WO2021251806A1

WO2021251806A1 - Content recommendation method and system

Info

Publication number: WO2021251806A1
Application number: PCT/KR2021/095065
Authority: WO
Inventors: Qiugan SHI; Peng Zhou
Original assignee: Samsung Electronics Co., Ltd.
Priority date: 2020-06-10
Filing date: 2021-05-28
Publication date: 2021-12-16
Also published as: CN111694971A

Abstract

The present application discloses a content recommendation method, including: obtaining, by an edge server, a score of a viewing content according to a behavior of a user towards the viewing content; updating, by the edge server, an interest feature vector of the user and sending the score and an updated interest feature vector of the user to a cloud server; determining, by the cloud server, a user cluster most matched with the user among all user clusters on the cloud server by using the received updated interest feature vector of the user; updating, by the cloud server, the initial recommended queue by adding a content corresponding to the user cluster and deleting a content towards which the user has generated a behavior from the initial recommendation queue; sequencing the updated initial recommendation queue; and recommending first N contents in the sequenced initial recommendation queue to the user. By applying the present application, the computing pressure of the cloud server can be greatly reduced.

Description

CONTENT RECOMMENDATION METHOD AND SYSTEM

The present application relates to the computer technology and, more particularly, to a content recommendation method and system.

With the rise of content platforms such as short video platforms and news information platforms, users like to browse various short video applications and news applications through clients in leisure time. When different users browse a same content platform, the platform needs to push different contents to the different users because the different users have different interests. In addition, for a user, his interest is changing from time to time. Due to the short viewing time of short videos and news, if the server cannot quickly respond to the change of the user's interest in a short time, the user's experience will be poor.

Recommendation systems in different industries are similar in nature although they have different application fields and scenarios. All recommendation systems rely on the features of both content and user interest to decide what to recommend. At present, all the features (content features and user interest features) are calculated at the cloud server end, and therefore with the increasing number of users, the cloud server will suffer from increasingly higher computing pressure. With the network speed increasing year by year and the user demands increasing, adopting the existing content recommendation methods will bring more and more computing pressure to the cloud server, so how to relieve the computing pressure of the cloud server becomes an urgent problem yet to be solved in the new era.

The present application provides a content recommendation method and system, to significantly reduce the computing pressure of the cloud server.

The technical problem to be achieved by an embodiment of the present disclosure is not limited to the above-described technical problem, and other technical problems may be inferred from the following embodiments.

To achieve the above object, the present application adopts the following technical solutions.

Provided herein is a content recommendation method, including: obtaining, by an edge server, a score of a viewing content according to a behavior of a user towards the viewing content, wherein the score is used to indicate the user's level of interest in the viewing content; determining, by the edge server, an interest feature vector of the user after obtaining the score, and inquiring a content feature vector of the viewing content from a cloud server; updating, by the edge server, the interest feature vector of the user according to the acquired content feature vector of the viewing content and the obtained score; sending, by the edge server, the obtained score and an updated interest feature vector of the user to the cloud server; determining, by the cloud server, a user cluster most matched with the user among all user clusters on the cloud server by using the received updated interest feature vector of the user; determining, by a cloud server, an initial recommendation queue including a content corresponding to the determined user cluster; updating the initial recommendation queue by deleting a content towards which the user has generated a behavior from the initial recommendation queue; sequencing the updated initial recommendation queue; and recommending first N contents in the sequenced initial recommendation queue to the user, wherein N is a preset positive integer.

Also provided herein is a content recommendation system including: an edge server and a cloud server; wherein, the edge server is configured to obtain a score of a viewing content according to a behavior of a user towards the viewing content, wherein the score is used to indicate the user's level of interest in the viewing content; determine an interest feature vector of the user after obtaining the score, and inquire a content feature vector of the viewing content from the cloud server; update the interest feature vector of the user according to the acquired content feature vector of the viewing content and the obtained score; and send the obtained score and an updated interest feature vector of the user to the cloud server; and the cloud server is configured to determine a user cluster most matched with the user among all user clusters on the cloud server by using the received updated interest feature vector of the user; determine an initial recommendation queue including a content corresponding to the determined user cluster; update the initial recommendation queue by deleting a content towards which the user has generated a behavior from the initial recommendation queue; sequence the updated initial recommendation queue; and recommend first N contents in the sequenced initial recommendation queue to the user; wherein N is a preset positive integer.

Also provided herein is a content recommendation method processed by a cloud server, comprising: determining a content feature vector of a viewing content; sending the content feature vector of the viewing content to an edge server; receiving, from the edge server, an interest feature vector of a user updated based on the content feature vector of the viewing content, and a score of the viewing content indicating the user's level of interest in the viewing content; determining a user cluster most matched with the user among all user clusters on the cloud server by using the received interest feature vector of the user; determining an initial recommendation queue including a content corresponding to the determined user cluster; updating the initial recommendation queue by deleting a content towards which the user has generated a behavior from the initial recommendation queue; sequencing the updated initial recommendation queue; and recommending first N contents in the sequenced initial recommendation queue to the user, wherein N is a preset positive integer.

Fig. 1 is a schematic diagram of a basic flow of a content recommendation method in the present application;

Fig. 2 is a schematic diagram of a basic structure of a content recommendation system in the present application;

Fig. 3 is a flow chart of a process at an intelligent terminal in the content recommendation method of the present application;

Fig. 4 is a flow chat of a process at an edge server in the content recommendation method of the present application;

Fig. 5 is a schematic diagram of a process at a cloud server in the content recommendation method of the present application;

Fig. 6 is a flow chart of recommending contents in the cloud server according to the method of the present application;

Fig. 7 is a schematic diagram of contents associated with a user cluster in the cloud server according to the method of the present application; and

Fig. 8 is a schematic diagram of a content cluster in the cloud server according to the method of the present application.

In a first embodiment, a content recommendation method includes: generating, by an intelligent terminal, a score of a viewing content according to a behavior of a user towards the viewing content and sending the score to an edge server, in which the score is used to indicate the user's level of interest in the viewing content; determining, by the edge server, an interest feature vector of the user after receiving the score, and inquiring a content feature vector of the viewing content from a cloud server; updating the interest feature vector of the user according to the interest feature vector of the user, the content feature vector of the viewing content and the score, and sending the score and an updated interest feature vector of the user to the cloud server; determining, by the cloud server, a user cluster most matched with the user among all user clusters on the cloud server by using the updated interest feature vector of the user received, and adding a content corresponding to the user cluster into an initial recommendation queue, when the cloud server needs to recommend a network content to the user; and deleting a content towards which the user has generated a behavior from the initial recommendation queue, and after sequencing an updated initial recommendation queue, recommending first N contents in the updated initial recommendation queue to the user, in which N is a preset positive integer.

Alternatively, after the cloud server receives the score and the updated interest feature vector of the user, the method further includes: updating, by the cloud server, the content feature vector of the viewing content according to the received score, the content feature vector of the viewing content and the interest feature vector of the user before being updated/the updated interest feature vector of the user; and determining, by the cloud server, a content cluster most matched with the viewing content among all content clusters on the cloud server by using an updated content feature vector of the viewing content, and adding a content corresponding to the content cluster into the initial recommendation queue, when the cloud server determines that the user is interested in the viewing content according to the score.

Alternatively, determining the user cluster most matched with the user among all the user clusters on the cloud server includes: calculating a distance between the user and a center of each user cluster by using the updated interest feature vector of the user, and finding out a user cluster with a closest distance as the user cluster most matched with the user.

Alternatively, determining the content cluster most matched with the viewing content among all the content clusters on the cloud server includes: calculating a distance between the viewing content and a center of each content cluster by using the updated content feature vector of the viewing content, and finding out a content cluster with a closest distance as the content cluster most matched with the viewing content.

Alternatively, when the user is interested in the viewing content, the method further includes: adding the viewing content into a content queue corresponding to the user cluster most matched with the user.

Alternatively, after updating the content feature vector of the viewing content, the method further includes: removing the viewing content from a content cluster where the viewing content is currently located, and adding the viewing content into the content cluster most matched with the viewing content.

Alternatively, a way in which the cloud server divides user clusters includes: dividing, by the cloud server, all users into different clusters by utilizing a k-means algorithm according to interest feature vectors of all the users, in which a difference between interest feature vectors of different users in each cluster is within a first set range.

Alternatively, a way in which the cloud server divides content clusters includes: dividing, by the cloud server, all contents into different clusters by using a k-means algorithm according to content feature vectors of all the contents, wherein a difference between content feature vectors of different contents in each cluster is within a second set range.

Alternatively, determining, by the edge server, the interest feature vector of the user after receiving the score includes: inquiring locally, by the edge server, the interest feature vector of the user; and if the edge server fails to find the interest feature vector of the user, acquiring, by the edge server, the interest feature vector of the user from the cloud server.

In a second embodiment, a content recommendation system includes: an intelligent terminal, an edge server and a cloud server; in which the intelligent terminal is configured to generate a score of a viewing content according to a behavior of a user towards the viewing content and send the score to the edge server, in which the score is used to indicate the user's level of interest in the viewing content; the edge server is configured to determine an interest feature vector of the user after receiving the score, and inquire a content feature vector of the viewing content from the cloud server; and update the interest feature vector of the user according to the interest feature vector of the user, the content feature vector of the viewing content and the score, and send the score and an updated interest feature vector of the user to the cloud server; and the cloud server is configured to determine a user cluster most matched with the user among all user clusters on the cloud server by using the updated interest feature vector of the user received, and add a content corresponding to the user cluster into an initial recommendation queue, when the cloud server needs to recommend a network content to the user; and the cloud server is further configured to delete a content towards which the user has generated a behavior from the initial recommendation queue, and after sequencing an updated initial recommendation queue, recommend first N contents in the updated initial recommendation queue to the user; wherein N is a preset positive integer.

Alternatively, the cloud server includes: a receiver, a learning engine, a storage and a recommendation engine; in which the receiver is configured to receive the score and the updated interest feature vector of the user; the learning engine is configured to update the content feature vector of the viewing content according to the score, the content feature vector of the viewing content and the interest feature vector of the user before being updated/the updated interest feature vector of the user stored in the storage, and store an updated content feature vector of the viewing content in the storage; the storage is configured to store interest feature vectors of all users and content feature vectors of all contents; and the recommendation engine is configured to determine a user cluster most matched with the user among all the user clusters on the cloud server by using the updated interest feature vector of the user received, and add a content corresponding to the user cluster into an initial recommendation queue, when it is needed to recommend a network content to the user; the recommendation engine is further configured to determine a content cluster most matched with the viewing content among all content clusters on the cloud server by using an updated content feature vector of the viewing content, and add a content corresponding to the content cluster into the initial recommendation queue, when it is determined that the user is interested in the viewing content according to the score; and the recommendation engine is further configured to delete a content towards which the user has generated a behavior from the initial recommendation queue, and after sequence an updated initial recommendation queue, recommend first N contents in the updated initial recommendation queue to the user.

As can be seen from the above technical solutions, in the present application, the terminal generates a score of a viewing content according to a behavior of a user towards the viewing content and sends the score to an edge server; the edge server determines an interest feature vector of the user after receiving the score, and inquires a content feature vector of the viewing content from a cloud server; updates the interest feature vector of the user according to the interest feature vector of the user, the content feature vector of the viewing content and the score, and sends the score and an updated interest feature vector of the user to the cloud server; the cloud server determines a user cluster most matched with the user among all user clusters on the cloud server by using the updated interest feature vector of the user received, and adds a content corresponding to the user cluster into an initial recommendation queue, when the cloud server needs to recommend a network content to the user; and deletes a content towards which the user has generated a behavior from the initial recommendation queue, and after sequencing an updated initial recommendation queue, recommends first N contents in the updated initial recommendation queue to the user. Through the above processing, updating the interest feature vector of the user is carried out by the edge server, and the cloud server updates the content feature vectors and matches and selects contents to recommend, which significantly reduces the computing pressure of the cloud server.

To make the objects, technical schemes and advantages of the present application more clearly understood, the present application will be further described in detail with reference to the accompanying drawings.

According to a content recommendation method and system provided by the present application, a new way of content recommendation is provided through the cooperation of an edge server and a cloud server, in which the edge server is configured to calculate interest features of users, and the cloud server is configured to calculate features of contents, and can provide personalized contents to users through operations such as aggregation, filtering, sequencing and the like. By putting the calculation of the interest features of the users to the edge server, the pressure of the cloud server is relieved, thus the changes of the users' interests are quickly responded to, and the contents in which the users are interested are pushed to the users in time.

Fig. 1 is a schematic diagram of a basic flow of a content recommendation method in the present application, and Fig. 2 is a schematic diagram of a basic structure of a content recommendation system in the present application. Fig. 1 may be implemented in the system shown in Fig. 2. As shown in Fig. 1 and Fig. 2, the method includes the following steps.

At step 101, an intelligent terminal 201 may generate a score of a viewing content according to a behavior of a user towards the viewing content and may send the score to an edge server 203.

The processing of this step may be carried out by the intelligent terminal 201. When a user A views a content B (e.g. a short video), the intelligent terminal 201 collects a behavior of the user A towards the viewing content B and generates a score based thereon. The score is used to indicate a level at which the user A is interested in the viewing content B, i.e. user A's level of preference for the viewing content B, and preferably the greater the score is, the more the user A prefers for the viewing content B.

At step 102, after receiving the score, the edge server 203 may determine an interest feature vector of the user A, and may inquire a content feature vector of the viewing content B from a cloud server 202.

The interest feature vector of the user is used to characterize interest features of the user, and may include interests of various aspects of the user, and thus the interest feature vector is represented in vector form. The content feature vector is used to characterize specific features of the content, and may also include features of various aspects of the content, and thus the content feature vector is represented in vector form. For example, when a content is a certain movie, its content feature vector may include a movie title, a movie type, leading roles of the movie etc.

The edge server 203 firstly inquires the interest feature vector of the user A locally, and if failing to find it locally, the edge sever 203 acquires the interest feature vector of the user A from the cloud server 202. Further, the edge server 203 needs to inquire the content feature vector of the viewing content B from the cloud server 202.

At step 103, the edge server 203 may update the interest feature vector of the user A based on an old interest feature vector of the user A, the inquired content feature vector of the viewing content B, and the received score, and may send the score and an updated new interest feature vector of the user A to the cloud server 202.

At step 103, the interest feature vector of the user A may be updated at the edge server 203, in which the interest feature vector of the user A determined by the edge server 203 is called the old interest feature vector, and the updated interest feature vector of the user A is called the new interest feature vector. The edge server 203 may calculate the new interest feature vector of the user A using the old interest feature vector of the user A, the content feature vector of the viewing content B, and the received score. Specifically, the new interest feature vector may be computed by a collaborative filtering algorithm. After the new interest feature vector is obtained through the calculation, the new interest feature vector and the score received by the edge server 203 are sent to the cloud server 202, so that the cloud server 202 can update and keep the interest feature vector of the user.

At step 104, when the cloud server 202 needs to recommend a network content to the user A, the cloud server 202 may determine a user cluster most matched with the user A among all user clusters on the cloud server 202 by using the updated interest feature vector of the user A received, and may add contents corresponding to the user cluster into an initial recommendation queue.

In the cloud server 202, all users are divided into a plurality of user clusters, and different users in a same user cluster may have similar interests. Meanwhile, in the cloud server 202, all contents are also divided into a plurality of content clusters and different contents in a same content cluster may have similar features.

According to an embodiment, a method in which the cloud server 202 divides user clusters may include: dividing, by the cloud server 202, all users into different clusters by utilizing a k-means algorithm according to interest feature vectors of all the users, in which a difference between interest feature vectors of different users in each cluster is within a preset first range, which ensures that different users in each user cluster have similar interest feature vectors. Each user cluster has a content queue to keep contents in which users within the user cluster are interested.

According to an embodiment, a method in which the cloud server 202 divides content clusters may include: dividing, by the cloud server 202, all contents into different clusters by using a k-means algorithm according to content feature vectors of all the contents, in which a difference between content feature vectors of different contents in each cluster is within a preset second range, which ensures that different contents in each content cluster have similar content feature vectors.

When a recommended content is selected for the user A, basically, based on the user clusters, a user cluster most matched with the user A among all user clusters on the cloud server 202 is determined by using the updated interest feature vector of the user A received. Contents corresponding to the determined user cluster is then added into an initial recommendation queue. The determining the user cluster most matched with the user A is to calculate a distance between the user A and a center of each user cluster by using the updated interest feature vector of the user A, and to find out a user cluster with the closest distance as the user cluster most matched with the user A. Of course, other existing matching ways may be selected and will not be described in detail herein.

The initial recommendation queue, here a simplest initial recommendation queue, can be obtained in the way described above, and step 106 can then be performed directly. Alternatively, on the basis of the above initial recommendation queue, it is also preferable to further enrich the initial recommendation queue by the following step 105 and then perform the step 106.

At step 105, the cloud server 202 may update the content feature vector of the viewing content B according to the received score, the content feature vector of the viewing content B, and the interest feature vector of the user A. The cloud server 202 may also determine a content cluster most matched with the viewing content among all content clusters on the cloud server 202 by using an updated content feature vector of the viewing content B. The cloud server 202 may then add contents corresponding to the determined content cluster into the initial recommendation queue, when the cloud server 202 determines that the user A is interested in the viewing content B according to the score.

Here, after receiving the scores from the edge server 203, the cloud server 202 first updates the content feature vector of the viewing content B. Specifically, the content feature vector of the viewing content B is updated based on the received score, the content feature vector of the viewing content B before updating, and the interest feature vector of the user A before updating.

After the content feature vector is updated, if it is determined that the user A is interested in the viewing content B according to the score, the content cluster most matched with the viewing content B among all content clusters on the cloud server 202 is determined by using the updated content feature vector of the viewing content B, and the contents corresponding to the determined content cluster are added into the initial recommendation queue. The determining the content cluster most matched with the watching content B is to calculate the distance between the watching content B and the center of each content cluster by using the updated content feature vector of the watching content B, and to find out a content cluster with the closest distance as the content cluster most matched with the watching content B.

Still further preferably, when the user A is interested in the viewing content B, the viewing content B can be further added to the content queue corresponding to the user cluster most matched with the user A, and the viewing content B is removed from a content cluster where the viewing content B is currently located and added into the content cluster most matched with the viewing content B.

At step 106, a content towards which the user A has generated a behavior may be deleted from the initial recommendation queue, an updated initial recommendation queue may be sequenced, and first N contents in the updated initial recommendation queue may be recommended to the user A.

The processing of this step may be performed in the cloud server 202. The initial recommendation queue is filtered to remove contents on which the user A has generated behaviors. Further, the updated initial recommendation queue can be sequenced, and the sequencing may be set according to needs, such as sequencing according to novelty, popularity, diversity and the like. Finally the first N contents in the initial recommendation queue are recommended to the user A, namely being sent to the intelligent terminal 201.

Fig. 3 is a flow chart of a process at an intelligent terminal 301 in the content recommendation method of the present application.

The processing of step 101 of Fig. 1 may be performed at the intelligent terminal 301. A processing flow at the intelligent terminal 301 may be as shown in Fig. 3. The intelligent terminal 301 may include a receiver, a behavior collector, a behavior convertor, and a transmitter.

The receiver of the intelligent terminal 301 is mainly configured to receive a content pushed to a client by a cloud server 302 from the cloud server 302.

The behavior collector is mainly configured to collect a behavior of a user 305, such as whether the user 305 repeatedly views the content, whether the user 305 gives a "like" to the content, whether the user 305 forwards the content, whether the user 305 leaves after only viewing the content for one or two seconds, and so on.

The behavior convertor converts the behavior of the user 305 into a grade score to generate the user 305's level of preference for the content, in which the greater the score, the more the user 305 prefers for the content.

The transmitter of the intelligent terminal 301 is configured for the client to send the score of the content to the edge server 303.

Fig. 4 is a flow chat of a process at an edge server 403 in the content recommendation method of the present application.

The processing of steps 102and 103 of Fig. 1 described above may be performed at the edge server 403, as shown in Fig. 4. The edge sever 403 may include a receiver, a learning engine, and a transmitter.

The receiver of the edge server 403 receives the score of the user A towards the viewing content B from the intelligent terminal 401, searches for the old interest feature vector of the user A locally, and if failing to find it, the receiver inquires the old interest feature vector of the user A from the cloud server 402.

The receiver of the edge server 403 also inquires the content feature vector of the viewing content B viewed by the user A from the cloud server 402.

The learning engine of the edge server 403 calculates the new interest feature vector of the user A through the collected score of the content, the content feature vector, and the old interest feature vector of the user A using the collaborative filtering algorithm.

The transmitter of the edge server 403 sends the calculated new interest feature vector of the user A and the collected score of the content to the cloud server 402.

Fig. 5 is a schematic diagram of a process at a cloud server 502 in the content recommendation method of the present application.

The processing of the

steps

104, 105, and 106 of Fig. 1 may be performed at the cloud server 502. An algorithm model for calculating the content feature vector and a recommendation model are shown in Fig. 5. The cloud server 502 may include a receiver, a learning engine, a storage, and a recommendation engine.

The receiver of the cloud server 502 receives the score of the user towards the viewed content and the new interest feature vector of the user from the edge server 503.

The learning engine of the cloud server 502 learns the new content feature vector according to the score of the user towards the content, the user interest feature vector and the old content feature vector transmitted from the edge server 503 by using the collaborative filtering algorithm. The interest feature vector of the user used when the content feature vector is updated may be the old interest feature vector of the user or the new interest feature vector of the user, and experiments show that the new content feature vector obtained by using the old interest feature of the user is more accurate.

The storage of the cloud server 502 stores the interest feature vector of the user, and the content feature vector.

The recommendation engine divides the users into p clusters using the k-means algorithm according to interest feature vectors of all the users and each cluster has similar interest feature vectors; and divides the contents into k clusters by using the k-means algorithm according to the content feature vectors, and the contents of each cluster have similar feature vectors.

Fig. 6 is a flow chart of recommending contents in the cloud server 502 of Fig. 5, according to the method of the present application.

When a content is recommended to the user A, the recommendation engine calculates the distances between the user A and the centers of all p user clusters 630 by using the new interest feature vector 611 of the user sent by the edge server 503, finds a user cluster q 631 with the closest distance, and adds contents 651 corresponding to the user cluster q 631 into the initial recommendation queue 660, as shown in Fig. 6.

Whether the user A likes a content d is determined according to the score of the user A's level of preference for the content d received by the receiver, and if the user A likes the content d, the distances between the content d and the centers of all k content clusters 640 are calculated by using the new content feature vector 621 of the content d calculated by the learning engine of the cloud server 502, a content cluster e 641 with the closest distance to the content d is determined, and contents 652 corresponding to the content cluster e 641 are added into the initial recommendation queue 660, as shown in Fig. 6.

Fig. 7 is a schematic diagram of contents associated with a user cluster in the cloud server 502 according to the method of the present application.

If the user A 705 likes the content b 709, the distances between the user A 705 and the centers of all p user clusters 730 are calculated by using the new interest feature vector 711 of the user A 705 calculated by the edge server 503, the user cluster q 731 with the closest distance is found out, and the content b 709 is added into the content queue 751 corresponding to the user cluster q 731, as shown in Fig. 7.

Fig. 8 is a schematic diagram of a content cluster in the cloud server 502 according to the method of the present application.

The distances between the content d 809 and the centers of all k content clusters 840 are calculated by using the new content feature vector 821 of the content d 809 calculated by the learning engine of the cloud server 502, the content cluster e 841 with the closest distance is found out, and the content d 809 is removed from a content cluster where the content d 809 was previously located and is added into the content queue 852 corresponding to the content cluster e 841, as shown in Fig. 8.

Referring back to the Fig. 6, at step 670, the initial recommendation queue 660 is filtered to remove contents on which the user A has generated behaviors. At step 680, the contents in the initial recommendation queue 660 are then sequenced according to novelty, popularity and diversity. Then, the first N contents of the initial recommendation queue 660 are added into a final recommendation queue 690.

At this point, the flow of the content recommendation method in the present application may end. Through the processing according to the present application, each intelligent terminal may transmit the scores of a user towards contents (such as short videos) to the edge server, the edge server finishes the learning of the new interest feature vectors of the users based on a machine learning algorithm, the cloud server does not need to calculate the interest feature vectors of the users, and only needs to push new contents (such as short videos) to the users according to the new interest feature vectors transmitted by the edge server using the recommendation algorithm, which optimizes the response time.

The foregoing description is merely the preferred embodiments of the present invention and is not intended to limit the present invention, and any modification, equivalent replacement and improvement within the spirit and principle of the present invention should be contained in the protection scope of the present invention.

Claims

A content recommendation method, comprising:

obtaining, by an edge server, a score of a viewing content according to a behavior of a user towards the viewing content, wherein the score is used to indicate the user's level of interest in the viewing content;

determining, by the edge server, an interest feature vector of the user after obtaining the score, and inquiring a content feature vector of the viewing content from a cloud server;

updating, by the edge server, the interest feature vector of the user according to the acquired content feature vector of the viewing content and the obtained score;

sending, by the edge server, the obtained score and an updated interest feature vector of the user to the cloud server;

determining, by the cloud server, a user cluster most matched with the user among all user clusters on the cloud server by using the received updated interest feature vector of the user;

determining, by a cloud server, an initial recommendation queue including a content corresponding to the determined user cluster;

updating the initial recommendation queue by deleting a content towards which the user has generated a behavior from the initial recommendation queue;

sequencing the updated initial recommendation queue; and

recommending first N contents in the sequenced initial recommendation queue to the user, wherein N is a preset positive integer.
The method of claim 1, further comprising:

updating, by the cloud server, the content feature vector of the viewing content according to the received score and the received interest feature vector of the user; and

determining, by the cloud server, a content cluster most matched with the viewing content among all content clusters on the cloud server by using an updated content feature vector of the viewing content;

updating the initial recommendation queue by adding a content corresponding to the determined content cluster into the initial recommendation queue, when the cloud server determines that the user is interested in the viewing content according to the score.
The method of claim 1, wherein the determining of the user cluster most matched with the user among all the user clusters on the cloud server comprises: calculating a distance between the user and a center of each user cluster by using the updated interest feature vector of the user, and determining a user cluster with a closest distance as the user cluster most matched with the user.
The method of claim 2, wherein the determining of the content cluster most matched with the viewing content among all the content clusters on the cloud server comprises: calculating a distance between the viewing content and a center of each content cluster by using the updated content feature vector of the viewing content, and determining a content cluster with a closest distance as the content cluster most matched with the viewing content.
The method of claim 1 further comprising: adding the viewing content into a content queue corresponding to the user cluster most matched with the user, when the user is determined to be interested in the viewing content.
The method of claim 2 further comprising: removing, after updating the content feature vector of the viewing content, the viewing content from a content cluster where the viewing content is currently located, and adding the viewing content into the content cluster most matched with the viewing content.
The method of claim 1, wherein a way in which the cloud server divides user clusters comprises:

dividing, by the cloud server, all users into different clusters by using a k-means algorithm according to interest feature vectors of all the users, wherein a difference between interest feature vectors of different users in each cluster is within a preset first range.
The method of claim 2, wherein a way in which the cloud server divides content clusters comprises:

dividing, by the cloud server, all contents into different clusters by using a k-means algorithm according to content feature vectors of all the contents, wherein a difference between content feature vectors of different contents in each cluster is within a preset second range.
The method of claim 1, wherein the determining of the interest feature vector of the user after obtaining the score comprises:

inquiring locally, by the edge server, the interest feature vector of the user; and

acquiring, by the edge server, when the edge server fails to locally find the interest feature vector of the user, the interest feature vector of the user from the cloud server.
A content recommendation system, comprising: an edge server and a cloud server; wherein,

the edge server is configured to obtain a score of a viewing content according to a behavior of a user towards the viewing content, wherein the score is used to indicate the user's level of interest in the viewing content; determine an interest feature vector of the user after obtaining the score, and inquire a content feature vector of the viewing content from the cloud server; update the interest feature vector of the user according to the acquired content feature vector of the viewing content and the obtained score; and send the obtained score and an updated interest feature vector of the user to the cloud server; and

the cloud server is configured to determine a user cluster most matched with the user among all user clusters on the cloud server by using the received updated interest feature vector of the user; determine an initial recommendation queue including a content corresponding to the determined user cluster; update the initial recommendation queue by deleting a content towards which the user has generated a behavior from the initial recommendation queue; sequence the updated initial recommendation queue; and recommend first N contents in the sequenced initial recommendation queue to the user; wherein N is a preset positive integer.
The recommendation system according to claim 10, wherein the cloud server comprises: a receiver, a learning engine, a storage, and a recommendation engine; wherein,

the receiver is configured to receive the score and the updated interest feature vector of the user;

the learning engine is configured to update the content feature vector of the viewing content according to the received score and the received interest feature vector of the user stored in the storage, and store an updated content feature vector of the viewing content in the storage;

the storage is configured to store interest feature vectors of all users and content feature vectors of all contents; and

the recommendation engine is configured to: determine a user cluster most matched with the user among all the user clusters on the cloud server by using the received updated interest feature vector of the user; add a content corresponding to the determined user cluster into an initial recommendation queue; determine a content cluster most matched with the viewing content among all content clusters on the cloud server by using an updated content feature vector of the viewing content; add a content corresponding to the determined content cluster into the initial recommendation queue, when it is determined that the user is interested in the viewing content according to the score; delete a content towards which the user has generated a behavior from the initial recommendation queue; sequence the updated initial recommendation queue; and recommend first N contents in the sequenced initial recommendation queue to the user.
The recommendation system according to claim 11, wherein the recommendation engine is further configured to:

calculate a distance between the user and a center of each user cluster by using the updated interest feature vector of the user, and determine a user cluster with a closest distance as the user cluster most matched with the user; and

calculate a distance between the viewing content and a center of each content cluster by using the updated content feature vector of the viewing content, and determine a content cluster with a closest distance as the content cluster most matched with the viewing content.
The recommendation system according to claim 10, wherein the edge server comprises: a receiver, a learning engine, and a transmitter; wherein,

the receiver is configured to obtain the score and inquire the content feature vector of the viewing content from the cloud server;

the learning engine is configured to determine the interest feature vector of the user, and update the interest feature vector of the user according to the acquired content feature vector of the viewing content and the obtained score; and

the transmitter is configured to send the obtained score and an updated interest feature vector of the user to the cloud server.
A content recommendation method processed by a cloud server, comprising:

determining a content feature vector of a viewing content;

sending the content feature vector of the viewing content to an edge server;

receiving, from the edge server, an interest feature vector of a user updated based on the content feature vector of the viewing content, and a score of the viewing content indicating the user's level of interest in the viewing content;

determining a user cluster most matched with the user among all user clusters on the cloud server by using the received interest feature vector of the user;

determining an initial recommendation queue including a content corresponding to the determined user cluster;

updating the initial recommendation queue by deleting a content towards which the user has generated a behavior from the initial recommendation queue;

sequencing the updated initial recommendation queue; and

recommending first N contents in the sequenced initial recommendation queue to the user, wherein N is a preset positive integer.
The method of claim 14, further comprising:

updating the content feature vector of the viewing content according to the received score and the received interest feature vector of the user; and

determining a content cluster most matched with the viewing content among all content clusters on the cloud server by using an updated content feature vector of the viewing content;

updating the initial recommendation queue by adding a content corresponding to the determined content cluster into the initial recommendation queue, when the cloud server determines that the user is interested in the viewing content according to the score.