The fifth Chapter hybrid recommendation method
Recommendation questions can be converted to effects using functions, the input of functions for users and items, the output for the utility of the user and the item-utility refers to the ability of an object to satisfy an abstract goal, such as satisfying a user's needs, or satisfying a retail conversion rate, and so on, any recommendation system that chooses n items from a multitude of items. The sum of the utility of these n items is maximum
1. Integrated hybrid design from the bottom of the features to consider the integration of 1.1 features hybrid scheme if there are many features can be used, such as: User browsing, click, Search, purchase, etc. behavior, to predict what the user next to buy what, obviously search and purchase behavior than browsing, click behavior More important. For the use of features of different importance (important to use important, otherwise unimportant, or weighted synthesis), coupled, outside the popular algorithm framework, such as: Collaborative filtering 1.2 features complementary program not too understand, seemingly and the above--when a feature OK, use this feature; , the weaker features are used
2. Parallel hybrid design Multiple recommendation engines, how to fuse together? 2.1 Cross-mixing multiple results of multiple recommendation engines, cross-merge into one result: first engine first result ranked first, second engine first result ranked second ... 2.2 Weighted mixed linear weighted combination, one weight per engine, weight normalization 2.3 switching mix when in some cases with the first classifier (for example: data is full when using co-filtering), in some cases with another classifier (for example: Cold start with content-based or knowledge-based recommendations)
3. pipelining Hybrid Design This section is a bit of a RIP. 3.1 The output of a single recommendation engine in tandem is entered as another recommendation engine. For example: first with a relatively rough recommendation engine, select n items, and then, with a relatively delicate engine, from the N items to further filter out M (m<n) items, further, there may be a recommendation engine, the M-items re-ordered 3.2 gradation mixed with the front of no big difference, Don't look.
Finish.
"Book Notes" recommendation System (Recommender systems An introduction) Fifth Hybrid recommendation method
