Machine Learning Algorithms Excellence: Metrics

Evaluation algorithm Excellent program, commonly used a series of indicators to measure, mainly including: Precision,recall,f-1 score, why design these values? Can't I use precision alone?

1, what is precision?

Precison, accuracy, mainly indicates how much of the detected alert is the correct judgment (True POSITIVE,TP).

In practice, due to the different proportions of normal and abnormal data in the sample, accuracy can not reflect the real algorithm performance, for example:

Cancer detection: It is assumed that 99.5% of the people are healthy, and 0.5% of the people suffer from cancer.

Error detection algorithm: Simple all patients are judged to be healthy.

Accuracy: The above algorithm accuracy is 99.5%!! Only 0.5% of unhealthy people are wrongly judged to be healthy.

2, why use recall?

An obvious error detection algorithm in the above example can easily achieve 99.5% accuracy, so accuracy does not fully measure the algorithm excellence, to set up a recall value to measure.

Recall, summon rate, anomaly data is detected in proportion. Recall = tp/(TP+FN)

In the example above, recall is 0 and no outliers can be detected.

3, why use F-1 score?

Combining precision and recall can be a more comprehensive measure of how good an algorithm is, and F-1 is one such indicator:

F-1 score = 2*precision*recall/(Precision+recall)

Note: It is not easy to Precision and Recall mean (Average), the above example of the obvious error, Precision = 0.995, Recall = 0, the mean value of 0.5, obviously unreasonable.

And the use of F-1SCORE,F1 is still 0, more reasonable.

Machine Learning Algorithms Excellence: Metrics