aiacademy: 機器學習 logistci regression & 實作:ch04

logistic regression

gradient ascent

和大神教的有點不一樣　ＬＯＬ…

靠一樣！只是 負號 提出來

• 我的筆記

• Similarly, we can apply gradient descent, stochastic gradient descent, and mini-batch gradient descent to solve logistic regression

Multi-calss logistic regression

這邊介紹到 softmax function
就要看這篇

Q & A

• Quize:

  • What is logistic regression?

  • Why do we usually maximize log-likelihood function (instead of likelihood function)?

  • What is the cross entropy loss (for logistic regression)

• Answer:

Evaluation (classification)

看看 Coursera 我的筆記

• Classification metrics
  • Accuracy
  • Precision
  • Recall
  • F1 score
  • Precision recall curve
  • ROC curve and AUC (Area Under Curve)

Accuracy

Confusino matrix

Precision

Recall

Precision and recall tradeoff

• if we want a very high precision
  • return only the most confident positive instances
• if we wnat a very high recall
  • return all the instance
• the two metrics are ususally a tradeoff

F1 Score

ROC Curve

• Precisions, recalls (TPRs), and FPRs of different thresholds

• ROC Curve

Logistic Regression

• 1

• 2

Logistic Regression in Scikit-Leran

C 值越大 對weight 的控制力越弱

multi_class: ovr, multinominal

short summary

F1-Score and AUC ROC

Threshold

F1 Score

Prcision vs. Recall / ROC:TPR vs. FPR