A marketing campaign indicator prediction model establishment method

Abstract

The application provides a marketing activity index prediction model establishment method, and belongs to the technical field of large models.The application collects multi-modal data such as text descriptions, promotional images and feature parameters of marketing activities, adopts a BERT model, a ResNet-50 model and a structured encoder to respectively extract text, visual and activity feature vectors, constructs a four-path bidirectional attention mechanism to realize deep interactive fusion of multi-modal features, establishes a static prediction module to realize total index regression prediction by stacking a residual multilayer perceptron, constructs a dynamic prediction module to realize time series index dynamic prediction by adopting unified naming multi-scale time series feature extraction and a double-branch collaborative mechanism, and designs a dynamic fusion output layer to generate a dynamic weight matrix to adaptively fuse static and dynamic prediction results, so that the technical problem of low prediction accuracy caused by insufficient fusion of multi-modal data features of marketing activities is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of large models, and in particular, relates to a marketing activity index prediction model establishment method. BACKGROUND

[0002] In the field of intelligent marketing prediction, traditional marketing activity index prediction methods mainly rely on single modal data analysis techniques, such as natural language processing methods based only on text descriptions, computer vision techniques based only on image content, or statistical analysis models based only on structured data. When dealing with comprehensive marketing scenarios containing text descriptions, promotional images, and feature parameters, these methods often use a strategy of independent processing followed by simple splicing. However, existing multi-modal fusion techniques have significant defects, mainly manifested in the lack of an effective cross-modal semantic alignment mechanism, the inability to establish deep associative relationships between different modalities, and the inability to fully utilize complementary information between text, images, and structured features. Moreover, existing methods lack multi-scale time modeling capabilities when dealing with temporal dynamic features. In current marketing activity prediction applications, due to the multi-modal characteristics of marketing data, such as rich text descriptions, diverse visual content, and complex temporal features, existing techniques are unable to achieve deep fusion and collaborative modeling of information from each modality, resulting in the inability of prediction models to accurately capture comprehensive feature representations of marketing activities, which in turn affects the accuracy of index prediction. SUMMARY

[0003] Therefore, the application provides a marketing activity index prediction model establishment method, which can solve the technical problem of low prediction accuracy caused by insufficient fusion of marketing activity multi-modal data features in the prior art.

[0004] The application is implemented in the following manner: a marketing activity index prediction model establishment method is provided, multi-modal data of a marketing activity is collected, including marketing activity text description, marketing activity promotional image and marketing activity characteristic parameter; a multi-modal feature extraction module is constructed, a text feature vector is obtained by encoding the marketing activity text description through a BERT model, a visual feature vector is obtained by encoding the marketing activity promotional image through a ResNet-50 model, and an activity feature vector is obtained by encoding the marketing activity characteristic parameter through a structured encoder; a cross-modal information dynamic fusion mechanism is designed, a four-path bidirectional attention mechanism is constructed to realize the interactive fusion between different modal feature vectors, different modal feature vectors are constructed as bipartite graph nodes through a maximum weight bipartite graph matching algorithm, and the optimal cross-modal feature pairing weight is solved to generate a multi-modal fusion representation; a static prediction module is established, the multi-modal fusion representation is input into a stacked residual multilayer perception network to realize the regression prediction of the total index of the marketing activity, and a static prediction result is obtained; a dynamic prediction module is constructed, multi-scale sliding windows are used to extract time sequence features, and a dynamic prediction of the marketing activity index sequence is realized through a double-branch collaborative mechanism of a time sequence evolution branch and a cross-modal interaction branch; a dynamic fusion output layer is designed, a dynamic weight matrix is generated through cross-modal attention, and the output of the time sequence evolution branch and the output of the cross-modal interaction branch are adaptively fused to obtain a dynamic prediction result; and historical marketing activity data is used to train and optimize the marketing activity index prediction model.

[0005] The marketing activity characteristic parameter specifically includes time characteristics and activity strategy characteristics, the time characteristics capture the natural time sequence rules of weeks, months and holidays through discrete coding, and the activity strategy characteristics depict the marketing strategy attributes of activity budget intensity, activity type and duration cycle through structured representation.

[0006] The cross-modal information dynamic fusion mechanism specifically realizes fine-grained cross-modal semantic alignment and information enhancement through differentiated query key-value pair configuration, four independent attention paths in the four-path bidirectional attention mechanism model asymmetric dependency relationships between different modalities, and the text feature vector is used as a query vector to actively align the visual content to ensure the semantic consistency of the activity description and the picture theme.

[0007] The four-path bidirectional attention mechanism specifically realizes the interactive fusion between the text feature vector and the visual feature vector, the activity feature vector and the visual feature vector, the activity feature vector and the text feature vector, and the text feature vector and the activity feature vector, generates query key-value pairs through independent projection matrices, and then calculates the attention distribution of the feature vector to other modal content through scaled dot-product attention.

[0008] The static prediction module is specifically a processing chain composed of a batch normalization layer, a self-attention layer and a residual connection, which is used for regression prediction of the total index of the marketing activity. A stacked residual multi-layer perception network includes a feature transformation path and a residual direct connection path. The feature transformation path captures nonlinear relationships through progressive mapping from 512 dimensions to 384 dimensions and then to 256 dimensions. The residual direct connection path preserves original information to prevent gradient vanishing.

[0009] The multi-scale sliding window is specifically a parallel multi-scale convolution kernel group used to extract different granularity time series patterns. A 3-day short kernel is used to capture sudden promotional fluctuations, a 7-day medium kernel is used to analyze periodic consumption rules, and a 30-day long kernel is used to identify trend-based billing patterns.

[0010] The time series evolution branch is specifically a sequence pattern capture based on a causal Transformer. A causal mask mechanism is used to constrain the attention range, ensuring that the prediction time step only accesses historical information. Time series features are aggregated and learnable position encodings are generated to generate time series representations.

[0011] The cross-modal interaction branch is specifically a feature enhancement achieved through the interaction of the time series global context and the multi-modal fusion representation. A query vector is generated through a pooling operation, and a key-value pair is generated through a multi-modal fusion representation. The interaction features are output for subsequent fusion processing.

[0012] The dynamic weight matrix is specifically an element that reflects the attention intensity of different prediction steps to the time series pattern, achieving dynamic adjustment of the importance of the mode in the prediction scenario. The dynamic weight is generated through the interaction of the projection features of the time series branch and the cross-modal interaction features.

[0013] The visual feature vector encoding is specifically a deep residual network used to extract fine-grained spatial features. The convolutional layer group of ResNet-50 is used to extract abstract features step by step. The visual feature vector is obtained through spatial attention pooling and linear projection.

[0014] The text feature vector encoding is specifically a dynamic semantic extractor based on a pre-trained language model. The BERT model is used to generate context-aware semantic vectors. The text projection matrix and the marketing activity text description are input for encoding processing.

[0015] The environment perception gating mechanism is specifically a four-path bidirectional attention mechanism that concatenates the output features of the four independent attention paths with the original activity feature vector. A learnable parameter matrix is used to generate adaptive fusion weights for the four paths. The final multi-modal fusion representation is generated through weighted combination.

[0016] The multi-scale feature extraction formula is where i represents the index of the different scale convolution kernels. For input time series data, the time series feature generation formula is Wherein is the aggregated sequence feature is the learnable position encoding.

[0017] Wherein, the query vector generation formula is The key-value pair generation formula is And The static prediction result calculation formula is The dynamic weight matrix generation formula is .

[0018] Wherein, the visual feature vector encoding formula is Wherein is the trainable projection matrix is the bias term, and the text feature vector encoding formula is Wherein is the text projection matrix is the marketing activity text description input.

[0019] Wherein, the game model includes an upper model with the goal of maximizing marketing effect and a lower model with the goal of optimizing resource allocation, and the upper model objective function is The lower model objective function is .

[0020] The present application realizes the deep interactive fusion of multi-modal features by constructing a four-channel bidirectional attention mechanism, designs a cross-modal information dynamic fusion mechanism to establish the semantic alignment relationship between text, visual and activity features, and effectively solves the technical defects of insufficient multi-modal data fusion in traditional methods. The present application uses a multi-scale sliding window to extract time series features of different granularities, realizes the organic combination of static features and dynamic features through the double-branch collaborative mechanism of the time series evolution branch and the cross-modal interaction branch, overcomes the shortcomings of the prior art in time series modeling, and significantly improves the ability to capture the complex spatio-temporal patterns of marketing activities. In summary, the present application establishes a complete multi-modal fusion prediction framework, and fundamentally solves the technical problem of low prediction accuracy caused by insufficient fusion of marketing activity multi-modal data features in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the flowchart of the method of the present application.

[0022] Figure 2 is the structure diagram of the model established by the present application.

[0023] Figure 3 is the residual multilayer perceptron block schematic diagram.

[0024] Figure 4 is a static prediction module structure diagram.

[0025] Figure 5 is a dynamic prediction module structure diagram. DETAILED DESCRIPTION

[0026] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0027] As Figure 1 shown is a flowchart of a marketing activity index prediction model establishment method provided by the present application, and the method comprises the following steps:

[0028] S01, collecting multi-modal data of a marketing activity, wherein the multi-modal data comprises marketing activity text description, marketing activity promotional image and marketing activity feature parameter, and the marketing activity feature parameter comprises time feature and activity strategy feature;

[0029] S02, constructing a multi-modal feature extraction module, encoding the marketing activity text description by a BERT model to obtain a text feature vector, encoding the marketing activity promotional image by a ResNet-50 model to obtain a visual feature vector, and encoding the marketing activity feature parameter by a structured encoder to obtain an activity feature vector;

[0030] S03, designing a cross-modal information dynamic fusion mechanism, constructing a four-path bidirectional attention mechanism to realize interactive fusion between the text feature vector and the visual feature vector, the activity feature vector and the visual feature vector, the activity feature vector and the text feature vector, and the text feature vector and the activity feature vector, constructing different modal feature vectors as bipartite graph nodes by a maximum weight bipartite graph matching algorithm and solving optimal cross-modal feature pairing weights to generate a multi-modal fusion representation;

[0031] S04, establishing a static prediction module, inputting the multi-modal fusion representation into a stacked residual multi-layer perception network, realizing regression prediction of a marketing activity total index by a processing chain composed of a batch normalization layer, a self-attention layer and a residual connection to obtain a static prediction result;

[0032] S05, constructing a dynamic prediction module, extracting time sequence features by a multi-scale sliding window, and realizing dynamic prediction of a marketing activity index sequence by a double-branch collaborative mechanism of a time sequence evolution branch and a cross-modal interaction branch;

[0033] S06, designing a dynamic fusion output layer, generating a dynamic weight matrix by cross-modal attention, adaptively fusing the output of the time sequence evolution branch and the output of the cross-modal interaction branch to obtain a dynamic prediction result.

[0034] S07, training and optimizing the marketing activity index prediction model by using historical marketing activity data, balancing the dual goals of maximizing marketing effect and optimizing resource allocation through a game model, updating model parameters using a back propagation algorithm, and using the static prediction result and the dynamic prediction result to accurately predict key indicators of marketing activities.

[0035] The structure of the model constructed by the application is as shown in Figures 2 to 5 .

[0036] The cross-modal information dynamic fusion mechanism realizes fine-grained cross-modal semantic alignment and information enhancement through differentiated query key-value pair configuration, and four independent attention paths in the four-path bidirectional attention mechanism model asymmetric dependency between different modalities, wherein the text feature vector is used as a query vector to actively align the visual content to ensure the semantic consistency of the activity description and the picture theme, the activity feature vector is used as a query to drive the attention to the key area of the picture, so that the parameter setting can locate the corresponding expression in the image, the activity feature vector actively checks the compatibility of the description text, and the text feature vector actively queries related parameter information. The multi-scale sliding window uses a parallel multi-scale convolution kernel group to extract different granularity time series patterns, captures sudden promotional fluctuations through a 3-day short kernel, analyzes periodic consumption rules through a 7-day medium kernel, and identifies trend bill patterns through a 30-day long kernel, and the multi-scale feature extraction formula is wherein i represents the index of different scale convolution kernels, is input time series data. The time series evolution branch captures sequence patterns based on causal Transformer, and adopts a causal mask mechanism to constrain the attention range, so that the prediction time step only accesses historical information, and the time series representation generation formula is wherein is the aggregated sequence feature is the learnable position encoding. The cross-modal interaction branch realizes feature enhancement through the interaction of the time series global context and the multi-modal fusion representation, and the query vector generation formula is The key-value pair generation formula is and , wherein is the multi-modal fusion representation, and the output interaction feature is used for subsequent fusion processing. The stacked residual multilayer perceptron network includes a feature transformation path and a residual direct connection path, the feature transformation path captures the nonlinear relationship through progressive mapping from 512 dimensions to 384 dimensions to 256 dimensions, and the residual direct connection path retains the original information to prevent gradient disappearance, and the static prediction result calculation formula is wherein represents the Lth residual multilayer perceptron block, is the multi-modal fusion representation. The dynamic weight matrix elements reflect the attention intensity of different prediction steps to the time sequence mode, realize the dynamic adjustment of the importance of the mode in the prediction scene, and the dynamic weight matrix generation formula is wherein is the time sequence branch projection feature is the cross-modal interaction feature.

[0037] The visual feature vector encoding adopts a deep residual network to extract fine-grained spatial features, and abstract features are extracted through the convolutional layer groups of ResNet-50 in stages, and the visual feature vector is obtained through spatial attention pooling and linear projection, and the encoding formula is wherein is the trainable projection matrix is the bias term is the marketing activity promotional image input. The text feature vector encoding is based on a dynamic semantic extractor of a pre-trained language model, and a BERT model is used to generate a context-aware semantic vector, and the encoding formula is wherein is the text projection matrix is the marketing activity text description input. The activity feature vector encoding decomposes the key factors affecting user behavior into two dimensions of time features and activity strategy features, the time features adopt discrete encoding to capture the natural time sequence law of weeks, months and holidays, and the activity strategy features depict the marketing strategy attributes of activity budget intensity, activity type and duration cycle through structured representation, and the encoding formula is wherein is the trainable parameter matrix is the time feature is the activity strategy feature.

[0038] The calculation process of the text feature vector to the visual feature vector path in the four-path bidirectional attention mechanism is to generate a query key-value pair through an independent projection matrix , , , and then calculate the attention distribution of the text feature vector to the visual content through the scaled dot product attention , and the other three paths respectively calculate the interaction representation of the activity feature vector to the visual feature vector, the activity feature vector to the text feature vector, and the text feature vector to the activity feature vector. The environmental perception gating mechanism splices the output features of the four independent attention paths in the four-path bidirectional attention mechanism with the original activity feature vector , generates adaptive fusion weights of the four paths through a learnable parameter matrix , and the final multi-modal fusion representation is generated through weighted combination .

[0039] The multi-scale feature extraction is aimed at the coupling characteristics of short-term promotion impulse and long-term periodic regularity in user consumption behavior, adopts a multi-scale causal convolution architecture to extract hierarchical time sequence features, extracts different granularity time sequence patterns through a parallel multi-scale convolution kernel group given an input sequence, and realizes cross-scale dynamic calibration based on multi-granularity features by introducing a learnable parameter matrix for dynamic feature aggregation where the weight matrix is multiplied with the feature map element by element after being broadcast along the time dimension. The dual-branch collaborative mechanism realizes dynamic and static feature decoupling by adopting a dual-branch collaborative mechanism, the time sequence evolution branch focuses on the periodic and trend changes of user behavior, and the cross-modal interaction branch models the immediate influence of marketing materials, and both realize dynamic coupling through a dynamic fusion output layer.

[0040] The dynamic fusion output layer realizes adaptive fusion of time sequence features and cross-modal information through three steps, first, the output of the time sequence evolution branch is projected in the feature space , second, a dynamic weight matrix is generated through cross-modal attention , and finally, the output layer fuses information and maps the output dimension to obtain a dynamic prediction result where the dominant mode of time sequence is strengthened and the cross-modal reference information is retained.

[0041] The game model includes an upper model with the objective of maximizing marketing effect and a lower model with the objective of optimizing resource allocation. The objective function of the upper model is , and the constraint condition is , , . The objective function of the lower model is , and the constraint condition is , . The marketing effect maximization function is used to optimize the overall revenue performance of the marketing activity, and the input includes user conversion rate derived from historical marketing activity data statistics, activity coverage derived from marketing activity feature parameters, user engagement derived from user behavior data analysis, market competition intensity derived from external market environment data, and the output is a marketing effect comprehensive score for guiding marketing strategy formulation. The resource allocation optimization function is used to minimize the resource input cost of the marketing activity, and the input includes advertising cost derived from marketing budget allocation data, human resource cost derived from personnel allocation plan, technical maintenance cost derived from system operation expenditure, risk control cost derived from risk assessment model and user conversion rate The coupling term is derived from the output of the upper model, and the output is a total cost control index for resource allocation decision. The coupling term represents the mutual influence relationship between user conversion rate and advertising cost.

[0042] Optionally, it further includes an adaptive learning rate adjustment function for adjusting the learning rate parameter of the neural network, the function calculates an adjustment factor value based on the training loss rate derived from model training process monitoring, the gradient norm derived from back propagation calculation, the model convergence speed derived from verification set performance tracking, and the historical performance index derived from model evaluation record, when the adjustment factor value belongs to the range of 0 to 0.3, it is used to adjust the learning rate parameter of the neural network by adopting an exponential decay adjustment strategy, when the adjustment factor value belongs to the range of 0.3 to 0.7, it is used to adjust the learning rate parameter of the neural network by adopting a linear decreasing adjustment strategy, when the adjustment factor value belongs to the range of 0.7 to 1.0, it is used to adjust the learning rate parameter of the neural network by adopting a cosine annealing adjustment strategy, and the adjusted learning rate parameter is used to optimize the model training effect.

[0043] Optionally, it further includes a gating weight function for adjusting the gating mechanism weight distribution of the neural network, the function calculates a balance value based on the modal feature importance score derived from attention weight analysis, the historical prediction accuracy derived from model performance record, the current training stage derived from training progress monitoring, and the model parameter complexity index derived from network structure analysis, when the balance value belongs to the range of 0 to 0.4, it is used to adjust the parameters of the gating by adopting a conservative weight adjustment function, when the balance value belongs to the range of 0.4 to 0.6, it is used to adjust the parameters of the gating by adopting an equal weight adjustment function, when the balance value belongs to the range of 0.6 to 1.0, it is used to adjust the parameters of the gating by adopting an aggressive weight adjustment function, and the adjusted gating parameters are used to optimize the multi-modal information fusion effect.

[0044] The specific implementation of the above steps is described in detail as follows.

[0045] The specific implementation of step S01 is to construct a multi-dimensional marketing activity data collection system, which realizes all-around information acquisition based on the principle of data source diversity. First, a text data collection module is established, and natural language processing technology is used to structureally extract text information such as title description, product introduction, and promotion copy of marketing activities. The key marketing elements are identified by using word segmentation algorithm and semantic analysis technology, and the unstructured text is converted into a processable digital format. Next, an image data collection module is constructed, and computer vision technology is used to automatically collect and preprocess visual content such as marketing poster, product display picture, and advertising creative material. Image enhancement algorithm is used to improve image quality and unify format specification. Then, a feature parameter collection module is established, which systematically collects time dimension features such as seasonal factor, holiday identifier, and activity cycle of marketing activities, and collects key parameters such as budget scale, activity type, target audience, and channel distribution. Finally, a data quality control mechanism is implemented to ensure the reliability of collected data through data integrity test, outlier detection, and data consistency check. A data standardization process is established to unify data from different sources to the same dimension and format standard.

[0046] The specific implementation of step S02 is to establish a three-way parallel multi-modal feature encoding architecture, which realizes the unified feature space mapping of heterogeneous data based on the principles of deep learning and representation learning. The text feature extraction path adopts the pre-trained bidirectional encoder representation technology, which captures the long-distance dependency between words in the text sequence through the multi-head self-attention mechanism of the BERT model, initializes the encoder parameters using the position encoding and mask language model pre-training weights, and maps the variable-length text sequence to a fixed-dimensional dense vector representation. The image feature extraction path adopts the deep residual network architecture, which extracts hierarchical visual features from edge texture to semantic concept through the convolution layer cascade structure of the ResNet-50 model, solves the gradient vanishing problem of deep network through the skip connection mechanism, and compresses the two-dimensional feature map to a one-dimensional feature vector through the global average pooling operation. The structured feature extraction path designs a hybrid encoding strategy for discrete and continuous parameters of marketing activities, which captures the periodic pattern of time series by combining one-hot encoding and periodic encoding for time features, and performs nonlinear transformation on activity strategy features by using dense connection network, and stabilizes the training process and accelerates convergence through batch normalization technology.

[0047] The specific implementation of step S03 is to construct an attention mechanism-based cross-modal information fusion architecture that realizes semantic alignment and information enhancement between different modal features based on the multi-sensory information integration principle in cognitive psychology. First, a four-channel bidirectional attention network is established, each channel using a triple query-key-value attention calculation mode. The input features are projected to the query space, key space and value space through a learnable linear transformation matrix, and the relevance weight distribution between different modal features is calculated using the scaled dot-product attention mechanism. The text-to-image channel uses the text features as the query vector and the image features as the key-value pair to realize active retrieval and semantic matching of the visual content by the text description, ensuring the theme consistency of the marketing copy and visual materials. The image-to-text channel uses the opposite query direction to drive the attention allocation of the text keywords through the visual features, strengthening the semantic resonance relationship between the image content and the text description. The two channels of activity features to text and image model the constraint and guidance of structured parameters on unstructured content, and realize the directional optimization of marketing strategies on creative content through parameterized attention weights. Then, an environment perception gating mechanism is constructed to concatenate the output features of the four attention channels with the original activity features, generate adaptive fusion weights through the softmax normalization function, and realize the dynamic balance and optimal combination of information in different channels. Finally, the maximum weight bipartite graph matching algorithm is used to optimize the cross-modal feature pairing, the feature vectors of different modalities are constructed as the nodes of the bipartite graph, and the Hungarian algorithm is used to solve the optimal matching scheme to ensure the maximum retention and optimal allocation of information in the feature fusion process.

[0048] The specific implementation of step S04 is to establish a static prediction architecture based on a stacked residual network, which realizes the regression prediction of marketing indicators based on the residual learning principle and the idea of ensemble learning in deep learning. First, a backbone structure of a multi-layer perception network is constructed, a progressive dimension mapping strategy of 512 dimensions to 384 dimensions and then to 256 dimensions is adopted, a nonlinear activation function is used to introduce the expression ability of the model, dropout regularization technology is added after each hidden layer to prevent overfitting, and the dropout probability is set between 0.2 and 0.3 to balance the model capacity and generalization ability. Then, a residual connection mechanism is introduced, a parallel structure of feature transformation path and identity mapping path is established in each multi-layer perception block, the feature transformation path is responsible for learning the complex nonlinear mapping relationship from input to output, the identity mapping path preserves the original feature information through direct connection, and the outputs of the two paths are fused through element-by-element addition operation to effectively alleviate the gradient vanishing problem in deep network training. Next, a processing chain composed of a batch normalization layer and a self-attention layer is added, the batch normalization layer stabilizes the activation value distribution in the training process through standardization operation, and the self-attention layer captures the correlation pattern between the internal elements of the feature vector through inner product operation to enhance the model's ability to identify important feature dimensions. Finally, a regression output layer is established, a linear transformation is used to map the high-dimensional feature vector to the numerical space of the marketing indicators, the mean square error loss function is used to optimize the prediction accuracy, and the static prediction result is output as the benchmark evaluation of the overall effect of the marketing activities.

[0049] The specific implementation of step S05 is to construct a dual-branch collaborative dynamic prediction architecture, which realizes the sequential prediction of marketing indicators based on time series analysis theory and multi-scale signal processing principles. The multi-scale sliding window module adopts a parallel convolution kernel group architecture, and one-dimensional convolution kernels of 3 days, 7 days and 30 days are designed to capture short-term fluctuations, medium-term cycles and long-term trends, respectively. Each convolution kernel adopts a causal convolution structure to ensure that only historical information is used at the prediction time, and a ReLU activation function is used to introduce a nonlinear transformation capability. The multi-scale features are fused through a weighted aggregation mechanism, and the weight matrix is broadcasted along the time dimension and then multiplied with the feature map element by element. The time evolution branch realizes long-distance dependence modeling of sequence patterns based on a causal Transformer architecture, uses a causal mask matrix to limit the range of attention calculation, ensures that each time step can only access current and historical information, and uses a multi-head attention mechanism to calculate the attention distribution of different representation subspaces in parallel. The position encoding module assigns a learnable position identifier to each time step in the sequence, enhancing the model's ability to perceive time series information. The cross-modal interaction branch establishes an interaction mechanism between the time series global context and the multi-modal fusion representation, compresses the time series features into a global context vector as the query information through pooling operation, and generates key-value pair information through linear transformation of the multi-modal fusion representation. The attention calculation result represents the contribution of different modal information to the time series prediction. The dual-branch collaborative mechanism realizes the information integration of the time evolution branch and the cross-modal interaction branch through a gating fusion strategy, dynamically adjusts the weight distribution of the two branches according to different prediction scenarios, and enhances the weight of the time series branch in the trend prediction scenario and improves the importance of the cross-modal branch in the sudden event prediction scenario.

[0050] The specific implementation of step S06 is to design a dynamic fusion output architecture of adaptive weight allocation, which realizes the optimal combination of multi-source prediction information based on attention mechanism and information fusion theory. First, the spatial projection transformation is performed on the output features of the time evolution branch, and a causal one-dimensional convolution network is used to map the time sequence features to a representation space matching the cross-modal features. The convolution kernel size is set to between 3 and 5 to balance the receptive field coverage and the computational complexity, and the zero padding strategy is used to maintain the consistency of the sequence length. Then, a cross-modal attention mechanism is constructed to generate a dynamic weight matrix. The projected time sequence features are used as the query vector, and the cross-modal interaction features are used as the key vector. The similarity scores between the features are calculated by inner product operation, and the attention weight distribution is obtained by normalization processing using the softmax function. Each element of the weight matrix reflects the importance degree of the corresponding time step and modal dimension. Next, a weighted fusion operation is implemented to perform matrix multiplication operation between the dynamic weight matrix and the time sequence projection features, highlighting the key information in the time sequence mode while retaining the cross-modal benchmark features as supplementary information. The linear combination of the two types of information is realized by element-wise addition operation. Finally, an output mapping layer is established to map the fused high-dimensional feature vector to the target prediction dimension using a multi-layer perceptron network. The fitting ability of the model is enhanced through a nonlinear activation function, and the output layer does not use an activation function to support numerical prediction in any range, obtaining the final dynamic prediction result.

[0051] The specific implementation of step S07 is to establish a multi-objective optimization training mechanism under the game theory framework, which realizes the collaborative optimization of marketing effect and resource allocation based on the double-layer optimization theory and reinforcement learning principle. The upper game model takes the maximization of marketing effect as the optimization target, and the objective function adopts a polynomial combination form to comprehensively consider four key factors of user conversion rate, activity coverage, user participation and market competition intensity. The logarithmic term and square root term are used to model the diminishing marginal effect rule, the quadratic term is used to depict the negative impact of competition intensity, the interaction term is used to capture the synergistic effect of conversion rate and coverage, and the coefficients of each term are determined by fitting historical data. The constraint conditions include the upper limit of budget, the lower limit of coverage and the upper limit of competition intensity, etc. The lower game model takes the optimization of resource allocation as the optimization target, and the objective function adopts the combination form of quadratic term, linear term, exponential term and interaction term to model the nonlinear relationship of advertising cost, human resource cost, technology maintenance cost and risk control cost. The quadratic term is used to depict the scale effect, the exponential term is used to model the rapid growth characteristics of technology cost, the interaction term is used to establish the coupling relationship between the upper and lower models, and the constraint conditions include the upper limit of total cost and the lower limit of risk control, etc. The game solution adopts an iterative optimization algorithm. The upper model finds the optimal solution of marketing effect by gradient ascent method, and the lower model finds the cost minimization scheme by gradient descent method. The two models exchange information through the coupling term, and dynamically adjust the decision variables during the iteration process until the Nash equilibrium state is reached. The back propagation algorithm is used to update the neural network parameters, and an adaptive learning rate adjustment strategy is adopted to calculate the adjustment factor according to the training loss change rate, gradient norm, model convergence speed and historical performance indicators. When the adjustment factor is in the range of 0 to 0.3, the exponential decay strategy is adopted, in the range of 0.3 to 0.7, the linear decreasing strategy is adopted, and in the range of 0.7 to 1.0, the cosine annealing strategy is adopted, to ensure the stability and convergence of the model training.

[0052] The key technical ideas of the present application include the following aspects. The multi-modal feature extraction and cross-modal information fusion technology encodes the features of text, image and structured parameters through deep neural networks respectively, realizes semantic alignment and information enhancement between different modalities by using a four-pass bidirectional attention mechanism, and compared with traditional single-modal prediction methods, can fully utilize the multi-dimensional information of marketing activities, overcome the limitations of incomplete information of single data source, and improve the richness of feature representation and prediction accuracy. The dynamic and static combined double-branch prediction architecture captures the overall effect benchmark of marketing activities through the static prediction module, and models the time-varying pattern through the dynamic prediction module, and the two complement each other through an adaptive fusion mechanism, compared with traditional single prediction mode, can process both steady-state features and time-varying features of marketing activities, and enhance the adaptability of the model to complex marketing environment. The multi-scale time series feature extraction technology uses a group of parallel multi-scale convolution kernels to capture short-term fluctuations, medium-term cycles and long-term trends of time series patterns, combined with the long-distance dependence modeling capability of causal Transformer, compared with traditional single-scale time series analysis method, can more comprehensively describe the multi-level time regularity of user consumption behavior, and improve the accuracy and robustness of time series prediction. The game theory driven multi-objective optimization framework optimizes the dual objectives of maximizing marketing effect and optimizing resource allocation through upper and lower game models, realizes the balanced decision-making of effect and cost in marketing strategy formulation, compared with traditional single-objective optimization method, can better reflect the complexity and multi-objective nature of actual marketing decision-making. The synergistic effect of these technical ideas forms a complete marketing prediction and optimization system, multi-modal information fusion provides a rich feature basis for prediction, dynamic and static combined prediction architecture ensures the comprehensiveness of the prediction results, multi-scale time series modeling enhances the sensitivity to time variation, and the game optimization framework guarantees the practicality of the prediction results, compared with existing technologies, realizes the whole process technical innovation from data processing, feature extraction, model prediction to decision optimization.

[0053] The detailed structure of the marketing activity index prediction model includes five main components: data input layer, multi-modal feature extraction layer, cross-modal information fusion layer, double-branch prediction layer, and output fusion layer. The data input layer is responsible for receiving the text description, promotional images, and feature parameters of the marketing activities, and through the data preprocessing module, it realizes format standardization and quality control. The multi-modal feature extraction layer uses BERT encoder to process text data, ResNet-50 network to process image data, and structured encoder to process parameter data, mapping heterogeneous input to a unified high-dimensional feature space. The cross-modal information fusion layer establishes the interaction between different modal features through a four-way bidirectional attention network, optimizes feature pairing using the maximum weight bipartite graph matching algorithm, and generates a fused multi-modal representation. The double-branch prediction layer includes a static prediction branch and a dynamic prediction branch. The static branch uses a stacked residual multilayer perceptron to achieve overall index prediction, while the dynamic branch combines multi-scale time series feature extraction and causal Transformer to achieve sequential prediction. The output fusion layer adaptively combines the prediction results of the two branches through a dynamic weight matrix to generate the final marketing index prediction value.

[0054] The detailed steps of model training first perform data preparation and preprocessing, collect historical marketing activity data and divide it into training set, validation set, and test set, perform word segmentation and vocabulary construction on text data, size normalization and enhancement processing on image data, and feature engineering and missing value processing on structured parameters. Then initialize the model parameters, initialize the BERT encoder with pre-trained weights, initialize the ResNet-50 network with ImageNet pre-trained weights, and initialize other network layers with Xavier or He initialization strategy. Next, perform multi-stage training, freeze the pre-trained layer parameters in the first stage and only train the newly added network layers, unfreeze all parameters in the second stage and perform end-to-end fine-tuning, and introduce a game optimization framework in the third stage for multi-objective joint training. During training, adaptive learning rate adjustment and gate weight adjustment strategies are used, the hyperparameters are dynamically adjusted according to the performance of the validation set, the early stopping mechanism is used to prevent overfitting, and finally the model performance is evaluated on the test set.

[0055] The marketing activity index prediction model established by the application can process multi-source heterogeneous data and complex time sequence dependent relationship in a marketing environment. Traditional marketing prediction methods are mainly based on statistical regression models or simple time sequence analysis, and can only process single type of structured data, are difficult to fuse text, image and other unstructured information, and cannot effectively model the interaction relationship between multi-modal data. Compared with the traditional method, the advantages of the model of the application are reflected in three aspects of multi-modal information fusion capability, dynamic and static combined prediction mechanism and game theory optimization framework. The multi-modal information fusion maps heterogeneous data to a unified representation space through deep learning technology, overcomes the problem of insufficient information utilization of the traditional method, and improves the comprehensiveness and accuracy of prediction. The dynamic and static combined prediction mechanism considers the steady state characteristics and time-varying characteristics of the marketing activity at the same time, compared with the single static prediction or dynamic prediction method, can more comprehensively depict the complex marketing law, and enhances the adaptability and robustness of the model. The game theory optimization framework jointly optimizes the maximization of marketing effect and the optimization of resource allocation as a dual objective, compared with the traditional single objective optimization method, is more in line with the multi-objective characteristics of actual marketing decision, and improves the practical value and decision guidance significance of the prediction result.

[0056] It should be noted that the application also solves the following technical problems: first, the application solves the technical problem that the modeling of static characteristics and dynamic characteristics is separated in the marketing activity prediction, resulting in inaccurate time sequence dependent relationship capture. The traditional prediction method usually processes the static properties and dynamic time sequence characteristics of the marketing activity independently, and cannot effectively model the interaction relationship between static and dynamic characteristics. The application designs a double-branch collaborative mechanism, wherein the time sequence evolution branch captures the periodic and trend changes of user behavior, and the cross-modal interaction branch models the immediate influence of marketing materials, and the two branches realize adaptive coupling through a dynamic fusion output layer, so that the static marketing activity properties can be organically combined with the dynamic time sequence evolution mode, thereby accurately capturing the complex dependent relationship between static and dynamic characteristics in the marketing activity. Secondly, the application solves the technical problem that the fixed weight distribution in the multi-modal feature fusion process leads to poor adaptability in different prediction scenarios. The existing multi-modal fusion method usually adopts a fixed weight distribution strategy, and cannot dynamically adjust the importance of each modality according to the specific prediction scene and data characteristics. The application introduces an environment perception gating mechanism and a dynamic weight matrix, which can adaptively adjust the fusion weights of text, visual and activity characteristics according to the current prediction scene, wherein the elements of the dynamic weight matrix reflect the attention intensity of different prediction steps to the time sequence mode, realizing dynamic adjustment of the importance of the modality in the prediction scene, and significantly improving the adaptability and prediction accuracy of the model in different marketing environments.

[0057] Specifically, the principle of the present application is that the present application can solve the technical problem of insufficient fusion of multi-modal data features of marketing activities, mainly based on the following technical principles: first, the text description, promotional images and feature parameters of marketing activities have inherent relevance in the semantic space. A four-channel bidirectional attention mechanism can establish a bidirectional interaction channel between different modalities, so that the text features can actively align the visual content to ensure the semantic consistency of the activity description and the picture theme, and the activity features can drive the attention to the key areas of the picture, thereby realizing fine-grained cross-modal semantic alignment. Second, marketing activity data has multi-scale time regularity, and user behavior follows natural time cycles and cognitive psychological laws. A multi-scale sliding window can capture short-term promotional pulses, medium-term periodic regularities and long-term trend changes at the same time, and multi-scale feature extraction conforms to the hierarchical time structure of user consumption behavior. Third, the static prediction module realizes nonlinear feature transformation through a stacked residual multilayer perception network, and the residual connection ensures effective gradient propagation. The dynamic prediction module uses causal Transformer to ensure the causality constraint of time series modeling, and the dual-branch collaborative mechanism realizes decoupled modeling and adaptive fusion of static features and dynamic features. Finally, the maximum weight bipartite graph matching algorithm constructs different modal features into a graph structure, solves the optimal cross-modal feature pairing weight, realizes the global optimization and fusion of multi-modal information, and the dynamic fusion output layer further realizes the adaptive adjustment of the importance of the modal in the prediction scene through the dynamic weight matrix generated by the cross-modal attention, ensuring the effectiveness of the multi-modal fusion representation and the accuracy of the prediction result.

[0058] A specific embodiment 1 of the present application is provided below, and the specific implementation of each step in embodiment 1 is described in detail as follows.

[0059] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be described in detail here.

[0060] The specific implementation of step S02 is to construct a multi-modal feature extraction module, which encodes and transforms different types of data through a deep neural network. The visual feature vector encoding adopts a deep residual network to extract fine-grained spatial features, and is specifically represented as follows:

[0061] ;

[0062] In the formula, is a visual feature vector; is a trainable projection matrix with a dimension of 512x256; is a feature extraction function of the input image by the ResNet-50 network; is the input of the promotional image of the marketing activity, with a size of 224x224x3; is the bias term vector with dimension 256 1. The text feature vector encoding is based on a pre-trained language model dynamic semantic extractor, which is specifically represented as follows:

[0063]

[0064] where, is the text feature vector; is the text projection matrix with dimension 768 256; is the encoding function of the BERT model for the input text; is the input of the marketing campaign text description; is the text bias term with dimension 256 1. The activity feature vector encoding decomposes the key factors that affect user behavior into two dimensions of time features and activity strategy features, which is specifically represented as follows:

[0065]

[0066] where, is the activity feature vector; is the trainable parameter matrix with dimension 128 256; is the time feature, including week, month, holiday, etc. discrete encoding with dimension 64; is the activity strategy feature, including budget intensity, activity type, duration cycle, etc. parameters with dimension 64; denotes the feature concatenation operation; is the activity feature bias term with dimension 256 1.

[0067] The specific implementation of step S03 is to design a cross-modal information dynamic fusion mechanism to realize the interaction and fusion between different modal features through a four-path bidirectional attention mechanism. The calculation process of the text feature vector to the visual feature vector path first generates a query key-value pair through an independent projection matrix, which is specifically represented as follows:

[0068]

[0069] where, is the text-to-image query matrix; is the image key matrix; is the image value matrix; are the projection weight matrices of the query, key, and value respectively, all with dimension 256 64. Then the attention distribution of the text feature vector to the visual content is calculated through the scaled dot product attention, which is specifically represented as follows:

[0070] ;​​​​​​​

[0071] wherein, is the text-to-image attention output feature; is a model dimension parameter, taking a value of 64; is a normalization function. The environmental perception gating mechanism concatenates the output features of the four independent attention paths in the four-path bidirectional attention mechanism with the original activity feature vector, and is specifically expressed as follows:

[0072] ;

[0073] wherein, is the concatenated comprehensive feature matrix; is the activity feature-to-image attention output; is the activity feature-to-text attention output; is the text-to-activity feature attention output; is an environmental perception projection matrix, with a dimension of 256x64. An adaptive fusion weight of the four paths is generated through a learnable parameter matrix, and is specifically expressed as follows:

[0074] ;

[0075] wherein, is an adaptive fusion weight vector, with a dimension of 4x1; is a weight generation matrix, with a dimension of 320x4. The final multi-modal fusion representation is generated by weighted combination, and is specifically expressed as follows:

[0076] ;

[0077] wherein, is a multi-modal fusion representation; is a fusion weight of the kth path; is an output feature of the kth attention path, wherein , , , correspond to the output of the text-to-image, activity feature-to-image, activity feature-to-text, and text-to-activity feature four paths, respectively.

[0078] The specific implementation of step S04 is to establish a static prediction module, input the multi-modal fusion representation into a stacked residual multi-layer perception network to realize regression prediction of the marketing activity total index, and is specifically expressed as follows:

[0079] ;

[0080] wherein, is a static prediction result; is a multi-layer perception function; represents the Lth residual multilayer perception block, L takes 3; represents the function composition operation; is a multimodal fusion representation input.

[0081] The specific implementation of step S05 is to construct a dynamic prediction module, and a multi-scale sliding window is used to extract time sequence features. The multi-scale feature extraction uses a parallel multi-scale convolution kernel group to extract different granularity time sequence patterns, which is specifically represented as follows:

[0082] ;

[0083] In the formula, is the time sequence feature of the ith scale; represents the index of different scale convolution kernels, taking values 1, 2, and 3 corresponding to 3-day, 7-day, and 30-day windows respectively; is a rectified linear unit activation function; is the ith one-dimensional convolution operation; is an input time sequence data sequence, wherein is the current time step index. The dynamic feature aggregation introduces a learnable parameter matrix based on multi-granularity features to realize cross-scale dynamic calibration, which is specifically represented as follows:

[0084] ;

[0085] In the formula, is the aggregated sequence feature; is the number of scales, taking a value of 3; is the weight matrix of the ith scale, which has the same dimension as ; represents an element-wise multiplication operation. The time sequence evolution branch is based on the sequence pattern capture of the causal Transformer, and the time sequence representation generation is specifically represented as follows:

[0086] ;

[0087] In the formula, is the time sequence evolution feature; is a causal transformer function; is a learnable position encoding matrix, which has the same dimension as . The cross-modal interaction branch realizes feature enhancement through the interaction of the time sequence global context and the multimodal fusion representation, and the query vector generation is specifically represented as follows:

[0088] ;

[0089] In the formula, is the cross-modal interaction query vector; is a global pooling operation; To query the projection matrix, the dimension is 256x64. The key-value pair generation is specifically represented as follows:

[0090] , ;

[0091] In the formula, and are the key matrix and value matrix of cross-modal interaction respectively; and are the projection weight matrices of the key and value respectively, both with a dimension of 256x64. The cross-modal interaction feature is calculated through an attention mechanism, and the specific representation is as follows:

[0092] ;

[0093] In the formula, is the cross-modal interaction feature; is the interaction feature dimension, with a value of 64.

[0094] The specific implementation of step S06 is to design a dynamic fusion output layer to generate a dynamic weight matrix through cross-modal attention. First, the output of the time evolution branch is projected in the feature space, and the specific representation is as follows:

[0095] ;

[0096] In the formula, is the time branch projection feature; is a causal one-dimensional convolution function, with a convolution kernel size of 3. A dynamic weight matrix is generated through cross-modal attention, and the specific representation is as follows:

[0097] ;

[0098] In the formula, is the dynamic weight matrix; is the cross-modal interaction feature; is the feature dimension, with a value of 256. The dynamic fusion output layer fuses information and maps the output dimension to obtain a dynamic prediction result, and the specific representation is as follows:

[0099] ;

[0100] In the formula, is the dynamic prediction result; is the output layer multi-layer perception; reinforces the time sequence dominant mode; retains the cross-modal benchmark information.

[0101] The specific implementation of step S07 is to train and optimize the marketing activity index prediction model using historical marketing activity data, and balance the dual goals of maximizing marketing effect and optimizing resource allocation through a game model. The upper model takes maximizing marketing effect as the goal, and the objective function is specifically represented as follows:

[0102] ;

[0103] In the formula, is a marketing effect maximization function; is a user conversion rate, with a value range of 0.01-0.5; is an activity coverage, with a value range of 0.1-1.0; is a user engagement, with a value range of 0.05-0.8; is market competition intensity, with a value range of 0.1-1.0; , , , are undetermined coefficients, respectively taking values of 2.5, 1.8, 0.3, and 1.2. The constraint conditions include , , , wherein is a budget constraint parameter, taking a value of 1.2; is a lower limit of engagement, taking a value of 0.1; is an upper limit of competition intensity, taking a value of 0.9. The lower model takes optimizing resource allocation as the goal, and the objective function is specifically represented as follows:

[0104] ;

[0105] In the formula, is a resource allocation optimization function; is an advertising cost, with a value range of 1000-50000; is a human resource cost, with a value range of 500-20000; is a technology maintenance cost, with a value range of 200-10000; is a risk control cost, with a value range of 100-5000; , , , are cost weight coefficients, respectively taking values of 0.001, 1.0, 0.5, and 2.0; is a coupling term coefficient, taking a value of 10000. The constraint conditions include , , wherein is a total cost upper limit, taking a value of 60000; is a lower limit of risk control, taking a value of 200.

[0106] The principles and effects of each formula are explained as follows. Visual feature vector encoding formula Based on the feature extraction principle of deep convolutional neural network, the hierarchical feature learning mechanism of residual network is used to extract the spatial semantic information of the image, the linear projection operation is used to map the high-dimensional visual features to a unified feature space, and the bias term is used to adjust the mean shift of the feature distribution. Compared with the traditional hand-crafted feature extraction method, it can automatically learn more rich visual representation, improve the utilization efficiency of image information and the semantic understanding ability of marketing materials, and provide high-quality visual feature basis for subsequent cross-modal fusion. Text feature vector encoding formula Based on the context-aware encoding principle of pre-trained language model, the bidirectional attention mechanism of BERT model is used to capture the complex semantic relationship between words in the text sequence, and the projection matrix is used to realize the conversion from the pre-trained feature space to the task-adaptive feature space. Compared with the traditional bag-of-words model or TF-IDF method, it can better understand the semantic content and emotional tendency of marketing text, enhance the expression ability and semantic consistency of text information, and provide deep semantic support for text understanding of marketing activities. Activity feature vector encoding formula Based on the encoding principle of multi-dimensional feature fusion, the heterogeneous information in time dimension and strategy dimension is uniformly encoded through feature splicing operation, and the linear transformation matrix is used to learn the interaction relationship between different feature dimensions. Compared with the single-dimensional feature encoding method, it can more comprehensively describe the multi-faceted attributes of marketing activities, improve the information density and prediction relevance of structured parameters, and provide rich marketing environment context information for the model.

[0107] Query-key pair generation formula in four-way bidirectional attention mechanism 、 、 Based on the information retrieval principle of attention mechanism, different modal features are mapped to the query search space through independent projection transformation, realizing accurate alignment and relevance calculation of cross-modal information. Compared with the simple feature splicing method, it can establish more fine-grained semantic correspondence between modalities, and improve the accuracy and effectiveness of multi-modal information fusion. Scaling dot product attention calculation formula Based on the weighted aggregation principle of soft attention, the similarity score between query and key is calculated through inner product operation The scaling factor prevents gradient vanishing problem, normalization ensures the probabilistic nature of weight distribution, compared with hard attention mechanism, it can realize smoother information integration and more stable training process, and enhance the robustness and generalization ability of cross-modal attention. Multi-modal fusion representation formula Based on the fusion principle of adaptive weight aggregation, the learned weight coefficient By dynamically balancing the contributions of different attention pathways, the optimal combination of multimodal information is achieved. Compared with fixed-weight fusion methods, it can adaptively adjust the fusion strategy according to the characteristics of the input content, thereby improving the expressive power and task adaptability of multimodal representation.

[0108] Multi-scale feature extraction formula Based on the principle of multi-resolution time series analysis, time series patterns of short-term fluctuations, medium-term cycles, and long-term trends are extracted in parallel using convolutional kernels of different window sizes. Activation functions introduce nonlinear transformation capabilities, enabling a more comprehensive capture of multi-level temporal patterns in user consumption behavior compared to single-scale time series analysis methods. This improves the accuracy of time series prediction and the ability to model complex time patterns. Dynamic feature aggregation formula. Based on the principle of multi-scale information fusion, features at different scales are adaptively weighted and aggregated using a learnable weight matrix, with element-wise multiplication operations performed. This method achieves fine-grained feature tuning, which, compared to simple averaging or maximum aggregation methods, better balances the importance of different time scales, enhancing the expressive richness and predictive relevance of temporal features. (Cross-modal interactive feature calculation formula) Information retrieval based on the attention mechanism principle, through query vectors... AND key matrix The inner product operation is used to calculate the relevance score. Normalization ensures the probabilistic characteristics of the weight distribution, and compared with simple feature concatenation methods, it can achieve more accurate cross-modal information alignment and dynamic weight allocation, thereby improving the effectiveness and prediction accuracy of multimodal information fusion.

[0109] Dynamic weight matrix generation formula Based on the dynamic weight allocation principle of the self-attention mechanism, it operates through the inner product operation of temporal features and cross-modal features. Calculate the correlation matrix. Normalization ensures the probability distribution characteristics of the weights, scaling factor Maintaining numerical stability, compared to fixed-weight fusion methods, this approach dynamically adjusts the importance of different information sources based on the prediction scenario, improving the predictive model's adaptability and responsiveness to complex marketing environments. (Dynamic prediction result formula) Based on the prediction principle of weighted information fusion, a dynamic weight matrix is ​​used. Strengthening the time-series dominant mode At the same time, cross-modal reference information is preserved. As a supplement The final nonlinear mapping achieves a better integration of complementary information from temporal evolution and cross-modal interaction compared to prediction methods based on a single information source, thus improving the accuracy and stability of dynamic prediction.

[0110] Objective function of upper-level game model Based on the economic principle of maximizing multi-objective utility, The modeling conversion rate and coverage exhibit diminishing marginal returns. The concave function characteristic of participation is described by the term. The item represents the negative impact of competitive intensity. Interaction terms capture the synergistic effect of conversion rate and reach. Compared with single-objective linear optimization methods, this approach more accurately reflects the complex nonlinear relationship and multi-factor interaction of marketing effectiveness, thus improving the scientific rigor and practicality of marketing strategy optimization. The objective function of the lower-level game model... Based on the principle of resource allocation that minimizes costs, The modeling of the diseconomies of scale in advertising campaigns The item represents the proportional relationship of labor costs. The rapidly increasing maintenance costs of engraving technology This item reflects the cost-saving effect of risk control. The coupling term establishes the mutual influence relationship between the upper and lower level models. Compared with the traditional single-level optimization method, it can better balance the dual constraints of marketing effectiveness and resource input, and achieve overall optimization of marketing decisions and high efficiency of resource allocation.

[0111] To better understand and implement this invention, the following is an embodiment 2 of a specific application scenario: A technical team established an indicator prediction model for a mobile phone bill recharge marketing campaign. This campaign is a "Spring Festival Phone Bill Discount Recharge Campaign," with the text description stating "Recharge 100 yuan during the Spring Festival and receive 20 yuan in phone credit, limited time 7 days." The promotional image is a poster containing Spring Festival elements and recharge discount information. The campaign's characteristic parameters include time characteristics (Spring Festival period, 7 days) and campaign strategy characteristics (recharge 100 yuan, receive 20 yuan bonus, 80% discount rate). The technical team used the marketing campaign indicator prediction model establishment method to perform static and dynamic predictions of the number of users participating in the campaign and the recharge amount.

[0112] The technical team first collected multi-modal data of historical marketing activities as training samples. The text description data contains 3000 historical activity promotion texts such as "charge 50 yuan on weekends and get 10 yuan of call fee", "charge 200 yuan on National Day and get 50 yuan of call fee", etc. After BERT model encoding, a 768-dimensional text feature vector is obtained. The image data contains corresponding 2800 promotional posters, with a size of 1080x1920 pixels, and a 2048-dimensional visual feature vector is obtained by ResNet-50 model extraction. The activity feature parameters include time features (day of the week, month, holiday or not) and activity strategy features (recharge amount, gift ratio, activity type, duration), which are processed by a structured encoder to obtain a 256-dimensional activity feature vector.

[0113] In the multi-modal feature extraction stage, the technical team maps the feature vectors of the three modalities to a 512-dimensional shared latent space through a trainable projection matrix. For the current Spring Festival recharge activity, the text feature vector is obtained by encoding "Recharge 100 yuan during the Spring Festival and get 20 yuan of call fee, limited to 7 days" through the BERT model, the visual feature vector is obtained by encoding the Spring Festival theme poster through the ResNet-50 model, and the activity feature vector is obtained by structuring the time feature (Spring Festival, 7 days) and the activity strategy feature (100 yuan recharge, 20 yuan gift, 20% discount rate).

[0114] The cross-modal information dynamic fusion mechanism realizes feature interaction through a four-path bidirectional attention mechanism. The text feature vector to visual feature vector path calculates the attention weight matrix, where the "Spring Festival" keyword in the text and the Spring Festival decoration element in the poster form a high attention weight of 0.85, and "Recharge 100 yuan" and the amount display area in the poster form an attention weight of 0.78. In the activity feature vector to visual feature vector path, the time feature "Spring Festival" and the holiday atmosphere design of the poster produce a correlation weight of 0.82, and the activity strategy feature "20% discount" and the promotional sign in the poster obtain an attention weight of 0.76. The activity feature vector to text feature vector path verifies the compatibility of the activity parameters and the text description, and the matching degree of the recharge amount parameter and "100 yuan" in the text reaches 0.91. The text feature vector to activity feature vector path extracts the time parameter information corresponding to "limited to 7 days" in the text, and the matching weight is 0.88.

[0115] The environmental perception gating mechanism adaptively fuses the outputs of the four paths, as shown in Table 1:

[0116] Table 1 Four-path attention weight distribution table

[0117]

[0118] The multi-modal fusion representation generated by the weighted combination The dimension is 512, which contains the comprehensive information of text semantics, visual content and activity parameters.

[0119] The static prediction module uses the multi-modal fusion representation to predict the total index of the activity. The technical team builds a stacked network containing 4 residual multi-layer perceptron blocks, each block internally uses a 512→384→256 dimensional progressive feature transformation channel and a residual direct channel. The input multi-modal fusion features pass through the batch normalization layer in turn to constrain the feature distribution to zero mean and unit variance, and the self-attention layer dynamically selects key information and assigns an attention weight of 0.73 to the recharge amount related features, 0.68 to the time node features, and 0.52 to the visual attraction features. After processing by the stacked residual network, the static prediction module outputs the expected number of participants in the Spring Festival recharge activity as 28500 people, and the expected total recharge amount as Yuan.

[0120] The dynamic prediction module uses a multi-scale time series modeling method to process the historical activity data sequence. The technical team collects historical activity index data for the past 180 days, including daily participant numbers and recharge amounts. The multi-scale sliding window uses 3-day, 7-day and 14-day convolution kernels to extract time series features of different granularities. The 3-day short kernel captures the user behavior fluctuations within 72 hours of the weekend promotion, and the extracted features show that the recharge amount on Friday night increases by 65% compared to weekdays, and reaches a peak on Saturday and falls on Sunday. The 7-day medium kernel analyzes the periodic consumption pattern and identifies that the peak of recharge occurs on Friday and Saturday, accounting for 42% of the weekly recharge amount. The 14-day long kernel identifies the influence of the bi-weekly salary cycle on recharge behavior, and finds that the recharge activity in the first and middle of the month is 23% and 18% higher than other periods, respectively.

[0121] As shown in Table 2, the weight distribution of multi-scale feature extraction:

[0122] Table 2 Multi-scale convolution kernel weight distribution table

[0123]

[0124] The double-branch collaborative mechanism includes a time series evolution branch and a cross-modal interaction branch. The time series evolution branch is based on the causal Transformer architecture, which inputs the aggregated multi-scale features and learnable position encoding to generate time series representation . This branch identifies three stage features of user recharge behavior during the Spring Festival: recharge amount gradually increases in the week before the Spring Festival, recharge activity is relatively flat from the New Year's Eve to the third day of the New Year, and recharge amount quickly rises after the fourth day of the New Year as work resumes. The cross-modal interaction branch generates query vectors, key vectors and value vectors through the interaction of time series global context and multi-modal fusion representation, and calculates the interaction features , which reflects the influence effect of the active material on the instant user behavior.

[0125] The dynamic fusion output layer realizes the adaptive fusion of the time sequence features and the cross-modal information through three steps. The output of the time sequence evolution branch is projected into the feature space through causal convolution to obtain the projection feature . The cross-modal attention mechanism generates a dynamic weight matrix , and the weight elements of the matrix reflect the attention intensity of different prediction steps to the time sequence mode. In the 7-day prediction period of the Spring Festival recharge activity, the weights of the first 3 days are biased towards the short-term promotion effect, and the weight values are 0.82, 0.79, and 0.74, respectively. The last 4 days pay more attention to the long-term trend mode, and the weight values are 0.65, 0.58, 0.52, and 0.47, respectively.

[0126] The technical team optimizes the model parameters through historical data training, and uses a game model to balance the dual goals of maximizing marketing effect and optimizing resource allocation. The upper model takes maximizing marketing effect as the goal, and the input parameters include user conversion rate 0.156, activity coverage 0.734, user participation 0.682, market competition intensity 0.423, and the marketing effect comprehensive score calculated by the objective function is 4.23. The lower model takes optimizing resource allocation as the goal, and the input parameters include advertising cost , human resource cost , technical maintenance cost , risk control cost , and the total cost control index calculated is .

[0127] An adaptive learning rate adjustment function is used in the model training process, and based on the training loss change rate 0.023, the gradient norm 1.45, the model convergence speed 0.67, and the historical performance index 0.84, the adjustment factor value 0.54 is calculated. Since the adjustment factor value is within the range of 0.3 to 0.7, the system uses a linear decreasing adjustment strategy to gradually adjust the learning rate from the initial value 0.001 to 0.0003. The gating weight function calculates the balance value 0.65 according to the modal feature importance score 0.76, the historical prediction accuracy 0.82, the current training stage 0.65, and the model parameter complexity index 0.58, and uses an aggressive weight adjustment function to optimize the gating parameters and improve the multi-modal information fusion effect.

[0128] ​​​​After 200 rounds of iterative training, the performance of the model on the validation set is shown in Table 3:

[0129] Table 3 Model validation set performance index table

[0130]

[0131] The technical team applied the trained model to the actual prediction of the Spring Festival recharge activity. The static prediction results show that the number of expected participating users during the activity is 28500, the expected total amount of recharge is , the daily average number of recharge users is 4071, and the daily average amount of recharge is . The dynamic prediction results show the trend of the indicators during the 7-day activity: the first day is affected by the activity start-up effect, with a predicted number of participating users of 5200 and a recharge amount of ; the second and third days maintain high levels, with a predicted daily average number of participating users of 4800; the fourth and fifth days fall slightly, with a daily average number of participating users of 4200; the sixth and seventh days approach the end of the activity, with a last-minute surge, and the predicted number of participating users rises to 4600.

[0132] The model also predicts user behavior patterns at different times. The proportion of recharge users is predicted to be 26% during the 10:00-12:00 period, 22% during the 14:00-16:00 period, 31% during the 19:00-21:00 period, and 21% during other periods. Age distribution prediction shows that the 25-35 age group has the highest participation, accounting for 38% of the total participating users, the 36-45 age group accounts for 32%, the 18-25 age group accounts for 21%, and other age groups account for 9%.

[0133] Compared with the traditional marketing activity prediction method, the model building method brings significant technological progress. The traditional prediction method is usually based on single modal data, such as only using historical numerical data for time series analysis, or only relying on activity parameters for rule matching, which is difficult to capture the complex interaction between text semantics, visual content and activity characteristics. The invention realizes the deep semantic alignment of heterogeneous data through multi-modal feature extraction and cross-modal information dynamic fusion mechanism, which can simultaneously utilize the language expression of promotional copy, the visual appeal of promotional poster and the structured information of activity parameters, and form a more comprehensive marketing scene understanding ability. The time series modeling of the traditional method often uses fixed window or single scale analysis, which cannot capture the coupling effect of short-term promotion pulse and long-term consumption rule. The multi-scale sliding window and double-branch collaborative mechanism of the invention can accurately model the multi-level time series features from hour-level promotion response to monthly consumption cycle through differentiated time granularity extraction and dynamic and static feature decoupling. The traditional prediction model usually separates static planning and dynamic regulation, resulting in a large deviation between the expected of the activity planning stage and the actual performance of the execution stage. The static prediction and dynamic prediction collaborative architecture of the invention realizes the whole cycle integrated prediction from activity planning to execution monitoring through shared multi-modal fusion representation and consistent feature encoding system, which improves the continuity and reliability of the prediction results.

[0134] It should be noted that the variables involved in the present application are explained in detail as shown in Table 4.

[0135] Table 4 Variable explanation table

[0136]

[0137] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for establishing a marketing activity indicator prediction model, characterized in that, The application relates to a marketing activity index prediction method based on multi-modal dynamic fusion mechanism. S01, collecting multi-modal data of a marketing activity, wherein the multi-modal data comprises a marketing activity text description, a marketing activity promotional image and marketing activity characteristic parameters; S02, constructing a multi-modal feature extraction module, encoding the marketing activity text description through a BERT model to obtain a text feature vector, encoding the marketing activity promotional image through a ResNet-50 model to obtain a visual feature vector, and encoding the marketing activity characteristic parameters through a structured encoder to obtain an activity feature vector; S03, designing a cross-modal information dynamic fusion mechanism, constructing a four-path bidirectional attention mechanism to realize interactive fusion between the text feature vector and the visual feature vector, the activity feature vector and the visual feature vector, the activity feature vector and the text feature vector, and the text feature vector and the activity feature vector, constructing different modal feature vectors into bipartite graph nodes through a maximum weight bipartite graph matching algorithm and solving optimal cross-modal feature pairing weights to generate a multi-modal fusion representation; S04, establishing a static prediction module, inputting the multi-modal fusion representation into a stacked residual multi-layer perception network, realizing regression prediction of a marketing activity total index through a processing chain composed of a batch normalization layer, a self-attention layer and a residual connection, and obtaining a static prediction result; S05, constructing a dynamic prediction module, extracting time sequence features through a multi-scale sliding window, and realizing dynamic prediction of a marketing activity index sequence through a double-branch collaborative mechanism of a time sequence evolution branch and a cross-modal interaction branch; S06, designing a dynamic fusion output layer, generating a dynamic weight matrix through cross-modal attention, adaptively fusing the time sequence evolution branch output and the cross-modal interaction branch output, and obtaining a dynamic prediction result; S07, training and optimizing a marketing activity index prediction model through historical marketing activity data, balancing a double target of maximum marketing effect and optimal resource allocation through a game model, updating model parameters through a back propagation algorithm, and using the static prediction result and the dynamic prediction result to realize accurate prediction of a key index of a marketing activity. The multi-scale sliding window specifically extracts different granularity time sequence modes through a parallel multi-scale convolution kernel group, captures sudden promotion fluctuations through a 3-day short kernel, analyzes periodic consumption rules through a 7-day medium kernel, and identifies trend bill patterns through a 30-day long kernel.

2. The marketing activity metric prediction model establishment method of claim 1, wherein, The marketing activity characteristic parameters specifically include time characteristics and activity strategy characteristics, the time characteristics capture natural time sequence rules of weeks, months and holidays through discrete coding, and the activity strategy characteristics depict marketing strategy attributes of activity budget intensity, activity type and duration through structured representation. 3.The method of claim 2, wherein, The cross-modal information dynamic fusion mechanism specifically realizes fine-grained cross-modal semantic alignment and information enhancement through a differentiated query key-value pair configuration, four independent attention paths in the four-path bidirectional attention mechanism model asymmetric dependency relationships between different modalities, and the text feature vector actively aligns visual content as a query vector to ensure semantic consistency between activity description and picture theme.

4. The marketing activity metric prediction model establishment method of claim 3, wherein, The four-way bidirectional attention mechanism specifically realizes the interaction and fusion between the text feature vector and the visual feature vector, the activity feature vector and the visual feature vector, the activity feature vector and the text feature vector, and the text feature vector and the activity feature vector, generates a query key-value pair through an independent projection matrix, and then calculates the attention distribution of the feature vector to other modal content through a scaled dot-product attention. 5.The method of claim 4, wherein, The static prediction module specifically realizes the regression prediction of the marketing activity total index through a processing chain composed of a batch normalization layer, a self-attention layer and a residual connection, a stacked residual multi-layer perception network includes a feature transformation path and a residual direct connection path, the feature transformation path captures a nonlinear relationship through progressive mapping from 512 dimensions to 384 dimensions and then to 256 dimensions, and the residual direct connection path retains original information to prevent gradient vanishing.

6. The method of claim 5, wherein, The time series evolution branch specifically realizes sequence pattern capture based on a causal Transformer, adopts a causal mask mechanism to constrain the attention range, ensures that a prediction time step only accesses historical information, and generates time series representation by aggregating sequence features and learnable position encoding.

7. The method of claim 6, wherein, The cross-modal interaction branch specifically realizes feature enhancement through the interaction of a time series global context and a multi-modal fusion representation, generates a query vector through a pooling operation, generates a key-value pair through the multi-modal fusion representation, and outputs interaction features for subsequent fusion processing. 8.The method of claim 7, wherein, The dynamic weight matrix specifically reflects the attention intensity of different prediction steps to time series patterns, realizes the dynamic adjustment of the importance of modes in a prediction scenario, and generates dynamic weights through the interaction calculation of the projection features of the time series branch and the cross-modal interaction features.

Images