R2值低但MAPE高？ANN回归模型指标异常原因咨询

Hey there, let's unpack this confusing result with your ANN model. First, a quick correction: when you mention the "MAPE corresponding accuracy is surprisingly high," that’s likely a small mix-up. Looking at your code, accuracy = 100 - mape—so with an 88.73% MAPE, your ANN’s actual accuracy is 11.27%, which is way lower than the ~21% accuracy your decision tree, bagging, and random forest models are getting. What you’re probably puzzled by is: why does the ANN have a far worse R² and RMSE, but also a much higher (worse) MAPE than the tree-based models?

Here are the key reasons behind this discrepancy:

1. Each Metric Measures Something Totally Different #

Regression metrics prioritize different aspects of your model’s performance, which leads to conflicting results when your model fails in specific ways:

• R² Score: Tracks how much of the data’s overall variance your model explains. It’s sensitive to how well the model fits the entire dataset—if your ANN is off-target across most samples, the R² will plummet quickly.
• RMSE: Penalizes large errors heavily (thanks to squaring the error values). A handful of samples with huge absolute errors will blow up your RMSE, even if most predictions are decent.
• MAPE: Calculates the average of |prediction - true value| / true value as a percentage. This metric’s biggest quirk is that it’s hyper-sensitive to samples with small true values. Even a tiny absolute error becomes massive in percentage terms when the true value is close to zero, and those outliers can skew the entire MAPE result.

2. Small True-Value Samples Are Likely the Culprit #

The most probable cause here is that your test set has some samples where the true label is very small. Let’s say you have a sample with a true value of 1:

• If your ANN predicts 10, that’s an absolute error of 9, which translates to a 900% percentage error.
• Your random forest might predict 2, an absolute error of 1 (100% percentage error).

That single small-value sample would jack up your ANN’s MAPE dramatically, while the impact on RMSE (9² vs. 1²) and R² (which looks at overall variance) would be noticeable but not as extreme as the MAPE swing. Tree-based models often handle small-value samples better because they split data based on feature thresholds, which can capture edge cases more reliably than an untuned ANN.

3. Your ANN Might Be Misconfigured #

Don’t rule out issues with how you’re training the ANN either:

• Did you forget to normalize/standardize your data? ANNs are notoriously sensitive to feature and label scales. If your features have wildly different ranges, the model might struggle to converge to a good fit, leading to poor R² and RMSE across the board.
• Is your model structure off? Too few neurons/layers mean underfitting (can’t capture patterns), too many mean overfitting (performs well on training data but bombs on test).
• Training parameters: A learning rate that’s too high can make the model oscillate instead of converging, or not training for enough epochs could leave the model underdeveloped.

How to Verify This #

Try these quick checks to confirm the root cause:

• Pull out the test samples with the smallest true values (top 10% smallest) and compare the ANN’s predictions vs. other models’ on these samples. You’ll almost certainly see the percentage errors here are driving the high MAPE.
• Recalculate MAPE after excluding samples where the true value is near zero, or use a weighted MAPE that downweights extreme percentage errors. This should make the MAPE align more closely with R² and RMSE.
• Plot your ANN’s training loss curve. If it’s still decreasing at the end of training, you need more epochs; if it’s low on training but high on test, you’re overfitting.

内容的提问来源于stack exchange，提问作者Tanuj Chakraborty