神经网络回归任务（连续变量预测）能否像分类任务一样计算Accuracy？

Great question! While accuracy is traditionally tied to classification tasks where you're predicting discrete labels, you absolutely can adapt similar "correct/incorrect" framing for regression (continuous prediction) work—you just need to reframe what counts as a "correct" prediction.

First, let’s clarify: raw accuracy (matching the exact true value) is useless for continuous targets, since hitting an exact value is nearly impossible. Instead, we define a "good enough" standard to turn continuous predictions into a binary or multi-class "correct/incorrect" judgment. Here are the most practical approaches:

1. Threshold-Based "Accuracy" (Binary Correctness) #

This is the most common workaround. You set a tolerance range around the true value—if your prediction falls within that range, you count it as a "correct" prediction. Then calculate accuracy just like in classification:

Accuracy = (Number of predictions where |y_pred - y_true| ≤ ε) / Total number of samples

• How to choose ε: This depends entirely on your use case. For example:
  • In medical dose prediction, ε might be 1% (since even small errors can have serious consequences)
  • In retail sales forecasting, ε could be 10% (since minor misses don’t derail business decisions)
• For data with wide value ranges (like house prices), use relative error instead: |(y_pred - y_true)/y_true| ≤ ε

2. Binning (Turn Regression into Multi-Class Classification) #

If your continuous target has meaningful segments, split it into discrete bins and treat the problem as classification. For example:

• Instead of predicting exact house prices, split prices into "Low ($0-$200k)", "Medium ($200k-$500k)", "High ($500k+)"
• Train your model to predict these bins, then calculate standard classification accuracy
• Pro tip: Use quantile-based bins (e.g., quartiles) to ensure balanced classes, or bins aligned with specific business goals (like profit margin tiers)

3. Rank-Based Accuracy (For Order-Aware Regression) #

If your regression task focuses on relative order (e.g., predicting user ratings to rank products), you can measure how well your predictions match the true ranking:

• Top-N Accuracy: For each user/sample, check if the top N predicted items match the top N true preferred items
• Pairwise Accuracy: Count the share of sample pairs where the predicted order matches the true order (e.g., if true value A > B, does your model predict A > B?)

Key Caveats #

• These metrics are supplementary, not replacements for standard regression metrics (MSE, MAE, RMSE). Accuracy tells you how often you’re "close enough," but it doesn’t capture how wrong you are when you miss—two models could have the same accuracy, but one has tiny misses and the other has massive errors.
• Threshold or bin choices are subjective. Always tie them to real-world business needs, not arbitrary numbers.

内容的提问来源于stack exchange，提问作者Kareem