For example, when you open Netflix, "Recommended for you" Is on the top screen. This "recommended for you" An engine that extracts.
The recommended engine system greatly changes the sales of the entire site, making it one of the places where data scientists can play an active role. There is surprise etc. as a framework to make a recommend engine.
It seems that there are about 3 types.
Simple engine We simply recommend the ones with high evaluation. We do not evaluate individual user preferences. For example, when the rating of "Your Name" is 5 stars, I recommend it to all users.
advantage Very simple and easy to implement
Disadvantage It is not recommended to reflect the tastes of individual users. Therefore, SF is recommended for women, and romance movies popular with women are recommended for men, and the accuracy is low.
Content-based engine We simply recommend products with similar content. For example, Predator and Alien are similar in content. If you like Predator, I recommend aliens.
advantage The implementation is uncertain and easy to explain to people. Strong when recommending movies by genre (SF, romance, etc.)
Disadvantage I tend to recommend old movies because I can't keep up with trends. Since it only recommends items with similar contents, it is recommended regardless of whether it is low or high.
Collaborative Filtering engine I recommend to you what users like you like.
For example, suppose you like Ponyo on the cliff (Ghibli works in general). Other than you, users who like Ponyo on the cliff (general Ghibli works) also like "Your Name" for some reason. So the recommend engine recommends you "Your Name".
advantage High accuracy because it recommends to individual user's taste.
Disadvantage It's a little difficult to implement and difficult to explain to people. The accuracy is low without a large amount of data.
Data-savvy companies such as Netflix use collaborative filtering engines. The difficulty of mounting increases from top to bottom, but the impression that accuracy also increases.
Recommended Posts