Video live broadcast method and system

Abstract

The invention discloses a video live broadcast method and system, and relates to the technical field of video live broadcast, and the method comprises the following steps: S1, constructing an index value model; s2, dividing user values; s3, recommending live broadcast according to the viewing preference; s4, acquiring viewing features; and S5, generating a recommendation list. According to the method, users with different values are distinguished by constructing the index value model and dividing the user values, the loyalty of the users is further improved while high-quality live broadcast content is provided for the users, the accuracy of the recommendation result is improved, benefits are brought to a live broadcast platform while welfare is provided for the users with higher values, and the user experience is improved. Through preference estimation and preference prediction construction, viewing preferences distinguished by user values are obtained, preferences of different user types are mapped to corresponding virtual channels, and broadcast characteristics of each channel can be reflected based on a time period division mode, so that a channel scoring matrix is constructed, and the recommendation effect is further improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video live broadcast, in particular to a video live broadcast method and system. BACKGROUND

[0002] With the development of the live broadcast industry gradually stabilizing, major live broadcast platforms begin to focus on fine operation and strive to provide high-quality live broadcast content for users. In a Chinese patent with application number 202110591353.3, a live broadcast video recommendation method and device are disclosed, which includes: when it is detected that a live broadcast room is in a live broadcast state, obtaining video frames and audio of a live broadcast video of the live broadcast room; performing image recognition on the currently obtained video frames; and performing speech recognition on the currently obtained audio; based on the image recognition result and the speech recognition result, determining whether there is a target label in a candidate label set that matches the live broadcast room; if there is, adding the target label to the live broadcast room, and determining a user set corresponding to the target label; the user set contains users who have set the target label; determining whether each user in the user set meets the recommendation condition; if there is a user who meets the recommendation condition, recommending the live broadcast video of the live broadcast room to the user who meets the recommendation condition.

[0003] The above file can recommend live broadcast content of interest to the user, however, in addition to recommending content of interest, the live broadcast platform should also consider its own interests, improve user loyalty, and encourage user behavior participation. First, the live broadcast platform provides high-quality live broadcast content, and second, according to user preferences and user needs, it recommends suitable live broadcast content to the user, reduces the user's interest discovery cost, and also retains users who value the live broadcast platform. In addition, the recommendation of live broadcast video is not accurate enough, and it is also difficult to quickly and efficiently obtain the user's preferences. For the user viewing process, there is also a feature judgment that affects the result. If there is an error in feature acquisition, the utilization efficiency of the viewing data is reduced, and the quality effect of the live broadcast video recommendation is reduced. SUMMARY

[0004] The purpose of the present application is to provide a video live broadcast method and system to solve the problems raised in the background art.

[0005] To achieve the above purpose, the present application provides the following technical solution: a video live broadcast method, comprising the following steps:

[0006] S1, constructing an index value model: setting three indexes representing the time interval between the user's last live broadcast consumption and the statistical time, the number of user evaluations, and the user's live broadcast viewing time, calculating the values of the three indexes according to the user's behavior records to construct an index value model;

[0007] S2, user value grade division: obtaining the value score of the user according to the index value model, and then classifying the user according to different values to classify and recommend the user;

[0008] S3, recommending live broadcast according to viewing preference: obtaining the viewing preference based on the user value division through preference estimation and preference prediction, and traversing the live broadcast video sample data, obtaining the number of users who like the same type of live broadcast video by counting the number of users who like each live broadcast video, and then quickly recommending the live broadcast video interested by the user;

[0009] S4, obtaining viewing features; triggering live broadcast recommendation at the moment when the user selects the channel, which corresponds to the start time of the user's viewing behavior, positive feedback viewing refers to the user's formal viewing behavior of the live broadcast video, negative feedback viewing refers to the user's obtaining of the live broadcast video information from the short viewing process and the waiting for the advertisement, the process of negative feedback viewing means that the user does not like the current live broadcast video, and the viewing distance refers to the number of days from the end time of the last positive feedback viewing to the current time, so the start time, positive feedback viewing, negative feedback viewing and viewing distance of the user to each channel are taken as viewing features;

[0010] S5, generating a recommendation list: taking the viewing features affecting the viewing results as the input of the recurrent neural network, then taking the viewing results at the recommendation time as the output of the recurrent neural network, finally, adopting the Softmax function as the activation function for training, in the recommendation stage, taking the viewing features at the recommendation time as the input, and the output result is the vector representation of the viewing probability of each channel, and finally the output result is sorted from high to low according to the viewing probability to generate a recommendation list.

[0011] Preferably, the index value model construction process is as follows:

[0012] D i =min(d end -d r ) ij

[0013]

[0014] The index value model is represented as {D i ,F i ,T i |i∈1,2,···,n};

[0015] Wherein, d end and d r represent the statistical day and the last live broadcast consumption time respectively, (d end -d r ) ijrepresents the time interval of the user i's last consumption in the live broadcast j, min represents the minimum function, a id represents the evaluation data of the user i on the date d, t ij represents the time length of the user i watching the live broadcast j, j represents the video live broadcast number, N represents the total number of video live broadcasts, i represents the user watching the live broadcast, n represents the total number of users watching the live broadcast, D i , F i and T i represent three index values.

[0016] Preferably, the step S2 comprises the following steps:

[0017] S201, standardizing the indexes;

[0018]

[0019] In the formula, x d , x f , x t respectively represent the standardized values of the user's last live consumption time interval, the user's evaluation live video frequency and the user's live watching time length, and respectively represent the sample means of the three indexes, S d , S f and S t respectively represent the variances of the data samples of the three indexes, x d , x f and x t respectively represent three index value variables.

[0020] S202, four scales are set and correspond to value 1, value 2, value 3 and value 4 respectively, and the larger the value is, the more important the index is. More than 50 people in the live broadcast field are collected to select the scales of the three indexes, and the weights of the three indexes are determined according to the total proportion of the scales of each index.

[0021] Preferably, the step S2 further comprises the following steps:

[0022] S203, determining the user value score in combination with the weight and the standardized value;

[0023] W i = p d x id + p f x if + p t x it

[0024] wherein, xid , x if , and x it respectively represent the standardized values of the user's last evaluation time distance statistical time interval, the user's evaluation frequency, and the user's live viewing duration of the three indicators, p d , p f , and p t respectively represent the weights of the three indicators, W i represents the value score of user i;

[0025] S204, select m users from the user sample data set {x1, x2, ···, x n} as initial cluster centers, calculate the Euclidean distance of user x j from each cluster center, and mark the cluster to which user x j belongs with the nearest distance cluster center, and divide user x j into the corresponding cluster, where j ∈ [1, n], for the divided cluster, calculate the sample point mean of each cluster as a new cluster center, and repeat the clustering to divide the user value.

[0026] Preferably, the step S3 comprises the following steps:

[0027] S301, convert the viewing behavior of the user into a viewing score, and quantify the viewing behavior of the user each time into the interval of 0 to 1, and the viewing score of the user about the live video is represented as r ui ;

[0028] S302, take the start time of the viewing record of all users related to the channel C as input, divide one broadcast period of the channel C into k sub-periods, each sub-period corresponds to a virtual channel of the channel C, and the division of the time period can effectively reflect the broadcast characteristics of each channel;

[0029] S303, construct a channel score matrix E = (e uv ) m×n , where e

[0030] wherein, represents that the start time of the viewing behavior of the user u is in the live video set corresponding to the Sth sub-period of the channel C, F(r ui , h i ) represents the score of the user u about the live video i after h i hours of decay, r ui represents the score of the user u about the live video i, and m × n represents the matrix order;

[0031] S304, convert the user's preference for the live video into a user's score for the k virtual channels, thereby performing preference estimation.

[0032] Preferably, the step S3 further comprises the following steps:

[0033] S305, fill in the empty items in the channel score matrix E using the collaborative filtering technique, thereby obtaining a preference prediction matrix E', at the recommendation time t, any candidate live video of the user is mapped into a virtual channel, and corresponds to a preference score in the preference prediction matrix E', the preference score is used as the score of the candidate live video;

[0034] S306, by calculating the similarity between the candidate live video and the historical live video in the same virtual channel, and then combining the user's score for the historical live video, the user's score for the candidate live video is predicted, thereby performing preference prediction;

[0035] S307, the video similarity is used to predict the user's score for the live video, and finally the user's viewing preference for the live video i at time t is obtained.

[0036] Preferably, the step S5 comprises the following steps:

[0037] S501, set α as the viewing time length ratio threshold, take 0.1, w represents the viewing time length, the unit is second, T l represents the available viewing time length, that is, the time interval from the beginning of the user watching to the end of the playing, which satisfies And the viewing behavior of w < 300 seconds is regarded as negative feedback viewing, and the remaining viewing behavior is regarded as positive feedback viewing;

[0038] S502, given the viewing time period set {d1, d2, ···, d k}, wherein d i represents a sub-period, k represents the total number of sub-periods, i ∈ [1, k], and the vector PF of the positive feedback viewing record of the user u starting at time t is represented as:

[0039] PF = {pf1, pf2, ···, pf k}

[0040] Wherein,

[0041] S503, given the channel set {m1, m2, ···, m n}, wherein m i represents a channel, n represents the total number of channels, i ∈ [1, n], and the live video set I u involved in the negative feedback viewing behavior of the user u between time t-1 and time t is obtained, and the live video set It Then, the vector NW representing the negative feedback behavior of user u before time t is:

[0042] NW = {nw1,nw2,...,nw} n}

[0043] in,

[0044] S504, Suppose user u is in time period d i The set of start times T of all positive feedback viewership records for the starting channel. c User u during time period d i About channel m i The number of days from the end time t' of the last positive feedback viewing record to the current time t is dw. r , where t, t' ∈ d i The vector DW representing the viewing distance of user u across channels at time t is:

[0045] DW = {dw1, dw2, ..., dw} p}

[0046] in, p represents the total number of channels, and max represents the maximum value function.

[0047] A video live streaming system includes the following units:

[0048] The value model construction unit constructs an indicator value model by setting three indicators, which respectively represent the time interval between the user's most recent live broadcast consumption and the statistical time, the number of user reviews, and the duration of the user's live broadcast viewing.

[0049] The rating division unit obtains the user's value score based on the indicator value model, and then divides the user according to different values ​​to perform user classification and recommendation.

[0050] The preference recommendation unit obtains viewing preferences based on user value segmentation through preference estimation and preference prediction, and obtains the number of users who like the same type of live video by counting the number of users who like each live video, thereby quickly recommending live videos that users are interested in.

[0051] The feature acquisition unit uses the start time of the selected channel that triggers the live broadcast recommendation, positive feedback viewership, negative feedback viewership, and the user's viewing distance to each channel as viewership features.

[0052] The recommendation generation unit takes a viewing feature affecting a viewing result as an input of a recurrent neural network, and then takes the viewing result at a recommended time as an output of the recurrent neural network.

[0053] Compared with the prior art, the present application has the following advantages:

[0054] 1、The present application divides users with different values by constructing an index value model and dividing user value, provides high-quality live content for users, and further improves user loyalty. Since high-value users are beneficial to the revenue of live video, locking high-value users can improve the popularity of live video and encourage user behavior participation. Through standardized processing, data can be unified, and the weight is determined by collecting opinions of people in the live field, which is more in line with the actual value recognition of live video. Different value users are subdivided by clustering method to avoid data sparseness and improve the accuracy of recommendation results. The recommended results meet the value preferences of users, and provide benefits for users with higher value while also bringing benefits to the live platform;

[0055] 2、The present application obtains the viewing preference of the user value division by constructing the preference estimation and preference prediction, converts the user's preference for live video into the user's score for k virtual channels, thereby performing preference estimation, maps the preferences of different user types to the corresponding virtual channels, and based on the division method of the time period, the broadcast characteristics of each channel can be reflected, thereby constructing a channel score matrix, efficiently obtaining the value of live video in the user's mind, and further improving the recommendation effect;

[0056] 3、The present application obtains the viewing features of the start time of the selected channel triggering live recommendation, positive feedback viewing, negative feedback viewing, and user viewing distance for each channel, regards the live video recommendation problem as a classification problem, and inputs into a recurrent neural network to output the recommendation result, thereby improving the utilization efficiency of viewing data and reducing the possibility of obtaining incorrect data. BRIEF DESCRIPTION OF DRAWINGS

[0057] Fig. 1 The overall method flowchart provided by the embodiment of the present application is shown in the figure;

[0058] Fig. 2 The overall system structure diagram provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0059] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0060] Please refer to Figs. 1-2 The present application provides a technical solution: a video live streaming method, comprising the following steps:

[0061] S1, constructing an index value model: setting three indexes, respectively representing the time interval between the user's last live streaming consumption and the statistical time, the number of user evaluations and the duration of the user's live streaming watching, calculating the values of the three indexes according to the user's behavior records to construct an index value model;

[0062] S2, user value division: obtaining the value score of the user according to the index value model, and then dividing the user according to different values to classify and recommend the user;

[0063] S3, recommending live streaming according to viewing preference: obtaining the viewing preference based on the user value division through preference estimation and preference prediction, and traversing the live streaming video sample data, obtaining the number of users who like the same type of live streaming video by counting the number of users who like each live streaming video, and thus quickly recommending the live streaming video of interest to the user;

[0064] S4, obtaining viewing features; triggering live streaming recommendation at the moment when the user selects a channel, which corresponds to the start time of the user's viewing behavior, positive feedback viewing refers to the user's formal live streaming video watching behavior, negative feedback viewing refers to the user's obtaining of live streaming video information from the short viewing process and waiting for the advertisement, the process of negative feedback viewing means that the user does not like the current live streaming video, and viewing distance refers to the number of days from the end time of the last positive feedback viewing to the current time, therefore, the start time, positive feedback viewing, negative feedback viewing and the viewing distance of the user to each channel are taken as viewing features;

[0065] S5, generating a recommendation list: taking the viewing features affecting the viewing results as the input of the recurrent neural network, then taking the viewing results at the recommendation moment as the output of the recurrent neural network, finally, adopting the Softmax function as the activation function for training, in the recommendation stage, taking the viewing features at the recommendation moment as the input, and the output result is the vector representation of the viewing probability of each channel, finally, the output result is sorted from high to low according to the viewing probability to produce a recommendation list.

[0066] The recurrent neural network comprises an input layer, a hidden layer and an output layer, the input layer is a full connection layer for vector representation of input features, the hidden layer adopts a suitable neural network to complete network fitting, and the output layer adopts a full connection and uses a Softmax function as an activation function.

[0067] The viewing probability formula is specifically:

[0068]

[0069] P(w t |u,E) represents the content w t watched by the user u at the t time under the viewing environment E vi is a probability of a live video p, e vp represents a score corresponding to the live video i, e i represents a score corresponding to the live video p, and R represents all candidate live videos.

[0070] The index value model construction process formula is as follows:

[0071] D end = min(d r -d ij

[0072]

[0073] The index value model is represented as {D i , F i , T i |i∈1,2,···,n}.

[0074] wherein d end and d r represent a statistical day and a live consumption time of the last time, (d end -d r ) ij represents a time interval of the user i in the live j for the last time consumption from the statistical day, min represents a minimum value function, a id represents evaluation data of the user i on the date d, t ij represents a duration of the user i watching the live j, j represents a video live number, N represents a total number of video lives, i represents a user watching the live, n represents a total number of users watching the live, D i , F i and T i represent three index values.

[0075] The step S2 comprises the following steps:

[0076] S201, standardizing the index;

[0077]

[0078] wherein x d , x f , and x t respectively represent the standardized value of the user's last live consumption time distance from the statistical time interval, the standardized value of the user's evaluation of the live video times, and the standardized value of the three indicators of the user's live viewing duration, and respectively represent the sample mean of the three indicators, S d , S f , and S t respectively represent the variance of the data sample of the three indicators, x d , x f , and x t respectively represent the three indicators of the numerical variable;

[0079] S202, four scales are set and correspond to the numerical value 1, the numerical value 2, the numerical value 3, and the numerical value 4 respectively, and the larger the numerical value, the more important the indicator is. More than 50 people in the live field are collected for the scale selection of the three indicators, and the weight of each indicator is determined according to the total proportion of the scale.

[0080] The step S2 further comprises the following steps:

[0081] S203, the weight and the standardized value are combined to determine the user value score;

[0082] W i = p d x id + p f x if + p t x it

[0083] wherein x id , x if , and x it respectively represent the standardized value of the user's last live consumption time distance from the statistical time interval, the standardized value of the user's evaluation of the live video times, and the standardized value of the three indicators of the user's live viewing duration, p d , p f , and p t respectively represent the weight of the three indicators, W i represents the value score of the user i;

[0084] S204, select m users from the user sample data set {x1, x2, ···, x n} as the initial clustering center, and calculate the user x jEuclidean distance to each cluster center, and the user x is marked with the nearest distance cluster center j The user x is divided into the corresponding cluster, where j∈[1,n], and the sample point mean of each cluster is calculated as a new cluster center, and the user value is divided; j

[0085] Step S3 includes the following steps:

[0086] S301, the viewing behavior of the user is converted into a viewing score, and the viewing behavior of the user each time is quantified to the interval of 0 to 1, and the viewing score of the user about the live video is represented as r ui ;

[0087] S302, the start time of the viewing record of all users related to the channel C is taken as input, and one broadcast period of the channel C is divided into k sub-periods, each of which corresponds to a virtual channel of the channel C;

[0088] S303, a channel score matrix E=(e uv ) m×n is constructed, where the score of the user u to the channel v is represented as e

[0089] Wherein, represents that the start time of the viewing behavior of the user u is in the live video set corresponding to the S-th sub-period of the channel C, F(r ui ,h i ) represents the score of the user u about the live video i after the score is decayed by h i hours, r ui represents the score of the user u about the live video i, and m×n represents the matrix order;

[0090] S304, the preference of the user to the live video is converted into the score of the user to the k virtual channels, so as to perform preference estimation;

[0091] Step S3 further includes the following steps:

[0092] S305, since it cannot be guaranteed that all users have an evaluation to the virtual channel, there is a null value phenomenon in the channel score matrix E, so the collaborative filtering technology is used to fill the null item in the channel score matrix E, so as to obtain a preference prediction matrix E', at the recommendation t time, any candidate live video of the user is mapped into a virtual channel, and corresponds to a preference score in the preference prediction matrix E', the preference score is taken as the score of the candidate live video;

[0093] ​S306, preference prediction is performed by calculating the similarity of the candidate live video and the historical live video in the same virtual channel, and then combining the score of the historical live video by the user to predict the score of the candidate live video by the user;

[0094] S307, the score of the live video by the user is predicted by using the video similarity, and finally the viewing preference of the user for the live video i at t is obtained;

[0095]

[0096] Wherein, s ui represents the score of the live video by the user, i and j represent the live video and the historical live video in the virtual channel v respectively, sim represents similarity calculation, F(r uj , h j ) represents the score of the historical live video j by the user u after h j hours of decay, represents the set of historical live videos j viewed by the user u from the virtual channel v, S uR represents the score of all candidate live videos j respectively, min represents the minimum value function;

[0097] Specifically, the video similarity is calculated by using cosine similarity;

[0098] Normalization processing is performed, and the value range of the normalized score of the live video by the user is adjusted to the closed interval of 1 to 2;

[0099]

[0100] Wherein, s ui ' represents the normalized score of the live video by the user, max represents the maximum value function, and min represents the minimum value function;

[0101] Step S5 includes the following steps:

[0102] S501, set α as the viewing time length ratio threshold, take 0.1, w represents the viewing time length, the unit is second, T l represents the viewable time length, that is, the time interval from the start of watching to the end of playing, which satisfies And the viewing behavior of w<300 seconds is regarded as negative feedback viewing, and the remaining viewing behavior is regarded as positive feedback viewing;

[0103] S502, given the viewing time period set {d1, d2, ···, d k}, wherein, d idenotes a sub-period, k denotes the total number of divided sub-periods, i∈[1, k], and a vector PF representing the positive feedback viewing records of the user u starting at the t moment is denoted as:

[0104] PF={pf1,pf2,···,pf k}

[0105] wherein,

[0106] S503, a channel set {m1, m2, ···, m n} is known, wherein m i denotes a channel, n denotes the total number of channels, i∈[1, n], a live video set I u involved in the negative feedback viewing behavior of the user u between the t-1 moment and the t moment is denoted as I t , and a live video set I n playing at the t moment is denoted as I

[0107] NW={nw1,nw2,···,nw i}

[0108] wherein,

[0109] S504, assuming that the user u starts to view the channel at the d c set T i of the user u in the time period d i , the ending time t' of the last positive feedback viewing record of the channel m r about the channel m i , and a vector DW representing the channel viewing distance of the user u at the t moment is denoted as:

[0110] DW={dw1,dw2,···,dw p}

[0111] wherein, p denotes the total number of channels, and max denotes a maximum value function;

[0112] A video live system comprises the following units:

[0113] A value model construction unit, which constructs an index value model by setting three indexes, the three indexes respectively representing the interval of the time distance between the latest live consumption of the user and the statistical time, the number of user evaluation times, and the length of time of the user watching live;

[0114] The grade dividing unit obtains the value score of the user according to the index value model, and then divides the user according to different values, so as to classify and recommend the user;

[0115] The preference recommendation unit obtains the viewing preference based on the user value division through preference estimation and preference prediction, and obtains the number of users who like the same type of live video by counting the number of live videos liked by users, so as to quickly recommend the live video interested by the user;

[0116] The feature acquisition unit takes the start time of the selected channel triggering live recommendation, positive feedback viewing, negative feedback viewing and user viewing distance of each channel as viewing features;

[0117] The recommendation generation unit takes the viewing features affecting the viewing result as the input of the recurrent neural network, and then takes the viewing result at the recommendation moment as the output of the recurrent neural network.

[0118] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0119] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method of live video streaming, the method comprising: The method comprises the following steps: S1, constructing an index value model: setting three indexes, respectively representing the time interval between the user's last live consumption and the statistical time, the number of user evaluations, and the length of time the user watches live, calculating the values of the three indexes according to the user's behavior records to construct an index value model; S2, user value division: obtaining the user's value score according to the index value model, and then dividing the users according to different values to classify and recommend the users; S3, recommending live according to viewing preference: obtaining the viewing preference based on user value division through preference estimation and preference prediction, and traversing the live video sample data, obtaining the number of users who like the same type of live video by counting the number of users who like each live video, and thus quickly recommending the live video of interest to the user; S4, obtaining viewing features: triggering live recommendation at the moment when the user selects a channel, which corresponds to the start time of the user's viewing behavior, positive feedback viewing refers to the user's formal viewing behavior of live video, negative feedback viewing refers to the user's obtaining of live video information from a short viewing process and waiting for advertisements, the process of negative feedback viewing means that the user does not like the current live video, and viewing distance refers to the number of days from the end time of the last positive feedback viewing to the current time, therefore, the start time, positive feedback viewing, negative feedback viewing, and viewing distance of the user for each channel are taken as viewing features; S5, generating a recommendation list: taking the viewing features that affect the viewing results as the input of the recurrent neural network, then taking the viewing results at the recommendation time as the output of the recurrent neural network, finally, using the Softmax function as the activation function for training, in the recommendation stage, taking the viewing features at the recommendation time as the input, and the output result is a vector representation of the viewing probability of each channel, finally, the output result is sorted from high to low according to the viewing probability to produce a recommendation list.

2. The method of claim 1, wherein: In the step S1, the index value model construction process is as follows: D i = min(d end -d r ) ij The index value model is represented as {D i ,F i ,T i |i∈1,2,···,n}. where d end and d r represent the statistical day and the last live consumption time, respectively, (d end -d r ) ij represents the time interval between the last consumption of user i in live j and the statistical day, min represents the minimum function, a id represents the evaluation data of user i on date d, t ij represents the duration of user i watching live j, j represents the video live number, N represents the total number of video live, i represents the user watching live, n represents the total number of users watching live, D i , F i , and T i represent three index values.

3. The method of claim 1, wherein: The step S2 comprises the following steps: S201, standardizing the indexes; wherein x d , x f , and x t respectively represent the standardized values of the user's last live consumption time distance from the statistical time interval, the user's evaluation of the live video number of times, and the user's live viewing time, respectively, and respectively represent the sample mean of the three indicators, S d , S f , and S t respectively represent the variance of the data sample of the three indicators, x d , x f , and x t respectively represent the three indicator numerical variables. S202, setting four scales corresponding to the values 1, 2, 3, and 4 respectively, and the larger the value, the higher the importance of the index, collecting the scale selection of not less than 50 people on the three indexes from the live field, and determining the weight of each of the three indexes according to the total proportion of the scale of each index.

4. The method of claim 3, wherein: The step S2 further comprises the following steps: S203, determining the user value score in combination with the weight and the standardized value; W i = p d x id + p f x if + p t x it wherein x id , x if and x it respectively represent the standardized values of the user's last evaluation time distance statistical time interval, the user's evaluation frequency and the user's live viewing duration of the three indicators, p d , p f and p t respectively represent the weights of the three indicators, and W i represents the value score of the user i. S204. From the user sample dataset {x1,x2,...,x...} n Select m users as initial cluster centers and calculate user x. j User x is labeled with the Euclidean distance to each cluster center, using the cluster center with the closest distance. j The cluster to which user x belongs j The data is assigned to the corresponding clusters, where j∈[1,n]. For each cluster, the mean of the sample points is calculated as the new cluster center. This process is repeated to divide the user value.

5. The method of claim 1, wherein: The step S3 comprises the following steps: S301, convert the viewing behavior of the user into a viewing score, and quantify the viewing behavior of the user each time to the interval of 0 to 1, the viewing score of the user about the live video is expressed as r ui ; S302, taking the start time of the viewing record of all users related to the channel C as the input, dividing one broadcast period of the channel C into k sub-periods, and each sub-period corresponds to a virtual channel of the channel C; S303, constructing a channel score matrix E=(e uv ) m×n wherein the score of the user u to the channel v wherein, denotes the live video set corresponding to the Sth sub-period of channel C where the start time of the viewing behavior of user u is located, F(r ui ,h i ) denotes the score of user u on live video i after the score is decayed by h i hours, r ui denotes the score of user u on live video i, and m x n denotes the matrix order; S304, converting the user's preference for live video into the user's score for the k virtual channels, thereby performing preference estimation.

6. The method of claim 5, wherein: The step S3 further comprises the following steps: S305, filling the empty items in the channel score matrix E by using the collaborative filtering technology to obtain the preference prediction matrix E', at the recommendation time t, any candidate live video of the user is mapped into a virtual channel, and corresponds to a preference score in the preference prediction matrix E', the preference score is used as the score of the candidate live video; S306, by calculating the similarity of the candidate live video and the historical live video in the same virtual channel, and then combining the score of the historical live video by the user, the score of the candidate live video by the user is predicted, so as to carry out the preference prediction; S307, the score of the live video by the user is predicted by using the video similarity, and finally the viewing preference of the user at t time for the live video i is obtained.

7. The method of claim 1, wherein: The step S5 comprises the following steps: S501, set alpha as the viewing time length ratio threshold, take value 0.1, w represents the viewing time length, unit is second, T l The viewing time length represents the time interval from the start of viewing to the end of playing, which satisfies And the viewing behavior of w<300 seconds is regarded as negative feedback viewing, and the remaining viewing behavior is regarded as positive feedback viewing; S502. Given the set of viewing time periods {d1, d2, ..., d...} k }, where d i Let k represent the total number of sub-time periods, i∈[1,k]. The vector PF for positive feedback viewing records of user u starting at time t is represented as: PF = {pf1, pf2, ···, pf k} wherein S503, Given the channel set {m1, m2, ..., m...} n }, where m i Let I represent the number of channels, n represent the total number of channels, and i ∈ [1, n]. The set of live videos involved in user u's negative feedback viewing behavior between time t-1 and time t is I. u The set of live videos playing at time t is I. t Then, the vector NW representing the negative feedback behavior of user u before time t is: NW = { nw1, nw2, ···, nw n} wherein, S504, set user u in time period d i Start time set T of all positive feedback viewing records of the starting of the video channel c , user u in time period d i The last positive feedback viewing record of channel m i The number of days from the current time t to the end time t' of the last positive feedback viewing record of channel m r , wherein t, t'∈d i The vector DW of the channel viewing distance of user u at time t is represented as: DW = {dw1, dw2, ···, dw p} wherein p denotes the total number of channels, max denotes the maximum function.

8. A video live streaming system, characterized by, The video live system is suitable for the video live method of claims 1-7, comprising the following units: The value model construction unit constructs the index value model by setting three indexes, the three indexes respectively represent the interval of the time distance statistics time of the user's last live consumption, the user's evaluation times and the user's live watching time; The grade division unit obtains the value score of the user according to the index value model, and then divides the user according to different values, so as to carry out user classification recommendation; The preference recommendation unit obtains the viewing preference based on the value division of the user by preference estimation and preference prediction, and obtains the number of users who like the same type of live video by counting the number of live videos liked by the user, so as to quickly recommend the live video interested by the user; The feature acquisition unit takes the start time of the selected channel triggering live recommendation, positive feedback viewing, negative feedback viewing and user viewing distance of each channel as the viewing features; The recommendation generation unit takes the viewing features affecting the viewing results as the input of the recurrent neural network, and then takes the viewing results at the recommendation time as the output of the recurrent neural network.

Images