Imbalanced datasets are a persistent challenge in supervised machine learning. In many real-world problems such as fraud detection, medical diagnosis, and fault prediction, the minority class carries higher importance but appears far less frequently than the majority class. Traditional classifiers tend to bias predictions toward the majority class, leading to misleading accuracy metrics and poor generalisation. To address this issue, Synthetic Minority Over-sampling Technique (SMOTE) was introduced as a data-level solution that generates synthetic minority samples instead of duplicating existing ones.
Over time, researchers observed that standard SMOTE does not perform equally well across all data distributions. This led to the development of refined variants such as Adaptive Synthetic Sampling (ADASYN) and Borderline-SMOTE. Understanding these methods is essential for practitioners and learners building strong foundations through a data analyst course in Delhi, where handling skewed datasets is a core analytical skill.
Limitations of Standard SMOTE
Standard SMOTE works by interpolating new minority samples between a data point and its k-nearest minority neighbours. Mathematically, a synthetic point is generated as:
x_new = x_i + λ(x_nn − x_i), where λ ∈ [0,1]
While this approach reduces overfitting caused by simple replication, it treats all minority samples equally. This uniform treatment introduces two key problems. First, synthetic points may be generated in regions dominated by the majority class, increasing class overlap. Second, it ignores the varying difficulty levels of minority samples, some of which are already well represented while others are hard to learn.
These limitations motivated more adaptive approaches that consider local data distribution before generating synthetic samples.
Adaptive Synthetic Sampling (ADASYN): Mathematical Intuition
ADASYN extends SMOTE by dynamically adjusting the number of synthetic samples generated for each minority instance. The core idea is simple: harder-to-learn minority samples should receive more synthetic support.
For each minority instance x_i, ADASYN computes the ratio r_i of majority class neighbours within its k-nearest neighbours. This ratio reflects the local learning difficulty. The total number of synthetic samples G is distributed proportionally:
g_i = (r_i / Σ r_i) × G
This ensures that samples surrounded by many majority instances receive more synthetic points. Like SMOTE, the actual synthetic data generation still relies on linear interpolation between minority neighbours.
From a mathematical perspective, ADASYN shifts the sampling distribution toward high-density decision boundaries. This improves classifier sensitivity but can introduce noise if the dataset contains outliers. Learners in a data analyst course in Delhi often encounter ADASYN when studying advanced resampling strategies and bias-variance trade-offs.
Borderline-SMOTE: Focusing on Decision Boundaries
Borderline-SMOTE addresses imbalance by targeting only those minority samples that lie near the class boundary. The method classifies minority points into three categories based on their k-nearest neighbours: safe, danger, and noise.
If most neighbours belong to the minority class, the point is considered safe and ignored. If the majority of neighbours belong to the majority class but not entirely, the point lies in the danger zone. Borderline-SMOTE generates synthetic samples only for these danger points.
Mathematically, interpolation is restricted to minority neighbours that also lie close to the boundary. This focused strategy reduces unnecessary oversampling in well-separated regions and minimises noise amplification. Compared to ADASYN, Borderline-SMOTE is more conservative and often performs better when class boundaries are well defined but overlapping.
Comparative Insights and Practical Considerations
ADASYN and Borderline-SMOTE differ primarily in how they allocate synthetic samples. ADASYN adapts sample counts based on local density, while Borderline-SMOTE selectively oversamples boundary regions. ADASYN is effective when learning difficulty varies significantly across samples, whereas Borderline-SMOTE excels when misclassification risk is concentrated near decision boundaries.
However, both methods assume that the minority class structure is meaningful and that interpolation is valid. In high-dimensional or noisy datasets, improper use may degrade performance. Therefore, resampling should always be paired with cross-validation and suitable evaluation metrics such as precision-recall curves.
These nuances are frequently discussed in a data analyst course in Delhi, where learners practice choosing resampling techniques based on data geometry rather than relying on default solutions.
Conclusion
SMOTE variants represent an important evolution in handling imbalanced learning problems. ADASYN introduces adaptive weighting to focus on difficult samples, while Borderline-SMOTE concentrates synthetic generation near classification boundaries. Both approaches address weaknesses of standard SMOTE but must be applied with care to avoid noise amplification or class overlap.
A strong theoretical understanding of these methods enables practitioners to make informed preprocessing decisions rather than relying on trial and error. For those developing analytical expertise through a data analyst course in Delhi, mastering SMOTE variants builds a solid foundation for tackling real-world datasets where imbalance is the norm rather than the exception.