CN109800325A

CN109800325A - Video recommendation method, device and computer readable storage medium

Info

Publication number: CN109800325A
Application number: CN201811599231.3A
Authority: CN
Inventors: 蔡锦龙
Original assignee: Beijing Dajia Internet Information Technology Co Ltd
Current assignee: Beijing Dajia Internet Information Technology Co Ltd
Priority date: 2018-12-26
Filing date: 2018-12-26
Publication date: 2019-05-24
Anticipated expiration: 2038-12-26
Also published as: CN109800325B

Abstract

The application is about a kind of video recommendation method, device and computer storage medium.The video recommendation method includes: to establish clicking rate prediction model based on neural network algorithm；The similarity between the attributive character of target user and the attributive character of multiple historical users is calculated separately, is obtained and most similar first historical user of the attributive character of the target user；Based on the clicking rate prediction model, first historical user is obtained to the clicking rate of associated video；And based on first historical user to the clicking rate of the associated video, multiple video recommendations in the associated video are given to the target user.In the video recommendation method, the clicking rate for making full use of the attributive character of user to estimate user to associated video, new user or the less user of user behavior are estimated to the accuracy of the clicking rate of associated video to improve, and then improve the accuracy that video recommendations are carried out to the less user of new user or user behavior.

Description

Video recommendation method, device and computer readable storage medium

Technical field

The application belongs to computer software application field, especially video recommendation method and device.

Background technique

With science and technology be showing improvement or progress day by day and internet it is universal, more and more people transmitted by video information with share Life, the video recommendations of magnanimity personalization, which seem, to become more and more important.Application at present is more widely the method by machine learning To estimate user to targets such as the clicking rates of video.

In individualized video recommendation, be generally divided into two stages: the first stage is video triggering, according to the attribute of user And behavior, user is found out from video collection may interested video collection.Second stage is video sequence, passes through nerve net The video Ordering and marking that network model or strategy trigger out to the first stage, finds out the possible most interested video exhibition of user Now give user.Similar video is usually found according to the video that user browsed, or is found out according to the author that user pays close attention to Similar author, then similar video or/and the video recommendations of similar author to user.This method can not be used and be used The attribute at family carries out associated video recommendation, when a user be new user either user historical behavior it is less when, will so that Individualized video triggering is insufficient, reduces the accuracy rate of video recommendations.

Summary of the invention

The individualized video of the new user either less user of historical behavior is touched present in the relevant technologies to overcome The problem of hair deficiency causes the accuracy rate of video recommendations to reduce, the application discloses a kind of video recommendation method and device, based on Most similar first historical user of the attributive character of target user is to the clicking rate of associated video, by multiple views in associated video Frequency recommends target user, to realize accurate video recommendations.

According to the embodiment of the present application in a first aspect, providing a kind of video recommendation method, comprising:

Based on neural network algorithm, clicking rate prediction model is established；

Calculate separately the similarity between the attributive character of target user and the attributive character of multiple historical users, obtain with Most similar first historical user of the attributive character of the target user；

Based on the clicking rate prediction model, first historical user is obtained to the clicking rate of associated video；And

Based on first historical user to the clicking rate of the associated video, by multiple videos in the associated video Recommend the target user.

Optionally, the attributive character of user includes: User ID feature, static nature and behavioral characteristics.

Optionally, the static nature of user includes at least one of following characteristics: the age, gender, geographical location, The application list that IP address, mobile phone model, mobile phone are installed；

The behavioral characteristics of user include at least one of following characteristics: user clicks history feature, user thumbs up History feature and user pay close attention to list characteristics.

Optionally, video has video features, and the video features include at least one of following characteristics: video ID is special It seeks peace video author's ID feature.

Optionally, the phase between the attributive character for calculating separately target user and the attributive character of multiple historical users Like degree, obtain with most similar first historical user of the attributive character of the target user, including；

Extract the attributive character of the target user and the attributive character of the multiple historical user；

Calculate separately the target user the attributive character and the multiple historical user the attributive character it Between distance；And

The distance is ranked up, obtains using with most similar first history of the attributive character of the target user Family.

Optionally, described to be based on the clicking rate prediction model, first historical user is obtained to the point of associated video Hit rate, comprising:

Extract the attributive character of first historical user and the video features of the associated video；

The video ID feature of the static nature and the associated video to first historical user carry out to Amount summation, obtains fisrt feature；

The video author ID feature of the static nature and the associated video to first historical user into Row vector summation, obtains second feature；

By the User ID feature and institute of the fisrt feature and the second feature and first historical user The behavioral characteristics of the first historical user are stated as third feature；And

The third feature is inputted into the clicking rate prediction model, obtains first historical user to the related view The clicking rate of frequency.

Optionally, it is described based on first historical user to the clicking rate of the associated video, by the associated video In multiple video recommendations give the target user, comprising:

According to first historical user to the clicking rate of the associated video, to the clicking rate backward of the associated video Sequence；

It is sorted according to the clicking rate backward of the associated video, by multiple views of front in the associated video Frequency recommends the target user.

Optionally, described to be based on neural network algorithm, establish clicking rate prediction model, comprising:

Extract the attributive character of sample of users；

It extracts the video features of Sample video and marks video tab for the Sample video；

Based on neural network algorithm, establishes clicking rate and estimate object module；

The video features of the attributive character and the Sample video based on the sample of users, to the click Rate is estimated before object module carries out to learning；And

The video features of the attributive character and the Sample video based on the sample of users, to the click Rate estimates object module and carries out backward learning.

Optionally, described to be based on neural network algorithm, establish clicking rate prediction model, further includes:

The video ID feature of the static nature and the Sample video to the sample of users carries out vector and asks With obtain fourth feature；

The video author ID feature of the static nature and the Sample video to the sample of users carry out to Amount summation, obtains fifth feature；And

By the User ID feature of the fourth feature and the fifth feature and the historical user and described go through The behavioral characteristics of history user are as sixth feature.

Optionally, the video of the attributive character based on the sample of users and the Sample video is special Sign estimates before object module carries out to learning the clicking rate, comprising:

The sixth feature is inputted into the clicking rate and estimates object module；

Object module is estimated in the clicking rate from bottom to top successively to convert the sixth feature, obtains the point The rate of hitting estimates the top layer vector of object module；

The top layer vector is converted to the probability of clicking rate.

Optionally, the video of the attributive character based on the sample of users and the Sample video is special Sign estimates object module to the clicking rate and carries out backward learning, comprising:

According to the video tab of the probability of the clicking rate and the Sample video, calculates the clicking rate and estimate target mould The loss function of type；

The loss function that the clicking rate estimates object module is minimized using stochastic gradient descent method；

Solve the gradient that the clicking rate estimates the loss function of object module；

From top to bottom successively update the network parameter that the clicking rate estimates object module；And

With the newly corresponding network parameter of the sixth feature.

It is optionally, described to mark video tab for the Sample video, comprising:

If the sample of users clicks the Sample video of operation pages displaying, the Sample video is marked For positive sample；

If the sample of users does not have the Sample video of clicking operation page presentation, by the Sample video mark Note is negative sample.

According to the second aspect of the embodiment of the present application, a kind of video recommendations device is provided, comprising: include:

Model foundation unit establishes clicking rate prediction model for being based on neural network algorithm；

Nearest _neighbor retrieval unit, for calculating separately the attributive character of target user and the attributive character of multiple historical users Between similarity, obtain and most similar first historical user of the attributive character of the target user；

Clicking rate estimates unit, for being based on the clicking rate prediction model, obtains first historical user to correlation The clicking rate of video；And

Video recommendations unit, for the clicking rate based on first historical user to the associated video, by the phase Multiple video recommendations in video are closed to the target user.

Optionally, the nearest _neighbor retrieval unit, comprising:

Fisrt feature extraction unit, for extract the target user the attributive character and the multiple historical user The attributive character；

Metrics calculation unit, for calculate separately the target user the attributive character and the multiple historical user The distance between the attributive character；And

First sequencing unit obtains the attributive character with the target user for being ranked up to the distance Most similar first historical user.

Optionally, the clicking rate estimates unit, comprising:

Second feature extraction unit, for extract first historical user the attributive character and the associated video The video features；

Fisrt feature integrated unit, for the static nature and the associated video to first historical user The video ID feature carries out vector summation, obtains fisrt feature；

Unit is estimated, for the third feature to be inputted the clicking rate prediction model, first history is obtained and uses Clicking rate of the family to the associated video.

Optionally, the video recommendations unit, comprising:

Second sequencing unit, for the clicking rate according to first historical user to the associated video, to the phase Close the clicking rate backward sequence of video；

Recommendation unit will be suitable in the associated video for being sorted according to the clicking rate backward of the associated video The forward multiple video recommendations of sequence give the target user.

Optionally, the model foundation unit, comprising:

Third feature extraction unit, for extracting the attributive character of sample of users；

Object module establishes unit, for being based on neural network algorithm, establishes clicking rate and estimates object module；

Forward direction unit, the view for the attributive character and the Sample video based on the sample of users Frequency feature estimates before object module carries out to learning the clicking rate；And

Backward learning unit, the view for the attributive character and the Sample video based on the sample of users Frequency feature estimates object module to the clicking rate and carries out backward learning.

Optionally, the model foundation unit, further includes:

Second feature integrated unit, for described in the static nature and the Sample video to the sample of users Video ID feature carries out vector summation, obtains fourth feature；

The top layer vector is converted to the probability of clicking rate.

With the newly corresponding network parameter of the sixth feature.

It is optionally, described to mark video tab for the Sample video, comprising:

According to a third aspect of the embodiments of the present invention, a kind of video recommendations device is provided, comprising:

Processor；

Memory for storage processor executable instruction；

Wherein, the processor is configured to executing video recommendation method described in above-mentioned any one.

According to a fourth aspect of the embodiments of the present invention, a kind of computer readable storage medium is provided, which is characterized in that described Computer-readable recording medium storage has computer instruction, and the computer instruction, which is performed, realizes above-mentioned video recommendations side Method.

The technical solution that embodiments herein provides can include the following benefits:

Calculate separately the similarity between the attributive character of target user and the attributive character of each historical user, obtain with Most similar first historical user of the attributive character of the target user.The N layer neural net layer of the clicking rate prediction model is to this The attributive character and user behavior characteristics of first historical user and the video features of associated video are converted layer by layer, final To first historical user to the clicking rate of associated video.According to first historical user to the clicking rate of associated video, by this First historical user gives the target user to the higher multiple video recommendations of the clicking rate of associated video.Make full use of user's Attributive character estimates user to the clicking rate of associated video, estimates new user or the less user of user behavior to improve To the accuracy of the clicking rate of associated video, and then improves to the less user of new user or user behavior and carry out video recommendations Accuracy.

Calculate separately the distance between attributive character and the attributive character of multiple historical user of the target user.To this Distance is ranked up, obtain with most similar first historical user of the attributive character of the target user, improve lookup with the mesh The accuracy for marking most similar first historical user of attributive character of user, to further improve the accurate of video recommendations Property.

It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not The application can be limited.

Detailed description of the invention

Fig. 1 is the flow chart of video recommendation method shown according to an exemplary embodiment.

Fig. 2 is the flow chart of video recommendation method shown according to an exemplary embodiment.

Fig. 3 is the flow chart of video recommendation method shown according to an exemplary embodiment.

Fig. 4 is the flow chart of video recommendation method shown according to an exemplary embodiment.

Fig. 5 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.

Fig. 6 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.

Fig. 7 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.

Fig. 8 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.

Fig. 9 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.

Figure 10 is a kind of block diagram of device for executing video recommendation method shown according to an exemplary embodiment.

Figure 11 is a kind of block diagram of device for executing video recommendation method shown according to an exemplary embodiment.

Specific embodiment

Example embodiments are described in detail here, and the example is illustrated in the accompanying drawings.Following description is related to When attached drawing, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements.Following exemplary embodiment Described in embodiment do not represent all embodiments consistent with the application.On the contrary, they be only with it is such as appended The example of the consistent device and method of some aspects be described in detail in claims, the application.

Fig. 1 is the flow chart of video recommendation method shown according to an exemplary embodiment, specifically includes the following steps:

In step s101, it is based on neural network algorithm, establishes clicking rate prediction model.

In this step, it is based on neural network algorithm, establishes clicking rate prediction model.The clicking rate prediction model is to connect entirely Connect neural network.Such as the clicking rate prediction model shares N layers of neural net layer.N-1 layers any one node, all with N All node connections of layer.

In step s 102, it calculates separately between the attributive character of target user and the attributive character of multiple historical users Similarity obtains and most similar first historical user of the attributive character of the target user.

In this step, the phase between the attributive character of target user and the attributive character of each historical user is calculated separately Like degree, obtain and most similar first historical user of the attributive character of the target user.E.g. by calculating target user's Euclidean distance between attributive character and the attributive character of each historical user carries out arest neighbors inspection using the Euclidean distance Rope, obtains and most similar first historical user of the attributive character of the target user.

In step s 103, it is based on the clicking rate prediction model, obtains first historical user to associated video Clicking rate.

In this step, based on the clicking rate prediction model established in step S101, by by first historical user's The video features of attributive character, user behavior characteristics and associated video input the clicking rate prediction model.The clicking rate is estimated The N layer neural net layer of model carries out layer by layer the video features of the attributive character, the user behavior characteristics and associated video Transformation, finally obtains first historical user to the clicking rate of associated video.

In step S104, based on first historical user to the clicking rate of the associated video, by the related view Multiple video recommendations in frequency give the target user.

In this step, according to first historical user in step S103 to the clicking rate of associated video, by this first Historical user gives the target user to the higher multiple video recommendations of the clicking rate of associated video.

According to the embodiment of the present application, calculate separately target user attributive character and each historical user attributive character it Between similarity, obtain and most similar first historical user of the attributive character of the target user.The N of the clicking rate prediction model Layer neural net layer to the video features of the attributive character and user behavior characteristics of first historical user and associated video into Row converts layer by layer, finally obtains first historical user to the clicking rate of associated video.According to first historical user to correlation The clicking rate of video uses the higher multiple video recommendations of the clicking rate of first historical user to associated video to the target Family.The clicking rate for making full use of the attributive character of user to estimate user to associated video, thus improve estimate new user or The less user of user behavior improves less to new user or user behavior the accuracy of the clicking rate of associated video User carry out video recommendations accuracy.

Fig. 2 is the flow chart of video recommendation method shown according to an exemplary embodiment, the step S102 of specifically Fig. 1 In, calculate separately the similarity between the attributive character of target user and the attributive character of multiple historical users, obtain with it is described The process of most similar first historical user of the attributive character of target user, comprising the following steps:

In step s 201, the attributive character of the target user and the category of the multiple historical user are extracted Property feature.

The feature of one user includes attributive character and the behavioural characteristic of user.The attributive character include User ID feature, Static nature and behavioral characteristics.The static nature of user includes at least one of following characteristics: age, gender, geographical position It sets, the application list of IP address, mobile phone model, mobile phone installation.The behavioral characteristics of user include in following characteristics extremely One of few: user clicks history feature, user thumbs up history feature and user pays close attention to list characteristics.

The attributive character of the target user and the attributive character of multiple historical users are extracted in this step.

In step S202, the attributive character of the target user and the institute of the multiple historical user are calculated separately State the distance between attributive character.

Several distance calculating methods of common characterization similarity include: Euclidean distance, manhatton distance, mahalanobis distance, Cosine similarity and Hamming distance.Wherein, Euclidean distance is point-to-point transmission in Euclidean space " common " (i.e. straight line) distance.It is graceful Hatton's distance be Euclidean space fixation rectangular co-ordinate fasten two o'clock be formed by projection that line segment generates axis away from From summation.The covariance distance of mahalanobis distance expression data.The cosine value that cosine similarity passes through the angle of two vectors of measurement To measure the similitude between them.In information theory, the Hamming distance between two isometric character strings is two character strings pair Answer the number of the kinds of characters of position.

In this step, such as target user is X, and m historical user is respectively Y¹,Y²,Y³,…,Y^m.Calculate separately this Euclidean distance between the attributive character of target user X and the attributive character of the m historical user.In Euclid's sky Between in, the attributive character of target user X is x=(x₁,x₂,x₁,..,x_n) and i-th of historical user YⁱAttributive characterBetween Euclidean distance be

Wherein, x₁,x₂,x₁,..,x_nIt is the attributive character of target user X,It is i-th of historical user YⁱAttributive character.

In step S203, the distance is ranked up, obtains the attributive character most phase with the target user The first close historical user.

In this step, the category of the attributive character to the target user obtained in step S202 and multiple historical user Property the distance between feature be, for example, that Euclidean distance is ranked up, obtain it is the smallest this apart from corresponding historical user.The distance The smaller attributive character characterized between the target user and the historical user is more close.This is the smallest apart from corresponding history use Family is the first historical user.

Attribute according to the embodiment of the present application, the attributive character and multiple historical user that calculate separately the target user is special The distance between sign.The distance is ranked up, obtain with most similar first historical user of the attributive character of the target user, The accuracy searched with most similar first historical user of attributive character of the target user is improved, to further improve The accuracy of video recommendations.

Fig. 3 is the flow chart of video recommendation method shown according to an exemplary embodiment, step in specifically Fig. 1 S103- step S104 is based on the clicking rate prediction model, obtain first historical user to the clicking rate of associated video and Based on first historical user to the clicking rate of the associated video, by multiple video recommendations in the associated video to institute State the process of target user.The following steps are included:

In step S301, the attributive character of first historical user and the view of the associated video are extracted Frequency feature.

The feature of first historical user includes attributive character and the behavioural characteristic of user.The attributive character includes User ID Feature, static nature and behavioral characteristics.The static nature of first historical user includes at least one of following characteristics: year The application list that age, gender, geographical location, IP address, mobile phone model, mobile phone are installed.This of first historical user is dynamic State feature includes at least one of following characteristics: user clicks history feature, user thumbs up history feature and user pays close attention to column Table feature.The video features of the associated video include at least one of following characteristics: video ID feature and video author ID are special Sign.

The attributive character of first historical user and the attributive character of the associated video are extracted in this step.

In step s 302, the video of the static nature to first historical user and the associated video ID feature carries out vector summation, obtains fisrt feature.

In this step, the video ID feature of the static nature to first historical user and each associated video carry out to Amount summation, obtains the fisrt feature about each associated video of first historical user.

In step S303, the video of the static nature and the associated video to first historical user Author's ID feature carries out vector summation, obtains second feature.

In this step, the video author ID feature of the static nature to first historical user and each associated video into Row vector summation, obtains the second feature about each associated video of first historical user.

In step s 304, by the fisrt feature and the second feature and the use of first historical user The behavioral characteristics of family ID feature and first historical user are as third feature.

In this step, by the fisrt feature obtained in step S302 and step S303 and the second feature and this The User ID feature of one historical user and the behavioral characteristics of first historical user are as first historical user about each The third feature of a associated video.

In step S305, the third feature is inputted into the clicking rate prediction model, first history is obtained and uses Clicking rate of the family to the associated video.

In this step, by the third about each associated video of first historical user obtained in step s304 Feature inputs the clicking rate prediction model.By the transformation layer by layer of the clicking rate prediction model, first historical user couple is obtained The clicking rate of each associated video.

In step S306, according to first historical user to the clicking rate of the associated video, to the related view The clicking rate backward of frequency sorts.

In this step, by first historical user obtained in step S305 to the clicking rate of each associated video into The sequence of row backward, the backward for obtaining the clicking rate sort the recommendation order of corresponding associated video.

In step S307, sorted according to the clicking rate backward of the associated video, it will be suitable in the associated video The forward multiple video recommendations of sequence give the target user.

In this step, the backward of the clicking rate according to obtained in step S306 sorts the recommendation of corresponding associated video Sequentially, one or more video recommendations of front in the associated video are given to the target user.

According to the embodiment of the present application, the video ID feature of static nature and each associated video to first historical user Vector summation is carried out, the fisrt feature about each associated video of first historical user is obtained.To first historical user Static nature and each associated video video author's ID feature carry out vector summation, obtain first historical user about The second feature of each associated video.By the fisrt feature and the second feature and the User ID feature of first historical user The third feature about each associated video with the behavioral characteristics of first historical user as first historical user.Base In the third feature about each associated video of first historical user, pass through the change layer by layer of the clicking rate prediction model It changes, obtains first historical user to the clicking rate of each associated video.First historical user regards each correlation The clicking rate of frequency carries out backward sequence, and the backward for obtaining the clicking rate sorts the recommendation order of corresponding associated video.By the phase One or more video recommendations of front in video are closed to the target user.By the attributive character of first historical user Fusion treatment is carried out with the video features of associated video, improves and estimates first historical user to the clicking rate of the associated video Accuracy, to further improve the accuracy of video recommendations.

Fig. 4 is the flow chart of video recommendation method shown according to an exemplary embodiment, step S101 in specifically Fig. 1 In, it is based on neural network algorithm, establishes the process of clicking rate prediction model.The following steps are included:

In step S401, the attributive character of sample of users and the video features of Sample video are extracted and for institute State Sample video mark video tab.

The feature of the sample of users includes attributive character and the behavioural characteristic of user.The attributive character includes User ID spy Sign, static nature and behavioral characteristics.The static nature of the sample of users includes at least one of following characteristics: age, property Not, the application list that geographical location, IP address, mobile phone model, mobile phone are installed.The behavioral characteristics of the sample of users include At least one of following characteristics: user clicks history feature, user thumbs up history feature and user pays close attention to list characteristics.The sample The video features of this video include at least one of following characteristics: video ID feature and video author's ID feature.

In this step, the attributive character of sample of users and the video features of Sample video are extracted.It and is the sample This video labeling video tab, comprising: if the sample of users clicks the Sample video of operation pages displaying, the sample Video labeling is positive sample.If the sample of users does not have the Sample video of clicking operation page presentation, which is regarded Frequency marking note is negative sample.In one embodiment, the video tab of the positive sample is labeled as 1, by the video mark of the negative sample Label are labeled as 0.

In step S402, the video ID of the static nature and the Sample video to the sample of users is special Sign carries out vector summation, obtains fourth feature.

In this step, vector is carried out to the video ID feature of the static nature of the sample of users and each Sample video Summation, obtains the fourth feature about each Sample video of the sample of users.

In step S403, the video author of the static nature and the Sample video to the sample of users ID feature carries out vector summation, obtains fifth feature.

In this step, video author's ID feature of the static nature of the sample of users and each Sample video is carried out Vector summation, obtains the fifth feature about each Sample video of the sample of users.

In step s 404, by the fourth feature and the fifth feature and the User ID of the historical user The behavioral characteristics of feature and the historical user are as sixth feature.

In this step, by the fourth feature obtained in step S402 and step S403 and the fifth feature and the sample The behavioral characteristics of the User ID feature of this user and the sample of users, as the sample of users about each Sample video Sixth feature.

In step S405, it is based on neural network algorithm, clicking rate is established and estimates object module.

In this step, it is based on neural network algorithm, clicking rate is established and estimates object module.The clicking rate estimates target mould Type is full Connection Neural Network.Such as the clicking rate estimates object module and shares N layers of neural net layer.Any one of N-1 layers Node is all connected with all nodes of n-th layer.

In step S406, the video of the attributive character and the Sample video based on the sample of users is special Sign estimates before object module carries out to learning the clicking rate.

In this step, the sixth feature about each Sample video of the sample of users in step S404 is inputted The clicking rate in step S405 estimates object module.Object module is estimated by the sample of users about each in the clicking rate The sixth feature of a Sample video is from bottom to top successively converted, obtain the sample of users about each Sample video The clicking rate estimates the top layer vector of object module.By top layer vector be converted to the sample of users about each Sample video Clicking rate probability.

For example, clicking rate about each Sample video of the sample of users to be estimated to the top layer vector of object module The calculation formula for being converted to the probability of clicking rate is sigmoid function:

Wherein, aⁱFor the sample of users the clicking rate about i-th of Sample video estimate the top layer of object module to Amount, σ (aⁱ) it is top layer vector aⁱCorresponding probability, value range is (0,1).

In step S 407, the video of the attributive character based on the sample of users and the Sample video is special Sign estimates object module to the clicking rate and carries out backward learning.

In this step, according to the probability of the clicking rate about each Sample video of the sample of users in step S406 With the video tab of each Sample video, the clicking rate about each Sample video for calculating the sample of users is estimated The loss function of object module.Being somebody's turn to do about each Sample video of the sample of users is minimized using stochastic gradient descent method Clicking rate estimates the loss function of object module.The clicking rate about each Sample video for solving the sample of users is estimated The gradient of the loss function of object module.Target is estimated by clicking rate about each Sample video of the sample of users The gradient of the loss function of model from top to bottom successively updates the network parameter that the clicking rate estimates object module.Pass through the sample Clicking rate about each Sample video of this user estimates the gradient of the loss function of object module, uses with the new sample The corresponding network parameter of sixth feature about each Sample video at family.

For example, the loss function for estimating object module about the clicking rate of each Sample video of the sample of users The calculation formula of (Log Loss) are as follows:

L=-yⁱlogpⁱ-(1-yⁱ)log(1-pⁱ) (3)

Wherein, pⁱ=σ (aⁱ) be the sample of users the clicking rate about i-th of Sample video probability, σ is sigmoid Function, yⁱ∈ { 0,1 } is the video tab of i-th of Sample video of the sample of users.

According to the embodiment of the present application, to the video ID feature of the static nature of the sample of users and each Sample video into Row vector summation, obtains the fourth feature about each Sample video of the sample of users.It is special to the static state of the sample of users Seek peace each Sample video video author's ID feature carry out vector summation, obtain the sample of users about each sample regard The fifth feature of frequency.By the User ID feature and the sample of users of the fourth feature and the fifth feature and the sample of users Behavioral characteristics, the sixth feature about each Sample video as the sample of users.By the sample of users about every The sixth feature of one Sample video trains clicking rate to estimate object module, adjusts, optimizes the clicking rate and estimate object module With the network parameter of the sixth feature about each Sample video of the sample of users, to establish clicking rate prediction model. The accuracy of the clicking rate prediction model of foundation is improved, to further improve the accuracy rate of video recommendations.

Fig. 5 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.As shown in figure 5, the device 50 It include: that model foundation unit 501, nearest _neighbor retrieval unit 502, clicking rate estimate unit 503 and video recommendations unit 504.

Model foundation unit 501 establishes clicking rate prediction model for being based on neural network algorithm.

The unit is based on neural network algorithm, establishes clicking rate prediction model.The clicking rate prediction model is full connection mind Through network.Such as the clicking rate prediction model shares N layers of neural net layer.N-1 layers any one node, all with n-th layer institute There is node connection.

Nearest _neighbor retrieval unit 502, for calculating separately the attributive character of target user and the attribute of multiple historical users Similarity between feature obtains and most similar first historical user of the attributive character of the target user.

The unit calculates separately the similarity between the attributive character of target user and the attributive character of each historical user, It obtains and most similar first historical user of the attributive character of the target user.It is e.g. special by calculating the attribute of target user Euclidean distance between the attributive character of each historical user of seeking peace carries out nearest _neighbor retrieval using the Euclidean distance, obtains With most similar first historical user of attributive character of the target user.

Clicking rate estimates unit 503, for being based on the clicking rate prediction model, obtains first historical user to phase Close the clicking rate of video.

The clicking rate prediction model that the unit is established based on model foundation unit 501, by by first historical user's The video features of attributive character, user behavior characteristics and associated video input the clicking rate prediction model.The clicking rate is estimated The N layer neural net layer of model carries out layer by layer the video features of the attributive character, the user behavior characteristics and associated video Transformation, finally obtains first historical user to the clicking rate of associated video.

Video recommendations unit 504 will be described for the clicking rate based on first historical user to the associated video Multiple video recommendations in associated video give the target user.

The unit estimates first historical user that unit 503 obtains to the clicking rate of associated video according to clicking rate, will First historical user gives the target user to the higher multiple video recommendations of the clicking rate of associated video.

Fig. 6 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.Arest neighbors is examined in specifically Fig. 5 The schematic diagram of cable elements 502.As shown in fig. 6, the device 60 includes: fisrt feature extraction unit 601, metrics calculation unit 602 With the first sequencing unit 603.

Fisrt feature extraction unit 601, for extracting the attributive character and the multiple history of the target user The attributive character of user.

Fisrt feature extraction unit 601 extracts the attributive character of the target user and the attribute spy of multiple historical users Sign.

Metrics calculation unit 602, for calculating separately the attributive character and the multiple history of the target user The distance between described attributive character of user.

For example, target user is X, m historical user is respectively Y¹,Y²,Y³,…,Y^m.Metrics calculation unit 602 is counted respectively Calculate the Euclidean distance between the attributive character of the target user X and the attributive character of the m historical user.

In Euclidean space, the attributive character of target user X is x=(x₁,x₂,x₁,..,x_n) and i-th of history use Family YⁱAttributive characterBetween Euclidean distance be

Wherein, x₁,x₂,x₁,..,x_nIt is the X attributive character of target user,It is i-th of historical user YⁱAttributive character.

First sequencing unit 603 obtains and the attribute of target user spy for being ranked up to the distance Levy most similar first historical user.

First sequencing unit 603 adjust the distance the target user that computing unit 602 obtains attributive character and multiple go through The distance between attributive character of history user is, for example, that Euclidean distance is ranked up, and obtains the smallest this and uses apart from corresponding history Family.This is more close apart from the smaller attributive character characterized between the target user and the historical user.The smallest distance is corresponding The historical user be the first historical user.

Fig. 7 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.Clicking rate is pre- in specifically Fig. 5 Estimate the schematic diagram of unit 503.As shown in fig. 7, the device 70 includes: second feature extraction unit 701, fisrt feature integrated unit 702 and estimate unit 703.

Second feature extraction unit 701, for extract the attributive character of first historical user to it is described related The video features of video.

The feature of first historical user includes attributive character and the behavioural characteristic of user.The attributive character includes User ID Feature, static nature and behavioral characteristics.The static nature of first historical user includes at least one of following characteristics: year The application list that age, gender, geographical location, IP address, mobile phone model, mobile phone are installed.This of first historical user is dynamic State feature includes at least one of following characteristics: user clicks history feature, user thumbs up history feature and user pays close attention to column Table feature.The video features of the associated video include at least one of following characteristics: video ID feature, video author's ID feature With video tab feature.

Second feature extraction unit 701 extracts the attributive character of first historical user and the attribute of the associated video Feature.

Fisrt feature integrated unit 702, for first historical user the static nature to it is described it is related view The video ID feature of frequency carries out vector summation, obtains fisrt feature.To the static nature of first historical user Vector summation is carried out with the video author ID feature of the associated video, obtains second feature.By the fisrt feature and The dynamic of the User ID feature and first historical user of the second feature and first historical user Feature is as third feature.

Fisrt feature integrated unit 702 is special to the static nature of first historical user and the video ID of each associated video Sign carries out vector summation, obtains the fisrt feature about each associated video of first historical user.First history is used Video author's ID feature of the static nature at family and each associated video carries out vector summation, obtains the pass of first historical user In the second feature of each associated video.The fisrt feature and the second feature and the User ID of first historical user is special It seeks peace the behavioral characteristics of first historical user, the third about each associated video as first historical user is special Sign.

Unit 703 is estimated, for the third feature to be inputted the clicking rate prediction model, obtains first history Clicking rate of the user to the associated video.

Unit 703 is estimated by first historical user obtained in fisrt feature integrated unit 702 about each phase The third feature for closing video inputs the clicking rate prediction model.By the transformation layer by layer of the clicking rate prediction model, obtain this Clicking rate of one historical user to each associated video.

Fig. 8 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.Video recommendations in specifically Fig. 5 The schematic diagram of unit 504.As shown in figure 8, the device 80 includes: the second sequencing unit 801 and recommendation unit 802.

Second sequencing unit 801, for the clicking rate according to first historical user to the associated video, to described The clicking rate backward of associated video sorts.

Second sequencing unit 801 will estimate first historical user obtained in unit 703 to each associated video Clicking rate carries out backward sequence, and the backward for obtaining the clicking rate sorts the recommendation order of corresponding associated video.

Recommendation unit 802 will be in the associated video for being sorted according to the clicking rate backward of the associated video Multiple video recommendations of front give the target user.

The corresponding related view of the backward sequence of the clicking rate according to obtained in the second sequencing unit 801 of recommendation unit 802 The recommendation order of frequency gives one or more video recommendations of front in the associated video to the target user.

Fig. 9 is the schematic diagram of video recommendations device shown according to an exemplary embodiment.Model foundation in specifically Fig. 5 The schematic diagram of unit 501.As shown in figure 9, the device 90 includes: third feature extraction unit 901, second feature integrated unit 902, object module establishes unit 903, forward direction unit 904 and backward learning unit 905.

Third feature extraction unit 901 extracts the described of Sample video for extracting the attributive character of sample of users Video features simultaneously mark video tab for the Sample video.

Third feature extraction unit 901 extracts the attributive character of sample of users and the video features of Sample video.And And video tab is marked for the Sample video, comprising: if the sample of users clicks the Sample video of operation pages displaying, Then the Sample video is labeled as positive sample.It, will if the sample of users does not have the Sample video of clicking operation page presentation The Sample video is labeled as negative sample.In one embodiment, the video tab of the positive sample is labeled as 1, by the negative sample Video tab be labeled as 0.

Second feature integrated unit 902, for the static nature and the Sample video to the sample of users The video ID feature carries out vector summation, obtains fourth feature.To the static nature and the sample of the sample of users The video author ID feature of this video carries out vector summation, obtains fifth feature.By the fourth feature and the described 5th The behavioral characteristics of the User ID feature and the historical user of feature and the historical user are as sixth feature.

Video ID feature of the second feature integrated unit 902 to the static nature of the sample of users and each Sample video Vector summation is carried out, the fourth feature about each Sample video of the sample of users is obtained.To the static state of the sample of users Video author's ID feature of feature and each Sample video carries out vector summation, obtain the sample of users about each sample The fifth feature of video.By the User ID feature and the sample of users of the fourth feature and the fifth feature and the sample of users Behavioral characteristics, the sixth feature about each Sample video as the sample of users.

Object module establishes unit 903, for being based on neural network algorithm, establishes clicking rate and estimates object module.

Object module establishes unit 903 and is based on neural network algorithm, establishes clicking rate and estimates object module.The clicking rate is pre- Estimating object module is full Connection Neural Network.Such as the clicking rate estimates object module and shares N layers of neural net layer.N-1 layers Any one node is all connected with all nodes of n-th layer.

Forward direction unit 904, the institute for the attributive character and the Sample video based on the sample of users Video features are stated, the clicking rate is estimated before object module carries out to learning.

Forward direction unit 904, the sample of users that second feature integrated unit 902 is obtained about each sample The clicking rate that the sixth feature input object module of video establishes the foundation of unit 903 estimates object module.It is pre- in the clicking rate Estimate object module from bottom to top successively to convert the sixth feature about each Sample video of the sample of users, obtain Clicking rate about each Sample video of the sample of users estimates the top layer vector of object module.By the sample of users The top layer vector about each Sample video is converted to the clicking rate about each Sample video of the sample of users Probability.

For example, the top layer vector about each Sample video of the sample of users to be converted to the meter of the probability of clicking rate Calculation formula is sigmoid function:

Backward learning unit 905, the institute for the attributive character and the Sample video based on the sample of users Video features are stated, object module is estimated to the clicking rate and carries out backward learning.

Backward learning unit 905 is according to the preceding sample of users into unit 904 about each Sample video The video tab of the probability of clicking rate and each Sample video, calculate the sample of users about each Sample video The clicking rate estimates the loss function of object module.Using stochastic gradient descent method minimize the sample of users about each The clicking rate of Sample video estimates the loss function of object module.Solve the sample of users about each Sample video The clicking rate estimates the gradient of the loss function of object module.Pass through point about each Sample video of the sample of users The rate of hitting estimates the gradient of the loss function of object module, from top to bottom successively updates the network ginseng that the clicking rate estimates object module Number.The gradient of the loss function of object module is estimated by clicking rate about each Sample video of the sample of users, With the corresponding network parameter of the sixth feature about each Sample video of the new sample of users.

L=-yⁱlogpⁱ-(1-yⁱ)log(1-pⁱ) (3)

Figure 10 is a kind of block diagram of device 1200 for executing video recommendation method shown according to an exemplary embodiment.Example Such as, interactive device 1200 can be mobile phone, computer, digital broadcasting terminal, messaging device, game console, put down Panel device, Medical Devices, body-building equipment, personal digital assistant etc..

Referring to Fig.1 0, device 1200 may include following one or more components: processing component 1202, memory 1204, Power supply module 1206, multimedia component 1208, audio component 1210, the interface 1212 of input/output (I/O), sensor module 1214 and communication component 1216.

The integrated operation of the usual control device 1200 of processing component 1202, such as with display, telephone call, data communication, Camera operation and record operate associated operation.Processing component 1202 may include one or more processors 1220 to execute Instruction, to perform all or part of the steps of the methods described above.In addition, processing component 1202 may include one or more moulds Block, convenient for the interaction between processing component 1202 and other assemblies.For example, processing component 1202 may include multi-media module, To facilitate the interaction between multimedia component 1208 and processing component 1202.

Memory 1204 is configured as storing various types of data to support the operation in equipment 1200.These data Example includes the instruction of any application or method for operating on device 1200, contact data, telephone book data, Message, picture, video etc..Memory 1204 can by any kind of volatibility or non-volatile memory device or they Combination is realized, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), it is erasable can Program read-only memory (EPROM), programmable read only memory (PROM), read-only memory (ROM), magnetic memory, flash memory Reservoir, disk or CD.

Power supply module 1206 provides electric power for the various assemblies of device 1200.Power supply module 1206 may include power management System, one or more power supplys and other with for device 1200 generate, manage, and distribute the associated component of electric power.

Multimedia component 1208 includes the screen of one output interface of offer between described device 1200 and user.? In some embodiments, screen may include liquid crystal display (LCD) and touch panel (TP).If screen includes touch panel, Screen may be implemented as touch screen, to receive input signal from the user.Touch panel includes that one or more touch passes Sensor is to sense the gesture on touch, slide, and touch panel.The touch sensor can not only sense touch or sliding is dynamic The boundary of work, but also detect duration and pressure associated with the touch or slide operation.In some embodiments, more Media component 1208 includes a front camera and/or rear camera.When equipment 1200 is in operation mode, as shot mould When formula or video mode, front camera and/or rear camera can receive external multi-medium data.Each preposition camera shooting Head and rear camera can be a fixed optical lens system or have focusing and optical zoom capabilities.

Audio component 1210 is configured as output and/or input audio signal.For example, audio component 1210 includes a wheat Gram wind (MIC), when device 1200 is in operation mode, when such as call mode, recording mode, and voice recognition mode, microphone quilt It is configured to receive external audio signal.The received audio signal can be further stored in memory 1204 or via communication Component 1216 is sent.In some embodiments, audio component 1210 further includes a loudspeaker, is used for output audio signal.

I/O interface 1212 provides interface, above-mentioned peripheral interface module between processing component 1202 and peripheral interface module It can be keyboard, click wheel, button etc..These buttons may include, but are not limited to: home button, volume button, start button and Locking press button.

Sensor module 1214 includes one or more sensors, and the state for providing various aspects for device 1200 is commented Estimate.For example, sensor module 1214 can detecte the state that opens/closes of equipment 1200, the relative positioning of component, such as institute The display and keypad that component is device 1200 are stated, sensor module 1214 can be with detection device 1200 or device 1,200 1 The position change of a component, the existence or non-existence that user contacts with device 1200,1200 orientation of device or acceleration/deceleration and dress Set 1200 temperature change.Sensor module 1214 may include proximity sensor, be configured in not any physics It is detected the presence of nearby objects when contact.Sensor module 1214 can also include optical sensor, as CMOS or ccd image are sensed Device, for being used in imaging applications.In some embodiments, which can also include acceleration sensing Device, gyro sensor, Magnetic Sensor, pressure sensor or temperature sensor.

Communication component 1216 is configured to facilitate the communication of wired or wireless way between device 1200 and other equipment.Dress The wireless network based on communication standard, such as WiFi can be accessed by setting 1200, carrier network (such as 2G, 3G, 4G or 5G) or they Combination.In one exemplary embodiment, communication component 1216 receives via broadcast channel and comes from external broadcasting management system Broadcast singal or broadcast related information.In one exemplary embodiment, the communication component 1216 further includes near-field communication (NFC) module, to promote short range communication.For example, radio frequency identification (RFID) technology, Infrared Data Association can be based in NFC module (IrDA) technology, ultra wide band (UWB) technology, bluetooth (BT) technology and other technologies are realized.

In the exemplary embodiment, device 1200 can be by one or more application specific integrated circuit (ASIC), number Signal processor (DSP), digital signal processing appts (DSPD), programmable logic device (PLD), field programmable gate array (FPGA), controller, microcontroller, microprocessor or other electronic components are realized, for executing the above method.

In the exemplary embodiment, a kind of non-transitorycomputer readable storage medium including instruction, example are additionally provided It such as include the memory 1204 of instruction, above-metioned instruction can be executed by the processor 1220 of device 1200 to complete the above method.Example Such as, the non-transitorycomputer readable storage medium can be ROM, random access memory (RAM), CD-ROM, tape, soft Disk and optical data storage devices etc..

Figure 11 is a kind of block diagram of device 1300 for executing video recommendation method shown according to an exemplary embodiment.Example Such as, device 1300 may be provided as a server.Referring to Fig.1 1, device 1300 includes processing component 1322, is further wrapped One or more processors, and the memory resource as representated by memory 1332 are included, it can be by processing component for storing The instruction of 1322 execution, such as application program.The application program stored in memory 1332 may include one or one with On each correspond to one group of instruction module.In addition, processing component 1322 is configured as executing instruction, to execute above-mentioned letter Cease list display method.

Device 1300 can also include that a power supply module 1326 be configured as the power management of executive device 1300, and one Wired or wireless network interface 1350 is configured as device 1300 being connected to network and input and output (I/O) interface 1358.Device 1300 can be operated based on the operating system for being stored in memory 1332, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or similar.

Those skilled in the art after considering the specification and implementing the invention disclosed here, will readily occur to its of the application Its embodiment.This application is intended to cover any variations, uses, or adaptations of the application, these modifications, purposes or Person's adaptive change follows the general principle of the application and including the undocumented common knowledge in the art of the application Or conventional techniques.The description and examples are only to be considered as illustrative, and the true scope and spirit of the application are by following Claim is pointed out.

It should be understood that the application is not limited to the precise structure that has been described above and shown in the drawings, and And various modifications and changes may be made without departing from the scope thereof.Scope of the present application is only limited by the accompanying claims.

Claims

1. a kind of video recommendation method characterized by comprising

Calculate separately the similarity between the attributive character of target user and the attributive character of multiple historical users, obtain with it is described Most similar first historical user of the attributive character of target user；

Based on first historical user to the clicking rate of the associated video, by multiple video recommendations in the associated video To the target user.

2. video recommendation method according to claim 1, which is characterized in that the attributive character of user includes: user ID feature, static nature and behavioral characteristics；

Preferably, the static nature of user includes at least one of following characteristics: age, gender, geographical location, IP The application list that location, mobile phone model, mobile phone are installed；

The behavioral characteristics of user include at least one of following characteristics: user clicks history feature, user thumbs up history Feature and user pay close attention to list characteristics；

Preferably, video have video features, the video features include at least one of following characteristics: video ID feature and Video author's ID feature.

3. video recommendation method according to claim 2, which is characterized in that the attribute for calculating separately target user is special Similarity between the attributive character of multiple historical users of seeking peace obtains and the attributive character of the target user most similar One historical user, comprising:

It calculates separately between the attributive character of the target user and the attributive character of the multiple historical user Distance；And

The distance is ranked up, is obtained and most similar first historical user of the attributive character of the target user.

4. video recommendation method according to claim 3, which is characterized in that it is described to be based on the clicking rate prediction model, First historical user is obtained to the clicking rate of associated video, comprising:

The video ID feature of the static nature and the associated video to first historical user carries out vector and asks With obtain fisrt feature；

The video author ID feature of the static nature and the associated video to first historical user carry out to Amount summation, obtains second feature；

By the fisrt feature and the second feature and the User ID feature and described of first historical user The behavioral characteristics of one historical user are as third feature；And

The third feature is inputted into the clicking rate prediction model, obtains first historical user to the associated video Clicking rate.

5. video recommendation method according to claim 4, which is characterized in that described to be based on first historical user to institute The clicking rate for stating associated video gives multiple video recommendations in the associated video to the target user, comprising:

According to first historical user to the clicking rate of the associated video, the clicking rate backward of the associated video is arranged Sequence；

It is sorted according to the clicking rate backward of the associated video, multiple videos of front in the associated video is pushed away It recommends to the target user.

6. video recommendation method according to claim 5, which is characterized in that it is described to be based on neural network algorithm, it establishes a little Hit rate prediction model, comprising:

Extract the attributive character of sample of users；

The video features of the attributive character and the Sample video based on the sample of users, it is pre- to the clicking rate Estimate before object module carries out to learning；And

The video features of the attributive character and the Sample video based on the sample of users, it is pre- to the clicking rate Estimate object module and carries out backward learning；

Preferably, described to be based on neural network algorithm, establish clicking rate prediction model, further includes:

The video ID feature of the static nature and the Sample video to the sample of users carries out vector summation, obtains To fourth feature；

The video author ID feature of the static nature and the Sample video to the sample of users carries out vector and asks With obtain fifth feature；And

The User ID feature and the history of the fourth feature and the fifth feature and the historical user are used The behavioral characteristics at family are as sixth feature；

Preferably, the video features of the attributive character based on the sample of users and the Sample video, it is right The clicking rate is estimated before object module carries out to learning, comprising:

Object module is estimated in the clicking rate from bottom to top successively to convert the sixth feature, obtains the clicking rate Estimate the top layer vector of object module；

The top layer vector is converted to the probability of clicking rate；

Preferably, the video features of the attributive character based on the sample of users and the Sample video, it is right The clicking rate estimates object module and carries out backward learning, comprising:

According to the video tab of the probability of the clicking rate and the Sample video, calculates the clicking rate and estimate object module Loss function；

With the newly corresponding network parameter of the sixth feature；

It is preferably, described to mark video tab for the Sample video, comprising:

If the sample of users clicks the Sample video of operation pages displaying, Sample video mark is positive Sample；

If the sample of users does not have the Sample video of clicking operation page presentation, the Sample video is labeled as Negative sample.

7. a kind of video recommendations device characterized by comprising

Nearest _neighbor retrieval unit, for calculating separately between the attributive character of target user and the attributive character of multiple historical users Similarity, obtain and most similar first historical user of the attributive character of the target user；

Clicking rate estimates unit, for being based on the clicking rate prediction model, obtains first historical user to associated video Clicking rate；And

Video recommendations unit, for the clicking rate based on first historical user to the associated video, by the related view Multiple video recommendations in frequency give the target user.

8. video recommendations device according to claim 7, which is characterized in that the attributive character of user includes: user ID feature, static nature and behavioral characteristics；

Preferably, video have video features, the video features include at least one of following characteristics: video ID feature and Video author's ID feature；

Preferably, the nearest _neighbor retrieval unit, comprising:

Fisrt feature extraction unit, for extracting the attributive character of the target user and the institute of the multiple historical user State attributive character；

Metrics calculation unit, for calculating separately the attributive character of the target user and the institute of the multiple historical user State the distance between attributive character；And

First sequencing unit obtains the attributive character most phase with the target user for being ranked up to the distance The first close historical user；

Preferably, the clicking rate estimates unit, comprising:

Second feature extraction unit, for extracting the attributive character of first historical user and the institute of the associated video State video features；

Fisrt feature integrated unit, for described in the static nature and the associated video to first historical user Video ID feature carries out vector summation, obtains fisrt feature；

Unit is estimated, for the third feature to be inputted the clicking rate prediction model, obtains first historical user couple The clicking rate of the associated video；

Preferably, the video recommendations unit, comprising:

Second sequencing unit, for the clicking rate according to first historical user to the associated video, to the related view The clicking rate backward of frequency sorts；

Recommendation unit will sequentially be leaned on for being sorted according to the clicking rate backward of the associated video in the associated video Preceding multiple video recommendations give the target user；

Preferably, the model foundation unit, comprising:

Forward direction unit, the video for the attributive character and the Sample video based on the sample of users are special Sign estimates before object module carries out to learning the clicking rate；And

Backward learning unit, the video for the attributive character and the Sample video based on the sample of users are special Sign estimates object module to the clicking rate and carries out backward learning；

Preferably, the model foundation unit, further includes:

Second feature integrated unit, the video for the static nature and the Sample video to the sample of users ID feature carries out vector summation, obtains fourth feature；

The top layer vector is converted to the probability of clicking rate；

With the newly corresponding network parameter of the sixth feature；

It is preferably, described to mark video tab for the Sample video, comprising:

9. a kind of video recommendations device characterized by comprising

Processor；

Memory for storage processor executable instruction；

Wherein, the processor is configured to executing video recommendation method described in 1 to 6 any one of the claims.

10. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has computer to refer to It enables, the computer instruction is performed realization such as video recommendation method as claimed in any one of claims 1 to 6.