A self-media promotion content heat prediction method based on deep learning

Abstract

The application discloses a kind of self-media promotion content heat prediction methods based on deep learning, comprising: S1, acquisition multi-modal data and pre-process output text word vector sequence and standard size image;S2, construct semantic structure atlas and output text semantic feature vector;S3, carry out saliency target detection and output visual saliency feature map;S4, feature aggregation and interaction are carried out using MLP-Mixer model, and output graphic-text coupling feature vector;S5, construct improved DCCA network, introduce covariance reconstruction loss function to generate psychological resonance index feature;S6, carry out time series evolution prediction and output heat prediction result.The application realizes graphic-text feature deep coupling and audience psychological resonance quantization, effectively improves the accuracy and explainability of heat prediction.

Description

Technical Field

[0001] This invention relates to the field of self-media content analysis and prediction, and in particular to a method for predicting the popularity of self-media promotional content based on deep learning. Background Technology

[0002] Multimodal feature fusion and deep canonical correlation analysis techniques have been widely applied in recent years in fields such as public opinion analysis, content recommendation, and multimedia retrieval due to their significant advantages in cross-modal information association and latent feature extraction, becoming an important development direction for multimodal data processing. However, in practical applications, predicting the popularity of self-media promotional content faces many challenges, such as a large modal semantic gap, complex audience psychological mechanisms, and dynamic and ever-changing patterns of popularity evolution. The deployment effectiveness of existing technologies is still constrained by many factors.

[0003] Most current trend prediction methods rely on single-modal feature inputs or simple feature splicing, making it difficult to fully explore the deep complementary relationship between the semantic structure of the text and visually salient regions, resulting in a lack of completeness in the expression of the content theme. Some systems ignore the implicit mapping relationship between the audience's psychological dimension and content features, relying solely on historical statistical data for numerical regression, failing to construct quantitative indicators of psychological resonance in the common subspace, thus limiting the prediction model's ability to perceive the intrinsic driving force of "viral" content. Furthermore, the prediction logic process is often a "black box" regression, lacking matching and comparison with historical standard evolution patterns, making it difficult to provide operators with clear decision-making basis including trend peaks and durations, affecting the interpretability and practical value of the prediction results.

[0004] Furthermore, existing multimodal analysis methods often suffer from structural information loss during feature mapping. They fail to retain the original input distribution structure by constraining the projection matrix through covariance reconstruction loss function, resulting in deviations in the calculation of feature correlation after mapping. This makes it difficult to adapt to the rapid evolution of self-media content styles and audience interests, seriously affecting the model's generalization ability and prediction accuracy in real-world scenarios.

[0005] Therefore, how to provide a method for predicting the popularity of self-media promotion content based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method for predicting the popularity of self-media promotion content. This invention fully integrates key steps such as semantic structure graph construction, visual saliency extraction, MLP-Mixer model feature fusion, improved DCCA network mapping, and temporal evolution pattern matching. It constructs an intelligent prediction process with deep image-text coupling, quantitative projection of psychological space, and accurate matching of popularity curves, achieving comprehensive perception and accurate judgment of content popularity in complex multimodal environments. This invention possesses advantages such as sufficient feature interaction, accurate quantification of psychological resonance, reasonable evolution pattern matching, and strong interpretability of prediction results. In particular, the MLP-Mixer model breaks down modal barriers, achieving deep aggregation of image and text features; the improved DCCA network introduces a covariance reconstruction loss function, effectively constraining the loss of structural information during the mapping process, ensuring the maximization of feature correlation and the integrity of the distribution structure in the common subspace, thereby effectively solving problems such as superficial multimodal fusion, lack of audience psychological mapping, and difficulty in capturing the evolutionary patterns of popularity in existing methods.

[0007] A method for predicting the popularity of self-media promotion content based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect the original multimodal data of self-media promotion content, perform word segmentation and word vector mapping on the text data, output the text word vector sequence, preprocess the image data, and output standard size images. S2. Perform entity linking and semantic expansion on the text word vector sequence, construct a semantic structure graph, extract deep features, and output the text semantic feature vector; S3. Perform salient target detection and background stripping on standard-sized images, preserve the foreground subject area, extract features from the foreground subject area, and output a visual salient feature map. S4. The MLP-Mixer model is used to map the text semantic feature vector and the visual salient feature map into a feature sequence of the same dimension. Through the internally stacked Token-mixing layer and Channel-mixing layer, information aggregation and interaction are performed on the feature sequence along the spatial dimension and the channel dimension, respectively. After dimensionality reduction, the image-text coupled feature vector is output. S5. Construct an improved DCCA network, map the image-text coupled feature vectors and the pre-constructed audience psychological standard feature set to a common subspace, introduce a covariance reconstruction loss function, constrain the feature covariance matrix after mapping to retain the distribution structure of the original input, calculate the maximum relevance projection, construct a multi-dimensional audience psychological space, output the sensitivity projection values ​​of image-text features on each psychological dimension coordinate axis, and generate psychological resonance index features. S6. Perform time-series evolution prediction on the psychological resonance index features, match the standard curve template in the popularity evolution pattern library, and output the popularity prediction results.

[0008] Optionally, S1 specifically includes: S11. Obtain raw multimodal data containing text and image information from the self-media platform, segment the text information to obtain a word sequence, calculate the distributed vector representation of each word in the word sequence, convert the words into real number vectors, and output the text word vector sequence. S12. Read the original pixel values ​​of the image information, normalize and scale the pixel values ​​to between 0 and 1, perform interpolation calculations on the pixel matrix of the image information according to the target size, adjust the number of rows and columns of the image, and output a standard-sized image.

[0009] Optionally, S2 specifically includes: S21. Extract words one by one from the text word vector sequence, obtain the public knowledge base by downloading entity names and entity attributes through accessing the public data interface of the network, match the words with the entity list downloaded from the public knowledge base, select the successfully matched entities as the link results, and complete the entity linking; S22. Search for the attribute description text and associated entity name corresponding to the linked entity in the public knowledge base, perform word segmentation and word vector mapping on the attribute description text and associated entity name, and insert the obtained real number vector into the original text word vector sequence to complete semantic expansion. S23. Set all entities in the text as nodes of the graph, determine whether two entities appear in the same text content, if they do, determine that the two entities are related, draw a connecting line between the nodes corresponding to the two entities, and construct a semantic structure graph. S24. Read the vector value of each node in the semantic structure graph and the vector values ​​of all adjacent nodes directly connected by the corresponding node through the connecting line. Calculate the arithmetic mean of the vector values ​​of all adjacent nodes. Concatenate the calculated arithmetic mean with the vector value of the node itself and output the text semantic feature vector.

[0010] Optionally, S3 specifically includes: S31. Read a standard-sized image, calculate the sum of squared differences between the values ​​of each pixel in the red, green, and blue channels and the global average values ​​of the corresponding channels, mark the pixels whose sum of squared differences is greater than the average sum of squared differences of all pixels in the entire image as salient pixels, and determine the region formed by all salient pixels as the salient target region. S32. Set the pixel values ​​of all pixels in the standard-sized image except for the salient target area to 0, remove the background information, and retain the foreground subject area; S33. Calculate the difference between adjacent pixels in the horizontal direction of each pixel in the foreground main area as the horizontal gradient value, calculate the difference between adjacent pixels in the vertical direction as the vertical gradient value, calculate the gradient magnitude and gradient direction of the corresponding pixel based on the horizontal gradient value and the vertical gradient value, divide the angle range from 0 degrees to 360 degrees into preset angle intervals, count the cumulative sum of the gradient magnitudes of all pixels in each angle interval, and generate a gradient direction histogram. S34. Divide the foreground subject region into several image blocks of the same size, concatenate the gradient direction histograms of each image block into a vector, and connect the vectors of all image blocks in order to output a visual salient feature map.

[0011] Optionally, S4 specifically includes: S41. Read the text semantic feature vector and the visual salient feature map, expand the visual salient feature map into a one-dimensional pixel row vector, map the text semantic feature vector and the one-dimensional pixel row vector to the same dimension value through matrix multiplication, and concatenate the mapped text semantic feature vector and the one-dimensional pixel row vector to generate a feature sequence of the same dimension. S42. Input the feature sequence of uniform dimension into the MLP-Mixer model, calculate the mean and standard deviation of each row of the feature sequence, subtract the mean of the row for each value in the feature sequence and divide by the standard deviation of the row to output the standardized feature matrix, and transpose the rows and columns of the standardized feature matrix to obtain the transposed feature matrix. S43. Input the transposed feature matrix into the Token-mixing layer, calculate the first intermediate result by the linear transformation and GELU activation function of the first layer, and obtain the spatial aggregation feature matrix by the linear transformation of the first intermediate result. S44. Perform matrix addition on the spatial aggregated feature matrix and the transposed feature matrix to obtain the first residual feature matrix. Calculate the mean and standard deviation of each row of the first residual feature matrix. Subtract the mean of the row from each value in the first residual feature matrix and divide by the standard deviation of the row to output the first normalized feature matrix. S45. Input the first normalized feature matrix into the Channel-mixing layer, calculate the linear transformation of the first normalized feature matrix and the GELU activation function to calculate the second intermediate result, and calculate the linear transformation of the second intermediate result to obtain the channel interaction feature matrix. S46. Perform matrix addition on the channel interaction feature matrix and the first normalized feature matrix to obtain the second residual feature matrix, which is used as the mixed feature sequence. S47. Repeat steps S42 to S46, using the mixed feature sequence obtained in the previous round as the input for the next round, until the set number of iterations is reached. Read the mixed feature sequence output in the last round, calculate the arithmetic mean of all column vectors in the mixed feature sequence to obtain the global feature vector, input the global feature vector into the fully connected layer, and output the image-text coupling feature vector.

[0012] Optionally, S5 specifically includes: S51. Read the image-text coupling feature vector, obtain the audience psychological dimension score data obtained by statistically analyzing historical high-popularity content samples through questionnaire surveys, normalize the audience psychological dimension score data, arrange the processed data into a matrix form, obtain the audience psychological standard feature set, and randomly generate two matrices with the number of rows equal to the input feature dimension and the number of columns equal to the common subspace dimension value, which are used as the first initial projection matrix and the second initial projection matrix, respectively. S52. Calculate the matrix product of the image-text coupling feature vector and the first initial projection matrix, and output the first projection feature vector. Calculate the matrix product of the audience psychological standard feature set and the second initial projection matrix, and output the second projection feature vector. S53. Calculate the covariance matrix of the first projected eigenvector as the first covariance matrix, calculate the covariance matrix of the second projected eigenvector as the second covariance matrix, perform eigenvalue decomposition on the first covariance matrix and the second covariance matrix respectively, select the eigenvectors corresponding to the K largest eigenvalues ​​to construct an orthogonal basis, multiply the orthogonal basis with the corresponding transpose matrix to obtain a low-rank approximate matrix, and replace the original covariance matrix to complete the reconstruction of the first covariance matrix and the second covariance matrix and output it. S54. Introduce the covariance reconstruction loss function, calculate the first matrix difference between the reconstructed first covariance matrix and the original first covariance matrix, calculate the second matrix difference between the reconstructed second covariance matrix and the original second covariance matrix, calculate the sum of squares of all elements in the first matrix difference and the second matrix difference respectively, take the square root of the sum of squares to obtain the Frobenius norm, and add them together to obtain the covariance reconstruction loss value. S55. Calculate the Pearson correlation coefficient between the first projected eigenvector and the second projected eigenvector. Multiply the Pearson correlation coefficient by negative one to obtain the correlation loss value. Add the correlation loss value to the covariance reconstruction loss value and output the total loss value. S56. Calculate and output the psychological resonance index feature using the total loss value on the first initial projection matrix and the second initial projection matrix.

[0013] Optionally, S56 specifically includes: S561. Calculate the partial derivatives of the total loss value with respect to the first initial projection matrix and the second initial projection matrix as the gradient value. Subtract the product of the gradient value and the preset learning rate value from the first initial projection matrix and the second initial projection matrix respectively. Repeat the subtraction operation until the total loss value is less than the preset total loss threshold. Output the updated first target projection matrix and the second target projection matrix. S562. Multiply the image-text coupled feature vector with the first target projection matrix to obtain the image-text feature representation mapped to the common subspace. Multiply the audience psychological standard feature set with the second target projection matrix to obtain the psychological baseline feature representation mapped to the common subspace. Calculate the dot product value between the image-text feature representation and the psychological baseline feature representation, and output the sensitivity projection value of the image-text features on each psychological dimension coordinate axis. S563. Input the sensitivity projection value into the Sigmoid function, calculate the exponent of the negative sensitivity projection value with the natural constant e as the base, add one to the calculation result to get the denominator value, calculate one and divide by the denominator value to output the psychological resonance index feature.

[0014] Optionally, S6 specifically includes: S61. Read the popularity numerical sequence of known self-media promotion content samples from the historical database, normalize the popularity numerical sequence, calculate the similarity distance between normalized numerical sequences of different samples using the dynamic time warping algorithm, classify samples with similarity distance less than the preset distance threshold into the same evolutionary pattern cluster, calculate the arithmetic mean of all numerical sequences in the evolutionary pattern cluster to generate a standard curve template, label the peak popularity value, popularity duration and popularity change trajectory label, and summarize to build a popularity evolutionary pattern library; S62. Read the psychological resonance index features, arrange the psychological resonance index features into a feature sequence according to the preset time step, calculate the first difference of the values ​​of adjacent time points in the feature sequence, obtain the feature change rate, and concatenate the feature sequence and the feature change rate to generate a time-series evolution feature curve. S63. Retrieve the heat evolution model library, read several standard curve templates stored in the heat evolution model library, calculate the numerical difference between the time series evolution characteristic curve and each standard curve template at each time point, sum the squares of the numerical differences at all time points and take the square root to obtain the Euclidean distance between the time series evolution characteristic curve and each standard curve template. S64. Compare all the calculated Euclidean distance values, select the standard curve template corresponding to the Euclidean distance with the smallest value as the target template for successful matching, read the pre-bound label data of the target template, use the label data corresponding to the target template as the prediction output result, and output the heat prediction result.

[0015] The beneficial effects of this invention are: This invention constructs a semantic structure graph by linking entities and semantically expanding text word vector sequences. Addressing the issues of semantic sparsity and lack of background knowledge in self-media texts, it employs a graph structure to extract deep features and generate text semantic feature vectors. By performing salient object detection and background stripping on standard-sized images, it addresses the problem of complex backgrounds interfering with subject feature extraction, preserving the foreground subject region and extracting gradient direction histograms to output visually salient feature maps. An MLP-Mixer model maps the text semantic feature vectors and visually salient feature maps into a unified-dimensional feature sequence. Combining Token-mixing and Channel-mixing layers, it performs information aggregation and interaction operations across spatial and channel dimensions, outputting image-text coupled feature vectors. By constructing an improved DCCA network and introducing a covariance reconstruction loss function, it addresses the issues of easily losing feature distribution structure and weak psychological dimension correlation in traditional mapping methods by performing maximum correlation projection calculation between image-text features and psychological features, generating a psychological resonance index feature. In the popularity prediction stage, the psychological resonance index features are input into the temporal evolution prediction network, matched with standard curve templates in the popularity evolution pattern library, and the output includes the peak popularity value, the duration of popularity, and the trajectory of popularity change. Ultimately, this achieves deep semantic understanding of self-media promotion content, precise quantification of audience psychology, and accurate prediction of popularity evolution trends, effectively improving the tightness of multimodal feature fusion, the accuracy of psychological resonance calculation, and the practical value of popularity prediction results. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for predicting the popularity of self-media promotion content based on deep learning, as proposed in this invention. Figure 2 This is a flowchart of the deep coupling and interactive aggregation of image and text features based on the MLP-Mixer model proposed in this invention. Figure 3 This is a flowchart of the psychological resonance index feature generation process proposed in this invention, based on an improved DCCA network and covariance reconstruction loss. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A method for predicting the popularity of self-media promotion content based on deep learning includes the following steps: S1. Collect the original multimodal data of self-media promotion content, perform word segmentation and word vector mapping on the text data in the original multimodal data, output the text word vector sequence, and perform normalization and size scaling on the image data to output a standard size image. S2. Perform entity linking and semantic expansion on the text word vector sequence, construct a semantic structure graph containing entity attributes and relational dependencies, extract deep features, and output text semantic feature vectors. S3. Perform salient target detection and background stripping on standard-sized images, preserve the foreground subject area, extract features from the foreground subject area, and output a visual salient feature map. S4. The MLP-Mixer model is used to map the text semantic feature vector and the visual salient feature map into a feature sequence of the same dimension. Through the internally stacked Token-mixing layer and Channel-mixing layer, information aggregation and interaction are performed on the feature sequence along the spatial dimension and the channel dimension, respectively. After dimensionality reduction, the image-text coupled feature vector is output. S5. Construct an improved DCCA network, map the image-text coupled feature vectors and the pre-constructed audience psychological standard feature set to a common subspace, introduce a covariance reconstruction loss function, constrain the feature covariance matrix after mapping to retain the distribution structure of the original input, calculate the maximum correlation projection between the image-text feature representation and the psychological feature representation, construct a multi-dimensional audience psychological space, output the sensitivity projection values ​​of image-text features on each psychological dimension coordinate axis, and generate the psychological resonance index feature through a non-linear activation function. S6. Perform time-series evolution prediction on the psychological resonance index features, match the standard curve template in the heat evolution pattern library, and output the heat prediction results including the heat peak value, heat duration and heat change trajectory.

[0019] This invention significantly improves the accuracy and deep semantic understanding capabilities of predicting the popularity of self-media content. By constructing a semantic structure graph and extracting visually salient features, it achieves deep cleaning and key information enhancement of unstructured text and image data, effectively filtering out background noise interference. Utilizing an MLP-Mixer model for text and image feature coupling breaks down the heterogeneity barriers between multimodal data, ensuring sufficient feature interaction. Particularly in psychological resonance modeling, an improved DCCA network introduces covariance reconstruction loss, maximizing the correlation between content features and audience psychological standard features while preserving the integrity of the original data's distribution structure, successfully quantifying the deep mechanism by which content evokes audience psychological resonance. The psychological resonance index feature generated by this mechanism compensates for the shortcomings of traditional methods that neglect the audience's psychological dimension. Combining temporal evolution prediction and pattern matching, it can accurately output the peak popularity and its trajectory. The overall solution achieves intelligent processing throughout the entire process from data collection and feature fusion to psychological mapping, significantly improving the efficiency of promotion resource allocation and return on investment, demonstrating extremely high practical value in the complex and ever-changing online content ecosystem.

[0020] In this embodiment, S1 specifically includes: S11. Obtain raw multimodal data containing text and image information from the self-media platform, segment the text information to obtain a word sequence, calculate the distributed vector representation of each word in the word sequence, convert the words into real number vectors, and output the text word vector sequence. S12. Read the original pixel values ​​of the image information, normalize and scale the pixel values ​​to between 0 and 1, perform interpolation calculations on the pixel matrix of the image information according to the target size, adjust the number of rows and columns of the image, and output a standard-sized image with uniform resolution.

[0021] In this embodiment, S2 specifically includes: S21. Extract words one by one from the text word vector sequence, obtain the public knowledge base by downloading entity names and entity attributes through accessing the public data interface of the network, match the words with the entity list downloaded from the public knowledge base, select the successfully matched entities as the link results, and complete the entity linking; S22. Search for the attribute description text and associated entity name corresponding to the linked entity in the public knowledge base, perform word segmentation and word vector mapping on the attribute description text and associated entity name, and insert the obtained real number vector into the original text word vector sequence to complete semantic expansion. S23. Set all entities in the text as nodes of the graph, determine whether two entities appear in the same text content, if they do, determine that the two entities have a relationship, draw a connecting line between the nodes corresponding to the two entities, and construct a semantic structure graph containing entity attributes and relational dependencies. S24. Read the vector value of each node in the semantic structure graph and the vector values ​​of all adjacent nodes directly connected by the corresponding node through the connecting line. Calculate the arithmetic mean of the vector values ​​of all adjacent nodes. Concatenate the calculated arithmetic mean with the vector value of the node itself and output the concatenated numerical sequence as the text semantic feature vector.

[0022] In this embodiment, S3 specifically includes: S31. Read a standard-sized image, calculate the sum of squared differences between the values ​​of each pixel in the red, green, and blue channels and the global average values ​​of the corresponding channels, mark the pixels whose sum of squared differences is greater than the average sum of squared differences of all pixels in the entire image as salient pixels, and determine the region formed by all salient pixels as the salient target region. S32. Set the pixel values ​​of all pixels in the standard-sized image except for the salient target area to 0, remove the background information, and retain the foreground subject area; S33. Calculate the difference between adjacent pixels in the horizontal direction for each pixel in the foreground main area as the horizontal gradient value, calculate the difference between adjacent pixels in the vertical direction as the vertical gradient value, calculate the gradient magnitude and gradient direction of the corresponding pixel based on the horizontal and vertical gradient values, divide the angle range from 0 degrees to 360 degrees into a preset number of angle intervals, and calculate the cumulative sum of the gradient magnitudes of all pixels in each angle interval to generate a gradient direction histogram. The preset number of angle intervals is 9. S34. Divide the foreground subject region into several image blocks of the same size, concatenate the gradient direction histograms of each image block into a vector, and connect the vectors of all image blocks in order to output a visual salient feature map.

[0023] In this embodiment, S4 specifically includes: S41. Read the text semantic feature vector and the visual salient feature map, expand the visual salient feature map into a one-dimensional pixel row vector, map the text semantic feature vector and the one-dimensional pixel row vector to the same dimension value through matrix multiplication, and concatenate the mapped text semantic feature vector and the one-dimensional pixel row vector to generate a feature sequence of the same dimension. S42. Input the feature sequence of uniform dimension into the MLP-Mixer model, calculate the mean and standard deviation of each row of the feature sequence, subtract the mean of the row for each value in the feature sequence and divide by the standard deviation of the row to output the standardized feature matrix, and transpose the rows and columns of the standardized feature matrix to obtain the transposed feature matrix. S43. Input the transposed feature matrix into the Token-mixing layer, calculate the first intermediate result by the linear transformation and GELU activation function of the first layer, and obtain the spatial aggregation feature matrix by the linear transformation of the first intermediate result. S44. Perform matrix addition on the spatial aggregated feature matrix and the transposed feature matrix to obtain the first residual feature matrix. Calculate the mean and standard deviation of each row of the first residual feature matrix. Subtract the mean of the row from each value in the first residual feature matrix and divide by the standard deviation of the row to output the first normalized feature matrix. S45. Input the first normalized feature matrix into the Channel-mixing layer, calculate the linear transformation of the first normalized feature matrix and the GELU activation function to calculate the second intermediate result, and calculate the linear transformation of the second intermediate result to obtain the channel interaction feature matrix. S46. Perform matrix addition on the channel interaction feature matrix and the first normalized feature matrix to obtain the second residual feature matrix, which is used as the mixed feature sequence. S47. Repeat steps S42 to S46, using the mixed feature sequence obtained in the previous round as the input for the next round, until the set number of loops is reached. Read the mixed feature sequence output in the last round, calculate the arithmetic mean of all column vectors in the mixed feature sequence to obtain the global feature vector, input the global feature vector into the fully connected layer, and output the image-text coupled feature vector. The set number of loops is 8 times.

[0024] This implementation introduces the MLP-Mixer network as the core fusion technology, which has significant differences and advantages compared to traditional multimodal modeling methods such as CNN, RNN, and Transformer. Traditional CNNs are limited by local receptive fields, making it difficult to capture global semantic relationships; RNN-like models suffer from low efficiency in sequential computation and the vanishing gradient problem for long-range dependencies; while the standard Transformer has global modeling capabilities, the computational complexity of its self-attention mechanism increases quadratically with the sequence length, facing huge computational overhead and inference latency when processing high-dimensional image and text features. MLP-Mixer completely abandons convolution and self-attention mechanisms, achieving feature interaction only through matrix multiplication, significantly reducing computational complexity while ensuring performance.

[0025] This invention, through detailed steps S41 to S47, utilizes the unique stacked structure of MLP-Mixer's alternating Token-mixing and Channel-mixing layers to achieve efficient deep coupling of image and text features. The linear transformations and GELU activation functions in S43 and S45, combined with residual connections and layer normalization operations in S44 and S46, effectively alleviate the gradient vanishing problem in deep networks, ensuring the smooth transmission of feature information. In particular, the eight iterations set in S46 allow the model to repeatedly aggregate and interact information in both spatial and channel dimensions, fully exploring the potential correlation between textual semantics and visual salient features. Finally, the global average pooling in S47 generates a compact image-text coupled feature vector, which not only preserves rich multimodal information but also significantly improves the robustness and generalization ability of the features, laying a solid foundation for the accurate calculation of the subsequent psychological resonance index.

[0026] In this embodiment, S5 specifically includes: S51. Read the image-text coupling feature vector, obtain the audience psychological dimension score data obtained by statistically analyzing historical high-popularity content samples through questionnaire surveys, normalize the audience psychological dimension score data, arrange the processed data into a matrix form, obtain the audience psychological standard feature set, and randomly generate two matrices with the number of rows equal to the input feature dimension and the number of columns equal to the common subspace dimension value, which are used as the first initial projection matrix and the second initial projection matrix, respectively. S52. Calculate the matrix product of the image-text coupling feature vector and the first initial projection matrix, and output the first projection feature vector. Calculate the matrix product of the audience psychological standard feature set and the second initial projection matrix, and output the second projection feature vector. S53. Calculate the covariance matrix of the first projected eigenvector as the first covariance matrix, calculate the covariance matrix of the second projected eigenvector as the second covariance matrix, perform eigenvalue decomposition on the first covariance matrix and the second covariance matrix respectively, select the eigenvectors corresponding to the K largest eigenvalues ​​to construct an orthogonal basis, multiply the orthogonal basis with the corresponding transpose matrix to obtain a low-rank approximate matrix, and replace the original covariance matrix to complete the reconstruction of the first covariance matrix and the second covariance matrix and output it. The value of K is 10. S54. Introduce the covariance reconstruction loss function, calculate the first matrix difference between the reconstructed first covariance matrix and the original first covariance matrix, calculate the second matrix difference between the reconstructed second covariance matrix and the original second covariance matrix, calculate the sum of squares of all elements in the first matrix difference and the second matrix difference respectively, take the square root of the sum of squares to obtain the Frobenius norm, and add them together to obtain the covariance reconstruction loss value. S55. Calculate the Pearson correlation coefficient between the first projected eigenvector and the second projected eigenvector. Multiply the Pearson correlation coefficient by negative one to obtain the correlation loss value. Add the correlation loss value to the covariance reconstruction loss value and output the total loss value. S56. Calculate and output the psychological resonance index feature using the total loss value on the first initial projection matrix and the second initial projection matrix.

[0027] In this embodiment, S56 specifically includes: S561. Calculate the partial derivatives of the total loss value with respect to the first initial projection matrix and the second initial projection matrix as the gradient value. Subtract the product of the gradient value and the preset learning rate value from the first initial projection matrix and the second initial projection matrix respectively. Repeat the subtraction operation until the total loss value is less than the preset total loss threshold. Output the updated first target projection matrix and the second target projection matrix. The preset learning rate value is 0.001 and the preset total loss threshold is 0.01. S562. Multiply the image-text coupled feature vector with the first target projection matrix to obtain the image-text feature representation mapped to the common subspace. Multiply the audience psychological standard feature set with the second target projection matrix to obtain the psychological baseline feature representation mapped to the common subspace. Calculate the dot product value between the image-text feature representation and the psychological baseline feature representation, and output the sensitivity projection value of the image-text features on each psychological dimension coordinate axis. S563. Input the sensitivity projection value into the Sigmoid function, calculate the exponent of the negative sensitivity projection value with the natural constant e as the base, add one to the calculation result to get the denominator value, calculate one and divide by the denominator value, and output the psychological resonance index feature with a value range between 0 and 1.

[0028] This invention constructs an improved DCCA network and combines it with a covariance reconstruction loss mechanism to achieve deep correlation mining and quantification mapping between text / image content features and audience psychological features. The coupled text / image feature vectors and the audience's psychological standard feature set are mapped to a common subspace. A low-rank approximation matrix is ​​constructed using eigenvalue decomposition, effectively constraining the structural integrity of the mapped feature covariance matrix and solving the problem of traditional methods easily losing the original data distribution characteristics during projection. By introducing Frobenius norm quantification of reconstruction error and jointly optimizing it with Pearson correlation loss, the stability of the feature space is ensured while maximizing feature correlation. The target projection matrix obtained through gradient iteration updates can accurately capture the sensitivity projection of text / image features in a multi-dimensional psychological space. Finally, the psychological resonance index feature generated by the Sigmoid activation function achieves a standardized measure of the deep appeal of content, significantly improving the interpretability and robustness of the popularity prediction model in complex semantic scenarios.

[0029] The improved DCCA network of this invention is similar to the original DCCA network in that it retains the core idea of ​​canonical correlation analysis combined with a deep learning architecture. Specifically, it constructs two parallel neural network branches to process input data from different modalities and maps them to a common subspace. Both rely on projection matrices to project high-dimensional features into a low-dimensional common space, and their core objective is to maximize the linear correlation between the two view projection vectors. The covariance matrix of the projection vectors is used to measure the alignment of features, and the gradient descent algorithm is used to iteratively optimize the projection matrix parameters to achieve deep correlation mining of cross-modal features.

[0030] The difference lies in that this invention breaks through the limitation of the original DCCA network, which only focuses on maximizing the correlation of projected features while ignoring the preservation of the original data distribution structure, and introduces an innovative covariance reconstruction mechanism. Building upon the original DCCA model's direct use of the projection vector covariance matrix for eigenvalue decomposition to solve for correlation, this invention performs a low-rank approximation reconstruction operation in step S53. After eigenvalue decomposition of the covariance matrix, the first K eigenvalues ​​are selected to construct an orthogonal basis, generating a low-rank approximation matrix that replaces the original covariance matrix, thus completing the matrix reconstruction. Immediately following, in step S54, a covariance reconstruction loss function is constructed. The Frobenius norm is used to accurately calculate the difference between the reconstructed matrix and the original covariance matrix, and this difference is added as a loss term to the total loss function, forcing the model to minimize reconstruction error while optimizing correlation.

[0031] Based on the aforementioned improvements, the beneficial effect of this invention lies in the fact that by introducing covariance reconstruction loss constraints, the improved DCCA network can effectively constrain the destruction of the original feature distribution structure during the projection process, avoiding the feature distribution distortion or loss of key information that may occur in traditional DCCA when pursuing high relevance. This mechanism ensures that the mapped features, while maximally aligning with the audience's psychological dimension, still faithfully preserve the original manifold structure and detailed features of the text and image data, significantly enhancing the robustness and integrity of the feature representation. Combined with the Pearson correlation loss in step S55, this invention achieves dual optimization of feature alignment accuracy and structure preservation capability, making the final generated psychological resonance index feature more interpretable and discriminative, thereby greatly improving the accuracy of subsequent popularity prediction tasks.

[0032] In this embodiment, S6 specifically includes: S61. Read the popularity numerical sequence of known self-media promotion content samples from the historical database, normalize the popularity numerical sequence, calculate the similarity distance between normalized numerical sequences of different samples using the dynamic time warping algorithm, classify samples with similarity distance less than a preset distance threshold into the same evolutionary pattern cluster, calculate the arithmetic mean of all numerical sequences within the evolutionary pattern cluster to generate a standard curve template, label the peak popularity value, popularity duration and popularity change trajectory labels, and summarize to construct a popularity evolutionary pattern library. The preset distance threshold is set to 0.15. S62. Read the psychological resonance index features, arrange the psychological resonance index features into a feature sequence according to the preset time step, calculate the first difference of the values ​​of adjacent time points in the feature sequence, obtain the feature change rate, and concatenate the feature sequence and the feature change rate to generate a time-series evolution feature curve. S63. Retrieve the heat evolution model library, read several standard curve templates stored in the heat evolution model library, calculate the numerical difference between the time series evolution characteristic curve and each standard curve template at each time point, sum the squares of the numerical differences at all time points and take the square root to obtain the Euclidean distance between the time series evolution characteristic curve and each standard curve template. S64. Compare all the calculated Euclidean distance values, select the standard curve template corresponding to the Euclidean distance with the smallest value as the target template for successful matching, read the pre-bound tag data of the target template, the tag data includes the peak value of popularity, the duration of popularity and the trajectory of popularity change, take the tag data corresponding to the target template as the prediction output result, and output the popularity prediction result including the peak value of popularity, the duration of popularity and the trajectory of popularity change.

[0033] Example 1: To verify the feasibility of this invention in the field of self-media content popularity prediction and promotion strategy optimization, this invention was deployed in the intelligent promotion decision support system of the regional marketing service center of a well-known short video social platform in China. This platform covers a nationwide content distribution network with over 80 million daily active users and more than 150,000 new self-media promotion content entries daily, covering multiple vertical fields such as beauty, technology, food, finance, and tourism. In such a high-concurrency, high-data-volume environment, content creators and operators face significant decision-making dilemmas: traditional popularity prediction mainly relies on the subjective experience of operators or simple linear regression models, focusing only on the statistical patterns of single historical data such as likes and shares. It fails to deeply understand the complex relationship between textual emotional color, visual composition impact, and the audience's deep psychological needs, resulting in a large amount of high-quality content being buried due to inappropriate promotion strategies, while some low-quality content wastes valuable promotion resources due to misjudgment of popularity, leading to a consistently low overall return on investment in promotion.

[0034] In practical deployment, the method of this invention first constructs a deep feature extraction and fusion process for multimodal content. The system accesses the title text, keyframe images, and historical interaction data of the video to be published in real time. For the text portion, a semantic structure graph is constructed using entity linking technology to supplement background knowledge; for the image portion, saliency detection is used to remove complex backgrounds and extract core visual features. Particularly noteworthy is that this invention utilizes the MLP-Mixer model to break down the modal barriers between text and images. Through deep interaction between Token-mixing and Channel-mixing layers, deep coupling of text and image features is achieved, effectively solving the problem of superficial multimodal information fusion in traditional methods. Building upon this, the system introduces an improved DCCA network, innovatively incorporating a covariance reconstruction loss function to map high-dimensional features to the audience's psychological resonance subspace. This network not only maximizes the correlation between content features and audience psychological features (such as curiosity, empathy, and thirst for knowledge) but also preserves the distribution characteristics of the original data by constraining the structural integrity of the projection matrix, thereby generating an accurate "psychological resonance index feature." This feature quantifies the degree to which content resonates with the audience's psychology, becoming the core driving force for popularity prediction. Finally, the system inputs the psychological resonance index into the time-series evolution prediction network, matches it with the historical popularity evolution pattern library, and outputs a comprehensive prediction report containing the peak popularity, duration, and spread trajectory, providing a basis for decision-making on promotion strategies.

[0035] After three months of field operation and data tracking, this invention has achieved significant breakthroughs in key indicators such as content popularity prediction accuracy, promotion resource utilization, and return on investment. Table 1 below details the performance comparison data between this invention and the platform's existing traditional recommendation algorithm in content promotion across different vertical fields: Table 1. Performance Comparison Data Between the Invention and Traditional Heat Prediction Methods

[0036] Based on the comparative data shown in Table 1, it can be seen that the self-media promotion content popularity prediction method proposed in this invention, based on deep learning, shows significant performance advantages over traditional methods in predicting the performance of content in multiple fields. In particular, it has achieved comprehensive improvements in key indicators such as the accuracy of popularity peak prediction, the matching degree of psychological resonance index, user retention rate, and return on investment.

[0037] Regarding the accuracy of peak popularity prediction, this invention maintains a high level of over 89% across all five vertical content categories, significantly exceeding the average accuracy of traditional methods (approximately 68%). For example, in the "finance" field, traditional methods rely on simple linear extrapolation of readership, neglecting in-depth analysis of market sentiment and audience curiosity, achieving only 65.4% accuracy. In contrast, this invention constructs a psychological resonance subspace using an improved DCCA network, precisely quantifying the correlation between content and audience psychology, achieving 90.2% accuracy and effectively enhancing the ability to identify high-potential content.

[0038] In terms of matching accuracy of the psychological resonance index, this invention maintains a success rate above 86% across all scenarios, while traditional methods typically hover around 50%. Traditional methods lack effective modeling of the audience's psychological dimensions, making it difficult to capture the deep emotional resonance evoked by content. This invention, however, utilizes an MLP-Mixer model to achieve deep coupling of text and image features, and optimizes the projection structure using a covariance reconstruction loss function, resulting in more accurate quantification of psychological resonance and ensuring the interpretability of the popularity prediction results.

[0039] It also demonstrates significant advantages in terms of return on investment (ROI) and resource utilization efficiency. This invention increases the average resource utilization rate to approximately 85%, doubling the ROI. Traditional methods often employ a "scattershot" approach, resulting in significant resource waste and ROIs generally below 1:3.5. This invention uses a precise weighted targeting based on psychological resonance index; for example, in the "technology" sector, the ROI reaches as high as 1:6.2, significantly reducing ineffective promotion costs and improving the economic benefits of marketing conversion.

[0040] Overall, this invention, by integrating multimodal features and constructing a psychological resonance quantification mechanism, achieves efficient, accurate, and interpretable popularity prediction and strategy optimization in key content areas such as "beauty," "technology," "food," "finance," and "tourism," demonstrating broad practical value and commercial potential. Particularly noteworthy is its significantly improved accuracy and resource optimization capabilities in audience-driven prediction, filling the technical gaps in deep semantic understanding and psychological dimension modeling inherent in traditional methods.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the popularity of self-media promotion content based on deep learning, characterized in that, Includes the following steps: S1. Collect the original multimodal data of self-media promotion content, perform word segmentation and word vector mapping on the text data, output the text word vector sequence, preprocess the image data, and output standard size images. S2. Perform entity linking and semantic expansion on the text word vector sequence, construct a semantic structure graph, extract deep features, and output the text semantic feature vector; S3. Perform salient target detection and background stripping on standard-sized images, preserve the foreground subject area, extract features from the foreground subject area, and output a visual salient feature map. S4. The MLP-Mixer model is used to map the text semantic feature vector and the visual salient feature map into a feature sequence of the same dimension. Through the internally stacked Token-mixing layer and Channel-mixing layer, information aggregation and interaction are performed on the feature sequence along the spatial dimension and the channel dimension, respectively. After dimensionality reduction, the image-text coupled feature vector is output. S5. Construct an improved DCCA network, map the image-text coupled feature vectors and the pre-constructed audience psychological standard feature set to a common subspace, introduce a covariance reconstruction loss function, constrain the feature covariance matrix after mapping to retain the distribution structure of the original input, calculate the maximum relevance projection, construct a multi-dimensional audience psychological space, output the sensitivity projection values ​​of image-text features on each psychological dimension coordinate axis, and generate psychological resonance index features. S6. Perform time-series evolution prediction on the psychological resonance index features, match the standard curve template in the popularity evolution pattern library, and output the popularity prediction results.

2. The method for predicting the popularity of self-media promotion content based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Obtain raw multimodal data containing text and image information from the self-media platform, segment the text information to obtain a word sequence, calculate the distributed vector representation of each word in the word sequence, convert the words into real number vectors, and output the text word vector sequence. S12. Read the original pixel values ​​of the image information, normalize and scale the pixel values ​​to between 0 and 1, perform interpolation calculations on the pixel matrix of the image information according to the target size, adjust the number of rows and columns of the image, and output a standard-sized image.

3. The method for predicting the popularity of self-media promotion content based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Extract words one by one from the text word vector sequence, obtain the public knowledge base by downloading entity names and entity attributes through accessing the public data interface of the network, match the words with the entity list downloaded from the public knowledge base, select the successfully matched entities as the link results, and complete the entity linking; S22. Search for the attribute description text and associated entity name corresponding to the linked entity in the public knowledge base, perform word segmentation and word vector mapping on the attribute description text and associated entity name, and insert the obtained real number vector into the original text word vector sequence to complete semantic expansion. S23. Set all entities in the text as nodes of the graph, determine whether two entities appear in the same text content, if they do, determine that the two entities are related, draw a connecting line between the nodes corresponding to the two entities, and construct a semantic structure graph. S24. Read the vector value of each node in the semantic structure graph and the vector values ​​of all adjacent nodes directly connected by the corresponding node through the connecting line. Calculate the arithmetic mean of the vector values ​​of all adjacent nodes. Concatenate the calculated arithmetic mean with the vector value of the node itself and output the text semantic feature vector.

4. The method for predicting the popularity of self-media promotion content based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Read a standard-sized image, calculate the sum of squared differences between the values ​​of each pixel in the red, green, and blue channels and the global average values ​​of the corresponding channels, mark the pixels whose sum of squared differences is greater than the average sum of squared differences of all pixels in the entire image as salient pixels, and determine the region formed by all salient pixels as the salient target region. S32. Set the pixel values ​​of all pixels in the standard-sized image except for the salient target area to 0, remove the background information, and retain the foreground subject area; S33. Calculate the difference between adjacent pixels in the horizontal direction of each pixel in the foreground main area as the horizontal gradient value, calculate the difference between adjacent pixels in the vertical direction as the vertical gradient value, calculate the gradient magnitude and gradient direction of the corresponding pixel based on the horizontal gradient value and the vertical gradient value, divide the angle range from 0 degrees to 360 degrees into preset angle intervals, count the cumulative sum of the gradient magnitudes of all pixels in each angle interval, and generate a gradient direction histogram. S34. Divide the foreground subject region into several image blocks of the same size, concatenate the gradient direction histograms of each image block into a vector, and connect the vectors of all image blocks in order to output a visual salient feature map.

5. The method for predicting the popularity of self-media promotion content based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Read the text semantic feature vector and the visual salient feature map, expand the visual salient feature map into a one-dimensional pixel row vector, map the text semantic feature vector and the one-dimensional pixel row vector to the same dimension value through matrix multiplication, and concatenate the mapped text semantic feature vector and the one-dimensional pixel row vector to generate a feature sequence of the same dimension. S42. Input the feature sequence of uniform dimension into the MLP-Mixer model, calculate the mean and standard deviation of each row of the feature sequence, subtract the mean of the row for each value in the feature sequence and divide by the standard deviation of the row to output the standardized feature matrix, and transpose the rows and columns of the standardized feature matrix to obtain the transposed feature matrix. S43. Input the transposed feature matrix into the Token-mixing layer, calculate the first intermediate result by the linear transformation and GELU activation function of the first layer, and obtain the spatial aggregation feature matrix by the linear transformation of the first intermediate result. S44. Perform matrix addition on the spatial aggregated feature matrix and the transposed feature matrix to obtain the first residual feature matrix. Calculate the mean and standard deviation of each row of the first residual feature matrix. Subtract the mean of the row from each value in the first residual feature matrix and divide by the standard deviation of the row to output the first normalized feature matrix. S45. Input the first normalized feature matrix into the Channel-mixing layer, calculate the linear transformation of the first normalized feature matrix and the GELU activation function to calculate the second intermediate result, and calculate the linear transformation of the second intermediate result to obtain the channel interaction feature matrix. S46. Perform matrix addition on the channel interaction feature matrix and the first normalized feature matrix to obtain the second residual feature matrix, which is used as the mixed feature sequence. S47. Repeat steps S42 to S46, using the mixed feature sequence obtained in the previous round as the input for the next round, until the set number of iterations is reached. Read the mixed feature sequence output in the last round, calculate the arithmetic mean of all column vectors in the mixed feature sequence to obtain the global feature vector, input the global feature vector into the fully connected layer, and output the image-text coupling feature vector.

6. The method for predicting the popularity of self-media promotion content based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Read the image-text coupling feature vector, obtain the audience psychological dimension score data obtained by statistically analyzing historical high-popularity content samples through questionnaire surveys, normalize the audience psychological dimension score data, arrange the processed data into a matrix form, obtain the audience psychological standard feature set, and randomly generate two matrices with the number of rows equal to the input feature dimension and the number of columns equal to the common subspace dimension value, which are used as the first initial projection matrix and the second initial projection matrix, respectively. S52. Calculate the matrix product of the image-text coupling feature vector and the first initial projection matrix, and output the first projection feature vector. Calculate the matrix product of the audience psychological standard feature set and the second initial projection matrix, and output the second projection feature vector. S53. Calculate the covariance matrix of the first projected eigenvector as the first covariance matrix, calculate the covariance matrix of the second projected eigenvector as the second covariance matrix, perform eigenvalue decomposition on the first covariance matrix and the second covariance matrix respectively, select the eigenvectors corresponding to the K largest eigenvalues ​​to construct an orthogonal basis, multiply the orthogonal basis with the corresponding transpose matrix to obtain a low-rank approximate matrix, and replace the original covariance matrix to complete the reconstruction of the first covariance matrix and the second covariance matrix and output it. S54. Introduce the covariance reconstruction loss function, calculate the first matrix difference between the reconstructed first covariance matrix and the original first covariance matrix, calculate the second matrix difference between the reconstructed second covariance matrix and the original second covariance matrix, calculate the sum of squares of all elements in the first matrix difference and the second matrix difference respectively, take the square root of the sum of squares to obtain the Frobenius norm, and add them together to obtain the covariance reconstruction loss value. S55. Calculate the Pearson correlation coefficient between the first projected eigenvector and the second projected eigenvector. Multiply the Pearson correlation coefficient by negative one to obtain the correlation loss value. Add the correlation loss value to the covariance reconstruction loss value and output the total loss value. S56. Calculate and output the psychological resonance index feature using the total loss value on the first initial projection matrix and the second initial projection matrix.

7. The method for predicting the popularity of self-media promotion content based on deep learning according to claim 6, characterized in that, S56 specifically includes: S561. Calculate the partial derivatives of the total loss value with respect to the first initial projection matrix and the second initial projection matrix as the gradient value. Subtract the product of the gradient value and the preset learning rate value from the first initial projection matrix and the second initial projection matrix respectively. Repeat the subtraction operation until the total loss value is less than the preset total loss threshold. Output the updated first target projection matrix and the second target projection matrix. S562. Multiply the image-text coupled feature vector with the first target projection matrix to obtain the image-text feature representation mapped to the common subspace. Multiply the audience psychological standard feature set with the second target projection matrix to obtain the psychological baseline feature representation mapped to the common subspace. Calculate the dot product value between the image-text feature representation and the psychological baseline feature representation, and output the sensitivity projection value of the image-text features on each psychological dimension coordinate axis. S563. Input the sensitivity projection value into the Sigmoid function, calculate the exponent of the negative sensitivity projection value with the natural constant e as the base, add one to the calculation result to get the denominator value, calculate one and divide by the denominator value to output the psychological resonance index feature.

8. The method for predicting the popularity of self-media promotion content based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Read the popularity numerical sequence of known self-media promotion content samples from the historical database, normalize the popularity numerical sequence, calculate the similarity distance between normalized numerical sequences of different samples using the dynamic time warping algorithm, classify samples with similarity distance less than the preset distance threshold into the same evolutionary pattern cluster, calculate the arithmetic mean of all numerical sequences in the evolutionary pattern cluster to generate a standard curve template, label the peak popularity value, popularity duration and popularity change trajectory label, and summarize to build a popularity evolutionary pattern library; S62. Read the psychological resonance index features, arrange the psychological resonance index features into a feature sequence according to the preset time step, calculate the first difference of the values ​​of adjacent time points in the feature sequence, obtain the feature change rate, and concatenate the feature sequence and the feature change rate to generate a time-series evolution feature curve. S63. Retrieve the heat evolution model library, read several standard curve templates stored in the heat evolution model library, calculate the numerical difference between the time series evolution characteristic curve and each standard curve template at each time point, sum the squares of the numerical differences at all time points and take the square root to obtain the Euclidean distance between the time series evolution characteristic curve and each standard curve template. S64. Compare all the calculated Euclidean distance values, select the standard curve template corresponding to the Euclidean distance with the smallest value as the target template for successful matching, read the pre-bound label data of the target template, use the label data corresponding to the target template as the prediction output result, and output the heat prediction result.

Images