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

Abstract

The present application relates to the technical field of user behavior analysis marketing, in particular to an intelligent marketing system and method based on real-time calculation of customer behavior, which comprises an edge behavior acquisition unit, a cleaning desensitization unit, a space-time feature coding unit, a real-time prediction unit and a collaborative decision unit, the edge behavior acquisition unit in the present application captures original data such as page stay duration and touch track, the cleaning desensitization unit generates compliance time sequence segments, the space-time feature coding unit converts data into a space-time feature vector containing time decay weight, topological graph node and behavior integrity index, the real-time prediction unit fuses multi-dimensional vectors through a graph neural network, outputs real-time preference probability distribution of commodity category, and the collaborative decision unit dynamically allocates computing power to generate an optimal marketing action set, solving the problems of large deviation in capturing real-time intention of users and lag in computing power allocation in traditional marketing, improving marketing accuracy and timeliness, and reducing resource waste.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of user behavior analysis marketing, specifically to an intelligent marketing system and method based on real-time calculation of customer behavior. BACKGROUND

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

[0003] This technology captures user behavior data in real time on digital platforms, analyzes their preferences and intentions, and provides decision support for personalized marketing. It is of great significance to reduce marketing costs and improve resource utilization efficiency. Traditional marketing methods that rely on historical data or static models have been unable to cope with dynamic scenarios where user behavior changes rapidly.

[0004] However, traditional intelligent marketing technology has the core problems of user preference prediction lag and lack of precision in marketing actions. Existing solutions only analyze browsing records without integrating multi-dimensional real-time features such as page dwell time and touchpoint trajectory, resulting in a large deviation in capturing user instantaneous intentions. When users extend their page dwell time due to hesitation or interrupt transactions due to operational errors, the system cannot distinguish real demand changes, and the marketing content pushed is out of sync with user real-time preferences. At the same time, there is a lack of dynamic computing power allocation mechanism. When user behavior data surges, model calculation delay increases, further reducing prediction timeliness, leading to marketing action lag or misjudgment, affecting user experience and wasting marketing resources. It is difficult to meet the needs of real-time and personalized marketing on digital platforms. To solve this problem, we provide an intelligent marketing system and method based on real-time calculation of customer behavior. SUMMARY

[0005] The purpose of the present application is to provide an intelligent marketing system and method based on real-time calculation of customer behavior to solve the problems raised in the background technology.

[0006] 1. Since traditional marketing does not integrate multi-dimensional real-time behavior features, the deviation in capturing user instantaneous intentions is large. Therefore, this case converts page dwell time and other features into a spatiotemporal feature vector through a spatiotemporal feature encoding unit, and outputs a preference probability distribution combined with a graph neural network, which can accurately grasp user real-time needs and improve marketing content matching.

[0007] 2. Since traditional marketing lacks dynamic computing power allocation, data surges cause calculation delay. Therefore, this case dynamically allocates computing power according to data flow throughput through a collaborative decision unit, and generates a marketing action set using a constrained multi-armed bandit, which can ensure prediction timeliness and reduce marketing resource waste.

[0008] To achieve the above object, an intelligent marketing system based on real-time customer behavior calculation is provided, comprising:

[0009] 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 proxy module of the user terminal device, and the data stream contains the page stay duration, touch trajectory coordinate sequence and transaction interruption event identifier;

[0010] The cleaning and desensitization unit receives the native behavior data stream, performs real-time desensitization operation under the general data protection regulation standard, generates a compliant behavior data stream by replacing the sensitive field with an identifier, and cuts it into a standardized behavior timing segment according to a preset time window;

[0011] The space-time feature coding unit receives the compliant behavior data stream, maps the page stay duration to the time decay weight, converts the touch trajectory coordinate sequence to the topological graph node, and parses the transaction interruption event identifier to the behavior integrity indicator, and fuses to generate a unified space-time feature vector in dimension;

[0012] The real-time prediction unit is built-in with a lightweight graph neural network model, receives the space-time feature vector and performs:

[0013] Aggregate current user history vector, real-time environment vector and similar user group vector to construct a three-dimensional correlation matrix, and output the real-time preference probability distribution of the user to the preset commodity category through the field programmable gate array accelerated graph convolution layer operation;

[0014] The collaborative decision unit performs dynamic allocation of computing resources according to the behavior data stream throughput based on the real-time preference probability distribution and the current load state of the field programmable gate array calculation core, and generates an optimal marketing action set using the constrained multi-armed bandit algorithm.

[0015] The second object of the present application is to provide a method for implementing an intelligent marketing system based on real-time customer behavior calculation, comprising the following steps:

[0016] S1, capture the native behavior data stream in real time through the proxy module of the user terminal device, which contains the page stay duration, touch trajectory coordinate sequence and transaction interruption event identifier, perform real-time desensitization operation on the data stream using the GDPR standard, generate a compliant behavior data stream by replacing the sensitive field with an identifier hash, and cut it into a standardized behavior timing segment according to a preset time window, ensure data privacy compliance and processing timeliness;

[0017] S2, map the page dwell time in the compliance behavior data stream to a time decay weight with a decision hesitation coefficient, convert the touch point trajectory coordinate sequence into a topological graph node through gesture recognition technology, associate the transaction interruption event with the payment risk control system to parse into multi-layer behavior integrity indicators, input into the multi-head attention fusion module for feature channel priority weighting and dynamic alignment, generate a unified dimensional spatio-temporal feature vector, and the vector dimension maps the commodity category psychological cognitive level;

[0018] S3, aggregate the current user history vector, real-time environment vector and similar user group vector, construct a user-time-environment three-dimensional association matrix through cross-modal tensor splicing technology, use the hardware adaptive operator of the field programmable gate array to decompose the spatial graph convolution and time causal convolution channels, dynamically switch the operation precision according to the core junction temperature, and output the real-time preference probability distribution of the commodity category;

[0019] S4, generate an environment perception exploration coefficient based on the real-time preference probability distribution standard deviation, the field programmable gate array core temperature change rate and the historical return variance, generate a marketing action candidate set within the commodity category similarity threshold through the constraint multi-armed bandit algorithm, acquire the behavior data stream throughput according to the message queue back pressure mechanism, calculate the optimal computing power allocation point in the Riemannian manifold space combined with the interruption event weight, execute the marketing action through the clock gate register dynamic frequency division, and collect the conversion data to drive the incremental update of the graph neural network.

[0020] Compared with the prior art, the present application has the following advantages:

[0021] 1. The edge behavior acquisition unit cooperates with the spatio-temporal feature encoding unit, the former fully captures the page dwell time, touch point trajectory coordinate sequence and transaction interruption event and other original data through the terminal proxy module, the latter maps the dwell time to a time decay weight by introducing a decision hesitation coefficient, converts the touch point trajectory to a topological graph node by means of gesture recognition, and generates multi-layer behavior integrity indicators by associating the transaction interruption event with the payment risk control system, dynamically aligns the feature channels through the multi-head attention fusion module, generates a spatio-temporal feature vector that maps the commodity cognitive level, fully excavates the potential intention in user behavior, and solves the preference misjudgment problem caused by traditional single data dimension.

[0022] 2. The real-time prediction unit adopts a lightweight graph neural network, constructs a three-dimensional association matrix by fusing user history, real-time environment and similar user group vectors through cross-modal tensor splicing technology, accelerates graph convolution operation by means of field programmable gate array, dynamically adjusts the calculation precision according to the hardware resources, and at the same time maintains the model adaptability through incremental knowledge distillation, outputs the accurate commodity category preference probability distribution, avoids the prediction deviation caused by data processing lag in traditional models, and provides a reliable basis for marketing decision.

[0023] 3. The collaborative decision unit dynamically allocates computing power based on behavior data stream throughput, transaction interruption event weight, etc., realizes computing core frequency division through clock gate register, ensures processing efficiency when data surges, adopts constraint multi-armed bandit algorithm, combines environment perception exploration strategy and causal reasoning verification, generates optimal marketing action set adapting to user real-time preferences, reduces interference of excessive marketing on user experience, avoids waste of marketing resources, improves conversion rate and user stickiness, and meets personalized and real-time marketing needs of digital platforms. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 It is a whole block diagram of the present application;

[0025] Figure 2 It is a whole flow chart of the present application.

[0026] The meanings of various labels in the figure are as follows:

[0027] 1, edge behavior acquisition unit; 2, cleaning and desensitization unit; 3, spatiotemporal feature coding unit; 4, real-time prediction unit; 5, collaborative decision unit. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0029] The present application provides an intelligent marketing system based on real-time calculation of customer behavior, please refer to Figure 1 As shown in the figure, it comprises:

[0030] 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 proxy module of the user terminal device. The data stream contains page dwell time, touch trajectory coordinate sequence and transaction interruption event identifier;

[0031] The cleaning and desensitization unit 2 receives the native behavior data stream, performs real-time desensitization operation under the general data protection regulation standard, generates a compliant behavior data stream by replacing the identifier with a sensitive field, and cuts it into standardized behavior time sequence segments according to the preset time window;

[0032] The spatiotemporal feature coding unit 3 receives the compliant behavior data stream, maps the page dwell time to the time decay weight, converts the touch trajectory coordinate sequence to the topological graph node, and parses the transaction interruption event identifier to the behavior integrity index, and fuses to generate a unified spatiotemporal feature vector;

[0033] To accurately mine the potential intent in the user behavior data, the space-time feature encoding unit 3 maps the page stay time length into a time decay weight by combining the shopping psychology index and the multi-dimensional feature extraction technology through the behavior semantic analysis engine, and introduces the decision hesitation coefficient in the shopping psychology index;

[0034] An automatic negative exponential decay function is triggered to generate dynamic weight values when the stay duration exceeds a preset threshold. The page stay duration is processed by a behavior semantic analysis engine, and a decision hesitation coefficient is introduced to reflect the degree of hesitation of the user when selecting goods. A time decay weight is generated. According to the "nonlinear relationship between stay duration and purchase intention" in shopping psychology, when the user's stay duration on the product page is in a reasonable interval, i.e. 30 seconds-2 minutes, the decision hesitation coefficient is 1.0, indicating normal browsing. If the stay duration is too short, i.e. <30 seconds, it is marked as a false touch, and the coefficient is reduced to 0.5. If it is too long, i.e. >5 minutes, it is marked as hesitation due to indecision, and the coefficient is increased to 1.5. The preset stay duration threshold is 5 minutes. When the actual stay duration exceeds this threshold, an automatic negative exponential decay function is triggered to generate dynamic weight values. The longer the stay time, the slower the weight value decays. For example, the weight is 0.8 at 5 minutes and 0.75 at 6 minutes, avoiding feature weight distortion caused by excessive hesitation. For example, if a user stays on a mobile phone page for 7 minutes, the decision hesitation coefficient is 1.5, and the time decay weight is 0.65 after negative exponential decay calculation. This distinguishes between valid and invalid stays. The coefficient and decay function make the time feature more consistent with the true decision intention. At the same time, gesture trajectory recognition technology is used to convert the touch point trajectory coordinate sequence into a topological graph node. The intention feature vector in the finger sliding pattern is extracted by a convolutional autoencoder, and a weighted adjacency matrix is generated by combining the screen heat distribution map. The screen is divided into multiple functional areas, and the areas passed by the touch point trajectory are marked as topological graph nodes, such as "picture area -> price area -> evaluation area" corresponding to three nodes. The size of the node is weighted by the touch point stay time. The longer the stay, the larger the node. The convolutional autoencoder analyzes the finger sliding pattern to extract the intention feature vector. The screen heat map records the areas with high click frequency. The topological graph nodes are combined to generate a weighted adjacency matrix. The connection weight between nodes is calculated based on the trajectory jump frequency and the heat map 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 slowly slides their finger from the product picture area (node A) to the evaluation area (node B), and then quickly clicks the shopping cart button (node C), the generated adjacency matrix has a A-B connection weight of 0.6 (indicating attention to details due to slow sliding), and a B-C connection weight of 0.9 (indicating frequent clicking). The analysis of the transaction interruption event identifier is additionally associated with the real-time interception code library of the payment risk control system. When consecutive interruptions are detected, a multi-layer nested behavior integrity indicator is generated, which is associated with the real-time interception code library of the payment risk control system. The interruption reason is recorded, and the transaction interruption event identifier is analyzed. According to the interruption reason, a primary indicator is generated, such as "insufficient balance" marked as 0.3 (which can be remedied by recharge, with higher integrity), and "risk control interception" marked as 0.1 (higher risk, lower integrity). When consecutive interruptions are detected (such as two interruptions within 3 minutes), a secondary indicator is generated. If the two interruption reasons are the same (such as both being password errors), the secondary indicator is the product of the primary indicator (0.5*0.5=0.25, indicating that the integrity is reduced due to operation failure), if the reasons are different (such as insufficient balance for the first time, interception for the second time), the secondary index is the weighted sum of the primary index (0.3*0.6+0.1*0.4=0.22, which comprehensively reflects multiple problems), for example, the user interrupted the transaction for the first time due to "insufficient balance" (primary index 0.3), and interrupted again 5 minutes later due to "payment limit" (primary index 0.2), the secondary index is 0.3*0.5+0.2*0.5=0.25, indicating that the transaction integrity is medium, and there is still a possibility of remediation, which provides accurate input basis for subsequent real-time prediction of user preferences.

[0035] To integrate the time, space and behavior stability features into a unified vector, the process of generating the spatio-temporal feature vector is realized by a multi-head attention fusion module to dynamically align and compress multi-dimensional features. The specific implementation is as follows:

[0036] The time decay weight is taken as the time dimension scale, the adjacency matrix of the topological graph node is taken as the space relationship base, and the behavior integrity index is taken as the behavior stability correction factor. The three types of features obtained in the previous process are converted into a calculable vector form according to the preset rules and input into the multi-head attention fusion module. The time decay weight is taken as the scale, and the time sequence vector is formed according to the time sequence of user behavior (such as browsing, comparison, and order attempt). For example, the time vector of a certain user is [0.65, 0.72, 0.58], which corresponds to the decay weights of the three behavior stages respectively. The adjacency matrix of the topological graph node (reflecting the regional jump relationship of the touch track) is converted into a space vector, and the weight value of each element in the matrix is directly taken as the component of the vector (for example, the connection weight between A and B in the adjacency matrix is 0.6, and the connection weight between B and C is 0.9, which corresponds to the space vector [0.6, 0.9,...]). The multi-layer nested behavior integrity index is expanded into a vector according to the level (for example, the primary index is 0.3, and the secondary index is 0.25, which corresponds to the behavior stability vector [0.3, 0.25]). Different types of features are unified into a vector form, laying a foundation for subsequent similarity calculation and fusion, and ensuring that time, space and behavior stability features can be operated in the same dimensional space;

[0037] The multi-head attention fusion module calculates the cross-entropy similarity matrix of the three types of input vectors, assigns priority weights to the feature channels according to the similarity, and determines the priority of each feature channel by calculating the cross-entropy similarity of the three types of input vectors. The cross-entropy of the time vector and the space vector, the time vector and the behavior stability vector, and the space vector and the behavior stability vector is calculated respectively. The smaller the cross-entropy value, the stronger the relevance of the two types of features. For example, if the cross-entropy of the time vector and the space vector is 0.2, it means that the user's browsing time and touch trajectory matching degree are high. According to the cross-entropy value, the weight is assigned in reverse (the smaller the cross-entropy, the higher the weight). For example: the time-space cross-entropy 0.2 corresponds to the weight 0.4, the time-behavior stability cross-entropy 0.3 corresponds to the weight 0.3, and the space-behavior stability cross-entropy 0.5 corresponds to the weight 0.3. Ensure that the feature channels with strong relevance get higher priority. For example, a user's stay time on a mobile page (time feature) and frequent touch trajectory in the price area (space feature) have strong relevance (cross-entropy 0.15), so the priority weights of the time and space feature channels are set to 0.45 and 0.4 respectively, and the behavior stability feature (only once interrupted) weight is 0.15. Then, through the gating recurrent unit, the dynamic compression and alignment of the feature channels are realized, and finally the time-space feature vector with uniform dimension is output. For input vectors of different lengths (such as a time vector containing 3 components and a space vector containing 5 components), the gating recurrent unit discards redundant information (such as components with a weight less than 0.1 in the space vector) through the forget gate and retains core features (such as area jump relationships with a weight of 0.6 or higher). The compressed time, space, and behavior stability features are mapped to the same dimensional space (such as a 10-dimensional vector), ensuring that each dimension corresponds to a psychological cognitive level of the product category (such as the first dimension corresponding to "basic demand perception", the fifth dimension corresponding to "brand preference", and the tenth dimension corresponding to "purchase decision willingness"). For example, after compression, the time feature retains 2 core components "hesitation phase" and "decision phase", the space feature retains 2 core components "price area" and "evaluation area", and the behavior stability feature retains 1 core component "second interruption impact". The gating recurrent unit aligns them into a 10-dimensional vector, where the value of the eighth dimension (corresponding to "purchase hesitation degree") integrates the time decay weight and the behavior interruption indicator. The physical meaning of the vector dimension corresponds to the psychological cognitive level model of the product category. The final output time-space feature vector has a fixed dimension (such as 10 dimensions), and the physical meaning of each dimension corresponds to the psychological cognitive level model of the product category. Low dimension (1-3 dimensions) reflects the user's basic cognition of the product (such as initial perception of function and price), medium dimension (4-7 dimensions) reflects the user's comparison and evaluation process (such as comparison with similar products and brand trust), and high dimension (8-10 dimensions) reflects the user's decision tendency (such as purchase willingness intensity and sensitivity to promotion). For example, in a user's time-space feature vector, the value of the second dimension (price perception) is 0.8 (indicating high price attention), the 6th dimension (brand trust) value 0.3 (indicating weak brand influence), and the 9th dimension (purchase willingness) value 0.6 (indicating clear purchase tendency), which provides structured feature basis for subsequent prediction of product preferences and high-quality input features for the real-time prediction unit 4 to output preference probability distribution.

[0038] The real-time prediction unit 4 is built-in with a lightweight graph neural network model, which receives the spatio-temporal feature vector and performs:

[0039] The current user history vector, real-time environment vector, and similar user group vector are aggregated to construct a three-dimensional association matrix, and the graph convolution layer operation accelerated by the field programmable gate array is operated to output the real-time preference probability distribution of the user for the preset product category;

[0040] To comprehensively integrate user historical behavior, real-time environment, and similar group features, the real-time prediction unit 4 uses a cross-modal tensor splicing technology to construct a three-dimensional association matrix, and through orthogonal fusion of multi-dimensional features, the accuracy of user preference prediction is improved, and the specific implementation is as follows:

[0041] The historical vector extracted from the user behavior database contains the behavior characteristics of browsing, collecting, and purchasing in the past 7 days. The time slicing algorithm is used to reconstruct the time sequence cube. The historical vector is divided into multiple segments according to the time interval (e.g., one slice per day). Each slice contains the behavior characteristics in the corresponding time period (e.g., the browsing time distribution on the first day, the purchase category on the second day). A cube is constructed with "behavior type-time slice-feature value" as the three-dimensional dimension. For example, in the time sequence cube of a certain user, the feature value of "browsing behavior-3rd day-electronic product category" is 0.7, indicating that the user pays high attention to electronic products during this period. The feature value of "purchase behavior-5th day-clothing category" is 0.9, indicating that the user completed a clothing purchase during this period. The linear historical behavior data is converted into a three-dimensional structure, which not only preserves the behavior evolution trajectory in the time dimension but also clearly shows the feature distribution of different behavior types, providing a time reference for the subsequent fusion with real-time features. The real-time environment vector is generated by Kalman filter noise reduction of three groups of sensor data: geographic location fence, network delay gradient, and server load pulse. The geographic location fence data is obtained by the user terminal positioning function to mark the current area as one of the environmental characteristics. The network delay gradient records the network response speed change when the user accesses the platform. The server load pulse obtains the current platform server load fluctuation. The abnormal fluctuations in the three groups of data are smoothed to preserve the real environmental change trend. The three groups of noise-reduced data are integrated into the real-time environment vector (e.g., [0.8, 0.3, 0.6], representing high activity in the commercial district, low network delay, and medium server load, respectively) in the order of "geographic location-network state-server load". Environmental factors can affect user behavior (e.g., instant purchases are more likely to occur in commercial districts). Through noise reduction processing, the environment vector can truly reflect the user's current situation and provide a scenario-based basis for preference prediction. The similar user group vector is filtered through a dynamic topological similarity network. This network calculates the Fréchet distance of user behavior trajectories and the cosine similarity of behavior intentions in real time. Only the user subgroups with compact topological structures are retained. The Fréchet distance of user behavior trajectories measures the similarity of two trajectory shapes. For example, if the current user's browsing path shape is similar to user A's, the distance value is small. The cosine similarity of behavior intentions measures the matching degree of purchase preferences. For example, if the current user and user B both frequently browse electronic products, the similarity value is high. Double thresholds are set for distance and similarity. Only the user groups with a Fréchet distance less than the threshold and a cosine similarity greater than the threshold are retained to form a subgroup with a compact topological structure, ensuring that the behavior patterns are highly similar. The common features of the subgroup, such as preferred product categories and average dwell time, are extracted and integrated into the similar user group vector (e.g., [0.7, 0.2, 0.9] represents a sub-group with high preference for electronic products, low preference for daily necessities, and long average stay time), for example, the current user frequently browses smartphones and stays for a long time, and the dynamic topology similarity network filters out 100 users who also pay attention to smartphones and have similar browsing paths. The group vector has a feature value of 0.85 for the "electronic product category preference" feature, providing a group reference for predicting the current user's preferences. The three types of vectors are concatenated into a three-dimensional association matrix according to the user-time-environment three orthogonal dimensions, ensuring that the four types of vectors have consistent scales in the three orthogonal dimensions. The "user dimension" includes the current user and similar user group characteristics, the "time dimension" includes historical slices and real-time time points, and the "environment dimension" includes historical environment and current environment characteristics. The feature values of the four types of vectors are filled into the matrix according to the corresponding dimension positions, for example, the position of "current user-day 3- shopping district environment" in the matrix is filled with the user's behavior characteristics on the third day in the historical time sequence cube, the shopping district characteristics in the real-time environment vector, the behavior characteristics of similar user groups in the shopping district environment, and the corresponding values of the current spatio-temporal feature vector, forming a matrix unit with multiple feature superpositions. The final three-dimensional association matrix completely retains the multi-dimensional association of the user's own history, real-time state, group characteristics, and environmental impact, providing a structured input for the feature aggregation of the graph neural network, so that the subsequent output of the commodity category preference probability distribution can reflect the user's individual habits, real-time scene, and group commonality, improving the comprehensiveness and accuracy of the prediction.

[0042] To improve the operation efficiency of the graph convolution layer and adapt to the fluctuation of hardware resources, the field programmable gate array accelerated graph convolution layer operation realizes parallel processing of spatial and temporal features through a hardware adaptive operator, and outputs accurate commodity category preference probability distribution combined with a dynamic precision adjustment mechanism. The specific implementation is as follows:

[0043] The three-dimensional correlation matrix is decomposed into a spatial graph convolution path and a time convolution path to realize targeted extraction of features. The spatial path uses Chebyshev polynomial approximation to realize neighborhood node feature aggregation, and extracts the behavior correlation features of users and similar groups in the matrix (such as the commonality of users and subgroups in commodity category browsing). The “user-environment” dimension of the three-dimensional matrix is converted into a spatial topology graph (with users as nodes and similarity as edge weights), focusing on the behavior feature aggregation of different users in the same environment, and extracting the behavior evolution features of the “user-time” dimension in the matrix (such as the timing change of users from browsing to purchasing). The three-dimensional matrix is expanded into a time sequence according to time slicing, focusing on the behavior mode transition of a single user at different time points. The spatial graph convolution path uses Chebyshev polynomial approximation technology to realize efficient aggregation of neighborhood node features. The nodes of the spatial topology graph are clustered according to similarity, and users with high preference similarity are classified into one category to reduce computational complexity. The Chebyshev polynomial approximation is used to quickly calculate the weighted sum of the features of each node and its neighborhood nodes. The feature value of a certain user node is the weighted average of its own feature and the features of the three most similar users. The aggregation result reflects the influence of group behavior on individual preference. For example, the high attention of a sub-group to electronic products affects the current user. The neighborhood of the current user node contains three similar users, and after polynomial approximation calculation, the weight of the “electronic product category” in the aggregated feature is increased from 0.6 to 0.75, embodying the superimposed effect of group influence, the time path captures the evolution of behavior patterns through causal dilation convolution, using a convolution kernel with a dilation factor of 2, i.e. skipping 1 middle time point sampling, expanding the time receptive field, while covering the current, previous 2 steps, previous 4 steps of time slices, capturing long period behavior patterns, the user's weekly browsing frequency increases regularly, ensuring that convolution operations only rely on historical time point data, when calculating the current time feature, only using the behavior data of the past 1 hour, avoiding future information interference, conforming to the time sequence logic of real behavior, the past browsing behavior affects the current purchase decision, for example, the user has browsed a certain clothing commodity at 10:00, 10:10 and 10:30, the causal dilation convolution captures the association of these three time points through the dilation factor 2, the "clothing category preference" in the output time sequence feature gradually increases with time, the outputs of the two paths are input into the gate fusion unit on the high-speed interconnection bus on the field programmable gate array chip, for feature integration and precision adaptation, the gate fusion unit combines the outputs of the double-path into a comprehensive feature vector through weight distribution, the vector dimension corresponds to the preset commodity category, a 30-dimensional vector corresponds to 30 commodity categories, real-time monitoring of the hardware state of the field programmable gate array, when the resource occupancy rate is less than 70%, high-precision operation is adopted, i.e. retaining the last 4 digits of the feature vector, when the occupancy rate exceeds 90% or the junction temperature exceeds the critical value, i.e. 70℃, automatically switching to a low-bit quantization operation mode, retaining only the last 2 digits or the integer part, reducing the calculation load while ensuring basic accuracy, the fused feature vector is normalized so that the sum of the commodity category feature values is 1, converting it into a commodity category preference probability distribution, "electronic products" 0.35, "clothing" 0.25, "daily necessities" 0.4, directly reflecting the user's real-time preference degree for each category, for example, when the junction temperature of the field programmable gate array rises to 75℃, exceeding the critical value of 70℃, the gate fusion unit switches to a low-bit quantization mode, simplifying the comprehensive feature vector to generate a preference probability distribution, although the accuracy is slightly reduced, but ensures continuous and stable operation, avoiding prediction interruption caused by hardware overheating.

[0044] In order to keep the lightweight of the graph neural network model while continuously optimizing the prediction accuracy, an incremental knowledge distillation architecture is adopted, through the collaborative update of the cloud teacher model and the embedded student model, realizing efficient knowledge transfer and model scene adaptation, the specific implementation is as follows:

[0045] A "cloud teacher-terminal student" double-layer model architecture is constructed, and the function boundaries of the two are clearly defined. The teacher model is deployed on the cloud server, receives full user behavior data, including historical behavior, real-time features, and similar group characteristics for continuous training. For example, the platform receives all users' browsing records, transaction data, and environmental parameters every day, learns global behavior patterns through a complete graph neural network structure, and learns the differences in category preferences of users in different regions and seasonal consumption trends. The student model is embedded in a real-time prediction unit 4, which is optimized based on the hardware characteristics of a field programmable gate array. The convolution kernel of the model is processed by a special instruction set for sparsification, that is, the core calculation parameters are retained, and redundant parameters are removed. For example, 1000 convolution kernels are reduced to 300 key kernels to ensure that the model is small in size and fast in operation, and adapts to the power limit of the terminal device. The teacher model focuses on learning comprehensive and long-term behavior rules, and the student model focuses on real-time response and local prediction. The two achieve the transformation of "global knowledge to local ability" through knowledge distillation. The student model extracts knowledge from the teacher model at fixed time intervals to ensure timely updating of the student model without affecting real-time prediction. According to the change frequency of user behavior characteristics, the window length is set. If the platform user behavior fluctuates sharply, i.e. during the e-commerce promotion period, the window is set to 30 minutes, and if it is a daily period, the window is set to 2 hours. The balance between timeliness and system load is achieved. Each time the trigger is triggered, the teacher model generates an attention distribution heat map for the current popular commodity category, which intuitively shows the attention degree of the model to different features, such as the "price feature" in the electronic product category, the attention weight is 0.7, and the "evaluation feature" is 0.3. As a "soft label" passed to the student model, that is, an implicit knowledge signal. For example, the heat map generated by the teacher model at 10:00 shows that users pay more attention to the "delivery time" feature (weight 0.6) in the fresh food category. The student model will adjust its emphasis on this feature accordingly. The student model learns the soft label of the teacher model to match the hidden layer features of the two, as follows:

[0046] The student model and the teacher model respectively predict the same batch of samples, i.e., the user data with recent transaction behavior, and output respective hidden layer feature vectors reflecting the processing results of the features inside the model. The difference between the two sets of hidden layer feature vectors is measured by the KL divergence loss function. The smaller the difference, the closer the student model is to the cognitive mode of the teacher model, and the parameters of the student model are adjusted accordingly, increasing the weight of the "delivery timeliness" feature. The above comparison and adjustment process is repeated until the difference between the hidden layer features of the student model and the teacher model is reduced to a preset range, with a difference value less than 0.1, ensuring that the student model learns the judgment logic of the teacher model. For example, the initial weight of the "delivery timeliness" feature of the student model for fresh food categories is 0.3. After KL divergence calculation and parameter adjustment with the teacher model (weight 0.6), it gradually increases to 0.55, achieving feature alignment. In the knowledge distillation process, the weight parameters related to the real-time environment vector in the student model are frozen to avoid global knowledge covering the scene adaptability, and the model parameters corresponding to the real-time environment vector are determined. The values of these parameters are fixed during distillation and do not participate in iterative adjustment. Since the student model is deployed on the terminal and needs to respond quickly to local environment changes, the behavior switching of users moving from residential areas to commercial areas, freezing the environment-related weights can preserve their scene adaptation ability learned through local data. In the commercial environment, the high sensitivity of instant consumption categories is ensured, for example, the student model learns through local data that "when the user is in the commercial area, the prediction weight of fast food categories needs to be increased." During the distillation process, the environment-related weights are frozen to avoid being covered by the global rules of the teacher model, ensuring that the real-time prediction unit 4 can efficiently and accurately output the commodity category preference probability distribution on the terminal device, providing reliable support for marketing decisions.

[0047] The collaborative decision-making unit 5 dynamically allocates computing resources based on the real-time preference probability distribution and the current load state of the field programmable gate array computing core, and generates an optimal set of marketing actions using the constrained multi-armed bandit algorithm.

[0048] To achieve efficient utilization and real-time response of computing resources, the dynamic computing resource allocation of the collaborative decision-making unit 5 adopts an event-driven load balancing strategy, dynamically adjusting computing resources based on data stream characteristics and hardware state. The specific implementation is as follows:

[0049] By real-time acquisition of three key parameters, a three-dimensional resource allocation vector reflecting the demand for computing power and the state of hardware is constructed. Relying on the message queue back pressure mechanism of the streaming cleaning and desensitization unit 2, when the amount of data waiting for processing in the queue exceeds the preset value, an automatic feedback pressure signal is generated, and the throughput of the behavior data stream is obtained in real time. For example, when the back pressure signal shows that the amount of user behavior data accumulated in the queue reaches 1000 per second, the throughput parameter value is 0.8, and the full load is 1.0. According to the severity of the transaction interruption event, the weight is set. 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 for processing to recover the conversion possibility. The fluctuation range of the core voltage is collected by the hardware sensor and converted into a volatility parameter, reflecting the stability of hardware operation. The above three parameters are combined in the order of "throughput-interruption weight-voltage volatility" to form a three-dimensional resource allocation vector. Each dimension of the vector quantifies the key factors affecting the allocation of computing power. Based on the three-dimensional resource allocation vector, the optimal resource allocation point is located in the Riemannian manifold space, realizing the precise matching of computing power and demand. The three-dimensional vector is mapped to the Riemannian manifold space, a curved geometric space that better fits the nonlinear resource allocation rule. Each dimension corresponds to a coordinate axis in the space, for example, throughput corresponds to the "data processing demand axis", interruption weight corresponds to the "priority axis", and voltage volatility corresponds to the "hardware constraint axis". According to the vector values in the space, the coordinate point is determined. The location of this point directly reflects the current resource demand state, and the nearest resource allocation scheme to this point is found through a pre-set spatial distance algorithm. The Riemannian manifold space can better handle the nonlinear resource allocation relationship, avoiding the allocation deviation of traditional linear models in complex scenarios, ensuring that the allocation of computing power meets the data processing demand and does not exceed the hardware carrying capacity. By modifying the clock gate register of the field programmable gate array, dynamic frequency division of the computing core is realized, and resource preemption is executed when a transaction interruption event occurs. The clock gate register controls the operating frequency of the computing core. According to the results of the optimal resource allocation point, the frequency is adjusted, for example, when allocating 60% of the cores, the frequency of the corresponding core is reduced from 1GHz to 0.8GHz, and the remaining cores remain dormant. If 80% of the cores are allocated, the frequency is increased to 1.2GHz to speed up processing. When a transaction interruption event is detected, the preemption mechanism is triggered immediately, suspending the computing cores of low-priority tasks and adjusting their clock frequency to the highest, allocating them to real-time processing tasks related to the interruption event, ensuring that the interruption event is responded to within 100 milliseconds. For example, when the three-dimensional vector shows a high interruption weight, the system immediately preempts 30% of the dormant cores, adjusts their frequency to 1.2GHz, the feature extraction and marketing action generation of the interrupt event are preferentially processed, the user is avoided from abandoning the transaction due to the interrupt, and the event-driven load balancing strategy realizes the accurate allocation of the computing power through the three-dimensional vector and the Riemann manifold space, and ensures the real-time response of the key event through the dynamic frequency division and resource preemption, so that the system is efficiently operated, and the marketing conversion opportunity is maximized.

[0050] In order to enable the constrained multi-arm bandit algorithm to explore potential marketing opportunities while accurately matching real-time user demand and system running state, an environment-aware exploration strategy is adopted, the exploration intensity and range are dynamically adjusted, and a high-quality marketing action set is generated. The specific implementation is as follows:

[0051] The exploration coefficient dynamic adjustment function is defined by three parameters, which reflect the exploration demand and system carrying capacity in real time, and the standard deviation of the real-time preference probability distribution: the standard deviation is extracted from the commodity category preference probability distribution output by the real-time prediction unit 4, the larger the standard deviation, the smaller the difference in the preference probability of a user for electronic products and clothing, indicating that the user demand uncertainty is high, and the exploration needs to be strengthened. The FPGA core temperature change rate is obtained through a hardware sensor, and the temperature rises by 5°C within 10 seconds, indicating that the system load is too high, and the exploration needs to be weakened to reduce the computing pressure. The corresponding parameter value is reduced. The marketing action historical return variance is the fluctuation of the conversion effect of the same marketing action in the past 1 hour. The conversion rate of the pushed coupons is high and low, and the larger the variance, the lower the historical experience reference value, and the exploration needs to be strengthened. The corresponding parameter value is increased. The three parameters are combined into an exploration coefficient according to the preset weight, the higher the coefficient, the stronger the exploration, for example, 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 0.6x0.4+0.2x0.2+0.7x0.4=0.58, indicating that moderate-intensity exploration is needed. When it is detected that the user is in a high-value conversion scene, and has not settled after adding the shopping cart, and has browsed high-priced goods for more than 5 minutes, the exploration range is shrunk through the back propagation gradient of the graph neural network. Based on the user behavior integrity index and the commodity value, the high-value scene is determined, and the scene is marked as "precise exploration needed". The back propagation mechanism of the graph neural network is called, and the correlation gradient of different commodity categories and the current preference of the user is calculated (the larger the gradient, the more relevant the category and the current demand). Only the commodity categories with a gradient value exceeding the threshold value are retained, such as "phone accessories" and "protective film" with high preference correlation with "smartphone". The commodity category similarity threshold is set, and only marketing actions within the similarity range exceeding the value with the current preference category of the user are allowed to be explored. When the user pays attention to smartphones, only related categories such as "phone case" and "charger" are explored, and unrelated categories such as "clothing" and "food" are excluded. For example, a user has been browsing high-end cameras for a long time, which is a high-value scene. The graph neural network calculation shows that the similarity of "camera lens" and "storage card" to the camera category is 0.85, which exceeds the threshold value 0.7, The exploration action space is constrained in the marketing actions of the two types of related goods. The generated marketing action candidate set needs to be verified by the causal reasoning engine to ensure that there is a clear causal relationship with the user behavior integrity. In the constrained exploration space, specific marketing actions are generated, such as "camera lens coupon" and "storage card buy one get one free". Each action contains a trigger condition, i.e. user dwell time exceeds 3 minutes and expected effect. The causal reasoning engine analyzes the logical association between the action and user behavior, such as whether "pushing lens coupons" can directly reduce the probability of transaction interruption and improve behavior integrity. Actions without clear causality, such as pushing food coupons to camera browsing users, are excluded. The verified marketing actions are sorted by exploration coefficient and expected return, such as high return actions first, to form the optimal marketing action set. The set is sent to the user terminal simultaneously and the action execution time and parameters are recorded. For example, the candidate set "camera lens 50 yuan discount" can directly promote users to complete settlement after verification, while "attention store points" has no direct correlation with the current behavior integrity and is excluded. The final output includes the action set containing the discount, which ultimately improves marketing conversion efficiency and reduces user interference.

[0052] To cope with scenarios where marketing action return fluctuates too much or hardware state is abnormal, the environment-aware exploration strategy uses a dynamic confidence interval correction mechanism and a degradation decision mechanism to ensure the reliability of marketing actions and system stability. The specific implementation is as follows:

[0053] When the marketing action historical return variance exceeds the preset threshold, and the conversion rate of the same type of coupon push within the past 1 hour fluctuates beyond the preset range, switch to the Bayesian hierarchical model to calculate the return confidence interval, as follows:

[0054] The historical conversion data of various marketing actions is counted in real time, the return variance is calculated, if the conversion rate of a certain type of full-reduction activity fluctuates sharply between 10% and 30%, the variance exceeds the preset threshold, the confidence interval is corrected, the model includes two layers of parameters, the first layer is the individual return characteristics of specific marketing actions, the actual conversion effect of "full 200 minus 50", the second layer is the overall distribution characteristics of the same type of action, the average conversion rate of all full-reduction activities, the data fusion utilization is realized through hierarchical association, a large number of possible return values are extracted from the model using the sampling method, forming an expected return probability distribution band, the return distribution band of a certain activity is "15%-25%", indicating that there is a high probability of achieving a return within this interval, if 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%, leaving more exploration space, if the standard deviation is small, the distribution band is narrowed, focusing on high certainty actions, for example, the historical return variance of a certain skin care sample push exceeds the standard, the initial distribution band "8%-18%" is generated through model sampling, because the user's preference for skin care products has a large standard deviation, the band width is expanded to "5%-22%", ensuring that potential effective actions are not missed, when the FPGA core temperature abnormally fluctuates, a degraded decision topology network is used to replace the original constrained multi-armed bandit algorithm, the steps are as follows:

[0055] The core temperature is monitored in real time by a hardware sensor. When the temperature fluctuation amplitude exceeds the preset range, a single fluctuation exceeds 10℃ or continuously exceeds the critical value of 70℃, it is determined as an abnormal state, triggering a degradation mechanism. The commodity category similarity threshold and the behavior integrity index are subjected to tensor product operation, that is, the cross combination of the two types of parameters is taken, the marketing actions that meet the conditions are screened out, the action space scale is greatly reduced from 100 candidate actions to 20, and in the simplified space, only the actions with confidence exceeding the preset confidence threshold in the historical conversion scene are reserved, that is, the actions with stable and higher than 80% conversion rate in the past similar scene, such as "pushing a discount coupon to the user who has added the shopping cart but not settled", ensuring the high reliability of the action. For example, when the temperature is abnormal, the action space of "similarity 0.6 + integrity 0.7" is screened out through tensor product operation, and then 3 core actions with historical confidence exceeding 0.8 are reserved, such as "full-reduced coupon", "gift" and "time-limited discount". The calculation amount is reduced, and the marketing action set generated by the degradation decision needs to be verified by the causal reasoning engine and the compatibility of the user cognitive load model to ensure that it does not interfere with the user experience. The user cognitive load model includes the current operation complexity of the user, such as the number of commodities browsed at the same time and the interface interaction frequency, such as the number of clicks per minute. If the marketing action, such as a pop-up advertisement, will make the cognitive load exceed the threshold, the operation complexity will increase from 3 to 5, it is determined as incompatible, the actions that pass the compatibility verification are reserved, and are pushed after being sorted according to the historical conversion effect, such as preferentially pushing "full-reduced coupon". At the same time, the calculation resource occupation of each action is reduced, the rendering complexity of the push text is simplified, the hardware burden is reduced, for example, among the 3 core actions generated after degradation, "pop-up advertisement" is excluded because it will increase the user cognitive load, and finally the two low-interference actions of "silent full-reduced coupon" and "gift prompt" are executed. Through such dynamic confidence interval correction and degradation decision mechanism, the exploration strategy of environmental perception can not only maintain the flexibility of exploration when the return fluctuates greatly, but also ensure the stable operation of the system through simplifying the decision when the hardware is abnormal, and balance the marketing effect and system safety.

[0056] In the present application, the edge behavior acquisition unit 1 captures native data such as page stay duration and touch trajectory, the cleaning and desensitization unit 2 generates compliance time sequence segments, the space-time feature encoding unit 3 converts the data into a space-time feature vector containing time decay weight, topological graph node and behavior integrity index, the real-time prediction unit 4 fuses multi-dimensional vectors through a graph neural network, and outputs a commodity category real-time preference probability distribution. The collaborative decision unit 5 dynamically allocates computing power and generates an optimal marketing action set, solving the problem of large deviation in capturing user real-time intention and lag in computing power allocation in traditional marketing, improving marketing accuracy and timeliness, and reducing resource waste.

[0057] Please refer to Figure 2 The second object of the present application is to provide a method for implementing an intelligent marketing system based on customer behavior real-time calculation as described above.

[0058] S1, capturing native behavior data stream in real time through a proxy module of a user terminal device, containing page dwell time, touch trajectory coordinate sequence and transaction interruption event identifier, performing real-time desensitization operation on the data stream according to GDPR standard, generating compliant behavior data stream through identity identifier hash replacement and sensitive field encryption confusion, and cutting into standardized behavior time sequence fragments according to a preset time window, ensuring data privacy compliance and processing timeliness;

[0059] S2, mapping the page dwell time in the compliant behavior data stream into a time decay weight with a decision hesitation coefficient, converting the touch trajectory coordinate sequence into a topological graph node through gesture recognition technology, and correlating the transaction interruption event with a payment risk control system to parse into multi-layer behavior integrity indicators, inputting a multi-head attention fusion module for feature channel priority weighting and dynamic alignment to generate a unified dimensional spatio-temporal feature vector, and mapping the vector dimension to the psychological cognitive level of the commodity category;

[0060] S3, aggregating the current user history vector, real-time environment vector and similar user group vector, constructing a user-time-environment three-dimensional correlation matrix through cross-modal tensor splicing technology, decomposing spatial graph convolution and time causal convolution channels using hardware adaptive operators of a field programmable gate array, dynamically switching operation precision according to core junction temperature, and outputting a commodity category real-time preference probability distribution;

[0061] S4, generating an environment perception exploration coefficient 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 historical return variance, generating a marketing action candidate set within the commodity category similarity threshold through a constrained multi-armed bandit algorithm, simultaneously obtaining the behavior data stream throughput according to the message queue back pressure mechanism, calculating the optimal computing power allocation point in the Riemannian manifold space combined with the interruption event weight, executing the marketing action through clock gate register dynamic frequency division, and collecting conversion data to drive incremental updates of a graph neural network.

[0062] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application 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 calculation of customer behavior 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