A digital marketing system based on big data

Abstract

The application relates to the technical field of digital marketing, and discloses a digital marketing system based on big data, which comprises a multi-modal data acquisition module for collecting original behavior data of users on e-commerce platforms, social media platforms and search platforms; a cross-modal feature alignment module for performing space-time alignment and feature mapping on text, images and behavior sequences in the original behavior data to generate cross-platform alignment features; and a double-channel generative interest modeling module for extracting a time sequence mode through an explicit behavior channel based on the cross-platform alignment features, combining a latent vector channel to generate implicit interest representation, and separating long-term interest from short-term dynamic interest by using orthogonal constraints. The double-channel generative interest modeling module is adopted to improve the detection rate of implicit interest, and compared with a static label matching method based on collaborative filtering in the prior art, the technical bottlenecks of incomplete capture of potential user needs and recommendation deviation caused by long-term and short-term interest coupling are solved.

Description

Technical Field

[0001] The present invention relates to the field of digital marketing technology, and in particular to a digital marketing system based on big data. Background Art

[0002] In recent years, with the rapid development of e-commerce platforms, social media, and search engines, user-generated behavioral data has become massive, multi-source, and multi-modal. Leveraging heterogeneous data to achieve refined and personalized marketing recommendations has become a challenging issue for digital marketing systems. This paper addresses the challenges of traditional marketing systems, such as insufficient multi-source data integration, incomplete implicit interest mining, and delayed profile updates, by proposing a digital marketing system based on big data. The following is an analysis and logical deduction of existing technologies to reveal their shortcomings.

[0003] In existing technologies, mainstream interest modeling methods often use collaborative filtering or static tag-based matching. While these methods are effective for analyzing users' explicit behaviors, they lack the ability to mine their underlying implicit needs. Furthermore, these methods fail to distinguish between users' long-term preferences and short-term interests, the coupling of which can lead to biased recommendations.

[0004] Traditional systems use a fixed-cycle batch update mechanism, typically once a day or once a week, which is unable to respond to sudden user behavior. As a result, when users develop new interests in a short period of time, recommendations or push notifications are still based on the old profile, seriously affecting conversion rates.

[0005] To this end, the present invention proposes a digital marketing system based on big data to solve the above-mentioned problems. Summary of the Invention

[0006] In view of the deficiencies of the existing technology, the present invention provides a digital marketing system based on big data to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A digital marketing system based on big data, comprising:

[0008] Multimodal data collection module, used to collect raw user behavior data on e-commerce platforms, social media platforms, and search platforms;

[0009] The cross-modal feature alignment module performs spatiotemporal alignment and feature mapping on the text, images, and behavior sequences in the original behavior data to generate cross-platform alignment features;

[0010] The dual-channel generative interest modeling module, based on cross-platform alignment features, extracts temporal patterns through the explicit behavior channel, combines it with the latent vector channel to generate implicit interest representations, and uses orthogonal constraints to separate long-term interests from short-term dynamic interests;

[0011] The multi-granularity adversarial discriminant module verifies the distribution consistency of long-term interest representation and short-term interest representation through a global discriminator and compares fine-grained similarity through a local discriminator.

[0012] Dynamic interest fusion and portrait expansion module, which integrates the verified implicit interest representation with the user's basic portrait through the attention mechanism;

[0013] The real-time incremental update engine module monitors the change rate of user portraits and triggers model parameter updates when the change rate exceeds the preset threshold.

[0014] Preferably, the cross-modal feature alignment module further includes:

[0015] The time and space stamp cleaning unit calibrates the timestamps in the original behavior data across platforms, converts the geographic location coordinates into coordinate systems, and outputs the cleaned behavior sequence.

[0016] The multimodal encoding unit receives the cleaned behavior sequence and extracts the semantic vector h of the text review through the BERT model. t , extract the visual features of the image sharing content h through ResNet-50 v , extract the temporal features h of the behavior sequence through TCN s ;

[0017] The contrast alignment unit performs cross-modal contrast learning on semantic vectors, visual features, and temporal features. The loss function is:

[0018]

[0019] Among them, L align To compare the alignment loss function 0, B is the batch size, τ is the temperature coefficient, is the text feature vector of the i-th sample, is the visual feature vector of the jth sample, is the cosine similarity;

[0020] Output cross-platform alignment feature h align :

[0021] h align =[h t ⊕h v ⊕h s ],

[0022] Among them, h align To align features across platforms, h t is the semantic vector, h v is the visual feature, h s It is a time series feature.

[0023] Preferably, the dual-channel generative interest modeling module further includes:

[0024] The explicit feature encoding unit extracts multi-scale temporal patterns from the temporal behavior data in the cross-platform alignment features through a dilated convolutional network and outputs the explicit feature vector h e ;

[0025] Latent vector injection unit, generating Gaussian distributed latent vectors in, d z dimensional identity matrix, d z is the potential space dimension, N is the Gaussian distribution;

[0026] The dual-channel fusion unit transforms the explicit feature vector h e It is fused with the latent vector z through linear transformation and activation function:

[0027] h f =GELU(W e h e +W z z+b),

[0028] Among them, h f is the fusion feature, W e 、W z is the learnable weight matrix, b is the bias term, and GELU is the activation function;

[0029] Orthogonal decoupling unit, for fusion feature h f Separate the long-term interests h by orthogonal constraints l With short-term dynamic interest h s :

[0030] h l =tanh(W l h f ),

[0031] h s =σ(W s h f ),

[0032]

[0033] Among them, h l For long-term stable interest representation, h s For short-term dynamic interest representation, is the weight matrix transpose, W l 、W s is the decoupling layer weight matrix, ∈ is the constraint threshold, tanh is the hyperbolic tangent function, ||·|| Fis the Frobenius norm of a matrix, and σ is a sigmoid activation function.

[0034] An adaptive fusion unit dynamically fuses long-term and short-term interest representations:

[0035] h i =λh l +(1-λ)h s ,

[0036] where h i is the hidden interest representation, and λ is an adaptive weight calculated by the fusion feature h f through a fully connected layer:

[0037] λ=σ(W λ h f ), where W λ is a weight matrix.

[0038] Preferably, in the multi-granularity adversarial discrimination module, the long-term interest representation h l and the short-term interest representation h s output by the dual-channel generative interest modeling module are further verified, and the multi-granularity adversarial discrimination module further comprises:

[0039] A global discrimination unit calculates a global distribution similarity score through a three-layer fully connected network on the concatenated long-term and short-term interest representation h c =[h l ⊕h s ]:

[0040] D global (h c )=LeakyReLU(W2·LeakyReLU(W1h c +b1)+b2),

[0041] where W1 and b1 are the first layer parameters, W2 and b2 are the second layer parameters, h l is the long-term stable interest representation, h s is the short-term dynamic interest representation, h c is the long-term and short-term interest representation concatenation vector, and D global is the global discriminator output score.

[0042] A local discrimination unit extracts fine-grained features through a four-head self-attention mechanism on h l and h s :

[0043]

[0044] where Q is the query matrix, K is the key matrix, V is the value matrix, d k is the attention head dimension, and KT is the transpose of the key matrix;

[0045] The adversarial loss calculation unit combines the global discriminant score and the local similarity score to calculate the total adversarial loss:

[0046]

[0047] Among them, L adv is the total adversarial loss, For real data interest representation, To generate data interest representation, α is the balance coefficient, L cont is the local contrast loss:

[0048]

[0049] in, is the similarity of the positive sample pair, is the similarity of negative sample pairs, τ is the temperature coefficient, and exp is the exponential function;

[0050] Gradient penalty unit, imposes Lipschitz constraints on the discriminator parameters to stabilize training:

[0051]

[0052] in, is the interpolation sample, λ is the penalty coefficient, L gp is the gradient penalty loss.

[0053] Preferably, the dynamic interest fusion and portrait expansion module is used to process the implicit interest representation verified by the multi-granularity adversarial discrimination module, further comprising:

[0054] Credibility verification unit, implicit interest representation h i and user base portrait P base Perform distribution consistency test:

[0055]

[0056] When KL valid <δ, the output verification pass signal, where KL valid To generate the consistency test index between the interest distribution and the true distribution, δ is the preset threshold, is the generated interest distribution of the i-th user, is the real interest distribution of the i-th user;

[0057] Attention weight calculation unit, implicit interest representation h of verified i and user base portrait P base Calculate attention weights:

[0058]

[0059] Q=W q P base ,

[0060] K=W k h i ,

[0061] Among them, W q 、W k is the projection matrix, Q is the query matrix, K is the key matrix, d k is the attention head dimension, K T is the transpose of the key matrix;

[0062] Feature fusion unit, which performs weighted fusion of implicit interest representations based on attention weights:

[0063] h f =α·(W v h i )+P base ,

[0064] Among them, W v is the value projection matrix, α is the balance coefficient;

[0065] The image expansion unit will fuse the features h f Generate extended portraits by combining dimensions with the basic user portraits:

[0066] P new =[P base ⊕h f ]∈R 256 ,

[0067] Among them, P new is the expanded user portrait, and R is the vector dimension identifier.

[0068] Preferably, the real-time incremental update engine module is used to process the user portrait P output by the dynamic interest fusion and portrait expansion module. new , further including:

[0069] The change rate monitoring unit calculates the change rate of user profiles in adjacent time windows:

[0070]

[0071] in, The user extended portrait vector generated for the t-th time window, is the user extended portrait vector generated in the t-1th time window, ζ is the rate of change of the cosine similarity metric, and an update is triggered when ζ>0.3;

[0072] an elastic window scheduling unit dynamically divides a long window W according to a change rate ζ l and a short window W s :

[0073]

[0074] wherein W active is a dynamically activated sliding time window, W s is a short-term dynamic window, and W l is a long-term stable window;

[0075] a second-order optimization unit calculates a parameter increment based on samples in the sliding window W active :

[0076]

[0077] wherein Δθ is an incremental adjustment amount of the model parameters, η is a learning rate, H is a Hessian matrix, and θ is a parameter gradient;

[0078] a parameter updating unit performs incremental updating of weights of the dual-channel generative interest modeling module and the multi-granularity adversarial discriminant module: θ t+1 = θ t - Δθ,

[0079] wherein θ t is a model parameter set before the tth updating, and θ t+1 is a model parameter set before the (t+1)th updating.

[0080] Preferably, the multi-modal data acquisition and alignment module adopts a cross-modal contrast learning framework to jointly encode text, image and behavior sequence features, and optimizes feature space alignment through a cosine similarity loss function.

[0081] The dual-channel generative interest modeling module includes an interest decoupling network driven by orthogonal constraints, a long-term interest extraction adopts an LSTM network, and a short-term interest extraction adopts a dilated convolution time series network.

[0082] Preferably, the multi-modal data includes social media text comments, product sharing pictures and geographic location trajectory data, wherein the text data is encoded by a BERT model, and the image data is extracted for global features by a residual network.

[0083] A terminal device includes a processor and a memory, the memory stores a computer program, and the processor implements all functional modules of the digital marketing system when executing the program.

[0084] A storage medium stores computer-executable instructions, which, when executed by a processor, implement multimodal data alignment, interest generation, and dynamic update operations of the digital marketing system.

[0085] The present invention provides a digital marketing system based on big data. It has the following beneficial effects:

[0086] 1. This invention adopts a dual-channel generative interest modeling module to improve the detection rate of implicit interests. Compared with the static label matching method based on collaborative filtering in the existing technology, it solves the technical bottlenecks of incomplete capture of users' potential needs and recommendation bias caused by the coupling of long-term and short-term interests.

[0087] 2. The present invention adopts a real-time incremental update engine module to shorten the response time of sudden interest. Compared with the fixed-period batch update mechanism in the existing technology, it solves the key problem of marketing strategy failure caused by delayed user portrait updates. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is a system framework diagram of the present invention. DETAILED DESCRIPTION

[0089] To help those skilled in the art understand the present invention, the following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0090] The present invention is described in detail below with reference to the accompanying drawings:

[0091] Example:

[0092] Please see the attached Figure 1 , an embodiment of the present invention provides a digital marketing system based on big data, including:

[0093] Multimodal data collection module, used to collect raw user behavior data on e-commerce platforms, social media platforms, and search platforms;

[0094] The cross-modal feature alignment module performs spatiotemporal alignment and feature mapping on the text, images, and behavior sequences in the original behavior data to generate cross-platform alignment features;

[0095] The time and space stamp cleaning unit calibrates the timestamps in the original behavior data across platforms, converts the geographic location coordinates into coordinate systems, and outputs the cleaned behavior sequence.

[0096] The multimodal encoding unit receives the cleaned behavior sequence and extracts the semantic vector h of the text review through the BERT model. t , extract the visual features of the image sharing content h through ResNet-50 v , extract the temporal features h of the behavior sequence through TCN s ;

[0097] The contrast alignment unit performs cross-modal contrast learning on semantic vectors, visual features, and temporal features. The loss function is:

[0098]

[0099] Among them, L align To compare the alignment loss function 0, B is the batch size, τ is the temperature coefficient, is the text feature vector of the i-th sample, is the visual feature vector of the jth sample, is the cosine similarity;

[0100] Output cross-platform alignment feature h align :

[0101] h align =[h t ⊕h v ⊕h s ],

[0102] Among them, h align To align features across platforms, h t is the semantic vector, h v is the visual feature, h s is the time series feature;

[0103] The dual-channel generative interest modeling module, based on cross-platform alignment features, extracts temporal patterns through the explicit behavior channel, combines it with the latent vector channel to generate implicit interest representations, and uses orthogonal constraints to separate long-term interests from short-term dynamic interests;

[0104] The explicit feature encoding unit extracts multi-scale temporal patterns from the temporal behavior data in the cross-platform alignment features through a dilated convolutional network and outputs the explicit feature vector h e ;

[0105] Latent vector injection unit, generating Gaussian distributed latent vectors in, d z dimensional identity matrix, d z is the potential space dimension, N is the Gaussian distribution;

[0106] The dual-channel fusion unit transforms the explicit feature vector h eIt is fused with the latent vector z through linear transformation and activation function:

[0107] h f =GELU(W e h e +W z z+b),

[0108] Among them, h f is the fusion feature, W e 、W z is the learnable weight matrix, b is the bias term, and GELU is the activation function;

[0109] Orthogonal decoupling unit, for fusion feature h f Separate the long-term interests h by orthogonal constraints l With short-term dynamic interest h s :

[0110] h l =tanh(W l h f ),

[0111] h s =σ(W s h f ),

[0112]

[0113] Among them, h l For long-term stable interest representation, h s For short-term dynamic interest representation, is the weight matrix transpose, W l 、W s is the decoupling layer weight matrix, ∈ is the constraint threshold, tanh is the hyperbolic tangent function, ||·|| F is the Frobenius norm of the matrix, σ is the Sigmoid activation function;

[0114] Adaptive fusion unit, dynamically fusing long-term and short-term interest representations:

[0115] h i =λh l +(1-λ)h s ,

[0116] Among them, h i is the implicit interest representation, λ is the adaptive weight, and is obtained by fusion feature h f Calculated by the fully connected layer:

[0117] λ=σ(W λ h f ), where W λ is the weight matrix;

[0118] The multi-granularity adversarial discriminant module verifies the distribution consistency of long-term interest representation and short-term interest representation through a global discriminator and compares fine-grained similarity through a local discriminator.

[0119] Global discriminant unit, the long-term and short-term interest representation h after splicing c =[h l ⊕h s ]Compute the global distribution similarity score through a three-layer fully connected network:

[0120] D global (h c )=LeakyReLU(W2·LeakyReLU(W1h c +b1)+b2),

[0121] Among them, W1 and b1 are the first layer parameters, W2 and b2 are the second layer parameters, and h l For long-term stable interest representation, h s is a short-term dynamic interest representation, h c is the concatenated vector of long-term and short-term interest representation, D global Output score for the global discriminator;

[0122] Local discriminant unit, for h l and h s Extract fine-grained features through a four-head self-attention mechanism:

[0123]

[0124] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k is the attention head dimension, K T is the transpose of the key matrix;

[0125] The adversarial loss calculation unit combines the global discriminant score and the local similarity score to calculate the total adversarial loss:

[0126]

[0127] Among them, L adv is the total adversarial loss, For real data interest representation, To generate data interest representation, α is the balance coefficient, L cont is the local contrast loss:

[0128]

[0129] in, is the similarity of the positive sample pair, is the similarity of negative sample pairs, τ is the temperature coefficient, and exp is the exponential function;

[0130] Gradient penalty unit, imposes Lipschitz constraints on the discriminator parameters to stabilize training:

[0131]

[0132] in, is the interpolation sample, λ is the penalty coefficient, L gp is the gradient penalty loss;

[0133] Dynamic interest fusion and portrait expansion module, which integrates the verified implicit interest representation with the user's basic portrait through the attention mechanism;

[0134] Credibility verification unit, implicit interest representation h i and user base portrait P base Perform distribution consistency test:

[0135]

[0136] When KL valid <δ, the output verification pass signal, where KL valid To generate the consistency test index between the interest distribution and the true distribution, δ is the preset threshold, is the generated interest distribution of the i-th user, is the real interest distribution of the i-th user;

[0137] Attention weight calculation unit, implicit interest representation h of verified i and user base portrait P base Calculate attention weights:

[0138]

[0139] Q=W q P base ,

[0140] K=W k h i ,

[0141] Among them, W q 、W k is the projection matrix, Q is the query matrix, K is the key matrix, d k is the attention head dimension, K T is the transpose of the key matrix;

[0142] Feature fusion unit, which performs weighted fusion of implicit interest representations based on attention weights:

[0143] hf =α·(W v h i )+P base ,

[0144] Among them, W v is the value projection matrix, α is the balance coefficient;

[0145] The image expansion unit will fuse the features h f Generate extended portraits by combining dimensions with the basic user portraits:

[0146] P new =[P base ⊕h f ]∈R 256 ,

[0147] Among them, P new is the expanded user portrait, R is the vector dimension identifier;

[0148] The real-time incremental update engine module monitors the rate of change of user profiles and triggers model parameter updates when the rate of change exceeds a preset threshold;

[0149] The change rate monitoring unit calculates the change rate of user profiles in adjacent time windows:

[0150]

[0151] in, The user extended portrait vector generated for the t-th time window, is the user extended portrait vector generated in the t-1th time window, ζ is the rate of change of the cosine similarity metric, and an update is triggered when ζ>0.3;

[0152] The elastic window scheduling unit dynamically divides the long window W according to the change rate ζ l With short window W s :

[0153]

[0154] Among them, W active is a dynamically activated sliding time window, W s is a short-term dynamic window, W l For the long-term stability window;

[0155] Second-order optimization unit, based on the sliding window W active Calculate parameter increments for samples within:

[0156]

[0157] Among them, Δθ is the incremental adjustment of model parameters, η is the learning rate, H is the Hessian matrix, is the parameter gradient;

[0158] Parameter update unit, which incrementally updates the weights of the dual-channel generative interest modeling module and the multi-granularity adversarial discriminant module: θ t+1 =θ t -Δθ,

[0159] Among them, θ t is the model parameter set before the tth update, θ t+1 is the set of model parameters before the t+1th update.

[0160] The multimodal data collection module uses e-commerce platforms, social media platforms, and search platforms to simultaneously collect users' original behavioral data on clicks, browsing, comments, sharing, and searches. It has a wide coverage, a large amount of data, and takes into account both structured and unstructured information, and can provide comprehensive and heterogeneous input for subsequent analysis; the multimodal data collection module automatically captures and aggregates multi-source data through a unified interface and collection strategy, greatly reducing the cost of manual docking and the compatibility problems of heterogeneous systems, and building a solid data foundation support for the system.

[0161] The cross-modal feature alignment module uses a time-space stamp cleaning unit to ensure the precise synchronization and correction of behavioral data in different time zones and coordinate systems; the multimodal encoding unit uses BERT, ResNet-50 and TCN to extract text, image and time series features respectively, achieving multi-perspective coverage of feature expression; the contrast alignment unit maps multimodal features to the same latent space by designing a contrast loss centered on temperature coefficient and cosine similarity, eliminating modal differences and making subsequent feature fusion stable and efficient, thus breaking the limitations of traditional single modality or simple splicing methods.

[0162] In the explicit feature encoding unit, the dual-channel generative interest modeling module uses a dilated convolutional network to extract the temporal patterns of user behavior sequences at multiple scales, intuitively reflecting the user's observable preferences; the latent vector injection unit generates latent vectors through Gaussian distribution, which can explore the user's potential needs; the dual-channel fusion unit and the orthogonal decoupling unit work together to fuse explicit and implicit signals, and effectively separate long-term stable interests from short-term dynamic interests through orthogonal constraints; the adaptive fusion unit dynamically calculates the weighting coefficient based on the fused features to achieve a balance between long-term and short-term interests, ensure that the interest representation is stable and timely, and greatly enhance the model's ability to capture complex user preferences.

[0163] The multi-granularity adversarial discrimination module uses the global discriminant unit to check the distribution consistency of the spliced ​​long-term and short-term interest representations to ensure that the overall representation conforms to the real distribution; with the help of the multi-head self-attention mechanism in the local discriminant unit, it deeply explores the fine-grained similarities in the interest representation and accurately distinguishes between generated and real samples; the adversarial loss calculation unit integrates global and local losses and introduces a balance coefficient to continuously improve the quality of the generator in a joint optimization manner.

[0164] The dynamic interest fusion and portrait expansion module strictly verifies the consistency of implicit interest representation and user basic portrait distribution through the credibility verification unit. The attention weight calculation unit carefully calculates the query-key attention value and captures the correlation between representations with high-dimensional projection; the feature fusion unit then weightedly fuses the implicit interests based on the attention weight to ensure that the expanded portrait is comprehensive and accurate; the portrait expansion unit splices the fused features with the original portrait to enrich the user attribute dimension and provide deep and broad user insights for downstream recommendations and marketing strategies, achieving a dual improvement in user portrait coverage and sophistication.

[0165] The change rate monitoring unit of the real-time incremental update engine module can sensitively capture sudden changes in user behavior by calculating the change rate of the cosine similarity of portraits in adjacent time windows; the elastic window scheduling unit switches between short-term windows and long-term windows based on the change rate, balancing the needs of real-time response and stable modeling; the second-order optimization unit uses Hessian matrix information to perform efficient parameter incremental calculations, making the model update fast and convergent; the parameter update unit incrementally updates the weights of the generative interest modeling and adversarial discrimination modules under the action of the trigger mechanism, ensuring that the system can achieve minute-level adaptive adjustments when user interests change, significantly improving the timeliness and accuracy of portrait updates.

[0166] A terminal device includes a processor and a memory, wherein the memory stores a computer program and the processor implements all functional modules of a digital marketing system when executing the program.

[0167] A storage medium stores computer-executable instructions, which, when executed by a processor, implement multimodal data alignment, interest generation, and dynamic update operations of a digital marketing system.

[0168] The terminal devices tightly integrate the processor and memory, and by pre-installing all the functional modules of the digital marketing system, user behavior data acquisition, cleaning, encoding, model inference, and profile updates can all be completed locally. The terminal devices utilize local computing resources for low-latency, high-concurrency data processing and model calculations, effectively reducing reliance on the cloud and improving the response speed of real-time recommendations and personalized marketing. At the same time, through a unified hardware platform and unified interface standards, deployment and operation and maintenance costs are significantly reduced, providing efficient, scalable, secure, and stable hardware support for digital marketing in edge or near-source scenarios.

[0169] The storage medium, based on non-volatile storage technology, is pre-written with executable instructions and configuration parameters for each module of the digital marketing system. Combined with a file system and version management mechanism, this ensures that the system can be quickly loaded and restored to its latest operating state during device startup or updates. The storage medium is resistant to power outages, temperature and humidity, and tampering, ensuring long-term data and program integrity. Furthermore, through removable or network distribution, it enables one-click upgrades and rollbacks of system software, effectively simplifying the rollout and maintenance of the marketing system on large-scale terminal devices and ensuring consistency and high availability in deployment environments across all regions.

[0170] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A digital marketing system based on big data, characterized by: include: Multimodal data collection module, used to collect raw user behavior data on e-commerce platforms, social media platforms, and search platforms; The cross-modal feature alignment module performs spatiotemporal alignment and feature mapping on the text, images, and behavior sequences in the original behavior data to generate cross-platform alignment features; The dual-channel generative interest modeling module, based on cross-platform alignment features, extracts temporal patterns through the explicit behavior channel, combines it with the latent vector channel to generate implicit interest representations, and uses orthogonal constraints to separate long-term interests from short-term dynamic interests; The multi-granularity adversarial discriminant module verifies the distribution consistency of long-term interest representation and short-term interest representation through a global discriminator and compares fine-grained similarity through a local discriminator. Dynamic interest fusion and portrait expansion module, which integrates the verified implicit interest representation with the user's basic portrait through the attention mechanism; The real-time incremental update engine module monitors the change rate of user portraits and triggers model parameter updates when the change rate exceeds the preset threshold.

2. A digital marketing system based on big data according to claim 1, characterized in that: The cross-modal feature alignment module further includes: The time-space stamp cleaning unit calibrates the timestamps in the original behavior data across platforms, converts the geographic location coordinates, and outputs the cleaned behavior sequence. The multimodal encoding unit receives the cleaned behavior sequence and extracts the semantic vector h of the text review through the BERT model. t , extract the visual features of the image sharing content h through ResNet-50 v , extract the temporal features h of the behavior sequence through TCN s ; The contrast alignment unit performs cross-modal contrast learning on semantic vectors, visual features, and temporal features. The loss function is: Among them, L align To compare the alignment loss function 0, B is the batch size, is the temperature coefficient, is the text feature vector of the i-th sample, is the visual feature vector of the jth sample, is the cosine similarity; Output cross-platform alignment feature h align : Among them, h align To align features across platforms, h t is the semantic vector, h v is the visual feature, h s It is a time series feature.

3. The digital marketing system based on big data according to claim 1, characterized in that: The dual-channel generative interest modeling module further includes: The explicit feature encoding unit extracts multi-scale temporal patterns from the temporal behavior data in the cross-platform alignment features through a dilated convolutional network and outputs the explicit feature vector h e ; Latent vector injection unit, generating Gaussian distributed latent vectors in, d z dimensional identity matrix, d z is the potential space dimension, N is the Gaussian distribution; The dual-channel fusion unit transforms the explicit feature vector h e It is fused with the latent vector z through linear transformation and activation function: h f =GELU(W e h e +W z z+b), Among them, h f is the fusion feature, W e 、W z is the learnable weight matrix, b is the bias term, and GELU is the activation function; Orthogonal decoupling unit, for fusion feature h f Separate long-term interests h by orthogonal constraints l With short-term dynamic interest h s : h l =tanh(W l h f ), h s =σ(W s h f ), Among them, h l For long-term stable interest representation, h s For short-term dynamic interest representation, is the weight matrix transpose, W l 、W s is the decoupling layer weight matrix, ∈ is the constraint threshold, tanh is the hyperbolic tangent function, ||·|| F is the Frobenius norm of the matrix, σ is the Sigmoid activation function; Adaptive fusion unit, dynamically fusing long-term and short-term interest representations: h i =λh l +(1-λ)h s , Among them, h i is the implicit interest representation, λ is the adaptive weight, and is obtained by fusion feature h f Calculated by the fully connected layer: λ=σ(W λ h f ), where W λ is the weight matrix.

4. The digital marketing system based on big data according to claim 1, characterized in that: In the multi-granularity adversarial discrimination module, the long-term interest representation h output by the dual-channel generative interest modeling module is used to verify l and short-term interest representation h s , further including: Global discriminant unit, representing the long-term and short-term interests after splicing The global distribution similarity score is calculated through a three-layer fully connected network: D global (h c )=LeakyReLU(W2·LeakyReLU(W1h c +b1)+b2), Among them, W1 and b1 are the first layer parameters, W2 and b2 are the second layer parameters, and h l For long-term stable interest representation, h s is a short-term dynamic interest representation, h c is the concatenated vector of long-term and short-term interest representation, D global Output score for the global discriminator; Local discriminant unit, for h l and h s Extract fine-grained features through a four-head self-attention mechanism: Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k is the attention head dimension, K T is the transpose of the key matrix; The adversarial loss calculation unit combines the global discriminant score and the local similarity score to calculate the total adversarial loss: Among them, L adv is the total adversarial loss, For real data interest representation, To generate data interest representation, α is the balance coefficient, L cont is the local contrast loss: in, is the similarity of the positive sample pair, is the similarity of negative sample pairs, is the temperature coefficient, exp is the exponential function; Gradient penalty unit, imposes Lipschitz constraints on the discriminator parameters to stabilize training: in, is the interpolation sample, λ is the penalty coefficient, L gp is the gradient penalty loss.

5. The digital marketing system based on big data according to claim 1, characterized in that: The dynamic interest fusion and portrait expansion module is used to process the implicit interest representation verified by the multi-granularity adversarial discrimination module, further including: Credibility verification unit, implicit interest representation h i and user base portrait P base Perform distribution consistency test: When KL valid <δ, the output verification pass signal, where KL valid To generate the consistency test index between the interest distribution and the true distribution, δ is the preset threshold, is the generated interest distribution of the i-th user, is the real interest distribution of the i-th user; Attention weight calculation unit, implicit interest representation h of verified i and user base portrait P base Calculate attention weights: Q=W q P base , K=W k h i , Among them, W q 、W k is the projection matrix, Q is the query matrix, K is the key matrix, d k is the attention head dimension, K T is the transpose of the key matrix; Feature fusion unit, which performs weighted fusion of implicit interest representations based on attention weights: h f =α·(W v h i )+P base , Among them, W v is the value projection matrix, α is the balance coefficient; The image expansion unit will fuse the features h f Generate extended portraits by combining dimensions with the basic user portraits: Among them, P new is the expanded user portrait, and R is the vector dimension identifier.

6. The digital marketing system based on big data according to claim 1, characterized in that: The real-time incremental update engine module is used to process the user portrait P output by the dynamic interest fusion and portrait expansion module. new , further including: The change rate monitoring unit calculates the change rate of user profiles in adjacent time windows: in, The user extended portrait vector generated for the t-th time window, is the user extended portrait vector generated in the t-1th time window, ζ is the rate of change of the cosine similarity metric, and an update is triggered when ζ>0.3; The elastic window scheduling unit dynamically divides the long window W according to the change rate ζ l With short window W s : Among them, W active is a dynamically activated sliding time window, W s is a short-term dynamic window, W l For the long-term stability window; Second-order optimization unit, based on the sliding window W active Calculate parameter increments for samples within: Among them, Δθ is the incremental adjustment of model parameters, η is the learning rate, H is the Hessian matrix, is the parameter gradient; Parameter update unit, which incrementally updates the weights of the dual-channel generative interest modeling module and the multi-granularity adversarial discriminant module: θ t+1 =θ t -Δθ, Among them, θ t is the model parameter set before the tth update, θ t+1 is the set of model parameters before the t+1th update.

7. The digital marketing system based on big data according to claim 1, characterized in that: The multimodal data acquisition and alignment module adopts a cross-modal contrastive learning framework to jointly encode text, image and behavior sequence features, and optimizes feature space alignment through the cosine similarity loss function; The dual-channel generative interest modeling module includes an orthogonal constraint-driven interest decoupling network, an LSTM network is used for long-term interest extraction, and a dilated convolutional temporal network is used for short-term interest extraction.

8. The digital marketing system based on big data according to claim 1, characterized in that: The multimodal data includes social media text comments, product sharing pictures and geographic location trajectory data, among which the text data is encoded by the BERT model and the image data is extracted with global features by the residual network.

9. A terminal device, characterized in that: It includes a processor and a memory, the memory stores a computer program, and when the processor executes the program, it implements all functional modules of the digital marketing system as described in any one of claims 1 to 8.

10. A storage medium, characterized in that: Computer-executable instructions are stored, and when the instructions are executed by a processor, multimodal data alignment, interest generation and dynamic update operations of the digital marketing system as described in any one of claims 1 to 8 are implemented.

Images