Resampling Methods: SMOTE/ADASYN/Tomek

Resampling rebalances class distributions with data-level methods. SMOTE (Synthetic Minority Over-sampling Technique): for a minority sample $x_i$, randomly pick a k-nearest neighbor $x_{z_i}$ and interpolate along the segment to create $x_{new} = x_i + \lambda(x_{z_i} - x_i)$ with $\lambda \sim U[0,1]$; synthesizing new points reduces the overfitting caused by plain duplication. ADASYN (Adaptive Synthetic): the number of generated samples per minority point is proportional to how “hard” it is — points surrounded by more majority neighbors (closer to the boundary) get more synthetic samples, shifting the synthesis focus to the hard region. Tomek Links (cleaning): if $x$ and $y$ are mutual nearest neighbors from different classes, the pair is a Tomek Link (boundary pair); removing the majority-side sample cleans the boundary, often combined with SMOTE as SMOTE-Tomek: first synthesize boundary samples, then clean the overlapping boundary zone.

class imbalance with few minority samples and continuous numeric features (interpolation-able); interviews ask about the SMOTE formula, why it fails on text/high-dimensional sparse features, and why random duplication is bad.

random oversampling duplicates samples, so the model overfits on small repeated regions and its decision region is too concentrated; SMOTE interpolates in local feature space to “create” new samples and enlarges the minority decision region; but SMOTE does not address majority noise and boundary overlap (it may worsen overlap), while Tomek Links pushes the boundary away from overlapping zones — the combination (SMOTE-Tomek) mitigates both overfitting and boundary confusion.

Write the SMOTE formula $x_{new} = x_i + \lambda(x_{z_i} - x_i)$; why is $\lambda \in [0,1]$ and why does it reduce duplication overfitting?

SMOTE's limits: why is it unsafe for high-dimensional/text/sparse features? Is it sensitive to label noise?

ADASYN vs SMOTE? Where is the adaptivity (what drives the number of samples generated)?

What is a Tomek Link? Oversampling or cleaning? Why combine it with SMOTE?

Should SMOTE run before or after the train/validation split? What leakage occurs if it runs before?