Advertisement planning method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses an advertisement planning method and system based on artificial intelligence. Comprises: obtaining advertisement feature data; performing primary processing on the obtained advertisement feature data to obtain a primary advertisement feature data set; dividing audience groups by using a K-means algorithm based on the preliminary advertisement feature data set; extracting a keyword sequence based on audiences; based on the keyword sequence, utilizing a generative adversarial network model to generate an advertisement copywriting sequence; based on the advertisement copywriting sequence, similarity and fluency are calculated, a comprehensive score is obtained through weighted average, and the advertisement copywriting with the highest comprehensive score is selected as final output; according to the method, the correlation and attraction of the advertisement copywriting can be effectively improved, the advertisement putting effect is optimized, the click rate and the conversion rate of the user are finally improved, and higher investment return is brought to an advertiser. And the creativity and the delivery effect of the advertisement are greatly improved.

Description

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an advertising planning method and system based on artificial intelligence. Background Art

[0002] The patent with application publication number CN114820019A discloses an artificial intelligence advertising planning method and system, including: determining the name of the user's product and the user group of the product; evaluating the advertising features expected by consumers of the current product; inputting the planned product features; conducting data analysis on the field to which the planned product belongs and providing advertising strategies; generating slogans for the planned product for consumers to choose and then for users to choose; generating advertising pictures for the planned product for consumers to choose and then for users to choose.

[0003] The patent with application publication number CN116308533A discloses an advertising planning service platform based on artificial intelligence, including a service project assembly, which includes brand design, website construction, and data media. An information feedback is arranged on the right side of the service project assembly, and the information feedback includes telephone feedback, sampling feedback, and network feedback. A data proofreading is arranged on the right side of the information feedback, and an advertisement improvement is arranged on the right side of the data proofreading. Through the settings of information feedback, data proofreading, and advertisement improvement, after the cooperation party A reaches an agreement on brand design, website construction, and data media to produce advertisements in cooperation, the staff of the feedback department shall collect feedback data on the population covered by the advertisement by telephone feedback, sampling feedback, and network feedback, and proofread the data by collecting information and information collected before the advertisement is not covered, and improve the existing advertisement according to the collected information and the proofreading information.

[0004] However, existing advertising planning is often based on rough market segmentation and is unable to accurately divide the needs and preferences of different audiences, resulting in the inability of advertising to effectively reach the core user group, causing a disconnect between advertising content and user interests and behaviors; although audience segmentation is intended to improve the accuracy of advertising, the generated advertising copy is only based on a keyword sequence; after most advertising copy is generated, it is necessary to rely on manual or rough tools for effect evaluation. For example, the calculation of similarity and fluency is not perfect, which makes it difficult for advertising planners to objectively measure their effects and perform real-time optimization.

[0005] In view of this, the present invention proposes an advertising planning method and system based on artificial intelligence to solve the above problems. Summary of the invention

[0006] 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 solutions: an advertising planning method based on artificial intelligence, comprising: S1, obtaining advertising feature data;

[0007] S2. Preliminarily process the acquired advertisement feature data to obtain a preliminary advertisement feature data set;

[0008] S3. Based on the preliminary advertising feature data set, use the K-means algorithm to divide the audience groups; extract keyword sequences based on the audience groups;

[0009] S4. Generate advertising copy sequence based on keyword sequence using generative adversarial network model;

[0010] S5. Based on the sequence of advertising copy, calculate the similarity and fluency, obtain the comprehensive score through weighted average method, and select the advertising copy with the highest comprehensive score as the final output.

[0011] Furthermore, the advertisement characteristic data includes user behavior data and product characteristic data;

[0012] User behavior data includes user click behavior data, dwell time, and interaction frequency;

[0013] Product feature data includes: product price, ratings and sales data;

[0014] The advertising characteristic data includes user behavior data and product characteristic data;

[0015] User behavior data includes user click behavior data, dwell time, and interaction frequency;

[0016] Product feature data includes: product price, rating and sales data.

[0017] Furthermore, the method of acquiring the click behavior data includes:

[0018] Collect the user ID, number of clicks, click time, and click page data at the current time t;

[0019] Use one-hot encoding to convert user ID, click time, and click page into numerical data, and perform outlier, missing value, and normalization processing on the converted user ID, click time, click page, and click count data to obtain a preliminary feature data set;

[0020] The preliminary feature data set is input into the improved decision tree model to predict the click behavior data at the current time t.

[0021] Furthermore, the preliminary feature data set is input into the improved decision tree model to predict the click behavior data at the current time t, including:

[0022] In the attention mechanism, the input feature vector X is set to [x ID ,x cs ,x sj,x pl ], where x ID Indicates user ID, x cs Indicates the number of clicks, x sj Indicates click time, x pl Indicates a click on a page;

[0023] In the attention layer, the feature weights are calculated, and the input feature vector X is mapped through the weight matrix to obtain the query vector Q = W Q *X, and key matrix S = W S *X, W Q and W S The weight matrices for generating query vectors and building matrices are respectively used. The attention scores of query vectors and key matrices are calculated by dot product: score(Q,S)=Q*S T , and scale the attention score, and use the Softmax function on the scaled attention score to convert it into weight a v ;

[0024] Among them, the weight proof W is a learnable parameter, which linearly changes the input features to generate query vectors and build matrices;

[0025] The formula for scaling the attention score is: where d S is the dimension of the key vector feature; the formula for the weight is: where a v represents the attention weight of the current v-th feature, represents the scaling score of the current vth feature, and q represents the indexes of the four features: user ID, number of clicks, click time, and click page;

[0026] The feature weight a output by the attention layer v With the original feature x v Perform weighted summation to obtain weighted features. The calculation formula is: Among them, x ’ represents the weighted features;

[0027] According to the weighted feature x ’ , calculate the information gain of each feature, and select the feature with the largest information gain for splitting;

[0028] Repeat the above steps, recursively construct subnodes on the decision tree, introduce an attention layer in each subnode, calculate weighted features and split nodes according to information gain, stop recursive construction when the node sample is less than the set threshold, and output click behavior data.

[0029] Furthermore, the method of obtaining the preliminary advertisement feature data set includes:

[0030] The outliers and missing values ​​in the advertisement feature data are removed, and the advertisement feature data is normalized to obtain a preliminary advertisement feature data set.

[0031] Furthermore, the method of dividing the audience groups by using the K-means algorithm includes:

[0032] Initialization, randomly select a preliminary K value, calculate the clustering error sum of squares SSE for each K value, plot the K value and the corresponding clustering error sum of squares SSE to form an SSE-cluster number curve, find the elbow with a significant drop on the curve as a potential candidate K value, further determine the optimal K value, calculate the silhouette coefficient on the curve elbow and n1 candidate K values ​​around the elbow to measure the balance between intra-cluster similarity and inter-cluster separation, and select the K value with the highest silhouette coefficient as the optimal cluster number, and then randomly select K data points as the initial cluster centers, marked as C 1 , C 2 ,…,C K ;

[0033] Assign, for each data point X i =[d1 i ,d2 i ,d3 i ,d4 i ,d5 i ,d6 i ] calculates its distance from each initial cluster center and assigns the data point to the cluster center closest to it to form K clusters, where X i represents the i-th data point, d1 i , d2 i , d3 i , d4 i , d5 i and d6 i They are the click behavior, dwell time, interaction frequency, product price, rating and sales data of the i-th data point respectively. The distance calculation method uses the Euclidean distance, and the formula is: Among them, D(X i ,C K ) represents the data point X i to C K distance, 6 is the number of features, d ij For data point X i The jth feature of Kj is the cluster center C K The jth feature of

[0034] Update: According to the current K cluster allocation, calculate the mean of all data points in each cluster, and use the mean of all data points as the updated cluster center;

[0035] Iterate, repeat the assignment and update steps, and when the cluster center no longer changes, the algorithm converges and stops iterating.

[0036] Furthermore, the method of extracting the keyword sequence includes:

[0037] Each cluster represents an audience group. According to the updated cluster center in each cluster, the main features of each audience group are identified, and all feature values ​​in the updated cluster center are sorted from high to low. The change rate between two adjacent features is calculated. When the change rate is less than the set threshold, the selection is stopped. The main features of each audience group are determined and used as the keyword sequence. Y = [Y 1 , Y 2 , …, Y y ], where y represents the number of keywords.

[0038] Furthermore, the method of using the generative adversarial network model to generate advertising copy includes:

[0039] Construct a generative adversarial network and randomly initialize the discriminator parameters θ D and the parameters θ of the generator G ;

[0040] Perform discriminator training and collect real advertisements from real advertisement copy data sets for A real , through the extracted keyword sequence Y, the generator generates the advertising copy G(Y) and calculates the loss of the discriminator Among them, D(A real ) represents the discriminator's judgment on the real advertisement, D(G(Y) represents the discriminator's judgment on the generated advertisement, represents the expected value of the real advertisement, E Y represents the expected value of the keyword sequence, 1-D(G(Y) represents the probability that the generated advertisement is judged as false; the discriminator parameters are updated by the gradient descent method, Where η is the learning rate, Denotes the loss L of the discriminator D For the discriminator parameters θ D The gradient of , that is, the partial derivative of the loss function with respect to the discriminator parameters;

[0041] Train the generator and calculate the generator loss L G =-E Y [logD(G(Y))], where the generator parameters are updated by gradient descent,

[0042] After the set training cycle, the generator outputs the generated advertising copy and perturbs the keyword sequence using the random exchange method to obtain r keyword sequences Y r , and input into the discriminator and generator, and output the generated advertising copy sequence B = [B 1 , B 2 , …, B r ], where r is the number of generated advertising copies.

[0043] Furthermore, the similarity and fluency are calculated based on the sequence of advertisement texts, and a comprehensive score is obtained by weighted average, and the advertisement text with the highest comprehensive score is selected as the final output. The specific method includes:

[0044] Use TF-IDF technology to convert the generated advertising copy and keyword sequence into vector form and calculate the cosine similarity Among them, B b The vector representing the b-th advertising copy, b = 1, 2, ..., r, Y b Represents the vector of the b-th keyword, cosine similarity cs(B b ,Y b ) ranges from -1 to 1, where 1 means complete similarity, 0 means no similarity, and -1 means complete opposite;

[0045] Use N-gram technology in the advertising copy to obtain the probability P of each word in u1 , use perplexity to evaluate the fluency of advertising copy, the formula is: Among them, u1 represents the index of the number of words, u represents the total number of words, and Perplexity (B b ) represents the fluency value of generating advertising copy;

[0046] The cosine similarity value and fluency value are calculated using the weighted average method to obtain a comprehensive score, which is then sorted from high to low, and the advertising copy with the highest comprehensive score is selected as the final advertisement.

[0047] The technical effects and advantages of the artificial intelligence-based advertising planning method and system of the present invention are as follows:

[0048] The present invention can realize intelligent advertising planning, and combines multiple advanced technologies to realize efficient and accurate advertising copy generation and optimization. First, by acquiring user behavior data and product feature data, the system can deeply understand audience needs and market dynamics, thereby providing strong data support for subsequent advertising copy generation. Then, the application of the K-means algorithm effectively divides the audience groups and extracts keyword sequences, making the generated advertising copy more targeted and attractive. In addition, the introduction of the generative adversarial network (GAN) model ensures the quality and creativity of the generated advertising copy, and uses TF-IDF technology and N-gram method to evaluate the similarity and fluency of the copy, so that the final output advertising copy is more logical and natural. This comprehensive technical means not only improves the efficiency of advertising copy generation, but also greatly improves the relevance of the copy and the user's click-through rate, and promotes the optimization of advertising delivery effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic diagram of an advertising planning method based on artificial intelligence of the present invention;

[0050] Figure 2 This is a schematic diagram of an advertising planning system based on artificial intelligence in the present invention. DETAILED DESCRIPTION

[0051] 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.

[0052] Example 1

[0053] See also Figure 1 As shown, the advertising planning method based on artificial intelligence described in this embodiment includes: S1, obtaining advertising feature data;

[0054] S2. Preliminarily process the acquired advertisement feature data to obtain a preliminary advertisement feature data set;

[0055] S3. Based on the preliminary advertising feature data set, use the K-means algorithm to divide the audience groups; extract keyword sequences based on the audience groups;

[0056] S4. Generate advertising copy sequence based on keyword sequence using generative adversarial network model;

[0057] S5. Based on the advertising copy sequence, calculate the similarity and fluency, obtain the comprehensive score through weighted average, and select the advertising copy with the highest comprehensive score as the final output.

[0058] Advertising feature data includes user behavior data and product feature data;

[0059] User behavior data includes user click behavior data, dwell time, and interaction frequency;

[0060] Product feature data includes: product price, rating and sales data.

[0061] Dwell time refers to the length of time a user stays on a page or application. Front-end technologies such as JavaScript are used to monitor the start time and exit time of a user's page visit. For example, the timestamp when the user enters the page and the timestamp when the user leaves the page are recorded, and the dwell time is calculated by calculating the difference between the two. This data can be transmitted to the backend through the website or application's log system.

[0062] Interaction frequency usually refers to the number of times a user interacts with a specific element on a page, such as a button, form, or advertisement. Front-end technologies such as JavaScript are used to monitor user clicks, slides, and hovers on page elements, record these interaction events, and aggregate them into user interaction frequency. Interaction data can also be captured in real time through application APIs or third-party analysis tools such as Google Analytics and Mixpanel.

[0063] Product prices usually come from the merchant's product database or the backend system of the e-commerce platform. The data is directly retrieved and recorded through the API or database. For e-commerce advertising planning systems, these price data are usually obtained by directly connecting to the platform's product management system.

[0064] Product rating data usually comes from the user's comment system or evaluation module. The rating system collects user ratings through the rating form on the front-end page. The back-end system summarizes each user's rating of the product and stores it in the database. This data can also be captured through the API of third-party evaluation platforms such as Amazon, Taobao and other e-commerce platforms;

[0065] Sales data is usually collected directly from the merchant's sales system. By connecting with the merchant's order management system or ERP system, real-time sales quantities can be obtained. E-commerce platforms can regularly synchronize and update sales information through API interfaces.

[0066] The ways to obtain click behavior data include:

[0067] Collect the user ID, number of clicks, click time, and click page data at the current time t;

[0068] Use one-hot encoding to convert user ID, click time, and click page into numerical data, and perform outlier, missing value, and normalization processing on the converted user ID, click time, click page, and click count data to obtain a preliminary feature data set;

[0069] Input the preliminary feature data set into the improved decision tree model to predict the click behavior data at the current time t;

[0070] The user ID is collected through a unique identifier assigned when the user logs in or registers, such as an account ID or device ID, which can be automatically generated and recorded by the backend system of the website or application;

[0071] The number of clicks is monitored by the front-end page of the website or application. Usually, front-end technologies such as JavaScript monitor the user's interaction with the page and record the number of clicks in the back-end system.

[0072] Click time By recording the specific timestamp of each user click, front-end technologies such as JavaScript can obtain the exact time of the user's click event and store it in the database, usually in the format of UTC time or local time;

[0073] Click page data records the specific page URL or page ID visited when the user clicks. The website or application will use front-end technology to capture the URL, page name, element ID and other data of the current page, and transmit it to the back-end database for storage and analysis.

[0074] Input the preliminary feature data set into the improved decision tree model to predict the click behavior data at the current time t, including:

[0075] In the attention mechanism, the input feature vector X is set to [x ID ,x cs ,x sj ,x pl ], where x ID Indicates user ID, x cs Indicates the number of clicks, x sj Indicates click time, x pl Indicates a click on a page;

[0076] In the attention layer, the feature weights are calculated, and the input feature vector X is mapped through the weight matrix to obtain the query vector Q = W Q *X, and key matrix S = W S *X, W Q and W SThe weight matrices for generating query vectors and building matrices are respectively used. The attention scores of query vectors and key matrices are calculated by dot product: score(Q,S)=Q*S T , and scale the attention score, and use the Softmax function on the scaled attention score to convert it into weight a v ;

[0077] Among them, the weight proof W is a learnable parameter, which linearly changes the input features to generate query vectors and build matrices;

[0078] The formula for scaling the attention score is: where d S is the dimension of the key vector feature; the formula for the weight is: where a v represents the attention weight of the current v-th feature, represents the scaling score of the current vth feature, and q represents the indexes of the four features: user ID, number of clicks, click time, and click page;

[0079] The feature weight a output by the attention layer v With the original feature x v Perform weighted summation to obtain weighted features. The calculation formula is: Among them, x ’ represents the weighted features;

[0080] According to the weighted feature x ’ , calculate the information gain of each feature, and select the feature with the largest information gain for splitting;

[0081] Repeat the above steps, recursively construct subnodes on the decision tree, introduce an attention layer in each subnode, calculate weighted features and split nodes according to information gain, stop recursive construction when the node sample is less than the set threshold, and output click behavior data.

[0082] Among them, the dot product calculation is also called the inner product, which is used to measure the similarity or correlation between the query vector Q and the key matrix S;

[0083] The methods for obtaining the preliminary advertising feature dataset include:

[0084] Remove outliers in the advertising feature data using the Z-Score method; handle missing values ​​using the mean interpolation method; then normalize the advertising feature data using Min-Max normalization to ensure that each feature is within the same magnitude range, and obtain a preliminary advertising feature data set.

[0085] The ways to segment audiences using the K-means algorithm include:

[0086] Initialization, randomly select a preliminary K value, calculate the clustering error sum of squares SSE for each K value, plot the K value and the corresponding clustering error sum of squares SSE to form an SSE-cluster number curve, find the elbow with a significant drop on the curve as a potential candidate K value, further determine the optimal K value, calculate the silhouette coefficient on the curve elbow and n1 candidate K values ​​around the elbow to measure the balance between intra-cluster similarity and inter-cluster separation, and select the K value with the highest silhouette coefficient as the optimal cluster number, and then randomly select K data points as the initial cluster centers, marked as C 1 , C 2 ,…,C K ;

[0087] Assign, for each data point X i =[d1 i ,d2 i ,d3 i ,d4 i ,d5 i ,d6 i ] calculates its distance from each initial cluster center and assigns the data point to the cluster center closest to it to form K clusters, where X i represents the i-th data point, d1 i , d2 i , d3 i , d4 i , d5 i and d6 i They are the click behavior, dwell time, interaction frequency, product price, rating and sales data of the i-th data point respectively. The distance calculation method uses the Euclidean distance, and the formula is: Among them, D(X i ,C K ) represents the data point X i to C K distance, 6 is the number of features, d ij For data point X i The jth feature of Kj is the cluster center C K The jth feature of

[0088] Update: According to the current K cluster allocation, calculate the mean of all data points in each cluster, and use the mean of all data points as the updated cluster center;

[0089] Iterate, repeat the assignment and update steps, and when the cluster center no longer changes, the algorithm converges and stops iterating.

[0090] Methods for extracting keyword sequences include:

[0091] Each cluster represents an audience group. According to the updated cluster center in each cluster, the main features of each audience group are identified, and all feature values ​​in the updated cluster center are sorted from high to low. The change rate between two adjacent features is calculated. When the change rate is less than the set threshold, the selection is stopped. The main features of each audience group are determined and used as the keyword sequence. Y = [Y 1 , Y 2 , …, Y y ], where y represents the number of keywords;

[0092] For example, the updated clustering of a certain cluster is [click behavior = 0.7, dwell time = 0.6, interaction frequency = 0.8, product price = 0.5, rating = 0.9, sales data = 0.45], and the order is rating > interaction frequency > click behavior > dwell time > product price > sales data. The change rate between two adjacent features is calculated, and the change rate is 18.99%, 12.85%, 16.67%, 20% and 11.11%. When the threshold of the change rate is set to 10%, the selection stops, and the main features of the audience group are rating and interaction frequency.

[0093] Ways to use the generative adversarial network model to generate advertising copy include:

[0094] Construct a generative adversarial network and randomly initialize the discriminator parameters θ D and the parameters θ of the generator G ;

[0095] Perform discriminator training and collect real advertisements from real advertisement copy data sets for A real , through the extracted keyword sequence Y, the generator generates the advertising copy G(Y) and calculates the loss of the discriminator Among them, D(A real ) represents the discriminator's judgment on the real advertisement, D(G(Y) represents the discriminator's judgment on the generated advertisement, represents the expected value of the real advertisement, E Y represents the expected value of the keyword sequence, 1-D(G(Y) represents the probability that the generated advertisement is judged as false; the discriminator parameters are updated by the gradient descent method, Where η is the learning rate, Denotes the loss L of the discriminator D For the discriminator parameters θ D The gradient of , that is, the partial derivative of the loss function with respect to the discriminator parameters;

[0096] Train the generator and calculate the generator loss L G =-E Y [logD(G(Y))], where the generator parameters are updated by gradient descent,

[0097] After the set training cycle, the generator outputs the generated advertising copy and perturbs the keyword sequence using the random exchange method to obtain r keyword sequences Y r , and input into the discriminator and generator, and output the generated advertising copy sequence B = [B 1 , B 2 , …, B r ], where r is the number of generated advertising copy;

[0098] The gradient descent method calculates the gradient of the loss function with respect to the parameters and updates the parameters in the opposite direction of the gradient to minimize the loss function. First, the partial derivatives of the loss function with respect to each parameter are calculated. The vector composed of these partial derivatives is the gradient. Then, according to the learning rate and the gradient, the value of each parameter is updated to move the parameter in the direction where the loss function decreases. The above steps are repeated until the loss function converges to the minimum value or the preset number of iterations is reached.

[0099] Based on the sequence of advertisement texts, similarity and fluency are calculated, and a comprehensive score is obtained by weighted average. The specific method of selecting the advertisement text with the highest comprehensive score as the final output includes:

[0100] Use TF-IDF technology to convert the generated advertising copy and keyword sequence into vector form and calculate the cosine similarity Among them, B b The vector representing the b-th advertising copy, b = 1, 2, ..., r, Y b Represents the vector of the b-th keyword, cosine similarity cs(B b ,Y b ) ranges from -1 to 1, where 1 means complete similarity, 0 means no similarity, and -1 means complete opposite;

[0101] Use N-gram technology in the advertising copy to obtain the probability P of each word in u1 , use perplexity to evaluate the fluency of advertising copy, the formula is: Among them, u1 represents the index of the number of words, u represents the total number of words, and Perplexity (B b ) represents the fluency value of generating advertising copy;

[0102] Use the weighted average method to calculate the cosine similarity value and fluency value to obtain a comprehensive score, sort the comprehensive scores from high to low, and select the advertising copy with the highest comprehensive score as the final advertisement;

[0103] TF-IDF technology, which segments the copy and keywords into words to form a word set, calculates the word frequency TF of each word in the copy and the inverse document frequency IDF in the entire corpus, and multiplies them to get the TF-IDF value, which reflects the importance of the word in the copy. Finally, the TF-IDF values ​​of all words are combined into a vector through programming code. The semantic database is a database containing a large amount of text, which can be an advertising copy database;

[0104] N-gram technology, determine the value of N-gram, N = 1, 2, 3, 1 represents a single word, 2 represents a double word, and 3 represents a triple word. For example, determine N = 2, "an advertising planning method based on artificial intelligence" is an advertising copy, which contains a total of "a kind of, based on", "based on, artificial", "artificial, intelligent", "intelligent,", "advertisement", "advertisement, planning", "planning, method", a total of 7, "a kind of, based on" appears once, then the probability of "a kind of, based on" is 1 / 7, if "a kind of, based on" appears twice, then the probability is 2 / 7;

[0105] The random exchange method is a simple perturbation technique, which is mainly used to introduce randomness and diversity in keyword sequences. First, any two words in the keyword sequence are selected, and then their positions are exchanged to generate a new keyword sequence. This process can be repeated until the required number of different sequences are generated. In this way, different keyword combinations can be explored to help generate more creative advertising copy while maintaining the semantic information of the original keywords. This method is simple and easy to implement, and helps to enhance the generalization ability and robustness of the model.

[0106] The real advertising copy data set can be provided through the historical advertising copy database and the third-party advertising platform.

[0107] This implementation case can realize intelligent advertising planning and combines multiple advanced technologies to achieve efficient and accurate advertising copy generation and optimization. First, by obtaining user behavior data and product feature data, the system can deeply understand audience needs and market dynamics, thereby providing strong data support for subsequent advertising copy generation. Then, the application of the K-means algorithm effectively divides the audience groups and extracts keyword sequences to make the generated advertising copy more targeted and attractive. In addition, the introduction of the generative adversarial network model ensures the quality and creativity of the generated advertising copy, and uses TF-IDF technology and N-gram method to evaluate the similarity and fluency of the copy, making the final output advertising copy more logical and natural. This comprehensive technical means not only improves the efficiency of advertising copy generation, but also greatly improves the relevance of the copy and the user's click-through rate, and promotes the optimization of advertising delivery effects.

[0108] Example 2

[0109] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, and an advertising planning system based on artificial intelligence is provided, including: a data acquisition module: acquiring advertising feature data;

[0110] Data processing module: preliminarily processes the acquired advertisement feature data to obtain a preliminary advertisement feature data set;

[0111] Keyword extraction module: Based on the preliminary advertising feature data set, the K-means algorithm is used to divide the audience groups; keyword sequences are extracted based on the audience groups;

[0112] Advertising creative module: Based on the keyword sequence, the generative adversarial network model is used to generate the advertising copy sequence;

[0113] Creative evaluation module: Based on the advertising copy sequence, calculate the similarity and fluency, obtain the comprehensive score through weighted average, and select the advertising copy with the highest comprehensive score as the final output.

[0114] Example 3

[0115] 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 planning method and system is implemented.

[0116] Since the electronic device introduced in this embodiment is an electronic device used to implement an advertising planning method and system based on artificial intelligence in the embodiment of the present application, based on the advertising planning method and system based on artificial intelligence introduced in the embodiment of the present application, the technical personnel of the field can understand the specific implementation of the electronic device of the present embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present application is not described in detail here. As long as the technical personnel of the field implement the electronic device used by the advertising planning method and system based on artificial intelligence in the embodiment of the present application, it belongs to the scope of protection of this application.

[0117] 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.

[0118] 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 advertising planning method based on artificial intelligence, characterized in that: The advertising planning method based on artificial intelligence includes: S1, obtaining advertising feature data; S2. Preliminarily process the acquired advertisement feature data to obtain a preliminary advertisement feature data set; S3. Based on the preliminary advertising feature data set, use the K-means algorithm to divide the audience groups; extract keyword sequences based on the audience groups; S4. Generate advertising copy sequence based on keyword sequence using generative adversarial network model; S5. Based on the advertising copy sequence, calculate the similarity and fluency, obtain the comprehensive score through weighted average, and select the advertising copy with the highest comprehensive score as the final output.

2. The advertising planning method based on artificial intelligence according to claim 1, characterized in that: The advertising characteristic data includes user behavior data and product characteristic data; User behavior data includes user click behavior data, dwell time, and interaction frequency; Product feature data includes: product price, rating and sales data.

3. The advertising planning method based on artificial intelligence according to claim 2, characterized in that: The methods for obtaining the click behavior data include: Collect the user ID, number of clicks, click time, and click page data at the current time t; Use one-hot encoding to convert user ID, click time, and click page into numerical data, and perform outlier, missing value, and normalization processing on the converted user ID, click time, click page, and click count data to obtain a preliminary feature data set; The preliminary feature data set is input into the improved decision tree model to predict the click behavior data at the current time t.

4. The advertising planning method based on artificial intelligence according to claim 3 is characterized in that: The preliminary feature data set is input into the improved decision tree model to predict the click behavior data at the current time t, including: In the attention mechanism, the input feature vector X is set to [x ID ,x cs ,x sj ,x pl ], where x ID Indicates user ID, x cs Indicates the number of clicks, x sj Indicates click time, x pl Indicates a click on a page; In the attention layer, the feature weights are calculated, and the input feature vector X is mapped through the weight matrix to obtain the query vector Q = W Q *X, and key matrix S = W S *X, W Q and W S The weight matrices for generating query vectors and building matrices are respectively used. The attention scores of query vectors and key matrices are calculated by dot product: score(Q,S)=Q*S T , and scale the attention score, and use the Softmax function on the scaled attention score to convert it into weight a v ; Among them, the weight proof W is a learnable parameter, which linearly changes the input features to generate query vectors and build matrices; The formula for scaling the attention score is: where d S is the dimension of the key vector feature; the formula for the weight is: where a v represents the attention weight of the current v-th feature, represents the scaling score of the current vth feature, and q represents the indexes of the four features: user ID, number of clicks, click time, and click page; The feature weight a output by the attention layer v With the original feature x v Perform weighted summation to obtain weighted features. The calculation formula is: Among them, x ’ represents the weighted features; According to the weighted feature x ’ , calculate the information gain of each feature, and select the feature with the largest information gain for splitting; Repeat the above steps, recursively construct subnodes on the decision tree, introduce an attention layer in each subnode, calculate weighted features and split nodes according to information gain, stop recursive construction when the node sample is less than the set threshold, and output click behavior data.

5. The advertising planning method based on artificial intelligence according to claim 4, characterized in that: The method of obtaining the preliminary advertisement feature data set includes: The outliers and missing values ​​in the advertisement feature data are removed, and the advertisement feature data is normalized to obtain a preliminary advertisement feature data set.

6. The advertising planning method based on artificial intelligence according to claim 5, characterized in that: The method of using the K-means algorithm to divide the audience groups includes: Initialization, randomly select a preliminary K value, calculate the clustering error sum of squares SSE for each K value, plot the K value and the corresponding clustering error sum of squares SSE to form an SSE-cluster number curve, find the elbow with a significant drop on the curve as a potential candidate K value, further determine the optimal K value, calculate its silhouette coefficient on the curve elbow and n1 candidate K values ​​around the elbow, and select the K value with the highest silhouette coefficient as the optimal cluster number, and then randomly select K data points as the initial cluster centers, marked as C1, C2, ..., C K ; Assign, for each data point X i =[d1 i ,d2 i ,d3 i ,d4 i ,d5 i ,d6 i ] calculates its distance from each initial cluster center and assigns the data point to the cluster center closest to it to form K clusters, where X i represents the i-th data point, d1 i , d2 i , d3 i , d4 i , d5 i and d6 i They are the click behavior, dwell time, interaction frequency, product price, rating and sales data of the i-th data point respectively. The distance calculation method uses the Euclidean distance, and the formula is: Among them, D(X i ,C K ) represents the data point X i to C K distance, 6 is the number of features, d ij For data point X i The jth feature of Kj is the cluster center C K The jth feature of Update: According to the current K cluster allocation, calculate the mean of all data points in each cluster, and use the mean of all data points as the updated cluster center; Iterate, repeat the assignment and update steps, and when the cluster center no longer changes, the algorithm converges and stops iterating.

7. The advertising planning method based on artificial intelligence according to claim 6, characterized in that: The method of extracting the keyword sequence includes: Each cluster represents an audience group. According to the updated cluster center in each cluster, the main features of each audience group are identified, and all feature values ​​in the updated cluster center are sorted from high to low. The change rate between two adjacent features is calculated. When the change rate is less than the set threshold, the selection is stopped. The main features of each audience group are determined and used as the keyword sequence, Y = [Y1, Y2, ..., Y y ], where y represents the number of keywords.

8. The advertising planning method based on artificial intelligence according to claim 7, characterized in that: The method of using the generative adversarial network model to generate advertising copy includes: Construct a generative adversarial network and randomly initialize the discriminator parameters θ D and the parameters θ of the generator G ; Perform discriminator training and collect real advertisements from real advertisement copy data sets for A real , through the extracted keyword sequence Y, the generator generates the advertising copy G(Y) and calculates the loss of the discriminator Among them, D(A real ) represents the discriminator's judgment on the real advertisement, D(G(Y) represents the discriminator's judgment on the generated advertisement, represents the expected value of the real advertisement, E Y represents the expected value of the keyword sequence, 1-D(G(Y) represents the probability that the generated advertisement is judged as false; the discriminator parameters are updated by the gradient descent method, Where η is the learning rate, Denotes the loss L of the discriminator D For the discriminator parameters θ D The gradient of , that is, the partial derivative of the loss function with respect to the discriminator parameters; Train the generator and calculate the generator loss L G =-E Y [logD(G(Y))], where the generator parameters are updated by gradient descent, After the set training cycle, the generator outputs the generated advertising copy and perturbs the keyword sequence using the random exchange method to obtain r keyword sequences Y r , and input it into the discriminator and generator, and output the generated advertising copy sequence B = [B1, B2, ..., B r ], where r is the number of generated advertising copies.

9. The advertising planning method based on artificial intelligence according to claim 8, characterized in that: The specific method of calculating similarity and fluency based on the advertisement copy sequence, obtaining a comprehensive score by weighted average, and selecting the advertisement copy with the highest comprehensive score as the final output includes: Use TF-IDF technology to convert the generated advertising copy and keyword sequence into vector form and calculate the cosine similarity Among them, B b The vector representing the b-th advertisement copy, b = 1, 2, ..., r, Y b The vector representing the bth keyword; Use N-gram technology in the advertising copy to obtain the probability P of each word in u1 , use perplexity to evaluate the fluency of advertising copy, the formula is: Among them, u1 represents the index of the number of words, u represents the total number of words, and Perplexity (B b ) represents the fluency value of generating advertising copy; The cosine similarity value and fluency value are calculated using the weighted average method to obtain a comprehensive score, which is then sorted from high to low, and the advertising copy with the highest comprehensive score is selected as the final advertisement.

10. An artificial intelligence-based advertising planning system, used to implement the artificial intelligence-based advertising planning method described in any one of claims 1 to 9, characterized in that: include: Data acquisition module: obtains advertisement feature data; Data processing module: preliminarily processes the acquired advertisement feature data to obtain a preliminary advertisement feature data set; Keyword extraction module: Based on the preliminary advertising feature data set, the K-means algorithm is used to divide the audience groups; keyword sequences are extracted based on the audience groups; Advertising creative module: Based on the keyword sequence, the generative adversarial network model is used to generate the advertising copy sequence; Creative evaluation module: Based on the advertising copy sequence, calculate the similarity and fluency, obtain the comprehensive score through weighted average, and select the advertising copy with the highest comprehensive score as the final output.

Images