Class Imbalance Problem from a Geometric Perspective: A Threshold Adjustment Strategy Without Retraining Models

This article introduces the open-source project geom-imbalance, which addresses the class imbalance problem in machine learning from a geometric perspective. Its core idea is to optimize classification performance by adjusting decision thresholds without retraining the model, featuring non-intrusiveness and high efficiency—making it especially suitable for rapid optimization needs in production environments.

Class imbalance is a common challenge in machine learning, seen in scenarios like fraud detection (only hundreds of fraudulent transactions out of millions) and disease screening (far more healthy people than patients). Traditional solutions such as resampling (oversampling/undersampling) or adjusting class weights require retraining the model, which is time-consuming and labor-intensive, and may introduce bias.

The core concept of geom-imbalance: Class imbalance can be understood from the geometric distribution in the feature space, where the decision boundary position may not be optimal. Even after the model is trained, performance can still be optimized by adjusting the decision threshold. Geometrically, the default 0.5 threshold performs poorly in imbalanced scenarios (e.g., when the positive class accounts for only 1%, the model tends to predict negative). This project finds the optimal threshold by analyzing precision-recall trade-offs, ROC curves, etc.

The tool's features include:

1. Data loading and preprocessing: Supports CSV/Excel uploads and automatically identifies feature and target columns;
2. Resampling comparison: Integrates traditional methods such as random oversampling/undersampling and SMOTE as benchmarks;
3. Visualization tools: Class distribution histograms, ROC curves, precision-recall curves, threshold-performance relationship graphs;
4. Result export: Supports sharing in CSV/PDF formats.

Application scenarios:

• Financial risk control: Quickly adjust thresholds for fraud detection models to improve detection rates and respond to new fraud types;
• Medical diagnosis: Adjust disease screening thresholds to reduce missed diagnosis rates and visually explain adjustment reasons;
• Industrial quality inspection: Optimize defect detection thresholds to reduce the risk of missed inspections.

Comparison between threshold adjustment and resampling methods:

Dimension	Threshold Adjustment (geom-imbalance)	Resampling Methods
Computational Cost	Extremely low, no retraining needed	High, requires retraining the model
Implementation Difficulty	Simple, adjust a single parameter	Complex, need to handle data balance
Interpretability	High, threshold changes are intuitive	Medium, data changes are hard to track
Applicable Scenarios	Model already deployed, needs rapid optimization	Model training phase, sufficient data
Flexibility	High, can adjust in real-time	Low, adjustment requires retraining
Note: Threshold adjustment is not a panacea; retraining is still necessary for extremely imbalanced cases or poor-performing models.

Quick start steps:

1. Download and install: Get Windows/macOS versions from GitHub Releases;
2. Load data: Upload CSV/Excel files;
3. Select method: Compare resampling and threshold adjustment;
4. Run analysis: Click the Analyze button;
5. Interpret results: View visualization charts;
6. Export report: Save as CSV/PDF.

Limitations:

1. Relies on the model's discriminative ability; if the model has no effective features, the effect will be minimal;
2. Mainly for binary classification scenarios;
3. It is post-hoc optimization and cannot replace good data collection and model design. Future directions:

• Dynamic threshold adjustment for online learning;
• Combining with active learning to optimize scenarios with limited annotation resources.