A method for identifying high-value advertising users based on big data

Abstract

The present invention relates to the field of big data, and in particular to a method for identifying high-value advertising users based on big data, comprising obtaining an advertising information data set, a user information data set and a flow data set; dividing the flow data set to obtain multiple groups of user flow samples, pre-processing each group of user flow samples to obtain corresponding training samples, constructing an advertising high-value user identification model, and using the training sample set to train the advertising high-value user identification model; and optimizing the parameters of the advertising high-value user identification model through double threshold contrast loss and partial weight cross entropy loss until the model converges; obtaining user data to be identified and inputting it into the trained advertising high-value user identification model to obtain a user value identification result; the present invention improves identification accuracy.

Description

Technical Field

[0001] The present invention relates to the field of big data, and in particular to a method for identifying high-value advertising users based on big data. Background Art

[0002] Advertising is a technology derived from the dissemination of information. With the development of Internet technology and the advertising industry, advertisements are widely disseminated to users through various channels. For a certain advertisement, users who are more likely to purchase the advertised product or service are called high-value users of the advertisement. The original intention of advertisers to place advertisements is to attract high-value users to make purchases. Simple large-scale delivery methods can cover most users, but most users may not be interested in the advertisement, so it is impossible to efficiently convert users, and this method cannot timely feedback the user's preference level; in addition, too much push will cause disgust among some users, which runs counter to the original intention of advertisers. Therefore, the method of identifying high-value users of advertisements and then pushing them has attracted attention.

[0003] The identification of high-value advertising users involves a large amount of data, such as the browsing data, click data, purchase data, etc. of the users; in addition, the user's preference for the advertising products will change, especially for some consumable goods that need to be purchased multiple times, such as paper towels and laundry detergent. Users may need to buy them in a certain period of time, but not in other periods of time. The existing identification methods build efficient customer portraits and prediction models by deeply mining multi-dimensional data such as advertising clicks, browsing, and conversions, but they have the disadvantages of complex model structure, low data utilization, and low recognition accuracy. The present invention mainly optimizes data utilization, utilization efficiency, and recognition accuracy. Through the automated processing of pipeline data and the automated calculation of key model parameters, a suitable initial model architecture is used to adapt to business data, thereby reducing the gap between the model and data. Summary of the invention

[0004] To solve the above problems, the present invention provides a method for identifying high-value users of advertisements based on big data, comprising the following steps:

[0005] S1. Obtain advertising information dataset, user information dataset and flow dataset;

[0006] The flow data set is used to represent the interaction relationship between users and advertisements. Each flow data includes user ID, advertisement ID, interaction time between users and advertisements, interaction type between users and advertisements, and behavior conversion type between users and advertisements. The interaction types include browsing, liking, and receiving coupons. The behavior conversion type includes generated conversion behavior and non-conversion behavior.

[0007] S2. Classify the flow data set according to the user ID to obtain multiple user flow data sets, and then divide each user flow data set to obtain multiple groups of user flow samples; each group of user flow samples includes the flow data of the user within thirty-four days, of which the first thirty days are a feature interval data and the last four days are an evaluation interval data;

[0008] S3. Preprocess each group of user flow samples to obtain corresponding training samples, each group of training samples includes advertising information matrix, user information, actual conversion user advertising matching pair set and feature interval data;

[0009] Step S3 pre-processes any group of user flow samples to obtain corresponding training samples, including:

[0010] S31. Filter out all the flow data of the behavior conversion type that generates conversion behavior in the evaluation interval data, and match the user and advertisement corresponding to each flow data to obtain an actual conversion user-ad matching pair;

[0011] S32. Extract all advertisement IDs in the evaluation interval data to form a list of advertisements to be issued in the evaluation interval;

[0012] S33. Based on the advertising information data set, obtain an advertisement information corresponding to each advertisement ID in the advertisement list to be issued during the evaluation period to form an advertisement information matrix; based on the user information data set, the user information of the user ID corresponding to the user flow sample;

[0013] S34. The advertising information matrix, user information, actual conversion user advertising matching pair set and feature interval data form a set of training samples;

[0014] S4. Constructing an advertising high-value user identification model, the advertising high-value user identification model includes a feature extraction module, a behavior representation dynamic fusion module, and a multimodal feature fusion module;

[0015] S5. Use the training sample set to train the advertising high-value user identification model; and optimize the parameters of the advertising high-value user identification model through double threshold contrast loss and partial weighted cross entropy loss until the model converges;

[0016] S6. Obtain the user data to be identified and input it into the trained advertising high-value user identification model to obtain the user value identification result; the user data to be identified includes the user's transaction data, user information and advertising information for the past month.

[0017] Furthermore, in step S1, the advertising information data set is composed of multiple pieces of advertising information, each piece of advertising information includes an advertising ID, an advertising title, an advertising description, an advertising image, an advertising type, a delivery time, an advertising text length, keywords and image features; the user information data set is composed of multiple pieces of user information, each piece of user information includes a user ID, a user gender, a user age, a user geographic location, a user activity, a user portrait label and a timing analysis result of a user interaction sequence.

[0018] Furthermore, the image features include color histogram features and LBP texture statistical features, wherein the color histogram features include red color channel features, green color channel features and blue color channel features for the advertising image, and each color channel feature includes a histogram mean, a histogram variance and a histogram quartile; the LBP texture statistical features include an LBP histogram mean, an LBP histogram standard deviation and an LBP texture quartile.

[0019] Furthermore, the process of obtaining the time series analysis results of the user interaction sequence includes:

[0020] S11. Obtain the user's interactive sequence data and perform STL decomposition to obtain trend sequence, seasonal sequence and difference sequence;

[0021] S12. Convert the seasonal series into a spectrum graph, and take the highest peak in the spectrum graph as the interaction period;

[0022] S13. Use linear regression to fit the trend sequence and the difference sequence to obtain the trend sequence slope and the difference sequence slope, and use the trend sequence slope and the difference sequence slope as the interaction trend and the change trend respectively;

[0023] S14. The interaction cycle, interaction trend and change trend constitute the time series analysis results of the user interaction sequence.

[0024] Furthermore, the feature extraction module includes a plurality of transformer-encoder modules, and the number k of transformer-encoder modules is calculated as follows:

[0025]

[0026]

[0027] Among them, d represents the characteristic dimension of the flow data, Noise represents the noise level, and I represents the mean of the chi-square test between attribute features. Represents the average value of the correlation between attribute features.

[0028] Furthermore, the behavior representation dynamic fusion module is composed of multiple trainable dynamic parameters, where the number of dynamic parameters m is determined by the following formula:

[0029]

[0030] Among them, ACF represents the autocorrelation sequence, mean() represents the average value, t represents the autocorrelation lag order, and ACF(t) represents the t-th order autocorrelation coefficient.

[0031] Furthermore, in step S5, the specific process of inputting any group of training samples into the advertising high-value user recognition model for training includes:

[0032] S51. Input the feature interval data into the feature extraction module and output the feature vector;

[0033] S52. Input the feature vector into the behavior representation dynamic fusion module and output the fusion feature matrix;

[0034] S53. Using a multimodal feature fusion module, the fusion feature matrix, the user information, and the advertising information matrix are transformed into dimensions, and then a dual threshold contrast loss is calculated;

[0035] S54. Each fusion feature in the dimensionally converted fusion feature matrix is ​​spliced ​​with the dimensionally converted user information in pairs to obtain a first spliced ​​information matrix; each piece of advertising information in the dimensionally converted advertising information matrix is ​​spliced ​​with each piece of spliced ​​information in the first spliced ​​information matrix in pairs to obtain a second spliced ​​information matrix;

[0036] S55. Input the second concatenated information matrix into the linear layer, output a one-dimensional vector, use the sigmoid function to map the one-dimensional vector to the [0,1] interval to obtain the user value recognition result, and use the user value recognition result to calculate the partial weighted cross entropy loss.

[0037] Furthermore, the calculation formula of the dual threshold contrast loss is:

[0038] DistContrastiveLoss

[0039] =dist(a,b)*max(0,1-dist(a,b))+dist(b,c)

[0040] *max(0,1-dist(b,c))

[0041] Among them, dist() represents the calculation of Euclidean distance, a represents the user information after dimension transformation, b represents the fusion feature matrix after dimension transformation, and c represents the advertising information matrix after dimension transformation.

[0042] Furthermore, the calculation formula of partial weighted cross entropy loss is:

[0043]

[0044]

[0045] Among them, n represents the number of samples, w represents the loss weight, and y i Represents the label of the i-th sample. If the i-th sample is a positive sample, then y i =1, if the i-th sample is a negative sample, then y i =0; flag() indicates the negative sample labeling function, nflag() indicates the positive sample labeling function, p i Represents the model output of the i-th sample.

[0046] Beneficial effects of the present invention:

[0047] The high-value customer identification method for Internet advertising based on big data uses massive data and advanced technologies to achieve accurate identification of potential high-value customers and targeted advertising, thereby improving advertising effectiveness and commercial value. There are countless stream data on the Internet, and a large amount of data is suitable for fitting with a transformer model with a large parameter scale. This method uses a quantitative method to automatically calculate the noise level and data autocorrelation in the data set to determine the number of transformer layers and the length of the time series, thereby improving the rationality of data utilization; at the same time, through the combination of multimodal data, the efficiency of data utilization is improved; finally, through the partial weight loss function, data bias is eliminated to a certain extent, and the accuracy of model evaluation is improved.

[0048] Through the application of this method, advertisers can better understand the needs and behavioral characteristics of target users, achieve accurate delivery and personalized recommendations, increase ad click-through rate and conversion rate, and thus obtain higher commercial benefits. At the same time, advertising platforms can also use this method to optimize advertising delivery strategies and refine operations, improve advertising effects and commercial value, and maximize commercial benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the method for identifying high-value users of advertisements of the present invention;

[0050] Figure 2 This is a flow chart of the training of the advertisement high-value user identification model of the present invention;

[0051] Figure 3 This is a scoring flow chart of the advertising high-value user identification model of the present invention. DETAILED DESCRIPTION

[0052] 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.

[0053] The present invention provides a method for identifying high-value users of advertisements based on big data, such as Figure 1 As shown, the following steps are included:

[0054] S1. Obtain an advertisement information dataset, a user information dataset, and a flow dataset.

[0055] The flow data set is used to represent the interactive relationship between users and advertisements. Each flow data includes user ID, advertisement ID, interaction time between users and advertisements, interaction type between users and advertisements, and behavior conversion type between users and advertisements. The interaction types include browsing, liking, and receiving coupons. The behavior conversion types include behaviors that have generated conversions and behaviors that have not generated conversions.

[0056] Specifically, in step S1, the advertising information data set consists of multiple pieces of advertising information, each piece of advertising information includes an advertising ID, an advertising title, an advertising description, an advertising picture, an advertising type, a delivery time, an advertising text length, keywords and picture features; the user information data set consists of multiple pieces of user information, each piece of user information includes a user ID, a user gender, a user age, a user geographic location, a user activity, a user portrait label and a timing analysis result of a user interaction sequence.

[0057] Specifically, the image features include color histogram features and LBP texture statistical features, wherein the color histogram features include red color channel features, green color channel features and blue color channel features for the advertising image, and each color channel feature includes a histogram mean, a histogram variance and a histogram quartile; the LBP texture statistical features include an LBP histogram mean, an LBP histogram standard deviation and an LBP texture quartile.

[0058] Specifically, the process of obtaining the time series analysis results of the user interaction sequence includes:

[0059] S11. Obtain the user's interactive sequence data and perform STL decomposition to obtain trend sequence, seasonal sequence and difference sequence;

[0060] S12. Convert the seasonal series into a spectrum graph, and take the highest peak in the spectrum graph as the interaction period;

[0061] S13. Use linear regression to fit the trend sequence and the difference sequence to obtain the trend sequence slope and the difference sequence slope, and use the trend sequence slope and the difference sequence slope as the interaction trend and the change trend respectively;

[0062] S14. The interaction cycle, interaction trend and change trend constitute the time series analysis results of the user interaction sequence.

[0063] S2. Classify the flow data set according to user ID, and group the flow data with the same user ID into one user flow data set, thereby obtaining multiple user flow data sets; then use a sliding window to divide each user flow data set to obtain multiple groups of user flow samples; each group of user flow samples includes the flow data of the user within thirty-four days, of which the first thirty days are a feature interval data and the last four days are an evaluation interval data.

[0064] Specifically, the flow data set is divided into multiple groups of flow samples by sliding window division, and the number of sliding steps is 21 days.

[0065] S3. Preprocess each group of user flow samples to obtain corresponding training samples. Each group of training samples includes an advertising information matrix, user information, an actual conversion user-ad matching pair set, and feature interval data.

[0066] Step S3 pre-processes any group of user flow samples to obtain corresponding training samples, including:

[0067] S31. Filter out all the flow data of the behavior conversion type that generates conversion behavior in the evaluation interval data, and match the user and advertisement corresponding to each flow data to obtain an actual conversion user-ad matching pair;

[0068] S32. Extract all advertisement IDs in the evaluation interval data to form a list of advertisements to be issued in the evaluation interval;

[0069] S33. Based on the advertising information data set, obtain an advertisement information corresponding to each advertisement ID in the advertisement list to be issued during the evaluation period, and form an advertisement information matrix corresponding to all advertisement IDs;

[0070]

[0071] Among them, X is the advertising information matrix, x n , n={1,2,…,N} represents the advertisement information corresponding to the nth advertisement ID, where N represents the number of advertisement IDs in the list of advertisements to be released during the evaluation period, and x n =[x n1 ,x n2 ,…,x ng ], x n1It represents the first attribute feature of the nth advertisement information, and g represents the number of attribute features of the advertisement information. In the present invention, each advertisement information includes 9 attribute features, namely, advertisement ID, advertisement title, advertisement description, advertisement picture, advertisement type, delivery time, advertisement text length, keywords and picture features, so g is 9.

[0072] Based on the user information data set, the user information of the user ID corresponding to the user flow sample is obtained, where z is used i =[z i1 ,z i2 ,…,z ih ] represents the user information corresponding to the i-th user ID, z i1 represents the first attribute feature of the i-th user information, and h represents the number of attribute features of the advertising information; in the present invention, each piece of user information includes 7 attribute features, namely, user ID, user gender, user age, user geographic location, user activity, user portrait label and time series analysis result of user interaction sequence, so h is 7. ;

[0073] S34. The advertisement information matrix, user information, actual conversion user advertisement matching pair set and feature interval data are combined into a set of training samples.

[0074] S4. Construct an advertising high-value user identification model, which includes a feature extraction module, a behavior representation dynamic fusion module, and a multimodal feature fusion module.

[0075] Specifically, the feature extraction module is composed of multiple transformer-encoder modules connected together, wherein the noise in the interval feature data will affect the learning ability of the transformer-encoder module, and the deep structure, i.e., multiple transformer-encoder modules, can extract abstract features and resist the interference of noise, and separate the noise well, so as to better match the characteristics of the interval feature data. Therefore, the present invention sets the number of transformer-encoder modules to be positively correlated with the noise of the interval feature data. Among them, the calculation formula for the number k of transformer-encoder modules is:

[0076]

[0077]

[0078] Among them, d represents the characteristic dimension of each stream data, Noise represents the noise level, and I represents the mean of the chi-square test between attribute features. Indicates the average value of the correlation between attribute features; in the embodiment of the present invention, a piece of flow data includes five attribute features, namely user ID, advertisement ID, interaction time between user and advertisement, interaction type between user and advertisement, and behavior conversion type between user and advertisement, so d is 5. In the present invention, the chi-square test value and correlation value between every two attribute features in the flow data set are calculated, and then the average value is calculated.

[0079] Specifically, the behavior representation dynamic fusion module consists of multiple trainable dynamic parameters, where the number of dynamic parameters m is determined by the following formula:

[0080]

[0081] Among them, ACF represents the autocorrelation sequence, mean() represents the average value, t represents the autocorrelation lag order, and ACF(t) represents the t-th order autocorrelation coefficient. Specifically, if a stream data has 5 attribute features, then assuming that the feature interval data includes K stream data, each attribute feature corresponds to a 1×K feature autocorrelation sequence, and the ACF of the autocorrelation sequence can be obtained by horizontally averaging the 5 feature autocorrelation sequences.

[0082] The length of the dynamic parameter group is equal to the number of running data in the interval feature data of the input feature extraction module, so the number m of dynamic parameters is also the number of running data in the interval feature data of the input feature extraction module. The number m is positively correlated with the autocorrelation coefficient, so that the interval feature data can fully accommodate event-related information and structural information within the interval feature data (that is, the input sequence can better characterize user behavior). At the same time, the length is negatively correlated with the data noise level, which enhances the ability to filter noise from the data and reduces the risk of overfitting. At the same time, the dynamic parameter group can be dynamically adjusted according to the characteristics of the data, enhancing the expression ability of the model and better adapting to the scene.

[0083] Specifically, the multimodal feature fusion module includes three linear layers whose parameters are not shared, which are used to perform dimensional transformation on the outputs of the user information, advertising information and behavior representation dynamic fusion modules, and all are transformed and mapped to 32 dimensions.

[0084] S5. Use the training sample set to train the advertising high-value user identification model; and optimize the parameters of the advertising high-value user identification model through double threshold contrast loss and partial weighted cross entropy loss until the model converges.

[0085] Specifically, Figure 2 As shown, the specific process of inputting any group of training samples into the advertising high-value user recognition model for training in step S5 includes:

[0086] S51. Input the feature interval data into the feature extraction module, and take the output vector of the last transformer-encoder module in the feature extraction module as the feature vector. The dimension of the feature vector is [B, L, d], where B, L, and d represent batch_size, the number of pipeline data items in the feature interval data, and the feature dimension, respectively; the number of pipeline data items L in the feature interval data is equal to the number of dynamic parameters m of the behavior representation dynamic fusion module.

[0087] S52. Input the feature vector into the behavior representation dynamic fusion module and output the fusion feature matrix;

[0088] Specifically, the feature vector input behavior representation dynamic fusion module is dynamically fused with m dynamic parameters, which is expressed as:

[0089]

[0090] Among them, matrix i represents the feature matrix with the i-th dimension [B, d] in the feature vector (i.e., the i-th [B, d] dimension corresponding to the L dimension in the feature vector with the dimension [B, L, d]), weight i represents the i-th dynamic parameter, and result represents the fused feature matrix, whose dimension is [B, d].

[0091] S53. Using a multimodal feature fusion module, the fusion feature matrix, the user information, and the advertising information matrix are transformed into dimensions, and then a dual threshold contrast loss is calculated;

[0092] Specifically, the calculation formula of the dual threshold contrast loss is:

[0093] DistContrastiveLoss

[0094] =dist(a,b)*max(0,1-dist(a,b))+dist(b,c)

[0095] *max(0,1-dist(b,c))

[0096] Where dist represents the Euclidean distance, a represents the user information after dimension transformation, b represents the fusion feature matrix after dimension transformation, and c represents the advertising information matrix after dimension transformation. Optimizing this loss can map several different matrices to approximate spaces, so that the features represented by different modules are mapped in similar semantic spaces.

[0097] S54. Each fusion feature in the dimensionally converted fusion feature matrix is ​​spliced ​​with the dimensionally converted user information in pairs to obtain a first spliced ​​information matrix; each piece of advertising information in the dimensionally converted advertising information matrix is ​​spliced ​​with each piece of spliced ​​information in the first spliced ​​information matrix in pairs to obtain a second spliced ​​information matrix;

[0098] Specifically, each piece of second splicing information in the second splicing information matrix is ​​labeled, and a judgment is made based on the actual conversion user-advertisement matching pair set. If the user and the advertisement in the second splicing information are one of the actual conversion user-advertisement matching pairs, then the second splicing information is a positive sample, otherwise it is a negative sample.

[0099] S55. Input the second concatenated information matrix into the linear layer, output a one-dimensional vector, use the sigmoid function to map the one-dimensional vector to the interval [0,1] to obtain the user value recognition result (that is, obtain the score for the user), and use the user value recognition result to calculate the partial weighted cross entropy loss, which is expressed as:

[0100]

[0101]

[0102]

[0103]

[0104] Among them, n represents the number of samples, w represents the loss weight, and y i Represents the label of the i-th sample. If the i-th sample is a positive sample, then y i =1, if the i-th sample is a negative sample, then y i = 0; flag() represents the negative sample marking function, where if the i-th sample is a negative sample, flag(y i )=flag(0)=1;nflag() represents the positive sample labeling function, p i Represents the model output of the i-th sample.

[0105] S6. Obtain the user data to be identified and input it into the trained advertising high-value user identification model to obtain the user value identification result; the user data to be identified includes the user's transaction data, user information and advertising information for the past month.

[0106] Figure 3 The process of obtaining the user data to be identified and inputting the trained advertising high-value user identification model to obtain the user value identification result.

[0107] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0108] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying high-value users of advertisements based on big data, characterized in that: The following steps are involved: S1. Obtain an advertisement information data set, a user information data set, and a flow data set; the flow data set is used to represent the interaction relationship between the user and the advertisement, and each flow data includes a user ID, an advertisement ID, the interaction time between the user and the advertisement, the interaction type between the user and the advertisement, and the behavior conversion type between the user and the advertisement; the interaction type includes browsing, liking, and receiving coupons, and the behavior conversion type includes generated conversion behavior and no conversion behavior; S2. Classify the flow data set according to the user ID to obtain multiple user flow data sets, and then divide each user flow data set to obtain multiple groups of user flow samples; each group of user flow samples includes the flow data of the user within thirty-four days, of which the first thirty days are a feature interval data and the last four days are an evaluation interval data; S3. Preprocess each group of user flow samples to obtain corresponding training samples, each group of training samples includes advertising information matrix, user information, actual conversion user advertising matching pair set and feature interval data; Step S3 pre-processes any group of user flow samples to obtain corresponding training samples, including: S31. Filter out all the flow data of the behavior conversion type that generates conversion behavior in the evaluation interval data, and match the user and advertisement corresponding to each flow data to obtain an actual conversion user-ad matching pair; S32. Extract all advertisement IDs in the evaluation interval data to form a list of advertisements to be issued in the evaluation interval; S33. Based on the advertising information data set, obtain an advertisement information corresponding to each advertisement ID in the advertisement list to be issued during the evaluation period to form an advertisement information matrix; based on the user information data set, obtain the user information of the user ID corresponding to the user flow sample; S34. The advertising information matrix, user information, actual conversion user advertising matching pair set and feature interval data form a set of training samples; S4. Constructing an advertising high-value user identification model, the advertising high-value user identification model includes a feature extraction module, a behavior representation dynamic fusion module, and a multimodal feature fusion module; The feature extraction module includes multiple transformer-encoder modules, and the number k of transformer-encoder modules is calculated as follows: Among them, d represents the characteristic dimension of the flow data, Noise represents the noise level, and I represents the mean of the chi-square test between attribute features. Represents the average value of the correlation between attribute features; The behavior representation dynamic fusion module consists of multiple trainable dynamic parameters, where the number of dynamic parameters m is determined by the following formula: Among them, ACF represents the autocorrelation sequence, mean() represents the average value, t represents the autocorrelation lag order, and ACF(t) represents the t-th order autocorrelation coefficient; S5. Use the training sample set to train the advertising high-value user identification model; and optimize the parameters of the advertising high-value user identification model through double threshold contrast loss and partial weighted cross entropy loss until the model converges; S6. Obtain the user data to be identified and input it into the trained advertising high-value user identification model to obtain the user value identification result; the user data to be identified includes the user's transaction data, user information and advertising information for the past month.

2. According to the method for identifying high-value users of advertisements based on big data according to claim 1, it is characterized in that: In step S1, the advertisement information data set is composed of multiple advertisement information, each advertisement information includes advertisement ID, advertisement title, advertisement description, advertisement picture, advertisement type, delivery time, advertisement text length, keywords and picture features; The user information data set consists of multiple pieces of user information, each piece of user information includes user ID, user gender, user age, user geographic location, user activity, user portrait label and time series analysis results of user interaction sequence.

3. The method for identifying high-value advertising users based on big data according to claim 2, characterized in that: The image features include color histogram features and LBP texture statistical features, wherein the color histogram features include red color channel features, green color channel features and blue color channel features for the advertising image, and each color channel feature includes a histogram mean, a histogram variance and a histogram quartile; the LBP texture statistical features include an LBP histogram mean, an LBP histogram standard deviation and an LBP texture quartile.

4. The method for identifying high-value users of advertisements based on big data according to claim 2, characterized in that: The process of obtaining the time series analysis results of the user interaction sequence includes: S11. Obtain the user's interactive sequence data and perform STL decomposition to obtain trend sequence, seasonal sequence and difference sequence; S12. Convert the seasonal series into a spectrogram, and take the highest peak in the spectrogram as the interaction period; S13. Use linear regression to fit the trend sequence and the difference sequence to obtain the trend sequence slope and the difference sequence slope, and use the trend sequence slope and the difference sequence slope as the interaction trend and the change trend respectively; S14. The interaction cycle, interaction trend and change trend constitute the time series analysis results of the user interaction sequence.

5. The method for identifying high-value users of advertisements based on big data according to claim 1, characterized in that: Step S5: The specific process of inputting any group of training samples into the advertising high-value user recognition model for training includes: S51. Input the feature interval data into the feature extraction module and output the feature vector; S52. Input the feature vector into the behavior representation dynamic fusion module and output the fusion feature matrix; S53. Use a multimodal feature fusion module to perform dimension transformation on the fusion feature matrix, user information, and advertising information matrix, and then calculate the dual threshold contrast loss; S54. Each fusion feature in the dimensionally converted fusion feature matrix is ​​spliced ​​with the dimensionally converted user information in pairs to obtain a first spliced ​​information matrix; each piece of advertising information in the dimensionally converted advertising information matrix is ​​spliced ​​with each piece of spliced ​​information in the first spliced ​​information matrix in pairs to obtain a second spliced ​​information matrix; S55. Input the second concatenated information matrix into the linear layer, output a one-dimensional vector, use the sigmoid function to map the one-dimensional vector to the [0,1] interval to obtain the user value recognition result, and use the user value recognition result to calculate the partial weighted cross entropy loss.

6. The method for identifying high-value users of advertisements based on big data according to claim 5, characterized in that: The calculation formula of dual threshold contrast loss is: DistContrastiveLoss =dist(a,b)*max(0,1-dist(a,b))+dist(b,c)*max(0,1-dist(b,c)) Among them, dist() represents the calculation of Euclidean distance, a represents the user information after dimension transformation, b represents the fusion feature matrix after dimension transformation, and c represents the advertising information matrix after dimension transformation.

7. The method for identifying high-value users of advertisements based on big data according to claim 5, characterized in that: The calculation formula for partial weighted cross entropy loss is: Among them, n represents the number of samples, w represents the loss weight, and y i Represents the label of the i-th sample. If the i-th sample is a positive sample, then y i =1, if the i-th sample is a negative sample, then y i =0; flag() indicates the negative sample labeling function, nflag() indicates the positive sample labeling function, p i Represents the model output of the i-th sample.

Images