The recommendation system is a mechanism that ** selects the ones that are valuable to the user from multiple options and supports decision making **.
For example, when shopping on Amazon's website, it seems to suit users ** like "Products related to this product" and "People who checked this product also checked such products". They will recommend other products **. In addition to Amazon, many services such as Netflix (movie recommendation) and Twitter (user and post recommendation) use the recommendation system.
The first difference in the recommendation system classification is whether it is roughly personalized or not.
--Rule base --Content-based filtering --Collaborative filtering --Knowledge base filtering --Hybrid filtering
Content is presented by setting certain rules on the site side. For example, ** What's New **, ** Sale Items **, or ** Recommend this item to anyone who takes this action **. This is a non-personalized recommendation system.
It presents highly similar content based on the attributes (variables) of the content. For example, we recommend movies of the same genre, or books of the same author.
It is recommended based on the user's browsing history and behavior history, and is a method that is personalized from the content base. There are two types of collaborative filtering: item-based and user-based.
--Item base
It is a method of recommending on a product basis based on the action history.
For example, when Mr. A purchases product A, other people who purchased product A also purchase product B, so he recommends product B to Mr. A as well.
--User base
It is a method of recommending on a user basis based on the action history.
It calculates the similarity between users (similar purchase history, etc.) and recommends products purchased by users with similar similarities.
It is a recommendation method that reflects the user's desire "what do you want now" by having the user show their preference.
The difference from search is that it is ** presented in a narrower range in the form of recommendation **. In a general search, hundreds or thousands of hits that meet the conditions are hit, but knowledge base filtering makes suggestions according to the situation.
A combination of content-based filtering and collaborative filtering, which is more effective when there is a lot of data. For example, suppose Mr. A and Mr. B have similar tastes, and the attributes of product A and product B are similar. At this time, if Mr. A purchases product A and product B, and Mr. B purchases product A, it is highly likely that Mr. B likes product B and purchases it. This is a hybrid recommendation that has a high degree of similarity whether it is content-centric or user-centric.
Each has its advantages and disadvantages, so it is necessary to select the system according to the desired effect and cost.
For example, content-based content has the advantage of being able to accommodate new users, but it is not fresh because it recommends only similar items. In addition, while collaborative filtering tends to increase user satisfaction by unexpected product recommendations, it has the disadvantage of not being able to make good recommendations for new products and new users (cold start). While knowledge-based filtering can better reflect user needs, it is not available to users who do not have explicit needs.
There is also a difference between batch processing and real-time recommendation from a different angle than the classification by the above algorithm.
Professor Kamishima's explanatory material explains the recommendations from various angles. It was easy to understand!
Recommended Posts