An artificial intelligence-based cross-platform video promotion system and method

Abstract

The application relates to the field of video promotion, in particular to a cross-platform video promotion system and method based on artificial intelligence, which comprises a platform data module, a video association module, a cross-platform transmission module, a video recommendation module and an advertisement insertion module; the platform data module is used for recording and storing user information; the video association module is used for generating a joint influence function; the cross-platform transmission module is used for classifying user groups and performing cross-platform classification of classified data through a cloud; the video recommendation module is used for judging the group type of cross-platform users; and the advertisement insertion module is used for inserting advertisements in videos and generating video streams; the application can protect the personal privacy of a single user, improve the intelligence and security of an algorithm, intelligently adjust content strategies and video pushing arrangements, realize higher platform popularity, effectively improve the exposure rate and conversion rate of video advertisements, save promotion costs and increase benefits.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of video promotion, in particular to a cross-platform video promotion system and method based on artificial intelligence. BACKGROUND

[0002] Video promotion is a promotion method of promoting products, services or brands to target audiences through online video media. It realizes the goal of commodity promotion through inserting advertisements in video content or playing advertisements in the platform. Since video promotion can combine various data analysis tools for effect tracking and evaluation, it can achieve precise user group conversion.

[0003] There are many platforms that can play videos. If data intercommunication between platforms is not performed, a lot of algorithm resources will be wasted, so cross-platform promotion algorithms have emerged. Cross-platform promotion needs to involve multiple platforms and data sources, but there are certain difficulties in data synchronization and integration between different platforms, making it difficult to accurately determine user preferences on different platforms. In addition, user preferences involve personal privacy, and cross-platform transmission of private information also has the risk of data leakage.

[0004] In addition, advertisements will affect the video viewing experience of users and need to be evenly inserted into the viewing process. However, there are differences in the viewing time of users on different videos, making it difficult to accurately determine the advertisement placement window. The current cross-platform video promotion algorithm may have the problem of excessive advertisement placement, thereby affecting the effective reach and conversion rate of advertisements and leading to poor advertisement effect. SUMMARY

[0005] The purpose of the present application is to provide a cross-platform video promotion system and method based on artificial intelligence to solve the problems raised in the background art.

[0006] In order to solve the above technical problems, the present application provides the following technical scheme: a cross-platform video promotion system based on artificial intelligence, comprising: a platform data module, a video association module, a cross-platform transmission module, a video recommendation module and an advertisement insertion module;

[0007] The platform data module is composed of a database of a video platform, platform software and a local server, and is used for recording and storing user information, video element information, interaction data and playback history. A neural network model is set in the local server, and the interaction information of a single user is stored in the data pool of the server as a training library of the neural network model;

[0008] The video association module is configured to extract historical playback data of a user by taking the username of the user as a primary key, to store the historical playback data of the user into a training pool of a local server to train a neural network model for video playback behavior of a single user, to obtain a loss model, and to fit an association model between interactive data and each video element by taking video elements as input and taking the interactive data of the user as output, and to generate a joint influence function between video playback duration and each video element.

[0009] The cross-platform transmission module is configured to cut the joint influence function by taking a preset definition domain interval of each video element, to classify user groups according to the cutting result, and to upload the joint influence function of all users in a group to the cloud for each type of user group, and to classify and aggregate data packets from each platform by taking a group tag as a primary key, and to receive cross-platform data packet calls.

[0010] The video recommendation module is configured to parse the cloud data packet on a second platform, to fit the joint influence function in each group tag data for consistency, to generate a typical influence function corresponding to the group tag, and to input a video element into each typical function to output a prediction sequence of user interactive data by taking video playback duration as a constraint condition, and to accumulate a playback coefficient of an element corresponding to a group tag according to a distance between actual interactive data and each element in the sequence.

[0011] The advertisement insertion module is configured to extract a video from a video library under each video element classification according to the playback coefficient, to calculate a video arrangement mode with a maximum advertisement length of each video, and to generate a video stream according to the video arrangement mode to automatically recommend the video to a user.

[0012] Further, the platform data module comprises a database unit and an artificial intelligence unit.

[0013] The database unit is configured to store user information and video playback elements, wherein the video playback elements comprise a label, duration, creator, picture quality, and advertisement length of a video, and the interactive data comprises a click rate and video playback duration.

[0014] The artificial intelligence unit is arranged in a local database and is configured to analyze video playback data, and the neural network model adopted by the artificial intelligence unit comprises an R-CNN model, a DeepLab model, a U-Net model, and a Watershed model.

[0015] Further, the video association module comprises a video classification unit, a loss model unit, and a joint influence unit.

[0016] The video classification unit is configured to extract historical playback data of a user, and to classify the historical playback data according to video elements.

[0017] The loss model unit is configured to train a neural network model, and fit the influence relationship of each video element on the user interaction data.

[0018] The joint influence unit is configured to represent the fitting result in a function form, and generate a joint influence function with the interaction data as the dependent variable and each video element as the independent variable.

[0019] Further, the cross-platform transmission module includes a group preference unit and a cloud communication unit.

[0020] The group preference unit is configured to divide the user groups according to the video element range, and compress the joint influence function corresponding to the group into a data packet.

[0021] The cloud communication unit is configured to classify and aggregate the data packets of each platform according to the group label, and is responsible for the cloud uploading and downloading of the data packets.

[0022] Further, the video recommendation module includes a user identification unit, a group analysis unit, and a play accumulation unit.

[0023] The user identification unit is configured to fit the joint influence function in each data packet to generate a typical influence function, which represents the average preference level of the group label.

[0024] The group analysis unit is configured to obtain the operation behavior of the user in the second platform, and determine the group label of the user.

[0025] The play accumulation unit is configured to score the group label to which the user belongs according to the difference between the actual interaction data of the user and the predicted interaction data of each typical influence function.

[0026] Further, the advertisement insertion module includes a video extraction unit and a constraint conversion unit.

[0027] The video extraction unit is configured to store the video in each partition according to the interval of the video element, and extract the video from each partition according to the group label of the user.

[0028] The constraint conversion unit is configured to control the arrangement of the extracted video by taking the video play duration as a constraint condition, so that the sum of the lengths of the video advertisements is maximum.

[0029] A cross-platform video promotion method based on artificial intelligence, comprising the following steps:

[0030] Step S1. Collecting user interaction data and video play elements in a first platform, for each user ID in the platform, generating a data table in a local server with the interaction data as the table header and the video play elements as the table items.

[0031] Step S2. Generate a training library from the data table of all users in a preset period, train the neural network model with the data in the training library, obtain a loss model, and use the loss model to fit the influence of the video playback elements on the user interaction data to obtain a joint influence function, wherein the loss model takes the video playback elements as input and the user interaction data as output;

[0032] Step S3. According to the video classification according to the video playback elements, substitute the video playback elements in each classification into the joint influence function, and classify the joint influence function according to the video type where the joint influence function takes the maximum value;

[0033] Step S4. Package the joint influence function according to the classification result, upload the obtained data packet after marking the video type, and upload the obtained data packet to the cloud end, and the cloud end classifies and summarizes the data packets from each platform according to the video type as the keyword, and receives cross-platform data packet calls;

[0034] Step S5. The second platform parses the data packet, and performs consistency fitting on the joint influence function of each video type to obtain a typical function. Substitute the playback elements of the video watched by the user in the second platform into each typical function, and accumulate scores for each group according to the difference between the predicted results and the actual interaction data. According to the score proportion, extract the video to generate a recommended video stream.

[0035] Further, step S1 includes:

[0036] Step S11. Collect the interaction data and video playback elements of the user in the first platform, wherein the interaction data is the video click rate or the video playback time, and the video playback elements include the label, duration, creator, picture quality and advertisement length of the video;

[0037] Step S12. For each login behavior of each user ID, a data chain table is constructed, which has n rows and m columns, n is the total number of interaction data and video playback elements, and m is the total number of videos played by the user during the login. The interaction data is fixed as the table header of the data chain table, and the video playback elements are the table items of the data chain table. Store all generated data chain tables under the user ID in the local server.

[0038] Further, step S2 includes:

[0039] Step S21. The local server generates a new data space every statistical period as a storage location for the newly generated data chain table in the statistical period.

[0040] Step S22. Use a neural network model to learn a training library composed of all data chain tables under a single user ID;

[0041] Taking the video playing elements in the training library as input items and the user interaction data as output items, a function relationship F(x1, x2, …, x n-1 ) between the output variable and the input variable is fitted for each data chain table, wherein x n-1 represents the n-1th input variable, and the neural network model comprises an R-CNN model, a DeepLab model, a U-Net model and a Watershed model.

[0042] Step S23. According to the selected model category, an elastic network regression is used to fit the loss function:

[0043]

[0044] wherein LOSS is the loss function, SE is the root mean square error of the model, Y is the regularization strength of the model, k is the number of data chain tables in the training library, wi represents the preset weight of the ith data chain table in the training library, r is the proportion of the regularization term, F(x1, x2, …, x n-1 ) represents the function relationship between the output variable and the input variable in the ith data chain table, represents the linear regression function of the function F(x1, x2, …, x n-1 ).

[0045] Step S24. The size of the regularization strength Y and the proportion r of the regularization term when the loss function LOSS takes the minimum value is calculated, and a positive definite matrix P is generated according to the values of Y and r, so that G(x1, x2, …, x n-1 )=P·[x1,x2,…,x n-1 ] T , wherein T is a transpose symbol, and G(x1, x2, …, x n-1 ) is a joint influence function.

[0046] Further, step S3 comprises:

[0047] Step S31. The playing elements of the video are arranged and combined to obtain h video types, the playing elements corresponding to each video type are respectively substituted into the joint influence function, the function values obtained constitute a prediction set, and the joint influence function is classified into the video type corresponding to the maximum value in the prediction set.

[0048] Step S32. The joint influence function corresponding to each user ID is classified, and the user ID is hidden in the classification result, and only the function formula is kept.

[0049] Further, step S4 comprises:

[0050] Step S41. Pack all joint influence functions contained in each video type and upload to the cloud, and the cloud obtains the data packets sent by each platform, and classifies and summarizes the data packets according to the video type as the keyword;

[0051] Step S42. When the user logs in for the first time on the second platform, the second platform downloads all data packets from the cloud, and stores them in the local server in isolation according to the video type.

[0052] Further, step S5 includes:

[0053] Step S51. The second platform generates a typical function of each video type in the local server through a consistency equation:

[0054]

[0055] Wherein, B(x1,x2,…,x n-1 ) represents the typical function of the video type, v represents the number of joint influence functions in the video type, Gj(x1,x2,…,x n-1 ) represents the jth joint influence function in the video type.

[0056] Solving the consistency equation, the typical function corresponding to each video type is obtained;

[0057] Step S52. Obtain the video playback elements of the videos played by the user in the second platform, and substitute the video playback elements into the typical function of each video type, and the obtained value is denoted as a value set {A1,A2,…,At,…,Ah}, wherein At represents the function value of the typical function in the tth video type.

[0058] Step S53. Obtain the actual interaction data A0 of the user to the played video, calculate the difference AS between each element in the value set and the actual interaction data, the AS=|At-A0|, and accumulate the score UR of the corresponding video type, the UR=1 / AS.

[0059] Step S54. According to the cumulative score, extract videos from the video library corresponding to each video type, generate a video stream, and push the video stream to the user.

[0060] Compared with the prior art, the beneficial effects achieved by the present application are:

[0061] 1. According to the video playback history of the user on the platform, the present application obtains the type elements and interaction parameters of the video, and according to the video playback behavior of a single user, the present application uses a loss model to fit the correlation model of the interaction parameters and the video elements, obtains the joint influence function between the cumulative playback time and each factor, can understand the preferences and reactions of the user to the video content, and thus better grasps the audience demand, adjusts the content strategy and video pushing arrangement, and realizes higher platform popularity.

[0062] 2. The application can cut the joint influence function with each video element as the domain, thereby classifying users according to the factor influence dimension, sending the classification results to the cloud in a cross-platform communication protocol, and managing cross-platform big data with group preferences instead of individual preferences, protecting the personal privacy of individual users, and improving the intelligence and security of the algorithm.

[0063] 3. The application accumulates the play coefficients of each type according to the distance between the actual viewing time and the set element, extracts videos according to the play coefficients, and generates a video stream with the video arrangement mode of the maximum video advertisement length as the constraint condition, which can effectively improve the exposure rate and conversion rate of video advertisements, save promotion costs and increase benefits. BRIEF DESCRIPTION OF DRAWINGS

[0064] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0065] Figure 1 is a structural schematic diagram of a cross-platform video promotion system based on artificial intelligence according to the application;

[0066] Figure 2 is a step schematic diagram of a cross-platform video promotion method based on artificial intelligence according to the application. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0068] Please refer to Figure 1 , the application provides a technical solution: a cross-platform video promotion system based on artificial intelligence, comprising: a platform data module, a video association module, a cross-platform transmission module, a video recommendation module and an advertisement insertion module;

[0069] The platform data module is composed of a database of a video platform, platform software and a local server, and is used for recording and storing user information, video element information, interactive data and play history. A neural network model is set in the local server, and the interactive information of a single user is stored in the data pool of the server as a training library of the neural network model.

[0070] The platform data module comprises a database unit and an artificial intelligence unit.

[0071] The database unit is configured to store user information and video playing elements, wherein the video playing elements include tags, time lengths, creators, picture qualities and advertisement lengths of videos, and the interaction data includes click rates and video playing time lengths.

[0072] The artificial intelligence unit is arranged in the local database and is configured to analyze video playing data, and the neural network model adopted includes an R-CNN model, a DeepLab model, a U-Net model and a Watershed model.

[0073] The video correlation module is configured to extract historical playing data of a user by taking the username of the user as a primary key, to store the historical playing data of the user into a training pool of a local server for training a neural network model for a single user's video playing behavior, to obtain a loss model, to take video elements as inputs and user interaction data as outputs, to fit an association model of the interaction data and the video elements, and to generate a joint influence function between the video playing time length and the video elements.

[0074] The video correlation module includes a video classification unit, a loss model unit and a joint influence unit.

[0075] The video classification unit is configured to extract historical playing data of a user and classify the historical playing data according to video elements.

[0076] The loss model unit is configured to train a neural network model and fit an influence relationship of each video element on user interaction data.

[0077] The joint influence unit is configured to represent a fitting result in a function form, to generate a joint influence function with interaction data as a dependent variable and each video element as an independent variable.

[0078] The cross-platform transmission module is configured to cut the joint influence function according to a preset definition domain interval of each video element, to classify user groups according to a cutting result, and to pack and upload joint influence functions of all users in each type of user group to a cloud end, and the cloud end classifies and summarizes data packets from each platform according to a group tag as a primary key and receives cross-platform data packet calls.

[0079] The cross-platform transmission module includes a group preference unit and a cloud end communication unit.

[0080] The group preference unit is configured to divide user groups according to a video element range and to compress joint influence functions of the groups into data packets.

[0081] The cloud end communication unit is configured to classify and summarize data packets of each platform according to a group tag, and to be responsible for cloud end uploading and downloading of the data packets.

[0082] The video recommendation module is configured to parse the cloud data packet in the second platform, perform consistent fitting on the joint influence function in each group label data, generate a typical influence function corresponding to the group label, input the video elements into each typical function for each video playing of the user in the second platform, output a predicted sequence of the user interaction data, and accumulate a playing coefficient of the group label corresponding to each element according to the distance between the actual interaction data and the sequence elements.

[0083] The video recommendation module comprises a user identification unit, a group analysis unit and a playing accumulation unit.

[0084] The user identification unit is configured to perform fitting on the joint influence function in each data packet, and generate a typical influence function representing the average preference level of the group label.

[0085] The group analysis unit is configured to obtain the operation behavior of the user in the second platform, and determine the group label of the user.

[0086] The playing accumulation unit is configured to score the group label to which the user belongs according to the difference between the actual interaction data of the user and the predicted interaction data of each typical influence function.

[0087] The advertisement insertion module is configured to extract videos from the video library under each video element category according to the playing coefficient, calculate a video arrangement mode in which the sum of the advertisement lengths of the videos is maximum under the constraint of the video playing time length, generate a video stream according to the video arrangement mode, and automatically recommend the video stream to the user.

[0088] The advertisement insertion module comprises a video extraction unit and a constraint conversion unit.

[0089] The video extraction unit is configured to store the videos in each partition according to the interval of the video elements, and extract the videos from each partition according to the group label of the user.

[0090] The constraint conversion unit is configured to control the arrangement mode of the extracted videos under the constraint of the video playing time length, so that the sum of the advertisement lengths of the videos is maximum.

[0091] As shown in Figure 2 A cross-platform video promotion method based on artificial intelligence comprises the following steps:

[0092] Step S1. Collect user interaction data and video playing elements in a first platform, generate a data table in a local server for each user ID in the platform, with the interaction data as the table header and the video playing elements as the table items.

[0093] Step S1 comprises:

[0094] Step S11. Collect user interaction data and video playback elements within the first platform. The interaction data includes video click-through rate or video playback duration. The video playback elements include: video tags, duration, creator, picture quality, and advertisement length.

[0095] Step S12. For each login action of each user ID, construct a data linked list. The data linked list has n rows and m columns, where n is the total number of interactive data and video playback elements, and m is the total number of times the user plays videos on the platform during the current login. Fix the interactive data as the head of the data linked list, and the video playback elements as the table items of the data linked list. Store all the data linked lists generated under this user ID in the local server.

[0096] Step S2. Generate a training library from the data tables of all users within a preset period, train a neural network model with the data in the training library to obtain a loss model. The loss model takes video playback elements as input and user interaction data as output. The loss model is used to fit the influence of video playback elements on each user's interaction data to obtain a joint influence function.

[0097] Step S2 includes:

[0098] Step S21. Preset the platform statistical cycle. The local server generates a new data space every statistical cycle, which serves as the storage location for the newly generated data linked list within the statistical cycle.

[0099] Step S22. Use a neural network model to learn a training library consisting of all data linked lists under a single user ID;

[0100] Using video playback elements from the training library as input and user interaction data as output, for each data list, fit a functional relationship F(x1,x2,…,x) between the output variable and the input variable. n-1 ), where x n-1 Representing the (n-1)th input variable, the neural network models include: R-CNN model, DeepLab model, U-Net model, and Watershed model;

[0101] Step S23. Fit the loss function using elastic network regression according to the selected model category:

[0102]

[0103] Where LOSS is the loss function, SE is the root mean square error of the model, Y is the regularization strength of the model, k is the number of data lists in the training library, wi represents the preset weight of the i-th data list in the training library, r is the proportion of the regularization term, and Fi(x1,x2,…,x) n-1) represents a function relationship between the output variable and the input variable in the i th data chain table, representing the function Fi(x1, x2, …, x n-1 ) of linear regression function;

[0104] Step S24. Calculate the size of the regularization strength Y and the regularization term ratio r when the loss function LOSS takes the minimum value, generate a positive definite matrix P according to the value of Y and r, and make G(x1, x2, …, x n-1 ) = P·[x1, x2, …, x n-1 ] T , where T is the transpose symbol, G(x1, x2, …, x n-1 ) is the joint influence function.

[0105] Step S3. According to the classification of videos according to video playing elements, substitute the video playing elements in each classification into the joint influence function, and classify the joint influence function according to the video type where the joint influence function takes the maximum value;

[0106] Step S3 includes:

[0107] Step S31. Arrange and combine the playing elements of the video to obtain h video types, substitute the playing elements corresponding to each video type into the joint influence function respectively, and the function values obtained constitute a prediction set, and the joint influence function is classified into the video type corresponding to the maximum value in the prediction set;

[0108] Step S32. Classify the joint influence function corresponding to each user ID, and hide the user ID in the classification result, only keeping the function.

[0109] Step S4. Pack the joint influence function according to the classification result, upload the obtained data packet to the cloud after marking the video type, and the cloud classifies and summarizes the data packets from each platform according to the video type as the keyword, and receives cross-platform data packet calls;

[0110] Step S4 includes:

[0111] Step S41. Pack and upload all joint influence functions contained in each video type to the cloud, and the cloud obtains the data packets sent by each platform, and classifies and summarizes the data packets according to the video type as the keyword;

[0112] Step S42. When a user logs in to the second platform for the first time, the second platform downloads all data packets from the cloud, and stores them in the local server according to the isolation of video types.

[0113] Step S5. The second platform parses the data packet, performs consistency fitting on the joint influence function of each video type to obtain a typical function, substitutes the playback elements of the user watching the video in the second platform into each typical function, accumulates scores for each type of group label according to the difference between the prediction result and the actual interaction data, extracts videos according to the score proportion, and generates a recommended video stream.

[0114] Step S5 includes:

[0115] Step S51. The second platform generates a typical function of each video type in the local server through a consistency equation:

[0116]

[0117] wherein B(x1, x2, …, x n-1 ) represents the typical function of the video type, v represents the number of joint influence functions in the video type, Gj(x1, x2, …, x n-1 ) represents the jth joint influence function in the video type;

[0118] Solving the consistency equation obtains the typical function corresponding to each video type;

[0119] Step S52. The video playback elements of the video played by the user in the second platform are obtained, the video playback elements are substituted into the typical function of each video type, and the obtained value is recorded as a value set {A1, A2, …, At, …, Ah}, wherein At represents the function value of the typical function in the tth video type;

[0120] Step S53. The actual interaction data A0 of the user to the played video is obtained, the difference AS between each element in the value set and the actual interaction data is calculated, the AS = |At-A0|, the score UR of the video type corresponding to the element is accumulated, and the UR = 1 / AS;

[0121] Step S54. According to the accumulated score, the video is extracted from the video library corresponding to each video type to generate a video stream, and the video stream is pushed to the user.

[0122] Embodiment: there are 2 users and 4 video elements in the first platform, user 1 watches the video elements (1, 2, 1, 4) for 20 minutes in one login, user 2 watches the video elements (2, 1, 3, 3) for 50 minutes, and the rest of the videos are not watched, then the joint influence functions F1 and F2 are generated for user 1 and user 2 respectively, wherein F1 outputs 20 when the input is (1, 2, 1, 4), and outputs 0 for the rest, F2 outputs 50 when the input is (2, 1, 3, 3), and outputs 0 for the rest, after uploading to the cloud, the user logs in the second platform and watches the video elements (1, 2, 1, 4) for 10 minutes, then it is judged that the viewer belongs to the group of user 1, and the video stream is generated according to the group preference.

[0123] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0124] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An artificial intelligence-based cross-platform video promotion method, characterized by, The method comprises the following steps: Step S1. Collecting user interaction data and video playback elements in the first platform, for each user ID in the platform, taking the interaction data as the table header and the video playback elements as the table items, generating a data table in the local server; Step S2. Generating a training library from the data tables of all users in a preset period, training the neural network model with the data in the training library, obtaining a loss model, the loss model taking the video playback elements as the input and the user interaction data as the output, the video playback elements including the label, duration, creator, picture quality and advertisement length of the video, and the interaction data including the click rate and video playback duration, fitting the influence of each user interaction data on the video playback elements using the loss model to obtain a joint influence function; Step S3. Classifying the videos according to the video playback elements, substituting the video playback elements in each classification into the joint influence function, and classifying the joint influence function according to the video type where the maximum value of the joint influence function is located; Step S4. Packaging the joint influence function according to the classification result, uploading the obtained data package after marking the video type to the cloud, and classifying and summarizing the data packages from each platform in the cloud according to the video type as the keyword, and receiving cross-platform data package calls; Step S5. The second platform parses the data package, fits the joint influence function of each video type for consistency to obtain a typical function, substitutes the playback elements of the video watched by the user in the second platform into each typical function, accumulates the scores of each group according to the difference between the predicted result and the actual interaction data, extracts the video according to the score proportion, and generates a recommended video stream.

2. The cross-platform video promotion method based on artificial intelligence according to claim 1, characterized in that: Step S1 comprises: Step S11. Collecting user interaction data and video playback elements in the first platform, the interaction data being the video click rate or video playback duration, and the video playback elements including the label, duration, creator, picture quality and advertisement length of the video; Step S12. Constructing a data linked list for each login behavior of each user ID, the data linked list having n rows and m columns, n being the total number of interaction data and video playback elements, and m being the total number of videos played by the user during the login in the platform, the interaction data being fixed as the table header of the data linked list, the video playback elements being the table items of the data linked list, and storing all the generated data linked lists under the user ID in the local server.

3. The cross-platform video promotion method based on artificial intelligence according to claim 2, characterized in that: Step S2 comprises: Step S21. Preparing a platform statistical period, and the local server generating a new data space every other statistical period as the storage location of the newly generated data linked list in the statistical period; Step S22. Learning the training library constituted by all the data linked lists under a single user ID using the neural network model; Using video playback elements from the training library as input and user interaction data as output, for each data list, fit a functional relationship F(x1,x2,…,x) between the output variable and the input variable. n-1 ), where x n-1 Representing the (n-1)th input variable, the neural network models include: R-CNN model, DeepLab model, U-Net model, and Watershed model; Step S23. According to the selected model category, the elastic network regression is used to fit the loss function: where LOSS is the loss function, SE is the root mean square error of the model, Y is the regularization strength of the model, k is the number of data chains in the training library, wi represents the preset weight of the i th data chain in the training library, r is the proportion of the regularization term, and Fi(x1, x2, …, xi) represents the functional relationship between the output variable and the input variable in the i th data chain. n-1 represents the linear regression function of the function Fi(x1, x2, …, xi). n-1 ) represents the linear regression function of the function Fi(x1, x2, …, xi).​ Step S24. Calculate the magnitude of the regularization strength Y and the regularization term proportion r at the time when the loss function LOSS takes the minimum value, generate a positive definite matrix P according to the values of Y and r, and make G(x1, x2, …, x n-1 )=P·[x1,x2,…,x n-1 ] T , where T is a transpose symbol, and G(x1, x2, …, x n-1 ) is a joint influence function.

4. The cross-platform video promotion method based on artificial intelligence according to claim 3, characterized in that: Step S3 comprises: Step S31. Arranging and combining the playback elements of the video to obtain h video types, substituting the playback elements corresponding to each video type into the joint influence function to obtain a function value, and classifying the joint influence function into the video type corresponding to the maximum value in the prediction set; Step S32. Classifying the joint influence function corresponding to each user ID, and hiding the user ID in the classification result and only keeping the function. Step S4 includes: Step S41. Pack all joint influence functions contained in each video type and upload to the cloud, and the cloud obtains the data packets sent by each platform, classifies and summarizes the data packets according to the video type as the keyword; Step S42. When the user logs in for the first time on the second platform, the second platform downloads all data packets from the cloud, and stores them in the local server according to the video type.

5. The cross-platform video promotion method based on artificial intelligence according to claim 4, characterized in that: Step S5 includes: Step S51. The second platform generates a typical function of each video type in the local server through a consistency equation: wherein B(x1,x2,…,x n-1 ) represents the typical function of the video type, v represents the number of joint influence functions in the video type, and Gj(x1,x2,…,x n-1 ) represents the jth joint influence function in the video type. Solve the consistency equation to obtain the typical function corresponding to each video type; Step S52. Obtain the video playback elements of the videos played by the user in the second platform, and substitute the video playback elements into the typical function of each video type. The obtained value is denoted as a value set {A1, A2, …, At, …, Ah}, wherein At represents the function value of the typical function in the tth video type; Step S53. Obtain the actual interaction data A0 of the played videos by the user, calculate the difference AS between each element in the value set and the actual interaction data, wherein AS = |At-A0|, and accumulate the score UR of the corresponding video type, wherein UR = 1 / AS; Step S54. According to the cumulative score, extract videos from the video library corresponding to each video type, generate a video stream, and push the video stream to the user. 6.A cross-platform video promotion system based on artificial intelligence, characterized in that, The system includes the following modules: platform data module, video association module, cross-platform transmission module, video recommendation module and advertisement insertion module; The platform data module is composed of the database of the video platform, the platform software and the local server, and is used for recording and storing user information, video element information, interaction data and playback history. A neural network model is set in the local server, and the interaction information of a single user is stored in the data pool of the server as a training library of the neural network model; The video association module is used to extract the historical playback data of the user with the user name as the primary key. For the video playback behavior of a single user, the historical playback data of the user is stored in the training pool of the local server to train the neural network model, and a loss model is obtained. The loss model takes video elements as input and user interaction data as output. The video playback elements include: video label, duration, creator, picture quality and advertisement length. The interaction data includes: click rate and video playback time. The association model between the interaction data and each video element is fitted, and the joint influence function between the video playback time and each video element is generated; The cross-platform transmission module is used to cut the joint influence function according to the preset definition domain interval of each video element, classify the user groups according to the cutting result, and for each type of user group, pack all joint influence functions of the users in the group and upload to the cloud. The cloud classifies and summarizes the data packets from each platform according to the group label as the primary key, and receives the cross-platform data packet call. The video recommendation module is used for parsing the cloud data packet on the second platform, fitting the joint influence function in each group label data for consistency, generating the typical influence function corresponding to the group label, inputting the video elements into each typical function for each video playing of the user on the second platform, outputting the predicted sequence of the user interaction data, and accumulating the play coefficients of the group label corresponding to each element according to the distance between the actual interaction data and the sequence elements. The advertisement insertion module is used for extracting the video from the video library under each video element category according to the play coefficient, calculating the video arrangement mode with the maximum video advertisement length as the constraint condition, generating the video stream according to the video arrangement mode, and automatically recommending the user.

7. The cross-platform video promotion system based on artificial intelligence according to claim 6, characterized in that: The platform data module comprises a database unit and an artificial intelligence unit. The database unit is used for storing user information and video playing elements, wherein the video playing elements comprise the label, length, creator, picture quality and advertisement length of the video, and the interaction data comprises the click rate and video playing length. The artificial intelligence unit is arranged in the local database and is used for analyzing the video playing data, and the neural network model adopted comprises an R-CNN model, a DeepLab model, a U-Net model and a Watershed model. 8.The cross-platform video promotion system based on artificial intelligence of claim 7, wherein: The video correlation module comprises a video classification unit, a loss model unit and a joint influence unit. The video classification unit is used for extracting the historical playing data of the user, and classifying the historical playing data according to the video elements. The loss model unit is used for training the neural network model, fitting the influence relationship of each video element on the user interaction data. The joint influence unit is used for representing the fitting result in the form of function, generating the joint influence function with the interaction data as the dependent variable and each video element as the independent variable. 9.The cross-platform video promotion system based on artificial intelligence of claim 8, wherein: The cross-platform transmission module comprises a group preference unit and a cloud communication unit. The group preference unit is used for dividing the user group according to the range of the video elements, and compressing the joint influence function corresponding to the group into a data packet. The cloud communication unit is used for classifying and summarizing the data packets of each platform according to the group label, and is responsible for the cloud uploading and downloading of the data packets. The video recommendation module comprises a user identification unit, a group analysis unit and a play accumulation unit. The user identification unit is used for fitting the joint influence function in each data packet, generating the typical influence function representing the average preference level of the group label. The group analysis unit is used for obtaining the operation behavior of the user in the second platform, and judging the group label of the user. The play accumulation unit is used for scoring the group label to which the user belongs according to the difference between the actual interaction data of the user and the predicted interaction data of each typical influence function. 10.The cross-platform video promotion system based on artificial intelligence of claim 9, wherein: The advertisement insertion module comprises a video extraction unit and a constraint conversion unit. The video extraction unit is used for storing the video in each partition according to the interval of the video element, and extracting the video from each partition according to the group label of the user. The constraint conversion unit is used for controlling the arrangement mode of the extracted video with the video playing length as the constraint condition, so that the sum of the video advertisement lengths is maximum.

Images