The algorithm model greatly improves the prediction efficiency of various results.
[Essence of the algorithm model]
The essence of the algorithm model is to obtain prediction results based on various input variable factors through calculation rules (model or formula).
Typical prediction results such as:
1. (Through historical behavior & preference prediction) the possibility of a user clicking on a piece of information
2. (Through historical behavior & preference prediction) natural demographic attributes of users such as gender, etc.
[How to determine the quality of the model]
The evaluation of accuracy and recall is one of the most commonly used means to verify the quality of the algorithm model.
Now suppose you and the model are playing a question-and-answer game, take a sample each time, tell him some information about this person, and let him find out all the boys.
Accuracy = predicted accuracy/recall (find out).
For example: in all samples, the model pre-finds 50 people and says they are all men, but only 40 of the people found are actually men, and the accuracy rate = 40/50 = 80%, which is used to measure the part of the found out. Accuracy.
Recall = Amount of Accuracy in Recall / Amount of Objective Correctness.
It is to compare the real results with the predicted results. For example: there are actually 60 males in total, and the model only finds 50, then the recall rate = 50/60 = 83.3%, which is used to measure the coverage of the found part to the actual real part.
[Quasi-call judgment, what will happen? 】
The evaluation of the predicted results thus constitutes the following four sets.
As shown in the figure below, the four types of collections just now look like this when the graphics are spread out.
The figure below shows a case where the model is overly conservative.
The model can be very conservative, with an accuracy rate of 100%, but due to excessive pursuit of accuracy, a large number of correct results are missed.
The figure below represents over-recall.
The recall rate is 100%, which ensures that the correct set is recalled, but the precision rate is very low due to the recall of a large number of wrong sets.
The image below represents the ideal situation—many and accurate!
