Precision marketing intelligent application system based on AI training reasoning and data fusion

Abstract

The invention relates to a precision marketing intelligent application system based on AI training reasoning and data fusion, in particular to the field of artificial intelligence, realizes low-delay localized preliminary fusion through real-time feature extraction and event detection of multi-modal flow data by an edge side, and realizes real-time feature extraction and event detection of multi-modal flow data at a cloud end. A time-varying attention mechanism of multi-dimensional information is constructed and fused based on a dynamic tense map, the user-event association strength changing in real time can be accurately represented, and the online prediction and uncertainty perception ability of the time-varying attention mechanism are combined with a meta-learning rapid tuning mechanism triggered based on specific conditions, so that the user-event association strength can be accurately represented. The prediction adaptability and reliability in the sudden market mode are remarkably improved, finally, the prediction confidence is fed back to the edge node to dynamically optimize the data processing strategy, and a complete and adaptive optimization closed loop from cloud intelligent decision to edge data perception is formed. Therefore, the collaborative improvement of marketing response speed, precision and resource efficiency is realized on the whole.

Description

Precision Marketing Intelligent Application System Based on AI Training Inference and Data Fusion Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a precision marketing intelligent application system based on AI training inference and data fusion. Background Technology

[0002] In today's digital marketing environment, the dynamics and complexity of consumer behavior have significantly increased, especially during new product launches, major shopping festivals, or sudden social hot topics. Massive amounts of interactive data flood in from multiple channels, including social media, e-commerce platforms, offline smart devices, and customer service systems. This data comes in various forms, including structured click and purchase records as well as unstructured comments, live video streams, and emotional voice messages. Together, they form a valuable but chaotic source of information reflecting market trends and user intentions. Market opportunities are often fleeting, and the speed of capturing and responding to consumer intentions directly determines the success or failure of marketing campaigns. This requires back-end systems to have near real-time situational awareness and decision-making capabilities.

[0003] However, existing marketing intelligence systems face severe technical bottlenecks in processing multimodal, cross-domain, and high-concurrency streaming data. Firstly, at the data access and fusion level, existing architectures struggle to efficiently synchronize, semantically align, and assess the credibility of heterogeneous data streams. For example, there is a dynamic time-latency correlation between emotional outbursts on social media and changes in search behavior on e-commerce platforms, and simple window aggregation or weighted average fusion methods lose crucial temporal causal information and cannot effectively distinguish noise and anomalies. Secondly, at the model inference and adaptation level, widely deployed predictive models trained based on historical cyclical patterns struggle to identify and adapt to new, unseen behavioral patterns triggered by sudden events, despite the challenges posed by streaming data. While learning technologies are partially applied, their stability and rapid generalization capabilities during drastic changes in data distribution remain insufficient. This results in a lag in the overall system response, making it impossible to complete the closed loop from data perception to strategy optimization within minutes or even seconds. The direct consequence is that marketing actions lag significantly behind market changes, not only missing golden marketing windows but also potentially causing user resentment due to untimely push notifications. At the same time, system resources are heavily consumed in processing low-value or conflicting data, leading to inefficiency and cost waste. Therefore, how to achieve real-time and accurate fusion of multimodal streaming data and build a predictive inference engine with online mutation perception and rapid adaptive capabilities has become a core technical challenge hindering further improvement in the intelligence level of precision marketing. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a precision marketing intelligent application system based on AI training inference and data fusion. It solves the problems mentioned in the background art through an edge feature processing module, a graph construction and association module, an online prediction and optimization module, and a feedback optimization module.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: an edge feature processing module, a graph construction and association module, an online prediction and optimization module, and a feedback optimization module connected in sequence, wherein: the edge feature processing module is deployed at each edge node and is used to respond to receiving real-time multimodal stream data from the connected data source, perform feature extraction and event detection on the unstructured data in the real-time multimodal stream data, generate potential event markers with timestamps, and aggregate the structured behavioral data in the real-time multimodal stream data within a time window, aligning and binding the potential event markers and the aggregated structured behavioral data in the time dimension, generating and outputting a primary fusion data packet; the graph construction and association module is used to respond to receiving... Multiple primary fusion data packets from various edge feature processing modules are used to dynamically construct and update a global temporal graph based on entity, relation, timestamp, and tag information in the primary fusion data packets. The edge weights connecting nodes in the temporal graph are dynamically calculated using a time-varying attention weight calculation function. This function integrates node feature similarity, event occurrence time proximity, and event abruptness information, thus outputting a dynamic temporal graph that reflects real-time edge weights. The online prediction and optimization module, in response to the current state of the received dynamic temporal graph, uses an online prediction unit to infer the behavior prediction probability for each user node based on the dynamic temporal graph, and calls an uncertainty quantization unit for each node. The behavior prediction probability generates a confidence assessment value. When the edge weight growth rate of the event topic node in the dynamic temporal graph exceeds a preset threshold and the prediction confidence assessment value of the related user node is lower than a set threshold, a meta-learning tuning unit is activated. The meta-learning tuning unit takes the low-confidence user node and its local neighborhood structure in the dynamic temporal graph as input, performs small-sample fast tuning on the parameters of the online prediction unit, and finally outputs the behavior prediction result with the corresponding confidence assessment value. The feedback optimization module is used to respond to the received behavior prediction result and its confidence assessment value, and sends the confidence assessment value for different users and the associated edge node identifier as a feedback signal to trigger the corresponding edge feature processing module to dynamically adjust its edge feature based on the confidence assessment value. The internal feature extraction model selection and event detection threshold parameters form a closed-loop optimization from the cloud to the edge. In a preferred embodiment, the specific process of performing feature extraction and event detection on unstructured data in real-time multimodal streaming data to generate potential event markers with timestamps in the edge feature processing module is as follows: the received real-time multimodal streaming data is separated into a data part containing unstructured data and a data part containing structured behavioral data according to the data format and structural attributes; for each data unit in the data part containing unstructured data, a lightweight feature extraction model is used for calculation, and the input data unit is encoded and converted according to the internal parameters of the feature extraction model to generate a dense feature vector of fixed dimension.Maintain a sliding time window consisting of dense feature vectors generated at the current time and several consecutive times immediately preceding it; obtain all dense feature vectors within the sliding time window except for the dense feature vector at the current time, forming a historical feature vector set; perform principal component analysis on the historical feature vector set to obtain the corresponding principal component subspace; project the dense feature vector at the current time onto the principal component subspace to obtain a reconstructed feature vector; calculate the difference vector between the dense feature vector at the current time and the reconstructed feature vector; multiply the difference vector element-wise by a preset feature dimension importance mask vector, where the weights corresponding to different dimensions in the feature dimension importance mask vector are not fully defined. All identical features are multiplied to obtain a weighted difference vector. The Euclidean norm of the weighted difference vector is calculated, and the calculated norm value is defined as the burst deviation degree at the current time. When the burst deviation degree exceeds an adaptive threshold, it is determined as the starting point of a potential event. Monitoring continues until the burst deviation degree falls back below the adaptive threshold, and this interval is defined as a potential event segment. A globally unique identifier is generated for the potential event segment, and the start and end timestamps of the potential event segment are recorded. All dense feature vectors within the potential event segment are aggregated and calculated to generate a representative feature vector representing the potential event segment. The final output is a potential event marker containing a globally unique identifier, start and end timestamps, and a representative feature vector.

[0006] In a preferred embodiment, the process of aggregating structured behavioral data in real-time multimodal streaming data within a time window specifically involves: dividing the data portion containing structured behavioral data into a pre-defined, fixed-duration aligned time window; within each aligned time window, performing aggregation operations (counting or summing) on ​​behavioral records with the same user identifier in the structured behavioral data according to their behavioral type to generate an aggregated behavioral vector, and recording the time corresponding to the midpoint of the aligned time window as the center timestamp; outputting the aggregated behavioral vector and the center timestamp in pairs as the result of aggregating the structured behavioral data within the time window.

[0007] In a preferred embodiment, the specific process of aligning and binding the potential event markers with the aggregated structured behavioral data in the time dimension to generate and output a primary fusion data packet is as follows: For a generated potential event marker, it includes start and end timestamps; based on a preset time tolerance value, find all aggregated behavioral vectors in the time dimension whose center timestamps fall within the time interval jointly formed by the time obtained by subtracting the time tolerance value from the start time of the start and end timestamps and adding the time tolerance value to the end time of the start and end timestamps; bind and associate the globally unique identifier, start and end timestamps, and representative feature vectors contained in the potential event marker with all aggregated behavioral vectors found in the time interval and their corresponding user identifiers and center timestamps, and package them to generate a primary fusion data packet.

[0008] In a preferred embodiment, the specific process of dynamically calculating the edge weights connecting nodes in the temporal graph using a time-varying attention weight calculation function in the graph construction and association module is as follows: Parse the primary fusion data packet, extracting the start and end timestamps corresponding to the globally unique identifier and used to represent the event subject entity, the user entity corresponding to the user identifier, the center timestamp associated with the user behavior, the representative feature vector corresponding to the event subject entity, and the aggregated behavior vector corresponding to the user entity; based on the extracted representative feature vector and aggregated behavior vector, calculate the dot product between the representative feature vector and the aggregated behavior vector, and perform scaling and nonlinear activation processing on the dot product result to obtain the first attention score representing the feature similarity between nodes; obtain the center timestamp associated with the user behavior in the primary fusion data packet, calculate the absolute time difference between the center timestamp and the current time, and convert the absolute time... The difference is substituted into an exponential decay function with a natural constant as the base and a negative decay coefficient as the exponent to obtain the second decay factor characterizing the time decay effect. From the information provided by the edge feature processing module and associated with the potential event markers in the primary fusion data packet, an event abrupt change intensity scalar value is obtained. The event abrupt change intensity scalar value originates from the burst deviation calculated by the edge feature processing module during the generation of potential event markers. The event abrupt change intensity scalar value is first scaled, then subjected to a nonlinear compression transformation, and then the compression transformation result is multiplied by a modulation coefficient. Finally, the product result is added to a constant to obtain the third modulation factor characterizing the event abrupt change modulation effect. The first attention score, the second decay factor, and the third modulation factor are continuously multiplied, and the final product result is used as the edge weight connecting the user entity and the event topic entity at the current time.

[0009] In a preferred embodiment, the specific operation of outputting a dynamic temporal graph that reflects real-time edge weights is as follows: maintaining a global temporal graph, which includes a node set and an edge set; for each received primary fusion data packet, performing the following operations to incrementally update the temporal graph: in response to the absence of an event topic entity node in the node set of the temporal graph corresponding to the globally unique identifier extracted from the primary fusion data packet, creating an event topic entity node in the temporal graph identified by the globally unique identifier, and storing the representative feature vector and start and end timestamps extracted from the primary fusion data packet as attributes of the node; in response to the absence of a user entity node in the node set of the temporal graph corresponding to the user identifier extracted from the primary fusion data packet, creating a user entity node in the temporal graph identified by the user identifier, and storing the aggregation behavior vector extracted from the primary fusion data packet as attributes of the node; establishing or locating a directed edge between the user entity node and the event topic entity node, and using the center timestamp extracted from the primary fusion data packet as the time of the first association of the directed edge. The time-of-flight attribute is stored; if the directed edge already exists, the stored first association time attribute is retained unchanged; based on the current time, the time-varying attention weight calculation function is called to calculate a real-time edge weight, and this real-time edge weight is stored as the weight attribute of the directed edge; new primary fusion data packets are continuously processed, and for each new primary fusion data packet, the operations of creating a node in response to the absence of a corresponding event topic entity node in the node set of the temporal graph, and creating a node in response to the absence of a corresponding user entity node in the node set of the temporal graph are repeatedly executed. The system performs operations such as establishing or locating directed edges between user entity nodes and event topic entity nodes and storing the attribute of the first association time, as well as calling the time-varying attention weight calculation function to calculate and store the real-time edge weights. This allows for incremental updates and real-time recalculation of edge weights in the temporal graph, maintaining a dynamic temporal graph where edge weights evolve over time. The graph construction and association module generates and outputs a graph state snapshot reflecting all nodes, edges, and their real-time weights at a given time, based on the current state of the maintained dynamic temporal graph, at fixed time intervals or in response to external prediction requests.

[0010] In a preferred embodiment, the online prediction and optimization module utilizes an online prediction unit to perform inference based on a dynamic temporal graph, generating behavioral prediction probabilities for user nodes, and calls an uncertainty quantification unit to generate a confidence assessment value for each behavioral prediction probability. The specific operation is as follows: the online prediction unit, based on the current state of the received dynamic temporal graph, performs inference for each user node in the dynamic temporal graph and outputs the behavioral prediction probability distribution for each user node; the uncertainty quantification unit calculates a confidence assessment value for each user node. This calculation process is as follows: first, based on the edge connection relationships and real-time edge weights between the user node and other nodes in the dynamic temporal graph, it obtains all nodes connected to the user node through the same event topic entity node in the dynamic temporal graph. Other user nodes form a set of associated user nodes. Based on the real-time edge weights between the user node, each node in the associated user node set, and the commonly connected event topic entity nodes in the dynamic temporal graph, the behavior prediction probability distributions of each node in the associated user node set are weighted and aggregated to obtain the local neighborhood prediction consensus distribution of the user node. Next, the graph-aware prediction divergence between the user node's own behavior prediction probability distribution and the local neighborhood prediction consensus distribution is calculated, and this divergence value is amplified by a modulation factor related to the variance of the edge weights. Then, the information entropy of the user node's own behavior prediction probability distribution is calculated. Finally, the information entropy and the modulated graph-aware prediction divergence are weighted and summed, and the negative exponential function value of the summation result is used as the confidence evaluation value of the user node.

[0011] In a preferred embodiment, the process of outputting the behavior prediction result with corresponding confidence evaluation value specifically includes: for each event topic node in the dynamic temporal graph, calculating its edge weight growth rate within the most recent preset time window; specifically: calculating the weighted degree centrality of the event topic node at the calculation time, where the weighted degree centrality is the sum of the real-time edge weights between the event topic node and all connected user nodes; calculating the weighted degree centrality of the event topic node at the time before the preset time window; subtracting the weighted degree centrality of the previous time from the current time's weighted degree centrality, and then dividing by the length of the preset time window, to obtain the instantaneous change rate of the edge weights of the event topic node, which is used as its edge weight growth rate; for For each event topic node in the dynamic temporal graph, a set of all user nodes directly connected to that event topic node via edges is obtained. The number of user nodes in this set whose confidence assessment value is lower than a preset low confidence threshold is counted, and the proportion of this number to the total number of user nodes in the set is calculated as the low confidence user association ratio of that event topic node. When both the edge weight growth rate of an event topic node exceeds a preset first growth rate threshold and the low confidence user association ratio of that event topic node exceeds a preset second ratio threshold, the meta-learning optimization unit is activated. The meta-learning optimization unit uses all user nodes associated with this event topic node corresponding to the trigger condition whose confidence assessment value is lower than the low confidence threshold. The support set for a few-shot learning task is formed by the support set and its neighborhood subgraphs defined within a preset hop count range in the dynamic temporal graph. The meta-learning tuning unit calculates the mean of the feature vectors of all user nodes in this support set after transformation by the intermediate layer of the online prediction unit, thus obtaining the prototype feature representation of the few-shot learning task. Based on this support set, the meta-learning tuning unit calculates the current gradient of the parameters of the online prediction unit. Simultaneously, it retrieves the gradients of several historical tasks most similar to the current task's prototype feature representation from a memory bank storing historical task gradient information, including the task prototype feature representations and their corresponding parameter gradients recorded during meta-learning tuning when activation conditions were met. The average direction of the historical task gradient is obtained. The meta-learning tuning unit weights and fuses the current gradient with the average direction of the historical task gradient to obtain an aligned fused gradient. This fused gradient is used to perform a gradient update on the parameters of the online prediction unit, i.e., subtracting the product of this fused gradient and a preset learning rate from the current parameter value to obtain the tuned parameters. This gradient update is specifically performed on the parameters of the online prediction unit. Using the online prediction unit with updated parameters, the behavior prediction probability inference is re-performed for all target user nodes to generate a new behavior prediction probability distribution for each user node. At the same time, the uncertainty quantization unit is used to recalculate the confidence evaluation value of each user node based on this new behavior prediction probability distribution.Output the pair of the new behavior prediction probability distribution and the recalculated confidence score for each user node, as the final behavior prediction result with the confidence score.

[0012] In a preferred embodiment, the feedback optimization module sends out the confidence assessment values ​​and associated edge node identifiers for different users as feedback signals, triggering the corresponding edge feature processing module to dynamically adjust its internal feature extraction model selection and event detection threshold parameters based on the confidence assessment values. The specific process is as follows: The feedback optimization module receives the binary pair of behavior prediction probability distribution and confidence assessment values ​​for each user from the online prediction and tuning module, and obtains the edge node identifiers associated with each user; based on these edge node identifiers, it aggregates the confidence assessment values ​​of all users belonging to the same edge node to obtain the comprehensive confidence performance index for each edge node; the aggregation calculation uses a time decay-based approach. The weighted harmonic average method proceeds as follows: For an edge node, a recent time window is defined; for each processing moment within this time window, the following operations are performed: The number of users receiving data from the edge node at that processing moment is obtained and recorded; the reciprocal of the confidence assessment value for each user is calculated; a weight value is assigned to both the user count and the reciprocal of the confidence assessment value for each user; after completing the above operations for all processing moments within the time window, the weighted user counts for each processing moment are summed to obtain the numerator; the weighted reciprocals of the confidence assessment values ​​for each user at each processing moment are summed to obtain the denominator; the quotient obtained by dividing the numerator by the denominator is used as the quotient. The quotient is used as the comprehensive confidence performance index for the edge node. Based on the comprehensive confidence performance index, combined with preset high-performance and low-performance thresholds, an adaptive adjustment strategy is generated for each edge node. The specific strategy generation process is as follows: the comprehensive confidence performance index of the edge node is compared with the high-performance and low-performance thresholds; if the comprehensive confidence performance index is higher than or equal to the high-performance threshold, a strategy to simplify or maintain the existing processing is generated for the edge node; if the comprehensive confidence performance index is lower than or equal to the low-performance threshold, a strategy to enhance the processing depth is generated for the edge node. This enhanced processing strategy includes indicating a switch to a feature extraction model with higher computational complexity, and calculating a negative adjustment amount to reduce event detection. An adaptive threshold parameter is generated. If the overall confidence performance index is between the low-performance threshold and the high-performance threshold, a fine-tuning strategy is generated for the edge node to slightly adjust the adaptive threshold parameter for event detection. Before finally determining the adjustment instruction, the current resource utilization rate of the edge node is also obtained. If the resource utilization rate exceeds the safety threshold, the adjustment action involving increased computing load is restricted or postponed. Based on the generated adaptive adjustment strategy, a feedback control instruction is generated. The feedback control instruction is encapsulated in a structured data format, which includes the target edge node identifier, the type of action to be performed, and the specific parameter values ​​related to the action. The encapsulated feedback control signal is sent to the corresponding edge feature processing module through a message queue.

[0013] In a preferred embodiment, the specific process of forming a closed-loop optimization from the cloud to the edge is as follows: The generated feedback control command is sent to the edge feature processing module corresponding to the target edge node specified in the command. The edge feature processing module parses the received feedback control command and dynamically adjusts its operating configuration according to the command content. Dynamic adjustment includes: if the command requires switching the feature extraction model, loading the specified new model from the preset model library and replacing the currently running model; if the command requires adjusting the adaptive threshold parameter of event detection, updating the locally stored threshold variable with the new parameter value contained in the command; after completing the configuration adjustment, the edge feature processing module sends a response to the feedback optimization module. A status confirmation signal indicating that the adjustment has been completed is sent. The feedback optimization module receives this status confirmation signal and associates and stores the issued adjustment instruction, the corresponding target edge node identifier, and its comprehensive confidence performance index as a basis for historical adjustments. After completing the configuration adjustment, the edge feature processing module uses the new configuration parameters to process the subsequent real-time multimodal stream data and uploads the generated new primary fusion data package to the graph construction and association module. The prediction confidence evaluation value generated by the subsequent online prediction and optimization module based on the new data will be received and evaluated by the feedback optimization module again, thus forming a continuous edge data processing strategy optimization closed loop based on cloud prediction performance feedback.

[0014] The beneficial effects of this invention are as follows: By extracting features and detecting events in multimodal streaming data in real time at the edge, low-latency localized preliminary fusion is achieved. In the cloud, relying on a dynamic temporal graph to construct and fuse multi-dimensional information, a time-varying attention mechanism can accurately characterize the real-time changing user-event association strength. Its online prediction and uncertainty perception capabilities, combined with a meta-learning fast optimization mechanism triggered by specific conditions, significantly improve the predictive adaptability and reliability in sudden market modes. Finally, by feeding back the prediction confidence to the edge nodes to dynamically optimize the data processing strategy, a complete and adaptive optimization closed loop from cloud-based intelligent decision-making to edge data perception is formed, thereby achieving a synergistic improvement in marketing response speed, accuracy, and resource efficiency as a whole. Attached Figure Description

[0015] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1

[0020] This embodiment provides a precision marketing intelligent application system based on AI training inference and data fusion, as shown in Figures 1-2. Specifically, it includes: an edge feature processing module, a graph construction and association module, an online prediction and optimization module, and a feedback optimization module connected sequentially. The edge feature processing module is deployed at each edge node and is used to respond to real-time multimodal stream data received from connected data sources (i.e., connected sensors, application logs, and API interfaces). It performs feature extraction and event detection on unstructured data in the real-time multimodal stream data, generates potential event markers with timestamps, and performs feature extraction and event detection on the real-time multimodal stream data. The structured behavioral data in the temporal stream is aggregated within a time window. Potential event markers are aligned and bound to the aggregated structured behavioral data along the time dimension, generating and outputting a primary fusion data packet. The graph construction and association module responds to multiple primary fusion data packets received from various edge feature processing modules. Based on the entity, relationship, timestamp, and marker information in the primary fusion data packets, it dynamically constructs and updates a global temporal graph. The edge weights connecting nodes in the temporal graph are dynamically calculated using a time-varying attention weight calculation function, which integrates node features. The system analyzes similarity, the time proximity of events, and the mutability of events to output a dynamic temporal graph that reflects real-time edge weights. The online prediction and tuning module, responding to the current state of the received dynamic temporal graph, uses an online prediction unit to infer the behavior prediction probabilities for user nodes. It then calls an uncertainty quantification unit to generate a confidence assessment value for each behavior prediction probability. When the edge weight growth rate of the event topic node in the dynamic temporal graph exceeds a preset threshold and the prediction confidence assessment value of the relevant user nodes falls below a set threshold, a meta-learning tuning module is activated. The optimization unit, or meta-learning optimization unit, takes low-confidence user nodes and their local neighborhood structure in the dynamic temporal graph as input to perform small-sample rapid optimization of the parameters of the online prediction unit, and finally outputs the behavior prediction result with the corresponding confidence evaluation value. The feedback optimization module is used to respond to the received behavior prediction result and its confidence evaluation value, and sends the confidence evaluation value for different users and the associated edge node identifier as a feedback signal to trigger the corresponding edge feature processing module to dynamically adjust its internal feature extraction model selection and event detection threshold parameters according to the confidence evaluation value, forming a closed-loop optimization from the cloud to the edge.

[0021] In this embodiment, the specific process of performing feature extraction and event detection on unstructured data in real-time multimodal streaming data to generate potential event markers with timestamps in the edge feature processing module is as follows: The received real-time multimodal streaming data is separated into a data part containing unstructured data and a data part containing structured behavioral data according to data format and structural attributes; for each data unit in the data part containing unstructured data, a lightweight feature extraction model is used for calculation, and the input data unit is encoded and transformed according to the internal parameters of the feature extraction model to generate a dense feature vector of fixed dimension; the lightweight feature extraction model can be a miniature T-model that has undergone model pruning and quantization compression. A Ransformer encoder or a depthwise separable convolutional neural network, with a fixed dimension, such as 128 or 256, is used. A sliding time window is maintained, consisting of dense feature vectors generated at the current time step and several consecutive time steps immediately preceding it. The number of consecutive time steps included in the sliding time window can be preset according to the data stream frequency, for example, including features from the current time step and nine previous time steps, forming a sequence of length 10. All dense feature vectors within the sliding time window, excluding the dense feature vector at the current time step, are obtained, forming a historical feature vector set. Principal component analysis is performed on the historical feature vector set to obtain the corresponding principal component subspace. The cumulative variance can be selected as the principal component subspace. The top k principal component directions with a contribution rate exceeding 85% are used to construct the feature vector. The dense feature vector at the current time step is projected onto the principal component subspace to obtain a reconstructed feature vector. The difference vector between the dense feature vector at the current time step and the reconstructed feature vector is calculated. A pre-defined feature dimension importance mask vector is multiplied element-wise with the difference vector. The weights for different dimensions in the feature dimension importance mask vector are not entirely the same. After multiplication, a weighted difference vector is obtained. The feature dimension importance mask vector can be obtained through domain knowledge pre-setting or small-sample learning, and is used to amplify the contribution of feature dimensions with higher relevance to marketing intent (such as the correlation between sentiment polarity and specific keywords). The Euclidean norm of the weighted difference vector is calculated. The calculated norm value is defined as the burst deviation at the current moment. When the burst deviation value exceeds an adaptive threshold, it is determined as the starting point of a potential event. The adaptive threshold can be dynamically adjusted according to the statistics of the historical burst deviation sequence. For example, the initial value is set to the average value of the sequence plus twice the standard deviation, and it is fine-tuned according to the false alarm rate during the operation. The monitoring continues until the burst deviation value falls back below the adaptive threshold. This interval is defined as a potential event segment. A globally unique identifier is generated for the potential event segment, the start and end timestamps of the potential event segment are recorded, and all dense feature vectors in the potential event segment are aggregated and calculated to generate a representative feature vector representing the potential event segment.Aggregation calculations can employ mean pooling or max pooling operations, with the representative feature vector having the same dimension as the dense feature vector. The final output includes a globally unique identifier, start and end timestamps, and a potential event marker for the representative feature vector. This marker is represented as a structured data object (such as JSON format) for easy subsequent transmission and parsing. The aggregation process within a time window for structured behavioral data in real-time multimodal streaming data specifically involves dividing the data portion containing structured behavioral data according to a preset aligned time window of fixed or dynamic duration. The fixed duration of the aligned time window can be set according to the business scenario, such as 1 second, 5 seconds, or 10 seconds; the dynamic duration can be adjusted according to the data flow. The system adapts by shortening the window during peak traffic periods and extending it during off-peak periods. Within each aligned time window, behavior records with the same user identifier in the structured behavioral data are aggregated by counting or summing according to their behavior type, generating an aggregated behavior vector. Behavior types include clicks, browsing, adding to cart, and payments. Each dimension of the aggregated behavior vector corresponds to a statistical value for a behavior type, and the time corresponding to the midpoint of the aligned time window is recorded as the center timestamp. The center timestamp is used to accurately represent the time position of the aggregated behavior. The aggregated behavior vector and the center timestamp are output in pairs as the result of aggregating the structured behavioral data within the time window. This result is also structured... The data is organized into object form, including fields such as user identifier, aggregated behavior vector, and center timestamp. The specific process of aligning and binding potential event markers with the aggregated structured behavior data along the time dimension to generate and output the initial fusion data package is as follows: For a generated potential event marker, it includes start and end timestamps; based on a preset time tolerance value, all aggregated behavior vectors whose center timestamps fall within the time interval formed by subtracting the time tolerance value from the start time of the start and end timestamps and adding the time tolerance value to the end time of the start and end timestamps; the time tolerance value is used to compensate for data transmission and processing... Minor delays and reasonable time lags between actions and events are typically set to half the length of the alignment time window, such as 0.5 seconds or 2.5 seconds. The globally unique identifier, start and end timestamps, and representative feature vectors contained in the potential event marker are bound and associated with all aggregated behavior vectors found within the time interval, along with their corresponding user identifiers and center timestamps, and packaged into a primary fusion data packet. This binding and association forms a "one-to-many" relationship structure, where one potential event marker is associated with zero or more time-proximity aggregated user behavior results. The primary fusion data packet has a standardized data pattern; for example, it contains an event header object and a list of multiple behavior objects.The primary fusion data packet, as the module's output, represents potential events detected within a specific time period and their associated user behavior aggregation information in the vicinity of that time period. This packet format design eliminates the need for complex time alignment and data cleaning in subsequent graph construction modules, allowing direct parsing of the "user-event-behavior" triples and their attributes, significantly improving the efficiency of data fusion and processing.

[0022] In this embodiment, it is specifically necessary to explain the process by which the edge weights connecting nodes in the temporal graph are dynamically calculated using a time-varying attention weight calculation function in the graph construction and association module. The process involves parsing the primary fusion data packet, extracting the start and end timestamps corresponding to the globally unique identifier and used to represent the event subject entity, the user entity corresponding to the user identifier, the central timestamp associated with the user behavior, the representative feature vector corresponding to the event subject entity, and the aggregated behavior vector corresponding to the user entity. Based on the extracted representative feature vector and aggregated behavior vector, the dot product between the representative feature vector and the aggregated behavior vector is calculated, and the dot product result is scaled and nonlinearly activated to obtain the feature similarity between nodes. The first attention score is calculated by scaling the dot product result using a scaling factor, which is the square root of the common dimension of the representative feature vector and the aggregated behavior vector. For example, when both the representative feature vector and the aggregated behavior vector have a dimension of 128, the scaling factor is √128. This operation helps stabilize the gradient before subsequent activation function processing, preventing numerical instability during training due to excessively large vector dimensions or dot product values. Non-linear activation processing uses the Sigmoid function to map the scaled dot product result to a numerical range between zero and one. This function smoothly compresses any real number input to the (0,1) interval, giving the first attention score a clear probabilistic interpretation. When the score approaches positive infinity, it approaches 1, indicating a perfect semantic match; when it approaches negative infinity, it approaches 0, indicating complete irrelevance. The system obtains the central timestamp associated with user behavior in the primary fusion data packet, calculates the absolute time difference between the central timestamp and the current time, and substitutes this absolute time difference into an exponential decay function with a base of the natural constant and an exponent of a negative decay coefficient. This yields a second decay factor representing the time decay effect. The decay coefficient is a preset, configurable constant greater than zero, used to control the rate at which weights decay over time. Its typical value range is 0.1 to 0.5, determined through experimental calibration. For example, when the decay coefficient is set to 0.3, the edge weight 24 hours after the association occurs will decay to... The initial value is approximately 5%, which effectively simulates the natural decay process of the correlation between user interest and events. From the information provided by the edge feature processing module and associated with the potential event markers in the primary fusion data packet, an event mutation intensity scalar value is obtained. The event mutation intensity scalar value originates from the burst deviation calculated by the edge feature processing module during the generation of potential event markers. It is usually taken as the maximum value or time-weighted average of the burst deviation at all calculation times within the event segment. The event mutation intensity scalar value is first scaled, then nonlinearly compressed, and then the compression transformation result is multiplied by a modulation coefficient. Finally, the product result is added to a constant to obtain the third modulation factor characterizing the event mutation modulation effect.Scale scaling uses a preset, positive-zero scaling factor to divide the scalar value of the event catastrophicity intensity. This normalizes catastrophicity intensities from different data sources or with different dimensions to a standard range, facilitating subsequent processing. Nonlinear compression transform uses a hyperbolic tangent function to map the scaled result to a numerical range between -1 and +1. This function compresses the input to a finite range of (-1, 1), ensuring that the modulation effect does not fluctuate drastically due to extreme abrupt changes. For example, a scaled value of 0.8 will be approximately 0.66 after the hyperbolic tangent transformation. The modulation factor is a preset, configurable constant greater than or equal to zero, used to control the amplification of the event catastrophicity's edge weights. Its value range is usually set between [0,1]. When the modulation coefficient is 0, the third modulation factor is always 1, that is, the impact of event mutation is ignored; when the modulation coefficient is 0.5, the modulation contribution to the transformation result of 0.66 in the above example is 0.33, making the third modulation factor 1.33, which means that the weight of the associated edge is enhanced by 33%; the first attention score, the second attenuation factor and the third modulation factor are continuously multiplied, and the final product is used as the edge weight between the user entity and the event subject entity at the current time. This calculation method realizes the coupling of semantic relevance, temporal freshness and event impact, so that the edge weight can comprehensively and dynamically represent the user-event association at any time. The specific operations for outputting a dynamic temporal graph that reflects real-time edge weights are as follows: Maintain a global temporal graph, which includes a set of nodes and a set of edges; For each received primary fusion data packet, perform the following operations to incrementally update the temporal graph: In response to the absence of an event topic entity node in the node set of the temporal graph corresponding to the globally unique identifier extracted from the primary fusion data packet, create an event topic entity node in the temporal graph identified by the globally unique identifier, and store the representative feature vector and start and end timestamps extracted from the primary fusion data packet as attributes of the node; In response to the absence of an event topic entity node in the node set of the temporal graph corresponding to the globally unique identifier extracted from the primary fusion data packet, create an event topic entity node in the temporal graph identified by the globally unique identifier, and store the representative feature vector and start and end timestamps extracted from the primary fusion data packet as attributes of the node; For the user entity node corresponding to the user identifier extracted from the primary fusion data packet, a user entity node with that user identifier is created in the temporal graph, and the aggregated behavior vector extracted from the primary fusion data packet is stored as an attribute of the node. A directed edge is established or located between the user entity node and the event topic entity node, and the center timestamp extracted from the primary fusion data packet is stored as the first association occurrence time attribute of the directed edge. If the directed edge already exists, the stored first association occurrence time attribute is retained unchanged. Based on the current time, the time-varying attention weight calculation function is called to calculate a real-time edge weight, and the real-time edge weight is stored as the weight attribute of the directed edge.The value of the weight attribute depends on the current time of calculation, causing the weight attribute value of the same directed edge to change over time. That is, the topology of the graph (the existence of nodes and edges) is relatively stable, but the weight signal carried on its edges is a dynamic value that changes continuously over time. New primary fusion data packets are continuously processed. For each new primary fusion data packet, the following operations are repeatedly executed: creating a node if no corresponding event subject entity node is found in the node set of the temporal graph; creating a node if no corresponding user entity node is found in the node set of the temporal graph; establishing or locating a directed edge between the user entity node and the event subject entity node and storing the attribute of the first association time; and calling the time-varying attention weight calculation function to calculate and store the real-time edge weights. This allows for incremental updates to the temporal graph and real-time recalculation of edge weights, maintaining a dynamic time-varying edge weight. The graph construction and association module generates and outputs a graph state snapshot reflecting all nodes, edges, and their real-time weights at a given time, based on the current state of the maintained dynamic temporal graph, at fixed time intervals or in response to external prediction requests. This snapshot is then provided to the subsequent online prediction and optimization modules. The fixed time interval can be configured according to the real-time requirements of the business, for example, set to 1 second or 0.5 seconds. The graph state snapshot is a complete data structure containing all event-themed entity nodes and their representative feature vectors and start / end timestamp attributes, all user entity nodes and their aggregate behavior vector attributes, and all directed edges and their corresponding first association times and real-time edge weight attributes at the time of snapshot generation. This snapshot freezes the complete state of the dynamically evolving graph at a specific moment, providing stable and up-to-date input for downstream prediction algorithms based on static graph structures.

[0023] In this embodiment, it is specifically necessary to explain that in the online prediction and optimization module, the online prediction unit performs inference based on the dynamic temporal graph to generate behavioral prediction probabilities for user nodes, and calls an uncertainty quantization unit to generate a confidence evaluation value for each behavioral prediction probability. The specific operation is as follows: the online prediction unit, based on the current state of the received dynamic temporal graph, performs inference for each user node in the dynamic temporal graph and outputs the behavioral prediction probability distribution for each user node; the uncertainty quantization unit calculates a confidence evaluation value for each user node. The calculation process is as follows: first, based on the edge connection relationship and real-time edge weight between the user node and other nodes in the dynamic temporal graph, the confidence evaluation value is obtained in the dynamic temporal graph. All other user nodes connected to the current user node through the same event theme entity node constitute a set of associated user nodes. This set reflects a group of users with similar behavioral relationships to the current user under the current marketing event theme. Based on the real-time edge weights between the current user node, each node in the associated user node set, and the commonly connected event theme entity node in the dynamic temporal graph, the behavioral prediction probability distribution of each node in the associated user node set is weighted and aggregated to obtain the local neighborhood prediction consensus distribution of the current user node. In the weighted aggregation, the associated user node with the higher the edge weight contributes more to the consensus distribution of its prediction probability distribution, which reflects the influence of association strength on the consistency of group behavior. Secondly, the calculation... The graph-aware prediction divergence between the user node's own behavior prediction probability distribution and the local neighborhood prediction consensus distribution is amplified by a modulation factor related to the edge weight variance; the larger the edge weight variance, the greater the amplification. The edge weight variance reflects the stability of the interest consistency between the user and the associated user group. The larger the variance, the less stable the interest consistency or the greater the intra-group differences. In this case, the credibility of the user's individual prediction deviating from the group consensus is lower, thus requiring a larger divergence value as a warning. Next, the information entropy of the user node's own behavior prediction probability distribution is calculated. Finally, the information entropy is weighted and calculated with the modulated graph-aware prediction divergence. The negative exponential function value of the summation result is used as the confidence evaluation value of the user node. The lower the confidence evaluation value, the less reliable the behavior prediction of the user node. This confidence evaluation value integrates the uncertainty of individual prediction and the degree of deviation between individual prediction and local group consensus, providing a reliable reliability measure for downstream decision-making. The process of outputting the behavior prediction result with the corresponding confidence evaluation value is as follows: For each event topic node in the dynamic temporal graph, calculate its edge weight growth rate in the most recent preset time window; specifically, calculate the weighted degree centrality of the event topic node at the calculation time. The weighted degree centrality is the sum of the real-time edge weights between the event topic node and all connected user nodes.Calculate the weighted degree centrality of the event topic node at the moment preceding the preset time window; subtract the weighted degree centrality of the previous moment from the current moment's weighted degree centrality, and then divide by the length of the preset time window to obtain the instantaneous change rate of the edge weights of the event topic node, which is used as its edge weight growth rate; the preset time window can be set according to the dynamic nature of the marketing scenario, such as 5 minutes or 10 minutes; the edge weight growth rate reflects the instantaneous change speed at which the marketing event or topic attracts the overall attention of users, and is one of the leading indicators for detecting the formation of market hotspots; for each event topic node in this dynamic temporal graph, obtain the edges directly connected to the event topic node. For all connected user nodes, count the number of user nodes in the set whose confidence evaluation value is lower than a preset low confidence threshold, and calculate the proportion of this number to the total number of user nodes in the set. This proportion is taken as the low confidence user association ratio of the event topic node. The low confidence threshold can be set, for example, between 0.5 and 0.7, to define the range of users with insufficient prediction reliability. The low confidence user association ratio reflects the model's overall grasp of the predicted user group associated with the event. The higher the ratio, the more difficult it is for the model to understand the user behavior patterns under the influence of the event. When the growth rate of the edge weight of an event topic node exceeds a preset first growth rate, the association ratio is considered complete. When the rate threshold is reached and the proportion of users associated with the low-confidence topic node of the event exceeds a preset second proportion threshold, the meta-learning tuning unit is activated. The first growth rate threshold can be set to, for example, a 20% increase in the total weight per hour, to identify rapidly emerging hot topics. The second proportion threshold can be set to, for example, 30%, to identify the group range where model predictions fail. These two conditions together constitute the intelligent triggering mechanism for model adaptation, ensuring that tuning resources are accurately deployed to key scenarios that are "emerging and model-confused". The meta-learning tuning unit uses all user nodes associated with the event topic node corresponding to the triggering condition whose confidence evaluation value is lower than the low-confidence threshold. Together with its neighborhood subgraph defined within a preset hop count range in the dynamic temporal graph, the support set for a few-shot learning task is formed. The preset hop count range typically includes a 1-hop or 2-hop neighborhood to ensure sufficient capture of relevant contextual information. This support set focuses on "hard samples" where the model is unsure of the predictions and their local environments, providing high-quality input for targeted learning. The meta-learning tuning unit calculates the mean of the feature vectors of all user nodes in this support set after transformation by the intermediate layer of the online prediction unit, obtaining the prototype feature representation of this few-shot learning task. The prototype feature representation abstracts the core pattern of the current learning task and is used for comparison with historical tasks.The meta-learning tuning unit calculates the current gradient of the parameters of the online prediction unit based on this support set. Simultaneously, it retrieves the gradients of several historical tasks most similar to the current task's prototype feature representation from a memory storing historical task gradient information. This historical task gradient information includes the task prototype feature representations and their corresponding parameter gradients recorded during meta-learning tuning when activation conditions were previously met. The unit then calculates the average direction of these historical task gradients. The similarity of historical tasks can be measured by calculating the cosine similarity between prototype feature representations. This step enables the acquisition of transferable tuning directions from historical experience. The meta-learning tuning unit then compares this current gradient with the gradients of these historical tasks. The average directions of the gradients are weighted and fused to obtain an aligned fused gradient. This weighted fusion is performed as follows: the current gradient is multiplied by a first coefficient, the average direction of the historical task gradients is multiplied by a second coefficient, and the two products are summed to obtain the fused gradient. The sum of the first and second coefficients is a constant, and the value of the second coefficient is between zero and one. Its specific value is adaptively determined based on the number of sample nodes in the current support set and the similarity between the average directions of the current gradient and the historical task gradients. The smaller the number of support set samples or the higher the similarity, the larger the value of the second coefficient. This adaptive weighting mechanism ensures that even with small... In cases with sufficient samples, historical experience is relied upon more for stable learning, while in cases with sufficient samples or when the current data differs significantly from historical experience, the current data is relied upon more. This fused gradient is used to perform a gradient update on the parameters of the online prediction unit, i.e., subtracting the product of this fused gradient and a preset learning rate from the current parameter values ​​to obtain the optimized parameters. This gradient update is specifically performed on the parameters of the online prediction unit. The preset learning rate is usually set to a small value, such as 0.001 or 0.0001, to ensure the stability of rapid optimization. This update aims to quickly adapt to the new pattern with minimal parameter perturbation. The online prediction unit with updated parameters is then used for all target user nodes. The behavior prediction probabilistic inference is re-performed to generate a new behavior prediction probability distribution for each user node. Simultaneously, using the uncertainty quantification unit, the confidence assessment value for each user node is recalculated based on this new behavior prediction probability distribution. The output is a tuple of this new behavior prediction probability distribution and the recalculated confidence assessment value for each user node, serving as the final behavior prediction result with the confidence assessment value. This tuple output provides downstream precision marketing decisions (such as personalized push notifications and resource allocation) with complete information input that combines predictive directionality and reliability assessment. Furthermore, through a confidence feedback mechanism, it influences upstream edge data processing strategies, forming a global optimization closed loop.

[0024] In this embodiment, it is specifically necessary to explain the process by which the feedback optimization module sends out the confidence assessment values ​​and associated edge node identifiers for different users as feedback signals, triggering the corresponding edge feature processing module to dynamically adjust its internal feature extraction model selection and event detection threshold parameters based on the confidence assessment values. The feedback optimization module receives the tuple of the behavior prediction probability distribution and confidence assessment values ​​for each user from the online prediction and tuning module, and obtains the edge node identifiers associated with each user. Based on these edge node identifiers, the confidence assessment values ​​of all users belonging to the same edge node are aggregated to obtain the comprehensive confidence performance index for each edge node. The aggregation calculation uses a time-decay weighted tuning... The averaging method, specifically, involves the following steps: For an edge node, a recent time window is set, for example, with a length of one hour or one processing cycle (e.g., 100 data processing batches). For each processing moment within this time window, the following operations are performed: Identify the users who received data from the edge node at that processing moment and record the number of these users; calculate the reciprocal of the confidence assessment value for each of these users; assign a weight value that decays over time to both the user count and the reciprocal of the confidence assessment value for each user, where the closer the processing moment is to the current moment, the greater the assigned weight value. For example, the weight value can be calculated using an exponential decay function, with newer data receiving a higher weight. After completing all processing within the time window... After the above operations at each processing time, the weighted user counts at each processing time are summed up as the numerator; the reciprocals of the weighted confidence assessment values ​​for each processing time and each user are summed up as the denominator; the quotient obtained by dividing the numerator by the denominator is calculated, and this quotient is used as the comprehensive confidence performance index for that edge node. This calculation method makes the comprehensive confidence performance index particularly sensitive to samples with low confidence assessment values ​​(i.e., high prediction uncertainty). When the data transmitted by an edge node frequently leads to low confidence predictions, its performance index will decrease significantly, thus more accurately exposing quality problems of the data source or problems such as feature extraction mismatch. This performance index is sensitive to samples with low confidence assessment values ​​and can effectively reflect the impact of the data provided by the edge node on cloud predictions. The overall level of uncertainty is measured. Based on the comprehensive confidence performance index, combined with preset high-performance and low-performance thresholds, an adaptive adjustment strategy is generated for each edge node. The specific strategy generation process is as follows: the comprehensive confidence performance index of the edge node is compared with the high-performance threshold and the low-performance threshold. For example, the high-performance threshold and the low-performance threshold can be set to 0.8 and 0.4 respectively. The specific values ​​need to be calibrated according to the prediction confidence distribution in the actual business scenario. If the comprehensive confidence performance index is higher than or equal to the high-performance threshold, a strategy to simplify or maintain the existing processing is generated for the edge node. For example, it can be instructed to switch to a lightweight feature extraction model with lower computational overhead, or to maintain the current configuration when resources are sufficient, so as to save edge computing resources.If the overall confidence performance index is lower than or equal to the low-performance threshold, a strategy to enhance the processing depth is generated for the edge node. This enhancement strategy includes instructing a switch to a feature extraction model with higher computational complexity, and calculating a negative adjustment to reduce the adaptive threshold parameter for event detection. The absolute value of the adjustment is positively correlated with the difference between the low-performance threshold and the overall confidence performance index; that is, the worse the performance, the greater the threshold reduction, in order to capture weaker event signals and provide richer features. If the overall confidence performance index is between the low-performance threshold and the high-performance threshold, a fine-tuning strategy is generated for the edge node to make small adjustments to the adaptive threshold parameter for event detection. For example, based on the performance index... Linear interpolation is performed relative to the threshold to determine the adjustment direction and magnitude. Before finalizing the adjustment instruction, the current resource utilization of the edge node, such as CPU utilization or memory usage, is also obtained. If the resource utilization exceeds a safety threshold, such as exceeding 80%, adjustments involving increased computational load are restricted or postponed to prioritize system stability. Based on the generated adaptive adjustment strategy, feedback control instructions are generated. These instructions are encapsulated in a structured data format, such as JSON or Protobuf, which includes the target edge node identifier, the type of action to be performed (e.g., "switch model" or "adjust threshold"), and specific parameter values ​​related to the action (e.g., target model). The encapsulated feedback control signal is sent to the corresponding edge feature processing module via a message queue or dedicated communication channel to ensure reliable and asynchronous transmission of instructions. The specific process of forming a closed-loop optimization from the cloud to the edge is as follows: the generated feedback control instruction is sent to the edge feature processing module corresponding to the target edge node specified in the instruction. The edge feature processing module parses the received feedback control instruction and dynamically adjusts its running configuration according to the instruction content. Dynamic adjustment includes: if the instruction requires switching the feature extraction model, the specified new model is loaded from the preset model library and replaced with the currently running model to realize dynamic upgrading or downgrading of processing capabilities; if the instruction requires adjusting the adaptive threshold of event detection... If the parameter value is changed, the locally stored threshold variable is updated using the new parameter value contained in the instruction, thereby changing the sensitivity of event detection. After completing the configuration adjustment, the edge feature processing module sends a status confirmation signal indicating that the adjustment is complete to the feedback optimization module. The feedback optimization module receives the status confirmation signal and associates and stores the issued adjustment instruction, the corresponding target edge node identifier, and its comprehensive confidence performance index as a historical adjustment basis for subsequent analysis of the effectiveness of the adjustment strategy or optimization of threshold setting. After completing the configuration adjustment, the edge feature processing module uses the new configuration parameters to process the subsequent real-time multimodal stream data and uploads the generated new primary fusion data packet to the map construction and association module.The subsequent online prediction and optimization module, based on new data, generates prediction confidence assessment values, which are then fed back to the optimization module for further evaluation. This forms a continuous, cloud-based prediction performance feedback-driven edge data processing strategy optimization loop. This loop enables the system to reverse-engineer the front-end data processing stage based on the final marketing prediction results, achieving end-to-end adaptive optimization from decision-making outcomes to data collection and processing. This improves the system's overall robustness and marketing accuracy in dynamic environments.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A precision marketing intelligent application system based on AI training inference and data fusion, characterized in that, Specifically, it includes: The module consists of an edge feature processing module, a map construction and association module, an online prediction and optimization module, and a feedback optimization module, which are connected sequentially. Edge Feature Processing Module: Deployed at each edge node, this module responds to real-time multimodal streaming data received from connected data sources. It performs feature extraction and event detection on unstructured data within the real-time multimodal streaming data, generates potential event markers with timestamps, and aggregates structured behavioral data within a time window. It aligns and binds the potential event markers with the aggregated structured behavioral data along the time dimension, generating and outputting a primary fusion data package. Graph Construction and Association Module: Responds to multiple primary fusion data packages received from each edge feature processing module. Based on the entity, relationship, timestamp, and marker information in the primary fusion data packages, it dynamically constructs and updates a global temporal graph. The edge weights connecting nodes in the temporal graph are dynamically calculated using a time-varying attention weight calculation function. This function integrates node feature similarity, event occurrence time proximity, and event abruptness information, thus outputting a dynamic temporal graph that reflects real-time edge weights. Online Prediction and Tuning Module: Used for... In response to receiving the current state of the dynamic temporal graph, an online prediction unit performs inference based on the dynamic temporal graph to generate behavioral prediction probabilities for user nodes. An uncertainty quantification unit is then invoked to generate a confidence assessment value for each behavioral prediction probability. When the edge weight growth rate of the event topic node in the dynamic temporal graph exceeds a preset threshold and the prediction confidence assessment value of the related user node is lower than a set threshold, a meta-learning tuning unit is activated. This unit uses low-confidence user nodes and their local neighborhood structure in the dynamic temporal graph as input to perform small-sample rapid tuning of the parameters of the online prediction unit, ultimately outputting behavioral prediction results with corresponding confidence assessment values. The feedback optimization module, in response to receiving the behavioral prediction results and their confidence assessment values, sends out the confidence assessment values ​​for different users and the associated edge node identifiers as feedback signals. This triggers the corresponding edge feature processing module to dynamically adjust its internal feature extraction model selection and event detection threshold parameters based on the confidence assessment values, forming a closed-loop optimization from the cloud to the edge.

2. The precision marketing intelligent application system based on AI training inference and data fusion as described in claim 1, characterized in that: The edge feature processing module performs feature extraction and event detection on unstructured data in real-time multimodal stream data to generate potential event markers with timestamps. The specific process is as follows: The received real-time multimodal stream data is separated into a data portion containing unstructured data and a data portion containing structured behavioral data according to data format and structural attributes. For each data unit in the data portion containing unstructured data, a lightweight feature extraction model is used for calculation. The input data unit is encoded and converted according to the internal parameters of the feature extraction model to generate a dense feature vector of fixed dimension. A sliding time window is maintained, consisting of dense feature vectors generated at the current time and several consecutive times immediately preceding the current time. All dense feature vectors within the sliding time window, except for the dense feature vector at the current time, are obtained to form a historical feature vector set. Principal component analysis is performed on the historical feature vector set to obtain the corresponding principal component subspace. The dense feature vector at the current time is then mapped to the principal component subspace. The system performs a projection operation on the space to obtain a reconstructed feature vector. It then calculates the difference vector between the current dense feature vector and the reconstructed feature vector. An element-wise multiplication is performed between the difference vector and a predefined feature dimension importance mask vector. Since the weights for different dimensions in the feature dimension importance mask vector are not identical, a weighted difference vector is obtained after multiplication. The Euclidean norm of the weighted difference vector is calculated, and the calculated norm value is defined as the burst deviation at the current time. When the burst deviation exceeds an adaptive threshold, it is identified as the starting point of a potential event. Monitoring continues until the burst deviation falls below the adaptive threshold, and this interval is defined as a potential event segment. A globally unique identifier is generated for each potential event segment, and the start and end timestamps of the segment are recorded. All dense feature vectors within the potential event segment are aggregated to generate a representative feature vector representing the segment. Finally, the output includes a potential event marker containing a globally unique identifier, start and end timestamps, and a representative feature vector.

3. The precision marketing intelligent application system based on AI training inference and data fusion as described in claim 2, characterized in that: The process of aggregating structured behavioral data in real-time multimodal streaming data within a time window is as follows: the data portion containing structured behavioral data is divided according to a preset, fixed-duration alignment time window; within each alignment time window, behavioral records with the same user identifier in the structured behavioral data are aggregated by counting or summing according to their behavioral types to generate an aggregated behavioral vector, and the time corresponding to the midpoint of the alignment time window is recorded as the center timestamp; The aggregated behavior vector is output in pairs with the center timestamp, serving as the result of aggregating structured behavior data within a time window.

4. The precision marketing intelligent application system based on AI training inference and data fusion according to claim 3, characterized in that: The specific process of aligning and binding potential event markers with aggregated structured behavioral data in the time dimension to generate and output a primary fusion data package is as follows: For a generated potential event marker, it includes start and end timestamps; based on a preset time tolerance value, find all aggregated behavioral vectors in the time dimension whose center timestamps fall within the time interval formed by the time obtained by subtracting the time tolerance value from the start time of the start and end timestamps and adding the time tolerance value to the end time of the start and end timestamps; bind and associate the globally unique identifier, start and end timestamps, and representative feature vectors contained in the potential event marker with all aggregated behavioral vectors found in the time interval and their corresponding user identifiers and center timestamps, and package them to generate a primary fusion data package.

5. The precision marketing intelligent application system based on AI training inference and data fusion as described in claim 4, characterized in that: In the graph construction and association module, the edge weights connecting nodes in the temporal graph are dynamically calculated using a time-varying attention weight calculation function. The specific process is as follows: Parse the primary fusion data packet and extract the start and end timestamps corresponding to the globally unique identifier, the user entity corresponding to the user identifier, the central timestamp associated with the user behavior, the representative feature vector corresponding to the event entity, and the aggregated behavior vector corresponding to the user entity. Based on the extracted representative feature vector and aggregated behavior vector, calculate the dot product between the representative feature vector and the aggregated behavior vector, and perform scaling and nonlinear activation processing on the dot product result to obtain the first attention score representing the feature similarity between nodes. Obtain the central timestamp associated with user behavior in the primary fusion data packet, calculate the absolute time difference between the central timestamp and the current time, substitute the absolute time difference into an exponential decay function with a natural constant as the base and a negative decay coefficient as the exponent, and obtain the second decay factor characterizing the time decay effect. From the information provided by the edge feature processing module that is associated with the potential event markers in the primary fusion data packet, an event mutation intensity scalar value is obtained. The event mutation intensity scalar value is derived from the burst deviation calculated by the edge feature processing module during the generation of potential event markers. The event mutation intensity scalar value is first scaled, then nonlinearly compressed, and then the compression result is multiplied by a modulation coefficient. Finally, the product is added to a constant to obtain the third modulation factor characterizing the event mutation modulation effect. The first attention score, the second attenuation factor and the third modulation factor are multiplied continuously, and the final product is used as the edge weight between the user entity and the event topic entity at the current time.

6. The precision marketing intelligent application system based on AI training inference and data fusion as described in claim 5, characterized in that: The specific operation of outputting a dynamic temporal graph that reflects real-time edge weights is as follows: maintain a global temporal graph, which includes a set of nodes and a set of edges; for each received primary fusion data packet, perform the following operations to incrementally update the temporal graph: in response to the failure to find an event topic entity node in the node set of the temporal graph that corresponds to the globally unique identifier extracted from the primary fusion data packet, create an event topic entity node in the temporal graph with the globally unique identifier as its identifier, and store the representative feature vector and start and end timestamps extracted from the primary fusion data packet as attributes of the node; In response to the failure to find a user entity node in the node set of the temporal graph that corresponds to the user identifier extracted from the primary fusion data packet, a user entity node with the user identifier as the identifier is created in the temporal graph, and the aggregation behavior vector extracted from the primary fusion data packet is stored as an attribute of the node. A directed edge is established or located between the user entity node and the event topic entity node, and the center timestamp extracted from the primary fusion data packet is stored as the first association occurrence time attribute of the directed edge; if the directed edge already exists, the stored first association occurrence time attribute is retained unchanged. Based on the current moment, the time-varying attention weight calculation function is called to calculate a real-time edge weight, and this real-time edge weight is stored as the weight attribute of the directed edge. The system continuously processes new primary fusion data packets. For each new primary fusion data packet, it repeatedly performs the following operations: creating a node if no corresponding event subject entity node is found in the node set of the temporal graph; creating a node if no corresponding user entity node is found in the node set of the temporal graph; establishing or locating directed edges between user entity nodes and event subject entity nodes and storing the attribute of the first association time; and calling the time-varying attention weight calculation function to calculate and store the real-time edge weights. This incrementally updates the temporal graph and recalculates the edge weights in real time, maintaining a dynamic temporal graph where the edge weights evolve over time. The graph construction and association module generates and outputs a graph state snapshot reflecting all nodes, edges, and their real-time weights at a given time, based on the current state of the maintained dynamic temporal graph, at fixed time intervals or in response to external prediction requests.

7. The precision marketing intelligent application system based on AI training inference and data fusion according to claim 6, characterized in that: In the online prediction and optimization module, the specific operation of using an online prediction unit to infer based on the dynamic temporal graph to generate behavioral prediction probabilities for user nodes, and calling an uncertainty quantification unit to generate a confidence evaluation value for each behavioral prediction probability is as follows: The online prediction unit, based on the current state of the received dynamic temporal graph, infers for each user node in the dynamic temporal graph and outputs the behavioral prediction probability distribution for each user node; the uncertainty quantification unit calculates a confidence evaluation value for each user node. The calculation process is as follows: First, based on the edge connection relationship and real-time edge weight between the user node and other nodes in the dynamic temporal graph, all other user nodes connected to the user node in the dynamic temporal graph through the same event topic entity node are obtained. A set of associated user nodes is formed. Based on the real-time edge weights between the user node in the dynamic temporal graph, each node in the associated user node set, and the commonly connected event topic entity nodes, the behavior prediction probability distribution of each node in the associated user node set is weighted and aggregated to obtain the local neighborhood prediction consensus distribution of the user node. Next, the graph-aware prediction divergence between the user node's own behavior prediction probability distribution and the local neighborhood prediction consensus distribution is calculated, and this divergence value is amplified by a modulation factor related to the variance of the edge weights. Then, the information entropy of the user node's own behavior prediction probability distribution is calculated. Finally, the information entropy and the modulated graph-aware prediction divergence are weighted and summed, and the negative exponential function value of the summation result is used as the confidence evaluation value of the user node.

8. The precision marketing intelligent application system based on AI training inference and data fusion according to claim 7, characterized in that: The process of outputting behavior prediction results with corresponding confidence scores is as follows: For each event topic node in the dynamic temporal graph, calculate its edge weight growth rate within the most recent preset time window; specifically: calculate the weighted degree centrality of the event topic node at the calculation time, which is the sum of the real-time edge weights between the event topic node and all connected user nodes; calculate the weighted degree centrality of the event topic node at the time before the preset time window; subtract the weighted degree centrality of the previous time from the current weighted degree centrality, and then divide by the length of the preset time window to obtain the edge weight of the event topic node. The instantaneous change rate is used as the growth rate of its edge weights. For each event topic node in the dynamic temporal graph, a set of all user nodes directly connected to the event topic node through edges is obtained. The number of user nodes in the set whose confidence evaluation value is lower than a preset low confidence threshold is counted, and the proportion of this number to the total number of user nodes in the set is calculated as the low confidence user association ratio of the event topic node. When the edge weight growth rate of an event topic node exceeds a preset first growth rate threshold and the low confidence user association ratio of the event topic node exceeds a preset second ratio threshold, the meta-learning tuning unit is activated. The meta-learning tuning unit uses all user nodes associated with the event topic node corresponding to the triggering condition and whose confidence evaluation value is lower than the low confidence threshold, as well as their neighborhood subgraphs defined by a preset hop count range in the dynamic temporal graph, to form a support set for a few-shot learning task. The meta-learning tuning unit calculates the mean of the feature vectors of all user nodes in this support set after transformation by the intermediate layer of the online prediction unit to obtain the prototype feature representation of this few-shot learning task. The meta-learning tuning unit calculates the current gradient of the parameters of the online prediction unit based on this support set. Simultaneously, it retrieves the gradients of several historical tasks most similar to the current task's prototype feature representation from a memory storing historical task gradient information. This historical task gradient information includes the task prototype feature representations and their corresponding parameter gradients recorded during meta-learning tuning when activation conditions were previously met. The unit then calculates the average direction of these historical task gradients. The meta-learning tuning unit weightedly fuses this current gradient with the average direction of these historical task gradients to obtain an aligned fused gradient. This fused gradient is then used to perform a gradient update on the parameters of the online prediction unit. The process involves subtracting the product of the fusion gradient and a preset learning rate from the current parameter values ​​to obtain the optimized parameters. This gradient update is specifically performed on the parameters of the online prediction unit. Using the updated parameters, the online prediction unit re-performs the behavior prediction probability inference for all target user nodes, generating a new behavior prediction probability distribution for each user node. Simultaneously, using the uncertainty quantization unit, the confidence evaluation value for each user node is recalculated based on this new behavior prediction probability distribution. The output is a tuple of the new behavior prediction probability distribution and the recalculated confidence evaluation value for each user node, serving as the final behavior prediction result with the confidence evaluation value.

9. The precision marketing intelligent application system based on AI training inference and data fusion according to claim 8, characterized in that: In the feedback optimization module, the confidence assessment values ​​for different users and the associated edge node identifiers are sent as feedback signals to trigger the corresponding edge feature processing module to dynamically adjust its internal feature extraction model selection and event detection threshold parameters based on the confidence assessment values. The specific process is as follows: The feedback optimization module receives the tuple of the behavior prediction probability distribution and confidence assessment value for each user from the online prediction and tuning module, and obtains the edge node identifier associated with each user; based on the edge node identifier, the confidence assessment values ​​of all users belonging to the same edge node are aggregated to obtain the comprehensive confidence performance index for each edge node; the aggregation calculation uses a time decay-based approach. The weighted harmonic average method works as follows: For an edge node, a recent time window is set; for each processing moment within this time window, the following operations are performed: The number of users receiving data from the edge node at that processing moment is obtained and recorded; the reciprocal of the confidence assessment value for each user is calculated; a weight value is assigned to both the user count and the reciprocal of the confidence assessment value for each user; after completing the above operations for all processing moments within the time window, the weighted user counts for each processing moment are summed to obtain the numerator; the weighted reciprocals of the confidence assessment values ​​for each user at each processing moment are summed to obtain the denominator. Calculate the quotient obtained by dividing the numerator by the denominator, and use the quotient value as the comprehensive confidence performance index of the edge node; based on the comprehensive confidence performance index, combined with the preset high performance threshold and low performance threshold, generate an adaptive adjustment strategy for each edge node; The specific strategy generation process is as follows: the comprehensive confidence performance index of the edge node is compared with the high performance threshold and the low performance threshold; if the comprehensive confidence performance index is higher than or equal to the high performance threshold, a strategy to simplify or maintain the existing processing is generated for the edge node. If the overall confidence performance index is lower than or equal to the low performance threshold, a strategy to enhance the processing depth is generated for the edge node. This enhancement strategy includes instructing a switch to a feature extraction model with higher computational complexity, and calculating a negative adjustment to reduce the adaptive threshold parameter for event detection. If the overall confidence performance index is between the low performance threshold and the high performance threshold, a fine-tuning strategy to slightly adjust the adaptive threshold parameter for event detection is generated for the edge node. Before finalizing the adjustment instructions, the current resource utilization rate of the edge node is also obtained. If the resource utilization rate exceeds a safety threshold, adjustments involving increased computational load are restricted or postponed. Based on the generated adaptive adjustment strategy, a feedback control instruction is generated. The feedback control instruction is encapsulated in a structured data format, which includes the target edge node identifier, the type of action to be performed, and the specific parameter values ​​related to the action. The encapsulated feedback control signal is sent to the corresponding edge feature processing module through a message queue.

10. The precision marketing intelligent application system based on AI training inference and data fusion according to claim 9, characterized in that: The specific process of forming a closed-loop optimization from the cloud to the edge is as follows: The generated feedback control command is sent to the edge feature processing module corresponding to the target edge node specified in the command. The edge feature processing module parses the received feedback control command and dynamically adjusts its operating configuration according to the command content. Dynamic adjustment includes: if the command requires switching the feature extraction model, loading the specified new model from the preset model library and replacing the currently running model; if the command requires adjusting the adaptive threshold parameter of event detection, updating the locally stored threshold variable with the new parameter value contained in the command; after completing the configuration adjustment, the edge feature processing module sends an adjustment instruction to the feedback optimization module. The completed status confirmation signal is received by the feedback optimization module, which then associates and stores the issued adjustment instruction, the corresponding target edge node identifier, and its comprehensive confidence performance index as a basis for historical adjustments. After completing the configuration adjustment, the edge feature processing module uses the new configuration parameters to process the subsequent real-time multimodal stream data and uploads the generated new primary fusion data package to the graph construction and association module. The prediction confidence evaluation value generated by the subsequent online prediction and tuning module based on the new data will be received and evaluated by the feedback optimization module again, thus forming a continuous, cloud-based prediction performance feedback-based edge data processing strategy optimization closed loop.

Images