Advertisement material classification method and device based on deep learning

Abstract

The application provides an advertisement material classification method and device based on deep learning, which comprises the following steps: obtaining low-order features of advertisement materials, wherein the low-order features comprise visual features and text features of the advertisement materials; obtaining high-order features of the advertisement materials, wherein the high-order features represent the correlation features between the advertisement materials and other advertisement materials; and determining the category to which the advertisement materials belong according to the low-order features and the high-order features of the advertisement materials. The advertisement material classification method based on deep learning has significant advantages starting from the whole data processing process. The method fuses visual and text low-order features, avoids one-sidedness of single-modal classification, automatically extracts features by using deep learning, breaks through the limitation of manual annotation, adapts to dynamic changes of materials, obtains high-order correlation features, excavates the relationship between materials, helps popular trend identification and prediction of delivery effect, and comprehensively improves classification accuracy, generalization ability and practicability.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing, and in particular to a method and apparatus for classifying advertising materials based on deep learning. Background Technology

[0002] With the rapid development of the global digital marketing industry, the quantity and types of advertising creatives are growing exponentially, covering various formats such as images, videos, and copy. In overseas marketing scenarios, advertisers need to deliver diverse advertising creatives tailored to different regions, cultures, and audiences. Therefore, efficient and accurate classification and management of massive amounts of advertising creatives has become a key aspect of improving campaign efficiency and optimizing marketing strategies. Precise classification of advertising creatives can help businesses quickly locate target creatives, enabling intelligent recommendations and personalized delivery, thereby improving ad conversion rates and user engagement.

[0003] Existing advertising creative classification technologies primarily employ unimodal analysis or simple machine learning methods. On one hand, traditional solutions often rely on manual annotation or rule matching for classification, filtering advertising creatives using pre-defined keywords and tag templates, such as identifying ad types based on specific words in the copy. On the other hand, some technologies only utilize shallow machine learning models to analyze single visual or textual features of advertising creatives, such as using traditional image classification algorithms to identify product categories in ad images. These methods typically treat advertising creatives as independent entities, lacking the ability to uncover potential relationships between creatives.

[0004] However, existing technologies have significant limitations: First, single-modal analysis cannot comprehensively utilize the multi-dimensional information of advertising creatives. For example, relying solely on text features may miss key information conveyed by visual elements, leading to biased classification results. Second, manual annotation or rule-based matching methods are difficult to adapt to the dynamic changes in advertising creatives, and are prone to classification errors or omissions when faced with emerging advertising formats or semantically ambiguous copy. Third, existing solutions ignore the correlation between advertising creatives, failing to capture common features of similar creatives or potential connections between different creatives, resulting in an inability to effectively identify popular trends or predict the effectiveness of creative campaigns. These shortcomings make existing technologies suffer from low classification efficiency, insufficient accuracy, and poor generalization ability when processing large-scale, multimodal advertising creatives, making it difficult to meet the needs of the digital marketing field for refined management and intelligent analysis of creatives. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and apparatus for classifying advertising materials based on deep learning, so as to at least partially solve the above problems.

[0006] A deep learning-based method for classifying advertising creatives, comprising:

[0007] Obtain low-level features of the advertising creative, including the visual features and text features of the advertising creative itself;

[0008] Obtain high-level features of the advertising creative, wherein the high-level features represent the correlation features between the advertising creative and other advertising creatives;

[0009] The category to which the advertising material belongs is determined based on its low-level and high-level features.

[0010] Optionally, the visual features include at least one of the following: color distribution features, compositional structure features, and subject object features; the text features include: keyword features and semantic sentiment features.

[0011] Optionally, obtaining the low-level features of the advertising material includes: extracting the visual features of the advertising material using a convolutional neural network (CNN) and extracting the text features of the advertising material using a natural language processing (NLP) model.

[0012] Optionally, the NLP model includes: a BERT model based on the Transformer architecture.

[0013] Optionally, the higher-order features include at least one of the following: the similarity features of the advertising material among the materials deployed in the same batch, and the correlation features between the advertising material and historically popular materials.

[0014] Optionally, obtaining the high-order features of the advertising material includes at least one of the following: calculating the similarity features of the advertising material using a Siamese neural network based on the low-order features of the advertising material and the low-order features of other advertising materials; or obtaining the correlation features of the advertising material using a graph neural network based on the low-order features of the advertising material and the features of historically popular materials.

[0015] Optionally, determining the category to which the advertising material belongs based on its low-order and high-order features includes: concatenating the low-order and high-order features of the advertising material to form the total features of the advertising material; inputting the total features into a deep learning-based classifier; and determining the category to which the advertising material belongs based on the output of the classifier.

[0016] Optionally, the deep learning-based classifier is a hybrid model combining a multilayer perceptron or random forest with deep learning.

[0017] A deep learning-based advertising material classification device, comprising:

[0018] The first acquisition module is adapted to acquire low-level features of advertising materials, wherein the low-level features include the visual features and text features of the advertising materials themselves;

[0019] The second acquisition module is adapted to acquire high-level features of advertising materials, wherein the high-level features represent the correlation features between the advertising materials and other advertising materials;

[0020] The processing module is adapted to determine the category to which the advertising material belongs based on its low-level and high-level features.

[0021] Optionally, the visual features include at least one of the following: color distribution features, compositional structure features, and subject object features; the text features include: keyword features and semantic sentiment features; the higher-order features include at least one of the following: similarity features of the advertising material among materials released in the same batch, and correlation features between the advertising material and historically popular materials.

[0022] Optionally, the first acquisition module includes: a visual feature extraction unit, which uses a convolutional neural network to extract the visual features of the advertising material; and a text feature extraction unit, which uses a BERT model based on the Transformer architecture to extract the text features of the advertising material.

[0023] Optionally, the second acquisition module includes: a similarity calculation unit, which calculates the similarity features of the advertising material with other advertising materials through a Siamese neural network; and a correlation analysis unit, which obtains the correlation features of the advertising material with historical popular materials through a graph neural network.

[0024] Optionally, the processing module includes: a feature fusion unit, which concatenates the low-order and high-order features of the advertising material to form a total feature; and a classification decision unit, which inputs the total feature into a deep learning-based classifier and determines the category to which the advertising material belongs based on the output result.

[0025] The deep learning-based advertising creative classification method proposed in this application provides a systematic solution for the entire data processing process, and has the following technical advantages:

[0026] 1. By acquiring the visual and textual features of the advertising materials themselves as low-level features, the current situation of relying solely on information from a single modality is changed. By comprehensively utilizing multi-dimensional data, the classification bias caused by relying solely on textual or visual information is avoided, making the classification results more comprehensive and accurate.

[0027] 2. It uses deep learning to automatically acquire low-order and high-order features, eliminating the need to pre-set a large number of keywords and tag templates. It can automatically adapt to the dynamic changes of advertising materials, effectively solving the problem of errors or omissions in manual annotation and rule matching when facing emerging advertising formats or ambiguous copy.

[0028] 3. By acquiring high-level features that indicate the relationship between advertising creatives and other creatives, we can overcome the limitations of existing technologies that treat creatives as independent entities. This allows us to capture the commonalities of similar creatives and the potential connections between different creatives, which helps to identify popular trends, provides data support for predicting the effectiveness of creative campaigns, and improves the generalization ability and practicality of classification. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0030] Figure 1 This is a flowchart illustrating a deep learning-based advertising material classification method according to an embodiment of this application.

[0031] Figure 2 This invention provides a schematic diagram of the structure of an advertising material classification device based on deep learning. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0033] It should be understood that the terms "first," "second," and "third," etc., in the claims, specification, and drawings of this disclosure are used to distinguish different objects, not to describe a specific order. The terms "comprising" and "including" as used in the specification and claims of this disclosure indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or sets thereof.

[0034] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any and all combinations of one or more of the associated listed items, and includes such combinations.

[0035] Figure 1 This is a flowchart illustrating a deep learning-based advertising material classification method according to an embodiment of this application. Figure 1 As shown, a deep learning-based method for classifying advertising creatives includes:

[0036] Obtain low-level features of the advertising creative, including the visual features and text features of the advertising creative itself;

[0037] Obtain high-level features of the advertising creative, wherein the high-level features represent the correlation features between the advertising creative and other advertising creatives;

[0038] The category to which the advertising material belongs is determined based on its low-level and high-level features.

[0039] Optionally, the visual features include at least one of the following: color distribution features, compositional structure features, and subject object features; the text features include: keyword features and semantic sentiment features.

[0040] Optionally, obtaining the low-level features of the advertising material includes: extracting the visual features of the advertising material using a convolutional neural network (CNN) and extracting the text features of the advertising material using a natural language processing (NLP) model.

[0041] Optionally, the NLP model includes: a BERT model based on the Transformer architecture.

[0042] Optionally, the higher-order features include at least one of the following: the similarity features of the advertising material among the materials deployed in the same batch, and the correlation features between the advertising material and historically popular materials.

[0043] Preferably, when performing parallel extraction of multimodal low-order features, for advertising materials (covering multiple forms of data such as images, videos, and text), a convolutional neural network (CNN) is used to perform convolution operations on the visual part of the advertising materials to extract color distribution features, composition structure features, and subject object features; a BERT model based on the Transformer architecture is used to perform semantic encoding on the text part of the advertising materials to extract keyword features and semantic sentiment features, thereby obtaining a low-order feature set of advertising materials (a feature set integrating visual and textual multimodal information).

[0044] Preferably, when constructing high-order correlation features across materials, for the low-order feature set of advertising materials and the low-order feature data of advertising materials in the same batch / historical batch, the cosine similarity of the low-order features of the current advertising material and the materials in the same batch is calculated by using a Siamese neural network to generate similarity features; and a graph neural network is used to construct a correlation graph between the low-order features of the current advertising material and the features of historical popular materials to mine correlation features in order to obtain the high-order feature set of advertising materials (feature combinations that represent the correlation between materials).

[0045] Preferably, when making feature fusion classification decisions, the low-order feature set and the high-order feature set of the advertising material are dimensionally concatenated to form a complete advertising material feature vector; the feature vector is input into the multilayer perceptron (MLP) classification model, and the probability of each preset category is output through the Softmax function. The category with the highest probability is selected as the final classification result to obtain the advertising material classification result (a judgment conclusion that clearly belongs to a certain advertising type or label).

[0046] Optionally, obtaining the high-order features of the advertising material includes at least one of the following: calculating the similarity features of the advertising material using a Siamese neural network based on the low-order features of the advertising material and the low-order features of other advertising materials; or obtaining the correlation features of the advertising material using a graph neural network based on the low-order features of the advertising material and the features of historically popular materials.

[0047] Preferably, when calculating similarity features based on Siamese neural networks, for the low-order features of the current advertising creative and the low-order features of other advertising creatives in the same batch, the low-order features of the current advertising creative and the low-order features of other advertising creatives in the same batch are respectively input into two branches of the Siamese neural network, and feature vectors are extracted through structures such as convolutional layers and pooling layers; the cosine similarity or Euclidean distance between the two feature vectors is calculated to obtain the similarity value of the current advertising creative in the same batch of creatives; the similarity value is compared with a preset threshold, and the results with high similarity are selected to obtain a set of similarity features (containing the similarity feature data of the current advertising creative and other creatives in the same batch).

[0048] Preferably, when acquiring correlation features based on graph neural networks, a graph structure is constructed for the low-order features of the current advertising creative and the features of historical popular creatives. The current advertising creative and historical popular creatives are used as nodes in the graph, and the node attributes are their respective feature data. The graph structure is processed by a graph neural network (such as GNN or GAT) to perform message passing and feature aggregation, and the correlation between nodes is learned. The correlation features between the current advertising creative and historical popular creatives are output, such as quantitative indicators like influence weight and correlation strength, thereby obtaining a set of correlation features (feature data representing the correlation between the current advertising creative and historical popular creatives).

[0049] When performing high-order feature integration, the similarity feature set and the correlation feature set are merged to form a unified high-order feature vector. According to the needs of different application scenarios, the merged feature vector is weighted to highlight key correlation information, thereby obtaining a complete high-order feature set (including comprehensive feature data on the relationship between advertising creatives and other creatives).

[0050] Optionally, determining the category to which the advertising material belongs based on its low-order and high-order features includes: concatenating the low-order and high-order features of the advertising material to form the total features of the advertising material; inputting the total features into a deep learning-based classifier; and determining the category to which the advertising material belongs based on the output of the classifier.

[0051] Preferably, during the multi-dimensional feature fusion and splicing, for the low-order features (visual features, text features) and high-order features (similarity features, correlation features) of the advertising material, the low-order features and high-order features are aligned in dimensions and then spliced ​​along the feature axis to form a composite feature vector that includes the material's own attributes and correlation relationships; according to the importance of features, different modal features are weighted (e.g., visual features weight 0.4, text features weight 0.3, high-order features weight 0.3) to obtain the total feature vector of the advertising material (a complete feature representation that integrates multi-dimensional information).

[0052] Preferably, when making inference decisions using a deep classification model, the total feature vector of the advertising material is input into a deep learning-based classifier (such as a multilayer perceptron MLP or Transformer classification model), and the probability distribution of each category is calculated through forward propagation. The output result is normalized using the softmax function, and the category with the highest probability is selected as the prediction result, thereby obtaining the classification result of the advertising material (clearly defining the label and confidence level of the category).

[0053] Optionally, the deep learning-based classifier is a hybrid model combining a multilayer perceptron or random forest with deep learning.

[0054] Preferably, during the multilayer perceptron (MLP) classification process, the total feature vector (a composite feature vector resulting from concatenating low-order and high-order features) of the advertising material is input into the MLP model, and the features are nonlinearly transformed through multiple fully connected layers. Activation functions (such as ReLU) are used in the hidden layers to enhance the model's expressive power, and the softmax function is used in the output layer to calculate the probability values ​​of the advertising material belonging to each category. The category with the highest probability value is selected as the preliminary classification result, thus obtaining the preliminary MLP classification result (including data on the probability of each category and the predicted category).

[0055] Preferably, in the fusion processing of the random forest and deep learning hybrid model, based on the preliminary classification results of the MLP and the original total feature vector, the probability value output by the MLP is combined with the original total feature vector as the input of the hybrid model; the random forest part performs tree-structure-based rule partitioning of features, explores the nonlinear relationships and decision boundaries between features, and the deep learning part (such as CNN or Transformer) further extracts the abstract semantics of the features; the output results of the two parts are weighted and fused through the model fusion layer, and the probability of each classification category is recalculated; finally, the fused advertising material classification result is output, thus obtaining the final classification result of the advertising material (the clear and optimized category label and confidence score of the advertising material).

[0056] Figure 2 An embodiment of this application provides an advertising material classification device based on deep learning, comprising:

[0057] The first acquisition module is adapted to acquire low-level features of advertising materials, wherein the low-level features include the visual features and text features of the advertising materials themselves;

[0058] The second acquisition module is adapted to acquire high-level features of advertising materials, wherein the high-level features represent the correlation features between the advertising materials and other advertising materials;

[0059] The processing module is adapted to determine the category to which the advertising material belongs based on its low-level and high-level features.

[0060] Optionally, the visual features include at least one of the following: color distribution features, compositional structure features, and subject object features; the text features include: keyword features and semantic sentiment features; the higher-order features include at least one of the following: similarity features of the advertising material among materials released in the same batch, and correlation features between the advertising material and historically popular materials.

[0061] Preferably, when extracting low-order features in multimodal modes, the first acquisition module can achieve collaborative extraction of visual and textual features through a cross-modal aligned feature fusion network. The specific technical implementation is as follows:

[0062] An improved ResNet-50 architecture is employed to extract multi-scale features at different convolutional layers (e.g., conv2_x to conv5_x). Shallow features (e.g., conv2_x) preserve color distribution and edge texture information, while deep features (e.g., conv5_x) capture the main object and abstract semantics. Feature fusion is performed through a Feature Pyramid Network (FPN): high-level features are upsampled and added element-wise with low-level features to form a feature pyramid containing information at different scales. A Convolutional Block Attention Module (CBAM) is added after each residual block, comprising two sub-modules: channel attention and spatial attention. Channel attention compresses the spatial dimension of the feature map through global average pooling and global max pooling, and generates channel weights via a multilayer perceptron; spatial attention compresses the channel dimension and generates spatial weights through convolution. The multiplication of the two modules achieves feature enhancement.

[0063] Preferably, during keyword extraction, a pre-trained BERT-base model is used. The advertising copy is segmented into words by a tokenizer and input into the model. The output of the [CLS] tag is taken as the overall semantic representation. A bidirectional LSTM network is used to process the token representation output by BERT. The hidden states at each time step are weighted and summed through an attention layer, and finally, the sentiment polarity (positive / negative / neutral) and sentiment intensity are output through a fully connected layer. At the same time, word-level features (such as the vector representation of each token), sentence-level features (such as the [CLS] vector), and document-level features (such as the sentiment analysis results) are extracted and dynamically fused through a gating mechanism.

[0064] Preferably, this application uses a visual encoder (based on ResNet) and a text encoder (based on BERT) to map two types of features to the same semantic space through adversarial training. The discriminator distinguishes the feature source (visual or textual), and the encoder learns cross-modal invariant representations by minimizing the discriminator's accuracy. Predefined semantic anchors in the marketing domain (such as "promotion," "high-end," and "environmentally friendly") are used to align visual and textual features to these anchors through contrastive learning. For example, when a visual image shows a red background with a discount label, and the text contains "limited-time offer," the association between both features and the "promotion" anchor is strengthened.

[0065] Preferably, the second acquisition module models the complex relationships between materials using a Dynamic Graph Neural Network (Dynamic GNN), and the specific technical implementation is as follows:

[0066] The basic network uses Siamese ResNet. Visual features of two ad creatives are input, and visual similarity is calculated using cosine similarity. A temporal embedding vector is introduced: the ad creative's placement time is encoded as a vector, concatenated with the visual features, and then input into the Siamese network, allowing the similarity calculation to consider the time factor. A time decay function is calculated to attenuate the similarity score of historical creatives, using the formula `score = original_score × exp(-λ×Δt)`, where λ is the decay coefficient and Δt is the time difference. Sentence-BERT is then used to calculate textual semantic similarity. The ad copy is first encoded as a fixed-length vector, and then cosine similarity is calculated. The Word Mover's Distance (WMD) algorithm is combined to consider the semantic distance between words, solving the matching challenges posed by synonyms and near-synonyms.

[0067] A heterogeneous graph network is introduced, with node types including: content nodes (containing low-level features), user nodes (containing user profile information), and effect nodes (such as click-through rate and conversion rate). Edge types include: content-content (similarity), content-user (targeting relationship), and user-effect (interaction relationship). Edge weights are dynamically learned through an attention mechanism; for example, the edge weights between content are determined by both similarity scores and time decay. Furthermore, a Graph Attention Network (GAT) is used to process the heterogeneous graph, where each node aggregates information from its neighbors through an attention mechanism. When new content is added, only the affected local graph structure is updated, and model parameters are quickly adjusted through incremental GNN training.

[0068] In this application, similarity features (scalar values) and correlation features (graph embedding vectors) are fused using a multilayer perceptron to output a final high-order feature vector. Then, L1 regularization is used to automatically filter important features and reduce redundant information.

[0069] In this application, the processing module employs an Adaptive Mixture of Experts (AMoE) model to achieve intelligent classification. The specific technical implementation is as follows:

[0070] Taking the feature statistics of the input material (such as visual complexity and text length) as input, the output is the fusion weight of low-order and high-order features. Visual complexity is evaluated using image entropy and edge density, while text complexity is evaluated using word frequency distribution and sentence length. A Transformer layer is designed, with low-order features as the query and high-order features as the key and value, respectively. A self-attention mechanism is used to explore the correlation between the two types of features. Residual connections and layer normalization are added to ensure smooth information flow.

[0071] Additionally, based on the CNN architecture, including a ResNet backbone and multiple fully connected layers, it focuses on processing visually driven ads. Based on the BERT architecture, a classification head is added to the pre-trained model, making it adept at handling text-driven ads. Based on the GNN architecture, the embedding representation of the content in the graph is input into a multilayer perceptron to mine relational value.

[0072] Configure reinforcement learning gating networks:

[0073] The state of the advertising creative classification device based on deep learning: the feature vector of the current creative and the historical classification results.

[0074] Deep learning-based advertising creative classification device action: expert weight allocation vector.

[0075] Rewards for a deep learning-based ad creative classification device: a weighted sum of classification accuracy and inference efficiency.

[0076] The deep learning-based advertising creative classification device trains a gating network using the Proximal Policy Optimization (PPO) algorithm, enabling it to dynamically assign expert weights based on input characteristics.

[0077] This application introduces a temperature scaling technique: the logits output by the classifier are scaled by temperature, using the formula `scaled_logits = logits / T`, where T is the temperature parameter. This scaling is learned by minimizing the cross-entropy loss on the validation set. Platt scaling or Isotonic regression is then used to map the classification probabilities to more accurate confidence values.

[0078] In this application, FPN is used to fuse features at different levels, while preserving shallow color / texture details and deep semantic information, thereby improving the ability to recognize small elements in advertisements (such as logos and price tags). The CBAM module dynamically adjusts channel and spatial attention weights, enabling the model to automatically focus on key areas (such as the product body and promotional text) and suppress background noise interference.

[0079] In this application, BERT extracts keyword features while BiLSTM+Attention network captures the sentiment of the copy, comprehensively understanding the advertising intent. For example, it distinguishes the different semantics of "limited-time offer" (promotion) and "quality assurance" (brand). The dynamic fusion of word-level, sentence-level, and document-level features adapts to advertising copy of different lengths and structures. For short copy, word-level features are emphasized, while for long copy, document-level semantic understanding is strengthened.

[0080] In this application, adversarial training is used to map visual and textual features to a shared semantic space, addressing the modal discrepancy problem. For example, the visual feature of "red background" and the textual features of "festive" and "promotion" form a strong association in the semantic space. Predefined marketing domain anchors (such as "promotion" and "high-end") are used, and contrastive learning is employed to strengthen the alignment between features and anchors, thereby improving cross-modal semantic consistency.

[0081] In this application, the temporal embedding vector and decay function enable the model to prioritize recently similar materials, adapting to the time-sensitive nature of the advertising industry. For example, summer ads are more similar to ads from the same period last year than winter ads. Sentence-BERT, combined with the WMD algorithm, considers both textual semantic similarity and semantic distance between words, accurately identifying synonyms (such as "special offer" and "discount") and near-synonyms (such as "high-end" and "luxury").

[0082] This application comprehensively captures the complex relationships within the advertising ecosystem across three node types: creatives, users, and performance. For example, it uncovers target audience preferences by following the path of "creatives → high-converting users → high-click-through-rate performance." When new creatives are added, only a partial graph structure is updated, significantly improving system response speed. For instance, newly launched popular ads are integrated in real-time, and related features are updated rapidly.

[0083] In this application, similarity features (scalars) and correlation features (vectors) are fused using MLP to retain complementary information of different types of features. Redundant features are automatically removed, reducing computational overhead while improving the model's generalization ability.

[0084] In this application, the fusion ratio of low-order and high-order features is automatically adjusted based on the complexity of the material (such as image entropy and text length). For example, the weight of low-order features is enhanced for visually complex advertisements. Implicit relationships between low-order and high-order features are mined using Transformer, such as aligning the visual feature of "red background" with the high-order knowledge of "high relevance to promotion".

[0085] In this application, CNN handles visually driven ads, BERT analyzes text-driven ads, and GNNs uncover association patterns, leveraging the strengths of each model. For example, CNN accurately identifies product types, BERT understands promotional language, and a gating network trained using the PPO algorithm balances classification accuracy and inference efficiency. For simple samples, only some experts are activated, while for complex samples, joint decision-making is employed.

[0086] In this application, the temperature parameter T is optimized to make the model's predicted probability closer to the true confidence level, reducing overconfident misclassifications. Platt scaling or Isotonic regression further improves classification reliability, providing a quantifiable confidence reference for ad review.

[0087] This application simultaneously captures the spatial features (such as image structure and text semantics) and temporal evolution (such as campaign cycle and seasonality) of advertising creatives, adapting to the rapidly changing nature of the advertising industry. The attention mechanism provides visualization of feature importance, and the expert selection process records the contribution of each branch, supporting the traceability of classification results and meeting the compliance requirements of advertising review. The modular design supports independent optimization and replacement of each component (such as upgrading the BERT model and expanding GNN functionality), improving system maintainability. Incremental learning and feature selection reduce computational costs, supporting real-time classification of large-scale advertising creatives.

[0088] Optionally, the first acquisition module includes: a visual feature extraction unit, which uses a convolutional neural network to extract the visual features of the advertising material; and a text feature extraction unit, which uses a BERT model based on the Transformer architecture to extract the text features of the advertising material.

[0089] Preferably, a multi-scale attention enhancement network is used for the visual feature extraction unit. Through an improved convolutional neural network architecture, high-precision extraction of visual features of advertising materials is achieved. The specific technology is as follows.

[0090] In this application, during multi-branch feature extraction: Basic feature branch: A ResNet-50 backbone network is used to extract basic visual features of the advertising image. Multi-scale feature branches: Shallow branches (e.g., conv2_x): Capture low-level features such as color distribution and edge texture. Mid-level branches (e.g., conv3_x): Extract mid-level features such as shape and local structure. Deep branches (e.g., conv4_x and conv5_x): Capture high-level features such as the main object and scene semantics.

[0091] In this application, the feature fusion mechanism is based on a Feature Pyramid Network (FPN): high-level features are upsampled and then added element-wise to low-level features to form a feature pyramid containing information at different scales. An attention mechanism is used to learn the importance weights of each branch feature, dynamically adjusting the fusion ratio.

[0092] In this application, attention weights are calculated for each location in the feature map, focusing on key areas in the advertisement (such as the main product and promotional labels). A spatial attention map is generated through convolution operations and multiplied with the original feature map to achieve feature enhancement.

[0093] In this application, attention weights are calculated for each channel of the feature map to highlight important feature channels (such as color and texture-related channels). Channel selection is achieved by generating channel attention vectors through global average pooling and fully connected layers, and then multiplying them with the original feature map.

[0094] In this application, advertising images are randomly cropped, rotated, and flipped to increase the diversity of training data. Color jitter is applied by adjusting parameters such as brightness, contrast, and saturation to enhance the model's robustness to color changes. Adversarial example training is introduced: adversarial examples are generated using the Fast Gradient Sign Method (FGSM) to improve the model's resistance to interference. Mixing original and adversarial examples during training enhances the model's generalization ability.

[0095] In this application, for the text feature extraction unit, a deep analysis of advertising text features is achieved through an improved BERT architecture. The specific technical details are as follows.

[0096] Using a pre-trained BERT-based model, the ad copy is segmented into words by a tokenizer and then input into the model. The output of the [CLS] tag is taken as the overall semantic representation to capture the core meaning of the ad text. A sentiment classification head is added on top of the BERT model to predict the sentiment polarity (positive / negative / neutral) and sentiment intensity of the ad text. Multitask learning allows the BERT model to learn semantic and sentiment features simultaneously, improving its text understanding capabilities.

[0097] In this application, the hidden state vector of each token is obtained to capture semantic information at the word level. Important keywords (such as brand names and promotional terms) receive special attention, and their weights are increased through an attention mechanism. The output vector of the [CLS] tag represents the semantic information of the entire sentence. Token-level features are integrated through pooling operations (such as average pooling and max pooling) to form a sentence-level representation. A bidirectional LSTM network is used to process the token sequence output by BERT to capture the contextual dependencies of the text. The hidden states of the LSTM are weighted and summed through an attention layer to generate a document-level feature representation.

[0098] In this application, word-level, sentence-level, or document-level features are dynamically selected based on the length and complexity of the advertising text. For short texts (such as titles), word-level and sentence-level features are primarily used; for long texts (such as product descriptions), the weight of document-level features is increased. A fully connected network is used to learn the importance weights of features at each granularity, automatically selecting the most discriminative feature combination.

[0099] In summary, this application utilizes a multi-branch network to simultaneously extract low-level, mid-level, and high-level visual features, combined with FPN and attention mechanisms, to comprehensively capture the details and semantic information of advertising images. Data augmentation and adversarial training strategies enhance the model's adaptability and robustness to diverse advertising images.

[0100] In this application, a multi-task learning framework enables the BERT model to learn semantic and sentiment features simultaneously, enhancing its understanding of advertising text. Multi-granularity feature fusion and dynamic selection mechanisms adapt to advertising copy of varying lengths and complexities, optimizing feature representation.

[0101] In this application, the outputs of the visual feature extraction unit and the text feature extraction unit share a consistent semantic space, providing a solid foundation for subsequent cross-modal fusion. The application of an attention mechanism in both units enables the extracted features to focus more on the key information of the advertisement, enhancing the accuracy of cross-modal alignment.

[0102] Optionally, the second acquisition module includes: a similarity calculation unit, which calculates the similarity features of the advertising material with other advertising materials through a Siamese neural network; and a correlation analysis unit, which obtains the correlation features of the advertising material with historical popular materials through a graph neural network.

[0103] Preferably, in a spatiotemporally aware Siamese neural network, a high-precision similarity calculation between advertising creatives is achieved through an improved Siamese neural network architecture. Specific technical details are as follows:

[0104] Visual branch: ResNet-50 is used as the backbone network, and a spatial pyramid pooling (SPP) layer is added after the last convolutional layer to extract visual features at different scales.

[0105] Text branch: The ad copy is encoded using DistilBERT, and the text representation is obtained through average pooling.

[0106] Temporal embedding: Convert the delivery time of the material into a periodic code (such as a sine / cosine function) to generate a temporal feature vector.

[0107] Visual, textual, and temporal features are fused using a gating mechanism: the importance weight of each modality is calculated, and then output = ×visual + ×text + ×time yields the comprehensive feature representation. Weights , , It is dynamically generated by the attention network based on the input.

[0108] In this application, a multi-scale contrastive learning approach is used in the similarity calculation mechanism: during feature extraction, both local feature similarity (e.g., visual regions, text phrases) and global feature similarity are calculated simultaneously. Local similarity is achieved using the PatchMatch algorithm, which divides the ad image into multiple regions and calculates the feature matching degree between regions. For time-aware similarity, a time decay factor is introduced: recent materials contribute more to the similarity calculation, while the contribution of historical materials decays exponentially over time. A time window mechanism is designed: similarity calculations are only performed on materials within a specific time window, such as similar ads from the last 30 days.

[0109] When implementing the above model training strategy, the triplet loss is optimized as follows:

[0110] Positive sample pairs: ad creatives of the same category and launched at similar times.

[0111] Negative sample pairs: ad creatives from different categories, or similar creatives from different categories but with a longer time interval.

[0112] The network is trained using Triplet Margin Loss, making the distance between positive sample pairs smaller than the distance between negative sample pairs by one margin value.

[0113] Online difficult case mining:

[0114] During training, difficult sample samples (i.e., negative sample pairs that are close to each other and positive sample pairs that are far apart) are dynamically selected to accelerate model convergence.

[0115] In this application, a dynamic heterogeneous graph neural network is used for the correlation analysis unit. This unit constructs a dynamic heterogeneous graph network to mine the deep correlation between advertising creatives and historically popular creatives. Specific technical details are as follows:

[0116] Graph structure construction:

[0117] Node type design: Material nodes: Contain low-level features (visual, textual) and high-level similarity features. User nodes: Represent the target audience, including profile information such as age, gender, and interests. Performance nodes: Represent the effectiveness of ad campaigns, such as click-through rate, conversion rate, and dwell time.

[0118] Edge Types and Weights: Creative-Creative Edge: Weight is determined by the similarity value output by the similarity calculation unit. Creative-User Edge: Weight is determined by the audience targeting strategy and actual reach of the ad. User-Performance Edge: Weight is determined by the intensity of user interaction with the ad.

[0119] In this application, a heterogeneous graph attention mechanism is introduced into the graph neural network design: independent attention weights are designed for different types of edges, and information from different neighbor nodes is aggregated through a multi-head attention mechanism. For example, when aggregating neighbor node information for material nodes, different attention parameters are used for user nodes and effect nodes.

[0120] This application introduces temporal evolution modeling: a timestamp attribute is added to each node, and a time-gated cyclic unit (GRU) is used to capture the evolution of the graph structure over time. When new material is added or historical data is updated, the graph structure and node embedding are adjusted through an incremental update mechanism.

[0121] In this application, when extracting relevance features, meta-path mining is determined: multiple meta-paths are defined (such as creatives → users → effects → popular creatives), and different types of relationships are captured through meta-path traversal algorithms. For example, potential success patterns are mined through the path "creatives → high-converting users → high-click-through-rate effects → historically popular creatives".

[0122] In this application, the GraphSAGE algorithm is used to generate node embeddings during graph embedding generation. By sampling and aggregating neighbor information, each element is mapped to a low-dimensional vector space. A self-supervised learning task is also added to predict the existence of edges between nodes, enhancing the expressive power of the graph embeddings.

[0123] In this application, multi-scale visual features are extracted using ResNet+SPP, combined with DistilBERT's textual semantic understanding and temporal periodic encoding to achieve deep fusion of visual, textual, and temporal information. For example, for seasonal promotional ads, time encoding can enhance the relevance matching with historical materials from the same period. The attention network automatically adjusts the weights of each modality based on the characteristics of the input material (such as visual complexity and text length), avoiding the limitations of traditional fixed-weight fusion. For example, visual weights are increased for pure image ads, and the influence of textual features is enhanced for long-copy ads.

[0124] This application uses the PatchMatch algorithm to calculate local region similarity, capturing subtle differences in advertising details (such as product placement and copywriting layout), while combining this with global feature similarity to achieve comprehensive matching from micro to macro levels. For example, comparing local features of advertising images of the same product from different angles can effectively identify similarities. The time decay factor and time window mechanism make the model focus more on recent similar patterns and filter out historical noise. For example, during new product launches, it prioritizes matching successful cases from the past 7 days, rather than data from the entire year.

[0125] This application enhances the model's ability to learn spatiotemporal correlations by designing special positive and negative sample pairs (positive for similar classes and close time intervals, and negative for different classes or large time intervals). For example, summer T-shirt advertisements and winter T-shirt advertisements are set as negative samples to improve seasonal sensitivity. Difficult-to-classify samples (such as visually similar but different categories of advertisements) are dynamically selected for focused training to accelerate model convergence and improve the clarity of classification boundaries.

[0126] This application simultaneously models three types of nodes—artifacts, users, and performance—and their relationships to comprehensively capture the complex interactions within the advertising ecosystem. For example, by analyzing the "artifact-user-performance" path, it identifies the actual feedback from specific audiences to advertisements and uncovers potential user preferences. Independent attention weights are designed for different types of edges, enabling the model to adjust its information aggregation strategy based on the relationship type. For instance, when aggregating user information at the artifact node, it focuses more on the feedback from high-conversion users rather than ordinary browsing users.

[0127] This application uses GRU units and timestamp attributes to track the changing trends of advertising performance over time and identify seasonal and periodic patterns. For example, it discovers that the click-through rate of a certain type of product advertisement is significantly higher on weekends than on weekdays. When new creatives are added or historical data is updated, only local adjustments to the graph structure and embedding are needed, without retraining the entire network, significantly improving the system's response speed. For example, it can integrate newly launched popular advertisements in real time and update related features.

[0128] This application uses predefined meta-paths (such as "artwork → high-converting users → high click-through rate → popular artwork") to specifically uncover successful advertising patterns. For example, it discovers that ads reaching specific user groups are more likely to achieve high conversion rates, guiding precise ad targeting. By predicting the existence of edges between nodes, it strengthens the graph embedding's ability to capture potential relationships and improves the model's generalization ability to unseen relationships.

[0129] In this application, the temporal similarity calculation of the Siamese neural network and the temporal evolution modeling of the graph neural network complement each other. The former focuses on the temporal correlation between materials, while the latter captures the temporal dynamics of the entire advertising ecosystem, jointly constructing a complete spatiotemporal information processing closed loop. The similarity value output by the similarity calculation unit is directly used as the weight of the material-material edges in the graph neural network, achieving seamless connection between the two modules and enhancing the depth of feature interaction. For example, materials with high similarity form strong connections in the graph, promoting information dissemination. Similarity features provide the direct matching degree between materials, while correlation features reveal the indirect influence of materials in the ecosystem. The combination of the two provides a multi-dimensional decision-making basis for ad classification, recommendation, and placement. For example, for materials with low similarity but high correlation, potential cross-category promotion opportunities can be explored.

[0130] The proposed solution utilizes spatiotemporal similarity matching and deep correlation mining to accurately recommend ad creatives to the target audience, thereby improving click-through rates and conversion rates. It analyzes the similarity and correlation patterns of historically popular creatives to provide data support for new ad creatives, such as recommending complementary visual elements or copywriting strategies. Real-time monitoring of the similarity between ad creatives and violations, combined with correlation network analysis of their propagation paths, provides early warnings of potential risks. By analyzing the correlation of ad performance across different time windows, it dynamically adjusts the timing, channels, and budget allocation to improve ad ROI.

[0131] Optionally, the processing module includes: a feature fusion unit, which concatenates the low-order and high-order features of the advertising material to form a total feature; and a classification decision unit, which inputs the total feature into a deep learning-based classifier and determines the category to which the advertising material belongs based on the output result.

[0132] In this application, for the feature fusion unit, a hybrid attention mechanism and a semantic bridging network are used to achieve deep fusion of low-order and high-order features. The specific technical details are as follows:

[0133] Dimensional alignment: Dimensional transformation is performed on low-order features (visual / textual) and high-order features (similarity / association), mapping them to the same dimensional space through a fully connected layer. For example, visual features (2048 dimensions), text features (768 dimensions), and high-order features (512 dimensions) are uniformly mapped to 1024 dimensions.

[0134] Feature standardization: Batch normalization is performed on features of each modality to eliminate training instability caused by differences in feature distribution.

[0135] When performing hybrid attention fusion, modal attention is calculated: the importance weight of each modality is calculated, and the fusion ratio of low-order and high-order features is dynamically adjusted by analyzing the characteristics of the input material (such as visual complexity and text length) through an attention network. For example, for visually driven ads, the weight of visual features is increased; for ads that rely on association-based recommendations, the weight of high-order features is increased. Channel attention is configured: after feature concatenation, the importance of each feature channel is evaluated through the Squeeze-and-Excitation (SE) module to suppress redundant channels and enhance the expression of key features.

[0136] For the semantic bridging network, a Transformer layer is designed as a bridge between low-order and high-order features: low-order features serve as queries, and high-order features serve as keys and values, using a self-attention mechanism to uncover the implicit relationships between the two types of features. For example, the visual low-order feature of "red background + discount label" is semantically aligned with the high-order feature of "high relevance to promotional materials".

[0137] Furthermore, residual connections and layer normalization are added to ensure smooth information flow and avoid the gradient vanishing problem.

[0138] In this application, a dynamic expert selection mechanism and uncertainty calibration are used to improve classification accuracy for classification decision units. The specific techniques are as follows:

[0139] This application employs a multi-branch classifier. For the visual branch: based on a CNN architecture, it classifies visual features, excelling at identifying product types and scenarios. For the text branch: based on a BERT architecture, it classifies text features, excelling at analyzing copywriting intent and sentiment. For the association branch: based on a GNN architecture, it classifies high-order association features, excelling at discovering group patterns between materials. In the integration decision layer, the results of each branch are merged through weighted voting, with the weights dynamically determined by the characteristics of the input materials. For example, for advertisements with simple visuals but complex text, the weight of the text branch is increased.

[0140] This application introduces a dynamic expert selection mechanism, which employs a gating network: The input is the fused total features, and the output is the weight allocation for each classifier. The gating network is trained using reinforcement learning, and the reward function combines classification accuracy and inference efficiency.

[0141] This application introduces an expert expansion strategy: when encountering difficult samples (such as materials with low classification confidence), more expert branches are dynamically activated for joint decision-making. For example, for semantically ambiguous advertisements, visual, textual, and relational branches are activated simultaneously for cross-validation.

[0142] Additionally, for uncertainty calibration: temperature scaling is applied to the logits output by the classifier, and the optimal temperature parameters are learned by minimizing the cross-entropy loss on the validation set. This makes the model's predicted probabilities closer to the true confidence levels, improving classification reliability. A rejection option mechanism is configured: when classification uncertainty exceeds a threshold, automatic classification is rejected, and a manual review process is triggered. Uncertainty is calculated using the entropy value or mutual information of the predicted probabilities.

[0143] In summary, this application adaptively adjusts the fusion ratio of features across different modalities by analyzing the characteristics of the input materials (such as visual complexity and text length). For example, for image-based advertisements, it automatically increases the weight of visual features, avoids interference from text features, and improves classification accuracy. The SE module learns the dependencies between feature channels, suppresses redundant information (such as background noise), enhances the expression of key features (such as product elements and promotional labels), and improves the model's sensitivity to subtle differences.

[0144] The Transformer layer acts as a bridge, semantically aligning low-order features (such as visual elements) with high-order features (such as relationships). For example, it associates the visual feature of "red background + discount tag" with the high-order knowledge of "high relevance to promotional materials," uncovering implicit semantic relationships and addressing the semantic gap that traditional splicing methods cannot capture. Residual connections and layer normalization ensure the effective propagation of gradients in deep networks, supporting the construction of deeper and more complex fusion networks and enhancing the model's expressive power.

[0145] This application maps features from different sources to the same dimensional space, avoiding information loss due to dimensionality mismatch and providing a unified feature representation for subsequent fusion. It eliminates training instability caused by differences in feature distributions across different modalities, accelerates model convergence, and reduces the risk of overfitting.

[0146] This application designs dedicated classifiers for different feature types (visual, textual, and relational), fully leveraging the strengths of each model. For example, CNN excels at processing spatial structure information, BERT excels at capturing textual semantics, and GNN excels at mining relational networks. By training the gating network through reinforcement learning, the weights of each branch are dynamically adjusted based on the characteristics of the input material, achieving an optimal balance between accuracy and efficiency. For example, for visually simple advertisements, the computational load of CNN branches is reduced, improving inference speed.

[0147] In this application, when encountering samples with low classification confidence, more expert branches are automatically activated for joint decision-making, and classification accuracy in complex scenarios is improved through multi-model cross-validation. For example, semantically ambiguous advertisements are judged comprehensively using visual, textual, and relational information simultaneously. The application supports on-demand expansion of expert branches (such as adding an audio classifier) ​​to improve the system's adaptability to new modalities without requiring a redesign of the overall architecture.

[0148] This application's temperature scaling technology optimizes temperature parameters to make the model's predicted probabilities closer to the true confidence distribution, reducing erroneous predictions due to overconfidence. The rejection option mechanism automatically identifies high-uncertainty samples and triggers manual review, forming a closed-loop optimization system of "automatic model classification → manual calibration of difficult samples → calibration data feedback training," continuously improving model performance.

[0149] This application helps improve advertising efficiency: through precise categorization, ad creatives are automatically matched to the most suitable delivery channels and audiences, increasing click-through rates and conversion rates.

[0150] This application helps reduce content moderation costs by automatically identifying high-risk advertisements (such as false advertising), reducing the workload of manual review, and improving review efficiency.

[0151] This application helps enhance personalized recommendations: based on material classification results and user history behavior, it generates more accurate personalized recommendations, thereby improving the user experience.

[0152] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims. The application, device, module, or unit described in the above embodiments can be specifically implemented by a computer chip or entity, or by a product with a certain function.

[0153] For ease of description, the above apparatus is described in terms of its functions, divided into various units. Of course, in implementing this invention, the functions of each unit can be implemented in one or more software and / or hardware components.

[0154] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, this application, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0155] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (this application), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0158] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, a network interface, and memory. Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0159] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0160] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0161] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, this application, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0162] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific transactions or implement specific abstract data types. This invention can also be practiced in distributed computing environments where transactions are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0163] The various embodiments in this specification are described in a progressive manner, with identical or similar parts between the embodiments referred to or substituted for each other. For the embodiments of this application, since they are basically similar to the method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0164] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims. The present application, device, module, or unit described in the above embodiments is specifically implemented by a computer chip or entity, or by a product having a certain function.

[0165] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, this application, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A deep learning-based method for classifying advertising creatives, characterized in that, include: Obtain low-level features of the advertising creative, including the visual features and text features of the advertising creative itself; Obtain high-level features of the advertising creative, wherein the high-level features represent the correlation features between the advertising creative and other advertising creatives; The category to which the advertising material belongs is determined based on its low-level and high-level features; The higher-order features include at least one of the following: similarity features of the advertising material among materials released in the same batch, and correlation features between the advertising material and historically popular materials; when performing parallel extraction of multimodal lower-order features, for the advertising material, a convolutional neural network is used to perform convolution operations on the visual part of the advertising material to extract color distribution features, composition structure features, and subject object features; the BERT model based on the Transformer architecture is used to perform semantic encoding on the text part of the advertising material to extract keyword features and semantic sentiment features, thereby obtaining the lower-order feature set of the advertising material; When constructing high-order correlation features across different creative materials, the cosine similarity between the low-order feature set of the current creative material and the low-order feature data of the same batch / historical creative materials is calculated using a Siamese neural network to generate similarity features. Then, a graph neural network is used to construct a correlation graph between the low-order features of the current creative material and the features of historical popular creative materials to mine correlation features and obtain the high-order feature set of the creative material. When making feature fusion classification decisions, the low-order feature set and the high-order feature set of the advertising material are concatenated dimensionally to form a complete advertising material feature vector. This feature vector is then input into the multilayer perceptron classification model, and the Softmax function outputs the probability of each preset category. The category with the highest probability is selected as the final classification result to obtain the advertising material classification result.

2. A deep learning-based advertising material classification device for any method of claim 1, characterized in that, include: The first acquisition module is adapted to acquire low-level features of advertising materials, wherein the low-level features include the visual features and text features of the advertising materials themselves; The second acquisition module is adapted to acquire high-level features of advertising materials, wherein the high-level features represent the correlation features between the advertising materials and other advertising materials; The processing module is adapted to determine the category to which the advertising material belongs based on its low-level and high-level features.

Images