A video big data recommendation training method, system, device and storage medium

Abstract

The present invention discloses a video big data recommendation training method, system, device and storage medium, and relates to the technical field of video big data recommendation. The method obtains an interactive data set between a video and a user, and the interactive data set includes: a comment data set, a viewing data set, a like data set, a collection data set and a forwarding data set; real heat training is performed on the comment data set, the viewing data set, the like data set, the collection data set and the forwarding data set respectively to obtain corresponding real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate; the method is used to solve the problem in the prior art that whether a video is recommended is directly judged by obtaining comment, viewing set, like, collection and forwarding data, which may cause malicious data brushing, resulting in insufficient validity of analyzed data and inaccurate video recommendation.

Description

Technical Field

[0001] The present invention relates to the technical field of video big data recommendation, and specifically to a video big data recommendation training method, system, device and storage medium. Background Art

[0002] Video recommendation uses artificial intelligence, big data, and machine learning technologies to analyze and understand massive video content to achieve personalized recommendations. It has become particularly important in the context of the popularization of smart devices and video as a mainstream communication method. It pushes precise content through algorithms based on content, collaborative filtering, and deep learning. Overall, the video recommendation system is becoming an important bridge connecting users and content, constantly promoting the innovation and development of information consumption models.

[0003] Video big data recommendation is to classify videos, label them with different tags, and recommend them to users who like these tags. The quality of video content under the same tag is different, which leads to different user preferences. Therefore, it is necessary to analyze the video itself and the user interaction data to judge the quality of the video. In the existing technology, whether a video is recommended is directly judged by obtaining comments, viewing sets, likes, favorites and forwarding data. There will be malicious data brushing, resulting in inaccurate recommendations based on comments, viewing sets, likes, favorites and forwarding data. For example, in a patent application with application publication number CN116992159A, a content recommendation method and system are disclosed. In the process of collecting and calculating the interaction information between users and videos, this method only directly calculates the number of likes and comments of users, lacks the effectiveness screening of the interaction information between users and videos, and This recommendation method is only used to recommend information to the analyzed users, and cannot analyze the popularity of the video itself. In the process of evaluating and recommending the popularity of videos or news information, the existing analysis method only directly superimposes and summarizes the number of interactions when using interactive information. For example, in a patent application with application publication number CN115964574A, a method for evaluating the popularity of public opinion on smart traffic safety based on data mining is disclosed. This method sets weights for four indicators, namely, the number of single-article forwarding, the number of single-article comments, the number of single-article likes, and the sentiment score, according to different user levels. The validity of the interactive data is not screened, and the weight setting process is relatively subjective. The weight setting is not based on the analysis of the browsing data of the network information itself, and there is a lack of evaluation of the validity of the interactive information, resulting in inaccurate heat analysis of the video itself, which ultimately affects the quality of the recommended video. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems in the prior art to a certain extent, by obtaining a data set of interactions between videos and users, performing real heat training on each data in the interaction data, obtaining corresponding real rates and summing them up to obtain the overall real heat; and recommending videos from large to small according to the values ​​of the overall real heat, so as to solve the problem in the prior art that whether a video should be recommended is directly judged by obtaining comments, viewing episodes, likes, favorites and forwarding data, which may lead to malicious data brushing, resulting in insufficient validity of the analyzed data and inaccurate video recommendations.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a video big data recommendation training method, comprising the following steps:

[0006] Obtain the interactive dataset between videos and users. The interactive dataset includes: comment dataset, viewing dataset, like dataset, favorite dataset, and forwarding dataset.

[0007] Perform real heat training on the comment dataset, viewing dataset, like dataset, collection dataset, and forwarding dataset respectively to obtain the corresponding real comment rate, real viewing rate, real like rate, real collection rate, and real forwarding rate;

[0008] The real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate are summed up to obtain the overall real heat; the videos are recommended from large to small according to the value of the overall real heat.

[0009] Furthermore, the process of performing real heat training on the comment dataset and obtaining the real comment rate includes the following sub-steps:

[0010] Get all reviews from the review dataset;

[0011] Compare the comment to the ASCII code table to obtain the pure symbol comment characters, and then obtain the pure symbol comment character length. If the comment character length is less than or equal to two characters, delete the comment;

[0012] The comments that directly comment on the video are set as the first-level comments, the comments that directly comment on the first-level comments are set as the second-level comments, and so on. The comments that directly comment on the i-1th level comments are set as the i-th level comments. The total number of comments from the first level to the i-th level is counted respectively, and they are represented by N1, N2, ..., N i express;

[0013] The comments are weighted and summed. The specific formula is:

[0014] M=N1+a2*N2+a3*N3+...+a i *N i ;

[0015] Where M is the total number of real hot comments, a2 to a i They are the second level weight to the i-th level weight respectively;

[0016] Get the total number of users Y who watched the video;

[0017] The calculated true review rate is: M / Y.

[0018] Furthermore, the process of performing real heat training on the comment dataset to obtain the real comment rate also includes the following sub-steps:

[0019] Get the video label, select the first number of historical videos with the same label for weighted simulation training, and the weighted simulation training is used to obtain the value from a2 to a through simulation. i , weight simulation training includes:

[0020] Count the total number of comments K of the first historical number of videos with the same tag, and the total number of commenting users L;

[0021] Get the average number of comments from commenting users: K / L;

[0022] Count the total number of second-level comments Q2 for videos of the same tag of the first historical number, and count the number of users who make second-level comments that are different from the users who make first-level comments, and set it as the second number of newly added users W2;

[0023] Calculate the second-level user comment ratio W2 / Q2;

[0024] The second-level weight is calculated by the formula a2 = (K / L)*(W2 / Q2);

[0025] Similarly, count the total number of level i comments Q of videos with the same label in the first historical number i , count the number of users who make comments at level i that are different from those who make comments at level i to level i-1, and set it as the number of new users W at level i i ;

[0026] Calculate the proportion of comments by level i users W i / Q i ;

[0027] The weight of the i-th level is calculated by the formula: a i =(K / L)*(W i / Q i );

[0028] Then the weight data of each item in M ​​is: a2 = (K / L) * (W2 / Q2), a3 = (K / L) * (W3 / Q3), ..., a i =(K / L)*(W i / Q i );

[0029] Every time the first historical number of videos with the same label is added, the first historical number of videos with the same label is re-trained through weight simulation to update a2 to a i .

[0030] Furthermore, the process of performing real heat training on the comment dataset and obtaining the real comment rate includes the following sub-steps:

[0031] Obtain the total duration Tz of the video, the total viewing duration Tgz of each user, the user's fast-forward speed Vb, and the viewing duration Tb at the corresponding speed from the viewing data set;

[0032] The actual viewing time of each user is calculated as: (Tgz-Tb)+Vb*Tb;

[0033] The average of the actual viewing time of all users is calculated and marked as the average user viewing time Tpg;

[0034] The actual viewing rate is calculated as: Tpg / Tz.

[0035] Furthermore, the process of performing real heat training on the likes dataset to obtain the real like rate includes the following sub-steps:

[0036] Get the number of likes Dz for the video from the likes dataset; get the actual viewing time of the user who liked the video, and set it as the actual viewing time Tdz of the user who liked the video;

[0037] Multiply Tz by the first ratio to get the effective viewing time;

[0038] When Tdz is less than the effective viewing time, delete the number of likes of the user who liked the post and get the filtered number of likes Dsz;

[0039] The actual like rate is calculated as: Dsz / Y.

[0040] Furthermore, the process of performing real heat training on the collection data set and obtaining the real collection rate includes the following sub-steps:

[0041] The number of video collections Sz is obtained from the collection data set; the actual viewing time of the user who collected the video is obtained and set as the actual viewing time Tsc of the user who collected the video;

[0042] When Tsc is less than the effective viewing time, delete the collection number of the collection user and obtain the filtered collection number Ssz;

[0043] Calculate the actual collection rate: Ssz / Y. Further,

[0044] Furthermore, the process of performing real heat training on the forwarding data set to obtain the real forwarding rate includes the following sub-steps:

[0045] Obtain the number of video forwardings Zz from the forwarding data set; obtain the actual viewing time of the forwarding user, and set it as the actual viewing time Tzf of the forwarding user;

[0046] When Tzf is less than the effective viewing time, delete all forwarding numbers of the forwarding user to obtain the filtered forwarding number Zsz;

[0047] Calculate the actual forwarding rate: Zsz / Y;

[0048] In a second aspect, the present invention provides a video big data recommendation training system, including: a data acquisition module, a real heat training module and a video recommendation module;

[0049] The data acquisition module is used to acquire the interactive data set between the video and the user, and the interactive data set includes: comment data set, viewing data set, like data set, favorite data set and forwarding data set;

[0050] The real heat training module is used to perform real heat training on the comment data set, the viewing data set, the like data set, the collection data set and the forwarding data set respectively, to obtain the corresponding real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate;

[0051] The video recommendation module sums up the real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate to obtain the overall real popularity; and recommends videos from large to small according to the value of the overall real popularity.

[0052] In a third aspect, the present invention provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps in the above method are performed.

[0053] In a fourth aspect, the present invention provides a storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps in the above method are performed.

[0054] The beneficial effects of the present invention are as follows: the present invention obtains an interaction data set between a video and a user, and performs real heat training on the comment data set, viewing data set, like data set, favorite data set and forwarding data set in the interaction data, obtains the corresponding real rates and sums them up to obtain the overall real heat; the video is recommended from large to small according to the value of the overall real heat. A single high real heat does not represent a good video. The design can train the interaction data separately and then combine the corresponding real rates to make an overall heat judgment for the video heat recommendation. Invalid data and data with a high heat value are eliminated during training, and video recommendations are made based on valid data.

[0055] The present invention obtains the labels of the videos and selects the first historical number of videos with the same labels for weight simulation training. The weight simulation training is used to obtain the weights from a2 to a through simulation. i ,This design can perform weight simulation training through historical data to make data weights more accurate and universal.

[0056] The present invention obtains the total duration Tz of the video, the total viewing duration Tgz of each user, the user's fast-forward speed Vb and the viewing duration Tb at the corresponding speed from the viewing data set, and calculates the actual viewing duration of each user as: (Tgz-Tb)+Vb*Tb. This design can prevent the user from using a single user viewing duration to judge whether the video reaches the effective viewing duration when watching at a certain speed, resulting in the problem of insufficient statistical validity of the actual viewing duration. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A flowchart of the steps of the present invention;

[0058] Figure 2 is a functional block diagram of the system of the present invention;

[0059] Figure 3 A schematic diagram of grading reviews according to the present invention;

[0060] Figure 4 It is a connection block diagram of the electronic device of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] Example 1, please refer to Figure 2As shown, the present invention provides a video big data recommendation training system, including: a data acquisition module, a real heat training module and a video recommendation module;

[0063] The data acquisition module is used to obtain the interactive data set between videos and users. The interactive data set includes: comment data set, viewing data set, like data set, collection data set and forwarding data set; the interactive data between videos and users obtained are data that have been reviewed and labeled accordingly, and the tags are recommended to users with corresponding preferences; if you want to judge the quality of a video by collecting comments, views, likes, collections and forwarding, comments, views, likes, collections and forwarding can represent the key data for the quality of video recommendations. For example, short video platforms will record users' behaviors such as playing, staying, following, commenting, liking and forwarding of a video in real time. Among them, there is data that has no obvious relationship with the popularity of the video, such as following, and the attention data and the popularity of the video itself are not analyzed;

[0064] The real heat training module is used to perform real heat training on the comment data set, viewing data set, like data set, collection data set and forwarding data set respectively to obtain the corresponding real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate; the real heat training module includes a real comment heat training unit, a real viewing heat training unit, a real collection heat training unit, a real like heat training unit and a real forwarding heat training unit.

[0065] The real comment heat training unit is configured with a real comment heat training strategy, and the real comment heat training strategy includes: obtaining all comments from the comment data set;

[0066] The comments are compared with the ASCII encoding table to obtain the pure symbol comment characters, and then the pure symbol comment character length is obtained. If the comment character length is less than or equal to two characters, the comment will be deleted. When the user clicks on the comment area, the user may accidentally send a shorter comment, and the pure symbol data has no correlation with the popularity of the video itself. By deleting these invalid comment data, such as "," and ".", the data can be made more accurate.

[0067] The comments that directly comment on the video are set as the first-level comments, the comments that directly comment on the first-level comments are set as the second-level comments, and so on. The comments that directly comment on the i-1th level comments are set as the i-th level comments. The total number of comments from the first level to the i-th level is counted respectively, and they are represented by N1, N2, ..., N iSome users usually have the habit of watching video comments, and the quality of comments will also affect the popularity of the video. For example, if a user understands the connotation of a video through comments and likes the video more, the popularity of the video brought by the comments is not the same, and the popularity brought by further comments on the comments is also different. Therefore, the popularity brought by the comments will be graded, that is, the popularity will be graded and analyzed according to different levels;

[0068] The comments are weighted and summed. The specific formula is: M = N1 + a2*N2 + a3*N3 + ... + a i *N i ;

[0069] Where M is the total number of real hot comments, a2 to a i They are the second-level weight to the i-th-level weight respectively; by weighted summation, the real hot comments are more accurate, while the single statistics of the total number of comments has poor accuracy;

[0070] Get the total number of users Y who watched the video; calculate the true comment rate: M / Y.

[0071] Calculate the real comment rate, and unify the data types of the real like rate, real collection rate, and real forwarding rate to facilitate unified data processing;

[0072] In the specific implementation, a comment scenario is set. In the specific application, there are many levels of comments on the data. Here, for the convenience of display, there are three levels. The number of first-level comments N1=100, the number of second-level comments N2=60, and the number of third-level comments N3=20; because the comments are related to the video, the comments on the video are generally greater than the comments on the comments, and the comments on the comments decrease in turn. The total number of users who watch the video is Y=1000; the second-level weight a2=0.5, the third-level weight a3=0.3, the weight setting is obtained through weight model training, and the weight cannot be set alone. The above data can be used to obtain the real comment popularity of the data;

[0073] M=100+60*0.5+20*0.3=136; the real comment popularity is more accurate than the sum of the comments. The total number of users who watched the video, 1000, is obtained from the comment data set; the real comment rate is calculated to be: 0.136. Calculating the comment rate facilitates unified processing of data.

[0074] Get the video label, select the first number of historical videos with the same label for weighted simulation training, and the weighted simulation training is used to obtain the value from a2 to a through simulation. i , weight simulation training includes:

[0075] By selecting the first historical number of videos with similar tags for weight simulation training, the existing weight setting process is also relatively subjective. The weight setting is not based on the analysis of the browsing data of the network information itself, and lacks the evaluation of the effectiveness of the interactive information, resulting in inaccurate heat analysis of the video itself, which ultimately affects the quality of the recommended video;

[0076] Count the total number of comments K of videos with the same tag in the first historical number, and the total number of commenting users L; if the data obtained by using a single video is not adaptable to other videos, a large amount of data needs to be counted;

[0077] The average number of comments from commenting users is K / L. Each user may make different numbers of comments on different types of videos. The average value of a large amount of data is obtained to obtain the number of comments for each user.

[0078] Count the total number of first-level comments Q2 for videos of the same tag of the first historical number, and count the number of users who make second-level comments that are different from the users who make first-level comments, and set it as the second number of newly added users W2;

[0079] Calculate the second-level user comment ratio W2 / Q2; the obtained ratio can be used to roughly calculate the new users brought by the first-level comments if the second-level comments are known. The new users brought by the first comments can represent the new popularity brought to the video;

[0080] The second-level weight is calculated by the formula a2 = (K / L) * (W2 / Q2); the second-level comments bring some valid comments by multiplying the number of new users by the average number of comments by commenting users;

[0081] Similarly, count the total number of level i comments Q of videos with the same label in the first historical number i , count the number of users who make comments at level i that are different from those who make comments at level i to level i-1, and set it as the number of new users W at level i i ;

[0082] Calculate the proportion of comments by level i users W i / Q i ;

[0083] The weight of the i-th level is calculated by the formula: a i =(K / L)*(W i / Q i );

[0084] Then the weight data of each item in M ​​is: a2 = (K / L) * (W2 / Q2), a3 = (K / L) * (W3 / Q3), ..., a i =(K / L)*(W i / Q i );

[0085] Every time the first historical number of videos with the same label is added, the first historical number of videos with the same label is re-trained through weight simulation to update a2 to a i By increasing the number of weight simulation training updates, the corresponding weight data is more adaptable to more videos, and the weights of video comments with different tags will also change over time and be more adaptable.

[0086] In the specific implementation, taking the weight of the second-level comments as an example, the total number of comments on videos with the same label of 100 in the first history is counted, which is 12,000, and the total number of commenting users is 10,000; the more data here is generally, the more universal the following calculation is, and the average number of comments by commenting users is obtained as: 1.2; the total number of second-level comments on videos with the same label of the first history is counted as 1,000, and the number of users who comment on the second level is different from the users who comment on the first level, which is set as the second number of new users 250; the proportion of second-level user comments is calculated as 0.25; the second-level weight is calculated by the formula a2=1.2*0.25=0.3, and only the second-level weight is calculated here, and all level weights must be calculated in practice;

[0087] The real viewing heat training unit is configured with a real viewing heat training strategy, which includes:

[0088] The total duration Tz of the video, the total viewing duration Tgz of each user, the user's fast-forward speed Vb, and the viewing duration Tb at the corresponding speed are obtained from the viewing data set; different users have different viewing speed habits, and also have different viewing speed habits for different types of videos. It is inaccurate to judge whether the video has reached the effective viewing time simply by the user's viewing time, so it is necessary to combine the speed data to make a judgment.

[0089] The actual viewing time of each user is calculated as: (Tgz-Tb)+Vb*Tb; since the normal speed viewing time is longer, the viewing time benchmark is set to the viewing time at normal speed, and the viewing time at speed is adjusted to the time required for normal video viewing;

[0090] The actual viewing time of all users is averaged and marked as the average user viewing time Tpg; the effective viewing time of a user is not universal, so the average viewing degree is calculated to make the data universal and accurate;

[0091] The actual viewing rate is calculated as: Tpg / Tz. Calculating the actual viewing rate facilitates unified data processing;

[0092] In the specific implementation, taking two users as an example, the total video duration of 30 seconds, the total viewing duration of one user of 12 seconds, the user's fast forward speed of 2x and the viewing time at the corresponding speed of 6 seconds are obtained from the viewing data set; the real viewing time of each user is calculated as: [(12-6)+2*6]=18 seconds; when the actual user watches the video, there may be a video that is watched at different speeds at different video nodes, which should be taken into account during the calculation, such as the total viewing time of another user of 6 seconds, the user's fast forward speed of 2x and the viewing time at the corresponding speed of 1 second, 4x and the viewing time at the corresponding speed of 1 second; the real viewing time of each user is calculated as: (6-1-1)+1*2+1*4=10 seconds; the average of the real viewing time of all users is taken, marked as the average user viewing time of 14 seconds; the real viewing rate is calculated as: 14 / 30=0.4.

[0093] The real likes heat training unit is configured with a real likes heat training strategy, which includes: obtaining the number of likes Dz for the video from the likes data set; obtaining the real viewing time of the user who likes the video, and setting it as the real viewing time Tdz of the user who likes the video;

[0094] Multiply Tz by the first ratio to obtain the effective viewing time; the first ratio is set to 0.15-0.3, and the setting of the first ratio is determined by referring to the existing setting standards for effective viewing time in the field of short videos, preferably 0.2. For example, a 15-second video is effectively viewed if at least 3 seconds are watched, and a 120-second video is effectively viewed if at least 24 seconds are watched. If the video is shorter, the first ratio can be appropriately increased. For example, for a 9-second video, the first ratio is 0.3, and at least 2.7 seconds are watched for effective viewing. If the video is longer, the first ratio can be appropriately reduced. For example, for a 1000-second video, the first ratio is 0.16, and at least 160 seconds are watched for effective viewing. Different adjustments are made according to different video lengths or types;

[0095] When Tdz is less than the effective viewing time, delete the number of likes of the user who liked the video and get the filtered number of likes Dsz. The likes of users who liked the video without watching the video content are poorly correlated with the popularity of the video content itself. This type of data is deleted to increase the authenticity of the data;

[0096] The actual like rate is calculated as: Dsz / Y. Calculating the actual like rate facilitates unified data processing;

[0097] In the specific implementation, 300 likes for the video are obtained from the likes data set; the actual viewing time of a user who likes the video is 2 seconds; the video length of 30 seconds is moderate, and the first ratio can be preferably set to 0.2, and 30 is multiplied by 0.2 to obtain an effective viewing time of 6 seconds; when 2 seconds is less than the effective viewing time of 6 seconds, the likes of the user who likes the video are deleted; if a total of 20 likes are deleted from the video, the actual like rate is calculated to be: 280 / 1000=0.28.

[0098] The real collection heat training unit is configured with a real collection heat training strategy, which includes:

[0099] The number of video collections Sz is obtained from the collection data set; the actual viewing time of the user who collected the video is obtained and set as the actual viewing time Tsc of the user who collected the video; obtaining the actual viewing time to compare with the effective time can prevent the user from using the time after watching at a double speed to make the comparison more accurate;

[0100] When Tsc is less than the effective viewing time, delete the collection number of the collection user and get the filtered collection number Ssz; directly collect the video without watching the video content, and the collection likes of such users are poorly correlated with the content popularity of the video itself, so delete such data to increase the authenticity of the data;

[0101] Calculate the real collection rate: Ssz / Y. Calculating the real collection rate facilitates unified processing; in specific implementation, obtain 200 video collections from the collection data set; obtain the real viewing time of a collection user, set to the real viewing time of the collection user as 3 seconds; 3 seconds is less than the effective viewing time of 6 seconds, delete the collection number of the collection user; if the real viewing time of the collection user is 7 seconds, and 7 seconds is greater than the effective viewing time of 6 seconds, do not delete, compare the real viewing time of all collection users, if the video deletes 20 collections in total, calculate the real collection rate: 180 / 1000=0.18.

[0102] The real forwarding heat training unit is configured with a real forwarding heat training strategy, which includes: obtaining the video forwarding number Zz from the forwarding data set; obtaining the real viewing time of the forwarding user, and setting it as the real viewing time Tzf of the forwarding user;

[0103] When Tzf is less than the effective viewing time, delete all forwarding numbers of the forwarding user and get the filtered forwarding number Zsz; prevent the situation of forwarding without watching the video content, and the forwarding of such users is poorly correlated with the popularity of the video content itself, so delete such data to increase the authenticity of the data; one user may have multiple forwardings, in which case all forwarding numbers should be calculated;

[0104] Calculate the real forwarding rate: Zsz / Y. Calculating the real forwarding rate facilitates unified processing;

[0105] In the specific implementation, the number of video forwardings, 230, is obtained from the forwarding data set; the actual viewing time of a forwarding user is obtained and set to the actual viewing time of the forwarding user, 4 seconds; when 4 seconds is less than the effective viewing time of 6 seconds, all forwardings of the forwarding user are deleted; if the actual viewing time of the forwarding user is 8 seconds, and 8 seconds is greater than the effective viewing time of 6 seconds, the video is not deleted, and the actual viewing time of all forwarding users is compared with , If the total number of forwardings deleted for the video is 30, it should be made clear here that it is 30 forwardings, not thirty forwarding users, and the actual forwarding rate is calculated: 200 / 1000=0.2.

[0106] The video recommendation module is used to sum the real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate to obtain the overall real heat; and recommend videos from large to small according to the overall real heat value. The higher the overall real heat value, the better the interaction data between the video and the user, which proves that the video is more popular and can be recommended to more users. However, a single data cannot prove that a video is popular. For example, if the real comment rate is good but other data is not good, the comments may all be saying that the video is bad.

[0107] In the specific implementation, an example is as follows: sum up the real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate to obtain the overall real heat, then the overall real heat is: 0.3+0.4+0.28+0.18+0.2=1.36; another overall real heat sum data of the same label is 1.23, 1.36 is greater than 1.23, so the video with an overall real heat of 1.36 is recommended first. Similarly, if there is more data, the videos are recommended from large to small according to the value of the overall real heat.

[0108] Example 2, please refer to Figure 1 As shown, the present invention provides a video big data recommendation training method, including: step S1, the acquired video and user interaction data is data that has passed the review and is marked with corresponding labels, and the labels are recommended to corresponding favorite users.

[0109] Step S2, respectively performing real heat training on the comment data set, viewing data set, like data set, collection data set and forwarding data set to obtain the corresponding real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate; Step S2 includes the following sub-steps:

[0110] Step S201, obtaining all comments from the comment data set;

[0111] Step S202, compare the comment with the ASCII code table to obtain the pure symbol comment characters, and then obtain the pure symbol comment character length. If the comment character length is less than or equal to two characters, delete the comment;

[0112] Step S203: set the comments that directly comment on the video as first-level comments, set the comments that directly comment on the first-level comments as second-level comments, and so on, set the comments that directly comment on the i-1th level comments as i-th level comments, and count the total number of comments from the first level to the i-th level, respectively, using N1, N2, ..., N i express;

[0113] Step S204: perform weighted calculation and summation on the comments. The specific formula is M=N1+a2*N2+a3*N3+...+a i *N i ; Where M is the total number of real hot comments, a2 to a i They are the second-level weight to the i-th-level weight respectively. By weighted summation, the real hot comments are more accurate, while the accuracy of a single statistical total number of comments is poor;

[0114] Step S205, obtaining the total number of users Y who watch the video, and calculating the true comment rate as: M / Y.

[0115] Step S204 includes the following sub-steps: Step S2041, obtaining the label of the video, selecting the first historical number of videos with the same label for weight simulation training, and the weight simulation training is used to obtain the weights from a2 to a through simulation. i , weight simulation training includes:

[0116] Step S2042, counting the total number of comments K of videos with the same tag of the first historical number, and the total number of commenting users L;

[0117] Step S2043, obtaining the average number of comments by commenting users: K / L; each user may make different numbers of comments on different types of videos, and the average value of a large amount of data is obtained to obtain the number of comments for each user.

[0118] Step S2044, counting the total number of first-level comments Q2 of the first historical number of videos with the same tag, and counting the number of users who made second-level comments that are different from the users who made first-level comments, and setting it as the second number of newly added users W2;

[0119] Step S2045, calculating the second-level user comment ratio W2 / Q2;

[0120] Step S2046, the second level weight is calculated by the formula a2 = (K / L) * (W2 / Q2). Similarly, the total number of level i comments Q of the videos with the same label of the first historical number is counted.i , count the number of users who make comments at level i that are different from those who make comments at level i to level i-1, and set it as the number of new users W at level i i ;

[0121] Step S2047, calculate the proportion of comments by the i-th level user W i / Q i , the weight of the i-th level is calculated as a by the formula i =(K / L)*(W i / Q i ), then the weight data of each item in M ​​is: a2 = (K / L) * (W2 / Q2), a3 = (K / L) * (W3 / Q3), ..., a i =(K / L)*(W i / Q i );

[0122] Step S2048: when the first historical number of videos with the same label are added, the newly added first historical number of videos with the same label are re-trained through weight simulation training to update a2 to a i .

[0123] Step S2 also includes the following sub-steps: Step S206, obtaining the total duration Tz of the video, the total viewing duration Tgz of each user, the user's fast-forward speed Vb, and the viewing duration Tb at the corresponding speed from the viewing data set; calculating the actual viewing duration of each user as: (Tgz-Tb)+Vb*Tb; taking the average of the actual viewing durations of all users, marked as the average user viewing duration Tpg; calculating the actual viewing rate as: Tpg / Tz.

[0124] Step S207, obtaining the number of likes Dz for the video from the likes data set; obtaining the actual viewing time of the user who likes the video, and setting it as the actual viewing time Tdz of the user who likes the video; multiplying Tz by the first ratio to obtain the effective viewing time; when Tdz is less than the effective viewing time, deleting the number of likes of the user who likes the video, and obtaining the filtered number of likes Dsz; calculating the actual like rate as: Dsz / Y.

[0125] Step S208, obtain the number of video collections Sz from the collection data set; set the actual viewing time of the user who collected the video as the actual viewing time Tsc of the user who collected the video; when Tsc is less than the effective viewing time, delete the number of collections of the user who collected the video, and obtain the filtered number of collections Ssz; calculate the actual collection rate: Ssz / Y.

[0126] Step S209, obtain the number of video forwardings Zz from the forwarding data set; set the actual viewing time of the forwarding user as the actual viewing time Tzf of the forwarding user; when Tzf is less than the effective viewing time, delete all forwarding numbers of the forwarding user to obtain the filtered forwarding number Zsz; calculate the actual forwarding rate: Zsz / Y.

[0127] Step S3, summing up the real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate to obtain the overall real popularity; and recommending the videos from large to small according to the value of the overall real popularity.

[0128] Example 3, please refer to Figure 4 As shown, Figure 4 The schematic diagram of the structure of an electronic device is illustrated, and the electronic device may include: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus. The memory stores computer-readable instructions, and the processor can call the instructions in the memory. When the computer-readable instructions are executed by the processor, the steps in a video big data recommendation training method are executed to achieve the following functions: by obtaining the interactive data set between the video and the user, each data in the interactive data is trained for real heat respectively, and the corresponding real rate is obtained and summed to obtain the overall real heat; according to the value of the overall real heat, the video is recommended from large to small.

[0129] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0130] Embodiment 4, the present application also provides a computer program product, the computer program product includes a computer program stored on a computer-readable storage medium, the computer program includes program instructions, when the program instructions are executed by the computer, the computer can execute a video big data recommendation training method provided by the above methods, the method includes: by obtaining the interaction data set between the video and the user, performing real heat training on each data in the interaction data respectively, obtaining the corresponding real rate and summing them up, to obtain the overall real heat; recommending the video from large to small according to the value of the overall real heat.

[0131] Embodiment 5, the present application also provides a computer-readable storage medium, and the present application provides a storage medium on which a computer program is stored. When the computer program is executed by the processor, the steps in the above video big data recommendation training method are executed to achieve the following functions: by obtaining the interaction data set between the video and the user, real heat training is performed on each data in the interaction data respectively, and the corresponding real rates are obtained and summed to obtain the overall real heat; the videos are recommended from large to small according to the value of the overall real heat.

[0132] Through the description of the above implementation methods, the embodiments of the present invention can be provided as methods, systems or computer program products. Based on such an understanding, the above technical solutions can be essentially or partly contributed to the prior art in the form of software products, which can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and include several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0133] In the embodiments provided in the present application, it should be understood that the disclosed system or method can be implemented in other ways. The embodiments described above are merely schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or units can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of systems, modules and units can be electrical, mechanical or other forms.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A video big data recommendation training method, characterized in that: include: Obtain the interactive dataset between videos and users. The interactive dataset includes: comment dataset, viewing dataset, like dataset, favorite dataset, and forwarding dataset. Perform real heat training on the comment dataset, viewing dataset, like dataset, collection dataset, and forwarding dataset respectively to obtain the corresponding real comment rate, real viewing rate, real like rate, real collection rate, and real forwarding rate; The real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate are summed up to obtain the overall real popularity; the videos are recommended from large to small according to the value of the overall real popularity; The process of training the comment dataset with real heat and obtaining the real comment rate includes the following sub-steps: Get all reviews from the review dataset; Compare the comment to the ASCII code table to obtain the pure symbol comment characters, and then obtain the pure symbol comment character length. If the comment character length is less than or equal to two characters, delete the comment; The comments that directly comment on the video are set as the first-level comments, the comments that directly comment on the first-level comments are set as the second-level comments, and so on. The comments that directly comment on the i-1th level comments are set as the i-th level comments. The total number of comments from the first level to the i-th level is counted respectively, and they are represented by N1, N2, ..., N i express; The comments are weighted and summed. The specific formula is: <h2 style=";text-align:left;direction:ltr">M = N1 + a2 * N2 + a3 * N3 + ... + a<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> *N<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> ; Where M is the total number of real hot comments, a2 to a i They are the second level weight to the i-th level weight respectively; Get the total number of users Y who watched the video; The calculated true review rate is: M / Y; The process of training the comment dataset for real heat and obtaining the real comment rate also includes the following sub-steps: Get the video label, select the first number of historical videos with the same label for weighted simulation training, and the weighted simulation training is used to obtain the value from a2 to a through simulation. i , weight simulation training includes: Count the total number of comments K of the first historical number of videos with the same tag, and the total number of commenting users L; Get the average number of comments from commenting users: K / L; Count the total number of second-level comments Q2 for videos of the same tag of the first historical number, and count the number of users who make second-level comments that are different from the users who make first-level comments, and set it as the second number of newly added users W2; Calculate the second-level user comment ratio W2 / Q2; The second-level weight is calculated by the formula a2 = (K / L)*(W2 / Q2); Similarly, count the total number of level i comments Q of videos with the same label in the first historical number i , count the number of users who make comments at level i that are different from those who make comments at level i to level i-1, and set it as the number of new users W at level i i ; Calculate the proportion of comments by level i users W i / Q i ; The weight of the i-th level is calculated by the formula: a i =(K / L)*(W i / Q i ); Then the weight data of each item in M ​​is: a2 = (K / L) * (W2 / Q2), a3 = (K / L) * (W3 / Q3), ..., a i =(K / L)*(W i / Q i ); Every time the first historical number of videos with the same label is added, the first historical number of videos with the same label is re-trained through weight simulation to update a2 to a i .

2. The video big data recommendation training method according to claim 1, characterized in that: The process of training the comment dataset with real heat and obtaining the real comment rate includes the following sub-steps: Obtain the total duration Tz of the video, the total viewing duration Tgz of each user, the user's fast-forward speed Vb, and the viewing duration Tb at the corresponding speed from the viewing data set; The actual viewing time of each user is calculated as: (Tgz-Tb)+Vb*Tb; The average of the actual viewing time of all users is calculated and marked as the average user viewing time Tpg; The actual viewing rate is calculated as: Tpg / Tz.

3. The video big data recommendation training method according to claim 2, characterized in that: The process of training the likes dataset with real heat and obtaining the real like rate includes the following sub-steps: Get the number of likes Dz for the video from the likes dataset; get the actual viewing time of the user who liked the video, and set it as the actual viewing time Tdz of the user who liked the video; Multiply Tz by the first ratio to get the effective viewing time; When Tdz is less than the effective viewing time, delete the number of likes of the user who gave the like, and get the filtered number of likes Dsz; The actual like rate is calculated as: Dsz / Y.

4. The video big data recommendation training method according to claim 3, characterized in that: The process of training the collection dataset for real popularity and obtaining the real collection rate includes the following sub-steps: The number of video collections Sz is obtained from the collection data set; the actual viewing time of the user who collected the video is obtained and set as the actual viewing time Tsc of the user who collected the video; When Tsc is less than the effective viewing time, delete the collection number of the collection user and obtain the filtered collection number Ssz; Calculate the actual collection rate: Ssz / Y.

5. The video big data recommendation training method according to claim 4, characterized in that: The process of performing real heat training on the forwarding data set to obtain the real forwarding rate includes the following sub-steps: Obtain the number of video forwardings Zz from the forwarding data set; obtain the actual viewing time of the forwarding user, and set it as the actual viewing time Tzf of the forwarding user; When Tzf is less than the effective viewing time, delete all forwarding numbers of the forwarding user to obtain the filtered forwarding number Zsz; Calculate the actual forwarding rate: Zsz / Y.

6. A video big data recommendation training system, used to implement a video big data recommendation training method according to any one of claims 1 to 5, characterized in that: Includes data acquisition module, real popularity training module and video recommendation module; The data acquisition module is used to acquire the interactive data set between the video and the user, and the interactive data set includes: comment data set, viewing data set, like data set, favorite data set and forwarding data set; The real heat training module is used to perform real heat training on the comment data set, the viewing data set, the like data set, the collection data set and the forwarding data set respectively, to obtain the corresponding real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate; The video recommendation module is used to sum the real comment rate, real viewing rate, real like rate, real collection rate and real forwarding rate to obtain the overall real popularity; and recommend videos from large to small according to the value of the overall real popularity.

7. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the steps in the method according to any one of claims 1 to 5 are executed.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are executed.

Images