Part 1 More general information about RecSys
(1) What do the recommendation system do?(Model perspective)
RecSys can be seen as use the known links to predict the unknown links.
There are two kinds of prediction a RecSys is going to do – rate prediction and behavior prediction. Someone says 80% of RecSys pprediction is behavior prediction and 20% is rate prediction.
To do the behavior prediction, we can predict the probability of the behavior, aka, CTR(Click Through Rate).
(2) The common problems in RecSys
There are three main problems that can cold start problem, explore and exploit (EE/Double E) problem and safety problem. Cold start problem: RecSys is data hungury application.
Part 2 The mindsets you should know
The components of RecSys and their importance:
UI/UE > Data > Domain knowledge > Algorithm (4>3>2>1)
Data: the food of the recommendation systems. “巧妇难为无米之炊”。
Target mindset(Quantify everything <- Data)
Uncertainty mindset(Probability)
Part 3 User portrait/ User profile / UP
- What is user profile?
The user profile is used for and to machine but not for human. User/Item embedding is what we called “user profile”, so we cannot say user profile is the target of the RecSys, it should be the by product of the process of building a RecSys.
- What is the key points of user profile?
The two key points to build a recommendation system are dimension and quantify.
- The three ways to build the user profile
Demographical data, history data and black box.
Ongoing