In the "Postscript" section of the book, the author lists the ten lessons of the recommendation system summarized by strand researchers, and summarizes their experience and lessons in designing the recommendation system.
1. Make sure you really need a recommendation system
Personal Evaluation: Recommendation is not for recommendation purposes. Recommendation is just a means. The key is to see where the user value is. From the user perspective, it can really help the user.
2. determine the relationship between business objectives and user satisfaction
3. select an appropriate developer
Generally, a large company should hire its own developers to develop a recommendation system.
4. Forget the Cold Start Problem
Continuous innovation, there is any data you want on the Internet. As long as users like your products, they will constantly contribute new data.
Personal comment: in fact, we put up the number of users of the product first. If we have users, we can naturally collect data.
5. balance the relationship between data and Algorithms
There is a good saying:Data analysis determines how to design a model, while algorithms only determine how to optimize the model.Personal rating:The data viewing process cannot be ignored.After you select a good feature, you can use a simple algorithm to achieve good results. On the contrary, sometimes the effect may not be the cause of the model, but the feature selection is too outrageous.
6. It is easy to find related items, but it is difficult to show them to users in what way. Do not recommend for recommendations, but be user-oriented and business-oriented.
Personal rating: When will the recommendation be made? In what way do I recommend? For example:The system frequently pops up the window to make the user dislike, even if the pop-up window content is actually the content that the user may consume.
7. users who do not waste time computing similarity can directly use the resources of social networks.
Personal rating:Is the relationship between usercf and social network algorithms.. I personally think that we should make full use of resources and make full use of existing data (such as social network user relationships), rather than Computing new data (usercf user relationships ).
8. Constantly Improve algorithm scalability.
9. select an appropriate user feedback method.
10. Design a reasonable evaluation system and keep an eye on the performance of all aspects of the recommendation system.
Recommendation System Practice (item bright)-postscript