Intelligent marketing system and method based on customer behavior real-time calculation

Abstract

The invention relates to the technical field of user behavior analysis marketing, in particular to an intelligent marketing system and method based on customer behavior real-time calculation, and the system comprises an edge behavior collection unit, a cleaning and desensitization unit, a spatial-temporal feature coding unit, a real-time prediction unit and a collaborative decision-making unit. According to the method, an edge behavior acquisition unit captures native data such as page staying duration and contact tracks, a cleaning and desensitization unit generates compliance time sequence fragments, and a spatial-temporal feature coding unit converts the data into spatial-temporal feature vectors containing time decay weights, topological graph nodes and behavior integrity indexes; the real-time prediction unit fuses multi-dimensional vectors through a graph neural network and outputs real-time preference probability distribution of commodity categories, and the collaborative decision-making unit dynamically allocates computing power and generates an optimal marketing action set, so that the problems of large user real-time intention capture deviation and lagging computing power allocation of traditional marketing are solved, the marketing accuracy and timeliness are improved, and the marketing efficiency is improved. And resource waste is reduced.

Description

Technical Field

[0001] This invention relates to the field of user behavior analysis marketing technology, and more specifically, to an intelligent marketing system and method based on real-time calculation of customer behavior. Background Technology

[0002] User behavior analysis marketing is an important technology. In the digital economy era, accurately grasping user needs and dynamically adjusting marketing strategies are the core means to improve conversion rates and enhance user stickiness.

[0003] This technology captures users' behavioral data on digital platforms in real time, analyzes their preferences and intentions, and provides decision support for personalized marketing. It is of great significance for reducing marketing costs and improving resource utilization efficiency. Traditional marketing methods that rely on historical data or static models are no longer able to cope with dynamic scenarios where user behavior changes rapidly.

[0004] However, traditional intelligent marketing technologies suffer from core problems such as lagging user preference prediction and insufficient accuracy in marketing actions. Existing solutions only analyze browsing history and do not integrate multi-dimensional real-time features such as page dwell time and touchpoint trajectory. This leads to significant deviations in capturing users' instantaneous intentions. When users extend their page dwell time due to hesitation or interrupt transactions due to operational errors, the system cannot distinguish the changes in real needs, and the pushed marketing content is out of sync with users' real-time preferences. At the same time, the lack of a dynamic computing power allocation mechanism increases the model calculation delay when user behavior data surges, further reducing the timeliness of predictions and causing marketing actions to be delayed or misjudged. This not only affects user experience but also wastes marketing resources and fails to meet the needs of digital platforms for real-time and personalized marketing. To solve this problem, we provide an intelligent marketing system and method based on real-time calculation of customer behavior. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent marketing system and method based on real-time calculation of customer behavior, so as to solve the problems mentioned in the background art.

[0006] 1. Because traditional marketing does not integrate multi-dimensional real-time behavioral characteristics, the instantaneous intent of users is greatly deviated. Therefore, this case uses a spatiotemporal feature encoding unit to convert page dwell time and other data into spatiotemporal feature vectors. Combined with a graph neural network to output the preference probability distribution, it can accurately grasp the real-time needs of users and improve the matching degree of marketing content.

[0007] 2. Because traditional marketing lacks dynamic computing power allocation and suffers from computational delays when data surges, this case study uses a collaborative decision-making unit to dynamically allocate computing power based on data flow throughput and adopts a constrained multi-armed gambling machine to generate marketing action sets, which can ensure the timeliness of predictions and reduce the waste of marketing resources.

[0008] To achieve the above objectives, an intelligent marketing system based on real-time calculation of customer behavior is provided, including: The edge behavior acquisition unit captures and outputs the native behavior data stream generated by the user on the digital platform in real time through the agent module of the user terminal device. The data stream includes page dwell time, touch point trajectory coordinate sequence and transaction interruption event identifier. The cleaning and desensitization unit receives the original behavior data stream, performs real-time desensitization operation under the General Data Protection Regulation standard, generates a compliant behavior data stream by replacing the identity identifier and obfuscating sensitive fields, and cuts it into standardized behavior time sequence segments according to a preset time window. The spatiotemporal feature encoding unit receives the compliance behavior data stream, maps the page dwell time into a time decay weight, converts the touch point trajectory coordinate sequence into a topology graph node, and parses the transaction interruption event identifier into a behavior integrity indicator, and integrates them to generate a spatiotemporal feature vector with unified dimensions. The real-time prediction unit incorporates a lightweight graph neural network model, receives the spatiotemporal feature vector, and executes: A three-dimensional correlation matrix is ​​constructed by aggregating the current user's historical vector, real-time environment vector, and similar user group vector. Through graph convolutional layer operations accelerated by field-programmable gate array, the real-time probability distribution of user preferences for preset product categories is output. Based on the real-time preference probability distribution and the current load status of the computing core of the field-programmable gate array, the collaborative decision-making unit dynamically allocates computing resources according to the throughput of behavioral data streams, and uses the constrained multi-armed gambling machine algorithm to generate the optimal marketing action set.

[0009] The second objective of this invention is to provide a method for implementing an intelligent marketing system including the above-mentioned real-time calculation based on customer behavior, comprising the following steps: S1. The agent module of the user terminal device captures the native behavior data stream in real time, including page dwell time, touch point trajectory coordinate sequence and transaction interruption event identifier. The data stream is desensitized in real time using the GDPR standard. The compliant behavior data stream is generated by hash replacement of identity identifier and encryption and obfuscation of sensitive fields. It is then cut into standardized behavior time sequence segments according to a preset time window to ensure data privacy compliance and processing timeliness. S2. The page dwell time in the compliance behavior data flow is mapped to a time decay weight with a decision hesitation coefficient. The touch point trajectory coordinate sequence is transformed into topology graph nodes through gesture recognition technology. The transaction interruption event is parsed into multi-layer behavior integrity indicators by the payment risk control system. The input is processed by the multi-head attention fusion module for feature channel priority weighting and dynamic alignment, generating a spatiotemporal feature vector with unified dimensions. Its vector dimension maps the psychological cognitive level of the product category. S3. Aggregate the current user's historical vector, real-time environment vector, and similar user group vector. Construct a three-dimensional correlation matrix of user-time-environment through cross-modal tensor splicing technology. Utilize the hardware adaptive operator of the field-programmable gate array to decompose the spatial graph convolution and temporal causal convolution paths. Dynamically switch the computational precision based on the core junction temperature to output the real-time preference probability distribution of product categories. S4. Based on the standard deviation of the real-time preference probability distribution, the core temperature change rate of the field-programmable gate array, and the variance of historical returns, an environmental perception exploration coefficient is generated. By constraining the multi-armed gambling machine algorithm, a candidate set of marketing actions is generated within the product category similarity threshold. At the same time, the throughput of behavioral data stream is obtained according to the message queue backpressure mechanism. The optimal computing power allocation point is calculated in the Riemann manifold space by combining the interrupt event weight. Marketing actions are executed by dynamic frequency division through the clock gating register, and conversion data is collected to drive the incremental update of the graph neural network.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The edge behavior acquisition unit and the spatiotemporal feature encoding unit work together. The former comprehensively captures native data such as page dwell time, touch point trajectory coordinate sequence and transaction interruption events through the terminal agent module. The latter introduces a decision hesitation coefficient to map the dwell time into a time decay weight. With the help of gesture recognition, the touch point trajectory is transformed into topology graph nodes. The payment risk control system analyzes the transaction interruption event to generate multi-layer behavior integrity indicators. The feature channels are dynamically aligned by the multi-head attention fusion module to generate a spatiotemporal feature vector that maps the product cognition level. This fully explores the potential intentions in user behavior and solves the problem of preference misjudgment caused by traditional single data dimensions.

[0011] 2. The real-time prediction unit adopts a lightweight graph neural network. It integrates user history, real-time environment and similar user group vectors through cross-modal tensor splicing technology to construct a three-dimensional correlation matrix. It accelerates graph convolution operations with the help of field-programmable gate arrays and dynamically adjusts the calculation accuracy according to hardware resources. At the same time, it maintains the model's adaptability through incremental knowledge distillation and outputs an accurate probability distribution of product category preferences. This avoids the prediction bias caused by the data processing lag of traditional models and provides a reliable basis for marketing decisions.

[0012] 3. The collaborative decision-making unit dynamically allocates computing power based on behavioral data stream throughput and transaction interruption event weights. It achieves core frequency division of computing through clock gating registers to ensure processing efficiency when data surges. It adopts a constrained multi-armed gambling machine algorithm, combined with environmental perception exploration strategy and causal reasoning verification, to generate the optimal marketing action set that adapts to users' real-time preferences. This reduces the interference of excessive marketing on user experience, avoids the waste of marketing resources, improves conversion rate and user stickiness, and meets the personalized and real-time marketing needs of digital platforms. Attached Figure Description

[0013] Figure 1 This is an overall block diagram of the present invention; Figure 2 This is the overall flowchart of the present invention.

[0014] The meanings of the labels in the diagram are as follows: 1. Edge behavior acquisition unit; 2. Cleaning and desensitization unit; 3. Spatiotemporal feature encoding unit; 4. Real-time prediction unit; 5. Collaborative decision-making unit. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] This invention provides an intelligent marketing system based on real-time calculation of customer behavior. Please refer to [link / reference]. Figure 1 As shown, it includes: Edge behavior acquisition unit 1 captures and outputs the native behavior data stream generated by the user on the digital platform in real time through the agent module of the user terminal device. The data stream includes page dwell time, touch point trajectory coordinate sequence and transaction interruption event identifier. Cleaning and desensitization unit 2 receives the native behavior data stream, performs real-time desensitization operation under the General Data Protection Regulation standard, generates compliant behavior data stream by replacing identity identifiers and obfuscating sensitive fields, and cuts it into standardized behavior time sequence segments according to a preset time window. Spatiotemporal feature encoding unit 3 receives the compliance behavior data stream, maps the page dwell time into time decay weight, converts the touch point trajectory coordinate sequence into topology graph nodes, and parses the transaction interruption event identifier into behavior integrity indicators, and integrates them to generate a spatiotemporal feature vector with unified dimensions. To accurately uncover the potential intent in user behavior data, the spatiotemporal feature encoding unit 3 combines shopping psychology indicators with multi-dimensional feature extraction technology through a behavioral semantic parsing engine. When mapping page dwell time to time decay weight, it introduces the decision hesitation coefficient from shopping psychology indicators. When the dwell time exceeds a preset threshold, a negative exponential decay function is automatically triggered to generate a dynamic weight value. The page dwell time is processed through a behavioral semantic analysis engine, introducing a decision hesitation coefficient to reflect the user's degree of hesitation when choosing a product, generating a time decay weight. Based on the "non-linear relationship between dwell time and purchase intention" in shopping psychology, coefficient rules are set. When the user's dwell time on the product page is within a reasonable range (30 seconds to 2 minutes), the decision hesitation coefficient is 1.0, indicating normal browsing. If the dwell time is too short (<30 seconds), it is marked as accidental touch, and the coefficient drops to 0.5. If it is too long (>5 minutes), it is marked as hesitation due to indecision, and the coefficient rises to 1.5. The preset dwell time threshold is 5 minutes. When the actual dwell time exceeds this threshold, a dynamic weight value is automatically triggered. A negative exponential decay function generates dynamic weight values, meaning the longer the dwell time, the slower the weight decays. For example, the weight is 0.8 after 5 minutes and 0.75 after 6 minutes, avoiding distortion of feature weights due to excessive hesitation. For instance, if a user stays on a mobile page for 7 minutes with a decision hesitation coefficient of 1.5, after negative exponential decay calculation, the time decay weight is 0.65, distinguishing between effective and ineffective dwell times. The coefficient and decay function make the time features more closely match the true decision intent. Simultaneously, gesture trajectory recognition technology transforms the touch point trajectory coordinate sequence into topological graph nodes. A convolutional autoencoder extracts the intent feature vector from the finger swipe pattern and fuses it with a screen heatmap to generate a weighted adjacency matrix, dividing the screen into multiple functional areas and touch point trajectories. The areas traversed by the trace are marked as nodes in the topology graph. For example, "image area → price area → review area" corresponds to three nodes. The size of the node is weighted by the touch dwell time; the longer the dwell time, the larger the node. A convolutional autoencoder analyzes the finger swipe pattern to extract intent feature vectors. A screen heatmap records areas with high user click frequency. These are combined with the topology graph nodes to generate a weighted adjacency matrix. The connection weight between nodes is calculated based on the trajectory jump frequency and the heatmap click intensity. The jump weight from the price area to the shopping cart button is 0.8, higher than other paths. For example, if a user's finger slowly swipes from the product image area (node ​​A) to the review area (node ​​B) and then quickly clicks the shopping cart button (node ​​C), the connection weight between A and B in the generated adjacency matrix is ​​0.6 (due to the slow swipe). (Sliding indicates attention to detail), BC connection weight 0.9 (due to frequent clicks), parsing of transaction interruption event identifiers, additionally linking to the real-time interception code library of the payment risk control system, when continuous interruption events are detected, generating multi-layered nested behavioral integrity indicators, linking to the real-time interception code library of the payment risk control system, recording the interruption reason, parsing the transaction interruption event identifier, generating a primary indicator based on the interruption reason, such as "insufficient balance" recorded as 0.3 (can be remedied by recharge, high integrity), "risk control interception" recorded as 0.1 (high risk, low integrity), when continuous interruptions are detected (such as two interruptions within 3 minutes), generating a secondary indicator, if the two interruptions have the same reason (such as both being password errors), the secondary indicator is the product of the primary indicator (0.5 × 0.5 = 0.25, indicating a decrease in transaction integrity due to operational errors. If the reasons differ (e.g., insufficient balance initially, secondary risk control interception), the secondary indicator is a weighted sum of the primary indicators (0.3 × 0.6 + 0.1 × 0.4 = 0.22, comprehensively reflecting multiple issues). For example, if a user's transaction is interrupted for the first time due to "insufficient balance" (primary indicator 0.3), and then interrupted again 5 minutes later due to "payment limit" (primary indicator 0.2), the secondary indicator is 0.3 × 0.5 + 0.2 × 0.5 = 0.25, indicating moderate transaction integrity and potential for remediation, providing accurate input for subsequent real-time prediction of user preferences.

[0017] To integrate temporal, spatial, and behavioral stability features into a unified vector, the process of generating a spatiotemporal feature vector is achieved through a multi-head attention fusion module, which dynamically aligns and compresses multi-dimensional features. The specific implementation method is as follows: Using time decay weights as the time dimension scale, the adjacency matrix of topology graph nodes as the spatial relationship basis, and the behavior integrity index as the behavior stability correction factor, these are input into the multi-head attention fusion module. The three types of features obtained from the pre-processing are converted into a computable vector form according to preset rules and used as input to the multi-head attention fusion module. Using time decay weights as the scale, the vectors are arranged according to the chronological order of user behavior (e.g., browsing, comparison, order attempt) to form a time series vector. For example, a user's time vector might be [0.65, 0.72, 0.58], corresponding to the decay weights of the three behavior stages. The adjacency matrix of the topology graph nodes (inverse...) is then used as the basis for the spatial relationship correction factor. The region jump relationship of the touch point trajectory is converted into a spatial vector. The weight value of each element in the matrix is ​​directly used as a component of the vector (e.g., in the adjacency matrix, the weight of AB connection is 0.6 and the weight of BC connection is 0.9, corresponding to the spatial vector [0.6, 0.9, ...]). The multi-layer nested behavioral integrity indicators are expanded into vectors according to the level (e.g., the first-level indicator is 0.3 and the second-level indicator is 0.25, corresponding to the behavioral stability vector [0.3, 0.25]). Different types of features are unified into vector form, laying the foundation for subsequent similarity calculation and fusion, and ensuring that time, space and behavioral stability features can be calculated in the same dimensional space. The multi-head attention fusion module calculates the cross-entropy similarity matrix of the three types of input vectors and assigns priority weights to feature channels based on the similarity. The module determines the priority of each feature channel by calculating the cross-entropy similarity of the three types of input vectors. It calculates the cross-entropy between the time vector and the space vector, the time vector and the behavioral stability vector, and the space vector and the behavioral stability vector, respectively. The smaller the cross-entropy value, the stronger the correlation between the two types of features. For example, a cross-entropy of 0.2 between the time vector and the space vector indicates a high degree of matching between the user's browsing time and touchpoint trajectory. Weights are then assigned inversely based on the cross-entropy value (the smaller the cross-entropy, the higher the weight). For example, a time-space cross-entropy of 0.2 corresponds to a weight of 0.4, and a time-behavioral stability cross-entropy of 0.3 corresponds to a weight of 0.4. The corresponding weight is 0.3, and the weight corresponding to the spatial-behavioral stability cross-entropy is 0.5, ensuring that highly correlated feature channels receive higher priority. For example, the duration a user spends on a mobile page (temporal feature) is highly correlated with the touchpoint trajectory of frequently viewing the price area (spatial feature) (cross-entropy 0.15). Therefore, the priority weights of the temporal and spatial feature channels are set to 0.45 and 0.4 respectively, and the weight of the behavioral stability feature (due to only one interruption) is 0.15. Then, the gating recurrent unit realizes the dynamic compression and alignment of feature channels, and finally outputs a spatiotemporal feature vector with uniform dimensions. For input vectors of different lengths (such as a time vector with 3 components and a spatial vector with 5 components), the gating recurrent unit discards redundant information through a forget gate. Information (such as components with weights below 0.1 in the spatial vector), retaining core features (such as region jump relationships with weights above 0.6), and mapping the compressed time, space, and behavioral stability features to the same dimensional space (such as unifying them into a 10-dimensional vector), ensuring that each dimension corresponds to the psychological cognitive level of the product category (such as the 1st dimension corresponding to "basic needs perception", the 5th dimension corresponding to "brand preference", and the 10th dimension corresponding to "purchase decision intention"). For example, after compression, the time feature retains two core components: "hesitation stage" and "decision stage"; the spatial feature retains two core components: "price zone" and "evaluation zone"; and the behavioral stability feature retains one core component: "secondary interruption impact". The gating loop unit aligns them into a 10-dimensional vector, where the 8th... The value of the dimension (corresponding to "purchase hesitation") combines time decay weights and behavioral interruption indicators. The physical meaning of its vector dimension corresponds to the psychological cognitive hierarchy model of the product category. The final output spatiotemporal feature vector has a fixed dimension (e.g., 10 dimensions). The physical meaning of each dimension corresponds one-to-one with the psychological cognitive hierarchy model of the product category. Low dimensions (1-3 dimensions) reflect the user's basic cognition of the product (e.g., initial perception of function and price), medium dimensions (4-7 dimensions) reflect the user's comparison and evaluation process (e.g., comparison with similar products, brand trust), and high dimensions (8-10 dimensions) reflect the user's decision-making tendency (e.g., intensity of purchase intention, sensitivity to promotions). For example, in a user's spatiotemporal feature vector, the value of the second dimension (price perception) is 0.The value of 8 (indicating high price focus), 0.3 for the 6th dimension (brand trust) (indicating weak brand influence), and 0.6 for the 9th dimension (purchase intention) (indicating a clear purchase tendency) provide structured feature basis for subsequent prediction of product preferences and provide high-quality input features for the preference probability distribution output by real-time prediction unit 4.

[0018] Real-time prediction unit 4 has a built-in lightweight graph neural network model that receives spatiotemporal feature vectors and executes them: A three-dimensional correlation matrix is ​​constructed by aggregating the current user's historical vector, real-time environment vector, and similar user group vector. Through graph convolutional layer operations accelerated by field-programmable gate array, the real-time probability distribution of user preferences for preset product categories is output. To comprehensively integrate user historical behavior, real-time environment, and similar group characteristics, real-time prediction unit 4 employs cross-modal tensor stitching technology to construct a three-dimensional correlation matrix. This improves the accuracy of user preference prediction through orthogonal fusion of multi-dimensional features. The specific implementation method is as follows: Historical vectors extracted from the user behavior database contain behavioral features such as browsing, favorites, and purchases over the past 7 days. These are reconstructed into a time-series cube using a time-slicing algorithm. The historical vectors are divided into multiple segments according to time intervals (e.g., one slice per day). Each slice contains behavioral features within the corresponding time period (e.g., the distribution of browsing time on the first day, the categories of goods purchased on the second day, etc.). The cube is constructed with "behavior type - time slice - feature value" as a three-dimensional dimension. For example, in a user's time-series cube, the feature value of "Browsing behavior - Day 3 - Electronics category" is 0.7, indicating a high level of interest in browsing electronic products during this period, while the feature value of "Purchase behavior - Day 5 - Clothing category" is 0. 9 indicates that clothing purchases were completed during this period. This transforms linear historical behavioral data into a three-dimensional structure, preserving the temporal evolution of behavior while clearly showcasing the characteristic distribution of different behavior types. This provides a time reference for subsequent fusion with real-time features. The real-time environment vector is generated from three sets of sensor data—geographic location fencing, network latency gradient, and server load pulse—after noise reduction using a Kalman filter. Geographic location fencing data: the current location is obtained through the user terminal's positioning function and marked as one of the environmental features. Network latency gradient: records changes in network response speed when the user accesses the platform. Server load pulse: obtains the load fluctuations of the current platform server. The vector is used to analyze the anomalies in the three sets of data. The data is smoothed to preserve the true trend of environmental changes. The three sets of denoised data are integrated into a real-time environment vector in the order of "geographic location-network status-server load" (e.g., [0.8, 0.3, 0.6], representing high-activity business district, low network latency, and medium server load, respectively). Environmental factors can affect user behavior (e.g., users are more likely to make immediate purchases when in a business district). The denoising process ensures that the environment vector can truly reflect the user's scenario, providing a contextual basis for preference prediction. Similar user group vectors are filtered through a dynamic topological similarity network. This network calculates the Friesian distance of user behavior trajectories and the cosine similarity of behavioral intentions in real time, retaining only the topological structure. The system identifies compact user subgroups by calculating the Fraser distance between user behavior trajectories to measure the similarity between two trajectories. For example, if the browsing paths of the current user and user A are similar, the distance value is small. The system also calculates the cosine similarity of behavioral intentions to measure the matching degree of purchasing preferences. For example, if the current user and user B both frequently browse electronic products, the similarity value is high. A dual threshold of distance and similarity is set, and only users with a Fraser distance less than the threshold and a cosine similarity greater than the threshold are retained to form subgroups with compact topological structures, ensuring highly similar behavioral patterns. Common features of the subgroups are extracted, such as preferred product categories and average dwell time, and integrated into a similar user group vector (e.g., [0.7, 0.2, ...[9] indicates that the subgroup has a high preference for electronic products, a low preference for daily necessities, and a longer average dwell time. For example, if the current user frequently browses smartphones and spends a long time on them, the dynamic topology similarity network will select 100 users who also pay attention to smartphones and have similar browsing paths. The feature value of "electronic product category preference" in their group vector reaches 0.85, providing a group reference for predicting the current user's preference. The three types of vectors and the current spatiotemporal feature vector are concatenated into a three-dimensional correlation matrix according to the three orthogonal dimensions of user-time-environment to ensure that the scales of the four types of vectors are consistent in the three orthogonal dimensions. The "user dimension" includes the current user and similar user group features, the "time dimension" includes historical slices and real-time time points, and the "environment dimension" includes historical environment and current environment. Environmental features are used to fill the matrix with the feature values ​​of the four types of vectors according to their corresponding dimensions. For example, the position of "Current User - Day 3 - Business District Environment" in the matrix is ​​filled with the user's behavioral features on day 3 from the historical time series cube, the business district features from the real-time environment vector, the behavioral features of similar user groups in the business district environment, and the corresponding values ​​of the current spatiotemporal feature vector. This forms a matrix unit with multiple superimposed features. The final generated three-dimensional correlation matrix fully preserves the multi-dimensional correlation of the user's own history, real-time status, group characteristics, and environmental influences. This provides structured input for feature aggregation in graph neural networks, enabling the subsequent output of the product category preference probability distribution to simultaneously reflect the user's individual habits, real-time scenario, and group commonalities, thus improving the comprehensiveness and accuracy of prediction.

[0019] To improve the computational efficiency of graph convolutional layers and adapt to fluctuations in hardware resources, the graph convolutional layer computation accelerated by field-programmable gate arrays (FPGAs) achieves parallel processing of spatial and temporal features through hardware adaptive operators. Combined with a dynamic precision adjustment mechanism, it outputs an accurate probability distribution of product category preferences. The specific implementation method is as follows: The 3D correlation matrix is ​​decomposed into spatial graph convolutional paths and temporal convolutional paths to achieve targeted feature extraction. The spatial path uses Chebyshev multinomial approximation to aggregate neighborhood node features, extracting behavioral association features between users and similar groups (e.g., commonalities between users and subgroups in product category browsing). The "user-environment" dimension of the 3D matrix is ​​transformed into a spatial topological graph (with users as nodes and similarity as edge weights), focusing on the aggregation of behavioral features of different users in the same environment. The "user-time" dimension of the matrix is ​​extracted to show behavioral evolution features (e.g., temporal changes from browsing to purchasing). The 3D matrix is ​​expanded into a temporal sequence by time slices, focusing on the behavioral pattern changes of a single user at different time points. The convolutional pathway employs Chebyshev polynomial approximation to achieve efficient aggregation of neighborhood node features. It clusters nodes in the spatial topology graph based on similarity, grouping user nodes with highly similar preferences into one category, thus reducing computational complexity. By approximating complex graph convolution operations using Chebyshev polynomials, it quickly calculates the weighted sum of features of each node and its neighboring nodes. The feature value of a user node is the weighted average of its own features and the features of its three most similar users. The aggregation result reflects the impact of group behavior on individual preferences, showing how the high attention of a subgroup to electronic products affects the current user. For example, if the current user node's neighborhood contains three users with similar preferences, after polynomial approximation calculation, the weight of "electronic product category" in the aggregated features increases from 0.6 to 0.75. Reflecting the cumulative effect of group influence, the temporal pathway captures behavioral pattern evolution through causal dilated convolution. It employs a convolution kernel with a dilation coefficient of 2, skipping one intermediate time point to expand the temporal receptive field. Simultaneously, it covers time slices of the current, previous two, and previous four steps, capturing long-term behavioral patterns. The increased browsing frequency on Fridays ensures that the convolution operation relies solely on historical time point data. When calculating current time features, only behavioral data from the past hour is used to avoid interference from future information, aligning with the temporal logic of real behavior. Past browsing behavior influences current purchasing decisions. For example, if a user browses a clothing item at 10:00, 10:10, and 10:30, the causal dilated convolution, with a dilation coefficient of 2, captures the correlation between these three time points. The "clothing category preference" in the output temporal features gradually increases over time. The outputs of the two pathways are input to a gating fusion unit via a high-speed interconnect bus on a field-programmable gate array (FPGA) chip for feature integration and accuracy adaptation. The gating fusion unit merges the outputs of the two pathways through weight allocation. The system integrates feature vectors, with each vector dimension corresponding to a preset product category. A 30-dimensional vector corresponds to 30 product categories. The system monitors the hardware status of the field-programmable gate array (FPGA) in real time. When resource utilization is below 70%, high-precision computation is used, retaining four decimal places of the feature vector. When utilization exceeds 90% or junction temperature exceeds a critical value of 70°C, it automatically switches to low-bit quantization, retaining only two decimal places or the integer part. This reduces computational load while maintaining basic accuracy. The fused feature vector is then normalized to ensure the sum of feature values ​​for each product category is 1, transforming it into a product category preference probability distribution: 0.35 for "electronic products," 0.25 for "clothing," and 0.4 for "daily necessities." This directly reflects the user's real-time preference for each category. For example, when the FPGA junction temperature rises to 75°C, exceeding the 70°C critical value, the gating fusion unit switches to low-bit quantization mode, simplifying the integrated feature vector to generate a preference probability distribution. Although accuracy is slightly reduced, this ensures continuous operational stability and avoids prediction interruptions due to hardware overheating.

[0020] To continuously optimize prediction accuracy while maintaining a lightweight graph neural network model, an incremental knowledge distillation architecture is adopted. Through the collaborative updates of the cloud-based teacher model and the embedded student model, efficient knowledge transfer and scenario adaptation of the model are achieved. The specific implementation method is as follows: A two-layer model architecture of "cloud teacher - terminal student" is constructed, clearly defining the functional boundaries of both. The teacher model is deployed on a cloud server, receiving full user behavior data, including historical behavior, real-time features, and similar group features for continuous training. For example, it receives browsing records, transaction data, and environmental parameters of all users on the platform daily, learning global behavioral patterns through a complete graph neural network structure, including differences in category preferences and seasonal consumption trends among users in different regions. The student model embeds a real-time prediction unit 4, optimized based on the hardware characteristics of a field-programmable gate array (FPGA). A dedicated instruction set is used to sparsify the model's convolutional kernels, retaining core computational parameters and removing redundant parameters, such as reducing 1000 convolutional kernels to 300 key kernels. This ensures a small model size, fast computation speed, and adaptability to the computing power limitations of terminal devices. The teacher model focuses on learning comprehensive, long-term behavioral patterns, while the student model focuses on real-time response and local prediction. Both achieve the transformation of "global knowledge into local capabilities" through knowledge distillation. The student model is triggered at fixed time intervals to extract knowledge from the teacher model, ensuring timely updates without affecting real-time predictions. The window duration is set based on the frequency of changes in user behavior characteristics. During peak e-commerce promotions, the window is set to 30 minutes, while during normal times, it is set to 2 hours, balancing update timeliness with system load. Each time the trigger occurs, the teacher model generates an attention distribution heatmap for currently popular product categories, visually demonstrating the model's focus on different features. For example, the attention weight for "price feature" in the electronics category is 0.7, and for "review feature" it is 0.3. This serves as a "soft label" passed to the student model, representing implicit knowledge signals. For instance, the heatmap generated by the teacher model at 10:00 shows that users' attention to the "delivery timeliness" feature (weight 0.6) in the fresh produce category. The student model will adjust its focus on this feature accordingly. By learning the teacher model's soft labels, the student model achieves matching of the hidden features of both models. The steps are as follows: The student model and the teacher model each predict the same batch of samples—user data with recent transaction behavior—and output their respective hidden feature vectors, reflecting the model's internal feature processing results. The KL divergence loss function is used to measure the difference between the two sets of hidden feature vectors. The smaller the difference, the closer the student model is to the teacher model's cognitive pattern. Based on this, the parameters of the student model are adjusted, increasing the weight of the "delivery timeliness" feature. This comparison and adjustment process is repeated until the difference between the student model's hidden features and the teacher model's is reduced to a preset range, with a difference value below 0.1, ensuring that the student model learns the teacher model's judgment logic. For example, initially, the student model's weight for the "delivery timeliness" feature in the fresh food category is 0.3. After KL divergence calculation and parameter adjustment with the teacher model (weight 0.6), it is gradually increased to 0.55, achieving feature alignment. During distillation, the weight parameters related to the real-time environment vector in the student model are frozen to avoid global knowledge coverage of scene adaptability. The model parameters corresponding to the real-time environment vector are clearly defined, and their values ​​are fixed during distillation, without participating in iterative adjustments. Since the student model is deployed on the terminal, it needs to respond quickly to changes in the local environment, such as the behavior switch when a user moves from a residential area to a business district. Freezing the environment-related weights can preserve its scene adaptability learned from local data and its high sensitivity to immediate consumption categories in the business district environment. For example, the student model learns from local data that "when a user is in a business district, the prediction weight of the fast food category needs to be increased." Freezing these environment-related weights during distillation avoids being covered by the global rules of the teacher model, ensuring that the real-time prediction unit 4 can efficiently and accurately output the probability distribution of product category preferences on the terminal device, providing reliable support for marketing decisions.

[0021] Based on the real-time preference probability distribution and the current load status of the computing core of the field-programmable gate array, the collaborative decision-making unit 5 dynamically allocates computing resources according to the throughput of behavioral data streams, and uses the constrained multi-armed gambling machine algorithm to generate the optimal marketing action set.

[0022] To achieve efficient utilization and real-time response of computing resources, the collaborative decision-making unit 5 adopts an event-driven load balancing strategy for dynamic computing resource allocation. This strategy dynamically adjusts computing resources based on data flow characteristics and hardware status. The specific implementation method is as follows: By collecting three types of key parameters in real time, a three-dimensional resource allocation vector reflecting computing power demand and hardware status is constructed. Relying on the message queue backpressure mechanism of the streaming cleaning and desensitization unit 2, when the amount of data waiting to be processed in the queue exceeds a preset value, a pressure signal is automatically fed back. The throughput of the behavioral data stream is obtained in real time. For example, if the backpressure signal shows that the amount of user behavioral data accumulated in the queue reaches 1,000 pieces per second, the corresponding throughput parameter value is 0.8, and the full load is 1.0. The weight is set according to the severity of the transaction interruption event. The weight of a single ordinary interruption (such as insufficient balance) is 0.3, and the weight of three consecutive high-risk interruptions (such as risk control interception) is 0.9. The higher the weight, the higher the priority. To mitigate the potential for conversion, hardware sensors are used to collect the core voltage fluctuation amplitude, converting it into a volatility parameter to reflect the stability of hardware operation. These three parameters are then combined in the order of "throughput - interrupt weight - voltage volatility" to form a three-dimensional resource allocation vector. Each dimension of the vector quantifies the key factors affecting computing power allocation. Based on this three-dimensional resource allocation vector, the optimal resource allocation point is located in the Riemannian manifold space, achieving precise matching between computing power and demand. Mapping the three-dimensional vector to the Riemannian manifold space—a curved geometric space—better aligns with the nonlinear resource allocation rules. Each dimension corresponds to a coordinate axis in this space; for example, throughput corresponds to "data processing demand." The algorithm identifies the "axis" (interrupt weights corresponding to the "priority axis") and the "hardware constraint axis" (voltage fluctuation rate corresponding to the "hardware constraint axis"). Coordinate points are determined in space based on vector values. The location of this point directly reflects the current resource demand. A preset spatial distance algorithm finds the resource allocation scheme closest to this point. Riemannian manifolds better handle nonlinear resource allocation relationships, avoiding allocation deviations in complex scenarios compared to traditional linear models. This ensures that computing power allocation meets data processing needs without exceeding hardware capacity. By modifying the clock gating register of the field-programmable gate array (FPGA), dynamic frequency division of the computing core is achieved, and resource preemption is executed when a transaction interruption event is triggered. The clock gating register... The controller manages the operating frequency of the computing cores, adjusting the frequency based on the optimal resource allocation. For example, when allocating 60% of the cores, the frequency of the corresponding cores is reduced from 1GHz to 0.8GHz, while the remaining cores remain dormant. If 80% of the cores are allocated, the frequency is increased to 1.2GHz to accelerate processing. When a transaction interruption event is detected, a preemption mechanism is immediately triggered, suspending the computing cores of low-priority tasks, adjusting their clock frequencies to the highest, and allocating them to real-time processing tasks related to the interruption event, ensuring a response to the interruption event within 100 milliseconds. For instance, when a 3D vector shows a high interrupt weight, the system immediately preempts 30% of the dormant cores, adjusting their frequencies to 1GHz.At 2GHz, priority is given to feature extraction and marketing action generation for this interruption event, preventing users from abandoning transactions due to interruption. The event-driven load balancing strategy achieves precise allocation of computing power through three-dimensional vectors and Riemannian manifold space, and ensures real-time response to critical events through dynamic frequency division and resource preemption, maximizing marketing conversion opportunities while ensuring efficient system operation.

[0023] To ensure that the constrained multi-armed gambling machine algorithm accurately matches real-time user needs with the system's operational status while exploring potential marketing opportunities, it employs an environment-aware exploration strategy. By dynamically adjusting the exploration intensity and scope, it generates a high-quality set of marketing actions. The specific implementation method is as follows: A dynamic adjustment function for the exploration coefficient, consisting of three parameters, is defined to reflect the exploration demand and system capacity in real time. The standard deviation of the real-time preference probability distribution is extracted from the product category preference probability distribution output by the real-time prediction unit 4. A larger standard deviation indicates a smaller difference in the probability of a user's preference for electronic products versus clothing, suggesting high uncertainty in user demand and requiring stronger exploration. The core temperature change rate of the field-programmable gate array (FPGA) is obtained through hardware sensors. A 5°C increase in temperature within 10 seconds indicates excessive system load, requiring a reduction in exploration to decrease computational pressure, resulting in a decrease in the corresponding parameter value. The historical return variance of marketing actions is calculated by statistically analyzing the conversion effect fluctuations of similar marketing actions over the past hour. The conversion rate of pushed coupons fluctuates, and a larger variance indicates low historical experience reference value, requiring stronger exploration, resulting in an increase in the corresponding parameter value. The three parameters are combined into an exploration coefficient according to preset weights. A higher coefficient indicates greater exploration intensity. For example, if a user has a large preference standard deviation (0.6), a low temperature change rate (0.2), and a large historical return variance (0.7), the calculated exploration coefficient is 0.6 × 0.4 + 0.2 × 0.2 + 0.7 × 0.4 = 0.58, indicating a need for moderate-intensity exploration. When a user is detected in a high-value conversion scenario, such as adding items to their cart but not checking out, or browsing high-priced items for more than 5 minutes, the exploration scope is narrowed using the backpropagation gradient of the graph neural network. Based on the user behavior integrity index and product value, high-value scenarios are determined and marked as "requiring precise exploration." The backpropagation mechanism of the graph neural network is invoked to calculate the correlation gradient between different product categories and the user's current preferences (the larger the gradient, the more relevant the category is to the current need). Only product categories with gradient values ​​exceeding a threshold are retained, such as "phone accessories" and "screen protectors," which are highly correlated with "smartphone" preferences. A product category similarity threshold is set, allowing exploration of marketing actions only within the range where the similarity to the user's current preferred category exceeds this value. When a user is interested in smartphones, only relevant categories such as "phone cases" and "chargers" are explored, excluding irrelevant categories such as "clothing" and "food." For example, if a user browses high-end cameras for a long time, which is a high-value scenario, the graph neural network calculation shows that the similarity between "camera lenses" and "memory cards" and the camera category reaches 0.85, exceeding the threshold of 0.7. The exploration space is constrained to marketing actions for these two related product categories. The generated marketing action candidate set needs to be verified by a causal inference engine to ensure a clear causal relationship with the integrity of user behavior. Within the constrained exploration space, specific marketing actions are generated, such as pushing "camera lens coupons" and "buy one get one free memory cards." Each action includes triggering conditions, namely, user dwell time exceeding 3 minutes and expected effect. The causal inference engine analyzes the logical relationship between actions and user behavior, such as whether "pushing lens coupons" can directly reduce the probability of user transaction interruption and improve behavior integrity. Actions without clear causality are excluded, such as pushing food coupons to users browsing the camera. Verified marketing actions are retained and sorted by exploration coefficient and expected return, such as prioritizing high-return actions, forming an optimal marketing action set. This set is simultaneously sent to the user's terminal, and the action execution time and parameters are recorded. For example, "50 yuan off camera lenses" in the candidate set is verified to directly prompt users to complete the checkout, while "follow the store to get points" is not directly related to the current behavior integrity and is excluded. The final output is an action set containing instant discounts, ultimately improving marketing conversion efficiency and reducing user interference.

[0024] To address scenarios involving excessively volatile marketing campaign returns or hardware malfunctions, the environmental awareness exploration strategy employs a dynamic confidence interval correction mechanism and a degradation decision mechanism to ensure the reliability of marketing campaigns and system stability. The specific implementation methods are as follows: When the historical return variance of a marketing campaign exceeds a preset threshold, or when the conversion rate fluctuation of similar coupon pushes exceeds a preset range within the past hour, the system automatically switches to a Bayesian hierarchical model to calculate the return confidence interval. The steps are as follows: The system analyzes historical conversion data for various marketing activities in real time and calculates the return variance. If the conversion rate of a certain type of discount activity fluctuates drastically between 10% and 30%, and the variance exceeds a preset threshold, confidence interval correction is triggered. The model contains two layers of parameters: the first layer represents the individual return characteristics of a specific marketing activity, such as the actual conversion effect of "spend 200, get 50 off"; the second layer represents the overall distribution characteristics of similar activities, such as the average conversion rate of all discount activities. Data fusion is achieved through hierarchical correlation. This sampling method extracts a large number of possible return values ​​from the model, forming an expected return probability distribution band. If the return distribution band for a certain activity is "15%-25%", it indicates a high probability of success. If the probability of obtaining a return within this range is high, and the standard deviation of the user's real-time preference probability distribution is large, the distribution band is widened from 15%-25% to 10%-30% to retain more exploration space. If the standard deviation is small, the distribution band is narrowed to focus on high-deterministic actions. For example, if the historical return variance of a skincare product trial pack push exceeds the standard, an initial distribution band of "8%-18%" is generated through model sampling. Because the standard deviation of user preferences for skincare products is large, the bandwidth is expanded to "5%-22%" to ensure that no potentially effective actions are missed. When abnormal fluctuations in the core temperature of the field-programmable gate array are detected, a degraded decision topology network is used to replace the original constrained multi-armed gambling machine algorithm. The steps are as follows: The core temperature is monitored in real time by hardware sensors. When the temperature fluctuation exceeds the preset range, a single fluctuation exceeding 10°C or a sustained temperature above the critical value of 70°C is considered an abnormal state, triggering a degradation mechanism. This mechanism performs a tensor product operation on the product category similarity threshold and the behavioral integrity index, essentially taking the cross-combination of these two parameters to filter out marketing actions that simultaneously meet the criteria. This significantly reduces the action space from 100 candidate actions to 20. Within this simplified space, only actions with a confidence level exceeding the preset confidence threshold in historical conversion scenarios are retained—actions with a stable conversion rate above 80% in similar past scenarios, such as "pushing instant discount coupons to users who have added items to their cart but haven't paid." This ensures high reliability of the actions. For example, when the temperature is abnormal, the action space with "similarity 0.6 + integrity 0.7" is filtered out through tensor product operation, and then three core actions with a historical confidence level exceeding 0.8, such as "discount coupons," "gifts," and "limited-time discounts," are retained. This reduces computational load. The marketing action set generated by the degradation decision must be verified by a causal reasoning engine. To ensure compatibility with the user cognitive load model and avoid interfering with user experience, the model incorporates the user's current operational complexity, such as the number of products viewed simultaneously and the frequency of interface interactions, such as clicks per minute. If a marketing action, such as a pop-up ad, causes the cognitive load to exceed a threshold, increasing the operational complexity from 3 to 5, it is considered incompatible. Actions that pass the compatibility verification are retained and pushed out in order of historical conversion performance, such as prioritizing "discount coupons." At the same time, the computational resource consumption of each action is reduced, the rendering complexity of the push copy is simplified, and the hardware burden is reduced. For example, among the three core actions generated after downgrading, "pop-up ads" are excluded because they increase the user's cognitive load. Ultimately, the two low-interference actions, "silently issuing discount coupons" and "gift reminders," are executed. Through this dynamic confidence interval correction and downgrading decision-making mechanism, the environmentally aware exploration strategy can maintain the flexibility of exploration when the return fluctuates greatly, and ensure the stable operation of the system by simplifying decisions when the hardware is abnormal, thus balancing marketing effectiveness and system security.

[0025] In this invention, the edge behavior acquisition unit 1 captures raw data such as page dwell time and touch point trajectory; the cleaning and desensitization unit 2 generates compliant time sequence segments; the spatiotemporal feature encoding unit 3 transforms the data into a spatiotemporal feature vector containing time decay weights, topology graph nodes, and behavior integrity indicators; the real-time prediction unit 4 fuses multi-dimensional vectors through a graph neural network to output the real-time preference probability distribution of product categories; and the collaborative decision-making unit 5 dynamically allocates computing power to generate the optimal marketing action set. This solves the problems of large deviations in capturing users' real-time intentions and lagging computing power allocation in traditional marketing, improves marketing accuracy and timeliness, and reduces resource waste.

[0026] The second objective of this invention is to provide a method for implementing an intelligent marketing system based on real-time calculation of customer behavior, including any of the above-mentioned features, comprising the following steps: S1. The agent module of the user terminal device captures the native behavior data stream in real time, including page dwell time, touch point trajectory coordinate sequence and transaction interruption event identifier. The data stream is desensitized in real time using the GDPR standard. The compliant behavior data stream is generated by hash replacement of identity identifier and encryption and obfuscation of sensitive fields. It is then cut into standardized behavior time sequence segments according to a preset time window to ensure data privacy compliance and processing timeliness. S2. The page dwell time in the compliance behavior data flow is mapped to a time decay weight with a decision hesitation coefficient. The touch point trajectory coordinate sequence is transformed into topology graph nodes through gesture recognition technology. The transaction interruption event is parsed into multi-layer behavior integrity indicators by the payment risk control system. The input is processed by the multi-head attention fusion module for feature channel priority weighting and dynamic alignment, generating a spatiotemporal feature vector with unified dimensions. Its vector dimension maps the psychological cognitive level of the product category. S3. Aggregate the current user's historical vector, real-time environment vector, and similar user group vector. Construct a three-dimensional correlation matrix of user-time-environment through cross-modal tensor splicing technology. Utilize the hardware adaptive operator of the field-programmable gate array to decompose the spatial graph convolution and temporal causal convolution paths. Dynamically switch the computational precision based on the core junction temperature to output the real-time preference probability distribution of product categories. S4. Based on the standard deviation of the real-time preference probability distribution, the core temperature change rate of the field-programmable gate array, and the variance of historical returns, an environmental perception exploration coefficient is generated. By constraining the multi-armed gambling machine algorithm, a candidate set of marketing actions is generated within the product category similarity threshold. At the same time, the throughput of behavioral data stream is obtained according to the message queue backpressure mechanism. The optimal computing power allocation point is calculated in the Riemann manifold space by combining the interrupt event weight. Marketing actions are executed by dynamic frequency division through the clock gating register, and conversion data is collected to drive the incremental update of the graph neural network.

[0027] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent marketing system based on real-time calculation of customer behavior, characterized in that: include: The edge behavior acquisition unit (1) captures and outputs the native behavior data stream generated by the user on the digital platform in real time through the agent module of the user terminal device. The data stream includes the page dwell time, the touch point trajectory coordinate sequence and the transaction interruption event identifier. The cleaning and desensitization unit (2) receives the original behavior data stream, performs real-time desensitization operation under the General Data Protection Regulation standard, generates compliant behavior data stream by replacing the identity identifier and obfuscating sensitive fields, and cuts it into standardized behavior time sequence segments according to a preset time window. The spatiotemporal feature encoding unit (3) receives the compliance behavior data stream, maps the page dwell time to time decay weight, converts the touch point trajectory coordinate sequence into topology graph nodes, and parses the transaction interruption event identifier into behavior integrity indicators, and integrates them to generate a spatiotemporal feature vector with unified dimensions. The real-time prediction unit (4) has a built-in lightweight graph neural network model, which receives the spatiotemporal feature vector and executes: A three-dimensional correlation matrix is ​​constructed by aggregating the current user's historical vector, real-time environment vector, and similar user group vector. Through graph convolutional layer operations accelerated by field-programmable gate array, the real-time probability distribution of user preferences for preset product categories is output. The collaborative decision-making unit (5) dynamically allocates computing resources based on the real-time preference probability distribution and the current load status of the computing core of the field programmable gate array, and generates the optimal marketing action set by using the constrained multi-armed gambling machine algorithm.

2. The intelligent marketing system based on real-time customer behavior calculation according to claim 1, characterized in that, When the spatiotemporal feature encoding unit (3) maps the page dwell time to time decay weight through the behavioral semantic parsing engine, it introduces the decision hesitation coefficient in the shopping psychology index. When the dwell time exceeds a preset threshold, a negative exponential decay function is automatically triggered to generate dynamic weight values. At the same time, gesture trajectory recognition technology is used to transform the touch point trajectory coordinate sequence into topological graph nodes. The intention feature vector in the finger swipe pattern is extracted by a convolutional autoencoder and fused with the screen heat map to generate a weighted adjacency matrix. The parsing of the transaction interruption event identifier is additionally linked to the real-time interception code library of the payment risk control system. When continuous interruption events are detected, a multi-layered nested behavior integrity index is generated.

3. The intelligent marketing system based on real-time customer behavior calculation according to claim 2, characterized in that, The process of fusing and generating spatiotemporal feature vectors with unified dimensions is as follows: The time decay weight is used as the time dimension scale, the adjacency matrix of the topology graph nodes is used as the spatial relationship basis, and the behavior integrity index is used as the behavior stability correction factor, which are then input into the multi-head attention fusion module. The multi-head attention fusion module calculates the cross-entropy similarity matrix of the three types of input vectors, assigns priority weights to feature channels based on similarity, and then achieves dynamic compression and alignment of feature channels through a gated loop unit, finally outputting a spatiotemporal feature vector with unified dimensions. The physical meaning of its vector dimension corresponds to the psychological cognitive hierarchy model of the product category.

4. The intelligent marketing system based on real-time calculation of customer behavior according to claim 3, characterized in that: The real-time prediction unit (4) uses cross-modal tensor stitching technology when constructing the three-dimensional correlation matrix: Historical vectors extracted from the user behavior database are reconstructed into time-series cubes using a time-slicing algorithm. Real-time environment vectors are generated from three sets of sensor data—geographic location fences, network latency gradients, and server load pulses—after noise reduction using a Kalman filter. Similar user group vectors are filtered through a dynamic topological similarity network, which calculates the Friesian distance between user behavior trajectories and the cosine similarity of behavioral intentions in real time, retaining only the features of user subgroups with compact topological structures. The three types of vectors are concatenated with the current spatiotemporal feature vectors along three orthogonal dimensions—user, time, and environment—to form a three-dimensional correlation matrix.

5. The intelligent marketing system based on real-time customer behavior calculation according to claim 4, characterized in that: The graph convolutional layer operation accelerated by the field-programmable gate array includes hardware adaptive operators: The three-dimensional correlation matrix is ​​decomposed into a spatial graph convolution path and a temporal convolution path. The spatial path uses Chebyshev multinomial approximation to aggregate neighborhood node features, while the temporal path captures behavioral pattern evolution through causal dilation convolution. The outputs of the two paths are input to the gating fusion unit via a high-speed interconnect bus on the field-programmable gate array (FPGA). This unit dynamically adjusts the calculation accuracy based on the current hardware resource utilization rate. When the junction temperature of the FPGA exceeds the critical value, it automatically switches to low-bit quantization operation mode, and finally outputs the probability distribution of product category preferences.

6. The intelligent marketing system based on real-time calculation of customer behavior according to claim 5, characterized in that: The graph neural network model employs an incremental knowledge distillation architecture: The teacher model is deployed in the cloud and receives full feature updates. The student model in the embedded real-time prediction unit (4) implements convolution kernel sparsity through the field programmable gate array dedicated instruction set. Every time a preset time window is set, the student model extracts the attention distribution heatmap from the teacher model as a soft label. The hidden layer features are aligned through the KL divergence loss function. During the distillation process, the weights related to the real-time environment vector are frozen to maintain scene adaptability.

7. The intelligent marketing system based on real-time calculation of customer behavior according to claim 1, characterized in that: The collaborative decision-making unit (5) implements an event-driven load balancing strategy for dynamic computing power allocation: The throughput of behavioral data stream is obtained in real time through the message queue backpressure mechanism of the streaming cleaning and desensitization unit (2). It is combined with the weight of the transaction interruption event and the core voltage fluctuation rate of the field programmable gate array to form a three-dimensional resource allocation vector. Based on this vector, the optimal resource allocation point is calculated in the Riemann manifold space. The dynamic frequency division of the computing core is realized by modifying the clock gate register of the field programmable gate array. When the transaction interruption event is triggered, the computing resources are immediately preempted.

8. The intelligent marketing system based on real-time calculation of customer behavior according to claim 7, characterized in that: The constrained multi-armed gambling machine algorithm employs an environment-aware exploration strategy: A dynamic adjustment function for the exploration coefficient is defined, consisting of the standard deviation of the real-time preference probability distribution, the temperature change rate of the field-programmable gate array core, and the variance of the historical returns of marketing actions. When a user is detected to be in a high-value conversion scenario, the exploration action space is constrained by the backpropagation gradient of the graph neural network. Exploration is only allowed in marketing actions where the similarity between product categories exceeds a threshold. The generated set of marketing actions is output after the causal relationship between the generated marketing action set and the integrity of the user behavior is verified by the causal inference engine.

9. The intelligent marketing system based on real-time calculation of customer behavior according to claim 8, characterized in that: The environmental perception exploration strategy further includes a dynamic confidence interval correction mechanism: When the historical return variance of a marketing action exceeds a preset threshold, it automatically switches to the return confidence interval calculation mode based on a Bayesian hierarchical model. It generates the expected return probability distribution band of the marketing action through Markov chain Monte Carlo sampling and dynamically adjusts the distribution bandwidth according to the standard deviation of the real-time preference probability distribution. During periods of abnormal temperature fluctuations in the core of the field-programmable gate array (FPGA), a degraded decision topology network is used to replace the original constrained multi-armed gambling machine algorithm. This network performs tensor product operations on the product category similarity threshold and the behavior integrity index to generate a simplified action space. Only core marketing actions with confidence levels exceeding the preset confidence threshold in historical conversion scenarios are retained. The output set of marketing actions is executed after the causal inference engine verifies its compatibility with the user cognitive load model.

10. A method for implementing an intelligent marketing system comprising any one of claims 1-9 based on real-time calculation of customer behavior, characterized in that: Includes the following steps: S1. The agent module of the user terminal device captures the native behavior data stream in real time, including page dwell time, touch point trajectory coordinate sequence and transaction interruption event identifier. The data stream is desensitized in real time using the GDPR standard. The compliant behavior data stream is generated by hash replacement of identity identifier and encryption and obfuscation of sensitive fields. It is then cut into standardized behavior time sequence segments according to a preset time window to ensure data privacy compliance and processing timeliness. S2. The page dwell time in the compliance behavior data flow is mapped to a time decay weight with a decision hesitation coefficient. The touch point trajectory coordinate sequence is transformed into topology graph nodes through gesture recognition technology. The transaction interruption event is parsed into multi-layer behavior integrity indicators by the payment risk control system. The input is processed by the multi-head attention fusion module for feature channel priority weighting and dynamic alignment, generating a spatiotemporal feature vector with unified dimensions. Its vector dimension maps the psychological cognitive level of the product category. S3. Aggregate the current user's historical vector, real-time environment vector, and similar user group vector. Construct a three-dimensional correlation matrix of user-time-environment through cross-modal tensor splicing technology. Utilize the hardware adaptive operator of the field-programmable gate array to decompose the spatial graph convolution and temporal causal convolution paths. Dynamically switch the computational precision based on the core junction temperature to output the real-time preference probability distribution of product categories. S4. Based on the standard deviation of the real-time preference probability distribution, the core temperature change rate of the field-programmable gate array, and the variance of historical returns, an environmental perception exploration coefficient is generated. By constraining the multi-armed gambling machine algorithm, a candidate set of marketing actions is generated within the product category similarity threshold. At the same time, the throughput of behavioral data stream is obtained according to the message queue backpressure mechanism. The optimal computing power allocation point is calculated in the Riemann manifold space by combining the interrupt event weight. Marketing actions are executed by dynamic frequency division through the clock gating register, and conversion data is collected to drive the incremental update of the graph neural network.

Images