WO2024051609A1

WO2024051609A1 - Advertisement creative data selection method and apparatus, model training method and apparatus, and device and storage medium

Info

Publication number: WO2024051609A1
Application number: PCT/CN2023/116575
Authority: WO
Inventors: 刘银星; 阮涛; 张政; 吕晶晶; 詹科
Original assignee: 北京沃东天骏信息技术有限公司
Priority date: 2022-09-09
Filing date: 2023-09-01
Publication date: 2024-03-14
Also published as: CN115564469A

Abstract

An advertisement creative data selection method and apparatus, a model training method and apparatus, and a device and a storage medium. The advertisement creative data selection method comprises: acquiring candidate advertisement creative data corresponding to a target article (S110); acquiring a sparse feature vector and a picture feature vector which correspond to the candidate advertisement creative data, and obtaining, on the basis of the sparse feature vector, the picture feature vector and a pre-trained creative selection model, a recommendation probability value corresponding to the candidate advertisement creative data (S120); and selecting target advertisement creative data according to the recommendation probability value (S130).

Description

Advertising creative data selection method and device, model training method and device, equipment, storage media

This application claims priority to the Chinese patent application with application number 202211104745.3, which was submitted to the China Patent Office on September 9, 2022. The entire content of this application is incorporated into this application by reference.

Technical field

This application relates to the field of artificial intelligence technology, such as advertising creative data selection methods and devices, model training methods and devices, equipment, and storage media.

Background technique

With the continuous development of artificial intelligence technology, image processing technology and natural language processing technology have been applied to the advertising industry.

The advertising creative selection method generally uses image processing technology to process image elements in advertising materials, and uses natural language processing models to identify and process text elements in advertising materials. These methods are targeted at specific fields and have poor generalizability. Moreover, these methods cannot directly select the best advertising creative elements for users, which reduces the user experience.

Contents of the invention

This application provides advertising creative data selection methods and devices, model training methods and devices, equipment, and storage media to automatically and accurately select optimal target advertising creative data from candidate advertising creative data.

In the first aspect, this application provides a method for selecting advertising creative data, including:

Obtain candidate advertising creative data corresponding to the target item; wherein the candidate advertising creative data includes advertising images and advertising copy;

Obtain the sparse feature vector and picture feature vector corresponding to the candidate advertising creative data, and obtain the recommendation probability value corresponding to the candidate advertising creative data based on the sparse feature vector, the picture feature vector and the pre-trained creative selection model ;

Select target advertising creative data according to the recommended probability value;

Wherein, the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and output the recommendation probability value based on the fusion result.

In a second aspect, this application also provides a model training method, which method includes:

Obtain training sample data, wherein the training sample data includes a sample range corresponding to the sample item The standard recommendation probability value corresponding to the advertising creative data and the sample advertising creative data, where the sample advertising creative data includes advertising images and advertising copy;

Obtain the sparse feature vector and picture feature vector corresponding to the sample advertising creative data, and obtain the predicted recommendation probability corresponding to the sample advertising creative data based on the sparse feature vector, the picture feature vector and the creative selection model to be trained. value;

Determine a loss function according to the standard recommendation probability value and the predicted recommendation probability value, adjust the network parameters in the creative selection model to be trained based on the loss function, and stop training when the preset iteration stop conditions are met ;

Wherein, the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and output the predicted recommendation probability value based on the fusion result.

In a third aspect, this application also provides an advertising creative data selection device, which includes:

The data acquisition module is configured to obtain candidate advertising creative data corresponding to the target item; wherein the candidate advertising creative data includes advertising pictures and advertising copy;

The probability value obtaining module is configured to obtain the sparse feature vector and picture feature vector corresponding to the candidate advertisement creative data, and obtain the candidate advertisement based on the sparse feature vector, the picture feature vector and the pre-trained creative selection model. Recommendation probability value corresponding to creative data;

A data selection module configured to select target advertising creative data according to the recommendation probability value;

In a fourth aspect, this application also provides a model training device, which includes:

A sample data acquisition module configured to obtain training sample data, wherein the training sample data includes sample advertising creative data corresponding to sample items and standard recommendation probability values corresponding to the sample advertising creative data, and the sample advertising creative data includes advertisements. images and advertising copy;

A vector acquisition module configured to obtain the sparse feature vector and the picture feature vector corresponding to the sample advertising creative data, and obtain the sample advertising creative based on the sparse feature vector, the picture feature vector and the creative selection model to be trained. The predicted recommendation probability value corresponding to the data;

A model training module configured to determine a loss function based on the standard recommendation probability value and the predicted recommendation probability value, adjust the network parameters in the creative selection model to be trained based on the loss function, and adjust the network parameters in the creative selection model to be trained, and perform Stop training when iterating the stop condition;

In a fifth aspect, this application also provides a computer device, including a memory, a processor and a storage device. A computer program is stored in a memory and can be run on a processor. When the processor executes the program, the above-mentioned advertising creative data selection method or model training method is implemented.

In a sixth aspect, the present application also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the above-mentioned advertising creative data selection method or model training method is implemented.

Description of the drawings

Figure 1 is a flow chart of an advertising creative data selection method provided by an embodiment of the present application;

Figure 2 is a schematic structural diagram of a creative selection model provided by an embodiment of the present application;

Figure 3 is a flow chart of a method for obtaining image feature vectors provided by an embodiment of the present application;

Figure 4 is a flow chart of a method for obtaining candidate advertising creative data provided by an embodiment of the present application;

Figure 5 is a flow chart of another method of obtaining candidate advertising creative data provided by an embodiment of the present application;

Figure 6 is a flow chart for optimizing candidate advertising creative data provided by an embodiment of the present application;

Figure 7 is a schematic flow chart of an advertising creative data selection and optimization process provided by an embodiment of the present application;

Figure 8 is a flow chart of a model training method provided by an embodiment of the present application;

Figure 9 is a schematic structural diagram of an advertising creative data selection device provided by an embodiment of the present application;

Figure 10 is a schematic structural diagram of a model training device provided by an embodiment of the present application;

Figure 11 is a schematic structural diagram of a computer device provided by an embodiment of the present application.

Detailed ways

The present application will be described below in conjunction with the drawings and embodiments. The specific embodiments described herein are merely illustrative of the application. For convenience of description, only parts relevant to the present application are shown in the drawings.

Similar reference numbers and letters refer to similar items in the following figures, so that once an item is defined in one figure, it does not need to be defined or explained in subsequent figures. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to differentiate the description and cannot be understood as indicating or implying relative importance.

Figure 1 is a flow chart of an advertising creative data selection method provided by an embodiment of the present application. This embodiment can automatically and accurately select optimal target advertising creative data from candidate advertising creative data. This method can be executed by the advertising creative data selection device in the embodiment of the present application. The device can be implemented in software and/or hardware. As shown in Figure 1, the method includes the following steps:

S110: Obtain candidate advertising creative data corresponding to the target item.

The target item represents the item for which corresponding advertising creative data needs to be generated or selected. There may be multiple candidate advertising creative data, and each candidate advertising creative data is used to describe an advertising creative plan for the target item. Among them, the candidate advertising creative data includes advertising images and advertising copy.

In the item pages displayed on some item websites and mobile applications (Application, APP), the advertising material data (advertising copy and advertising pictures) corresponding to the item can be obtained through the item details page. By performing text recognition and text extraction on the content in the details page, the advertising copy of the target item can be obtained; by performing image recognition and target detection on the content in the details page, the advertising image of the target item can be obtained. Alternatively, the advertising material data corresponding to the target item can be obtained from the advertising creative materials provided by advertising companies. Screen the advertising creative data corresponding to the target item according to the needs to obtain the candidate advertising creative data corresponding to the target item. For example, the target item is a mobile phone. On the mobile website, view the content of the detail pages of multiple mobile phones. There are pictures of mobile phones displayed at multiple angles at the top of the details page, and corresponding advertising copy below the pictures. Perform text recognition and text extraction on the content displayed on the mobile phone details page, and extract the advertising copy at the bottom of the mobile phone details page. Position the mobile phone pictures from multiple angles in the details page to obtain the advertising pictures. When the size of the extracted advertising image is not suitable, the image can be intelligently cropped to obtain the final advertising image. After obtaining the creative data, the creative data is filtered according to needs (such as click volume, sensitive words, etc.) to obtain candidate advertising creative data.

S120: Obtain the sparse feature vector and picture feature vector corresponding to the candidate advertising creative data, and obtain the recommendation probability value corresponding to the candidate advertising creative data based on the sparse feature vector, picture feature vector and the pre-trained creative selection model.

The creative selection model is used to: use the self-attention mechanism to fuse sparse feature vectors and image feature vectors; and output recommendation probability values based on the fusion results. In this solution, the creative selection model includes a multilayer perceptron (MLP) module, a self-attention module and an output module; among them, the MLP module is used to output the first feature vector based on sparse feature vectors; the self-attention module The output module is configured to output a second feature vector based on the sparse feature vector and the image feature vector; the output module is configured to output a recommendation probability value based on the first feature vector and the second feature vector.

A sparse feature vector is a vector used to reflect multiple types of sparse features. The sparse features of this solution include item features, user features and creative features. Item characteristics include information such as item identification number, advertising space identification number, brand identification number, and item type target identification number. User characteristics include the user’s age, gender, preferences, etc. Creative features include background template features, copywriting features and picture features. Background template features include background template identification number, template style, template layout, template main color and other features. Copywriting features include Features such as main copy, sub-copy, and bubble copy (the type of copy used to indicate that items are on sale or hot-selling), etc. The copy characteristics can be obtained from the advertising copy in the advertising creative data. Image features include whether there is text on the advertising image, whether there are people, and the creative type. The image feature vector is a vector used to reflect the image features of the advertising image. The image features can be obtained from the advertising images in the advertising creative data. The MLP module can map multiple input feature vectors to a single output feature vector. The self-attention module can quickly extract important features from sparse feature vectors. In this solution, the self-attention module includes the multi-head self-attention module in the Transformer model. Among them, the Transformer model is a neural network model. The Transformer model can learn the context of the data by tracking the relationships in the sequence data. The Transformer model includes a multi-head self-attention module. The multi-head self-attention module can extract feature information from multiple dimensions, and the multi-head self-attention module is highly parallel and can combine information from different dimensions to capture multiple ranges of dependencies within the sequence.

By inputting the sparse feature matrix and image feature vector into the creative selection model, the recommendation probability value corresponding to the candidate advertising creative data can be obtained. Figure 2 is a schematic structural diagram of a creative selection model provided by an embodiment of the present application. As shown in Figure 2, the sparse feature matrix and image feature vector are input into the creative selection model, and the sparse feature matrix can be converted into a sparse feature vector using the vector conversion table. Input sparse feature vectors and image feature vectors into the multi-head self-attention module, and input sparse feature vectors into the MLP module. Finally, based on the second feature vector output by the multi-head self-attention module and the first feature vector output by the MLP module, the recommendation probability value corresponding to the candidate advertising creative data (the output result of the output module) is predicted.

Before using the creative selection model, the creative selection model needs to be trained. Collect a large number of advertising creative plans and sort out advertising creative data (background template information, item information, copywriting information, picture information, etc.) from the advertising creative plans. The annotated sparse feature vectors and image feature vectors are used as sample data, and the recommended probability values (1 or 0) of the annotated sparse feature vectors and image feature vectors are used as sample labels. Input the sample data into the creative selection model, and obtain the recommendation probability value predicted by the creative selection model corresponding to the sample data. Then use the sample labels and predicted recommendation probability values to calculate the loss function, and continuously adjust and train the model parameters of the creative selection model based on the calculation results to obtain the trained creative selection model.

The acquisition, storage, use and processing of data in the technical solution of this application are all authorized by the user and comply with the relevant provisions of national laws and regulations.

S130: Select target advertising creative data according to the recommendation probability value.

The target advertising creative data is the optimal advertising creative data selected from multiple candidate advertising creative data. The target advertising creative data includes the target item's advertising copy, advertising pictures, background templates (template style, background color, layout, etc.) and other data related to the target item's advertising plan. The recommendation probability value is the probability value output by the creative selection model to recommend the corresponding candidate advertising creative data to the user. The larger the recommendation probability value is, the higher the creative selection model believes that the corresponding candidate advertising creative data is. A preset probability value can be set according to requirements. When the recommendation probability value is greater than the preset probability value, the corresponding candidate advertising creative data is determined as the target advertising creative data. Or, directly select the candidate advertising creative data with the largest recommendation probability value as the target advertising creative data.

After selecting the target advertising creative data, this plan also includes the following steps A1-Step A2:

Step A1: Obtain the first coding information of the advertising image and the second coding information of the advertising copy in the target advertising creative data, and generate a uniform resource locator (URL, Uniform) corresponding to the target advertising creative data based on the first coding information and the second coding information. Resource Locator).

URL is a concise representation of the location and access method of resources obtained from the Internet. It is the address of standard resources on the Internet. Every file on the Internet has a unique URL, which contains information indicating the location of the file and what the browser should do with it. The first encoding information can be obtained by URL encoding the advertising image, and the second encoding information can be obtained by URL encoding the advertising copy. A URL can be generated using the first encoding information and the second encoding information, that is, the URL corresponding to the target advertising creative data. Through the URL corresponding to the target advertising creative data, the address of the advertising creative data can be directly accessed.

Step A2: After receiving the access request for the URL sent by the client, obtain the advertising image and advertising copy in the target advertising creative data according to the URL, and perform a combined image operation on the obtained advertising image and advertising copy to obtain the target Advertising creative image; send the target advertising creative image to the client for display.

Through the URL corresponding to the target advertising creative data, the address of the advertising creative data can be directly accessed. When receiving an access request for a URL sent by the client, the advertising image and advertising copy pointed to by the URL are obtained based on the URL. The obtained advertising image and advertising copy are combined to obtain an advertising image, and the advertising image is used as the target advertising creative image. For example, image processing software is used to fuse the advertising copy and advertising images together with the category information of the target item. In the process of combining pictures, the position and size of the advertising copy and advertising image can be appropriately adjusted according to the needs and actual environment, and finally the target advertising creative image can be obtained.

Generating the URL corresponding to the target advertising creative data through the above steps can save the resources occupied by image storage, and the advertising image can also be updated at any time as the URL encoding changes, improving the efficiency of providing advertising creative data to users.

The technical solution of this embodiment is to obtain candidate advertising creative data corresponding to the target item; wherein the candidate advertising creative data includes advertising pictures and advertising copy. Obtain the sparse feature vector and picture feature vector corresponding to the candidate advertising creative data, and obtain the recommendation probability value corresponding to the candidate advertising creative data based on the sparse feature vector, picture feature vector and the pre-trained creative selection model; among which, the creative selection model includes MLP module, self-attention module and output module, MLP module is used based on sparse features The vector outputs a first feature vector, the self-attention module is used to output a second feature vector based on the sparse feature vector and the picture feature vector; the output module is used to output a recommendation probability value based on the first feature vector and the second feature vector. Select target advertising creative data based on recommendation probability values. The solution of this embodiment can use a creative selection model to automatically select target advertising creative data from the collected candidate advertising creative data. The self-attention module included in the creative selection model can identify (Identifier, ID) classes of sparse feature vectors. Features are fused with image feature vectors, so that images and text can be fused to solve the multi-modal creative selection problem in the advertising system, and it has good universality.

Figure 3 is a flow chart of a method for obtaining a picture feature vector provided by an embodiment of the present application. This embodiment explains the method of obtaining a picture feature vector based on the above embodiment. As shown in Figure 3, the method of this embodiment includes the following steps:

S210. Input the advertising images in the candidate advertising creative data into the pre-trained residual neural network model.

The candidate advertising creative data includes advertising copy and advertising images. As shown in Figure 2, the image feature vector needs to be input into the multi-head self-attention module. Therefore, it is necessary to input the advertising images in the candidate advertising creative data into the pre-trained residual neural network model to obtain the image feature vector. In this embodiment of the solution, training the residual neural network model includes the following steps B1 to B3:

Step B1: Obtain the sample image and the classification label corresponding to the sample image.

The classification label is the product word of the sample item contained in the sample picture. The product word is a vocabulary used to characterize the type of sample item and does not contain brand information. In related technologies, when training a residual neural network model, the category words of items are used as classification labels. The category word of an item includes the brand information of the item, and the product word can be used to indicate the type of item and does not contain brand information. For example, if the sample item is a mobile phone of brand A, the corresponding category words include mobile phones and brand A, and the corresponding product words only include mobile phones. A sample image is an advertising image in an advertising plan for a sample item.

You can obtain sample pictures and the classification labels corresponding to the sample pictures from the detail pages of some websites or mobile APPs. Or, obtain sample images and classification labels corresponding to the sample images from the creative library provided by advertisers or professionals. For example, you can sort all items from high to low based on the exposure of all items in the item's detail page/ad creative library, select the item category information corresponding to the top 10,000 exposure items among all items, and use this category information as a sample. Label. The method of using product words as sample labels is suitable for large-scale, multi-classification tasks. It can improve the generalization of image feature vectors generated by the residual neural network model and avoid insufficient expression of image content caused by excessive concentration of item category information. Moreover, in practical applications, for advertising data, item type information is often more important than item brand information. For example, there are big differences between mobile phone advertising plans and clothing advertising plans. An advertising plan for mobile phones may need to highlight the advertising copy (describing the performance of the phone, etc.), while an advertising plan for clothing may need to highlight advertising images. However, within the same category of items, even items of different brands, the advertising plans are mostly the same. For example, for different brands of mobile phones, the difference in their advertising plans may only be the content described in the advertising copy. Therefore, using the product words of sample items as classification labels can be closer to the actual situation and improve the accuracy of the recommendation probability value of the creative selection model.

Step B2: Input the sample image into the residual neural network model to obtain the predicted classification output by the residual neural network model.

The residual neural network model is a type of convolutional neural network model, such as the residual network (Residual Network, ResNet) model. Residual neural network models are mostly suitable for image classification and object recognition. The residual neural network model is easy to optimize, and the accuracy can be improved by increasing a certain depth of the network model. The internal residual block uses skip connections, which alleviates the vanishing gradient problem caused by increasing depth in deep neural networks. Input the sample image into the residual neural network model, and the residual neural network model predicts the type corresponding to the sample image through computational reasoning. In the embodiment of this solution, the tail of the residual neural network model includes three fully connected layers for outputting 32-dimensional vectors, 128-dimensional vectors and 256-dimensional vectors respectively. That is, these three fully connected layers are added to the tail of the ResNet model.

The residual neural network model includes convolutional layers, pooling layers, activation functions and fully connected layers. Among them, operations such as convolution layer, pooling layer and activation function map the original data to the feature space of the hidden layer to obtain the feature vector. The fully connected layer can map the feature vector of the distributed feature representation to the sample label space. The fully connected layer can extract the features of the feature vector and classify the sample images according to the feature vectors of the sample images. Depending on the size of the sample data and the classification requirements of the sample data, output vectors of different dimensions can be set for the residual neural network model. Utilizing a residual neural network model with three fully connected layers at the tail that output 32-dimensional vectors, 128-dimensional vectors and 256-dimensional vectors, the category of the sample image can be flexibly and accurately predicted based on business needs and the size of the sample data. .

Step B3: Determine the loss function based on the predicted classification and classification label, adjust the network parameters in the residual neural network model based on the loss function, and stop training when the preset iteration stop conditions are met.

Predictive classification refers to the category of the sample image predicted by the residual neural network model after inputting the sample image into the residual neural network model. The classification label is the true category of the annotated sample image.

The "gap" between the predicted classification and the sample label can be calculated through the loss function. According to the loss function determined based on the predicted classification and classification label, the network parameters in the residual neural network model can be continuously adjusted, so that the predicted classification and The classification labels get closer and closer until the training stops when the preset iteration stopping condition is met. The preset iteration stop condition includes that the prediction accuracy of the residual neural network model reaches the preset accuracy range. In the embodiment of this solution, the preset accuracy range includes [75%, 80%]. The higher the prediction accuracy of the residual neural network model, the more accurate the category of the predicted sample image. But at the same time, the computational complexity of the residual neural network model becomes greater, resulting in the residual neural network in practical applications. The calculation speed of the network model is slow and may cause over-fitting problems. Therefore, on the basis of improving the prediction accuracy of the residual neural network model, in order to avoid the excessive computational complexity of the residual neural network model, the preset accuracy range can be set to [75%, 80%]. When the prediction accuracy of the residual neural network model is not less than 75% and not greater than 80%, the training of the residual neural network model can be stopped.

In the above steps, using the product words of the sample items as classification labels can improve the accuracy of the recommendation probability value of the creative selection model. Utilizing a residual neural network model with three fully connected layers at the tail that output 32-dimensional vectors, 128-dimensional vectors and 256-dimensional vectors, the category of the sample image can be flexibly and accurately predicted based on business needs and the size of the sample data. , improving the accuracy of the recommended probability value of the creative selection model.

S220: Obtain the image feature vector output by the residual neural network model.

Image feature vectors are vectors that represent some features of the image in the form of vectors. In this solution, the image feature vector output by the residual neural network model can be used to represent the category of the image. For example, if a picture of a mobile phone is input into the residual neural network model, the residual neural network model predicts that the feature of the picture is "category: mobile phone picture" and outputs a feature vector of the picture with the characteristics of "mobile phone".

The technical solution of this embodiment is to input the advertisement pictures in the candidate advertisement creative data into the pre-trained residual neural network model; and obtain the picture feature vector output by the residual neural network model. The solution of this embodiment can flexibly and accurately predict the category of the sample picture based on business needs and the size of the sample data, and use the product words of the sample items as classification labels, improving the accuracy of the recommendation probability value of the creative selection model.

Figure 4 is a flow chart of a method for obtaining candidate advertisement creative data provided by an embodiment of the present application. This embodiment explains the method of obtaining candidate advertisement creative data based on the above embodiment. As shown in Figure 4, the method of this embodiment includes the following steps:

S310: Obtain multiple advertising material data corresponding to the target item, where the advertising material data includes advertising copy and advertising images.

In the item pages displayed on some item websites and mobile apps, the advertising material data (advertising copy and advertising pictures) corresponding to the item can be obtained through the item details page. By performing text recognition and text extraction on the content in the details page, the advertising copy is obtained; by performing image recognition and image cropping on the content in the details page, the advertising image is obtained. Alternatively, the advertising material data corresponding to the item can be obtained from the advertising creative materials provided by advertising companies.

S320: Select at least one advertisement copy and at least one advertisement image from the plurality of advertisement creative data according to the online click data corresponding to the plurality of advertisement creative data.

Online click data is used to represent the click volume of creative data. The click volume can reflect the user's How much you like the creative. The more clicks a creative gets, the more people are likely to like it. Therefore, at least one advertisement copy and at least one advertisement image can be selected from the plurality of advertisement creative data based on the online click data respectively corresponding to the plurality of advertisement creative data. In this embodiment of the solution, selecting at least one advertising copy and at least one advertising image from multiple advertising creative data includes the following steps C1 to C2:

Step C1: For each creative data, determine the score of the creative data based on the average number of online clicks of the creative data and the cumulative number of times the creative data is selected.

The number of online clicks on the creative data can be determined by viewing and clicking on the target item's details page. You can use the Multi-Armed Bandits (MAB) model and use the Upper Confidence Bound algorithm (Upper Confidence Bound, UCB) to calculate the offline statistics of the advertising material data in the past month. The score is:

Represents the average number of online clicks on the creative data, n _j is the cumulative number of times the current creative data has been selected, and n represents the number of creative data. The greater the average number of online clicks and the cumulative number of times the creative data is selected, the higher the score of the creative data. The smaller the average number of online clicks and the cumulative number of times the creative data is selected, the lower the score of the creative data.

According to the above formula, the scores of the advertising copy and advertising image in the creative are calculated respectively. After selecting at least one advertising copy and at least one advertising image from multiple creative data, the selected advertising copy forms a copy group, and the score of each advertising copy in the copy group is calculated. When calculating the score for ad copy, is the average number of online clicks for the advertising copy in the copywriting group, n _j is the cumulative number of times the current advertising copy has been selected, and n represents the number of advertising copywriting in the copywriting group. The selected advertising images form an image group, and the score of each advertising image in the image group is calculated. When calculating the score for an ad image, is the average number of online clicks on the advertising images in the image group, n _j is the cumulative number of times the current advertising image has been selected, and n represents the number of advertising images in the image group.

Step C2: Select at least one advertising copy and at least one advertising image from multiple creative data based on the score of each creative data.

After obtaining the score of the advertising image and the score of the advertising copy, at least one advertising image is selected from the advertising creative data based on the score of the advertising image. Select at least one ad copy from the creative data based on the ad copy's score.

Using the above steps, the score of the creative data can be accurately calculated, and the creative data can be selected based on the score of the creative data. The appropriate creative data can be accurately and quickly selected, and the process of combining the creative data can be avoided. The combinatorial explosion problem caused by

S330: Combine the selected advertising copy and advertising image to obtain at least one candidate advertising creative data.

After selecting at least one advertising copy and at least one advertising picture, the selected advertising copy and advertising pictures are combined in pairs to obtain at least one candidate advertising creative data. For example, based on the score of each creative data, the advertising copy selected from multiple creative data is copy A and copy B. The advertising image selected from multiple creative data is image C. Then the candidate advertising creative data can be obtained as AC and BC based on the advertising copy and advertising pictures.

The technical solution of this embodiment is to obtain multiple advertising material data corresponding to the target item, where the advertising material data includes advertising copy and advertising pictures, and according to the online click data corresponding to the multiple advertising material data, from the multiple advertising material data Select at least one advertising copy and at least one advertising picture, and combine the selected advertising copy and advertising pictures to obtain at least one candidate advertising creative data. The solution of this embodiment can accurately calculate the score of the advertising material data, select the advertising material data based on the score status of the advertising material data, accurately and quickly select the appropriate advertising material data, and avoid the occurrence of mutual interaction between the advertising material data. The combinatorial explosion problem caused by the combination process. Moreover, the candidate advertising creative data obtained by combining the selected advertising copy and advertising images is more accurate and more in line with the user's wishes.

Figure 5 is a flow chart of another method for obtaining candidate advertisement creative data provided by an embodiment of the present application. This embodiment explains the method of obtaining candidate advertisement creative data based on the above embodiment. As shown in Figure 5, the method of this embodiment includes the following steps:

S410: Identify and extract advertising copy from the item details page and/or advertising creative materials of the target item.

In the item pages displayed on some item websites and mobile apps, the advertising material data corresponding to the item can be obtained through the item details page. Ad copy can be extracted from creative data. In this embodiment of the solution, identifying and extracting advertising copy from the item details page and/or advertising creative materials of the target item includes the following steps D1 to D3:

Step D1: Based on the preset character recognition model, identify the candidate copy from the item details page and/or advertising creative materials of the target item.

Character recognition models include optical character recognition (Optical Character Recognition, OCR) models. In the item details page and/or advertising creative materials of the target item, the copywriting in the item details page and/or advertising creative materials can be identified and extracted through the OCR model to obtain the advertising copy, and the identified advertising copy can be used as the target item. Choose copy.

Step D2: Based on the first word list containing preset interest point words, filter out the benefits from the candidate copy. Click on copywriting.

Benefit point words are words used in advertising copy to express item characteristics, item benefits/advantages, consumer interests, emotions/values, etc. to users. The default interest point vocabulary is a table set according to needs and actual environment to record the interest point vocabulary of target items. For example, if the target item is a camera, the first word list of the interest point vocabulary contains:

Item characteristics: small size, high pixels; Item benefits/advantages: Easily take clear and beautiful photos; Consumer benefits: Easy to carry and easy to operate; Emotions/values: Record life and show the most real world.

After obtaining the candidate copy, the interest point vocabulary is selected from the candidate copy according to the first word list.

Step D3: Based on the preset word limit conditions and/or the second word list containing preset non-selling point words, screen out the selling point copy from the remaining copy after excluding the benefit point copy from the candidate copy.

Selling point copywriting is valuable copywriting that can increase users' purchasing interest and promote product sales. Selling point copy can describe the selling points of the product in simple language, so there is a certain word limit for selling point copy. Selling point copy can be screened out from the remaining copy after removing benefit point copy from the candidate copy based on the preset word count limit. However, in some cases, advertising copy that meets the word limit is not necessarily selling point copy. In this case, the selling point copy can be screened out from the remaining copy after excluding the benefit point copy from the candidate copy based on a second word list containing preset non-selling point words. In order to more accurately screen the selling point copywriting, you can filter the selling point copywriting from the remaining copywriting after excluding the benefit point copywriting from the candidate copywriting based on the preset word limit conditions and the second word list at the same time. For example, if the default word limit is 5, then first delete the advertising copy with more than 5 words from the remaining copy after excluding the benefit point copy from the candidate copy, and then filter the remaining advertising copy according to the second word list. Delete the ad copy that contains non-selling point words and finally get the selling point copy.

In the above steps, the OCR model can be used to accurately and quickly mine and identify the advertising copy in the item details page, and obtain the final interest point copy and selling point copy through the first word list and the second word list, which can be used to solve the problem. The problem of insufficient online copywriting materials.

S420: Position and crop the item image in the item details page of the target item to obtain an advertising image.

Position and represent the item picture in the item details page, and accurately find the location of the item picture in the item details page. In the item details page, the position of the image of the target item is uncertain, and the saliency algorithm can be used to divide the item details page into multiple specific areas with unique properties (such as text areas and picture areas). Identify the item pictures on the detail page in the segmented picture area, analyze the size of the item pictures, and intelligently crop the item pictures. For example, when the main body of the item in the item picture is too small and the user cannot clearly see the main body of the item from the item picture, the item picture can be modified. Crop to get an appropriately sized item image. Use the cropped item image as an advertising image.

S430: Select at least one advertisement copy and at least one advertisement image from the plurality of advertisement creative data according to the online click data respectively corresponding to the plurality of advertisement creative data.

S440: Combine the selected advertising copy and advertising image to obtain at least one candidate advertising creative data.

The technical solution of this embodiment is to identify and extract advertising copy from the item details page of the target item and/or advertising creative materials; position and crop the item pictures in the item details page of the target item to obtain the advertisement picture; According to the online click data corresponding to the multiple creative data, at least one advertising copy and at least one advertising image are selected from the multiple advertising creative data; the selected advertising copy and advertising image are combined to obtain at least one candidate advertising creative data. The technical solution of this embodiment can accurately and quickly mine and identify advertising copy in item detail pages, and solve the problem of insufficient online copy materials through selling point copy. Through intelligent cropping of advertising images, target item images that can highlight the target items are obtained, so that the advertising images can fully display the target items.

Figure 6 is a flow chart for optimizing candidate advertising creative data provided by an embodiment of the present application. This embodiment describes a method for optimizing candidate advertising creative data based on the above embodiment. As shown in Figure 6, the method in this embodiment includes the following steps:

S510: Combine the selected advertising copy and advertising images to obtain at least one copy and image combination.

After selecting at least one advertising copy and at least one advertising image from multiple creative data, the selected advertising copy and advertising images are combined in pairs to obtain at least one copy and image combination. For example, based on the score of each creative data, the advertising copy selected from multiple creative data is copy A and copy B. The advertising image selected from multiple creative data is image C. Then according to the advertising copy and advertising image, the copy image combination can be obtained as AC and BC.

S520: Combine at least one copywriting picture combination and at least one preset background template to obtain at least one creative combination.

The default background template is a layout template with a fixed style of copywriting and picture combination that is set in advance according to the characteristics and needs of the item. After obtaining the copywriting picture combination, you can select at least one background template for the copywriting picture combination based on the category information of the items and the characteristics of the items, etc., and combine the copywriting picture combination with at least one preset background template in pairs to obtain at least one creative combination. .

S530: Screen out at least one creative combination as candidate advertising creative data from at least one creative combination based on preset filtering factors.

The preset filtering factors include category information of the target items, and/or color information of the advertising images and background templates in each creative combination. To obtain advertising creative data that is more suitable for the target item, candidate advertising creative data needs to be filtered out based on the category information of the item and/or the color information of the advertising image and background template in each creative combination. The color information may include the main color. For example, the K-Means clustering algorithm can be used to perform cluster analysis and main color extraction on the color of the picture, and identify the main color of the picture.

When selecting at least one creative combination from at least one creative combination as candidate advertising creative data based on the category information of the target item, it may be based on the category information of the target item and the preset corresponding relationship between the item category and the background template style. , determine the background template style corresponding to the target item, and then select a creative combination that matches the background template style from at least one creative combination.

When selecting at least one creative combination as candidate advertising creative data from at least one creative combination based on the color information of the advertising image and background template in each creative combination, it may be based on the main color of the advertising image and the main color of the background template, using The Hue-Saturation-Value (HSV) color model selects at least one creative combination from at least one creative combination as candidate advertising creative data by giving priority to adjacent color matching and contrasting color matching.

S540: When the size of the target item area in the advertising image contained in the candidate advertising creative data is smaller than the preset threshold, crop the target item area, and use the cropped target item image to update the advertising image contained in the candidate advertising creative data. .

The preset threshold can be set in advance according to specific needs. When the size of the target item area in the advertisement image of the target item is smaller than the preset threshold, the target item in the image may be too small and the user cannot see the item intuitively and clearly. The target detection algorithm can be used to identify the target item area in the advertising image, and the advertising image can be intelligently cropped to obtain a target item image that can highlight the target item, and the advertising image contained in the candidate advertising creative data can be updated based on the target item image.

S550: Polish the target item area in the advertisement image according to the color information of the advertisement image in the candidate advertisement creative data.

The polishing process includes at least one of adjusting brightness, adjusting contrast, and adjusting saturation. After obtaining the advertising image contained in the updated candidate advertising creative data, image analysis is performed on the advertising image based on the color of the advertising image and the color of the target item area. When the color of the target item area is too dark, which may result in the target item area not being eye-catching enough, the color brightness of the target item area can be brightened to highlight the target item and attract users to click or trigger the target item. Similarly, when the color contrast between the target item area and the advertising image is weak, which may cause the target item area to blend into the background of the advertising image, the contrast can be adjusted to highlight the target item. When the color saturation of the target item area and the advertising image is weak, you can adjust the saturation to highlight the target item. In this embodiment of the solution, based on the color information of the advertising image in the candidate advertising creative data, polishing the target item area in the advertising image includes the following steps E1 - step E2:

Step E1: Determine whether the advertising image is a color image or a black image based on the pixel values of the pixels contained in the advertising image in the candidate advertising creative data.

The pixel information in the picture can reflect the color information of the image. Calculate the pixel values of pixels in advertising images. When there are pixels with a pixel value exceeding 190 (pixel value range is 0-255) among all pixels, and the number of pixels with a pixel value exceeding 190 exceeds the total At 50% of the pixels, the advertising image is determined to be a white image. When the number of pixels with a pixel value exceeding 190 does not exceed 15% of the total pixels, and the number of pixels with a pixel value not exceeding 55 exceeds 50% of the total pixels, the advertising image is determined to be a black image. In other cases, it is determined that the advertising image is a color image. When it is determined that the advertising image is a white image, the advertising image will not be retouched.

Step E2: When the advertising image is a color image, polish the target item area in the advertising image based on the first preset brightness parameter value, the first preset contrast parameter value and the first preset saturation parameter value. .

The first preset brightness parameter value, the first preset contrast parameter value and the first preset saturation parameter value can be set in advance according to needs. For example, when the advertising picture is a color picture, the first preset brightness parameter value can be set to 15, the first preset contrast parameter value can be set to 10, and the first preset saturation parameter value can be set to 10. According to the above-mentioned first preset brightness parameter value, first preset contrast parameter value and first preset saturation parameter value, the target item area in the advertising image is retouched.

Step E3: When the advertising image is a black image, polish the target item area in the advertising image based on the second preset brightness parameter value, the second preset contrast parameter value and the second preset saturation parameter value. .

The second preset brightness parameter value is greater than the first preset brightness parameter value, the second preset contrast parameter value is greater than the first preset contrast parameter value, and the second preset saturation parameter value is greater than the first preset saturation parameter value. The second preset brightness parameter value, the second preset contrast parameter value and the second preset saturation parameter value can be set in advance according to needs. When the advertising image is a black image, since the black image has no obvious effect on image retouching, strong retouching can be used on the black image. For example, the second preset brightness parameter value may be set to 20, the second preset contrast parameter value may be set to 15, and the second preset saturation parameter value may be set to 15. According to the above-mentioned second preset brightness parameter value, second preset contrast parameter value and second preset saturation parameter value, the target item area in the advertising image is retouched.

Through the above steps, the advertising image can be polished according to the color of the advertising image, making the advertising items in the advertising image more attractive and giving full play to the display function of the advertising image for the items.

Figure 7 is a schematic flow chart of an advertising creative data selection and optimization process provided by an embodiment of the present application. As shown in Figure 7, from the item details page and/or the advertising material library, the advertising copy of the target item is identified and extracted through the OCR model in the material mining module. Through the image segmentation model in the material mining module The model and target detection model identify advertising images of target items. Calculate the scores of advertising copy and advertising images through MAB. At least one advertising copy and at least one advertising image are selected from multiple creative data based on the scores. The candidate advertising creative data is obtained through the creative element combination module, the candidate advertising creative data is input into the creative selection module, and the target advertising creative data is obtained using the output value of the creative selection model in the creative selection module.

The solution of this embodiment is to obtain at least one copywriting and picture combination by combining the selected advertising copy and advertising pictures; to obtain at least one creative combination by combining at least one copywriting and picture combination with at least one preset background template; based on the preset Screening factors, selecting at least one creative combination from at least one creative combination as candidate advertising creative data; when the size of the target item area in the advertising image contained in the candidate advertising creative data is smaller than a preset threshold, the target item area is Crop, and use the cropped target item image to update the advertising image contained in the candidate advertising creative data; and polish the target item area in the advertising image according to the color information of the advertising image in the candidate advertising creative data. The solution of this embodiment can filter creative combinations according to the category information, color information, etc. of the target item, and obtain candidate advertising creative data that is more beautiful and whose color is closer to manual design. Finally, the advertising images are polished according to the color of the advertising images, making the advertising items in the advertising images more attractive and giving full play to the display role of the advertising images in displaying the items.

Figure 8 is a flow chart of a model training method provided by an embodiment of the present application. This embodiment can train an initial model to obtain a creative selection model. This method can be executed by the model training device in the embodiment of the present application. The device can It is implemented using software and/or hardware, as shown in Figure 8. The method includes the following steps:

S610, obtain training sample data.

The training sample data includes sample advertising creative data corresponding to the sample items and standard recommendation probability values corresponding to the sample advertising creative data. The sample advertising creative data includes advertising pictures and advertising copy. In one implementation, the sample data can be obtained from a historical material database that stores sample advertising creative data and standard recommendation probability values corresponding to the sample advertising creative data. For example, based on big data and data analysis algorithms, some advertising creative data and standard recommendation probability values corresponding to the advertising creative data can be determined, and these advertising creative data and their corresponding standard probability recommendation values are stored in the historical material database. Sample data can be obtained from the historical material database.

S620: Obtain the sparse feature vector and picture feature vector corresponding to the sample advertising creative data, and obtain the predicted recommendation probability value corresponding to the sample advertising creative data based on the sparse feature vector, picture feature vector and the creative selection model to be trained.

A sparse feature vector is a vector used to reflect multiple types of sparse features. The image feature vector is used is a vector that reflects the image features of the advertising image. The image features can be obtained from the advertising images in the sample advertising creative data. The predicted recommendation probability value is the recommendation probability value corresponding to the sparse feature vector and the picture feature vector output by the untrained creative selection model based on the sparse feature vector and the picture feature vector. Input sparse feature vectors and picture feature vectors into the creative selection model to be trained. The creative selection model to be trained can output the predicted recommendation probability value corresponding to the sample advertising creative data through model calculation.

S630: Determine the loss function based on the standard recommendation probability value and the predicted recommendation probability value, adjust the network parameters in the creative selection model based on the loss function, and stop training when the preset iteration stop conditions are met.

The loss function is a function that maps a random event or the value of its related random variable into a non-negative real number to represent the "risk" or "loss" of the random event. The network parameters in the model are configuration variables inside the model, and the values of the model parameters can be adjusted according to the loss function. The creative selection model is used to: use a self-attention mechanism to fuse sparse feature vectors and picture feature vectors, and output predicted recommendation probability values based on the fusion results. In this solution, the creative selection model includes a multi-layer perceptron neural network MLP module, a self-attention module and an output module. Among them, the MLP module is used to output the first feature vector based on the sparse feature vector; the self-attention module is used to output the second feature vector based on the sparse feature vector and the picture feature vector; the output module is used to output the first feature vector based on the first feature vector and the second feature vector. Output the predicted recommendation probability value.

Input the sparse feature vector and picture feature vector into the creative selection model to be trained, and obtain the corresponding predicted recommendation probability value. There is a big "gap" between the predicted recommendation probability value and the standard recommendation probability value at this time. The creative selection model to be trained is continuously optimized based on the loss function and the "gap". Adjust the network parameters of the creative selection model so that the "gap" between the predicted recommendation probability value and the standard recommendation probability value is continuously narrowed. When the iteration stop conditions are preset, the creative selection model after training can be obtained.

The technical solution of this embodiment can obtain training sample data, obtain the sparse feature vectors and picture feature vectors corresponding to the sample advertising creative data, and obtain the sample advertising creative data based on the sparse feature vectors, picture feature vectors and the creative selection model to be trained. The corresponding predicted recommendation probability value. Determine the loss function based on the standard recommendation probability value and the predicted recommendation probability value, adjust the network parameters in the creative selection model based on the loss function, and stop training when the preset iteration stop conditions are met. The technical solution of this embodiment can continuously optimize the creative selection model, making the predicted recommendation probability value output by the creative selection model closer to the standard recommendation probability value, and improving the accuracy of the predicted recommendation probability value.

The acquisition, storage, use and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations.

Figure 9 is a schematic structural diagram of an advertising creative data selection device provided by an embodiment of the present application. Book Embodiments can automatically and accurately select optimal target advertising creative data from candidate advertising creative data. The device can be implemented in the form of software and/or hardware. The device can be integrated in any device that provides the function of selecting advertising creative data. In the equipment, as shown in Figure 9, the devices for selecting advertising creative data include:

The data acquisition module 910 is configured to acquire the candidate advertising creative data corresponding to the target item; wherein the candidate advertising creative data includes advertising pictures and advertising copy; the probability value obtaining module 920 is configured to acquire the sparse advertising creative data corresponding to the candidate advertising creative data. feature vector and picture feature vector, and based on the sparse feature vector, the picture feature vector and the pre-trained creative selection model, obtain the recommendation probability value corresponding to the candidate advertising creative data; the data selection module 930 is set to The recommended probability value selects target advertising creative data; wherein the creative selection model is used to: use a self-attention mechanism to fuse based on the sparse feature vector and the picture feature vector, and output the recommended probability value based on the fusion result .

The creative selection model includes an MLP module, a self-attention module and an output module; where:

The MLP module is used to output a first feature vector based on the sparse feature vector; the self-attention module is used to output a second feature vector based on the sparse feature vector and the picture feature vector; the output module , used to output the recommendation probability value based on the first feature vector and the second feature vector.

In one embodiment, the probability value obtaining module 920 is configured as:

Input the advertising pictures in the candidate advertising creative data into the pre-trained residual neural network model; obtain the picture feature vector output by the residual neural network model.

In one embodiment, the probability value obtaining module 920 is further configured to:

Obtain a sample picture and a classification label corresponding to the sample picture; wherein the classification label is a product word of a sample item contained in the sample picture, and the product word is used to characterize the type of the sample item and does not contain Vocabulary of brand information; input the sample picture into the residual neural network model to obtain the predicted classification output by the residual neural network model; determine a loss function based on the predicted classification and the classification label, based on the loss The function adjusts the network parameters in the residual neural network model and stops training when the preset iteration stop conditions are met.

In one embodiment, the preset iteration stop condition includes that the prediction accuracy of the residual neural network model reaches a preset accuracy range, and the preset accuracy range may include [75%, 90%].

In one embodiment, the tail of the residual neural network model includes three fully connected layers for outputting 32-dimensional vectors, 129-dimensional vectors, and 256-dimensional vectors respectively.

In one embodiment, the self-attention module includes a multi-head self-attention module in the Transformer model.

In one embodiment, the data acquisition module 910 is configured as:

Obtain multiple advertising material data corresponding to the target item, wherein the advertising material data includes advertising copy and advertising pictures; according to the online click data corresponding to the multiple advertising material data, from the multiple advertising material data Select at least one advertising copy and at least one advertising picture; combine the selected advertising copy and advertising pictures to obtain at least one candidate advertising creative data.

In one embodiment, the data acquisition module 910 is further configured to:

Identify and extract advertising copy from the item details page of the target item and/or advertising creative materials; position and crop the item image in the item details page of the target item to obtain an advertising image.

In one embodiment, the data acquisition module 910 is further configured to:

Based on the preset character recognition model, the candidate copy is identified from the item details page and/or advertising creative material of the target item; based on the first word list containing the preset interest point vocabulary, the candidate copy is identified Screening out the benefit point copywriting; based on the preset word limit conditions and/or a second vocabulary list containing preset non-selling point words, screening out the selling point copywriting from the remaining copywriting after removing the benefit point copywriting from the candidate copywriting .

In one embodiment, the data acquisition module 910 is further configured to:

For each advertising material data, the score of the advertising material data is determined based on the average number of online clicks of the advertising material data and the cumulative number of times the advertising material data is selected; based on the score of each advertising material data, from At least one advertising copy and at least one advertising image are selected from the plurality of advertising material data.

In one embodiment, the data acquisition module 910 is further configured to:

Combine the selected advertising copy and advertising pictures to obtain at least one copy and picture combination; combine the at least one copy and picture combination with at least one preset background template to obtain at least one creative combination; based on the preset filtering factors, select Filter out at least one creative combination from the at least one creative combination as candidate advertising creative data; wherein the preset filtering factors include category information of the target item, and/or the characteristics of the advertising pictures and background templates in each creative combination. Color information.

In one embodiment, the data acquisition module 910 is further configured to:

When the size of the target item area in the advertising image contained in the candidate advertising creative data is smaller than the preset threshold, the target item area is cropped, and the cropped target item image is used to update the candidate advertising creative data. Advertising images included.

In one embodiment, the data acquisition module 910 is further configured to:

According to the color information of the advertising image in the candidate advertising creative data, the advertising image in the The target item area is subjected to retouching processing; wherein the retouching processing includes at least one of adjusting brightness, adjusting contrast, and adjusting saturation.

In one embodiment, the data acquisition module 910 is further configured to:

According to the pixel values of the pixels contained in the advertising images in the candidate advertising creative data, it is determined whether the advertising image is a color image or a black image; in the case where the advertising image is a color image, based on the first preset brightness parameter value , the first preset contrast parameter value and the first preset saturation parameter value, to polish the target item area in the advertising picture; when the advertising picture is a black picture, based on the second preset brightness The parameter value, the second preset contrast parameter value and the second preset saturation parameter value are used to polish the target item area in the advertising image; wherein the second preset brightness parameter value is greater than the first Preset brightness parameter value, the second preset contrast parameter value is greater than the first preset contrast parameter value, and the second preset saturation parameter value is greater than the first preset saturation parameter value.

In one embodiment, the device is further configured to:

Obtain the first coding information of the advertising picture and the second coding information of the advertising copy in the target advertising creative data, and generate the URL corresponding to the target advertising creative data according to the first coding information and the second coding information; after receiving In the case of an access request for the URL sent by the client, the advertising image and advertising copy in the target advertising creative data are obtained according to the URL, and the obtained advertising image and advertising copy are combined to obtain the target Advertising creative image; sending the target advertising creative image to the client for display.

The above-mentioned products can execute the methods provided by any embodiment of this application, and have corresponding functional modules and effects for executing the methods.

Figure 10 is a schematic structural diagram of a model training device provided by an embodiment of the present application. The device can be implemented in the form of software and/or hardware. The device can be integrated in any device that provides model training functions, as shown in Figure 10 As shown, the model training device includes:

The sample data acquisition module 1010 is configured to obtain training sample data. The training sample data includes sample advertising creative data corresponding to sample items and standard recommendation probability values corresponding to the sample advertising creative data. The sample advertising creative data includes advertising pictures. and advertising copy; the vector acquisition module 1020 is configured to obtain the sparse feature vector and picture feature vector corresponding to the sample advertising creative data, and based on the sparse feature vector, the picture feature vector and the creative selection model to be trained, obtain The predicted recommendation probability value corresponding to the sample advertising creative data; the model training module 1030 is configured to determine a loss function based on the standard recommendation probability value and the predicted recommendation probability value, and select the creative selection model based on the loss function. Adjust the network parameters and stop when the preset iteration stop conditions are met. Stop training; wherein, the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and output the predicted recommendation probability value based on the fusion result.

The MLP module is used to output a first feature vector based on the sparse feature vector; the self-attention module is used to output a second feature vector based on the sparse feature vector and the picture feature vector; the output module , used to output the predicted recommendation probability value based on the first feature vector and the second feature vector.

Figure 11 is a schematic structural diagram of a computer device provided by an embodiment of the present application. 11 illustrates a block diagram of an exemplary computer device 12 suitable for implementing embodiments of the present application. The computer device 12 shown in FIG. 11 is only an example and should not bring any limitations to the functions and scope of use of the embodiments of the present application.

As shown in Figure 11, computer device 12 is embodied in the form of a general purpose computing device. The components of computer device 12 may include, but are not limited to, one or more processors or processing units 16, memory 28, and a bus 18 connecting various system components, including memory 28 and processing unit 16.

Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics accelerated port, a processor, or a local bus using any of a variety of bus structures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, Enhanced ISA bus, Video Electronics Standards Association (Video Electronics Standards) Association, VESA) local bus and Peripheral Component Interconnect (PCI) bus.

Computer device 12 includes a variety of computer system readable media. These media can be any available media that can be accessed by computer device 12, including volatile and nonvolatile media, removable and non-removable media.

Memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and/or cache memory 32 . Computer device 12 may include other removable/non-removable, volatile/non-volatile computer system storage media. By way of example only, storage system 34 may be configured to read and write to non-removable, non-volatile magnetic media (not shown in Figure 11, commonly referred to as a "hard drive"). Although not shown in FIG. 11, a disk drive configured to read and write to a removable non-volatile disk (eg, a "floppy disk") may be provided, and a disk drive configured to read and write to a removable non-volatile disk may be provided. An optical disc drive that reads and writes optical discs (such as Compact Disc Read-Only Memory (CD-ROM), Digital Video Disc-Read Only Memory (DVD-ROM) or other optical media). In these cases, each drive may be connected to bus 18 through one or more data media interfaces. The memory 28 may include at least one program product having a set of (eg, at least one) program modules configured to perform the functions of embodiments of the present application.

A program/utility 40 having a set of (at least one) program modules 42, which may be stored, for example, in memory 28, such program modules 42 including, but not limited to, an operating system, one or more application programs, other programs Modules, as well as program data, each or a combination of these examples may include an implementation of a network environment. Program modules 42 generally perform functions and/or methods in the embodiments described herein.

Computer device 12 may also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), with one or more devices that enable a user to interact with computer device 12, and/or with Any device (eg, network card, modem, etc.) that enables the computer device 12 to communicate with one or more other computing devices. This communication may occur through an input/output (I/O) interface 22 . In addition, in the computer device 12 in this embodiment, the display 24 does not exist as an independent entity, but is embedded in the mirror. When the display surface of the display 24 is not displayed, the display surface of the display 24 and the mirror surface are visually integrated. Moreover, the computer device 12 can also communicate with one or more networks (such as a local area network (Local Area Network, LAN), a wide area network (Wide Area Network, WAN), and/or a public network, such as the Internet) through the network adapter 20. As shown, network adapter 20 communicates with other modules of computer device 12 via bus 18 . It should be understood that, although not shown in the figures, other hardware and/or software modules may be used in conjunction with the computer device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, disk arrays (Redundant Arrays). of Independent Disks, RAID) systems, tape drives, and data backup storage systems, etc.

The processing unit 16 executes a variety of functional applications and data processing by running programs stored in the memory 28, for example, implementing an advertising creative data selection method provided in the embodiment of the present application: obtaining candidate advertising creative data corresponding to the target item; Wherein, the candidate advertising creative data includes advertising pictures and advertising copy; the sparse feature vectors and picture feature vectors corresponding to the candidate advertising creative data are obtained, and based on the sparse feature vectors, the picture feature vectors and pre-trained creative Select a model to obtain the recommendation probability value corresponding to the candidate advertising creative data; wherein the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and based on the fusion result Output the recommended probability value, or implement a model training method provided by the embodiment of the present application: obtain training sample data, wherein the training sample data includes sample advertising creative data corresponding to the sample item and the sample advertising creative data Corresponding standard recommendation probability value, the sample Advertising creative data includes advertising pictures and advertising copy; obtain the sparse feature vectors and picture feature vectors corresponding to the sample advertising creative data, and obtain the sparse feature vectors, the picture feature vectors and the creative selection model to be trained. The predicted recommendation probability value corresponding to the sample advertising creative data; determine the loss function according to the standard recommendation probability value and the predicted recommendation probability value, and adjust the network parameters in the creative selection model to be trained based on the loss function , and stop training when the preset iteration stop conditions are met; wherein the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and based on the fusion result Output the predicted recommendation probability value.

Embodiments of the present application provide a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, an advertising creative data selection method as provided in all embodiments of the present application is implemented: obtaining the corresponding target item. Candidate advertising creative data; wherein, the candidate advertising creative data includes advertising pictures and advertising copy; obtain the sparse feature vector and picture feature vector corresponding to the candidate advertising creative data, and based on the sparse feature vector, the picture feature vector and a pre-trained creative selection model to obtain the recommendation probability value corresponding to the candidate advertising creative data; wherein the creative selection model is used to: use a self-attention mechanism to compare the sparse feature vector and the picture feature vector Perform fusion, and output the recommended probability value based on the fusion result, or implement a model training method provided by the embodiment of the present application: obtain training sample data, wherein the training sample data includes sample advertising creative data corresponding to the sample item The standard recommendation probability value corresponding to the sample advertising creative data, which includes advertising pictures and advertising copy; obtains the sparse feature vector and picture feature vector corresponding to the sample advertising creative data, and based on the sparse features vector, the picture feature vector and the creative selection model to be trained, to obtain the predicted recommendation probability value corresponding to the sample advertising creative data; determine a loss function based on the standard recommendation probability value and the predicted recommendation probability value, based on the The loss function adjusts the network parameters in the creative selection model to be trained, and stops training when the preset iteration stop conditions are met; wherein the creative selection model is used to: use a self-attention mechanism to The sparse feature vector and the picture feature vector are fused, and the predicted recommendation probability value is output based on the fusion result.

Any combination of one or more computer-readable media may be employed. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. Examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more conductors, portable computer disks, hard drives, RAM, ROM, Erasable Programmable Read-Only Memory Memory, EPROM or flash memory), optical fiber, CD-ROM, optical storage device, magnetic storage device, or any suitable combination of the above. As used in this document, a computer-readable storage medium may be any tangible medium that contains or stores a program that The program may be used by or in conjunction with an instruction execution system, apparatus, or device.

A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave carrying computer-readable program code therein. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device .

Program code embodied on a computer-readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wire, optical cable, radio frequency (Radio Frequency, RF), etc., or any suitable combination of the above.

Computer program code for performing the operations of the present application may be written in one or more programming languages, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages, or a combination thereof. A programming language, such as "C" or a similar programming language. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In situations involving remote computers, the remote computer may be connected to the user computer through any kind of network, including a LAN or WAN, or may be connected to an external computer (eg, through the Internet using an Internet service provider).

Claims

A method for selecting advertising creative data, including:

Obtain candidate advertising creative data corresponding to the target item; wherein the candidate advertising creative data includes advertising images and advertising copy;

Obtain the sparse feature vector and picture feature vector corresponding to the candidate advertising creative data, and obtain the recommendation probability value corresponding to the candidate advertising creative data based on the sparse feature vector, the picture feature vector and the pre-trained creative selection model ;

Select target advertising creative data according to the recommended probability value;

Wherein, the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and output the recommendation probability value based on the fusion result.
The method according to claim 1, wherein the creative selection model includes a multi-layer perceptron neural network MLP module, a self-attention module and an output module; wherein:

The MLP module is used to output a first feature vector based on the sparse feature vector;

The self-attention module is used to output a second feature vector based on the sparse feature vector and the picture feature vector;

The output module is configured to output the recommendation probability value based on the first feature vector and the second feature vector.
The method according to claim 1, wherein said obtaining the image feature vector corresponding to the candidate advertising creative data includes:

Input the advertising images in the candidate advertising creative data into the pre-trained residual neural network model;

Obtain the image feature vector output by the residual neural network model.
The method according to claim 3, wherein the training method of the residual neural network model includes:

Obtain a sample picture and a classification label corresponding to the sample picture; wherein the classification label is a product word of a sample item contained in the sample picture, and the product word is used to characterize the type of the sample item and does not contain vocabulary of brand messages;

Input the sample picture into the residual neural network model to be trained, and obtain the predicted classification output by the residual neural network model to be trained;

Determine a loss function based on the predicted classification and the classification label, adjust network parameters in the residual neural network model to be trained based on the loss function, and stop training when preset iteration stop conditions are met. .
The method according to claim 3, wherein the tail of the residual neural network model includes three fully connected layers for outputting 32-dimensional vectors, 128-dimensional vectors and 256-dimensional vectors respectively.
The method according to any one of claims 1-5, wherein said obtaining the candidate advertising creative data corresponding to the target item includes:

Obtain multiple advertising material data corresponding to the target item, wherein the advertising material data includes advertising copy and advertising pictures;

Select at least one advertising copy and at least one advertising image from the plurality of advertising material data according to the online click data corresponding to the plurality of advertising material data;

Combine the selected advertising copy and advertising images to obtain at least one candidate advertising creative data.
The method according to claim 6, wherein said obtaining multiple advertising material data corresponding to the target item includes:

Identify and extract advertising copy from at least one of the item details page and advertising creative material of the target item;

Position and crop the item image in the item details page of the target item to obtain an advertising image.
The method according to claim 7, wherein the advertising copy includes benefit point copy and selling point copy; and the advertisement is identified and extracted from at least one of the item details page of the target item and the advertising creative material. Copywriting, including:

Based on a preset character recognition model, identify the candidate copy from at least one of the item details page and advertising creative material of the target item;

Based on the first vocabulary list containing preset interest point vocabulary, filter out the interest point copywriting from the candidate copywriting;

Based on at least one of the preset word limit conditions and the second word list containing preset non-selling point words, the selling point copy is screened out from the remaining copy after excluding the benefit point copy from the candidate copy.
The method according to claim 6, wherein at least one advertising copy and at least one advertising image are selected from the plurality of advertising material data according to the online click data respectively corresponding to the plurality of advertising material data, include:

For each advertising material data, determine the score of the advertising material data based on the average number of online clicks of the advertising material data and the cumulative number of times the advertising material data is selected;

According to the score of each advertising material data, at least one advertising copy and at least one advertising image are selected from the plurality of advertising material data.
The method according to claim 6, wherein the selected advertising copy and advertising image are combined to obtain at least one candidate advertising creative data, including:

Combine the selected advertising copy and advertising pictures to obtain at least one copy copy and picture combination;

Combine the at least one copywriting picture combination and at least one preset background template to obtain at least one creative combination;

Based on preset screening factors, at least one creative combination is selected from the at least one creative combination as candidate advertising creative data; wherein the preset screening factors include category information of the target item, and the content of each creative combination. At least one of the color information of the ad image and background template.
The method according to claim 6, after obtaining at least one candidate advertising creative data, further comprising:

According to the color information of the advertising image in the candidate advertising creative data, the target item area in the advertising image is retouched; wherein the retouching process includes at least one of adjusting brightness, adjusting contrast, and adjusting saturation. .
The method according to claim 11, wherein polishing the target item area in the advertising image according to the color information of the advertising image in the candidate advertising creative data includes:

Determine whether the advertising image is a color image or a black image based on the pixel values of the pixels contained in the advertising image in the candidate advertising creative data;

When the advertising picture is a color picture, the target item area in the advertising picture is polished based on the first preset brightness parameter value, the first preset contrast parameter value and the first preset saturation parameter value. deal with;

In the case where the advertising image is a black image, the target item area in the advertising image is retouched based on the second preset brightness parameter value, the second preset contrast parameter value and the second preset saturation parameter value. deal with;

Wherein, the second preset brightness parameter value is greater than the first preset brightness parameter value, the second preset contrast parameter value is greater than the first preset contrast parameter value, and the second preset saturation parameter value is greater than the first preset contrast parameter value. The parameter value is greater than the first preset saturation parameter value.
The method according to any one of claims 1-5, after selecting the target advertising creative data, further comprising:

Obtain the first coding information of the advertising image and the second coding information of the advertising copy in the target advertising creative data, and generate a unified resource locator URL corresponding to the target advertising creative data based on the first coding information and the second coding information. ;

When an access request for the URL sent by the client is received, the advertising image and advertising copy in the target advertising creative data are obtained according to the URL, and the obtained advertising image and advertising copy are combined into a picture. , obtain the target advertisement creative image; send the target advertisement creative image to the client for display.
A model training method including:

Obtain training sample data, wherein the training sample data includes sample advertising creative data corresponding to the sample items and standard recommendation probability values corresponding to the sample advertising creative data, and the sample advertising creative data includes advertising pictures and advertising copy;

Obtain the sparse feature vector and picture feature vector corresponding to the sample advertising creative data, and obtain the predicted recommendation probability corresponding to the sample advertising creative data based on the sparse feature vector, the picture feature vector and the creative selection model to be trained. value;

Determine a loss function according to the standard recommendation probability value and the predicted recommendation probability value, adjust the network parameters in the creative selection model to be trained based on the loss function, and when the preset iteration stop conditions are met Stop training;

Wherein, the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and output the predicted recommendation probability value based on the fusion result.
An advertising creative data selection device, including:

The data acquisition module is configured to obtain candidate advertising creative data corresponding to the target item; wherein the candidate advertising creative data includes advertising pictures and advertising copy;

The probability value obtaining module is configured to obtain the sparse feature vector and picture feature vector corresponding to the candidate advertisement creative data, and obtain the candidate advertisement based on the sparse feature vector, the picture feature vector and the pre-trained creative selection model. Recommendation probability value corresponding to creative data;

A data selection module configured to select target advertising creative data according to the recommendation probability value;

Wherein, the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and output the recommendation probability value based on the fusion result.
A model training device including:

A sample data acquisition module configured to obtain training sample data, wherein the training sample data includes sample advertising creative data corresponding to sample items and standard recommendation probability values corresponding to the sample advertising creative data, and the sample advertising creative data includes advertisements. images and advertising copy;

The vector acquisition module is configured to obtain the sparse feature vector and picture feature vector corresponding to the sample advertising creative data, and based on the sparse feature vector, the picture feature vector and the creative to be trained Select a model intentionally to obtain the predicted recommendation probability value corresponding to the sample advertising creative data;

A model training module configured to determine a loss function based on the standard recommendation probability value and the predicted recommendation probability value, adjust the network parameters in the creative selection model to be trained based on the loss function, and adjust the network parameters in the creative selection model to be trained, and perform Stop training if the iteration stop condition is met;

Wherein, the creative selection model is used to: use a self-attention mechanism to fuse the sparse feature vector and the picture feature vector, and output the predicted recommendation probability value based on the fusion result.
An electronic device including:

at least one processor; and

a memory communicatively connected to the at least one processor; wherein,

The memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor, so that the at least one processor can perform any one of claims 1-13 The advertising creative data selection method, or the model training method of claim 14.
A computer-readable storage medium that stores computer instructions, and the computer instructions are used to implement the advertising creative data selection method described in any one of claims 1-13 when executed by a processor, Or the model training method according to claim 14.