One 、 Introduction to online learning

Online Learning It's not a model , It's a model training method ,Online Learning Based on online feedback data , Real time and fast model tuning , Make the model reflect the changes on the line in time , Improve the accuracy of online prediction .

Online Learning The process includes ： Show the prediction results of the model to users , And then collect feedback data from users , And then used to train the model , Closed loop system . As shown in the figure below ：

Two 、 The difference between online learning and traditional learning

Traditional training methods , After the model goes online , The update cycle will be longer （ It's usually one day , It's an hour when it's efficient ）, When this model goes online , It's usually static （ It won't change for a while ）, There's no interaction with the online situation , Suppose the prediction is wrong , Correction can only be done at the next update .Online Learning Different training methods , The model will be adjusted dynamically according to the results of online prediction . If the model predicts wrong , We'll fix it in time . therefore ,Online Learning Can reflect online changes more timely .

3、 ... and 、Online Learning The goal of optimization

As shown in the figure above ,Online Learning The training process also needs to optimize an objective function （ Marked in red ）, But unlike other training methods ,Online Learning It is required to find the optimal solution of the objective function quickly , It's better to have an analytic solution .