The rapid development of the Internet in modern society, people have bombarded daily by hundreds of thousands of information, APP Push, hot news, information flow and an effective advertising ...... "information filter" has become the people's daily lives just to be, also information provider in the fierce market environment stand out nirvana.
Recommended system is playing such a role, it is like a sieve in general, to help us find what most want. However, the recommended system is too high technical threshold and R & D costs to many companies in the door. The fourth paradigm based on machine learning technology product launch of the first intelligent recommendation recommended, focusing on personalized recommendations industry, by virtue of their technical superiority with an effective solution to this problem, it has served 36 krypton, petals, shell and other well-known media, continue to be widely praised in the industry.
In the next article, first recommended the system to explain the knowledge recommendation system, I hope to have more to technology enthusiasts recommendation system, more understanding. First, we will recommend the system workflow talk.
1. Information gathering phase
This phase collects information about the user to generate user portrait prediction task, the user attribute information includes resources, user behavior or user access. Only user portrait fully established, the recommendation system to start running. Recommended system needs to know as much about the user, so the results can provide reasonable recommendations from the outset for the user.
Recommendation systems rely on different types of input, for example, the most direct explicit feedback, i.e. user directly inputs the content of interest, or implicit feedback, i.e. user preferences inferred indirectly by observing user behavior, can also explicit and implicit feedback combination to obtain a mixed feedback.
To e-learning platform, for example, is a collection of portraits of user personal information associated with a particular user. This information includes the user's cognitive skills, intelligence, learning styles, interests and interaction behavior. User portrait is typically used when a user retrieves the information needed to build the model. In other words, the user portraits roughly reflects the user model. To want to do a successful recommendation system, it depends largely on its ability to characterize the user's interests. To obtain accurate recommendation results, accurate user model is essential.
1.1 explicit feedback
A website will typically prompt the user to make evaluations on the content on the user interface, in order to build and improve the user's user model. Recommended accuracy of the results depends on the number of users rating. The greater the number of users rating, the more accurate recommendation results. The only drawback explicit feedback is very dependent on the enthusiasm of the user rating, and that users are not always willing to make ratings. However, in contrast, does not involve displaying feedback from user behavior to obtain user preferences in this step, thus providing more reliable data, the entire recommendation process more transparent and better able to perceive the quality of the recommendation system, thereby improving customer satisfaction degree.
1.2 Implicit feedback
Site background by different monitor user behavior automatically inferred user interests preferences, such as purchase history, navigation history, on certain pages stay, the user clicks on a link, button, and email content. Implicit feedback infer user preferences from user behavior, reduce the burden on the user rating. Implicit positive feedback rating of less demanding users, the accuracy is lower.
There are also some people think, implicit user feedback data is actually more objective, in the case of implicit feedback, users do not need to expect the public to react the way, there is no need to maintain self-image, so the data provided more realistic.
1.3 mixed feedback
Advantages implicit and explicit feedback may be combined in a mixing system, to minimize both insufficient to achieve the best performance and the recommendation system. Specifically, the implicit feedback data to check data explicit feedback, or only allows the user to explicit feedback given by the expression is explicitly interested in.
2. The algorithm learning phase
At this stage, the system will pass learning algorithm, the user feedback filter stage obtained, and extracts user features. For details on this section, we will introduce in a future article.
3. prediction / recommendation Stage
At this stage, the system will predict the type of content the user might like. This step may be directly based on the data sets (based or model-based memory) collected in the information gathering phase is achieved, may also be implemented to monitor user behavior through the background.
In the next article, we will explain the details of the recommended filter technology, so stay tuned.