Advertisement putting screening system based on artificial intelligence

Abstract

The invention belongs to the technical field of advertisement putting, and discloses an advertisement putting screening system based on artificial intelligence. Comprising a data acquisition module for acquiring user data and external data of all users in a user group and integrating the user data and the external data of any user in the user group into combined user data; the preprocessing module is used for preprocessing the combined user data to obtain correct text data and accurate numerical data; the feature extraction module is used for carrying out feature extraction on the correct text data and the accurate numerical data to obtain user feature indexes, and carrying out clustering processing on user groups based on the user feature indexes to obtain target audience groups; the advertisement generation module is used for constructing an image generation model to generate an advertisement and obtaining high-quality advertisement content; the putting module is used for putting the user group on the basis of the target audience group and the corresponding high-quality advertisement content; the advertisement putting accuracy and efficiency are improved, and the advertisement click rate is indirectly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement placement, and more specifically, to an advertisement placement screening system based on artificial intelligence. Background Art

[0002] The patent application publication number CN107169785A discloses an advertisement delivery method and device, the method comprising: when the number of advertisements is at least two, screening out all follow-up advertisements that can be delivered in the advertisement delivery platform to obtain a first advertisement set; obtaining and screening out advertisements to be delivered that meet the delivery conditions of the advertisement request according to the information carried by the advertisement request to obtain a second advertisement set; sorting the second advertisement set in order to obtain a first sorted set of second advertisements; finding out whether each advertisement to be delivered in the first sorted set of second advertisements has a follow-up advertisement in the first advertisement set; if the advertisement to be delivered has a follow-up advertisement in the first advertisement set, and the priority of the follow-up advertisement is higher than that of the advertisement to be delivered, then adding the follow-up advertisement to the back of the advertisement to be delivered, and using the follow-up advertisement as the advertisement to be delivered of the advertisement request to obtain a second sorted set of second advertisements; and delivering the advertisements to be delivered that meet the number of advertisements according to the second sorted set of second advertisements.

[0003] In the field of traditional advertising technology, the accuracy of advertising is not high enough, and there is a lack of in-depth understanding and precise positioning of individual user preferences and behavior patterns, which leads to users' lack of interest in the delivered ads, directly resulting in deviations in the effectiveness of advertising. Traditional advertising technology fails to collect user data with a large enough coverage and to fully pre-process the data. User data may lack information in certain dimensions, such as user behavior data and attribute data such as the user's home address, income and expenditure level, or occupation. User data may be too redundant, making it difficult to extract user features. For example, there is too much useless text in user data, resulting in complex information. Traditional advertising content is not attractive enough to users, and the advertising content is not designed for the corresponding target audience, which indirectly reduces the click-through rate of ads.

[0004] In view of this, the present invention proposes an advertisement delivery screening system based on artificial intelligence to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: an advertisement placement screening system based on artificial intelligence, comprising:

[0006] A data collection module collects user data and external data of all users in the user group, and integrates the user data and external data of any user in the user group into combined user data;

[0007] A preprocessing module preprocesses the combined user data to obtain correct text data and accurate numerical data;

[0008] The feature extraction module extracts features from correct text data and precise numerical data to obtain user feature indicators, and clusters user groups based on the user feature indicators to obtain target audience groups;

[0009] Advertisement generation module, which builds an image generation model to generate advertisements and obtain high-quality advertising content;

[0010] The delivery module delivers advertisements to user groups based on the target audience groups and the corresponding high-quality advertisement contents; each module is connected to another via wired and / or wireless means.

[0011] Furthermore, the big data platform is used to collect user data and external data of all users in the user group within a preset time; the user data includes user behavior data and user attribute data; the external data includes user social media data and user credit score data.

[0012] Furthermore, the method of preprocessing the combined user data includes:

[0013] Classify the combined user data into text data and numerical data; preprocess the text data to obtain correct text data; perform data normalization on the numerical data and fill in missing values ​​to obtain accurate numerical data;

[0014] The method of preprocessing text data includes:

[0015] Integrate all text data into a long text data, filter out useless content, and use Chinese periods as identifiers to segment the long text data into sentences to obtain text data with clear structure; build a machine learning model, and use the built machine learning model to segment the text data with clear structure to obtain correct text data.

[0016] Furthermore, the method of constructing the machine learning model includes:

[0017] Collect historical combined user data and classify historical text data to obtain a historical text data set; perform sentence processing on the historical text data set to obtain historical text data with clear structure; and use each sentence in the historical text data with clear structure as an input character sequence;

[0018] Define the length of any input character sequence in the historical text data as n;

[0019] Using a word segmentation algorithm to perform word segmentation on the historical text data with clear structure, assigning a label to each word in the historical text data with clear structure, obtaining label text data, wherein any input character sequence in the label text data is a label sequence; using the label text data as a training label;

[0020] The machine learning model includes a bidirectional recurrent layer and a conditional random layer; the bidirectional recurrent layer includes two recurrent units, namely a forward recurrent unit and a reverse recurrent unit;

[0021] A forward loop unit receives any input character sequence in the historical text data with clear structure, and obtains a forward hidden sequence through forward loop processing; a reverse loop unit performs reverse loop processing on the input character sequence that is the same as the input character sequence received by the forward loop unit to obtain a reverse hidden sequence; the forward hidden sequence and the reverse hidden sequence are combined to obtain a bidirectional loop output sequence;

[0022] The bidirectional loop output sequence is input into the conditional random layer, the conditional random layer calculates the output conditional probability, and the word segmentation result is determined based on the conditional probability;

[0023] The calculation formula for the output conditional probability is:

[0024] Among them, P(Y i |M i ) means that when the input sequence is the i-th bidirectional cyclic output sequence M in the bidirectional cyclic output sequence set i When the label sequence is Y i The conditional probability of Y i represents the i-th input character sequence in the label text data; m represents the number of input bidirectional loop output sequences; T represents the local feature function; S represents the transfer feature function; y0 i represents the i-th label in any label sequence; φ represents the feature function;

[0025] Where A represents the label transfer score matrix; A(Y i ,Y i+1 ) represents the sequence Y from the label i Convert to label sequence Y i+1 The score of Y; D represents the label sequence score matrix; D(Y i ) indicates that the label sequence is Y i The score of the time;

[0026] Y=[Y1,Y2,Y3,...,Y m ]; where Y represents the set of all input character sequences in the label text data, Y1, Y2, Y3, ..., Y mRepresent each input character sequence in the set respectively;

[0027] M=[M1,M2,M3,...,M m ]; where M represents the set of bidirectional cyclic output sequences, M1, M2, M3, ..., M m Respectively represent each bidirectional loop output sequence in the set;

[0028] When the conditional probability is greater than or equal to the preset probability threshold, the bidirectional loop output sequence at this time is used as the post-word segmentation sequence. At this time, the parameters are fixed to obtain the constructed machine learning model; the constructed machine learning model is used to perform word segmentation on the structured text data to obtain the correct text data.

[0029] Furthermore, the method of combining the forward hidden sequence and the reverse hidden sequence to obtain a bidirectional cyclic output sequence includes:

[0030] Ways to obtain a forward hidden sequence through forward loop processing include:

[0031] Initialize time step t and hidden sequence h t ;

[0032] Calculate an updated forward loop output and a reset forward loop output, calculate a candidate forward hidden sequence based on the reset forward loop output, and calculate a forward hidden sequence based on the updated forward loop output and the candidate forward hidden sequence;

[0033] Update forward loop output Among them, σ represents the activation function; W0 Z0 Represents the weight matrix for updating the forward loop output; b0 Z0 represents the bias term for updating the output of the forward loop; h t-1 represents the hidden sequence when the time step is t-1; X t represents the input character sequence at time step t; [h t-1 ,X t ] represents the concatenation of the hidden sequence at time step t-1 and the input character sequence at time step t;

[0034] Reset forward loop output Among them, W0 R0 Represents the weight matrix for resetting the forward loop output; b0 R0 Represents the bias term that resets the forward loop output;

[0035] Candidate forward hidden sequence Where tanh represents the hyperbolic tangent activation function; W0 represents the weight matrix of the candidate forward hidden sequence; b0 represents the bias term of the candidate forward hidden sequence; [R0 t⊙h t-1 ,X t ] represents the concatenation of the reset forward loop output and the input character sequence at time step t; R0 T ⊙h t-1 represents the element-by-element product of the reset forward loop output and the hidden sequence at time step T-1; ⊙ represents element-by-element multiplication;

[0036] Hide Sequence T≤n;

[0037] When the time step t = n, h t =h0 n ; At this time, the forward hidden sequence h0 is obtained n ;

[0038] The method of performing reverse loop processing on the input character sequence that is the same as the input character sequence received by the forward loop unit through the reverse loop unit includes:

[0039] Calculate an updated reverse loop output and a reset reverse loop output, calculate a candidate reverse hidden sequence based on the reset reverse loop output, and calculate a reverse hidden sequence based on the updated reverse loop output and the candidate reverse hidden sequence;

[0040] Initialize the opposite hidden sequence

[0041] Update reverse loop output Among them, W1 Z1 Represents the weight matrix for updating the reverse loop output; b1 Z1 Represents the bias term for updating the reverse loop output; represents the opposite hidden sequence when the time step is t-1; Represents the concatenation of the opposite hidden sequence at time step t-1 and the input character sequence at time step t;

[0042] Reset reverse cycle output Among them, W1 R1 Represents the weight matrix for resetting the reverse loop output; b1 R1 Represents the bias term for resetting the reverse loop output;

[0043] Candidate reverse hidden sequence Wherein, W1 represents the weight matrix of the candidate reverse hidden sequence; b1 represents the bias term of the candidate reverse hidden sequence; Represents the concatenation of the reset reverse loop output and the input character sequence at time step t; represents the element-wise product of the reset reverse recurrent output and the opposite hidden sequence at time step t-1;

[0044] Conversely hide the sequence

[0045] When the time step t = n, At this time, the reverse hidden sequence h1 is obtained n ;

[0046] The forward hidden sequence and the reverse hidden sequence are merged element by element to obtain a bidirectional cyclic output sequence.

[0047] Furthermore, the method of extracting features from the correct text data and the accurate numerical data includes:

[0048] Calculate the word frequency CP and word probability DP of each word in the correct text data, and determine the characteristic words based on the word frequency and word probability values;

[0049] Word frequency Where N(word,b) represents the number of times word word appears in the bth sentence of the correct text data; N(word) represents the number of times word word appears in the correct text data; num b represents the number of words in the bth sentence of the correct text data; num represents the number of words in the correct text data;

[0050] Word Probability Wherein, U indicates that the correct text data includes U sentences; numN(word) indicates the number of times the word word appears in any sentence in the correct text data;

[0051] If the word frequency of word word is greater than or equal to the preset word frequency threshold and the word probability of word word is greater than or equal to the preset word probability threshold, then word word is determined as a feature word;

[0052] Extracting network links from the precise numerical data to obtain the URL or webpage information corresponding to the precise numerical data; classifying the precise numerical data based on the URL or webpage information corresponding to the precise numerical data, and assigning corresponding labels to each category of precise numerical data to obtain label-classified numerical data;

[0053] Take all the data in the label classification numerical data as a numerical data set L;

[0054] The numerical data set L includes q data points, each of which has E features;

[0055] Centralize the numerical data set L to obtain the centralized numerical data set L center ;

[0056] Based on the centralized numerical data set L center , construct the covariance matrix;

[0057] Covariance matrix

[0058] Based on the covariance matrix, calculate all the eigenvalues of the matrix and the eigenvectors corresponding to the eigenvalues, and sort the eigenvalues in descending order to obtain a descending sequence composed of eigenvalues;

[0059] Based on the descending sequence composed of eigenvalues, calculate the cumulative contribution rate of the first k eigenvalues;

[0060] Cumulative contribution rate k < E; where λ f represents any eigenvalue in the descending sequence composed of eigenvalues; a represents a constant;

[0061] When the cumulative contribution rate is greater than or equal to the preset cumulative contribution rate threshold, form a set of eigenvectors corresponding to these k eigenvalues to obtain eigenvalue data;

[0062] Based on the feature words and eigenvalue data, use a feature extraction algorithm to perform feature extraction processing to obtain user feature indicators; perform the same preprocessing and feature extraction processing on the combined user data of all users in the user group, and obtain the user feature indicators of each user in the user group respectively.

[0063] Furthermore, the method for clustering the user group based on the user feature indicators includes:

[0064] Define the combination of all user feature indicators of any user in the user group as a feature set, and use each feature set as a data point;

[0065] Calculate the distance between each data point, and calculate the silhouette coefficient based on the distance between each data point, the cohesion degree and the separation degree h ≤ 1000; where NA(h) represents the cohesion degree of the h-th data point; FL(h) represents the separation degree of the h-th data point;

[0066] Cohesion degree dis represents the distance between data points; f0 represents the cohesion degree coefficient;

[0067] Separation degree where f1 represents the separation degree coefficient;

[0068] Randomly select NO data points in the set of data points as the initial clustering centers; continuously adjust the number of clustering centers during the clustering process, calculate the silhouette coefficient and record the value of the silhouette coefficient each time an adjustment is made, and use the number of clustering centers corresponding to the largest value of the silhouette coefficient as the optimal number of clustering centers NI;

[0069] Reuse the clustering algorithm for clustering processing, randomly select NI different data points in the set of data points as initial cluster centers, and calculate the distance between each data point and each initial cluster center; based on the distance between each data point and each initial cluster center, assign each data point in the set of data points to the cluster corresponding to the nearest cluster center; at this time, count the number of clusters to obtain KI clusters; group all clusters into sets to obtain the target audience group.

[0070] Furthermore, the method of constructing an image generation model to generate advertisements includes:

[0071] Define a random noise vector;

[0072] By collecting advertising images that match the target audience group from the advertising resource library and assigning user classification labels to the advertising images based on different clusters in the target audience group, an advertising image training set is obtained; all advertising images in the advertising image training set are real advertising images;

[0073] The image generation model includes a label identifier, an image generator and an image discriminator;

[0074] The image generator includes a conditional embedding layer, a fully connected layer, a conditional attention layer, and a normalization layer;

[0075] The user classification label is input into the label identifier, and the user classification label is converted into a classification label vector through the label identifier; the classification label vector and the random noise vector are input into the image generator, and the classification label vector and the random noise vector are concatenated through the conditional embedding layer to obtain a label noise vector; the label noise vector is processed through the fully connected layer to generate an initial label feature map; the initial label feature map is input into the conditional attention layer, and the attention feature map is generated through the conditional attention layer; based on the initial label feature map and the attention feature map, an updated feature map is generated through the normalization layer.

[0076] Furthermore, the method of generating the attention feature map by processing the conditional attention layer includes:

[0077] The definition of conditional attention layer includes one type of attention structure and two types of attention structure;

[0078] The initial label feature map of the input conditional attention layer is converted into a first-class initial label feature vector through a first-class attention structure, and is converted into a second-class initial label feature vector through a second-class attention structure; based on the first-class initial label feature vector and the second-class initial label feature vector, an attention feature map is obtained through weighted calculation;

[0079] The calculation formula for converting the initial label feature map of the input conditional attention layer into a class of initial label feature vectors through a class of attention structures is:

[0080] Among them, F1 represents a class of initial label feature vectors; represents a type of attention transformation matrix; PI represents the initial label feature map; represents a type of attention coefficient;

[0081] Two-class initial label feature vector in, represents the two-class attention transformation matrix; represents the second-class attention coefficient;

[0082] The calculation formula of the attention feature map obtained by weighted calculation is:

[0083] Among them, PX represents the attention feature map; β represents the activation function; ε1 represents the first-class attention weight; ε2 represents the second-class attention weight.

[0084] Furthermore, the calculation formula for generating the updated feature map through the normalization layer processing is:

[0085] PL=PX⊙PI; where PL represents the updated feature map;

[0086] The updated feature map is adjusted through the residual connection in the normalization layer to obtain the target feature image and output it;

[0087] The image discriminator receives the target feature image generated by the image generator, identifies the user classification label of the target feature image; compares the user classification label of the target feature image with the user classification label of the real advertisement image, and outputs the probability of the label being true; compares the target feature image with the real advertisement image, and outputs the probability of the advertisement being true;

[0088] Repeat the process of using the image generator to generate the target feature image, using the image discriminator to judge the true probability of the label and the true probability of the advertisement, calculating the function value of the image generation loss function, and adjusting the parameters until the function value of the image generation loss function no longer decreases. At this time, the parameters are fixed to obtain the trained image generation model;

[0089] Use the trained image generation model to generate target feature images and obtain high-quality advertising content;

[0090] Based on the target audience and the corresponding high-quality advertising content, advertisements are delivered to the user groups through online channels.

[0091] The technical effects and advantages of the artificial intelligence-based advertising screening system of the present invention are as follows:

[0092] By collecting user data and external data of a certain number of user groups and extracting features from the data, we complete the clustering of user groups and obtain target audience groups with detailed category labels. Then, by building a model, we generate high-quality advertising content that matches the target audience groups, and achieve accurate advertising screening for user groups. Compared with existing experience, the collected data is more comprehensive and widely includes various data of user groups, laying the foundation for accurate classification of user groups. The data of user groups is pre-processed to obtain more accurate data, which indirectly improves the efficiency of subsequent classification of user groups. Based on the feature extraction of user data, user groups are clustered to obtain target audience groups with corresponding labels, which is convenient for screening of advertising populations. By building a model and generating high-quality advertising content that meets the label content based on the labels of the target audience groups, the user groups’ interest in the corresponding advertisements is enhanced, while the click-through rate of advertisements is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 A schematic diagram of an advertisement delivery screening system based on artificial intelligence of the present invention;

[0094] Figure 2 A schematic diagram of an advertisement placement screening method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0095] 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.

[0096] Example 1

[0097] See also Figure 1 As shown, the artificial intelligence-based advertisement delivery screening system described in this embodiment includes:

[0098] A data collection module collects user data and external data of all users in the user group, and integrates the user data and external data of any user in the user group into combined user data;

[0099] A preprocessing module preprocesses the combined user data to obtain correct text data and accurate numerical data;

[0100] The feature extraction module extracts features from correct text data and precise numerical data to obtain user feature indicators, and clusters user groups based on the user feature indicators to obtain target audience groups;

[0101] Advertisement generation module, which builds an image generation model to generate advertisements and obtain high-quality advertising content;

[0102] The delivery module delivers advertisements to user groups based on the target audience groups and the corresponding high-quality advertisement contents; each module is connected to another via wired and / or wireless means.

[0103] The user data and external data of all users in the user group (consisting of 1,000 randomly selected users) within a preset time are collected using a big data platform (such as a public database provided by a government agency); the user data includes user behavior data and user attribute data; the user behavior data includes the user's web browsing history, the user's purchase record on the trading platform, and the user's travel payment record (such as the user's purchase record on shopping platforms such as Taobao and JD.com; the payment record when taking the subway, bus or online car-hailing during daily travel); the user attribute data includes the user's name, gender, date of birth, occupation, home address, mobile phone number and Network IP address and other data; the external data includes user social media data (user social media data can be collected using social media software, such as TikTok) and user credit score data (user credit score data can be collected using relevant software, such as Alipay); user social media data includes posts, comments, likes, forwarded content, accounts followed, and topics discussed by users on online social platforms (for example, content posted by users when using TikTok, Bilibili, Tieba or Zhihu, content followed, and opinions and comments on certain content); user credit score data includes data such as the user's credit status, loan record, and credit card usage.

[0104] By collecting user behavior data, we can analyze users' online habits and interests, understand their consumption behaviors and habits, make personalized advertising recommendations, and push relevant promotional ads; collect user attribute data to supplement users' basic attribute information and deliver corresponding advertisements based on attributes such as age, gender, and income level; collect user social media data to understand users' preferences, social circles, and activity participation, so as to facilitate feature extraction from users; collect user credit score data to evaluate users' credit status and facilitate advertising for financial products and similar products.

[0105] The method of preprocessing the combined user data includes:

[0106] The combined user data is classified into text data (such as posts, comments and social software dynamics posted by users) and numerical data (such as users' payment records, online shopping platform consumption records, travel payment records, takeout consumption records and bank account details); the text data is preprocessed to obtain correct text data; the numerical data is normalized (such as normalization) and missing values ​​are filled (such as difference filling method or similarity filling method) to obtain accurate numerical data.

[0107] The method of preprocessing text data includes:

[0108] Integrate all text data into a long text data, filter out useless content (useless content refers to words without specific meanings that are often contained in text data, such as conjunctions or modal particles; this content can be filtered out by using stop word filtering technology), and use Chinese periods as identifiers to segment the long text data into sentences to obtain text data with clear structure; build a machine learning model, and use the built machine learning model to segment the text data with clear structure to obtain correct text data.

[0109] The method of building a machine learning model includes:

[0110] Collect historical combined user data (historical combined user data refers to combined user data collected from a certain period of time in the past), and classify historical text data to obtain a historical text data set; perform sentence processing on the historical text data set to obtain historical text data with a clear structure; and take each sentence in the historical text data with a clear structure as an input character sequence.

[0111] Define the length of any input character sequence in the historical text data as n;

[0112] Use a word segmentation algorithm (such as the forward maximum matching method) to perform word segmentation on the historical text data with clear structure, assign a label to each word in the historical text data with clear structure, and obtain label text data. Any input character sequence in the label text data is a label sequence; use the label text data as training labels.

[0113] The machine learning model includes a bidirectional recurrent layer and a conditional random layer; the bidirectional recurrent layer includes two recurrent units, namely a forward recurrent unit and a reverse recurrent unit (the forward recurrent unit is used to process an input character sequence from left to right, and the reverse recurrent unit is used to process an input character sequence that is the same as the input character sequence processed by the forward recurrent unit, but in the opposite direction).

[0114] A forward loop unit receives any input character sequence in the historical text data with clear structure, and obtains a forward hidden sequence through forward loop processing; a reverse loop unit performs reverse loop processing on the input character sequence that is the same as the input character sequence received by the forward loop unit to obtain a reverse hidden sequence; the forward hidden sequence and the reverse hidden sequence are combined to obtain a bidirectional loop output sequence.

[0115] The bidirectional loop output sequence is input into the conditional random layer, the conditional random layer calculates the output conditional probability, and the word segmentation result is determined based on the conditional probability.

[0116] The calculation formula for the output conditional probability is:

[0117] Among them, P(Y i |M i ) means that when the input sequence is the i-th bidirectional cyclic output sequence M in the bidirectional cyclic output sequence set i When the label sequence is Y i The conditional probability of Y i represents the i-th input character sequence in the label text data; m represents the number of bidirectional loop output sequences of the input; T represents the local feature function (calculates the score of a certain label to capture the various attributes of the input sequence related to a specific label); S represents the transfer feature function (calculates the transfer score from one label to another to capture the sequential dependency in the label sequence); y0 i represents the i-th label in any label sequence; φ represents the feature function (used to integrate multiple features to evaluate whether a specific label sequence is suitable for the input sequence);

[0118] Where A represents the label transfer score matrix; A(Y i ,Y i+1 ) represents the sequence Y from the label i Convert to label sequence Y i+1 The score of Y i ) indicates that the label sequence is Y i The score of the time;

[0119] Y=[Y1,Y2,Y3,...,Y m ]; where Y represents the set of all input character sequences in the label text data, Y1, Y2, Y3, ..., Y m Represent each input character sequence in the set respectively;

[0120] M=[M1,M2,M3,...,M m]; where M represents the set of bidirectional cyclic output sequences, M1, M2, M3, ..., M m Respectively represent each bidirectional loop output sequence in the set;

[0121] When the conditional probability is greater than or equal to the preset probability threshold, the bidirectional loop output sequence at this time is used as the post-word segmentation sequence. At this time, the parameters are fixed to obtain the constructed machine learning model; the constructed machine learning model is used to perform word segmentation on the structured text data to obtain the correct text data.

[0122] The method of combining the forward hidden sequence and the reverse hidden sequence to obtain a bidirectional cyclic output sequence includes:

[0123] Ways to obtain a forward hidden sequence through forward loop processing include:

[0124] Initialize time step t and hidden sequence h t (At this time t=0, h t =h0, h0 is an initialized zero vector);

[0125] An updated forward loop output and a reset forward loop output are calculated, a candidate forward hidden sequence is calculated based on the reset forward loop output, and a forward hidden sequence is calculated based on the updated forward loop output and the candidate forward hidden sequence.

[0126] Update forward loop output Among them, σ represents the activation function (outputs a value greater than 0 and less than 1, indicating the degree of update of each element in the input character sequence); W0 Z0 Represents the weight matrix for updating the forward loop output; b0 Z0 represents the bias term for updating the output of the forward loop; h t-1 represents the hidden sequence when the time step is t-1; X t represents the input character sequence at time step t; [h t-1 ,X t ] represents the concatenation of the hidden sequence at time step t-1 and the input character sequence at time step t;

[0127] Reset forward loop output Among them, W0 R0 Represents the weight matrix for resetting the forward loop output; b0 R0 Represents the bias term that resets the forward loop output;

[0128] Candidate forward hidden sequence Among them, tanh represents the hyperbolic tangent activation function (outputs a value greater than -1 and less than 1); W0 represents the weight matrix of the candidate forward hidden sequence; b0 represents the bias term of the candidate forward hidden sequence; [R0t ⊙h t-1 ,X t ] represents the concatenation of the reset forward loop output and the input character sequence at time step t; R0 t ⊙h t-1 Represents the element-by-element product of the reset forward loop output and the hidden sequence at time step t-1; ⊙ represents element-by-element multiplication (element-by-element multiplication is a calculation between two vectors or matrices, multiplying the corresponding elements of two vectors or matrices to generate a new matrix or vector);

[0129] Hide Sequence t≤n;

[0130] When the time step t = n, h t =h0 n ; At this time, the forward hidden sequence h0 is obtained n ;

[0131] The method of performing reverse loop processing on the input character sequence that is the same as the input character sequence received by the forward loop unit through the reverse loop unit includes:

[0132] An updated reverse loop output and a reset reverse loop output are calculated, a candidate reverse hidden sequence is calculated based on the reset reverse loop output, and a reverse hidden sequence is calculated based on the updated reverse loop output and the candidate reverse hidden sequence.

[0133] Initialize the opposite hidden sequence

[0134] Update reverse loop output Among them, W1 Z1 Represents the weight matrix for updating the reverse loop output; b1 Z1 Represents the bias term for updating the reverse loop output; represents the opposite hidden sequence when the time step is t-1; Represents the concatenation of the opposite hidden sequence at time step t-1 and the input character sequence at time step t;

[0135] Reset reverse cycle output Among them, W1 R1 Represents the weight matrix for resetting the reverse loop output; b1 R1 Represents the bias term for resetting the reverse loop output;

[0136] Candidate reverse hidden sequence Wherein, W1 represents the weight matrix of the candidate reverse hidden sequence; b1 represents the bias term of the candidate reverse hidden sequence; Represents the concatenation of the reset reverse loop output and the input character sequence at time step t; represents the element-wise product of the reset reverse recurrent output and the opposite hidden sequence at time step t-1;

[0137] Conversely hide the sequence

[0138] When the time step t = n, At this time, the reverse hidden sequence h1 is obtained n ;

[0139] The forward hidden sequence and the reverse hidden sequence are merged element by element to obtain a bidirectional cyclic output sequence.

[0140] The method of extracting features from correct text data and accurate numerical data includes:

[0141] Calculate the word frequency CP and word probability DP of each word in the correct text data, and determine the feature words based on the word frequency and word probability values; word frequency refers to the number of times a word appears in the text data. The larger the value, the higher the frequency of the word in the text and the more important it is. On the contrary, the lower the frequency of the word in the text, the lower its importance. Word probability refers to the probability of a word appearing in other texts. The lower the value, the more times the word appears in other texts and the less discriminative it is. On the contrary, the smaller the value, the less times the word appears in other texts and the more discriminative it is.

[0142] Word frequency Where N(word,b) represents the number of times word word appears in the bth sentence of the correct text data; N(word) represents the number of times word word appears in the correct text data; num b represents the number of words in the bth sentence of the correct text data; num represents the number of words in the correct text data;

[0143] Word Probability Wherein, U indicates that the correct text data includes U sentences; numN(word) indicates the number of times the word word appears in any sentence in the correct text data;

[0144] If the word frequency of word word is greater than or equal to a preset word frequency threshold and the word probability of word word is greater than or equal to a preset word probability threshold, word word is determined to be a feature word.

[0145] Perform network link extraction on the precise numerical data to obtain the URL or web page information corresponding to the precise numerical data (such as the product web page corresponding to the purchase record); classify the precise numerical data based on the URL or web page information corresponding to the precise numerical data, and assign corresponding labels to each category of precise numerical data to obtain label-classified numerical data.

[0146] Since the precise numerical data includes numerical data such as the user's payment and receipt records, online shopping platform consumption records, travel payment records, food delivery consumption records, and bank account details, extract the network links corresponding to these numerical data, match the precise numerical data with the corresponding network links, and classify these data; in order to summarize the content corresponding to a certain type of precise numerical data, use tags to summarize, for example, travel payment records correspond to the user's travel preferences, online shopping platform consumption records correspond to the user's shopping preferences, and bank account details correspond to the user's income and expenditure level; classifying the precise numerical data is conducive to analyzing the precise numerical data and extracting features.

[0147] Take all the data in the labeled classified numerical data as a numerical data set L;

[0148] Among them, the numerical data set L includes q data points, and each data point has E features (q data points represent the number of all data in the labeled classified numerical data is q; E features represent the number of label categories is E);

[0149] Perform centering processing on the numerical data set L to obtain the centered numerical data set L center ; The centering process is for the accurate calculation of the covariance matrix to be constructed later;

[0150] Based on the centered numerical data set L center , construct a covariance matrix;

[0151] Covariance matrix

[0152] Based on the covariance matrix, calculate all the eigenvalues of the matrix and the eigenvectors corresponding to the eigenvalues, and arrange the eigenvalues in descending order to obtain a descending sequence composed of eigenvalues.

[0153] Based on the descending sequence composed of eigenvalues, calculate the cumulative contribution rate of the first k eigenvalues;

[0154] Cumulative contribution rate k < E; where, λ f represents any eigenvalue in the descending sequence composed of eigenvalues; a represents a constant (a ∈ (-∞, +∞));

[0155] The cumulative contribution rate represents the ratio of the sum of the first k eigenvalues in the covariance matrix to the sum of all eigenvalues in the covariance matrix; the larger the cumulative contribution rate, the more these k eigenvalues can fully represent the information of the original data.

[0156] When the cumulative contribution rate is greater than or equal to the preset cumulative contribution rate threshold, form a set of eigenvectors corresponding to these k eigenvalues to obtain the characteristic numerical data.

[0157] Based on the characteristic words and characteristic numerical data, feature extraction processing is performed using a feature extraction algorithm (such as a linear discriminant analysis algorithm) to obtain user characteristic indicators; the combined user data of all users in the user group are subjected to the same preprocessing and feature extraction processing to obtain user characteristic indicators for each user in the user group.

[0158] User feature indicators can reflect users' interests and attributes in various aspects, making it easier to cluster correct text data and precise numerical data.

[0159] The method of clustering user groups based on user characteristic indicators includes:

[0160] Define all user feature indicators of any user in the user group as a feature set, and take each feature set as a data point (each feature set is equivalent to a vector, and each element in the vector corresponds to any user feature indicator of any user in the user group);

[0161] Calculate the distance between each data point and calculate the silhouette coefficient based on the distance, cohesion, and separation between each data point h≤1000; where NA(h) represents the cohesion of the hth data point; FL(h) represents the separation of the hth data point;

[0162] Cohesion dis represents the distance between data points; f0 represents the cohesion coefficient;

[0163] Separation Where f1 represents the separation coefficient;

[0164] Randomly select NO data points from the set of data points as initial cluster centers; continuously adjust the number of cluster centers during the clustering process, calculate the silhouette coefficient and record its value at each adjustment, and take the number of cluster centers corresponding to the largest silhouette coefficient as the optimal number of cluster centers NI.

[0165] Reuse the clustering algorithm for clustering processing, randomly select NI different data points in the set of data points as initial cluster centers, and calculate the distance between each data point and each initial cluster center; based on the distance between each data point and each initial cluster center, assign each data point in the set of data points to the cluster corresponding to the nearest cluster center (nearest means that the distance between the data point and the cluster center is less than or equal to the preset distance threshold); at this time, count the number of clusters to obtain KI clusters; group all clusters into sets to obtain the target audience group.

[0166] The method of constructing an image generation model to generate advertisements includes:

[0167] Define a random noise vector; as one of the input parts of the image generator, the random noise vector is usually obtained from a Gaussian distribution or some probability distribution, which provides randomness for the image generator and enables it to produce diversified outputs.

[0168] An advertising image training set is obtained by collecting advertising images that match the target audience group from an advertising resource library, and assigning user classification labels to the advertising images based on different clusters in the target audience group (user classification labels such as income and expenditure level, preference for a certain brand of electronic products, clothing purchase preferences, etc.); all advertising images in the advertising image training set are real advertising images; each cluster of the target audience group corresponds to a user characteristic indicator or a user characteristic indicator set composed of several related user characteristic indicators, and collecting advertising images that match the target audience group is to obtain advertising content related to the user characteristic indicators of the target audience group.

[0169] The image generation model includes a label identifier, an image generator and an image discriminator; the label identifier is used to enable the model to identify different user classification labels and convert the corresponding user classification labels into classification feature vectors, understand the theme of the image to be generated, and facilitate the subsequent image generator to generate images that meet the user classification labels; the image generator generates a feature map and the image discriminator makes a judgment, and generates higher quality advertising content that meets the user classification labels through continuous confrontation.

[0170] The image generator includes a conditional embedding layer, a fully connected layer, a conditional attention layer, and a normalization layer;

[0171] The user classification label is input into the label identifier, and the user classification label is converted into a classification label vector through the label identifier; the classification label vector and the random noise vector are input into the image generator, and the classification label vector and the random noise vector are concatenated through the conditional embedding layer to obtain a label noise vector; the label noise vector is processed through the fully connected layer to generate an initial label feature map; the initial label feature map is input into the conditional attention layer, and the attention feature map is generated through the conditional attention layer; based on the initial label feature map and the attention feature map, an updated feature map is generated through the normalization layer.

[0172] The method of generating an attention feature map by processing a conditional attention layer includes:

[0173] The definition of conditional attention layer includes one type of attention structure and two types of attention structure;

[0174] The initial label feature map of the input conditional attention layer is converted into a first-class initial label feature vector through a first-class attention structure, and is converted into a second-class initial label feature vector through a second-class attention structure; based on the first-class initial label feature vector and the second-class initial label feature vector, an attention feature map is obtained through weighted calculation.

[0175] The calculation formula for converting the initial label feature map of the input conditional attention layer into a class of initial label feature vectors through a class of attention structures is:

[0176] Among them, F1 represents a class of initial label feature vectors; represents a type of attention transformation matrix; PI represents the initial label feature map; represents a type of attention coefficient;

[0177] The initial label feature vector of the second class in, represents the two-class attention transformation matrix; represents the second-class attention coefficient;

[0178] The calculation formula of the attention feature map obtained by weighted calculation is:

[0179] Among them, PX represents the attention feature map; β represents the activation function; ε1 represents the first-class attention weight; ε2 represents the second-class attention weight;

[0180] The calculation formula for generating the updated feature map through the normalization layer processing is:

[0181] PL=PX⊙PI; where PL represents the updated feature map;

[0182] The updated feature map is adjusted through the residual connection in the normalization layer (such as the residual network ResNet) to obtain the target feature image and output it;

[0183] Residual connections are used to solve the gradient vanishing and gradient exploding problems in deep neural networks. They provide a direct gradient flow path in image generation models, reduce the possibility of gradient vanishing, and ensure the integrity and stability of information.

[0184] The image discriminator receives the target feature image generated by the image generator, identifies the user classification label of the target feature image; compares the user classification label of the target feature image with the user classification label of the real advertisement image, and outputs the probability that the label is true; compares the target feature image with the real advertisement image, and outputs the probability that the advertisement is true.

[0185] The image generator is repeatedly used to generate the target feature image, and the image discriminator is used to determine the true probability of the label and the true probability of the advertisement. The function value of the image generation loss function is calculated, and the parameters are adjusted at the same time until the function value of the image generation loss function no longer decreases. At this time, the parameters are fixed to obtain the trained image generation model.

[0186] Use the trained image generation model to generate target feature images and obtain high-quality advertising content;

[0187] Based on the target audience and the corresponding high-quality advertising content, advertisements are delivered to the user groups through online channels; the high-quality advertising content contains user classification labels corresponding to the target audience, and accurate advertising delivery screening is performed by matching the user classification labels with the target audience.

[0188] This embodiment collects user data and external data of a certain number of user groups and extracts features from the data to complete clustering processing of the user groups, obtains target audience groups with detailed category labels, and then generates high-quality advertising content that matches the target audience groups by building a model, thereby achieving accurate advertising screening for the user groups. Compared with existing experience, the collected data is more comprehensive and widely includes various data of the user groups, laying a foundation for accurately classifying user groups. The data of the user groups is pre-processed to obtain more accurate data, which indirectly improves the efficiency of subsequent classification of user groups. Based on the feature extraction of user data, the user groups are clustered to obtain target audience groups with corresponding labels, which is convenient for screening the advertising population. By building a model and generating high-quality advertising content that meets the label content based on the labels of the target audience groups, the user groups' interest in the corresponding advertisements is enhanced, while the click-through rate of the advertisements is improved.

[0189] Example 2

[0190] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides an advertisement delivery screening method based on artificial intelligence, including:

[0191] S1. Collect user data and external data of all users in the user group, and integrate the user data and external data of any user in the user group into combined user data;

[0192] S2. Combine user data for preprocessing to obtain correct text data and accurate numerical data;

[0193] S3. Extract features from the correct text data and precise numerical data to obtain user feature indicators, and cluster the user groups based on the user feature indicators to obtain the target audience groups;

[0194] S4. Build an image generation model to generate advertisements and obtain high-quality advertising content;

[0195] S5. Target user groups based on target audience groups and corresponding high-quality advertising content;

[0196] Example 3

[0197] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the above-mentioned artificial intelligence-based advertising screening method is implemented.

[0198] Since the electronic device introduced in this embodiment is an electronic device used to implement an artificial intelligence-based advertising screening method in the embodiment of this application, based on the artificial intelligence-based advertising screening method introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not introduced in detail here. As long as the technical personnel of this field implement the electronic device used in the artificial intelligence-based advertising screening method in the embodiment of this application, it belongs to the scope of protection of this application.

[0199] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0200] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. An artificial intelligence-based advertising screening system, characterized in that: include: A data collection module collects user data and external data of all users in the user group, and integrates the user data and external data of any user in the user group into combined user data; A preprocessing module preprocesses the combined user data to obtain correct text data and accurate numerical data; The feature extraction module extracts features from correct text data and precise numerical data to obtain user feature indicators, and clusters user groups based on the user feature indicators to obtain target audience groups; Advertisement generation module, which builds an image generation model to generate advertisements and obtain high-quality advertising content; The delivery module delivers advertisements to user groups based on the target audience groups and the corresponding high-quality advertisement contents; each module is connected to another via wired and / or wireless means.

2. The artificial intelligence-based advertising screening system according to claim 1, characterized in that: The big data platform is used to collect user data and external data of all users in the user group within a preset time; the user data includes user behavior data and user attribute data; the external data includes user social media data and user credit score data.

3. The artificial intelligence-based advertising screening system according to claim 2, characterized in that: The method of preprocessing the combined user data includes: Classify the combined user data into text data and numerical data; preprocess the text data to obtain correct text data; perform data normalization on the numerical data and fill in missing values ​​to obtain accurate numerical data; The method of preprocessing text data includes: Integrate all text data into a long text data, filter out useless content, and use Chinese periods as identifiers to segment the long text data into sentences to obtain text data with clear structure; build a machine learning model, and use the built machine learning model to segment the text data with clear structure to obtain correct text data.

4. The artificial intelligence-based advertising screening system according to claim 3, characterized in that: The method of building a machine learning model includes: Collect historical combined user data and classify historical text data to obtain a historical text data set; perform sentence processing on the historical text data set to obtain historical text data with clear structure; and use each sentence in the historical text data with clear structure as an input character sequence; Define the length of any input character sequence in the historical text data as n; Using a word segmentation algorithm to perform word segmentation on the historical text data with clear structure, assigning a label to each word in the historical text data with clear structure, obtaining label text data, wherein any input character sequence in the label text data is a label sequence; using the label text data as a training label; The machine learning model includes a bidirectional recurrent layer and a conditional random layer; the bidirectional recurrent layer includes two recurrent units, namely a forward recurrent unit and a reverse recurrent unit; A forward loop unit receives any input character sequence in the historical text data with clear structure, and obtains a forward hidden sequence through forward loop processing; a reverse loop unit performs reverse loop processing on the input character sequence that is the same as the input character sequence received by the forward loop unit to obtain a reverse hidden sequence; the forward hidden sequence and the reverse hidden sequence are combined to obtain a bidirectional loop output sequence; The bidirectional loop output sequence is input into the conditional random layer, the conditional random layer calculates the output conditional probability, and the word segmentation result is determined based on the conditional probability; The calculation formula for the output conditional probability is: Among them, P(Y i |M i ) means that when the input sequence is the i-th bidirectional cyclic output sequence M in the bidirectional cyclic output sequence set i When the label sequence is Y i The conditional probability of Y i represents the i-th input character sequence in the label text data; m represents the number of input bidirectional loop output sequences; T represents the local feature function; S represents the transfer feature function; y0 i represents the i-th label in any label sequence; φ represents the feature function; Where A represents the label transfer score matrix; A(Y i ,Y i+1 ) represents the sequence Y from the label i Convert to label sequence Y i+1 The score of Y; D represents the label sequence score matrix; D(Y i ) indicates that the label sequence is Y i The score of the time; Y=[Y1,Y2,Y3,...,Y m ]; where Y represents the set of all input character sequences in the label text data, Y1, Y2, Y3, ..., Y m Represent each input character sequence in the set respectively; M=[M1,M2,M3,...,M m ]; where M represents the set of bidirectional cyclic output sequences, M1, M2, M3, ..., M m Respectively represent each bidirectional loop output sequence in the set; When the conditional probability is greater than or equal to the preset probability threshold, the bidirectional loop output sequence at this time is used as the post-word segmentation sequence. At this time, the parameters are fixed to obtain the constructed machine learning model; the constructed machine learning model is used to perform word segmentation on the structured text data to obtain the correct text data.

5. The artificial intelligence-based advertising screening system according to claim 4, characterized in that: The method of combining the forward hidden sequence and the reverse hidden sequence to obtain a bidirectional cyclic output sequence includes: Ways to obtain a forward hidden sequence through forward loop processing include: Initialize time step t and hidden sequence h t ; Calculate an updated forward loop output and a reset forward loop output, calculate a candidate forward hidden sequence based on the reset forward loop output, and calculate a forward hidden sequence based on the updated forward loop output and the candidate forward hidden sequence; Update forward loop output Among them, σ represents the activation function; W0 Z0 Represents the weight matrix for updating the forward loop output; b0 Z0 represents the bias term for updating the output of the forward loop; h t-1 represents the hidden sequence when the time step is t-1; X t represents the input character sequence at time step t; [h t-1 , X t ] represents the concatenation of the hidden sequence at time step t-1 and the input character sequence at time step t; Reset forward loop output Among them, W0 R0 Represents the weight matrix for resetting the forward loop output; b0 R0 Represents the bias term that resets the forward loop output; Candidate forward hidden sequence Where tanh represents the hyperbolic tangent activation function; W0 represents the weight matrix of the candidate forward hidden sequence; b0 represents the bias term of the candidate forward hidden sequence; [R0 t ⊙h t-1 , X t ] represents the concatenation of the reset forward loop output and the input character sequence at time step t; R0 t ⊙h t-1 represents the element-by-element product of the reset forward loop output and the hidden sequence at time step t-1; ⊙ represents element-by-element multiplication; Hide Sequence When the time step t = n, h t =h0 n ; At this time, the forward hidden sequence h0 is obtained n ; The method of performing reverse loop processing on the input character sequence that is the same as the input character sequence received by the forward loop unit through the reverse loop unit includes: Calculate an updated reverse loop output and a reset reverse loop output, calculate a candidate reverse hidden sequence based on the reset reverse loop output, and calculate a reverse hidden sequence based on the updated reverse loop output and the candidate reverse hidden sequence; Initialize the opposite hidden sequence Update reverse loop output Among them, W1 Z1 Represents the weight matrix for updating the reverse loop output; b1 Z1 Represents the bias term for updating the reverse loop output; represents the opposite hidden sequence when the time step is t-1; Represents the concatenation of the opposite hidden sequence at time step t-1 and the input character sequence at time step t; Reset reverse cycle output Among them, W1 R1 Represents the weight matrix for resetting the reverse loop output; b1 R1 Represents the bias term for resetting the reverse loop output; Candidate reverse hidden sequence Wherein, W1 represents the weight matrix of the candidate reverse hidden sequence; b1 represents the bias term of the candidate reverse hidden sequence; Represents the concatenation of the reset reverse loop output and the input character sequence at time step t; represents the element-wise product of the reset reverse recurrent output and the opposite hidden sequence at time step t-1; Conversely hide the sequence When the time step t = n, At this time, the reverse hidden sequence h1 is obtained n ; The forward hidden sequence and the reverse hidden sequence are merged element by element to obtain a bidirectional cyclic output sequence.

6. The artificial intelligence-based advertising screening system according to claim 5, characterized in that: The method of extracting features from correct text data and accurate numerical data includes: Calculate the word frequency CP and word probability DP of each word in the correct text data, and determine the characteristic words based on the word frequency and word probability values; Word frequency Where N(word,b) represents the number of times word word appears in the bth sentence of the correct text data; N(word) represents the number of times word word appears in the correct text data; num b represents the number of words in the bth sentence of the correct text data; num represents the number of words in the correct text data; Word Probability Wherein, U indicates that the correct text data includes U sentences; numN(word) indicates the number of times the word word appears in any sentence in the correct text data; If the word frequency of word word is greater than or equal to the preset word frequency threshold and the word probability of word word is greater than or equal to the preset word probability threshold, then word word is determined as a feature word; Extracting network links from the precise numerical data to obtain the URL or webpage information corresponding to the precise numerical data; classifying the precise numerical data based on the URL or webpage information corresponding to the precise numerical data, and assigning corresponding labels to each category of precise numerical data to obtain label-classified numerical data; Take all the data in the label classification numerical data as a numerical data set L; The numerical data set L includes q data points, each of which has E features; Centralize the numerical data set L to obtain the centralized numerical data set L center ; Based on the centralized numerical data set L center , construct the covariance matrix; Covariance matrix Based on the covariance matrix, all eigenvalues ​​of the matrix and the eigenvectors corresponding to the eigenvalues ​​are calculated, and the eigenvalues ​​are arranged in descending order to obtain a descending sequence composed of the eigenvalues; Based on the descending sequence of eigenvalues, calculate the cumulative contribution rate of the first k eigenvalues; Cumulative contribution rate k<E; where λ f represents any eigenvalue in the descending sequence of eigenvalues; a represents a constant; When the cumulative contribution rate is greater than or equal to the preset cumulative contribution rate threshold, the eigenvectors corresponding to the k eigenvalues ​​are combined into a eigenvector set to obtain eigenvalue data; Based on the characteristic words and characteristic numerical data, feature extraction processing is performed using a feature extraction algorithm to obtain user characteristic indicators; the combined user data of all users in the user group are subjected to the same preprocessing and feature extraction processing to obtain user characteristic indicators for each user in the user group.

7. The artificial intelligence-based advertising screening system according to claim 6, characterized in that: The method of clustering user groups based on user characteristic indicators includes: Define all user feature indicators of any user in the user group as a feature set, and take each feature set as a data point; Calculate the distance between each data point and calculate the silhouette coefficient based on the distance, cohesion, and separation between each data point h≤1000; where NA(h) represents the cohesion of the hth data point; FL(h) represents the separation of the hth data point; Cohesion dis represents the distance between data points; f0 represents the cohesion coefficient; Separation Where f1 represents the separation coefficient; Randomly select NO data points from the set of data points as initial cluster centers; continuously adjust the number of cluster centers during clustering, calculate the silhouette coefficient and record the value of the silhouette coefficient at each adjustment, and take the number of cluster centers corresponding to the largest silhouette coefficient as the optimal number of cluster centers NI; Reuse the clustering algorithm for clustering processing, randomly select NI different data points in the set of data points as initial cluster centers, and calculate the distance between each data point and each initial cluster center; based on the distance between each data point and each initial cluster center, assign each data point in the set of data points to the cluster corresponding to the nearest cluster center; at this time, count the number of clusters to obtain KI clusters; group all clusters into sets to obtain the target audience group.

8. The artificial intelligence-based advertising screening system according to claim 7, characterized in that: The method of constructing an image generation model to generate advertisements includes: Define a random noise vector; By collecting advertising images that match the target audience group from the advertising resource library and assigning user classification labels to the advertising images based on different clusters in the target audience group, an advertising image training set is obtained; all advertising images in the advertising image training set are real advertising images; The image generation model includes a label identifier, an image generator and an image discriminator; The image generator includes a conditional embedding layer, a fully connected layer, a conditional attention layer, and a normalization layer; The user classification label is input into the label identifier, and the user classification label is converted into a classification label vector through the label identifier; the classification label vector and the random noise vector are input into the image generator, and the classification label vector and the random noise vector are concatenated through the conditional embedding layer to obtain a label noise vector; the label noise vector is processed through the fully connected layer to generate an initial label feature map; the initial label feature map is input into the conditional attention layer, and the attention feature map is generated through the conditional attention layer; based on the initial label feature map and the attention feature map, an updated feature map is generated through the normalization layer.

9. The artificial intelligence-based advertising screening system according to claim 8, characterized in that: The method of generating an attention feature map by processing a conditional attention layer includes: The definition of conditional attention layer includes one type of attention structure and two types of attention structure; The initial label feature map of the input conditional attention layer is converted into a first-class initial label feature vector through a first-class attention structure, and is converted into a second-class initial label feature vector through a second-class attention structure; based on the first-class initial label feature vector and the second-class initial label feature vector, an attention feature map is obtained through weighted calculation; The calculation formula for converting the initial label feature map of the input conditional attention layer into a class of initial label feature vectors through a class of attention structures is: Among them, F1 represents a class of initial label feature vectors; represents a type of attention transformation matrix; PI represents the initial label feature map; represents a type of attention coefficient; The initial label feature vector of the second class in, represents the two-class attention transformation matrix; represents the second-class attention coefficient; The calculation formula of the attention feature map obtained by weighted calculation is: Among them, PX represents the attention feature map; β represents the activation function; ε1 represents the first-class attention weight; ε2 represents the second-class attention weight.

10. The artificial intelligence-based advertisement screening system according to claim 9, characterized in that: The calculation formula for generating the updated feature map through the normalization layer processing is: PL=PX⊙PI; where PL represents the updated feature map; The updated feature map is adjusted through the residual connection in the normalization layer to obtain the target feature image and output it; The image discriminator receives the target feature image generated by the image generator, identifies the user classification label of the target feature image; compares the user classification label of the target feature image with the user classification label of the real advertisement image, and outputs the probability of the label being true; compares the target feature image with the real advertisement image, and outputs the probability of the advertisement being true; Repeat the process of using the image generator to generate the target feature image, using the image discriminator to judge the true probability of the label and the true probability of the advertisement, calculating the function value of the image generation loss function, and adjusting the parameters until the function value of the image generation loss function no longer decreases. At this time, the parameters are fixed to obtain the trained image generation model; Use the trained image generation model to generate target feature images and obtain high-quality advertising content; Based on the target audience and the corresponding high-quality advertising content, advertisements are delivered to the user groups through online channels.

Images