A statistical method of advertising data based on deep learning

Abstract

The present invention relates to the field of advertising data technology and discloses a method for statistically analyzing advertising data based on deep learning, comprising the following steps: Step 1: constructing a user-advertising interaction matrix for advertising data, wherein rows of the matrix represent different users, columns represent different advertisements, and each element of the matrix represents the degree of interaction between the user and the advertisement; Step 2: encoding the obtained user and advertisement features, converting the user and advertisement features into low-dimensional dense vectors, and mapping the discrete user and advertisement features through embedding technology so that the user and advertisement features can be expressed in a low-dimensional space. By adopting an adaptive matrix decomposition method and dynamically adjusting the matrix rank, the limitations of using a fixed rank in traditional methods are avoided, and the potential user-advertising relationship features can be effectively extracted from high-dimensional sparse advertising data, resulting in higher computational efficiency and more accurate advertising effect prediction.

Description

Technical Field

[0001] The present invention relates to the field of advertising data technology, and in particular to an advertising data statistics method based on deep learning. Background Art

[0002] In modern advertising, the statistics and analysis of advertising data are crucial for optimizing advertising effectiveness. With the diversification of advertising platforms, advertisers face challenges with diverse advertising data. Data is often highly dimensional and sparse, making traditional data analysis methods ineffective in extracting underlying characteristics of user-advertising relationships. Therefore, efficiently processing sparse data and accurately predicting advertising effectiveness have become key challenges in improving advertising efficiency and return on investment.

[0003] Current advertising data statistics methods primarily rely on traditional machine learning models and statistical methods, processing advertising data through simple feature engineering and matrix decomposition techniques. Traditional methods typically use a fixed rank for matrix decomposition, resulting in poor performance on sparse data. Furthermore, advertising effectiveness prediction models often ignore the differences between different advertising datasets and scenarios, lacking adaptability and, therefore, failing to achieve ideal prediction results in complex advertising data environments.

[0004] The development of deep learning, especially in its ability to process high-dimensional, sparse data, has enabled deep neural networks to demonstrate powerful feature learning capabilities, automatically extracting meaningful patterns and relationships from data. However, traditional deep learning methods still face challenges in effectively processing high-dimensional, sparse data and dynamically adjusting model structures to suit different advertising scenarios.

[0005] Therefore, those skilled in the art provide an advertising data statistics method based on deep learning to solve the above-mentioned problems. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the present invention provides an advertising data statistics method based on deep learning to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for statistically analyzing advertising data based on deep learning, comprising:

[0008] Step 1: Construct a user-ad interaction matrix of the ad data, where rows represent different users, columns represent different ads, and each element of the matrix represents the degree of interaction between the user and the ad;

[0009] Step 2: Encode the obtained user and ad features and convert them into low-dimensional dense vectors. Use embedding technology to map the discrete user and ad features so that they can be expressed in a low-dimensional space.

[0010] Step 3: Based on the obtained user and ad features, perform low-rank matrix decomposition and represent the user-ad interaction matrix as a matrix product. By decomposing and reducing the matrix dimension, the potential relationship between users and ads is extracted.

[0011] Step 4: Based on the low-rank matrix decomposition, regularization techniques are used to constrain the obtained user and ad matrices. By adding L1 regularization during the matrix decomposition process, the irrelevant feature terms in the matrix are pushed toward zero and the learning effect of the model is optimized.

[0012] Step 5: Use a neural network structure to process the low-rank matrix. The neural network performs nonlinear mapping on the user and ad embedding vectors extracted from the low-rank matrix, captures the nonlinear relationship, and outputs the final ad effect prediction value.

[0013] Step 6: Introduce the mutual information maximization strategy to maximize the mutual information between the feature vector and the advertising effect, so that the model can learn effective feature representation;

[0014] Step 7: Use an optimization algorithm to update all parameters, including the embedding vectors of users and ads and the weight matrix of the neural network. During the training process, optimization is performed using the backpropagation algorithm and gradient descent method, and the objective function is used to minimize the training error.

[0015] Step 8: After model training is complete, use the test data to evaluate the model. By comparing the predicted results with the real data, the accuracy of the model is evaluated and the model parameters are adjusted if necessary.

[0016] Step 9: Apply the trained model to real-time optimization of advertising delivery. Based on the model’s prediction results, adjust the advertising display strategy to optimize the audience, display frequency, and advertising creative to maximize the effectiveness of advertising delivery.

[0017] Preferably, in the low-rank matrix decomposition of step 3, an adaptive matrix decomposition method is used to dynamically adjust the matrix rank r, and the matrix decomposition is performed by optimizing the following objective function:

[0018]

[0019] Among them, L adapt is the loss function of adaptive matrix factorization,

[0020] A is the user-ad interaction matrix,

[0021] U and V are low-rank embedding matrices of users and ads,

[0022] r is the matrix rank, α is a hyperparameter that controls the matrix complexity,

[0023] λ1 and λ2 are regularization parameters,

[0024] ∥U∥1 is the L1 norm of the user feature matrix U,

[0025] ∥V∥1 is the L1 norm of the advertising feature matrix V,

[0026] Represents the sum of squares of the elements of a matrix.

[0027] Preferably, in the regularization process of step 4, L1 regularization is processed by the following formula:

[0028]

[0029] Among them, L total represents the comprehensive loss function, λ1 and λ2 are regularization parameters,

[0030] ∥U∥1 is the L1 norm of the user feature matrix U,

[0031] ∥V∥1 is the L1 norm of the advertising feature matrix V,

[0032] represents the sum of squares of the elements of the matrix,

[0033] A is the user-ad interaction matrix,

[0034] U and V are the low-rank embedding matrices of users and ads.

[0035] Preferably, in the neural network structure of step 5, a deep neural network is used to process the embedding vector obtained by low-rank matrix decomposition. The network structure includes multiple groups of fully connected layers. The output of each layer is nonlinearly mapped through an activation function, and finally the advertising effect prediction value y is output:

[0036] y=f(U,V)=σ(W3·σ(W2·σ(W1·(U·V T )))),

[0037] Among them, W1, W2 and W3 are the weight matrices of the fully connected layer,

[0038] y is the predicted value of advertising effect, U is the user feature matrix,

[0039] V is the advertising feature matrix, σ is the activation function,

[0040] (U·V T) is the product of the user feature matrix U and the advertisement feature matrix V obtained by low-rank matrix decomposition. Preferably, the mutual information maximization strategy in step 6 is maximized by the following objective function:

[0041] I(X,Y)=H(X)+H(Y)-H(X,Y),

[0042] Among them, I(X, Y) is the mutual information, H(X) is the entropy of the user and ad feature matrix X,

[0043] H(Y) is the entropy of advertising effect Y, and H(X, Y) is the joint entropy of user characteristics and advertising effect.

[0044] Preferably, the optimization algorithm in step 7 is an optimization algorithm based on gradient descent. During the training process, the parameters are updated. The parameters include the user embedding vector U, the advertisement embedding vector V, and the neural network weight matrices W1, W2, and W3. The gradient calculation formula is:

[0045]

[0046] Among them, A is the user-ad interaction matrix, U is the user feature matrix, and V is the ad feature matrix.

[0047] Indicates the direction and step size of optimizing the user feature matrix,

[0048] λ1 and λ2 are regularization parameters, sign(U) represents the sign function of matrix U,

[0049] sign(V) represents the sign function of matrix V.

[0050] Preferably, the optimization algorithm used is the Adam optimizer, which updates the parameters by calculating the gradient and adaptive learning rate to accelerate the convergence of the model. The specific optimization process is carried out according to the following update rules:

[0051]

[0052] Among them, θ t+1 represents the parameter value at the t+1th iteration, θ t represents the parameter value at the tth iteration, η is the learning rate, m t is the moving average of the gradient, v t is the moving average of the square of the gradient, and ∈ is a constant.

[0053] Preferably, L2 regularization is used in the training process, and the L2 regularization term is controlled by the following formula:

[0054]

[0055] Among them, L reg Represents the regularization loss of the neural network weight matrix, λ3 is the regularization parameter, are the elements of the neural network weight matrix W1, W2, and W3.

[0056] Preferably, the early stopping method is used during the training process to avoid overfitting. The early stopping method monitors the validation set loss. If the validation set loss fails to decrease in multiple consecutive training rounds, the training is stopped early to prevent overfitting of the model.

[0057] Preferably, after the training, the advertising effect prediction model obtained is applied to the real-time optimization of the advertising delivery strategy. According to the predicted advertising effect, the advertising display frequency, advertising creativity and audience positioning strategy are adjusted to improve the overall effect of advertising delivery.

[0058] The present invention provides an advertising data statistics method based on deep learning. It has the following beneficial effects:

[0059] 1. The present invention adopts an adaptive matrix decomposition method to dynamically adjust the matrix rank, avoiding the limitation of using fixed rank in traditional methods, and effectively extracting potential user-advertising relationship features in high-dimensional sparse advertising data, thereby achieving higher computational efficiency and more accurate advertising effect prediction.

[0060] 2. The present invention optimizes the objective function of matrix rank adjustment so that the matrix decomposition process can dynamically adjust the rank according to the actual changes in the data, thereby enhancing the adaptability of the model to different advertising data sets and scenarios, and obtaining a more flexible and efficient advertising data processing solution.

[0061] 3. This invention introduces a mutual information maximization strategy to enhance the model's ability to capture the deep relationship between input features and advertising effects, thereby achieving more precise and robust feature learning and obtaining more accurate advertising effect prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0063] To help those skilled in the art understand the present invention, the following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0064] The present invention is described in detail below with reference to the accompanying drawings:

[0065] Example:

[0066] Please see the attached Figure 1 , an embodiment of the present invention provides an advertising data statistics method based on deep learning, comprising:

[0067] Step 1: Construct a user-ad interaction matrix for the ad data. The rows of the matrix represent different users, the columns represent different ads, and each element of the matrix represents the degree of interaction between the user and the ad.

[0068] Step 2: Encode the obtained user and ad features and convert them into low-dimensional dense vectors. Use embedding technology to map the discrete user and ad features so that they can be expressed in a low-dimensional space.

[0069] Step 3: Based on the obtained user and ad features, perform low-rank matrix decomposition and represent the user-ad interaction matrix as a matrix product. By decomposing and reducing the matrix dimension, the potential relationship between users and ads is extracted.

[0070] Step 4: Based on the low-rank matrix decomposition, regularization techniques are used to constrain the obtained user and ad matrices. By adding L1 regularization during the matrix decomposition process, the irrelevant feature terms in the matrix are pushed toward zero and the learning effect of the model is optimized.

[0071] Step 5: Use a neural network structure to process the low-rank matrix. The neural network performs nonlinear mapping on the user and ad embedding vectors extracted from the low-rank matrix, captures the nonlinear relationship, and outputs the final ad effect prediction value.

[0072] Step 6: Introduce the mutual information maximization strategy to maximize the mutual information between the feature vector and the advertising effect, so that the model can learn effective feature representation;

[0073] Step 7: Use an optimization algorithm to update all parameters, including the embedding vectors of users and ads and the weight matrix of the neural network. During the training process, optimization is performed using the backpropagation algorithm and gradient descent method, and the objective function is used to minimize the training error.

[0074] Step 8: After model training is complete, use the test data to evaluate the model. By comparing the predicted results with the real data, the accuracy of the model is evaluated and the model parameters are adjusted if necessary.

[0075] Step 9: Apply the trained model to real-time optimization of advertising delivery. Based on the model’s prediction results, adjust the advertising display strategy to optimize the audience, display frequency, and advertising creative to maximize the effectiveness of advertising delivery.

[0076] The benefit of step 1 is that it can provide a clear basic data structure for subsequent advertising effect predictions, laying the foundation for the model to learn the potential relationship between user behavior and advertising.

[0077] The benefit of step 2 is to reduce the complexity of the feature space, avoid the computational burden brought by high-dimensional data processing, and improve the learning efficiency of the model.

[0078] The benefit of step 3 is that it can reveal the deep relationship between advertisements and users, while improving the efficiency of data processing and avoiding the computational bottleneck in traditional sparse matrix methods.

[0079] The benefit of step 4 is to reduce unnecessary features, improve the generalization ability of the model, make feature selection more accurate, and improve the stability and accuracy of the model.

[0080] The benefit of step 5 is that it automatically discovers potential complex patterns in the data through the deep learning model, improves the prediction accuracy of advertising effects, and enables the model to handle more complex data relationships.

[0081] The benefit of step 6 is that it enables the model to learn linear relationships and capture deep nonlinear information, thereby improving the accuracy of advertising effect prediction.

[0082] The benefit of step 7 is that the model can be continuously optimized based on feedback to ensure that the training process is efficient and accurate.

[0083] The benefit of step 8 is that it can ensure the generalization ability of the model, achieve better results in practical applications, and avoid overfitting and underfitting.

[0084] The benefit of step 9 is that it can adjust the strategy in real time based on the advertising effect, thereby improving the return on investment of advertising and ensuring that advertisers get the best results.

[0085] In the low-rank matrix decomposition of step 3, the adaptive matrix decomposition method is used to dynamically adjust the matrix rank r, and the matrix decomposition is performed through the following optimization objective function:

[0086]

[0087] Among them, L adapt is the loss function of adaptive matrix factorization,

[0088] A is the user-ad interaction matrix,

[0089] U and V are low-rank embedding matrices of users and ads,

[0090] r is the matrix rank, α is a hyperparameter that controls the matrix complexity,

[0091] λ1 and λ2 are regularization parameters,

[0092] ∥U∥1 is the L1 norm of the user feature matrix U,

[0093] ∥V∥1 is the L1 norm of the advertising feature matrix V,

[0094] Represents the sum of squares of the elements of a matrix.

[0095] Traditional low-rank matrix factorization methods use a fixed rank to decompose the user-ad interaction matrix. However, this method dynamically adjusts the matrix rank r, automatically selecting the most appropriate rank for different ad datasets and scenarios. This method ensures that the model can adapt to various data environments and effectively handle the complexity and size of different ad datasets, thereby improving the model's flexibility and generalization capabilities.

[0096] By optimizing the objective function, we ensure that the optimal rank is dynamically selected during the matrix factorization process based on the actual data, avoiding overfitting and oversimplification. A lower matrix rank reduces computational complexity, allowing the model to maintain high accuracy while improving computational efficiency. Furthermore, appropriate matrix rank selection helps reduce redundant features, avoid overfitting, and ultimately improve model stability and accuracy.

[0097] Adding an L1 regularization term to the matrix decomposition process forces irrelevant feature terms in the matrix to approach zero, thereby effectively selecting features. The introduction of the regularization term helps reduce feature redundancy and enhances the model's generalization ability, making the model more robust when processing new data.

[0098] Low-rank matrix factorization effectively extracts the underlying relationship features between users and ads. By adaptively adjusting the matrix rank, the model can more accurately capture the complex relationships between users and ads, ensuring that more valuable feature information is extracted from sparse data. This is particularly important for predicting advertising effectiveness, as the relationship between ads and users is often very complex. Low-rank factorization helps capture this underlying deep relationship.

[0099] During the regularization process of step 4, L1 regularization is processed by the following formula:

[0100]

[0101] Among them, L total represents the comprehensive loss function, λ1 and λ2 are regularization parameters,

[0102] ∥U∥1 is the L1 norm of the user feature matrix U,

[0103] ∥V∥1 is the L1 norm of the advertising feature matrix V,

[0104] represents the sum of squares of the elements of the matrix,

[0105] A is the user-ad interaction matrix,

[0106] U and V are the low-rank embedding matrices of users and ads.

[0107] The core advantage of L1 regularization is that it can force irrelevant feature terms in the matrix toward zero. By applying L1 regularization to the user and ad matrices U and V, the model automatically selects the most representative features and removes irrelevant ones, achieving the desired effect of feature selection.

[0108] Regularization helps reduce model complexity and effectively prevent overfitting. L1 regularization, in particular, reduces the model's degrees of freedom by compressing unimportant features, ensuring good performance on both training data and unseen data. This makes the model more generalizable and better able to adapt to changing data in real-world applications.

[0109] Because L1 regularization promotes matrix sparsity, it can reduce redundancy in calculations and improve the model's computational efficiency. In high-dimensional sparse data environments, L1 regularization makes the computational process more efficient, especially when processing large-scale advertising data, significantly reducing computational complexity.

[0110] In the neural network structure of step 5, a deep neural network is used to process the embedding vector obtained by low-rank matrix decomposition. The network structure includes multiple groups of fully connected layers. The output of each layer is nonlinearly mapped through an activation function, and the final output is the advertising effect prediction value y:

[0111] y=f(U,V)=σ(W3·σ(W2·σ(W1·(U·V T )))),

[0112] Among them, W1, W2 and W3 are the weight matrices of the fully connected layer,

[0113] y is the predicted value of advertising effect, U is the user feature matrix,

[0114] V is the advertising feature matrix, σ is the activation function,

[0115] (U·V T ) is the product of the user feature matrix U and the advertising feature matrix V obtained by low-rank matrix decomposition.

[0116] Neural networks, through a combination of multiple fully connected layers and activation functions, can automatically capture nonlinear relationships in advertising data. In predicting advertising effectiveness, the relationship between users and ads is complex and nonlinear. Deep neural networks, through layer-by-layer nonlinear mapping, can map features derived from low-rank matrix decomposition into a higher-dimensional space, improving the model's ability to model complex relationships.

[0117] Through deep neural networks and low-rank matrix decomposition, the embedding vectors of the user feature matrix U and the ad feature matrix V are generated to generate richer and more discriminative feature representations. After the features are mapped through a nonlinear activation function, they can enhance the learning of potential advertising effects and improve the accuracy of advertising effect prediction.

[0118] The hierarchical structure of deep neural networks enables the model to flexibly adapt to diverse datasets and advertising scenarios. The output of each layer is processed through an activation function, providing the model with enhanced feature extraction capabilities. The nonlinear combination of different layers helps the model automatically discover the most meaningful patterns in the data, adapting to the complexities of various advertising scenarios.

[0119] The mutual information maximization strategy in step 6 is maximized by the following objective function:

[0120] I(X,Y)=H(X)+H(Y)-H(X,Y),

[0121] Among them, I(X, Y) is the mutual information, H(X) is the entropy of the user and ad feature matrix X,

[0122] H(Y) is the entropy of advertising effect Y, and H(X, Y) is the joint entropy of user characteristics and advertising effect.

[0123] Maximizing mutual information helps the model learn more representative and higher-quality feature representations. By maximizing the mutual information between user features and ad performance, the model automatically identifies and enhances features closely related to ad performance, enabling better extraction of latent information from ad data, improving the expressiveness of features, and enhancing the accuracy of ad performance predictions.

[0124] Traditional models typically learn linear relationships, but by maximizing mutual information, the model can capture deeper, nonlinear connections between user characteristics and ad effectiveness. This process enables the model to understand the complex interactions between user behavior and ad content, improving the accuracy of ad performance predictions.

[0125] Maximizing mutual information enhances the robustness of the model by strengthening the dependency between features and advertising effectiveness. When dealing with incomplete and noisy data, the model can optimize feature learning by maximizing mutual information, reducing interference from irrelevant information and improving the stability and reliability of advertising effectiveness predictions.

[0126] The optimization algorithm in step 7 is based on gradient descent. During the training process, the parameters are updated. The parameters include the user embedding vector U, the ad embedding vector V, and the neural network weight matrices W1, W2, and W3. The gradient calculation formula is:

[0127]

[0128] Among them, A is the user-ad interaction matrix, U is the user feature matrix, and V is the ad feature matrix.

[0129] Indicates the direction and step size of optimizing the user feature matrix,

[0130] λ1 and λ2 are regularization parameters, sign(U) represents the sign function of matrix U,

[0131] sign(V) represents the sign function of matrix V.

[0132] The optimization algorithm used is the Adam optimizer, which updates the parameters by calculating the gradient and adaptive learning rate to accelerate the convergence of the model. The specific optimization process is carried out according to the following update rules:

[0133]

[0134] Among them, θ t+1 represents the parameter value at the t+1th iteration, θ t represents the parameter value at the tth iteration, η is the learning rate, m t is the moving average of the gradient, v t is the moving average of the square of the gradient, and ∈ is a constant.

[0135] Gradient descent-based optimization algorithms optimize the model by calculating gradients and adjusting parameters to continuously reduce the loss function. The gradient calculation formula effectively guides model parameter updates toward minimizing the loss function, improving model accuracy. The Adam optimizer accelerates training by dynamically adjusting the learning rate and combining a moving average of the gradient, avoiding local minima in the parameter space and ensuring rapid model convergence.

[0136] The Adam optimizer offers the advantage of an adaptive learning rate, adjusting the update step size based on the gradients of each parameter. Specifically, the algorithm automatically adjusts the learning rate of each parameter based on historical gradient information, enabling more precise optimization when faced with varying parameter update requirements. This adaptive learning rate makes parameter updates more flexible and efficient, maintaining stable convergence, particularly when dealing with complex and volatile advertising data.

[0137] During training, regularization helps control model complexity and reduce overfitting of the model to the training data. By controlling the sparsity of the user feature matrix U and the ad feature matrix V, the regularization parameter encourages the model to select more meaningful features, improving the model's generalization and stability.

[0138] L2 regularization is used during training, and the L2 regularization term is controlled by the following formula:

[0139]

[0140] Among them, L reg Represents the regularization loss of the neural network weight matrix, λ3 is the regularization parameter, are the elements of the neural network weight matrix W1, W2, and W3.

[0141] The early stopping method is used during the training process to avoid overfitting. The early stopping method monitors the validation set loss. If the validation set loss fails to decrease in multiple consecutive training rounds, the training is stopped early to prevent overfitting of the model.

[0142] After training, the resulting advertising effect prediction model is used for real-time optimization of advertising delivery strategies. Based on the predicted advertising effects, the advertising display frequency, advertising creativity, and audience targeting strategies are adjusted to improve the overall effectiveness of advertising delivery.

[0143] Through L2 regularization, early stopping, and real-time optimization, the present invention can effectively improve the performance of advertising data statistics methods. L2 regularization prevents overfitting and improves the generalization ability of the model by controlling model complexity; early stopping avoids overfitting during training and enhances model stability by monitoring validation set loss; and real-time optimization enables the model to dynamically adjust advertising strategies based on advertising effect prediction results, improving the accuracy and effectiveness of advertising delivery. This combination of technical means makes the advertising effect prediction model more efficient, stable, and accurate during training, evaluation, and application.

[0144] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for statistically analyzing advertising data based on deep learning, characterized in that: include: Step 1: Construct a user-ad interaction matrix of the ad data, where rows represent different users, columns represent different ads, and each element of the matrix represents the degree of interaction between the user and the ad; Step 2: Encode the obtained user and ad features and convert them into low-dimensional dense vectors. Use embedding technology to map the discrete user and ad features so that they can be expressed in a low-dimensional space. Step 3: Based on the obtained user and ad features, perform low-rank matrix decomposition and represent the user-ad interaction matrix as a matrix product. By decomposing and reducing the matrix dimension, the potential relationship between users and ads is extracted. Step 4: Based on the low-rank matrix decomposition, regularization techniques are used to constrain the obtained user and ad matrices. By adding L1 regularization during the matrix decomposition process, the irrelevant feature terms in the matrix are pushed toward zero and the learning effect of the model is optimized. Step 5: Use a neural network structure to process the low-rank matrix. The neural network performs nonlinear mapping on the user and ad embedding vectors extracted from the low-rank matrix, captures the nonlinear relationship, and outputs the final ad effect prediction value. Step 6: Introduce the mutual information maximization strategy to maximize the mutual information between the feature vector and the advertising effect, so that the model can learn effective feature representation; Step 7: Use an optimization algorithm to update all parameters, including the embedding vectors of users and ads and the weight matrix of the neural network. During the training process, optimization is performed using the backpropagation algorithm and gradient descent method, and the objective function is used to minimize the training error. Step 8: After model training is complete, use the test data to evaluate the model. By comparing the predicted results with the real data, the accuracy of the model is evaluated and the model parameters are adjusted if necessary. Step 9: Apply the trained model to real-time optimization of advertising delivery. Based on the model’s prediction results, adjust the advertising display strategy to optimize the audience, display frequency, and advertising creative to maximize the effectiveness of advertising delivery.

2. The advertising data statistics method based on deep learning according to claim 1, characterized in that: In the low-rank matrix decomposition of step 3, the matrix rank r is dynamically adjusted using an adaptive matrix decomposition method, and the matrix decomposition is performed by optimizing the following objective function: Among them, L adapt is the loss function of adaptive matrix factorization, A is the user-ad interaction matrix, U and V are low-rank embedding matrices of users and ads, r is the matrix rank, α is a hyperparameter that controls the matrix complexity, λ1 and λ2 are regularization parameters, ∥U∥1 is the L1 norm of the user feature matrix U, ∥V∥1 is the L1 norm of the advertising feature matrix V, Represents the sum of squares of the elements of a matrix.

3. The advertising data statistics method based on deep learning according to claim 2, characterized in that: During the regularization process of step 4, L1 regularization is performed using the following formula: Among them, L total represents the comprehensive loss function, λ1 and λ2 are regularization parameters, ∥U∥1 is the L1 norm of the user feature matrix U, ∥V∥1 is the L1 norm of the advertising feature matrix V, represents the sum of squares of the elements of the matrix, A is the user-ad interaction matrix, U and V are the low-rank embedding matrices of users and ads.

4. The advertising data statistics method based on deep learning according to claim 1, characterized in that: In the neural network structure of step 5, a deep neural network is used to process the embedding vector obtained by low-rank matrix decomposition. The network structure includes multiple groups of fully connected layers. The output of each layer is nonlinearly mapped through an activation function, and the final output is the advertising effect prediction value y: y=f(U,V)=σ(W3·σ(W2·σ(W1·(U·V T ))))), Among them, W1, W2 and W3 are the weight matrices of the fully connected layer, y is the predicted value of advertising effect, U is the user feature matrix, V is the advertising feature matrix, σ is the activation function, (U·V T ) is the product of the user feature matrix U and the advertising feature matrix V obtained by low-rank matrix decomposition.

5. The advertising data statistics method based on deep learning according to claim 1, characterized in that: The mutual information maximization strategy in step 6 is maximized by the following objective function: I(X,Y)=H(X)+H(Y)-H(X,Y), Among them, I(X, Y) is the mutual information, H(X) is the entropy of the user and ad feature matrix X, H(Y) is the entropy of advertising effect Y, and H(X, Y) is the joint entropy of user characteristics and advertising effect.

6. The advertising data statistics method based on deep learning according to claim 1, characterized in that: The optimization algorithm in step 7 is a gradient descent-based optimization algorithm. During the training process, the parameters are updated. The parameters include the user embedding vector U, the ad embedding vector V, and the neural network weight matrices W1, W2, and W3. The gradient calculation formula is: Among them, A is the user-ad interaction matrix, U is the user feature matrix, and V is the ad feature matrix. Indicates the direction and step size of optimizing the user feature matrix, λ1 and λ2 are regularization parameters, sign(U) represents the sign function of matrix U, sign(V) represents the sign function of matrix V.

7. The advertising data statistics method based on deep learning according to claim 6, characterized in that: The optimization algorithm used is the Adam optimizer, which updates the parameters by calculating the gradient and adaptive learning rate to accelerate the convergence of the model. The specific optimization process is carried out according to the following update rules: Among them, θ t+1 represents the parameter value at the t+1th iteration, θ t represents the parameter value at the tth iteration, η is the learning rate, m t is the moving average of the gradient, v t is the moving average of the square of the gradient, and ∈ is a constant.

8. The advertising data statistics method based on deep learning according to claim 6, characterized in that: L2 regularization is used in the training process, and the L2 regularization term is controlled by the following formula: Among them, L reg Represents the regularization loss of the neural network weight matrix, λ3 is the regularization parameter, are the elements of the neural network weight matrix W1, W2, and W3.

9. The advertising data statistics method based on deep learning according to claim 8, characterized in that: The early stopping method is used in the training process to avoid overfitting. The early stopping method monitors the validation set loss. If the validation set loss fails to decrease in multiple consecutive training rounds, the training is stopped early to prevent overfitting of the model.

10. The advertising data statistics method based on deep learning according to claim 9, characterized in that: After the training, the obtained advertising effect prediction model is applied to the real-time optimization of the advertising delivery strategy. According to the predicted advertising effect, the advertising display frequency, advertising creativity and audience positioning strategy are adjusted to improve the overall effect of advertising delivery.

Images