When does adding SMOTE samples make a classifier's decision boundary worse?

## The blind spot in the algorithm SMOTE picks a minority row, picks one of its k nearest **minority** neighbours, and interpolates between them. Nothing in that procedure consults the majority class. The method assumes that the space between two nearby minority points belongs to the minority — and when the classes are well separated, that assumption is fine. When the classes overlap, it fails in a specific way. Take two defective parts that are each other's nearest minority neighbours but sit on opposite sides of a dense cloud of good parts. The segment joining them cuts straight through that cloud, and every synthetic point drawn along it is a **positive label placed in the middle of negative data**. The learner is now being told that a region densely occupied by good parts should be predicted defective. Its boundary bulges outward to accommodate the contradiction. ## What that does to the metrics The predictable signature is **recall up, precision down**. Widening the minority region catches more true positives and a great many more false ones. Two operational consequences: - A single headline number can hide it entirely — a threshold-free summary may barely move while precision at the recall you actually operate at collapses. - The damage is concentrated exactly where decisions are hard, i.e. near the boundary, which is where the model's output is used to make the borderline calls that matter. ## Noise amplification SMOTE has no notion of a bad row. A mislabelled minority example, or a genuine freak, is treated as a seed like any other: it gets interpolated toward its neighbours, and its own neighbours interpolate toward it. One wrong row becomes a small cluster of wrong rows, and clusters are much harder for a learner to dismiss as noise than a lone outlier is. On rare-event data, where every minority row was expensive to obtain and label quality is often uneven, this matters. ## Where the risk is highest - **Heavy class overlap.** The features simply do not separate the classes; no synthetic point can fix that, and interpolation actively muddies it. - **Very few minority rows.** With 30 positives you are interpolating along a handful of segments; the synthetic cloud is a skeleton of the same 30 points, and the variance of everything downstream stays high. - **High dimensionality.** Distances concentrate, so the "nearest" neighbour is close to arbitrary and the interpolation direction carries little information. - **Aggressive ratios.** Pushing a 0.8% rate to 1:1 means the majority of the training set is manufactured, so the model is mostly fitting interpolation artefacts. ## The boundary-focused variants, and why they cut both ways Two well-known refinements target where the points go: - **Borderline-SMOTE** first classifies each minority row by how many of its m nearest neighbours (across the whole dataset) are majority. Rows with more than half but not all majority neighbours are "in danger" — near the boundary — and only those are used as seeds. Rows whose neighbours are *all* majority are treated as noise and skipped; rows deep inside minority territory are considered already safe. - **ADASYN** keeps every minority row as a possible seed but allocates the synthetic budget in proportion to how many majority neighbours each row has, so the hardest-to-learn rows get the most synthetic company. Both concentrate synthetic mass near the boundary. When the boundary is real and the minority side of it is simply undersampled, that is exactly right and beats uniform SMOTE. When the region is overlap or label noise, it is exactly wrong: they pour the most synthetic data into the area where the labels are least trustworthy. Borderline-SMOTE's explicit noise rule gives it some protection; ADASYN's density weighting gives it none, and a cluster of mislabelled rows can look identical to a genuinely hard boundary region. ## How to diagnose it - Always keep an **unresampled baseline** and compare like for like on evaluation data that was never resampled. - Look at **precision at a fixed recall**, not a single summary score. - Inspect where new false positives fall: if they cluster in a region full of synthetic points, the interpolation is the cause. - Try the cheaper interventions first, and treat the resampling ratio as a hyperparameter that can legitimately tune to "none". ## The senior framing Say plainly that SMOTE cannot create separability that the features do not contain. If the classes overlap, the honest fix is better features or better labels, not more manufactured rows — and resampling that improves a training curve while flattening held-out precision is a change that should be reverted.