| Title |
Handling class imbalance in k-nearest neighbors |
| Translation of Title |
Klasių Disbalanso sprendimas k Artimiausių Kaimynų Klasifikatoriui. |
| Authors |
Matvejevaitė, Jovita |
| Full Text |
|
| Pages |
45 |
| Keywords [eng] |
STIR-kNN, kNN, imbalanced classification, inference-time rebalancing, minority-class recall, asymmetric misclassification costs |
| Abstract [eng] |
This work introduces Stochastic Inference-Time Rebalancing kNN (STIR-kNN), an inference-time modification of the k-nearest neighbor classifier for imbalanced binary classification. The proposed method aims to improve minority-class recall while limiting the corresponding loss in majority-class precision, without modifying the training data. STIR-kNN combines adaptive neighborhood expansion with a stochastic vote-adjustment mechanism that selectively rebalances local class influence when minority-class evidence is detected. The approach is evaluated against standard kNN and kNN trained on SMOTE-augmented data across datasets with varying imbalance ratios, including numerical, heterogeneous, and time-dependent historical data. Experimental results demonstrate that STIR-kNN consistently enhances minority-class recall while maintaining competitive overall performance, highlighting the effectiveness of inference-time rebalancing for imbalanced kNN classification. These findings suggest that inference-time neighborhood rebalancing is an effective and flexible strategy for addressing class imbalance in kNN classifiers. STIR-kNN demonstrates that substantial gains in minority detection can be achieved without retraining or synthetic data generation, making it well-suited for applications where data integrity and deployment simplicity are critical. |
| Dissertation Institution |
Vilniaus universitetas. |
| Type |
Master thesis |
| Language |
English |
| Publication date |
2026 |
