Dynamic marketing decision system based on multi-modal time series fusion

Abstract

The application relates to the technical field of data processing and artificial intelligence application, and discloses a dynamic marketing decision system based on multi-modal time sequence fusion, which comprises a feature extraction layer, a sequence processing layer, a measurement layer, a fusion layer and a decision layer; the system decouples low-frequency graphic-text data and high-frequency behavior sequences, and stores static graphic-text features in a feature cache table after feature extraction; the sequence processing layer receives graphic network topology parameters, adaptively adjusts a time window and extracts a hidden state sequence; the system is provided with three reverse feedback links: the decision layer feeds back a loss probability parameter to the fusion layer to inversely calculate a dynamic trigger threshold, and controls the scheduling of cross-modal computing resources; the fusion layer feeds back a feature dimension weight vector to the measurement layer, which is used as a dimension mask for distribution offset calculation; and the decision layer issues an elimination mark to the front end according to a generated intervention action, and overwrites invalid text features in the cache. The application reduces invalid matrix operation, realizes dynamic allocation of computing power, and improves the response sensitivity of the system to user preference transfer.

Description

Technical Field

[0001] This invention relates to the fields of data processing and artificial intelligence application technology, and in particular to a dynamic marketing decision-making system based on multimodal time-series fusion. Background Technology

[0002] In typical digital marketing scenarios, decision-making systems need to comprehensively process data from advertising images, promotional text, and user browsing and clicking behaviors to predict user conversion status and output corresponding strategies. Existing multimodal decision-making systems generally employ a synchronous forward computation process when processing cross-modal data. Because the update frequency of advertising creative image and text data is low, while user behavior sequence data undergoes high-frequency dynamic changes, binding static and dynamic data and continuously inputting them into complex neural networks leads to the system repeatedly performing heavy matrix operations on the same image and text data when facing high concurrency requests, resulting in inefficient use of computing resources.

[0003] In processing time-series data, existing systems mostly rely on preset fixed time windows to capture user behavior sequences. User interaction activity typically fluctuates significantly, and fixed sequence lengths are difficult to adapt to this actual variation. When user interaction is intensive, long windows smooth out local short-term interest shifts; when user interaction is infrequent, fixed windows introduce excessive historical data and increase irrelevant computational overhead.

[0004] Furthermore, the existing decision-making system's data processing structure is primarily based on unidirectional forward propagation, lacking a reverse control loop from the decision-making end to the front-end computation stage. The system's output of churn risk prediction results cannot be fed back to the intermediate fusion network to guide the setting of computational thresholds, resulting in the system triggering feature interaction calculations with fixed standards regardless of risk level. Simultaneously, when the system issues specific marketing interventions that change the user's business context, the underlying feature library cannot perceive this change and continues to use the original advertising features in subsequent calculations. This disconnect between front-end and back-end data states prevents the system from adjusting feature benchmarks in a timely manner after user preference shifts or external interventions occur, causing the accuracy of decision results and the system's responsiveness to decline over time. Summary of the Invention

[0005] The technical problem solved by this invention is that existing marketing decision-making systems perform unified synchronous calculations on low-frequency graphic data and high-frequency behavioral data, which leads to repetitive matrix operations and wastes computing power; the use of fixed sequence time windows makes it difficult to adapt to fluctuations in user interaction frequency and cannot accurately extract short-term offset features; in addition, existing systems lack a reverse closed-loop control mechanism, and the decision output cannot effectively guide the front-end computing power scheduling and feature updates, resulting in a lag in feature benchmark adjustment when user preferences shift, and insufficient decision sensitivity and system robustness.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] The first aspect of the present invention provides a dynamic marketing decision system based on multimodal temporal fusion, comprising a feature extraction layer, a sequence processing layer, a measurement layer, a fusion layer and a decision layer cascaded in sequence;

[0008] The feature extraction layer is used to receive and process advertising image data and advertising text data that are updated infrequently, extract image features and text features and persist them in the internal feature cache table in the form of key-value pairs;

[0009] The sequence processing layer is used to receive and process high-frequency updated user behavior sequence data streams, and output the encoded extracted temporal aggregation features and the set of hidden state sequences.

[0010] The metric layer receives the set of hidden state sequences, measures the distribution shift of the user behavior sequence in the probability space, and outputs the quantized shift measurement result to the fusion layer.

[0011] The fusion layer is used to perform cross-modal feature interaction of multimodal data. It is connected to the feature cache table and the sequence processing layer to obtain static image and text features and time-series aggregated features, and then fuse them to generate a multimodal fusion feature vector.

[0012] The decision layer is used to perform graph network updates based on multimodal fusion feature vectors, predict churn probability parameters, and generate marketing intervention action sequences;

[0013] The system is also equipped with three reverse parameter feedback links, including:

[0014] The first feedback loop is used to feed back the churn probability parameter output by the decision layer to the fusion layer as the trigger control parameter for determining whether to initiate cross-modal feature interaction.

[0015] The second feedback link is used to feed back the feature dimension weight vector output by the fusion layer to the metric layer as the dimension mask vector when calculating its distribution offset.

[0016] The third feedback loop is used to feed back the marketing intervention action sequence output by the decision-making level to the feature cache table, which is used to trigger cache eviction under specific conditions.

[0017] Furthermore, the feature extraction layer internally includes a convolutional network module, a language model module, and a feature cache table. The convolutional network module and the language model module are used to perform feature encoding mapping on the original advertising image data and advertising text data, respectively. The resulting image feature vector and text feature vector are stored in the feature cache table in a key-value pair structure using the combination of user identifier and advertising material identifier as the index key. When a query request fails to find the target data, the feature extraction layer retrieves the original data again, performs feature encoding mapping, and refills the feature cache table.

[0018] Furthermore, the sequence processing layer contains a long short-term memory network module, and the decision layer contains a graph neural network module. The long short-term memory network module receives local topological feature parameters from the graph neural network module across layers, and adaptively adjusts its own sequence input time window accordingly. Within this window, it performs feature encoding on the user behavior sequence data stream and outputs a set of hidden state sequences to the metric layer.

[0019] In a preferred embodiment of the present invention, the local topological feature parameters include the rate of change of local clustering coefficients and the temporal edge decay weights; the long short-term memory network module calculates the sequence input time window at the current moment by combining the local clustering coefficient change rate and the temporal edge decay weights with the baseline time window length: when the local clustering coefficient change rate increases, the time window is reduced; when the temporal edge decay weights increase, the time window is expanded; and the sequence input time window is restricted to a preset safe range by a boundary truncation function.

[0020] Furthermore, the metric layer contains an offset calculation module, and the fusion layer contains an encoder module. The offset calculation module extracts the feature dimension weight vector of the encoder module's previous output through the second feedback link, and uses this feature dimension weight vector to perform dimension-wise weighting on the current hidden state sequence set and the historical hidden state sequence set. Based on the weighted set, the current empirical probability distribution and the historical empirical probability distribution are constructed respectively, and the optimal transmission distribution distance between the two distributions is calculated and output as the offset metric result to the fusion layer.

[0021] In a preferred embodiment of the present invention, the fusion layer is further provided with a scheduling module, and the decision layer is provided with a prediction module; the scheduling module obtains the churn probability parameter output by the prediction module in the cycle of the last global collaborative fusion calculation through the first feedback link, and calculates the dynamic trigger threshold by combining the trigger sensitivity scaling factor and the basic tolerance bias term; the scheduling module compares the optimal transmission distribution distance with the dynamic trigger threshold, and when the optimal transmission distribution distance is greater than or equal to the dynamic trigger threshold, it outputs an enable signal to activate the encoder module to perform cross-modal feature interaction.

[0022] Furthermore, after receiving the enable signal, the encoder module reads the image feature vector and text feature vector from the feature cache table using the combination of the current user identifier and the associated advertising material identifier as the key. These are then concatenated with the last hidden state vector in the hidden state sequence set to form a multimodal initial tensor. The multimodal initial tensor undergoes attention operations and dimensionality reduction through the internal Transformer encoder to generate a multimodal fusion feature vector. Simultaneously, the feature dimension gating unit generates feature dimension weight vectors representing the proportion of importance of each element in the corresponding dimension.

[0023] Furthermore, the graph neural network module uses the multimodal fusion feature vector as the initial feature embedding for the target user node. In the bipartite graph, it calculates the aggregation weight parameters of the neighboring nodes of the local subgraph by combining the time interval decay term and performs weighted summation to output the updated feature representation of the target user node. Based on the graph evolution, it calculates the ratio of the actual number of edges to the maximum possible number of edges at the current time to generate the current local clustering coefficient, and then generates the local clustering coefficient change rate and temporal edge decay weight required for backhaul.

[0024] Furthermore, the decision-making layer also includes a prediction module and a decision engine module. The prediction module predicts the churn probability parameter for the current period based on the updated target user node feature representation through nonlinear network mapping. The decision engine module generates a marketing intervention action sequence based on the same feature representation. When the churn probability parameter is determined to be greater than or equal to the preset intervention threshold, the decision engine module sends a cache eviction flag containing user identifiers and material identifiers to the feature cache table through the third feedback link to trigger cache eviction.

[0025] In a preferred embodiment of the present invention, after receiving the cache eviction flag, the feature cache table clears the corresponding historical text feature vectors while retaining the image feature vectors; the feature extraction layer extracts the structured text data in the marketing intervention action sequence, reassembles it to generate an intervention description text sequence, and triggers the language model module to perform forward propagation to reconstruct text features. The generated intervention text feature vectors are written back to the feature cache table to overwrite the original text feature vectors, thereby blocking the interference of invalid semantic features on subsequent feature interactions.

[0026] A second aspect of this invention provides a dynamic marketing decision-making method based on multimodal temporal fusion, comprising the following steps:

[0027] The feature extraction layer performs feature encoding mapping on the low-frequency updated advertising image data and advertising text data, and persists the resulting feature vectors into the feature cache table;

[0028] The sequence processing layer receives local topological feature parameters returned by the decision layer, adaptively calculates the sequence input time window at the current time, encodes the frequently updated user behavior sequence data stream within the window, and outputs a set of hidden state sequences.

[0029] The metric layer extracts the feature dimension weight vector retained from the previous forward computation, and uses this vector to perform dimension weighting on the two sets of hidden state sequences, namely the current and historical sets, to construct an empirical probability distribution and obtain the optimal transmission distribution distance.

[0030] The fusion layer obtains the churn probability parameter from the previous output of the decision layer, calculates the dynamic trigger threshold at the current moment, and compares the optimal transmission distribution distance with this threshold. When the trigger condition is met, the fusion layer reads the image and text features from the feature cache table, concatenates them with the temporal aggregated features output by the long short-term memory network module, performs cross-modal feature interaction, and generates a multimodal fusion feature vector.

[0031] The decision-making layer takes the multimodal fusion feature vector as input to perform local node aggregation and graph network update, and outputs the target user node feature representation. Based on this, it predicts the churn probability parameter and generates a marketing intervention action sequence.

[0032] When the churn probability parameter exceeds the preset intervention threshold, a cache eviction flag is issued to the feature cache table, the text associated with the intervention action is extracted and concatenated, the text features are reconstructed and the original cache entry is overwritten.

[0033] In summary, the present invention has at least one of the following beneficial technical effects:

[0034] 1. The system of this invention separates low-frequency static advertising materials from high-frequency user interaction events in terms of physical architecture. Static graphic features are extracted once and then persistently cached, avoiding repeated multi-layer network forward propagation calculations under high-concurrency requests, and effectively reducing the underlying computing power overhead caused by matrix operations.

[0035] 2. This invention utilizes the churn probability relayed from the upstream decision-making layer through a first feedback loop to adjust the initiation threshold of cross-modal feature interactions. In high-risk regions, the threshold is lowered to increase sensitivity to behavioral shifts; in low-risk regions, the threshold is raised to enter a computationally silent state. This mechanism avoids the blind operation of global feature interaction tensor concatenation and achieves dynamic allocation of computational resources.

[0036] 3. The sequence processing layer of this invention introduces clustering coefficients and time decay weights generated by graph structure evolution. By controlling the expansion and contraction of the input window of the Long Short-Term Memory network through the above parameters, the shortcomings of conventional models that rely on fixed step sizes are overcome. This allows the system to focus on recent short-range behaviors when local network connections are active, and to broaden the temporal perspective when interactions are subdued.

[0037] 4. This invention, through a third feedback loop, proactively eliminates invalid original advertising text features from the corresponding user's cache after generating an intervention action, and reconstructs semantic features based on the intervention description text. This design prevents old features from interfering with new computation cycles at the data storage level, allowing the underlying feature benchmark to iterate synchronously with the external intervention status. Attached Figure Description

[0038] Figure 1 This is a structural block diagram of the dynamic marketing decision-making system based on multimodal temporal fusion according to the present invention;

[0039] Figure 2 The flowchart shows the dynamic marketing decision-making method based on multimodal temporal fusion of the present invention.

[0040] Figure 3 This is an internal flowchart of the modal feature extraction and feature caching of the present invention;

[0041] Figure 4 This is a flowchart of the adaptive truncation of temporal features under topology sensing according to the present invention;

[0042] Figure 5 This is a flowchart of the attention mask weighted distribution distance metric of the present invention;

[0043] Figure 6 This is a flowchart of the risk probability-driven collaborative computing scheduling process of the present invention;

[0044] Figure 7 This is a flowchart of the node state update and global feature fusion process of the present invention;

[0045] Figure 8 This is a flowchart of the intervention action-driven cache eviction and feature reorganization process of the present invention.

[0046] The system consists of the following modules: 10. Feature Extraction Layer; 11. Convolutional Network Module; 12. Language Model Module; 13. Feature Cache Table; 20. Sequence Processing Layer; 21. Long Short-Term Memory Network Module; 30. Metric Layer; 31. Offset Calculation Module; 40. Fusion Layer; 41. Scheduling Module; 42. Encoder Module; 50. Decision Layer; 51. Graph Neural Network Module; 52. Prediction Module; and 53. Decision Engine Module. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0048] See attached document Figure 1 The present invention provides a dynamic marketing decision system based on multimodal temporal fusion, which includes a feature extraction layer 10, a sequence processing layer 20, a measurement layer 30, a fusion layer 40 and a decision layer 50 cascaded in sequence.

[0049] The feature extraction layer 10 is used to receive and process advertising image data and advertising text data that are updated infrequently. It contains a convolutional network module 11, a language model module 12, and a feature cache table 13. The convolutional network module 11 and the language model module 12 are used for feature encoding mapping of the image and text modalities, respectively. The outputs of both are connected to the input of the feature cache table 13, and the extracted feature vectors are stored in the feature cache table 13 in the form of key-value pairs.

[0050] The sequence processing layer 20 is used to receive and process the frequently updated user behavior sequence data stream. It has a long short-term memory network module 21 inside. The output of this module is connected to the metric layer 30 to send the encoded hidden state sequence set to the metric layer 30.

[0051] The metric layer 30 includes an offset calculation module 31, which measures the distribution offset of the user behavior sequence in the probability space. The output of the offset calculation module 31 is connected to the fusion layer 40.

[0052] The fusion layer 40 is used to schedule computing resources and perform cross-modal feature interaction of multimodal data. It contains a scheduling module 41 and an encoder module 42. The input of the encoder module 42 is connected to the feature cache table 13 and the long short-term memory network module 21, respectively, to simultaneously acquire static image and text features and temporal aggregated features.

[0053] The decision layer 50 is used to update the graph topology state and output the marketing intervention action sequence. It contains a graph neural network module 51, a prediction module 52 and a decision engine module 53. The input of the graph neural network module 51 is connected to the output of the encoder module 42.

[0054] In addition to the aforementioned forward path, the system architecture of this invention is further configured with three reverse parameter feedback links to feed back the outputs of the decision layer and fusion layer to the front-end stage, forming a cross-layer closed-loop control:

[0055] The first feedback link is located between the prediction module 52 and the scheduling module 41. The churn probability parameter output by the prediction module 52 is fed back to the scheduling module 41 as its trigger control parameter.

[0056] The second feedback link is located between the encoder module 42 and the offset calculation module 31. The feature dimension weight vector output by the encoder module 42 is fed back to the offset calculation module 31 as its dimension mask vector.

[0057] The third feedback link is located between the decision engine module 53 and the feature cache table 13. The marketing intervention action sequence output by the decision engine module 53 is fed back to the feature cache table 13 to trigger cache eviction.

[0058] See attached document Figure 2 This invention provides a dynamic marketing decision-making method based on multimodal temporal fusion, comprising the following steps:

[0059] S100, the feature extraction layer 10 performs feature encoding mapping and caching on the frequently updated advertising image data and advertising text data. When an advertising material is loaded for the first time, when a user requests the material for the first time, or when the material's attributes are updated, the convolutional network module 11 and the language model module 12 encode the image and text data respectively. The resulting feature vector is stored in the feature cache table 13 with the combination of the user identifier and the advertising material identifier as the key and the key-value pair structure.

[0060] S200, the sequence processing layer 20 performs topology-aware adaptive truncation of temporal features. The long short-term memory network module 21 receives the local topological feature parameters returned by the graph neural network module 51, adjusts its own sequence input time window accordingly, and encodes the frequently updated user behavior sequence data stream within the window, outputting a set of hidden state sequences;

[0061] S300, the metric layer 30 performs the distribution distance measurement under dimension mask weighting. The offset calculation module 31 extracts the feature dimension weight vector of the previous output of the encoder module 42, and uses it to weight the two sets of hidden state sequences of the current and historical data one dimension at a time. Then, it calculates the optimal transmission distribution distance based on the two empirical probability distributions constructed after weighting.

[0062] S400, the fusion layer 40 performs risk probability-driven collaborative computation scheduling. The scheduling module 41 takes the churn probability parameter output by the prediction module 52 in the previous cycle of global collaborative fusion computation, calculates the dynamic trigger threshold accordingly, and compares the optimal transmission distribution distance with the threshold to determine whether to trigger global collaborative fusion computation in the current cycle.

[0063] In step S500, the encoder module 42 and the graph neural network module 51 perform node state updates and global collaborative fusion calculations. When the triggering condition is met, the encoder module 42 reads graph and text features from the feature cache table 13, concatenates them with the temporal aggregated features output by the long short-term memory network module 21, and performs cross-modal feature interaction to generate a multimodal fusion feature vector. The graph neural network module 51 uses this vector as input to complete local node aggregation and graph network updates, and outputs the target user node feature representation.

[0064] S600, the decision layer 50 performs intervention-driven cache eviction and feature reorganization. The prediction module 52 predicts the churn probability parameter for the current period based on the target user node feature representation, and the decision engine module 53 generates a marketing intervention action sequence accordingly. When the churn probability parameter exceeds the intervention critical threshold, the decision engine module 53 issues a cache eviction flag to the feature cache table 13, extracts the associated text corresponding to the intervention action to complete the text sequence concatenation, and starts the forward propagation of the language model module 12 in the next calculation period to reconstruct the text features.

[0065] The key aspects and working principles of the technical solution of this invention will be explained one by one below, combining the specific execution logic and calculation formulas of each step.

[0066] See attached document Figure 3 Step S100 specifically includes the following steps:

[0067] S110, the feature extraction layer 10 receives the original material data and determines the triggering conditions for feature extraction.

[0068] The system continuously monitors the release and modification status of advertising materials and initiates extraction when one of the following three situations occurs: first, the system loads a new advertising material for the first time; second, a user requests to load an advertising material for the first time; third, existing advertising materials undergo attribute updates, such as product display images or promotional copy being modified.

[0069] Once any of the above conditions are met, the feature extraction layer 10 will send the corresponding original advertising image data and original advertising text data into the convolutional network module 11 and the language model module 12, respectively.

[0070] S120, the convolutional network module 11 performs feature encoding mapping on the original advertising image data.

[0071] Given that the update frequency of advertising images is much lower than that of interactive behaviors such as clicks, pre-extracting and independently calculating image features can significantly save computing power during high-concurrency business phases.

[0072] Let the input raw advertising image data be... The image is first preprocessed, including scaling and normalization, to transform it into a matrix with a fixed width, height, and number of channels. The convolutional network module 11 can employ a convolutional neural network architecture containing multiple residual blocks, each containing a convolutional layer, a batch normalization layer, and an activation function. After visual features are extracted from the image through multiple convolutional and pooling layers, a dimensionality-reduced, standardized image feature vector is output by a fully connected layer. The mathematical expression for this process is:

[0073] ;

[0074] in, For the forward propagation function of convolutional network module 11, Its network weight parameters, To map to Standardized image feature vectors in dimensional feature space, dimension The computing power can be configured to 256, 512, or 1024.

[0075] Its weight parameters Pre-training offline is possible: Based on a general image classification pre-trained model, historical ad images with different marketing conversion effects are collected as samples, labeled with the historical click-through rate of the image, and supervised updates are performed using loss functions such as cross-entropy. This enables the module to capture potential marketing visual attributes.

[0076] S130, the language model module 12 performs feature encoding mapping on the original advertising text data. The processing mechanism is similar to that of the image module, which independently extracts the semantic expression of fixed text descriptions in the short term.

[0077] Let the input original advertising text data be... The language model module 12 internally includes a word segmentation component and a multi-layer self-attention mechanism network. First, the word segmentation component divides the text into a sequence of words, and then concatenates a classification label [CLS] representing the semantics of the entire sentence at the beginning of the sequence. Next, the word sequence is fed into the multi-layer self-attention mechanism network, where multi-head attention calculation and a feedforward neural network layer are used to obtain the contextual representation of each word. The output corresponding to the classification label [CLS] is taken as the representation of the entire sentence, i.e., the standardized text feature vector. The mathematical expression for this process is:

[0078] ;

[0079] in, For the forward propagation function of language model module 12, Its network weight parameters, For the same mapping to Standardized text feature vectors in a 3D feature space.

[0080] Its weight parameters Based on an open-source general language model, and using advertising copy within a specific business scenario as corpus, the reconstruction loss can be calculated through a masked language model task for adaptive fine-tuning.

[0081] S140, the feature extraction layer 10 creates a storage entry for the material in the feature cache table 13.

[0082] The system uses the combination of the current user identifier and the advertising creative identifier (such as a product ID or advertising campaign number generated from the database) as the index key, and the image feature vector as the index key. With text feature vectors The combination of these values ​​is used to create data values, which are then persistently stored in memory or high-speed storage devices using key-value pairs.

[0083] S150, the system performs feature retrieval when responding to routine decision requests, and has fallback logic in case of a miss.

[0084] When the sequence processing layer 20 receives a user’s high-frequency interaction behavior with the advertising material and triggers subsequent global collaborative fusion calculation, the system extracts the user identifier and advertising material identifier carried by the interaction event, and uses the combination of the two as the key to query the feature cache table 13.

[0085] If the query matches, the cached image feature vector is read directly. With text feature vectors The signal is sent to the fusion layer 40, thus skipping the forward propagation of the convolutional network module 11 and the language model module 12.

[0086] If the query fails, for example, due to cache expiration and cleanup or system restart causing missing memory data, feature extraction layer 10 triggers an abnormal rollback mechanism, re-fetches the original ad image data and original ad text data of the material, returns to execution steps S110 to S140, and rewrites the data into feature cache table 13 after the calculation is completed.

[0087] Through the aforementioned read mechanism of "hit-and-fill" and "miss-and-fill", the system avoids repetitive heavy matrix operations under high-frequency requests, and achieves computational decoupling between modal data with different update frequencies.

[0088] See attached document Figure 4 Step S200 specifically includes the following steps:

[0089] S210, the sequence processing layer 20 receives the user behavior sequence data stream and obtains local topology feature parameters for feedback control across layers.

[0090] A fixed sequence length is difficult to adapt to fluctuations in user interaction frequency. Therefore, the Long Short-Term Memory (LSTM) network module 21 receives local topological feature parameters for the current user node output by the graph neural network module 51 during the last actual graph network update cycle. These parameters include the rate of change of local clustering coefficients. With time-series edge decay weights Two items.

[0091] Wherein, the local clustering coefficient is the ratio of the actual number of edges to the maximum possible number of edges in the subgraph where the current user node is located; its rate of change This reflects the evolution of the tightness of connections between a user's adjacent nodes over a period of time, representing the fluctuation in the user's recent exploration range within the associated material network. Temporal edge decay weights. The value is calculated using an exponential decay function combined with a time interval, representing the decay of user activity over time in historical interactions. The specific calculation formulas for both are generated by the graph neural network module 51 during graph network updates, as detailed in step S540 below.

[0092] It should be noted that during the first calculation cycle of the system, or when the target user first enters the observation and the graph neural network module 51 has not yet returned data, the sequence processing layer 20 takes a preset default initial value (e.g., Set to 0 (Set to 1), so that the initial truncation length is calculated based on the default parameter and the base time window length in the first cycle, and then dynamically adjusted after receiving the returned parameters in subsequent cycles.

[0093] S220, the sequence processing layer 20 calculates the sequence input time window of the long short-term memory network module 21 based on local topological feature parameters.

[0094] Regarding the current moment Sequence input time window The mapping calculation formula is as follows:

[0095] ;

[0096] in, The length of the action sequence extracted at the current moment, taken as a positive integer; The base time window length is typically 20 to 50 action sequence steps. , Adjust the attenuation weights according to the preset topology constraint coefficients. With cluster change rate The weights of the influence on window size are all in the range of 0 to 1 and can be determined by offline backtracking and grid optimization based on historical behavior observation data. This is a floor function to ensure that the sequence step size is a discrete integer; This is a boundary truncation function that limits the output to a safe interval. Inside, among which Take a smaller value to ensure the extraction of basic temporal features (e.g., 5). Choose a larger value to prevent memory overflow caused by an excessively large tensor dimension (e.g., 100).

[0097] The adjustment logic of this formula is: when the user interaction range shifts or fluctuates significantly ( (Increase), denominator increases, This shrinks the network, focusing it on recent, high-frequency behaviors within a small scope, capturing short-term shifts in interest; when interactions become less frequent ( (Increase), molecular increase, This expands the scope, extracting behavioral features over a longer span and reducing the computational overhead of high-frequency truncation. Boundary truncation ensures that the output window length remains valid even with extreme topology parameter inputs.

[0098] S230, Long Short-Term Memory Network Module 21 in an adaptive truncation time window Internal execution behavior sequence hidden state encoding.

[0099] System Extraction Window The system uses raw user behavior data, including interaction types, page dwell time, and corresponding ad creative identifiers. During preprocessing, discrete categorical features such as interaction types and ad creative identifiers are first mapped into dense vectors via an embedding layer. Then, continuous features such as page dwell time are normalized. Finally, these vectors are concatenated along the feature dimensions to obtain a vector with the following dimensions. The input sequence vector.

[0100] The Long Short-Term Memory (LSTM) network module 21 receives the input sequence vector and internally includes a forget gate, an input gate, a cell state update unit, and an output gate. The input vector of the current time step is concatenated with the hidden state vector of the previous time step, and then processed by their respective weight matrices and sigmoid activation to generate forget gate and input gate state vectors. The former determines the retention ratio of the cell state from the previous time step, and the latter determines the writing ratio of the current candidate state. After cell state updates are completed through dot multiplication and addition, the output gate and Tanh activation jointly regulate the output of the hidden state vector of the current time step.

[0101] After complete sequence encoding, the Long Short-Term Memory network module 21 is in the window Internal output hidden state sequence set , where each hidden state vector Carrying time steps And the accumulation of time-series information prior to that.

[0102] Regarding the training of the Long Short-Term Memory Network module 21, its initial weights can be obtained from the desensitized historical behavior sequence through unsupervised pre-training of the autoregressive prediction task. After the system is deployed, the weights of this module are no longer fixed, but instead form an end-to-end link with the subsequent graph neural network module and prediction module. The actual business conversion results (such as whether there is a loss) are used as labels, and the binary cross-entropy loss output by the decision layer is jointly fine-tuned by the time backpropagation algorithm.

[0103] After the Long Short-Term Memory network module 21 completes the encoding, it sets the hidden state sequence. It is sent to metric layer 30 to support subsequent distribution offset measurements.

[0104] See attached document Figure 5 Step S300 specifically includes the following steps:

[0105] S310, the offset calculation module 31 obtains the feature dimension weight vector of the previous output of the encoder module 42 across layers, and performs dimension mask weighting accordingly.

[0106] The contribution of each feature dimension of a high-frequency behavior sequence to business semantics is not equal. To suppress interference from non-core dimensions, the system introduces feature dimension weights obtained through cross-modal learning as a mask. The offset calculation module 31 extracts the feature dimension weight vector, representing the importance of each feature dimension, output by the encoder module 42 in the previous global collaborative fusion calculation. .

[0107] The vector The feature vector of the multimodal fusion is obtained by performing a fully connected mapping on the feature dimension gating unit inside the encoder module 42, followed by Softmax normalization. The length is related to the feature dimension. Consistent, each element represents the weight percentage of the corresponding dimension in the multimodal global interaction.

[0108] When the system is running for the first time and there are no historical weights available for extraction... Take a uniform vector with equal weights in each dimension as the default initial value.

[0109] S320, the offset calculation module 31 performs dimensional weighting on the current and historical sets of hidden state sequences using the Hadamard product.

[0110] The system retrieves the set of hidden state sequences output by the current Long Short-Term Memory network module 21. And the set of historical hidden state sequences cached during the last global collaborative fusion computation. Using feature dimension weight vector Multiply each eigenvector in both sets element-wise. The mathematical expression for this weighted calculation is:

[0111] ;

[0112] ;

[0113] in, Let be the hidden state vector in the sequence set. This is the Hadamard product. The weighted set of latent state sequences obtained through this weighting suppresses low-relevance dimensions while amplifying dimensions that are strongly correlated with the current marketing context. This is equivalent to weighting the high-dimensional behavioral features with prior knowledge of the importance of dimensions learned in the past.

[0114] It should be noted that this refers to the set of historical hidden state sequences during the system's first run, before any cache is available. At that time, the offset calculation module 31 will calculate the current set This is also used as a historical benchmark cache, so that the first cycle has a benchmark surface that can be used for subsequent measurements.

[0115] S330, the offset calculation module 31 constructs an empirical probability distribution based on the weighted set of hidden states.

[0116] The system assigns equal probability mass to each vector in the weighted hidden state sequence set. Let the weighted current and historical sets contain... and Given eigenvectors, whose cardinality may not be equal, then for each vector in its set... Empowerment Where N is the total number of vectors in the set, i.e., the current set contains... Historical collection Constructing an experience distribution ,in For The Dirac function centered at .

[0117] Thus, the current empirical probability distribution is obtained respectively. Probability distribution based on historical experience This allows the overall form of the two sets to be aligned and compared at the level of probability measure.

[0118] S340, Offset calculation module 31 calculates the optimal transmission distribution distance between two empirical probability distributions. .

[0119] The system uses the Wasserstein distance metric for distribution migration, which, compared to KL divergence and JS divergence, still provides a smooth distance gradient when the support sets of the two distributions do not overlap. Its mathematical expression is as follows:

[0120] ;

[0121] in, The marginal distributions are respectively equal to , The set of all joint distributions, i.e., all possible transmission plans; This refers to the specific transmission plan; x and y are respectively taken from... , eigenvectors; The Euclidean norm measures the cost of moving two points in the characteristic space. The goal is to find the minimum transmission cost, i.e., the optimal transmission distribution distance.

[0122] Directly solving the above problem with exact precision would be computationally too complex. Therefore, the offset calculation module 31 uses Sinkhorn approximation: an entropy regularization term is introduced into the objective function. and regularization coefficient In one embodiment Using values ​​between 0.01 and 0.1, an offline test can be conducted to balance accuracy and convergence speed. The problem is then transformed into a convex optimization problem and solved iteratively using the Sinkhorn-Knopp matrix alternating scaling algorithm. To prevent convergence during iteration, a maximum number of iterations (e.g., 50) and a convergence tolerance (e.g., 10) are set. -4 The output will terminate when the iteration increment is lower than the tolerance or the maximum number of iterations is reached.

[0123] To ensure it is comparable to the subsequent dynamic trigger threshold on the same scale, the transmission cost is normalized to the [0,1] interval using a scaling factor; this scaling factor can be taken from offline samples. The maximum observation value is used as a benchmark, and the actual cost is divided by this benchmark and truncated to [0,1].

[0124] Specifically, when the system is running for the first time and the historical set is temporarily used by the current set without a real benchmark for comparison, the offset calculation module 31 skips the above solution and directly sets... =1, to unconditionally trigger a global collaborative fusion calculation.

[0125] At this point, scalar This quantifies the deviation of the user's current behavior focus relative to the historical baseline, which is then used for comparison in subsequent scheduling processes.

[0126] See attached document Figure 6 Step S400 specifically includes the following steps:

[0127] S410, the scheduling module 41 acquires the churn probability parameters output by the prediction module 52 across cycles, and uses them as feedback control input.

[0128] The scheduling module 41 extracts the churn probability parameter output by the prediction module 52 for the current user node within the cycle of the last actual global collaborative fusion calculation. This parameter is a one-dimensional floating-point scalar with a value between 0 and 1, representing the expected risk of a user leaving the current marketing conversion link, as assessed by the system at the previous observation node based on the user's multimodal fusion characteristics and topological state.

[0129] In the first calculation cycle of the system, before the prediction module 52 has produced a result, the scheduling module 41 takes a preset prior default value (e.g., 0.5) as... Participate in the calculation.

[0130] S420, the scheduling module 41 adaptively calculates the dynamic trigger threshold based on the churn probability parameter.

[0131] The scheduling module 41 combines a preset sensitivity scaling factor with a basic bias term, based on the churn probability parameter. Calculate the dynamic trigger threshold at the current moment. The mapping formula is as follows:

[0132] ;

[0133] in, The dynamic trigger threshold for determining whether to initiate cross-modal feature interaction in this cycle is a one-dimensional floating-point scalar. The churn probability parameter returned by prediction module 52 for the previous execution of the global collaborative fusion calculation cycle; To trigger the sensitivity scaling factor, the response magnitude of risk changes to the threshold is adjusted; Based on the basic tolerance bias term, a minimum offset detection benchmark that must be maintained even under extremely high risks is set; , All are positive real numbers, which can be identified by offline grid search from historical logs covering different concurrency levels and conversion rates.

[0134] In some embodiments, Take a value of 0.1 to 0.5. The values ​​are taken to be between 0.05 and 0.2, and the sum of α and β is not greater than the normalized optimal transmission distribution distance. The upper bound of the value (i.e., 1) is used to ensure that the threshold still falls within the valid comparison range under boundary conditions.

[0135] The regulatory logic of this formula is: when the risk of loss is high ( When approaching 1), Approaching 0 Then lower and approach Even if the behavior sequence only shifts slightly at this point, It is also easier to overcome this low threshold, thereby improving the responsiveness of integration and intervention during crisis phases; when the risk of loss is low ( When approaching 0, Rise and approach This allows behavior to fluctuate within a wider range, thereby reducing unnecessary cross-modal multi-head attention forward computation and enabling differentiated allocation of computing power.

[0136] S430, the scheduling module 41 compares the optimal transmission distribution distance with the dynamic trigger threshold.

[0137] The scheduling module 41 receives the optimal transmission distribution distance from the offset calculation module 31. The optimal transmission distribution distance As a one-dimensional non-negative floating-point scalar, it is used to quantify the offset of the behavior sequence relative to the historical reference surface within the current observation window. The scheduling module 41 uses its internal logic operation unit to combine it with... Numerical comparison outputs a Boolean gating signal used to control the direction of subsequent data flow.

[0138] S440, the fusion layer 40 executes branch control according to the gating signal, covering the entire value range and leaving no decision dead zone, specifically divided into two mutually exclusive paths.

[0139] Firstly, when When the behavior deviation exceeds the tolerance of the risk state, the scheduling module 41 outputs a gating signal to enable the value, activates the encoder module 42, and causes it to read the image and text features from the feature cache table 13 and start the global collaborative fusion calculation.

[0140] Secondly, when When the behavior deviation is within the safe range, the scheduling module 41 outputs a gate signal to block the value acquisition. The system maintains the current feature state for silent monitoring and instructs the sequence processing layer 20 to continue to slide the time window according to the time step and collect the next batch of behavior sequence data.

[0141] Through the above full-range comparison, this embodiment can dynamically constrain the forward triggering conditions of the encoder module 42, avoid unnecessary multimodal tensor splicing and high-dimensional matrix operations, and improve operating efficiency.

[0142] See attached document Figure 7 Step S500 specifically includes the following steps:

[0143] S510, encoder module 42 receives the enable value of the gating signal and performs multimodal feature reading and tensor splicing.

[0144] After scheduling module 41 determines that the behavior sequence has shifted significantly and triggers global collaborative fusion calculation, the system uses the combination of the current user identifier and the advertising material identifier involved in the behavior sequence as the key to read the corresponding image feature vector from feature cache table 13. With text feature vectors .

[0145] Simultaneously, encoder module 42 receives the set of hidden state sequences output by long short-term memory network module 21 across layers. The hidden state vector of its last time step is taken as the temporal aggregation feature. This vector incorporates the cumulative evolution of behavioral preferences within the window, representing the user's current short-term focus of interest.

[0146] At the same time, encoder module 42 will receive the complete set of hidden state sequences. The offset calculation module 31 is cached back as a set of historical hidden state sequences required for the next distribution distance metric. .

[0147] Subsequently, the matrix transformation unit inside the encoder module 42 linearly concatenates the three vectors of the same dimension in the feature dimension direction to obtain a 3d multimodal initial tensor. It jointly represents the visual content, semantic information, and behavioral evolution that users pay attention to within the current cycle.

[0148] S520, encoder module 42 performs cross-modal feature interaction on the multimodal initial tensor.

[0149] Considering the non-linear coupling between visual, textual, and behavioral features in the conversion decision, encoder module 42 adopts a multi-layer Transformer encoder architecture to achieve fusion: multi-modal initial tensor After input, the signal passes sequentially through a multi-head self-attention layer and a feedforward neural network layer; in the multi-head self-attention layer, Multiplying by three learnable projection matrices respectively to generate the query matrix Q, key matrix K, and value matrix V, and then performing a scaled dot product attention operation, the mathematical expression of which is:

[0150] ;

[0151] in, The scaling factor for the head dimension is used to prevent the gradient vanishing caused by excessively large dot products. In practice, it can be 64 or 128 depending on the memory capacity. Softmax normalizes the inner product result into attention distribution weights.

[0152] After extracting cross-modal correlations through the aforementioned forward computation, encoder module 42 then undergoes adaptive dimensionality reduction via an internal linear mapping layer, resulting in an output dimension of [dimensionality value missing]. Multimodal fusion feature vector .

[0153] In output At the same time, the feature dimension gating unit of encoder module 42 controls... Perform a fully connected layer and then normalize it using Softmax to generate a feature dimension weight vector with the same length as the feature dimension d. Each element represents the importance ratio of the corresponding dimension in the multimodal global interaction; this vector is used as a buffer for the offset calculation module 31 to send back the inverse control parameters, and is used as the dimension mask vector for the next calculation cycle.

[0154] As for the basic components inside the Transformer encoder, such as layer normalization and residual connections, they can be implemented by directly calling the built-in interfaces of standard deep learning frameworks, and will not be elaborated here.

[0155] S530, the graph neural network module 51 takes the multimodal fused feature vector as input and performs time-aware local node aggregation.

[0156] The system maintains a dynamically evolving user-material bipartite graph in memory, with nodes divided into user nodes and material nodes, and edges representing the historical interactions between them. The graph neural network module 51 adopts a time-aware graph attention network architecture, using the multimodal fusion feature vector output by the encoder module 42. As a target user node The initial feature embedding is obtained, and its first and second-order neighbor nodes in the graph are taken. .

[0157] To represent the varying impacts of interactions across different time spans, the module introduces a time interval decay term in edge aggregation. For the set... any neighboring node Its nodes Aggregate weight parameters for:

[0158] ;

[0159] in, For nodes Multimodal fusion feature vector , The feature vectors of neighboring nodes; The learnable linear feature transformation matrix projects the node features onto a high-dimensional hidden graph representation space. The graph attention weight vector; This involves concatenating vectors. For nodes with neighboring nodes The time interval between interactions is quantified, and is generally made continuous in natural time or natural days. For example, it is mapped to a preset real number interval after taking the logarithm, so as to avoid the time dimension being too large and overwhelming other features. LeakyReLU is an activation function with a negative half-axis leakage rate. The leakage rate slope can be 0.2 to ensure that the gradient propagates in the negative interval.

[0160] Therefore, strongly correlated and recently active graph edges can obtain higher information transmission weight.

[0161] S540, the graph neural network module 51 performs a global mapping of graph topology updates and node states, and generates the topology parameters required for backhaul.

[0162] The module aggregates weight parameters The features of the neighbors in the local subgraph are weighted and summed, and then passed through a multilayer perceptron nonlinear mapping to output the updated feature representation of the target user node. This representation includes both the multimodal preferences of individual users and the group collaboration trends under the global topology at the graph level.

[0163] Meanwhile, the background process scans the graph for newly added and broken edges, calculates the ratio of the actual number of edges to the maximum possible number of edges in the subgraph where the target user node is located at the current moment, and obtains the current local clustering coefficient. This leads to the generation of the updated local clustering coefficient change rate. With time-series edge decay weights ; and cache it for calculation during the next graph network update. Required local clustering coefficients from the previous period .

[0164] Local clustering coefficient change rate Take the local clustering coefficient of the current period Compared with the local clustering coefficients obtained during the last graph network update The absolute value of the difference, i.e. Quantify the evolution of the tightness of the connection between adjacent nodes of a user between two consecutive updates;

[0165] Temporal edge decay weight Calculated using the exponential decay function, i.e. ,in The preset attenuation rate coefficient can be calibrated offline by combining the business activity distribution. The time interval between the current moment and the user's most recent interaction, quantified in calendar hours or calendar days; As the interval increases, it approaches 0; as it decreases, it approaches 1.

[0166] Subsequently, the graph neural network module 51 represents the features of the target user node. It is sent to the prediction module 52 and the decision engine module 53 to support risk quantification and intervention instruction generation; and will , The long short-term memory network module 21 in the reverse transmission sequence processing layer 20 is used to complete the closed-loop update of the local topology parameters for sequence truncation control.

[0167] For training this module, offline pre-training using contrastive learning can be employed before deployment: using de-identified user-material bipartite graphs within historical time slices as samples, locally connected positive sample node pairs are generated by random walks with restarts, and disconnected negative sample node pairs are generated by cross-domain random sampling. The cosine similarity between node features is calculated, and the linear transformation matrix is ​​updated using InfoNCE contrastive loss. With graph attention weight vector This enables the module to initially identify high-potential transformation nodes in a multidimensional topology.

[0168] See attached document Figure 8 Step S600 specifically includes the following steps:

[0169] S610, the decision-making level 50 predicts churn probability parameters and generates a sequence of marketing intervention actions.

[0170] Prediction module 52 receives the target user node feature representation transmitted by graph neural network module 51. This is then input into an internal multi-layer fully connected network. The input layer dimension of this network is... The system employs a consistent approach, with one to three hidden layers of decreasing dimensions in the middle. After feature compression, the final Sigmoid function is activated to output the churn probability parameter for the current period. ; It is a scalar in the range of 0 to 1, representing the quantitative risk of a user leaving the conversion link under the current observation window.

[0171] Decision engine module 53 synchronously receives the feature representation of the target user node. The input is then fed into an internal policy multilayer perceptron. This perceptron contains an input layer, several hidden layers, and an output layer. Features are projected onto the intervention action representation space through nonlinear mapping and ReLU activation in the hidden layers. The number of neurons in the output layer is consistent with the total number of candidate intervention actions in the library. A probability distribution for each candidate action is generated using Softmax, and the action combination corresponding to the highest probability is selected as the marketing intervention action sequence. For example, issuing specific discount coupons, pushing strong recall pop-up messages, or increasing the exposure priority of specific products are all execution instructions.

[0172] The two networks mentioned above can be trained offline: samples are constructed using actual click conversions and churn labels from historical marketing data. The binary churn task of prediction module 52 takes the real churn labels and optimizes them with binary cross-entropy loss. The multi-class action mapping task of decision engine module 53 takes historical successful intervention records as hard labels and optimizes them with multi-class cross-entropy loss. The two are jointly updated by gradient descent.

[0173] S620, the decision engine module 53 compares the intervention critical threshold and issues a cache elimination flag.

[0174] The system has a built-in intervention threshold. This is used to determine whether to actively update the underlying cache. In one embodiment... The value can be set between 0.6 and 0.85, and can be adjusted periodically based on peak business budget consumption and average customer retention cost. The decision engine module 53 compares the churn probability parameters using its internal logic processing unit. and .

[0175] when At this point, the predicted risk is high, and conventional feature fine-tuning is no longer sufficient to adapt to the change in user status. The system determines that the effectiveness of the original image and text features associated with the user has decreased. The decision engine module 53 then generates a cache eviction flag containing the current user identifier and the associated advertising material identifier, and sends it down to the feature cache table 13 along the reverse link. Based on this, the feature cache table 13 releases the corresponding memory address, clears the historical text feature vector associated with the user, and retains the image feature vector. Since the advertising image has not changed, there is no need to repeat the forward calculation of the convolutional network module 11, thereby blocking the interference of invalid semantic features on subsequent collaborative calculations at the data structure level.

[0176] when When the risk is within a controllable range, the system does not issue an elimination flag, and feature cache table 13 maintains its original storage state. Subsequent cycles will continue to use the existing multimodal features for collaborative inference. This branch, together with the previous branch, covers the entire probability domain, avoiding logical dead zones.

[0177] S630, after S620 triggers cache eviction, feature extraction layer 10 extracts intervention-related text and triggers language model module 12 to forward propagate to reconstruct text features.

[0178] Feature extraction layer 10 analyzes marketing intervention action sequences The structured text data is extracted, such as character information corresponding to the intervention rules, such as "discount amount", "category recall keywords" or "dedicated customer service prompts". Then, the character processing unit serializes and reassembles the data according to preset linguistic template splicing rules to generate an intervention description text sequence representing the new business context. .

[0179] Language model module 12 receives the intervention description text sequence This module employs a pre-trained language model with a multi-layer bidirectional Transformer architecture: The input identifier sequence is converted by a word segmenter, and after being superimposed with a word embedding tensor and a position encoding tensor, it is fed into a stacked multi-head self-attention layer. After multi-layer bidirectional context attention calculation and feedforward neural network mapping, the hidden layer output corresponding to the classification label [CLS] is taken, and then after a Tanh-activated fully connected layer mapping, the recombined intervention text feature vector is obtained. Its dimensions are consistent with the characteristics of the eliminated historical texts.

[0180] Language Model Module 12 will Write back the entry in feature cache table 13 corresponding to the combination of user identifier and advertising material identifier, overwriting only the text feature vector while keeping the image feature vector unchanged.

[0181] After this overwrite, the text features in the cached entries have been replaced by intervention phrase features from the original advertising copy features. These features are then read as text feature vectors and participate in cross-modal fusion in step S510 of the next calculation cycle. This vector constitutes the update reference surface when the system performs global collaborative fusion calculation and dimension mask weighting in the next cycle. This closes the data loop from decision output to low-level feature reconstruction within the architecture, which helps improve the system's robustness under feature shifts.

[0182] To facilitate understanding of the overall workflow of this invention, the following example of a user browsing a beauty advertisement, along with a set of exemplary numerical values, illustrates the end-to-end data flow within a single calculation cycle. It should be noted that the following numerical values ​​are for illustrative purposes only and do not constitute a limitation of this invention.

[0183] Let the feature space dimension be... Taking 512 as an example, when a certain beauty product advertisement material is loaded for the first time, the convolutional network module 11 and the language model module 12 encode the main product image and promotional text into image feature vectors, respectively. With text feature vectors The user identifier and the ad creative identifier are then stored in the feature cache table 13 using the combination of the user identifier and the ad creative identifier as the key. Subsequently, the user's high-frequency clicks, swipes, and other behaviors related to the ad creative continuously flow into the sequence processing layer 20.

[0184] In this cycle, the local topological feature parameters returned by the graph neural network module 51 during the last graph network update are: =0.6、 =0.4 indicates that the user's recent interaction range has fluctuated significantly and their activity level is relatively high. (Base window value) =30, Topological constraint coefficient =0.5 and =0.8, boundary =[5,100], then the sequence processing layer 20 calculates the sequence input time window according to the formula:

[0185] ;

[0186] This means encoding the most recent 24 behavioral steps within the current period. As a contrast, if the user transitions to a bland interaction (e.g., ...), ... =0.1、 =0.9), calculated using the same formula:

[0187] ;

[0188] The window size is adjusted accordingly to reflect the adaptive expansion and contraction of the window based on the topology. The Long Short-Term Memory (LSTM) network module 21 encodes within this 24-step window, outputting a set of hidden state sequences. .

[0189] Metric layer 30 extracts the feature dimension weight vector from the encoder module 42 in the previous cycle. ,right With the set of historical hidden state sequences After constructing two empirical probability distributions through dimension-wise weighting, the optimal transmission distribution distance for this period is obtained by Sinkhorn approximation and normalization. =0.7 indicates that the user's current behavior focus has significantly deviated from the historical baseline.

[0190] The scheduling module 41 of the fusion layer 40 retrieves the churn probability parameter output by the prediction module 52 during the previous global collaborative fusion calculation cycle. =0.8 (high churn risk), according to the formula =0.3、 =0.1 Reverse calculation of dynamic trigger threshold:

[0191] =0.3*0.2+0.1=0.16.

[0192] because =0.7≥ =0.16, scheduling module 41 outputs a gating signal to enable the value, triggering the global collaborative fusion calculation for this cycle. It can be seen that in a high-risk state, the threshold is adaptively lowered, making the system more sensitive to offsets; conversely, if... =0.2, then:

[0193] =0.3*0.8+0.1=0.34;

[0194] If at this time The value is only 0.3. Since 0.3 < 0.34, it will switch to silent monitoring to avoid unnecessary heavy calculations under low-risk normal conditions.

[0195] Upon triggering, encoder module 42 reads from feature cache table 13 and and take The final hidden state as a temporal aggregation feature The three are concatenated to form a multimodal initial tensor of dimension 1536. ; via multi-layer Transformer encoder (head dimension scaling factor) Take 64) cross-modal feature interactions and dimensionality reduction, and output a multimodal fusion feature vector. Simultaneously, the feature dimension weight vector is generated by the feature dimension gating unit. The offset calculation module 31 is sent back to be used as the dimension mask vector in the next cycle.

[0196] Graph Neural Network Module 51 For target user nodes The initial embedding is then weighted and aggregated with its first and second-order neighbors in a time-aware graph attention network to output the updated target user node feature representation. Meanwhile, the background process calculates the local clustering coefficients for this period. =0.55, compared to the value obtained in the last update. Compared to 0.45, the topology parameters generated for the next cycle are:

[0197] =|0.55-0.45|=0.10;

[0198] and according to Take the attenuation rate =0.05, since the most recent interaction =2 hours, so e=exp(-0.1)≈0.905; the two are reversed and sent back to sequence processing layer 20 to complete the closed loop of topology parameters for window truncation.

[0199] Prediction module 52 data Output the churn probability parameter for this period. =0.78, decision engine module 53 data Generate marketing intervention action sequences For example, distributing a 20 yuan discount coupon. Because... =0.78≥intervention critical threshold =0.7, the decision engine module 53 determines that the validity of the user's original image and text features has decreased, and issues a cache eviction flag to the feature cache table 13; the feature cache table 13 then clears the historical text feature vector associated with the user while retaining the image feature vector.

[0200] Subsequently, the feature extraction layer 10 is analyzed. The text on the coupon, such as "20 yuan off for purchases over a certain amount, limited-time offer," was pieced together to form a sequence of descriptive text for intervention. The forward propagation of language model module 12 is triggered to reconstruct the feature vector of the intervention text. The text feature vector is then written back to the corresponding entry in feature cache table 13, overwriting only the text feature vector. At this point, the user's semantic feature profile has been updated from the original advertising copy to the intervention message, and will be read and participate in cross-modal fusion in step S510 of the next calculation cycle.

[0201] As can be seen from the above single-cycle walkthrough, the system only becomes active when the risk of data loss increases ( =0.8) and the behavior is significantly deviated ( Heavy global collaborative fusion computation is only initiated when both conditions are met (=0.7), and based on the prediction results ( =0.78) Reverse reconstruction of underlying text features, while switching to silent monitoring in low-risk or low-offset states, thereby reducing the computing power overhead under normal conditions while maintaining decision sensitivity, demonstrating the overall technical effect of the synergistic effect of each feedback link in this invention.

[0202] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic marketing decision-making system based on multimodal temporal fusion, characterized in that, It includes a cascaded feature extraction layer, sequence processing layer, metric layer, fusion layer, and decision layer; The feature extraction layer is used to receive and process advertising image data and advertising text data that are updated infrequently, extract image features and text features and persistently store them in the internal feature cache table in the form of key-value pairs; The sequence processing layer is used to receive and process high-frequency updated user behavior sequence data streams, and output the encoded extracted temporal aggregation features and the set of hidden state sequences. The metric layer is used to receive the set of hidden state sequences, measure the distribution shift of the user behavior sequence in the probability space, and output the quantized shift metric result to the fusion layer. The fusion layer is used to perform cross-modal feature interaction of multimodal data. It is connected to the feature cache table and the sequence processing layer respectively to obtain static image and text features and the temporal aggregation features, and fuse them to generate a multimodal fusion feature vector. The decision layer is used to perform graph network updates based on the multimodal fusion feature vector, predict churn probability parameters, and generate a sequence of marketing intervention actions; The system is also configured with three reverse parameter feedback links, including: The first feedback link is used to feed back the churn probability parameter output by the decision layer to the fusion layer as a trigger control parameter for determining whether to initiate cross-modal feature interaction. The second feedback link is used to feed back the feature dimension weight vector output by the fusion layer to the metric layer as the dimension mask vector when calculating its distribution offset. The third feedback link is used to feed back the marketing intervention action sequence output by the decision-making layer to the feature cache table, which is used to trigger cache eviction under specific conditions.

2. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 1, characterized in that, The feature extraction layer internally includes a convolutional network module, a language model module, and a feature cache table. The convolutional network module and the language model module are respectively used to perform feature encoding mapping on the original advertising image data and advertising text data. The resulting image feature vector and text feature vector are stored in the feature cache table in a key-value pair structure using a combination of user identifier and advertising material identifier as the index key. When a query request fails to find the target data, the feature extraction layer retrieves the original data again, performs feature encoding mapping, and refills the feature cache table.

3. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 1, characterized in that, The sequence processing layer contains a long short-term memory network module, and the decision layer contains a graph neural network module. The long short-term memory network module receives local topological feature parameters returned by the graph neural network module across layers, and adaptively adjusts its own sequence input time window accordingly. Within the window, it performs feature encoding on the user behavior sequence data stream and outputs the set of hidden state sequences to the metric layer.

4. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 3, characterized in that, The local topological feature parameters include the rate of change of local clustering coefficients and the temporal edge decay weights. The long short-term memory network module calculates the sequence input time window at the current moment by combining the local clustering coefficient change rate and the temporal edge decay weights with the reference time window length: when the local clustering coefficient change rate increases, the time window is reduced; when the temporal edge decay weights increase, the time window is expanded; and the sequence input time window is restricted to a preset safe range by a boundary truncation function.

5. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 1, characterized in that, The metric layer includes an offset calculation module, and the fusion layer includes an encoder module. The offset calculation module extracts the feature dimension weight vector previously output by the encoder module through the second feedback link, and uses this feature dimension weight vector to perform dimension-wise weighting on the current hidden state sequence set and the historical hidden state sequence set. Based on the weighted set, the current empirical probability distribution and the historical empirical probability distribution are constructed respectively, and the optimal transmission distribution distance between the two distributions is calculated and output to the fusion layer as the offset metric result.

6. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 5, characterized in that, The fusion layer is further provided with a scheduling module, and the decision layer is provided with a prediction module; the scheduling module obtains the churn probability parameter output by the prediction module in the cycle of the last global collaborative fusion calculation through the first feedback link, and calculates the dynamic trigger threshold by combining the trigger sensitivity scaling factor and the basic tolerance bias term. The scheduling module compares the optimal transmission distribution distance with the dynamic trigger threshold. When the optimal transmission distribution distance is greater than or equal to the dynamic trigger threshold, it outputs an enable signal to activate the encoder module to perform the cross-modal feature interaction.

7. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 6, characterized in that, After the encoder module is activated, it reads the image feature vector and text feature vector from the feature cache table using the combination of the current user identifier and the associated advertising material identifier as the key, and concatenates them with the last hidden state vector in the hidden state sequence set to form a multimodal initial tensor. The multimodal initial tensor is then subjected to attention operation and dimensionality reduction by the internal Transformer encoder to generate the multimodal fusion feature vector. At the same time, the feature dimension gating unit generates the feature dimension weight vector, which represents the proportion of importance of each element in the corresponding dimension.

8. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 3, characterized in that, The graph neural network module uses the multimodal fusion feature vector as the initial feature embedding of the target user node, calculates the aggregation weight parameters of the neighbor nodes in the local subgraph in the bipartite graph by combining the time interval decay term, and performs weighted summation to output the updated feature representation of the target user node. Based on the graph evolution, the ratio of the actual number of edges to the maximum possible number of edges at the current moment is calculated to generate the current local clustering coefficients, and then the rate of change of the local clustering coefficients and the temporal edge decay weights required for backhaul are generated.

9. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 8, characterized in that, The decision-making layer also includes a prediction module and a decision engine module. The prediction module predicts the churn probability parameter for the current period based on the updated target user node feature representation through nonlinear network mapping. The decision engine module generates the marketing intervention action sequence based on the same feature representation. When the churn probability parameter is determined to be greater than or equal to a preset intervention threshold, the decision engine module issues a cache eviction flag containing user identifiers and material identifiers to the feature cache table through the third feedback link to trigger the cache eviction.

10. The dynamic marketing decision-making system based on multimodal temporal fusion according to claim 9, characterized in that, Upon receiving the cache eviction flag, the feature cache table clears the corresponding historical text feature vectors while retaining the image feature vectors; the feature extraction layer extracts the structured text data from the marketing intervention action sequence, reassembles it to generate an intervention description text sequence, and triggers the language model module to perform forward propagation to reconstruct text features. The generated intervention text feature vectors are written back to the feature cache table to overwrite the original text feature vectors, thereby blocking the interference of invalid semantic features on subsequent feature interactions.

Images