42
)
print("Before:", Counter(y_train))

tl = TomekLinks()
X_res, y_res = tl.fit_resample(X_train, y_train)

print("After:", Counter(y_res))

When to Use Tomek Links in Machine Learning

• Imbalanced classification problems
• Noisy datasets
• Overlapping distributions

Comparison vs SMOTE

SMOTE creates synthetic samples, while Tomek Links remove selected cross-class nearest neighbors, which are not necessarily noise.

They can be combined, but validate the result. SMOTETomek uses a TomekLinks cleaner with sampling_strategy=’all’ by default, so it can remove members of either class.

Combine SMOTE + Tomek Links

from imblearn.combine import SMOTETomek

smt = SMOTETomek(random_state=42)
X_res, y_res = smt.fit_resample(X_train, y_train)

Best Resources to Learn Machine Learning

To go deeper into machine learning, check these curated resources:

• Machine Learning Books
• Machine Learning Online Courses

FAQ

What is a Tomek Link?

A pair of samples from different classes that are each other’s nearest neighbors.

Do they remove minority samples?

In binary classification the default removes majority-class members; sampling_strategy=’all’ removes both members.

Are they better than random undersampling?

Not always. Tomek Links target boundary pairs; random undersampling removes examples without that condition. Compare them on the same validation data.

Final Thoughts

Tomek Links in Machine Learning are a targeted cleaning method worth testing on overlapping classes. Keep validation and test data unchanged, and retain the method only if it improves the metrics that matter for your task.