Intelligent Advertising Delivery System and Method Based on User Behavior Portrait

Abstract

The present invention relates to the technical field of advertising delivery. Specifically, it relates to an intelligent advertising delivery system and method based on user behavior portraits, including a user portrait construction module, an advertisement weight allocation module, a user portrait adjustment module, and an advertising delivery optimization module. In this intelligent advertising delivery system based on user behavior portraits, the user portrait construction module divides browsing data into current browsing data and historical browsing data, constructs a user portrait based on the commodity prices preferred by the user in the historical browsing data and excluding the commodity prices that do not meet the preferences when constructing the current user portrait. The advertisement weight allocation module assigns different advertisement delivery weights according to the number of times of browsing commodities in the current browsing data for advertising delivery, so that the advertisement can accurately reach those users who are truly interested in relevant commodities and meet their price expectations, reduce the number of advertisements pushed to users that they are not interested in or do not meet their consumption price levels, and avoid waste of advertising resources.

Description

Technical Field

[0001] The present invention relates to the technical field of advertising delivery, and more specifically, to an intelligent advertising delivery system and method based on user behavior portraits. Background Art

[0002] Advertising has always been an effective means for merchants to promote their brands. With the continuous popularization of the Internet, online advertising has gradually replaced traditional delivery methods such as print, television, and radio, and has become one of the most effective advertising delivery methods currently;

[0003] E-commerce platforms face challenges in the advertising delivery process. Because they have a vast amount of data resources, user operations such as browsing product detail pages, adding items to the shopping cart, completing purchases, and evaluating products will all leave detailed data traces. However, due to the storage of various browsing records on the platform, the data is intertwined and mixed, and the diverse browsing behavior data of users converges to form a complex data pool;

[0004] When advertising is delivered to users based on such a large and mixed browsing data, problems arise. If it is difficult to extract the true shopping intentions and core demand tendencies of the current user from the browsing data, the advertisements delivered at this time often do not match the products that the user actually wants to buy at the moment, ultimately resulting in a significant reduction in the accuracy of advertising delivery. In view of this, we propose an intelligent advertising delivery system and method based on user behavior portraits. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem of not being able to accurately provide users with advertisements that meet their current shopping needs.

[0006] To achieve the above purpose, the present invention provides an intelligent advertising delivery system based on user behavior portraits, including a user portrait construction module, an advertisement weight distribution module, a user portrait adjustment module, and an advertising delivery optimization module;

[0007] The user portrait construction module is used to divide user browsing data into current browsing data and historical browsing data, select the products browsed in the current browsing data and the price range of products preferred by the user in the historical browsing data, and eliminate the products that do not meet the preferred product price range in the user-browsed products. The pictures and prices of the user-browsed products are the user portraits;

[0008] The advertisement weight distribution module is used to calculate the first cosine similarity between the user portrait and the audience characteristics in the advertisement library, set a cosine similarity threshold, screen out the advertisements within the cosine similarity threshold, assign different weights according to the number of times the current user browses products, and deliver the screened advertisements to the user according to different weights. The audience characteristics in the advertisement library refer to the product pictures in the above user portrait construction module;

[0009] The user portrait adjustment module is used to sense the price and number of times of the products clicked by users in the placed advertisements, record the price of the products in the clicked advertisements > the price within the price range of the products preferred by users in the user portrait construction module, and when the number of times of the products in the clicked advertisements > the normal number of times of clicking products, calculate the average value of the prices of the products in the clicked advertisements. If the average value > the price range of the products preferred by users, select the adjustment coefficient from small to large, increase the upper limit of the price range of the products preferred by users, and the adjustment method for the lower limit of the price range of the products preferred by users is the same;

[0010] The advertisement placement optimization module is used to sense the product pictures clicked by users in the placed advertisements, set an adjustment threshold, calculate the second cosine similarity between the clicked product pictures and the product pictures in the historical browsing data through the advertisement weight distribution module. If the second cosine similarity > the adjustment threshold, determine that the clicked product is the same as the product pictures in the historical browsing data, transfer the number of times of browsing the product in the historical browsing data to the current number of times of browsing by the current user in the advertisement weight distribution module, and increase the weight of the corresponding product in the advertisement weight distribution module.

[0011] As a further improvement of this technical solution, the user portrait construction module senses the purchase behavior of users, uses the time node when the purchase behavior occurs as the division node, the browsing data corresponding after the division node is the current browsing data, and the browsing data corresponding before the division node is the historical browsing data;

[0012] As the user subsequently has a purchase behavior again, the time node of the new purchase behavior replaces the previous time node of the last purchase behavior and becomes the new division node.

[0013] As a further improvement of this technical solution, the price range of the products preferred by users in the user portrait construction module is calculated by the quantile method, retrieve the corresponding price for each product purchased in the historical browsing data, sort the prices of the purchased products in a descending order, divide the set of product price data into numerical points with five equal frequency parts, select the second and fourth numerical points from large to small, calculate the product prices corresponding to the second and fourth numerical points, and the amount between the product prices is the price range of the products preferred by users.

[0014] As a further improvement of this technical solution, the advertisement weight distribution module and the advertisement audience feature vector , , calculate the first cosine similarity between the two feature vectors according to the dimension of the vector;

[0015] The first cosine similarity ;

[0016] where and They are vectors respectively and among the elements indicating the th position in the vector

[0017] As a further improvement of this technical solution, the advertisement weight distribution module sets the cosine similarity threshold based on industry experience. Different industries have accumulated certain empirical data in the long-term practice of advertisement placement

[0018] As a further improvement of this technical solution, the advertisement weight distribution module allocates weights according to the number of times the current user browses the products. Since different browsing times reflect the user's interest in different products, advertisements related to products with higher user interest are assigned higher weights. The specific calculation formula for different weights of specific allocations is as follows

[0019] ;

[0020] where is the number of times the user browses the product represents the product browse times is used to distinguish different products is an index variable, used to traverse all products in the summation formula

[0021] As a further improvement of this technical solution, the user profile adjustment module senses the price of the product in the advertisement clicked by the user , where is the number of clicks. The formula for calculating the average price of the product in the clicked advertisement is , where represents the price of the product for the th click

[0022] Sense the price range of the products preferred by the user in the user profile construction module , where is the lower limit is the upper limit, and the adjustment coefficient is , and , is selected at equal intervals when selecting, and the upper limit of the increased price range is .

[0023] As a further improvement of this technical solution, the advertisement placement optimization module sets an adjustment threshold based on a large number of known identical product pictures. Specifically, through a deep learning model: the deep learning model divides the preprocessed picture data into a training set, a validation set, and a test set. During the training process, the identical pictures in the training set are input into the model. The deep learning model extracts the features of the pictures through the convolutional layer, gradually compresses the feature dimensions through the pooling layer, and outputs the prediction results for the identical pictures through the fully connected layer;

[0024] After the deep learning model is trained, the identical picture pairs in the validation set are input into the deep learning model, and the similarity scores output by the deep learning model are obtained. The mean value of the similarity scores is calculated as the adjustment threshold.

[0025] As a further improvement of this technical solution, before calculating the second cosine similarity of the product pictures, the advertisement placement optimization module uses picture recognition technology to extract the feature vectors in the product pictures;

[0026] The picture recognition technology maps the pixel values of the product pictures to the interval [0,1], adjusts the product pictures to the same size, and extracts the feature vectors through a convolutional neural network. The convolutional layer slides the convolutional kernel on the product pictures to perform convolutional operations on the local areas of the product pictures, and applies an activation function to introduce non-linear factors. The role of the activation function is to enable the neural network to learn complex non-linear relationships. The pooling layer performs downsampling on the feature map to reduce the amount of data while retaining the main feature information;

[0027] Through the combination of multiple convolutional layers, activation functions, and pooling layers, various features of the product pictures from low-level to high-level are extracted.

[0028] The intelligent advertisement placement method based on the user behavior portrait includes the following steps:

[0029] Step 1: Divide the current browsing data and historical browsing data to establish a user portrait;

[0030] Step 2: Place advertisements that match the user portrait, and allocate different advertisement placement weights according to the number of times the products are browsed in the user portrait to place advertisements;

[0031] Step 3: Adjust the user portrait according to the price of the product in the advertisement clicked by the user;

[0032] Step 4: Determine whether the product picture in the advertisement clicked by the user is the same as the product pictures in the historical browsing data. If they are the same, adjust the advertisement placement weight.

[0033] Compared with the prior art, the beneficial effects of the present invention:

[0034] In the intelligent advertising delivery system and method based on user behavior portraits, the user portrait construction module divides browsing data into current browsing data and historical browsing data, constructs a user portrait based on the commodity prices preferred by the user in the historical browsing data and excluding the commodity prices that do not conform to the preference when constructing the current user portrait. The advertising weight allocation module assigns different advertising delivery weights according to the number of times of browsing commodities in the current browsing data for advertising delivery. After the advertisement is delivered, the user portrait adjustment module adjusts the commodity prices preferred by the user in the user portrait construction module according to the price and number of times of the commodity corresponding to the user's click on the advertisement. The advertising delivery optimization module analyzes whether the picture of the commodity corresponding to the user's click on the advertisement is the same as the commodity pictures in the historical browsing data. If they are the same, the weights assigned in the advertising weight allocation module are adjusted, so that the advertisement can accurately reach those users who are really interested in the relevant commodities and meet their price expectations, reduce the number of advertisements pushed to users that they are not interested in or do not conform to their consumption price levels, and avoid the waste of advertising resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is the overall module schematic diagram of the present invention;

[0036] Figure 2 is the working step diagram of the present invention.

[0037] The meanings of the various reference numerals in the figure are as follows:

[0038] 100, user portrait construction module; 200, advertising weight allocation module; 300, user portrait adjustment module; 400, advertising delivery optimization module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0040] Such as Figure 1 - Figure 2 , the intelligent advertising delivery system based on user behavior portraits includes a user portrait construction module 100, an advertising weight allocation module 200, a user portrait adjustment module 300, and an advertising delivery optimization module 400;

[0041] In order to screen out the browsing data that conforms to the current user from the user browsing data, the user portrait construction module 100 is used to divide the user browsing data into current browsing data and historical browsing data, select the products browsed in the current browsing data and the price range of products that the user is inclined to in the historical browsing data, and eliminate the products that do not conform to the inclined product price range in the user-browsed products. At this time, the pictures and prices of the user-browsed products are the user portrait;

[0042] Specifically, the user portrait construction module 100 senses the user's purchase behavior. The time node when the user's purchase behavior occurs is used as the division node. The browsing data corresponding after the division node is the current browsing data, and the browsing data corresponding before the division node is the historical browsing data;

[0043] As the user's subsequent purchase behavior occurs again, the time node of the new purchase behavior replaces the time node of the previous purchase behavior and becomes the new division node.

[0044] Furthermore, the price range of products that the user is inclined to in the user portrait construction module 100 is calculated by the quantile method. The price corresponding to each product purchased in the historical browsing data is retrieved, and the prices of the purchased products are sorted in descending order. The price data set of the purchased products is divided into numerical points of five equal-frequency parts. The second and fourth numerical points are selected from largest to smallest, and the prices of the products corresponding to the second and fourth numerical points are calculated. The amount between the product prices is the price range of products that the user is inclined to;

[0045] For example, the price data set of the purchased products is divided into numerical points of five equal-frequency parts, and the purchase product prices corresponding to 25%-75% are respectively selected. The purchase product price range corresponding to 25%-75% is the price range of products that the user is inclined to.

[0046] In order to accurately deliver advertisements to users through the current user portrait, the advertisement weight distribution module 200 is used to calculate the first cosine similarity between the user portrait and the audience characteristics in the advertisement library, set a cosine similarity threshold, screen out the advertisements within the cosine similarity threshold, assign different weights according to the number of times the current user browses products, and deliver the screened advertisements to the user according to different weights. The audience characteristics in the advertisement library refer to the product pictures in the above-mentioned user portrait construction module 100;

[0047] Specifically, the advertisement weight distribution module 200 is based on the feature vector of the user portrait and the advertisement audience feature vector , The dimension of the feature vector, which is the number of different feature dimensions included in the user profile or the characteristics of the advertising audience, calculates the first cosine similarity between two feature vectors. The value of the first cosine similarity ranges from -1 to 1. The closer the value is to 1, the more similar the user profile is to the characteristics of the advertising audience.

[0048] The first cosine similarity ;

[0049] where and are the elements between the vectors and respectively, and represents the th position in the vector.

[0050] Furthermore, the advertising weight allocation module 200 sets the cosine similarity threshold based on industry experience. Different industries have accumulated certain empirical data in the long-term practice of advertising placement. In the fashion industry, for the advertising placement of clothing brands, if you want to accurately reach the target users, the cosine similarity threshold will be set between ;

[0051] Because the consumption of fashion products is closely related to factors such as the age, gender, and style preferences of users, a higher matching degree is required to arouse the interest of users. In some daily necessities industries, due to the strong versatility of products, the cosine similarity threshold is set relatively low, between . Therefore, the cosine similarity threshold can be set by referring to the advertising placement experience of similar products in the industry.

[0052] The advertising weight allocation module 200 allocates weights according to the number of times the current user browses products. Since different browsing times reflect the degree of interest of users in different products, higher weights are assigned to the advertisements related to the products that users are highly interested in, and there will be more opportunities to display the advertisements to users. The specific calculation formula for the different weights of the specific allocation is:

[0053] ;

[0054] where is the number of times the user browses products, represents the number of times of browsing the product , is used to distinguish different products, is an index variable used to traverse all products in the summation formula;

[0055] For example, if a user frequently browses a certain brand of mobile phone, it indicates that they have a strong interest in that mobile phone; while for other electronic products that are browsed only occasionally, the user's interest is relatively low. By assigning weights according to the number of views, this interest difference can be more accurately reflected in advertising placement. Higher weights are assigned to advertisements related to products that the user is highly interested in, giving these advertisements more opportunities to be shown to the user and increasing the probability of the user coming into contact with the advertisements they are interested in.

[0056] The present invention takes into account the situation where the price of the products that the current user favors may increase or decrease with the user's income. At this time, the price range of the products that the user favors set in the user portrait construction module 100 will be inaccurate. Therefore, the user portrait adjustment module 300 senses the price and number of times of the products clicked by the user in the advertisements being placed, and records that when the price of the product in the clicked advertisement > the price of the price range of the products that the user favors in the user portrait construction module 100, and the number of times of the product in the clicked advertisement > the normal number of times of clicking on products, calculate the average value of the price of the product in the clicked advertisement. If the average value > the price range of the products that the user favors, select the adjustment coefficient from small to large and increase the upper limit of the price range of the products that the user favors;

[0057] After increasing the upper limit of the price range of the products that the user favors, if the number of times of the product in the clicked advertisement > the number of times of the product in the clicked advertisement before the increase, then select and increase the upper limit of the price range of the products that the user favors again, otherwise stop adjusting the upper limit of the price range of the products that the user favors;

[0058] Similarly, when the average value < the price range of the products that the user favors, lower the lower limit of the price range of the products that the user favors through the above method;

[0059] Specifically, the user portrait adjustment module 300 senses the price of the products in the advertisements clicked by the user , where is the number of clicks, and the calculation formula for the average value of the price of the products in the clicked advertisement is: , where, represents the price of the product for the th click;

[0060] Senses the price range of the products that the user favors in the user portrait construction module 100 , where is the lower limit, is the upper limit, the adjustment coefficient is , and , When selecting, it is selected at equal intervals, and the upper limit of the increased price range is: .

[0061] If most of the prices of the products clicked by the user are higher than the upper limit of the original price range, it indicates that the user may have a new interest tendency towards products with higher prices. The user portrait adjustment module 300 will accordingly appropriately expand the upper limit of the price range or move the price range upward as a whole. On the contrary, if most of the clicked prices are lower than the lower limit of the original range, it may mean that the user's price preference has decreased, and the user portrait adjustment module 300 will lower the price range.

[0062] The advertisement placement optimization module 400 is used to sense the product pictures clicked by users in the placed advertisements, set an adjustment threshold, calculate the second cosine similarity between the clicked product pictures and the product pictures in the historical browsing data through the advertisement weight distribution module 200. If the second cosine similarity > the adjustment threshold, it is determined that the clicked product is the same as the product pictures in the historical browsing data, and the product browsing times in the historical browsing data are transferred to the current user browsing times in the advertisement weight distribution module 200, increasing the weight of the corresponding product in the advertisement weight distribution module 200. From the perspective of users' cognition and memory, people tend to pay more attention and attach more importance to things that are familiar and frequently contacted. When a product frequently appears in the user's browsing history and is clicked again, it indicates that the product has left a deep impression in the user's memory, and its attractiveness has withstood the test of time and multiple browsing comparisons.

[0063] The practice of the advertisement placement optimization module 400 integrating historical browsing times and adjusting weights conforms to the laws of users' cognition and memory, strengthens the display of these products that are deeply impressed in users' memory in advertisement placement, conforms to the attention tendency of users' subconsciousness to things they are interested in, is more likely to arouse users' resonance and positive response, and improves the influence and acceptance of advertisements in users' hearts.

[0064] Specifically, the advertisement placement optimization module 400 sets an adjustment threshold based on a large number of known identical product pictures. Specifically, through a deep learning model, the deep learning model divides the preprocessed picture data into a training set, a validation set, and a test set. During the training process, the identical pictures in the training set are input into the model. The deep learning model extracts the features of the pictures through the convolutional layer, gradually compresses the feature dimensions through the pooling layer, and outputs the prediction results for the identical pictures through the fully connected layer.

[0065] After the deep learning model is trained, the identical picture pairs in the validation set are input into the deep learning model, and the similarity scores output by the deep learning model are obtained. The average value of the similarity scores is calculated as the adjustment threshold.

[0066] Furthermore, before calculating the second cosine similarity of the product pictures, the advertisement placement optimization module 400 uses picture recognition technology to extract the feature vectors in the product pictures.

[0067] First, map the pixel values of the product images to the interval [0, 1], resize the product images to the same size, and extract feature vectors through a convolutional neural network. The convolutional layer slides a convolutional kernel over the product images to perform convolutional operations on local regions of the product images, and applies an activation function to introduce non-linearity. The role of the activation function is to enable the neural network to learn complex non-linear relationships. The pooling layer downsamples the feature maps to reduce the amount of data while retaining the main feature information;

[0068] Through the combination of multiple convolutional layers, activation functions, and pooling layers, various features of the product images from low-level to high-level are extracted.

[0069] An intelligent advertising placement method based on user behavior portraits includes the following steps:

[0070] Step 1: Divide the current browsing data and historical browsing data to build a user portrait;

[0071] Step 2: Place advertisements that match the user portrait, and allocate different advertising placement weights according to the number of times the products are browsed in the user portrait to place advertisements;

[0072] Step 3: Adjust the user portrait according to the price of the product in the advertisement clicked by the user;

[0073] Step 4: Determine whether the product image in the advertisement clicked by the user is the same as the product images in the historical browsing data. If they are the same, adjust the advertising placement weight.

[0074] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent advertising delivery system based on user behavior portraits, characterized in that, It includes a user portrait construction module (100), an advertisement weight allocation module (200), a user portrait adjustment module (300), and an advertisement placement optimization module (400); The user portrait construction module (100) is used to divide user browsing data into current browsing data and historical browsing data, select the browsed products in the current browsing data and the price range of the products preferred by the user in the historical browsing data, and eliminate the products in the user-browsed products that do not meet the preferred product price range. The pictures and prices of the user-browsed products are the user portrait; The advertisement weight allocation module (200) is used to calculate the first cosine similarity between the user portrait and the audience characteristics in the advertisement library, set a cosine similarity threshold, screen out the advertisements within the cosine similarity threshold, allocate different weights according to the current number of times the user browses products, and deliver the screened advertisements to the user according to different weights. The audience characteristics in the advertisement library refer to the product pictures in the above-mentioned user portrait construction module (100); The user profile adjustment module (300) is used to sense the price and number of times of the product clicked by the user in the displayed advertisement, and record the price of the product in the clicked advertisement > the price in the price range of the products preferred by the user in the user profile construction module (100); when the number of times of the product in the clicked advertisement > the normal number of times of clicking on the product, calculate the average value of the price of the product in the clicked advertisement. If the average value > the price range of the products preferred by the user, select the adjustment coefficient from small to large. The user profile adjustment module (300) senses the price of the product clicked by the user in the advertisement , where is the number of clicks, and the calculation formula for the average value of the price of the product in the clicked advertisement is: , where represents the price of the product for the th click; sense the price range of the products preferred by the user in the user profile construction module (100) , where is the lower limit, is the upper limit, and the adjustment coefficient is , and , is selected at equal intervals during selection. The upper limit of the increased price range is: , increasing the upper limit of the price range of the products preferred by the user. The method for adjusting the lower limit of the price range of the products preferred by the user is the same; The advertisement placement optimization module (400) is used to sense the product pictures clicked by the user in the delivered advertisements, set an adjustment threshold, calculate the second cosine similarity between the clicked product pictures and the product pictures in the historical browsing data through the advertisement weight allocation module (200). If the second cosine similarity > the adjustment threshold, it is determined that the clicked product is the same as the product pictures in the historical browsing data, and the number of times the product in the historical browsing data is browsed is transferred to the current number of times the user browses in the advertisement weight allocation module (200), and the weight of the corresponding product in the advertisement weight allocation module (200) is increased.

2. The intelligent advertising delivery system based on user behavior portraits according to claim 1, characterized in that: The user portrait construction module (100) senses the user's purchase behavior. The time node when the user's purchase behavior occurs is the division node. The browsing data corresponding after the division node is the current browsing data, and the browsing data corresponding before the division node is the historical browsing data; As the user's subsequent purchase behavior occurs again, the new purchase behavior time node replaces the previous last purchase behavior time node and becomes the new division node.

3. The intelligent advertising delivery system based on user behavior portraits according to claim 2, wherein: The price range of the products preferred by the user in the user portrait construction module (100) is calculated by the quantile method. The price corresponding to each purchased product in the historical browsing data is retrieved, and the purchased product prices are sorted in descending order. The set of purchased product price data is divided into five numerical points with equal frequencies. The second and fourth numerical points are selected from largest to smallest, and the product prices corresponding to the second and fourth numerical points are calculated. The amount between the product prices is the price range of the products preferred by the user.

4. The intelligent advertising delivery system based on user behavior portraits according to claim 1, wherein: The advertisement weight distribution module (200) calculates a first cosine similarity between two feature vectors based on the feature vector of the user profile and the feature vector of the advertisement audience , and the dimension of the direction vector First cosine similarity ; wherein and are respectively the elements between vectors and ; represents the th position in the vector.

5. The intelligent advertising delivery system based on user behavior portraits according to claim 4, wherein: The advertisement weight allocation module (200) sets the cosine similarity threshold based on industry experience. Different industries have accumulated certain empirical data in the long-term advertisement placement practice.

6. The intelligent advertising delivery system based on user behavior portraits according to claim 2, characterized in that: The advertisement weight allocation module (200) allocates weights according to the current number of times the user browses products. Since different browsing times reflect the user's interest levels in different products, advertisements related to products with high user interest are allocated higher weights. The specific calculation formula for the specific allocation difference of different weights is: ； Among them, is the number of times a user views a product, represents the product view count, used to distinguish different products, is an index variable used to iterate over all products in the summation formula.

7. The intelligent advertising delivery system based on user behavior portraits according to claim 6, wherein: The advertisement placement optimization module (400) sets an adjustment threshold based on a large number of known identical product images. Specifically, through a deep learning model: the deep learning model divides the preprocessed image data into a training set, a validation set, and a test set. During the training process, the identical images in the training set are input into the model. The deep learning model extracts the features of the images through convolutional layers, gradually compresses the feature dimensions through pooling layers, and outputs the prediction results for the identical images through fully connected layers; After the deep learning model is trained, the identical image pairs in the validation set are input into the deep learning model, and the similarity scores output by the deep learning model are obtained. The mean value of the similarity scores is calculated as the adjustment threshold.

8. The intelligent advertising delivery system based on user behavior portraits according to claim 7, wherein: Before calculating the second cosine similarity of the product image, the advertisement placement optimization module (400) extracts the feature vectors in the product image by using image recognition technology; The image recognition technology maps the pixel values of the product image to the [0,1] interval, adjusts the product image to the same size, and extracts the feature vectors through a convolutional neural network. The convolutional layer slides the convolutional kernel on the product image to perform convolutional operations on the local area of the product image, and applies an activation function to introduce non-linear factors. The role of the activation function is to enable the neural network to learn complex non-linear relationships. The pooling layer downsamples the feature map to reduce the amount of data while retaining the main feature information; Through the combination of multiple convolutional layers, activation functions, and pooling layers, various features of the product image from low-level to high-level are extracted.

9. The intelligent advertising delivery method based on user behavior portraits is applied to the intelligent advertising delivery system based on user behavior portraits according to any one of claims 1-8, and is characterized in that, It includes the following steps: Step 1: Divide the current browsing data and historical browsing data to establish a user profile; Step 2: Place advertisements that match the user profile, and allocate different advertisement placement weights according to the number of times the products in the user profile are browsed to place advertisements; Step 3: Adjust the user profile according to the price of the product in the advertisement clicked by the user; Step 4: Determine whether the product image in the advertisement clicked by the user is the same as the product images in the historical browsing data. If they are the same, adjust the advertisement placement weight.

Images