Closed Starsa closed 3 years ago
p-value:
the probability of getting results at least as extreme as what we observed, given the null hypotheses is true.
R2
How much variance in the target vector is explained by the features. Determines goodness of fit.
Adjusted R2
Modified version of R2, it imposes a penalty for too many features
AUC
AUC (Area under the ROC curve) is an aggregate measure of performance across all possible classification thresholds.
Measure of accuracy (True Positive/ TruePositives + False Positives)
True positive rate (True Positives/ TruePositives + False Negatives)
if we wanted to find all positives, absolutely, maximize the recall
Error- Expected error created by using a model to approximate a real-world function
Noise- The error from sensitivity to small fluctuations in the training set
A simple model has high bias and low variance and a complex model has high variance and low bias. The tradeoff is finding a good balance without overfitting or underfitting the data.
These are all great. Hopefully you won't get a grumpy statistician asking about p values
AUC
AUC (Area under the ROC curve) is an aggregate measure of performance across all possible classification thresholds.
Can you work out why I don't like AUC for model evaluation from this definition?
Precision
Measure of accuracy (True Positive/ TruePositives + False Positives)
Recall
True positive rate (True Positives/ TruePositives + False Negatives)
if we wanted to find all positives, absolutely, maximize the recall
According to Cassie Kozyrkov: Precision: "Don't waste my time. Missed opportunities are okay" Recall: "Collect 'em all. Duds are okay"
"All mistakes are equally bad"
"I can't choose between precision and recall"
Memorize statistical and machine learning concepts with one sentence answer