2. Logistic Regression - qihongl/MVPA_tutorial GitHub Wiki
Compare logistic loss versus least square loss (UNDER CONSTRUCTION)
[simulation code]
What I want to say:
- classification - as simple as least square can be used as a classifier - sign(X * beta)
- what is logistic regression
- logistic loss is more appropriate for classification (than least square linear prediction)
- all the regularization techniques work for logistic regression
Here, the OLS boundary is suboptimal (having a smaller margin), since OLS loss care about the distance between the prediction and the truth, even when the prediction is on the right side. On the other hand, logistic regression behaves more like an support vector machine. (explain more)
Thought question: Explain the plot above by comparing the logistic loss versus the squared error loss.
- Hint 1: logistic loss is less sensitive to far away points.
- Hint 2: square error loss has penalty for "easy" points.