We use four metrics (ie, precision §, recall ®, F1-score (F1) and mean intersection over union (Mean IoU)).
Wherein P, R, F1-score is calculated as follows:
Where TP (true positive) is represented by a positive cases had also been predicted number of positive cases, FP (false positive) would have a negative embodiment represented, are predicted to be the number of positive cases, FN represents a positive cases had been predicted is the number of negative cases.
Do not care about the name, it is important to define the P and R to be clear, on behalf of P is the identified goals, in the end how much is the recognition of the right. R represents the original need to be identified, in fact, how many are identified out.
I see a blog and write very well. Reference links:
https://www.cnblogs.com/xd502djj/p/6938747.html
Following is a brief list of some common recommendation system evaluation indicators:
1, precision and recall rate (Precision & Recall)
Precision and recall rate are two widely used measure of information retrieval and classification in the field of statistics, for quality evaluation results. Wherein accuracy is retrieved ratio related to the number of documents the total number of documents retrieved, measures the precision retrieval system; recall rate is the ratio of all the number of documents to retrieve relevant documents number and the document library, measure that recall retrieval system.
In general, Precision is retrieved entry (such as: documents, Web pages, etc.) how much is accurate, Recall is all correct entries out how many were retrieved.
(Just like their usual Baidu, as there are many recommendations, but only partly accurate. This can be calculated Precision and only a portion of the database is retrieved, and this can be calculated recall)
Correct, recall and F values are in cohabitation environment, a key indicator of selected targets. Take a look at the definition of these indicators first:
1. 正确率 = 提取出的正确信息条数 / 提取出的信息条数
2. 召回率 = 提取出的正确信息条数 / 样本中的信息条数
Both values between 0 and 1, the value closer to 1, precision or recall is higher.
3. F值 = 正确率 * 召回率 * 2 / (正确率 + 召回率) （F 值即为正确率和召回率的调和平均值）
You may wish to cite an example: there is a pond 1400 carp, shrimp, 300, 300 turtle. Now to catch carp for the purpose. Spread a large network, caught 700 carp, shrimp, 200, 100 turtle. So, these indicators are as follows:
Accuracy = 700 / (700 + 200 + 100) = 70%
Recall rate = 700/1400 = 50%
F value = 70% * 50% * 2 / (70% + 50%) = 58.3%
If we take a look at the pool of all the carp, shrimp and turtles are caught, but what happens to these indicators:
Accuracy = 1400 / (1400 + 300 + 300) = 70%
Recall rate = 1400/1400 = 100%
F value = 100% * 70% * 2 / (70% + 100%) = 82.35%
Thus, the correct rate is to assess the results of the capture of target achievement share pro rata; recall, by definition, from the areas of concern, the recall ratio target category; and F value, it is the comprehensive assessment of both indicators for a comprehensive reflection of the overall index.
Precision certainly hope that the higher the better search results, while Recall also the higher the better, but in fact the two are contradictory in some cases. Under such extreme circumstances, we search out only one result, and is accurate, Precision is 100%, but Recall very low; and if we put all the results are returned, then such Recall is 100%, but would Precision very low. Therefore, in different occasions the need to judge for themselves hope Precision Recall is relatively high or relatively high. If it is done experimental studies can draw Precision-Recall curves to help analyze.
2, a comprehensive evaluation (F-Measure)
The P and R indicators sometimes appear contradictory, so they must be considered, the most common method is the F-Measure (also known as the F-Score).
F-Measure is weighted harmonic mean of Recall and Precision:
When the parameter [alpha] = 1, is the most common F1, F1 understood combines P and R in result, when the high F1 Test Method described can be more effective.
3, E value
E P value indicates precision and recall of the weighted average of R, wherein when a is 0, E value. 1
B greater weight represents a greater weight of precision.
4, the average accuracy (Average Precision, AP)
Average accuracy rate indicates the average accuracy rate of the different check points on the full rate.
Original link: http://blog.csdn.net/taohuaxinmu123/article/details/9833001