|
탐색 건너뛰기 링크입니다.
|
|
Accuracy,Recall, F1 Score[자원전체][Resources][AI Study]
- F1 score for machine learnng system [link] [excel]
- for imbalanced dataset : F1 score
- for balanced dataset : Accuracy,F1 score
- F1 Score, precison,recall '
- How to compute F1 score
- metrics of classifier performance for balanced data
- metrics of classifier performance for imbalanced data
- Data preparation [link]
- Precision and Recall for multi-class classification [link] [link]
- The accuracy (of model prediction)
- is the proportion or percentage of correcly predicted labels over all predictions
- bad problem : relatively 'high' accuracy with the model predicting the 'not so important' class labels fairly accurately
- The precision and recall for each class label
- are used to analyze the individual performance on class labels or average the values to get the overall precision and recall
- The (prediction ) precision (of the class X)
- how many instances were correctly predicted?
- instance X가 X로 예측된 수 / Number of prediction X
- The recall (of class X)
- instance X가 X로 예측된 수 / Number of all instance X
- example, confusion matrix - help interpretation
 - precision=0.5 and recal=0.3 for label A
- precision for label A, 정밀도
- TP_A/(TP_A+FP_A)= TP_A / (Total predicted as A) = TP_A / TotlaPredicted_A = 30/60=0.5
- A라고 예측된 것 중에 50%만 정확하다.
- recall for label(class) A, 재현
- TP_A : instance A에 대하여 예측A(Positive)이 참(A)이다(T)
NF_A : instance A에 대하여 예측A가이닌것(B,C, Negative)가 A가아니다(F)
- TP_A / (TP_A + NF_A) = TP_A/(Total Gold for A) = TP_A/TotalGoldLabel_A = 30/100=0.3
- label A의 30%만 정확하게 예측된다.
- precision and recal for label B ?
- recall = TP_A/(TP_B + FN_B)=60/(60+20+20)=0.6
- precision = TP_B/(TP_B+FP_B) = 60/(60+50+10)=60/120=0.5
- the precision (exactness) and recall (completeness) of a model
- Accuracy Paradox Accuracy Paradox.
- Weighted Accuracy and Unweighted accuracy [link]
-
- therfore, for unblanced classes, WA is better
- Strategy to make dataset balanced
- Merge classes, i.e., weaker or smaller classes
- Reduce label space, i.e., consider juts happy, sad and nueral
- Drop extra data, e.g., we have hap # 15, sad # 20, and Neu # 70, then choose 15 each of hap,sad,neu.
-
LIST TOP
|