多分类器均预测“bad”阳性，模型评估术语及结果异常分析求助

Let’s get this foundational stuff sorted out to avoid confusion down the line—these terms are easy to mix up at first, so you’re not alone:

• Predicted Positive/Negative: This is just the label your model outputs for a sample. If you’re defining "bad" as the positive class, every time the model says "bad" that’s a predicted positive; "good" counts as predicted negative.
• True Positive (TP): Model predicts "bad" (positive), and the actual sample is "bad"—this is a correct prediction.
• True Negative (TN): Model predicts "good" (negative), and the actual sample is "good"—another correct call.
• False Positive (FP): Model incorrectly labels a "good" sample as "bad" (predicted positive, actual negative).
• False Negative (FN): Model incorrectly labels a "bad" sample as "good" (predicted negative, actual positive).

It sounds like you might have mixed up "Predicted Negative/True Positive" in your description— that would actually be a False Negative, since the model’s prediction doesn’t match the true label. But let’s dive into your core issue.

You noted all classifiers perform well on predicted negative cases (both TN and FN, I assume you mean the model rarely gets negative predictions wrong) but struggle with predicted positive cases, and all models are leaning hard into predicting "bad" (positive). Here are the most likely culprits:

1. Severe Class Imbalance #

This is the #1 suspect here. If your dataset has way more "bad" samples than "good" ones, models will learn that predicting "bad" for everything gives them a high overall accuracy—even though it’s useless for actually distinguishing classes.

• Why this matches your observation: The few times the model does predict "good" (negative), it’s only for samples it’s extremely confident about, so those predictions are mostly correct (high TN, low FN). But when it predicts "bad" (positive), most of those are actually "good" samples misclassified (high FP, low TP)—hence the poor performance in the positive prediction bucket.

2. Poor Feature Relevance #

If your features don’t actually capture meaningful differences between "good" and "bad" samples, the model can’t learn a useful pattern. Instead, it defaults to the majority class (bad) to minimize error.

3. Misaligned Decision Thresholds #

Most classifiers use a default threshold (like 0.5 for probability-based models) to turn raw scores into labels. If your model’s predicted probabilities for "bad" are consistently high (even for "good" samples), the threshold will trigger positive predictions too often. You might need to adjust this threshold based on your priorities.

• Check class distribution: Count how many "good" vs "bad" samples you have. If imbalance is the issue, try:
  • Over-sampling the minority class ("good") with techniques like SMOTE
  • Under-sampling the majority class ("bad")
  • Using weighted loss functions (most frameworks let you assign higher weights to minority classes)
• Audit your features: Use tools like feature importance scores (for tree-based models) or SHAP values to see which features are driving predictions. Remove irrelevant or redundant features, and add new ones that might capture the "good" vs "bad" difference better.
• Tune the decision threshold: Plot a ROC curve to find the threshold that balances your desired precision and recall. For example, if you want fewer false positives (not labeling "good" as "bad"), raise the threshold.
• Validate your labels: Double-check that your "good" and "bad" labels are correctly assigned—sometimes mislabeling can lead models to learn bad patterns.

内容的提问来源于stack exchange，提问作者Avba