Advertisement material recall model training method and device, equipment and storage medium

Abstract

The application discloses a training method and device of an advertisement material recall model, equipment and a storage medium, and belongs to the technical field of computers. Through the technical scheme provided by the embodiment of the application, the account features, text features and image features of sample accounts, sample advertisement texts and sample advertisement images can be extracted through the advertisement material recall model. Based on the first difference information between the account features and the text features, and the second difference information between the account features and the image features, the advertisement material recall model is trained, thereby realizing alignment of the three types of features, namely, the account features, the text features and the image features, and enabling the advertisement material recall model to have the capability of recalling advertisement materials based on accounts.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a training method, apparatus, device, and storage medium for an advertising creative recall model. Background Technology

[0002] With the development of computer technology, the number of advertising materials is increasing, and advertising materials come in many types, including advertising text, advertising video, and advertising image.

[0003] In related technologies, there are various types of advertising creative recall models. For example, there are text-image recall models that recall advertising images based on advertising text, and there are image-text recall models that recall advertising text based on advertising images.

[0004] However, various types of ad creative recall models can only achieve mutual recall between different types of ad creatives, and the recall methods are relatively limited. Summary of the Invention

[0005] This application provides a training method, apparatus, device, and storage medium for an advertising creative recall model, which can train an advertising creative recall model based on account recall of advertising creatives, enriching recall methods. The technical solution is as follows:

[0006] On the one hand, a training method for an advertising creative recall model is provided, the method comprising:

[0007] The sample account, sample ad text, and sample ad image are input into the ad creative recall model, which is used to recall at least one of the ad text and ad image based on the account.

[0008] The advertising material recall model extracts features from the sample account, the sample advertising text, and the sample advertising image to obtain the account features of the sample account, the text features of the sample advertising text, and the image features of the sample advertising image.

[0009] Based on the association between the sample account, the sample advertisement text, and the sample advertisement image, as well as the first difference information and the second difference information, the advertising material recall model is trained. The first difference information is the difference information between the account features of the sample account and the text features of the sample advertisement text, and the second difference information is the second difference information between the account features of the sample account and the image features of the sample advertisement image.

[0010] On the one hand, a training device for an advertising creative recall model is provided, the device comprising:

[0011] The input module is used to input sample accounts, sample advertising text, and sample advertising images into the advertising creative recall model, wherein the advertising creative recall model is used to recall at least one of the advertising text and advertising images based on the account.

[0012] The feature extraction module is used to extract features from the sample account, the sample advertisement text, and the sample advertisement image through the advertisement material recall model, respectively, to obtain the account features of the sample account, the text features of the sample advertisement text, and the image features of the sample advertisement image;

[0013] A training module trains the advertising creative recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, as well as first and second difference information. The first difference information is the difference between the account features of the sample account and the text features of the sample advertising text, and the second difference information is the difference between the account features of the sample account and the image features of the sample advertising image. In another aspect, a computer device is provided, comprising one or more processors and one or more memories, wherein at least one computer program is stored in the one or more memories, and the computer program is loaded and executed by the one or more processors to implement the training method for the advertising creative recall model.

[0014] In one possible implementation, the feature extraction module is used to: embed the sample account into the sample account using the advertising creative recall model to obtain the account features of the sample account; segment the sample advertising text using the advertising creative recall model to obtain multiple words in the sample advertising text; encode the multiple words and their positions based on an attention mechanism to obtain the lexical features of each word; fuse the lexical features of the multiple words to obtain the text features of the sample advertising text; split the advertising image into multiple image blocks using the advertising creative recall model; encode the multiple image blocks and their positions based on an attention mechanism to obtain the image block features of each image block; and fuse the image block features of the multiple image blocks to obtain the image features of the advertising image.

[0015] In one possible implementation, the feature extraction module is used to embed and encode the plurality of words and the position of each word to obtain a first fusion feature of each word, the first fusion feature including the word embedding feature and the position embedding feature of the corresponding word; input the first fusion feature of each word into the text encoding unit of the advertising material recall model, and through the text encoding unit, determine the attention weight between every two words in the plurality of words based on the first fusion feature of each word; and through the text encoding unit, output the word feature of each word based on the attention weight between every two words in the plurality of words and the first fusion feature of each word.

[0016] In one possible implementation, the feature extraction module is used to embed and encode the plurality of image blocks and the position of each image block to obtain a second fusion feature of each image block. The second fusion feature includes an image block embedding feature and a position embedding feature of the corresponding image block. The second fusion feature of each image block is input into the image encoding unit of the advertising material recall model. The image encoding unit determines the attention weight between every two image blocks in the plurality of image blocks based on the second fusion feature of each image block. The image encoding unit outputs the image block feature of each image block based on the attention weight between every two image blocks in the plurality of image blocks and the second fusion feature of each image block.

[0017] In one possible implementation, the training module is used to construct a joint loss function based on the first difference information and the second difference information; and to train the advertising creative recall model based on the association between the sample account, the sample advertising text, and the sample advertising image and the joint loss function.

[0018] In one possible implementation, the device further includes:

[0019] The dimension adjustment module is used to adjust the account features of the sample account, the text features of the sample advertisement text, and the image features of the sample advertisement image to the same dimension.

[0020] In one possible implementation, the device further includes:

[0021] The recall module is used to input the target account into the trained advertising material recall model; extract features of the target account through the advertising material recall model to obtain the account features of the target account; and recall the target advertising material based on the account features of the target account through the advertising material recall model to obtain the target advertising material, wherein the target advertising material includes at least one of advertising text and advertising image, and the target advertising material matches the target account.

[0022] In one possible implementation, the recall module is used to determine the similarity between the account characteristics of the target account and the indexes of multiple candidate advertising materials through the advertising material recall model, wherein the multiple candidate advertising materials include at least one of candidate advertising text and candidate advertising image; and to determine the candidate advertising materials whose indexes among the multiple candidate advertising materials meet the similarity condition between the indexes and the account characteristics of the target account as the target advertising material.

[0023] In one possible implementation, the device further includes:

[0024] The index determination module is used to input the plurality of candidate advertising materials into the advertising material recall model; extract features from the plurality of candidate advertising materials through the advertising material recall model to obtain the material features of each candidate advertising material; split the material features of each candidate advertising material into multiple sub-features; and determine the index of each candidate advertising material based on the multiple sub-features of each candidate advertising material.

[0025] In one possible implementation, the index determination module is used to cluster multiple sub-features of the multiple candidate advertising materials to obtain the index of each candidate advertising material.

[0026] On the one hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, the computer program being loaded and executed by a processor to implement a training method for the advertising creative recall model.

[0027] On the one hand, a computer program product or computer program is provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to perform the training method of the above-mentioned advertising material recall model.

[0028] The technical solution provided in this application allows for the extraction of account features, text features, and image features from sample accounts, sample ad text, and sample ad images using an ad creative retrieval model. Based on the first difference information between account features and text features, and the second difference information between account features and image features, the ad creative retrieval model is trained to align these three types of features, enabling the ad creative retrieval model to retrieve ad creatives based on accounts. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of the implementation environment of a training method for an advertising material recall model provided in an embodiment of this application;

[0031] Figure 2 This is a flowchart of a training method for an advertising material recall model provided in an embodiment of this application;

[0032] Figure 3 This is a flowchart of another training method for an advertising material recall model provided in an embodiment of this application;

[0033] Figure 4 This is an architecture diagram of a training method for an advertising creative recall model provided in an embodiment of this application;

[0034] Figure 5 This is a flowchart illustrating a method for recalling advertising materials, as provided in an embodiment of this application.

[0035] Figure 6 This is a flowchart illustrating the process of determining an index, as provided in an embodiment of this application.

[0036] Figure 7 This is a flowchart of another method for recalling advertising materials provided in an embodiment of this application;

[0037] Figure 8 This is a schematic diagram of the structure of a training device for an advertising material recall model provided in an embodiment of this application;

[0038] Figure 9 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0040] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0041] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0042] Machine Learning (ML) is a multidisciplinary field that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and many other disciplines. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge sub-models to continuously improve their performance.

[0043] Semantic features: Features used to represent the semantics expressed by advertising text. Different advertising texts can correspond to the same semantic features; for example, the advertising text "How's the weather today?" and the advertising text "How's the weather today?" can correspond to the same semantic feature. Computer devices can map the characters in the advertising text into character vectors, and combine and operate on the character vectors according to the relationships between characters to obtain the semantic features of the advertising text. For example, computer devices can use a bidirectional encoder representation from transformers (BERT).

[0044] Mask: A mask is a string of binary code multiplied by a target field to either hide or show a specific character within the target field. For example, if the target field is (1, 1, 0, 1) and the mask is (1, 0, 1, 0), multiplying the target field and the mask yields (1, 0, 0, 0). This means the first and third characters of the target field are preserved, while the second and fourth characters are "masked" and become 0. The mask reveals which characters in the target field are preserved and which are "masked."

[0045] Normalization: Mapping sequences of values ​​with different ranges to the interval (0, 1) to facilitate data processing. In some cases, normalized values ​​can be directly expressed as probabilities.

[0046] Dropout is a method for optimizing deep artificial neural networks. During the learning process, it reduces the interdependence between nodes by randomly setting some weights or outputs of the hidden layers to zero, thereby regularizing the neural network and reducing its structural risks. For example, in model training, given a vector (1, 2, 3, 4), after inputting this vector into a dropout layer, the dropout layer can randomly convert one of the numbers in the vector (1, 2, 3, 4) to 0. For example, converting 2 to 0 would change the vector to (1, 0, 3, 4).

[0047] Embedded coding, mathematically speaking, represents a correspondence, that is, mapping data in space X to space Y using a function F. This function F is injective, and the mapping result preserves the structure. An injective function means that the mapped data uniquely corresponds to the original data, and preserving the structure means that the order of the original data remains the same. For example, if there are data X1 and X2 before mapping, after mapping we get Y1 corresponding to X1 and Y2 corresponding to X2. If the original data X1 > X2, then correspondingly, the mapped data Y1 > Y2. For words, this means mapping words to another space to facilitate subsequent machine learning and processing.

[0048] Attention weights represent the importance of a piece of data during training or prediction. Importance indicates the magnitude of the influence of input data on output data. Data with high importance corresponds to higher attention weights, while data with low importance corresponds to lower attention weights. The importance of data varies in different scenarios, and training the model to assign attention weights is essentially the process of determining data importance.

[0049] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0050] Figure 1 This is a schematic diagram illustrating the implementation environment of a training method for an advertising creative recall model provided in this application embodiment. See also... Figure 1 The implementation environment may include terminal 110 and server 140.

[0051] Terminal 110 is connected to server 140 via a wireless or wired network. Optionally, terminal 110 may be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 110 has an application installed and running that supports advertising creative retrieval. In some embodiments, this application is provided to advertisers, who can use it to retrieve advertising creatives and thus complete the advertising creative.

[0052] Server 140 is a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. Server 140 provides background services for the applications running on terminal 110, namely, providing model training services and services for recalling advertising creatives using the models.

[0053] Those skilled in the art will understand that the number of terminals described above can be more or less. For example, there may be only one terminal, or there may be dozens or hundreds of terminals, or even more, in which case other terminals may also be included in the above implementation environment. This application does not limit the number of terminals or the type of device in its embodiments.

[0054] After introducing the implementation environment of the embodiments of this application, the application scenarios of the embodiments of this application will be introduced in conjunction with the above implementation environment. In the following description, the terminal is the terminal 110 in the above implementation environment, and the server is the server 140 in the above implementation environment.

[0055] The technical solution provided in this application can be applied to scenarios involving the recall of advertising creatives. By adopting the technical solution provided in this application, an advertising creative recall model based on accounts can be trained. Using this advertising creative recall model, advertising creative recall can be achieved based on accounts, providing richer advertising recall methods.

[0056] After introducing the implementation environment and application scenarios of the embodiments of this application, the training method of the advertising material recall model provided in the embodiments of this application will be described below. See also Figure 2 Taking the server as the executing entity as an example, the method includes the following steps.

[0057] 201. The server inputs the sample account, sample ad text, and sample ad image into the ad creative recall model, which is used to recall at least one of the ad text and ad image based on the account.

[0058] The sample accounts, sample ad text, and sample ad images serve as training samples for the ad creative recall model. The sample ad text can be either associated with or unrelated to the sample account. Ad text associated with the sample account refers to ad text that the sample account has interacted with, such as ad text that has been posted, saved, viewed, used, or liked by the sample account. Conversely, ad text unrelated to the sample account refers to ad text that the sample account has not interacted with. Similarly, ad images associated with the sample account refer to ad images that the sample account has interacted with, such as ad images that have been posted, saved, viewed, used, or liked by the sample account. Conversely, ad images unrelated to the sample account refer to ad images that the sample account has not interacted with.

[0059] 202. The server uses the advertising material recall model to extract features from the sample account, the sample advertising text, and the sample advertising image, respectively, to obtain the account features of the sample account, the text features of the sample advertising text, and the image features of the sample advertising image.

[0060] The process of feature extraction for sample accounts, sample advertisement text, and sample advertisement images involves abstracting and representing these elements to obtain account features, text features, and image features that represent the sample account, sample advertisement text, and sample advertisement image. Subsequent processing of these elements then focuses on these key features, improving processing efficiency.

[0061] 203. The server trains the advertising material recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, the first difference information, and the second difference information. The first difference information is the difference information between the account features of the sample account and the text features of the sample advertising text, and the second difference information is the second difference information between the account features of the sample account and the image features of the sample advertising image.

[0062] The association between the sample account and the sample ad text and image includes the association between the sample account and the sample ad text, and the association between the sample account and the sample ad image. The purpose of training the ad creative recall model based on the first and second difference information is to ensure that the account features of the sample account are as close as possible to the text features of the associated sample ad text, and as far away as possible from the text features of the unassociated sample ad text; similarly, the account features of the sample account are as close as possible to the image features of the associated sample ad image, and as far away as possible from the image features of the unassociated sample ad image. This achieves alignment of the three types of features: account features, text features, and image features.

[0063] The technical solution provided in this application allows for the extraction of account features, text features, and image features from sample accounts, sample ad text, and sample ad images using an ad creative retrieval model. Based on the first difference information between account features and text features, and the second difference information between account features and image features, the ad creative retrieval model is trained to align these three types of features, enabling the ad creative retrieval model to retrieve ad creatives based on accounts.

[0064] Steps 201-203 above are a brief introduction to the technical solutions provided in the embodiments of this application. The technical solutions provided in the embodiments of this application will be explained more clearly below with some examples. See [link to relevant documentation]. Figure 3 Taking the server as the executing entity as an example, the method includes the following steps.

[0065] 301. The server obtains sample accounts, sample advertisement text, and sample advertisement images.

[0066] The sample accounts, sample ad text, and sample ad images serve as training samples for the ad creative recall model. The sample ad text can be either associated with or unrelated to the sample account. Ad text associated with the sample account refers to ad text that the sample account has interacted with, such as ad text that has been posted, saved, viewed, used, or liked by the sample account. Conversely, ad text unrelated to the sample account refers to ad text that the sample account has not interacted with. Similarly, ad images associated with the sample account refer to ad images that the sample account has interacted with, such as ad images that have been posted, saved, viewed, used, or liked by the sample account. Conversely, ad images unrelated to the sample account refer to ad images that the sample account has not interacted with.

[0067] In one possible implementation, the server obtains a set of sample accounts, a set of sample advertisement texts, and a set of sample advertisement images. The server retrieves a sample account from the sample account set, a sample advertisement text from the sample advertisement text set, and a sample advertisement image from the sample advertisement image set. The sample account is any account in the sample account set, the sample advertisement text is any advertisement text in the sample advertisement text set, and the sample advertisement image is any advertisement image in the sample advertisement image set.

[0068] The set of sample accounts, sample ad texts, and sample ad images are collectively referred to as the training samples for the ad creative recall model. Training the ad creative recall model involves multiple iterations, and the sample account, sample ad text, and sample ad image are the training samples used in one training iteration.

[0069] It should be noted that all sample accounts in the sample account set were obtained after full authorization.

[0070] In this implementation, the server can sample from the sample account set, sample advertisement text set, and sample advertisement image set to obtain the sample account, the sample advertisement text, and the sample advertisement image, resulting in high efficiency in obtaining training samples.

[0071] For example, the server retrieves the sample account set from the corresponding maintained account database, the sample ad text set from the corresponding maintained ad text database, and the sample ad image set from the corresponding maintained ad image database. The server then retrieves the sample account from the sample account set, the sample ad text from the sample ad text set, and the sample ad image from the sample ad image set.

[0072] 302. The server inputs the sample account, sample ad text, and sample ad image into the ad creative recall model, which is used to recall at least one of the ad text and ad image based on the account.

[0073] Among them, at least one of the recall advertising text and advertising image includes three situations: recall advertising text, recall advertising image, and recall advertising text and advertising image.

[0074] In some embodiments, the advertising material recall model includes an account feature extraction unit, a text feature extraction unit, an image feature extraction unit, and a recall unit. The account feature extraction unit is used to extract features from the input account, the text feature extraction unit is used to extract features from the input advertising text, the image feature extraction unit is used to extract features from the input advertising image, and the recall unit is used to recall advertising materials based on the extracted features.

[0075] Accordingly, the server inputting the sample account, sample ad text, and sample ad image into the ad creative recall model means that the server inputs the sample account into the account feature extraction unit, the sample ad text into the text feature extraction unit, and the sample ad image into the image feature extraction unit.

[0076] In some embodiments, the text feature extraction unit and the image feature extraction unit are trained, meaning that the text features and image features extracted by the ad creative recall model are aligned. Training the ad creative recall model refers to training the account feature extraction unit of the ad creative recall model. When training the account feature extraction unit, a sample account, multiple sample ad texts, and multiple sample ad images are used.

[0077] 303. The server uses the advertising material recall model to extract features from the sample account, the sample advertising text, and the sample advertising image, respectively, to obtain the account features of the sample account, the text features of the sample advertising text, and the image features of the sample advertising image.

[0078] The process of feature extraction for sample accounts, sample advertisement text, and sample advertisement images involves abstracting and representing these elements to obtain account features, text features, and image features that represent the sample account, sample advertisement text, and sample advertisement image. Subsequent processing of these elements then focuses on these key features, improving processing efficiency.

[0079] In one possible implementation, the server uses the ad creative recall model to embed and encode the sample account, obtaining the account features of the sample account. The server then uses the ad creative recall model to segment the sample ad text, obtaining multiple words within the text. These words and their positions are encoded using an attention mechanism, obtaining the lexical features of each word. The lexical features of these multiple words are then fused to obtain the text features of the sample ad text. Finally, the server uses the ad creative recall model to split the sample ad image into multiple image blocks. These image blocks and their positions are encoded using an attention mechanism, obtaining the image block features of each image block. The image block features of these multiple image blocks are then fused to obtain the image features of the sample ad image.

[0080] The position of each word refers to its location within the advertising material, and the position of each image block refers to its location within the advertising material.

[0081] In this implementation, the server can use different methods to extract features from sample accounts, advertising text, and advertising images through the advertising material recall model, thereby obtaining the account features of the sample account, the text features of the advertising text, and the image features of the advertising image, thus achieving accurate and efficient feature extraction.

[0082] To provide a clearer explanation of the above embodiments, the following description will be divided into three parts.

[0083] Part 1: The server uses the advertising material recall model to embed and encode the sample account, thereby obtaining the account characteristics of the sample account.

[0084] In one possible implementation, the server uses the account feature extraction unit of the advertising material recall model to embed and encode the sample account to obtain the account features of the sample account.

[0085] In the second part, the server uses the ad creative recall model to segment the sample ad text, obtaining multiple words within the text. Based on an attention mechanism, these words and their positions are encoded to obtain lexical features for each word. These lexical features are then fused to obtain the textual features of the sample ad text.

[0086] In one possible implementation, the server segments the sample advertisement text to obtain multiple words. The server then embeds and encodes these words and their positions to obtain a first fusion feature for each word. This first fusion feature includes both lexical embedding features and positional embedding features. The server inputs these first fusion features into the text feature extraction unit of the ad creative recall model. Based on the first fusion features, the text feature extraction unit determines the attention weight between every two words. The server then outputs the lexical features of each word based on the attention weights and the first fusion features. Finally, the server fuses these lexical features to obtain the text features of the sample advertisement text.

[0087] The text feature extraction unit is a semantic feature extraction model that uses an attention mechanism to extract features from the advertising text, thereby obtaining the semantic features of the input advertising text. In some embodiments, the text feature extraction unit is the encoder of a BERT (Bidirectional Encoder Representations from Transformers) model, or other variant models based on BERT, such as Roberta (A Robustly Optimized BERT Pretraining Approach), AlBERT (A Lite BERT), or TinyBERT. Of course, with the development of science and technology, other text feature extraction units with different structures can also be used, and this application embodiment does not limit this.

[0088] In this implementation, the sample advertisement text can be segmented to obtain multiple words. These multiple words and their positions are then encoded using an attention mechanism to obtain lexical features for each word. Because the attention mechanism utilizes information from multiple words during encoding, the resulting lexical features reflect the semantics of the corresponding words. The material features obtained based on these lexical features can thus represent the semantics of the sample advertisement text.

[0089] For example, the server segments the sample advertisement text into words, obtaining multiple words. The server then embeds and encodes these words and their positions to obtain a first fusion feature for each word. This first fusion feature includes both word embedding and position embedding features. The server inputs these first fusion features into a text feature extraction unit. Using the first, second, and third transformation matrices of this unit, the server performs a linear transformation on these features, obtaining the query matrix, key matrix, and value matrix for each word. Based on the query and key matrices, the server determines the attention weight between every two words in the sample advertisement text. The server then multiplies this attention weight with the value matrix of each word, outputting the lexical features of each word. These lexical features represent the semantics of the corresponding word. Finally, the server adds or concatenates these lexical features to obtain the text features of the sample advertisement text.

[0090] Part Three: The server uses the ad creative recall model to split the sample ad image into multiple image blocks. Based on an attention mechanism, these multiple image blocks and their positions are encoded to obtain the image block features of each image block. The image block features of these multiple image blocks are then fused to obtain the image features of the sample ad image.

[0091] In one possible implementation, the server splits the sample advertisement image into multiple image blocks. Embedding encoding is performed on these multiple image blocks and their respective positions to obtain a second fusion feature for each image block. This second fusion feature includes both image block embedding features and position embedding features for the corresponding image block. The server inputs the second fusion features of each image block into an image feature extraction unit. Based on the second fusion features, the image feature extraction unit determines the attention weight between every two image blocks. Using the same image feature extraction unit, and based on the attention weights between every two image blocks and the second fusion features, the server outputs the image block features of each image block. The server then fuses these image block features to obtain the image features of the sample advertisement image.

[0092] The image feature extraction unit is a semantic feature extraction model that uses an attention mechanism to extract features from the advertising image, thereby obtaining the semantic features of the input advertising image. In some embodiments, the image feature extraction unit is an encoder of the BERT model, or other variant models based on the BERT model, such as the VIT (Vision Transformer) model, the AlBERT model, or the TinyBERT model. Of course, with the development of science and technology, other image feature extraction units with different structures can also be used, such as VGG (Visual Geometry Group) or MobileNet, etc., and this application embodiment does not limit this.

[0093] For example, the server splits the sample advertisement image into multiple image blocks, obtaining multiple image blocks within the sample advertisement image. The server embeds and encodes the multiple image blocks and their positions to obtain the first fusion feature of each image block. This first fusion feature includes the image block embedding feature and the position embedding feature of the corresponding image block. The server inputs the first fusion feature of each image block into an image feature extraction unit. Through the first transformation matrix, second transformation matrix, and third transformation matrix of the image feature extraction unit, the first fusion feature of each image block is linearly transformed to obtain the query matrix, key matrix, and value matrix of each image block. Based on the query matrix and key matrix of each image block, the server determines the attention weight between every two image blocks in the multiple image blocks using the image feature extraction unit. The server multiplies the attention weight between every two image blocks in the multiple image blocks with the value matrix of each image block using the image feature extraction unit, outputting the image block feature of each image block. The image block feature of each image block is used to represent the semantics of the corresponding image block. The server adds or concatenates the image block features of the multiple image blocks to obtain the image feature of the sample advertisement image.

[0094] 304. The server trains the advertising material recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, the first difference information, and the second difference information. The first difference information is the difference information between the account features of the sample account and the text features of the sample advertising text, and the second difference information is the second difference information between the account features of the sample account and the image features of the sample advertising image.

[0095] In one possible implementation, the server constructs a joint loss function based on the first difference information and the second difference information. The server then trains the ad creative recall model based on the association between the sample account, the sample ad text, and the sample ad image, and the joint loss function.

[0096] The association between a sample account and the sample ad text and image includes the association between the sample account and the sample ad text, and the association between the sample account and the sample ad image. The purpose of training the ad creative recall model based on the first and second difference information is to ensure that the account features of the sample account are as close as possible to the text features of the associated sample ad text, and as far away as possible from the text features of the unassociated sample ad text; similarly, the account features of the sample account are as close as possible to the image features of the associated sample ad image, and as far away as possible from the image features of the unassociated sample ad image. This achieves alignment of the three types of features: account features, text features, and image features. Sample ad text associated with a sample account is also called positive sample ad text, and sample ads unassociated with a sample account are also called negative sample ad text. Correspondingly, sample ad images associated with a sample account are also called positive sample ad images, and sample ads unassociated with a sample account are also called negative sample ad images. The joint loss function includes the contrastive loss function corresponding to the first difference information and the contrastive loss function corresponding to the second difference information.

[0097] In this implementation, the server can train the advertising creative recall model based on the first difference information and the second difference information, so that the account features, text features and image features extracted by the advertising creative recall model are aligned, thereby providing the advertising creative recall model with the function of recalling advertising text or advertising images based on the account.

[0098] Optionally, before step 304, the server can also adjust the dimensions of account features, advertising features, and image features in the following ways, so as to train the advertising creative recall model based on the dimension-adjusted account features, advertising features, and image features.

[0099] In one possible implementation, the server adjusts the account features of the sample account, the text features of the sample advertisement text, and the image features of the sample advertisement image to the same dimension.

[0100] In this implementation, by adjusting the account features, advertising features, and image features to the same dimension, the server can directly determine the first difference information and the second difference information, thereby improving the efficiency of training the advertising creative recall model based on the first difference information and the second difference information.

[0101] For example, the server inputs the account features of the sample account into the first linear layer of the ad creative recall model, and adjusts the dimension of the account features to the target dimension through the first linear layer. The server inputs the text features of the sample ad text into the second linear layer of the ad creative recall model, and adjusts the dimension of the text features to the target dimension through the second linear layer. The server inputs the image features of the sample ad image into the third linear layer of the ad creative recall model, and adjusts the dimension of the image features to the target dimension through the third linear layer. The target dimension is set by the technician according to the actual situation, such as 312 dimensions, etc., and this embodiment does not limit this.

[0102] The following will combine Figure 4 The above steps 303 and 304 will be explained.

[0103] See Figure 4 The server inputs the sample account, sample ad text, and sample ad image into the ad creative recall model. The server extracts features from the sample account using the account feature extraction unit 401 of the ad creative recall model, obtaining the account features. The server then adjusts the dimension of the account features to the target dimension using the first linear layer 402 of the ad creative recall model. The server extracts features from the sample ad text using the text feature extraction unit 403 of the ad creative recall model, obtaining the text features. The server then adjusts the dimension of the text features to the target dimension using the second linear layer 404 of the ad creative recall model. The server extracts features from the sample ad image using the image feature extraction unit 405 of the ad creative recall model, obtaining the image features. The server then adjusts the dimension of the image features to the target dimension using the third linear layer 406 of the ad creative recall model. The server inputs the target dimension's account features and text features into the first contrastive loss function, and inputs the target dimension's account features and image features into the second contrastive loss function. The advertising creative recall model is trained using the joint loss function composed of the first and second contrastive loss functions, which is to train the account feature extraction unit 430 of the advertising creative recall model.

[0104] It should be noted that the above explanation is based on the example of the server performing one round of iterative training on the advertising material recall model. The methods used by the server in other iterative training rounds are based on the same inventive concept as those used in this round of iterative training, and the implementation process will not be elaborated here.

[0105] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0106] The technical solution provided in this application allows for the extraction of account features, text features, and image features from sample accounts, sample ad text, and sample ad images using an ad creative retrieval model. Based on the first difference information between account features and text features, and the second difference information between account features and image features, the ad creative retrieval model is trained to align these three types of features, enabling the ad creative retrieval model to retrieve ad creatives based on accounts.

[0107] The following explains how to use the advertising creative recall model trained using steps 301-304 above. (See [link to documentation]). Figure 5 Taking the server as the executing entity as an example, the method includes the following steps.

[0108] 501. The server inputs the target account into the trained advertising creative recall model.

[0109] The target account is the account from which the advertising creative is to be recalled.

[0110] 502. The server uses the advertising material recall model to extract features from the target account and obtain the account features of the target account.

[0111] The process by which the server extracts features from the target account using the advertising material recall model is the same inventive concept as described in step 303 above. The implementation process is described in step 303 above and will not be repeated here.

[0112] 503. The server uses the advertising material recall model to recall based on the account characteristics of the target account, and obtains the target advertising material. The target advertising material includes at least one of advertising text and advertising image, and the target advertising material matches the target account.

[0113] In one possible implementation, the server uses the ad creative recall model to determine the similarity between the account characteristics of the target account and the indexes of multiple candidate ad creatives, which include at least one of candidate ad text and candidate ad image. The server then identifies the candidate ad creative whose index matches the similarity condition between the target account and the account characteristics of the target account as the target ad creative.

[0114] The candidate ad creative index is used to represent the corresponding candidate ad creative, and is used for rapid recall during ad creative retrieval. In some embodiments, the candidate ad creative index is determined based on the creative characteristics of the candidate ad creative. In some embodiments, similarity meets the similarity condition when it is among the top N similarity creatives (Top N), or when the similarity is greater than or equal to a similarity threshold, where N is a positive integer and the similarity threshold is set by technical personnel according to the actual situation. The target ad creative is the ad creative to be recalled.

[0115] In this implementation, the server can quickly recall ads by indexing candidate ad creatives, resulting in high recall efficiency.

[0116] For example, the server uses the ad creative recall model to break down the account characteristics of the target account into multiple sub-features. For any candidate ad creative among these multiple candidate ad creatives, the server determines the similarity between the multiple sub-features of the target account and the index of the candidate ad creative. If the similarity between the index of the candidate ad creative and the multiple sub-features of the target account meets the similarity condition, the server identifies the candidate ad creative as the target ad creative.

[0117] For example, the index of candidate ad creatives includes multiple sub-indexes, and the server breaks down the account characteristics of the target account into multiple sub-features. For any candidate ad creative among these multiple candidate ad creatives, the server determines multiple reference similarities between the multiple sub-features of the target account and the multiple sub-indexes of the candidate ad creative. The server then fuses these multiple reference similarities to obtain the similarity between the multiple sub-features of the target account and the index of the candidate ad creative.

[0118] The number of sub-indexes and sub-features is the same. There is a correspondence between sub-features and sub-indexes. The correspondence is reflected by the relative position of the account feature and the index. For example, there is a correspondence between the first sub-feature in the account feature and the first sub-index in the index.

[0119] For example, the index of candidate ad creatives includes multiple sub-indexes. The server breaks down the account characteristics of the target account into multiple sub-features. For any candidate ad creative among these multiple candidate ad creatives, the server determines multiple reference similarities between each sub-feature of the target account and the corresponding sub-index in the multiple sub-indexes of the candidate ad creative. Each reference similarity is the similarity between a sub-feature and a sub-index. The server adds up these multiple reference similarities to obtain the similarity between the multiple sub-features of the target account and the index of the candidate ad creative, which is also the similarity between the target account and the candidate ad creative.

[0120] Each sub-index of the candidate ad creative is obtained through clustering. One sub-index may correspond to multiple candidate ad creatives. For example, the index of a candidate ad creative is ABCD, which includes four sub-indexes: A, B, C, and D. Sub-index A can correspond to other candidate ad creatives besides this candidate ad creative. By having multiple candidate ad creatives correspond to one sub-index, the amount of computation can be reduced and the computational efficiency can be improved.

[0121] To provide a clearer explanation of the above implementation methods, the method for determining the index of the plurality of candidate advertising materials is described below.

[0122] In one possible implementation, the server extracts features from the multiple candidate ad creatives to obtain the creative features of each candidate ad creative. The server then breaks down the creative features of each candidate ad creative into multiple sub-features. Based on these multiple sub-features, the server determines the index of each candidate ad creative.

[0123] It should be noted that the method by which the server determines the index of the multiple candidate ad creatives is before step 501, that is, before using the ad creative recall model. The index of the multiple candidate ad creatives has already been determined and can be used directly in step 503. This process of determining the index can be called an offline process.

[0124] In this implementation, the server breaks down the material features of candidate advertising materials into multiple sub-features, determines the index of the candidate advertising materials based on these sub-features, achieves dimensionality reduction of the material features, and performs queries based on the index at a faster speed.

[0125] For example, the server extracts features from the multiple candidate ad creatives to obtain the creative features of each candidate ad creative. The method by which the server extracts features from the multiple candidate ad creatives belongs to the same inventive concept as step 303 above, and the implementation process is described in step 303 above, and will not be repeated here. The server breaks down the creative features of each candidate ad creative into multiple sub-features. The server clusters the multiple sub-features of the multiple candidate ad creatives to obtain the index of each candidate ad creative.

[0126] The following describes the method by which the server clusters multiple sub-features of the multiple candidate ad creatives.

[0127] For the first sub-feature of the multiple candidate ad creatives, the server clusters these first sub-features to obtain multiple first cluster centers. The server uses the first cluster center corresponding to the first sub-feature of each candidate ad creative as the first sub-index of each candidate ad creative. This process is repeated to obtain multiple sub-indexes for each candidate ad creative, which together constitute the index of each candidate ad creative.

[0128] For example, let's take the indexing of candidate ad creatives as an example, using the Facebook AI Similarity Search index. See [link to documentation]. Figure 6 The server inputs candidate ad text into the text feature extraction unit of the ad creative recall model to obtain candidate ad text features. This text feature extraction unit encodes the input candidate ad text into a 312-dimensional feature vector, which is also the creative feature of the candidate ad creative. Based on the candidate ad text features, the server generates the index (TEXT feature) of the candidate ad text. Similarly, the server inputs candidate ad images into the image feature extraction unit of the ad creative recall model to obtain candidate ad image features. This image feature extraction unit encodes the input candidate ad image into a 312-dimensional feature vector, which is also the creative feature of the candidate ad creative. Based on the candidate ad image features, the server generates the index (IMAGE feature) of the candidate ad image.

[0129] The following will combine Figure 7 The above steps 501-503 will be explained.

[0130] See Figure 7 The server inputs the target account into the ad creative recall model, which extracts features from the target account to obtain its account characteristics. Based on these account characteristics, the server uses the ad creative recall model to determine the N target ad images with the highest similarity to the account characteristics from the multiple candidate ad images, and the N target ad texts with the highest similarity to the account characteristics from the multiple candidate ad texts.

[0131] Figure 8 This is a schematic diagram of the structure of a training device for an advertising creative recall model provided in an embodiment of this application. See also... Figure 8 The device includes: an input module 801, a feature extraction module 802, and a training module 803.

[0132] Input module 801 is used to input sample account, sample advertising text and sample advertising image into advertising creative recall model, which is used to recall at least one of advertising text and advertising image based on account.

[0133] The feature extraction module 802 is used to extract features from the sample account, the sample advertisement text, and the sample advertisement image through the advertising material recall model, so as to obtain the account features of the sample account, the text features of the sample advertisement text, and the image features of the sample advertisement image.

[0134] Training module 803 trains the advertising creative recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, as well as first and second difference information. The first difference information is the difference between the account features of the sample account and the text features of the sample advertising text, and the second difference information is the difference between the account features of the sample account and the image features of the sample advertising image. In another aspect, a computer device is provided, comprising one or more processors and one or more memories, wherein at least one computer program is stored in the one or more memories, and the computer program is loaded and executed by the one or more processors to implement the training method for the advertising creative recall model.

[0135] In one possible implementation, the feature extraction module 802 is used to embed and encode the sample account using the advertising creative recall model to obtain the account features of the sample account. The advertising creative recall model is then used to segment the sample advertising text to obtain multiple words within the text. These multiple words and their positions are encoded using an attention mechanism to obtain the lexical features of each word. The lexical features of these multiple words are then fused to obtain the text features of the sample advertising text. Finally, the advertising image is split into multiple image blocks using the advertising creative recall model. These multiple image blocks and their positions are encoded using an attention mechanism to obtain the image block features of each image block. The image block features of these multiple image blocks are then fused to obtain the image features of the advertising image.

[0136] In one possible implementation, the feature extraction module 802 is used to embed and encode the multiple words and their positions to obtain a first fusion feature for each word. This first fusion feature includes both word embedding features and position embedding features. The first fusion features of each word are input into the text encoding unit of the advertising creative recall model. Based on the first fusion features, the text encoding unit determines the attention weight between every two words in the multiple words. Finally, based on the attention weights between every two words and the first fusion features, the text encoding unit outputs the word features of each word.

[0137] In one possible implementation, the feature extraction module 802 is used to embed and encode the plurality of image blocks and the position of each image block to obtain a second fusion feature for each image block. The second fusion feature includes an image block embedding feature and a position embedding feature for the corresponding image block. The second fusion feature of each image block is input into the image encoding unit of the advertising creative recall model. The image encoding unit determines the attention weight between every two image blocks in the plurality of image blocks based on the second fusion feature. The image encoding unit outputs the image block feature of each image block based on the attention weight between every two image blocks and the second fusion feature of each image block.

[0138] In one possible implementation, the training module 803 is used to construct a joint loss function based on the first difference information and the second difference information. The ad creative recall model is then trained based on the association between the sample account, the sample ad text, and the sample ad image, and the joint loss function.

[0139] In one possible implementation, the device further includes:

[0140] The dimension adjustment module is used to adjust the account features of the sample account, the text features of the sample advertisement text, and the image features of the sample advertisement image to the same dimension.

[0141] In one possible implementation, the device further includes:

[0142] The recall module is used to input the target account into the trained ad creative recall model. The ad creative recall model extracts features from the target account to obtain its account characteristics. Based on these account characteristics, the ad creative recall model retrieves target ad creatives, which include at least one of ad text and ad image, and these target ad creatives match the target account.

[0143] In one possible implementation, the recall module is used to determine the similarity between the account characteristics of the target account and the indexes of multiple candidate ad creatives through the ad creative recall model. The multiple candidate ad creatives include at least one of candidate ad text and candidate ad image. The candidate ad creatives whose indexes match the similarity criteria between the indexes and the account characteristics of the target account are identified as the target ad creative.

[0144] In one possible implementation, the device further includes:

[0145] The index determination module is used to input the multiple candidate ad creatives into the ad creative recall model. The ad creative recall model extracts features from the multiple candidate ad creatives to obtain the creative features of each candidate ad creative. The creative features of each candidate ad creative are then broken down into multiple sub-features. Based on these multiple sub-features, the index of each candidate ad creative is determined.

[0146] In one possible implementation, the index determination module is used to cluster multiple sub-features of the multiple candidate advertising materials to obtain the index of each candidate advertising material.

[0147] It should be noted that the training device for the advertising creative recall model provided in the above embodiments is only illustrated by the division of the above functional modules when training the advertising creative recall model. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the training device for the advertising creative recall model provided in the above embodiments and the training method embodiment for the advertising creative recall model belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0148] The technical solution provided in this application allows for the extraction of account features, text features, and image features from sample accounts, sample ad text, and sample ad images using an ad creative retrieval model. Based on the first difference information between account features and text features, and the second difference information between account features and image features, the ad creative retrieval model is trained to align these three types of features, enabling the ad creative retrieval model to retrieve ad creatives based on accounts.

[0149] The aforementioned computer equipment can be implemented as a server. The structure of a server is described below:

[0150] Figure 9This is a schematic diagram of a server structure provided in an embodiment of this application. The server 900 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 901 and one or more memories 902. The one or more memories 902 store at least one computer program, which is loaded and executed by the one or more processors 901 to implement the methods provided in the various method embodiments described above. Of course, the server 900 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server 900 may also include other components for implementing device functions, which will not be elaborated upon here.

[0151] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program that can be executed by a processor to complete the training method for the advertising creative recall model in the above embodiments. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0152] In an exemplary embodiment, a computer program product or computer program is also provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to perform the training method of the advertising material recall model described above.

[0153] In some embodiments, the computer program involved in the present application embodiments may be deployed and executed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0154] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0155] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A training method for an advertising creative recall model, characterized in that, The method includes: The sample account, sample ad text, and sample ad image are input into the ad creative recall model, which is used to recall at least one of the ad text and ad image based on the account. The advertising material recall model extracts features from the sample account, the sample advertising text, and the sample advertising image to obtain the account features of the sample account, the text features of the sample advertising text, and the image features of the sample advertising image. Based on the association between the sample account, the sample advertisement text, and the sample advertisement image, as well as the first difference information and the second difference information, the advertisement material recall model is trained. The first difference information is the difference information between the account features of the sample account and the text features of the sample advertisement text, and the second difference information is the second difference information between the account features of the sample account and the image features of the sample advertisement image. The step of training the advertising creative recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, as well as the first difference information and the second difference information, includes: A joint loss function is constructed based on the first difference information and the second difference information. The advertising creative recall model is trained based on the association between the sample account, the sample advertising text, and the sample advertising image, and the joint loss function. The joint loss function includes a contrastive loss function corresponding to the first difference information and a contrastive loss function corresponding to the second difference information.

2. The method according to claim 1, characterized in that, The step of extracting features from the sample account, the sample ad text, and the sample ad image using the ad creative recall model to obtain account features of the sample account, text features of the sample ad text, and image features of the sample ad image includes: The sample accounts are embedded and encoded using the advertising material recall model to obtain their account characteristics. The advertising material recall model is used to segment the sample advertising text to obtain multiple words in the sample advertising text; the multiple words and their positions are encoded based on an attention mechanism to obtain the lexical features of each word; the lexical features of the multiple words are fused to obtain the text features of the sample advertising text. The advertising image is divided into multiple image blocks using the advertising creative recall model; the multiple image blocks and their positions are encoded based on an attention mechanism to obtain the image block features of each image block; the image block features of the multiple image blocks are fused to obtain the image features of the advertising image.

3. The method according to claim 2, characterized in that, The encoding of the multiple words and their positions based on the attention mechanism to obtain the lexical features of each word includes: Embedding encoding is performed on the plurality of words and the position of each word to obtain a first fusion feature of each word. The first fusion feature includes the word embedding feature and the position embedding feature of the corresponding word. The first fusion feature of each of the words is input into the text encoding unit of the advertising creative recall model. The attention weight between each pair of words is determined by the text encoding unit based on the first fusion feature of each of the words. The text encoding unit outputs the lexical features of each word based on the attention weight between every two words and the first fusion feature of each word.

4. The method according to claim 2, characterized in that, The process of encoding the multiple image blocks and the position of each image block based on the attention mechanism to obtain the image block features of each image block includes: Embedding encoding is performed on the plurality of image blocks and the position of each image block to obtain a second fusion feature of each image block. The second fusion feature includes the image block embedding feature and the position embedding feature of the corresponding image block. The second fusion feature of each of the image blocks is input into the image encoding unit of the advertising creative recall model. The image encoding unit determines the attention weight between every two image blocks based on the second fusion feature of each of the image blocks. The image encoding unit outputs image block features for each image block based on the attention weights between every two image blocks and the second fusion features of each image block.

5. The method according to claim 1, characterized in that, The step of training the advertising material recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, as well as the first difference information and the second difference information, includes: A joint loss function is constructed based on the first difference information and the second difference information; The advertising creative recall model is trained based on the association between the sample account, the sample advertising text, and the sample advertising image, and the joint loss function.

6. The method according to claim 1, characterized in that, Before training the advertising creative recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, as well as the first difference information and the second difference information, the method further includes: The account features of the sample account, the text features of the sample advertisement text, and the image features of the sample advertisement image are adjusted to the same dimension.

7. The method according to claim 1, characterized in that, The method further includes: Input the target account into the trained advertising creative retrieval model; The target account's features are extracted using the advertising material recall model to obtain the target account's characteristics; The advertising material recall model is used to recall target advertising materials based on the account characteristics of the target account. The target advertising materials include at least one of advertising text and advertising image, and the target advertising materials are matched with the target account.

8. The method according to claim 7, characterized in that, The process of recalling target advertising materials using the advertising material recall model, based on the account characteristics of the target account, yields the following target advertising materials: The advertising creative recall model determines the similarity between the account characteristics of the target account and the indexes of multiple candidate advertising creatives, wherein the multiple candidate advertising creatives include at least one of candidate advertising text and candidate advertising image; The candidate ad creatives whose indexes and account characteristics match the similarity criteria of the target account are identified as the target ad creatives.

9. The method according to claim 8, characterized in that, The method for determining the index of the multiple candidate advertising creatives includes: Input the multiple candidate ad creatives into the ad creative recall model; The ad creative recall model is used to extract features from the multiple candidate ad creatives to obtain the creative features of each candidate ad creative. The material features of each candidate advertising material are broken down into multiple sub-features; The index of each candidate ad creative is determined based on multiple sub-features of each candidate ad creative.

10. The method according to claim 9, characterized in that, The process of determining the index of each candidate ad creative based on multiple sub-features includes: Clustering is performed on multiple sub-features of the candidate advertising materials to obtain the index of each candidate advertising material.

11. A training device for an advertising creative recall model, characterized in that, The device includes: The input module is used to input sample accounts, sample advertising text, and sample advertising images into the advertising creative recall model, wherein the advertising creative recall model is used to recall at least one of the advertising text and advertising images based on the account. The feature extraction module is used to extract features from the sample account, the sample advertisement text, and the sample advertisement image through the advertisement material recall model, respectively, to obtain the account features of the sample account, the text features of the sample advertisement text, and the image features of the sample advertisement image; The training module is used to train the advertising material recall model based on the association between the sample account, the sample advertising text, and the sample advertising image, as well as the first difference information and the second difference information. The first difference information is the difference information between the account features of the sample account and the text features of the sample advertising text, and the second difference information is the second difference information between the account features of the sample account and the image features of the sample advertising image. The training module is specifically used to: construct a joint loss function based on the first difference information and the second difference information, and train the advertising material recall model based on the association between the sample account, the sample advertising text, and the sample advertising image and the joint loss function; the joint loss function includes a contrastive loss function corresponding to the first difference information and a contrastive loss function corresponding to the second difference information.

12. A computer device, characterized in that, The computer device includes one or more processors and one or more memories, wherein at least one computer program is stored in the one or more memories, and the computer program is loaded and executed by the one or more processors to implement the training method of the advertising creative recall model as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the training method for the advertising creative recall model as described in any one of claims 1 to 10.

Images