An advertising recommendation method and system integrating user dynamic behavior modeling

Abstract

The present application provides an advertising recommendation method and system that integrates user dynamic behavior modeling. Among them, by obtaining the user's cross-device operation records on multiple terminal devices, the device unique identification code, high-frequency click area data and operation timestamp are synchronously collected. Based on the combined analysis of the device switching time interval and click behavior, a spatiotemporal relationship graph is constructed to generate a feature set reflecting the correlation of behaviors. The trajectory similarity is calculated using hardware acceleration at the edge computing node to generate migration correlation parameters. The spatiotemporal relationship graph is input into the behavior analysis model to extract coherent behavior features, and a multi-dimensional scene portrait is formed by combining geographic location changes and interface element distribution. Finally, by identifying the user's cross-device interest migration pattern, the advertising content that matches the current scene is dynamically screened to achieve the optimization of precise push strategies based on behavioral continuity. The technical solution provided by this application significantly improves the accuracy of cross-terminal advertising recommendations and the speed of scene response.

Description

Technical Field

[0001] The present application relates to the technical field of user behavior analysis and modeling, and in particular to an advertisement recommendation method and system integrating user dynamic behavior modeling. Background Art

[0002] In multi-terminal linkage scenarios, user behavior exhibits dynamic migration characteristics. For example, after browsing products on a mobile phone, a user may switch to a PC to complete the payment process, or immediately use a mobile phone to query detailed information after triggering a service through a smart speaker. Advertising recommendation systems need to capture cross-device behavior sequences in real time, model the migration patterns of user intentions, and predict their potential needs in specific device scenarios. However, the format difference between mobile click data and PC page jump paths makes it difficult to align features. The difficulty in modeling spatiotemporal correlations between devices also restricts the accuracy of recommendations. For example, it is difficult to achieve real-time correlation analysis between geographic location changes and interface interaction features. The conflict between the high computational overhead of multi-source data fusion and the need for low-latency recommendations further exacerbates the complexity of system implementation.

[0003] Currently, a representative solution is based on a hybrid strategy of device behavior clustering and collaborative filtering. This approach uses unique device identifiers to link user activity records across multiple devices and extract the independent behavioral characteristics of each device. An improved temporal clustering algorithm is then used to segment cross-device behavior into fixed time windows. The algorithm calculates behavioral similarity between devices and selects highly correlated device groups. Finally, ad candidate sets are matched based on common behavioral characteristics within these device groups. Summary of the Invention

[0004] The present application provides an advertising recommendation method and system that integrates user dynamic behavior modeling to solve the problem in the prior art that the recommended content lacks integration with the current device scene attributes.

[0005] In a first aspect, the present application provides an advertising recommendation method integrating user dynamic behavior modeling, comprising:

[0006] Obtain user operation records across multiple devices, synchronously record the unique identification code of each device, capture the user's high-frequency click area data on the device interface, and mark the time when each operation occurs;

[0007] Calculating the time interval between user switching between different terminal devices based on the unique identification code and the occurrence time, combining and analyzing the switching time interval with the high-frequency click area data to generate a feature set reflecting the correlation between the user's cross-device behavior, and constructing a spatiotemporal relationship graph containing device nodes and behavior paths based on the feature set;

[0008] In the edge computing node, a hardware acceleration unit is used to calculate the trajectory similarity of the cross-device operation records to generate a migration correlation parameter representing the consistency of the user's cross-device behavior;

[0009] Input the spatiotemporal relationship graph into a behavior analysis model to extract the user's consistent behavior characteristics across different terminals, combine the migration association parameters, and fuse the device's geographic location change data and interface element distribution characteristics to generate a multi-dimensional scene portrait;

[0010] Based on the fusion results of the coherent behavioral features and the multi-dimensional scene portraits, the interest migration pattern in the user's cross-device behavior is identified, and advertising content that matches the user's current cross-device scene is filtered from the advertising library according to the interest migration pattern, and the advertising push strategy is dynamically adjusted.

[0011] Optionally, in trajectory similarity calculation, a multimodal adversarial training network is used to perform adversarial learning on device operation records and interface click distribution data to optimize and generate migration correlation parameters with time alignment function;

[0012] Based on the optimized migration association parameters, a dynamic adjustment model for edge weights is constructed, node connection strengths in the spatiotemporal relationship graph are updated by adversarial learning loss values, and the updated spatiotemporal relationship graph is input into the behavior analysis model to optimize the extraction of coherent behavior features;

[0013] Establishing a joint learning framework between edge computing nodes, dynamically adjusting the weight parameters of the model output based on the edge weights, and iteratively optimizing the generation rules of the multi-dimensional scene portrait;

[0014] The advertising push strategy is input into the behavioral inference engine, and an adaptive feedback strategy is generated based on the noise characteristics output by adversarial learning. Based on the adaptive feedback strategy, the parameter update mechanism of the model is adjusted.

[0015] Optionally, an incremental graph convolution module is embedded in the spatiotemporal relationship graph to adjust the aggregation range of adjacent nodes based on the device switching frequency and the calibration value output by adversarial learning;

[0016] Construct a dual-path memory enhancement unit. The first path stores the optimized spatiotemporal features of high-frequency click areas and synchronizes them with the generation rules of the multi-dimensional scene portrait. The second path records abnormal events in cross-device behavior paths and triggers topology reconstruction based on weight parameters.

[0017] Integrating a time series prediction module into the hardware acceleration unit to predict the estimated value of the scene portrait in the future time period based on the migration correlation parameters generated by adversarial learning;

[0018] When the deviation between the estimated value of the scene portrait and the actual behavior of the user exceeds the set value three times in a row, topology reconstruction is triggered, and the generation rules of the multi-dimensional scene portrait are reset according to the reconstructed spatiotemporal relationship graph.

[0019] Optionally, according to the unique identification code, the user's operation record sequence on multiple terminals is matched, and the operation records of each device are sorted according to the occurrence time;

[0020] Traverse the operation record sequence of adjacent terminals and calculate the interval from the last operation time of the user on the current terminal to the first operation time of the next terminal as the switching time interval;

[0021] Extracting the interface click density distribution corresponding to each terminal in the high-frequency click area data, and counting the high-frequency click area coordinates of the previous and next terminals when each switching time interval occurs;

[0022] The switching time interval is combined with the coordinates of the high-frequency click area to generate a cross-device behavior correlation feature set. Based on the correlation feature set, multiple device operations in which the user's continuous switching time interval is less than a set value are combined into a hyperedge to construct a spatiotemporal relationship graph with hierarchical edges.

[0023] Optionally, an attribute vector is added to each node in the spatiotemporal relationship graph, including the physical position change of the corresponding terminal, the screen size, and the shape characteristics of the interface click area;

[0024] Through the node propagation mechanism in the behavior analysis model, the attribute vectors of adjacent nodes are aggregated along the hierarchical edge direction to generate the behavior code of each node;

[0025] Converting the migration association parameter into a dynamic scaling coefficient that is inversely associated with the edge weight in the spatiotemporal relationship graph, and adjusting the edge weight in real time according to the dynamic scaling coefficient when nodes are aggregated;

[0026] Extracting the spatial offset related to the device position in the behavior code, and calculating the cross-device behavior continuity index based on the shape similarity of the click area on the interface;

[0027] The cross-device behavior continuity indicator is nonlinearly fused with the adjusted edge weights to generate a multi-dimensional scenario portrait that characterizes the consistency of the user's multi-device behavior.

[0028] Optionally, in the edge computing node, the hardware acceleration unit is used to compare the high-frequency click area numbers on the device interface in the cross-device operation records of any two terminal devices, and extract the matching position distribution features of continuous matches in the operation records of the two devices;

[0029] Based on the matching position distribution characteristics, the hardware acceleration unit calculates a coverage density index of the operation traces between the two devices, where the coverage density index is formed by multiplying the ratio of the sum of the matching segment lengths to the total length of the longer operation record by the dispersion of the matching segment positions;

[0030] Applying a time decay factor to the coverage density index by the hardware acceleration unit to generate a cross-device trajectory similarity parameter, wherein the time decay factor is inversely proportional to the time interval between the operation records of the two devices;

[0031] After traversing all device combinations, device pairs whose cross-device trajectory similarity parameters exceed a preset threshold are selected to generate migration correlation parameters that characterize the consistency of user cross-device behavior.

[0032] Optionally, based on the fusion result of the coherent behavior feature and the multi-dimensional scene portrait, extracting the cross-device interaction hotspot area from the multi-dimensional scene portrait;

[0033] Performing a spatiotemporal overlay analysis of the device switching frequency and operation sequence repetition period in the coherent behavior features with the interaction hotspot areas to generate a probability matrix representing the user's cross-device behavior interest migration pattern. The probability matrix is ​​constructed by analyzing the hotspot area transfer paths and time distribution density of operation sequences between different devices;

[0034] Identifying the core migration path and associated device combinations of the interest migration pattern in the user's cross-device behavior based on the transition probability peaks of the continuous device switching paths in the probability matrix;

[0035] Based on the core migration path, screening advertising content that matches the usage scenario tag of the associated device combination from the advertising library, and calculating the matching weight between the advertising scenario feature vector and the corresponding migration path in the probability matrix;

[0036] Dynamically construct an advertising push priority queue, and the sorting weight of the advertising push priority queue is adjusted in real time by the superposition result of the matching weight and the activity of the high-frequency click area of ​​the current device interface. Based on the advertising push priority queue, adjust the advertising push strategy.

[0037] Optionally, the switching time interval is combined with the coordinates of the high-frequency click areas of the terminal before and after the switching, the inverse of the switching time interval is calculated as the temporal correlation strength value, and the plane distance between the coordinates of the high-frequency click areas of the terminal before and after the switching is calculated as the spatial correlation difference value;

[0038] The temporal correlation strength value and the spatial correlation difference value are fused into a multidimensional feature vector according to a preset ratio, wherein the proportion of the temporal correlation strength value increases as the switching time interval shortens, thereby generating a cross-device behavior correlation feature set;

[0039] Based on the associated feature set, extract a multi-device operation combination with a time interval less than a set value in the user's continuous device switching events; if the number of associated terminal devices is greater than or equal to three and the continuous switching time intervals are all less than the set threshold, then connect the multi-device operation combination through a hyperedge;

[0040] Assign a hierarchical weight to each hyperedge, and construct a spatiotemporal relationship graph with hierarchical edges based on the hierarchical weight and connection direction of the hyperedge.

[0041] Optionally, the migration-associated parameter of each device pair is converted into a dynamic scaling coefficient, wherein the dynamic scaling coefficient is inversely proportional to the value of the migration-associated parameter, and a larger value of the migration-associated parameter corresponds to a smaller dynamic scaling coefficient;

[0042] Extracting the migration association parameter of the device pair corresponding to the hyperedge according to the generation order and path direction of each hyperedge in the spatiotemporal relationship graph, and performing a product operation on the dynamic scaling coefficient and the path length of the device associated with the hyperedge to generate a hyperedge weight adjustment factor;

[0043] During the node aggregation process, when the attribute vector is transmitted along the hyperedge direction, the original hierarchical weight of the hyperedge is scaled in real time based on the hyperedge weight adjustment factor, so that the hyperedge hierarchical weight between the device pairs with high values ​​of the migration association parameter is suppressed, and the hyperedge hierarchical weight with low values ​​of the migration association parameter is enhanced.

[0044] In a second aspect, the present application provides an advertising recommendation system that integrates user dynamic behavior modeling, including:

[0045] The acquisition module obtains the user's cross-device operation records on multiple terminal devices, synchronously records the unique identification code of each device, captures the user's high-frequency click area data on the device interface, and marks the time when each operation occurs;

[0046] a calculation module, which calculates the time interval between user switching between different terminal devices based on the unique identification code and the occurrence time, combines and analyzes the switching time interval with the high-frequency click area data to generate a feature set reflecting the correlation between the user's cross-device behavior, and constructs a spatiotemporal relationship graph containing device nodes and behavior paths based on the feature set;

[0047] a generation module, which calculates the trajectory similarity of the cross-device operation records through a hardware acceleration unit in an edge computing node to generate a migration correlation parameter representing the consistency of the user's cross-device behavior;

[0048] An analysis module inputs the spatiotemporal relationship graph into a behavior analysis model, extracts the user's consistent behavior characteristics across different terminals, combines the migration association parameters, and integrates the device's geographic location change data and interface element distribution characteristics to generate a multi-dimensional scene portrait;

[0049] The adjustment module identifies the interest migration pattern in the user's cross-device behavior based on the fusion result of the coherent behavioral characteristics and the multi-dimensional scene portrait, filters the advertising content that matches the user's current cross-device scene from the advertising library according to the interest migration pattern, and dynamically adjusts the advertising push strategy.

[0050] In an embodiment of the present application, a user's cross-device operation records on multiple terminal devices are obtained, a unique identification code for each device is synchronously recorded, data on high-frequency click areas on the device interface of the user is captured, and the time of occurrence of each operation is marked; based on the unique identification code and the time of occurrence, the switching time interval between the user's different terminal devices is calculated, the switching time interval is combined and analyzed with the high-frequency click area data to generate a feature set reflecting the correlation between the user's cross-device behavior, and a spatiotemporal relationship graph containing device nodes and behavior paths is constructed based on the feature set; trajectory similarity calculation is performed on the cross-device operation records in an edge computing node through a hardware acceleration unit to generate migration association parameters that characterize the consistency of the user's cross-device behavior; the spatiotemporal relationship graph is input into a behavior analysis model to extract the user's coherent behavior features across different terminals, and the migration association parameters are combined with the device geographic location change data and interface element distribution features to generate a multi-dimensional scene portrait; based on the fusion result of the coherent behavior features and the multi-dimensional scene portrait, the interest migration pattern in the user's cross-device behavior is identified, and advertising content matching the user's current cross-device scenario is screened from an advertising library based on the interest migration pattern, and the advertising push strategy is dynamically adjusted.

[0051] This application has the following beneficial effects:

[0052] By obtaining unique identification codes, high-frequency click area data, and the time of operation, the device switching time interval is quantified and combined with click behavior to generate a feature set of cross-device behavior correlation. A spatiotemporal relationship graph of device nodes and behavior paths is constructed to solve the problem of multi-terminal intention dispersion. The hardware acceleration unit of the edge computing node calculates the similarity of cross-device trajectories in real time, generates migration correlation parameters that characterize behavioral consistency, reduces noise interference, and improves computing efficiency. The spatiotemporal relationship graph is input into the behavior analysis model to extract user coherent behavior characteristics, and integrates the device geographic location change data and interface element distribution characteristics to form a multi-dimensional scene portrait, enhancing the in-depth analysis of cross-device context. Finally, based on the fusion results of coherent behavior characteristics and multi-dimensional portraits, the user interest migration pattern is identified, the advertising library content is dynamically matched, and the push strategy is adjusted to achieve precise adaptation of advertising to dynamic cross-device scenarios, overcome the problem of identifying fragmented attention, and improve scenario response speed and advertising conversion rate.

[0053] Furthermore, adversarial learning enhances the ability to model temporal consistency across device behaviors, addressing parameter drift caused by differences in multimodal data distribution. A dynamic edge weight adjustment model improves the accuracy of spatiotemporal graphs in representing the dynamic evolution of user behavior. A joint learning framework enables collaborative optimization of profile generation rules across edge nodes, enhancing the global consistency of profiles across device scenarios. An adaptive feedback strategy, combined with a noise feature suppression mechanism, improves the robustness of ad push strategies and avoids local overfitting. This overall enhancement strengthens the model's generalization capabilities and real-time decision-making adaptability for complex cross-device scenarios, further optimizing the long-tail conversion rate of ad recommendations.

[0054] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0056] Figure 1 A flowchart of an advertisement recommendation method integrating user dynamic behavior modeling provided by the present application is shown;

[0057] Figure 2 A scenario diagram showing an advertisement recommendation method integrated with user dynamic behavior modeling provided by the present application is shown;

[0058] Figure 3 A schematic diagram of the structure of an advertisement recommendation system integrated with user dynamic behavior modeling provided by the present application is shown. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0060] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0061] Researchers have discovered that existing ad recommendation systems face multiple technical bottlenecks in cross-device scenarios: the heterogeneity of mobile and PC operational data makes it difficult to uniformly model user behavior characteristics; the spatiotemporal correlations generated by device switching are difficult to capture in real time due to dynamic geographic location changes and differences in interface interactions; and the fusion of multi-source heterogeneous data presents a conflict between high latency and high resource consumption. For example, the behavioral migration pattern of users browsing products on a tablet and then placing an order on a mobile phone often leads to delayed recommendations due to device feature alignment errors and insufficient real-time computing power. Therefore, an intelligent recommendation method that can dynamically analyze cross-device behavioral correlations is urgently needed.

[0062] In response to the above problems, the present invention proposes an advertising recommendation method that integrates user dynamic behavior modeling, the core of which is to construct a cross-device spatiotemporal relationship graph and migration association analysis accelerated by edge computing. Specifically, by synchronously collecting multi-terminal device identification codes and high-frequency click area data, a spatiotemporal relationship graph is generated in combination with the device switching time interval to characterize the user behavior trajectory; the hardware acceleration unit of the edge node is used to calculate the trajectory similarity in real time to generate migration association parameters; finally, the device geographic location offset and interface element distribution characteristics are integrated to construct a multi-dimensional scene portrait containing spatiotemporal coherent behavior characteristics. This method realizes cross-device behavior feature alignment through a spatiotemporal relationship graph, combines edge computing to reduce data transmission latency, and enables geographic location changes and interface interaction features to participate in interest migration pattern recognition in real time. While ensuring recommendation accuracy, the computational overhead of multi-source data fusion is reduced, effectively solving the system-level contradiction between data heterogeneity modeling between devices and real-time low-latency recommendation.

[0063] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0064] Figure 1 The present invention provides a flowchart of an advertisement recommendation method integrating user dynamic behavior modeling. Figure 1 As shown, the method includes:

[0065] 101. Obtain user operation records across multiple devices, synchronously record the unique identification code of each device, capture the user's high-frequency click area data on the device interface, and mark the time when each operation occurs;

[0066] In the above steps, cross-device operation records refer to the collection of historical operation behavior data performed by users on different terminal devices, including interactive actions such as clicking, browsing, and sliding; unique identification code refers to the unique identifier generated by device hardware or software, which is used to distinguish different terminal devices; high-frequency click area data refers to the screen coordinates or area identifiers that users repeatedly click on in the device interface, and their preferred areas are determined by counting the frequency of user operations; occurrence time refers to the specific timestamp of the user's operation, accurate to milliseconds or seconds, which is used to record the chronological order of behavior.

[0067] In the embodiments of the present application, first, a user's cross-device operation records are obtained on multiple devices. Specifically, device-side tracking technology or log collection tools are used to capture user interactions on different devices, such as mobile phones, tablets, and computers, in real time. The operation logs of different devices are aggregated into a unified data platform. Secondly, the unique identification code of each device is synchronously recorded. During this process, the interface provided by the device operating system is called to obtain the hardware-level unique identifier of each device. This identifier is then bound to the user account to ensure that cross-device operation records can be associated with the same user. Next, data on the user's frequently clicked areas on the device interface is captured. Front-end event monitoring technology or heat map analysis tools are used to record the screen coordinates or interface element IDs of user click events. By counting the number of clicks on the same area within a certain time window, areas with click frequencies exceeding a preset threshold are selected to serve as the basis for analyzing user interest preferences. Finally, the time of each operation is marked. When a user performs an operation, an accurate timestamp is obtained using the device's local clock or Network Time Protocol (NTP). This timestamp, along with the operation type, device identifier, and click area data, is stored in a behavior log database, forming a cross-device behavior sequence with time sequence information.

[0068] In practical applications, for example, in the implementation of a cross-device recommendation system on an e-commerce platform, the system uses device fingerprints to bind the user's smartphone, laptop, and smart speaker device identifiers, collecting multi-device operation sequences in real time. When a user browses a kitchen appliance product page on a tablet for 85 seconds and triggers the parameter comparison function five times, the system detects the user's evening query for the "auto-clean function" command on their home voice device. Combined with their subsequent visits to three related product pages on their office computer the following day, a time window aggregation strategy is used to construct a behavioral migration path from "mobile product screening - home appliance function verification - desktop purchasing decision." Based on the user's lunchtime search for "office supplies" on their mobile device, and considering the cross-device data showing a 68% share of office scenarios and a weekly search frequency of 3.8 home appliance searches, the system pushes real-time discounts on commercial equipment and consumables combinations. It also recommends smart home solutions when the user switches to their home large-screen device in the evening. This improves cross-device recommendation conversion rates, enhances behavioral path completeness, and enhances user behavior consistency across multiple device scenarios compared to traditional solutions.

[0069] In the overall solution of step 101 above, by collecting the operation records, interface interaction hot spots and timestamp data of multiple terminal devices in real time, combining the unique device identification code to build a cross-device behavior association network, and using deep learning to perform spatiotemporal modeling of user dynamic behavior, the law of interest migration across scenarios is accurately identified to solve the problem of fragmentation of single-device portraits. By integrating interface interaction heat maps with device context features, the system strengthens the scenario-based intent parsing capability and improves the accuracy and scenario adaptability of advertising recommendations. Based on the edge computing framework, real-time synchronization of cross-device behavior is achieved, which can dynamically predict the potential needs of users when switching devices. For example, mobile browsing behavior triggers collaborative advertising delivery of associated home terminals, optimizing cross-terminal experience and advertising effectiveness. The system supports parallel processing of massive device nodes, providing highly responsive and highly scalable intelligent recommendation services for multi-terminal linkage scenarios.

[0070] 102. Calculate the time interval between user switching between different terminal devices based on the unique identification code and the occurrence time, combine and analyze the switching time interval with the high-frequency click area data to generate a feature set reflecting the correlation between the user's cross-device behavior, and construct a spatiotemporal relationship graph containing device nodes and behavior paths based on the feature set;

[0071] Optionally, step 102 may specifically include the following steps:

[0072] 1021. Match the user's operation record sequences on multiple terminals according to the unique identification code, and sort the operation records of each device according to the occurrence time;

[0073] 1022. Traverse the operation record sequence of adjacent terminals and calculate the interval from the last operation time of the user on the current terminal to the first operation time of the next terminal as the switching time interval;

[0074] 1023. Extract the interface click density distribution corresponding to each terminal in the high-frequency click area data, and count the high-frequency click area coordinates of the previous and next terminals when each switching time interval occurs;

[0075] 1024. Combine the switching time interval with the high-frequency click area coordinates to generate a cross-device behavior correlation feature set. Based on the correlation feature set, combine multiple device operations in which the user's continuous switching time interval is less than a set value into a hyperedge, and construct a spatiotemporal relationship graph with hierarchical edges.

[0076] Among them, step 1024 may specifically include the following processes: combining the switching time interval with the high-frequency click area coordinates of the terminals before and after the switching, calculating the inverse of the switching time interval as the time correlation strength value, and calculating the plane distance of the high-frequency click area coordinates of the terminals before and after the switching as the spatial correlation difference value; fusing the time correlation strength value and the spatial correlation difference value into a multi-dimensional feature vector according to a preset ratio, wherein the proportion of the time correlation strength value increases with the shortening of the switching time interval, and generating a cross-device behavior correlation feature set; based on the correlation feature set, extracting a multi-device operation combination with a time interval less than a set value in the user's continuous device switching events, if the number of associated terminal devices is greater than or equal to three and the continuous switching time intervals are all less than the set threshold, then the multi-device operation combination is connected through a hyperedge; assigning a hierarchical weight to each hyperedge, and constructing a spatiotemporal relationship graph with hierarchical edges based on the hierarchical weight and connection direction of the hyperedge.

[0077] In the above steps, the switching time interval refers to the time difference between the end of the user's last operation on the current device and the beginning of the first operation on the next device; the interface click density distribution refers to the spatial clustering degree of the user's click positions on the interface of a single terminal device, which is obtained by counting the click frequency in different screen areas; the spatiotemporal relationship graph refers to a directed graph structure constructed by combining the time dimension and the spatial position of the device, in which the device node represents the terminal device, the behavior path describes the user's cross-device jump, and the hyperedge is used to connect multiple device nodes associated in the same time period; the hierarchical weight refers to the priority parameter assigned to the edge based on the device switching frequency and operation density; the time correlation strength value is an indicator of cross-device behavior continuity quantified in the form of the inverse of the switching time interval; the spatial correlation difference value is the Euclidean distance between the high-frequency click areas before and after the switch in the screen coordinate system; the multidimensional feature vector is a feature expression that integrates the time correlation strength and spatial difference according to the weight ratio.

[0078] In this embodiment of the present application, first, step 1021 is used to match the user's operation record sequences across multiple terminals based on the unique identification code. Specifically, the unique identification code is used as an index key to aggregate the operation logs of the same user across different devices, and the operation records in all devices are globally sorted in chronological order of occurrence, forming a cross-device operation chain structure with a timestamp.

[0079] Next, the operation record sequence of adjacent terminals is traversed in step 1022. Based on the operation chain structure, the operation records of two consecutive devices are extracted. The absolute difference is calculated by subtracting the timestamp of the last operation of the previous device from the timestamp of the first operation of the latter device. This difference is the switching interval, which is used to indicate the urgency of the user's active device switching.

[0080] Next, step 1023 extracts the interface click density distribution for each terminal within the high-frequency click area data. A kernel density estimation algorithm is used to create a two-dimensional distribution model of the click coordinates for each device, generating a probability density heat map. Density peaks are detected as high-frequency areas. When a user switches devices, the center coordinates of the high-frequency click areas within the corresponding time window are compared before and after the switch to quantify the degree of spatial shift.

[0081] Finally, step 1024 combines the switching time interval with the coordinates of the high-frequency click area for analysis. The time correlation strength value and spatial correlation difference value are calculated for each switching event, and the weight ratio of the time feature is dynamically adjusted according to the switching interval. For example, the time weight is 70% when the interval is 10 seconds, and 50% when the interval is 30 seconds. The normalized feature values ​​are spliced ​​into a multidimensional vector to form a set of associated features. When constructing the spatiotemporal relationship graph, the user's continuous device switching events are detected, such as clicking on a product on a mobile phone, searching for the same brand on a tablet, and completing a payment on a computer. If the switching interval of this series of operations is less than the set threshold and the number of devices involved is ≥3, a hyperedge is used to bind them into a behavior cluster, and the hierarchical weight of the hyperedge is assigned = the number of devices × the average time correlation strength. The direction of the arrow indicates the order of device switching, and finally a multi-level graph structure with spatiotemporal correlation is formed.

[0082] In a cross-device ad recommendation practice on an e-commerce platform, device fingerprints were used to correlate user activity on smartphones, tablets, and smart TVs. When a user browsed a running shoe product page on a tablet, the platform captured the user switching to their mobile device 19 seconds after finishing browsing. They then made five consecutive clicks on the waterproof performance description section of the product details page, with the click hotspot concentrated in the lower left corner of the screen (80 to 110 pixels on the X axis and 500 to 530 pixels on the Y axis). After a 31-second interval, they switched to their smart TV to watch a product review video. The algorithm calculated the time-correlation strength of the first switch as 0.053. Based on the planar distance between the center coordinates of the last click on the tablet (X95 / Y515) and the center coordinates of the first click on the phone (X102 / Y508), the spatial correlation difference was measured to be 9.7 pixels, generating a feature vector with a temporal weight of 71% and a spatial weight of 29%. When the user subsequently queries off-road terrain adaptation parameters on their in-car device at 22-second intervals and accesses the user review section on their laptop at 28-second intervals, the platform identifies that the intervals between these three consecutive device switches are all less than the 35-second threshold. It then connects the four terminal nodes via a hyperedge with a hierarchical weight of 0.85, forming a cross-device decision chain that spans browsing tablet core functions, verifying mobile phone parameters, and querying vehicle-mounted scene adaptation, all the way to word-of-mouth analysis on the computer. Based on this behavioral graph, the platform recommends professional-grade off-road running shoes when the user switches to their office desktop computer and a reflective night running kit when they use their home projector at night. Actual data shows that recommendations driven by spatiotemporal correlations increase cross-device ad click-through rates, shortening the user's decision-making time from watching reviews on their smart TV to placing an order on their mobile phone to 3 minutes and 48 seconds. Behavioral paths with hyperedge weights exceeding 0.8 contribute to cross-device conversions, improving the accuracy of multi-device behavior matching.

[0083] In the overall solution of step 102 above, by matching the unique identification codes and operation timestamps of multiple terminal devices, the user's cross-device behavior sequence is arranged in chronological order, and the time interval between the last and first operations of adjacent terminals is calculated as the switching interval. At the same time, the coordinates of the high-frequency click areas of the terminal interface before and after the switch are extracted, and the cross-device association features are generated by combining the time interval and the click density distribution. Specifically, the inverse of the time interval is converted into the time correlation intensity, and the plane distance of the click area is converted into the spatial correlation difference. The two are fused into a multidimensional feature vector according to the dynamic weight to form a cross-device behavior association feature set. Based on the continuous switching time interval threshold, the multi-device operation combination that meets the conditions is connected through a hyperedge, and the hyperedge is given a hierarchical weight based on the time interval and spatial difference. Finally, a spatiotemporal relationship graph with hierarchical edges and direction attributes is constructed to characterize the spatiotemporal association pattern of the user's cross-device behavior.

[0084] 103. In the edge computing node, a hardware acceleration unit is used to perform trajectory similarity calculation on the cross-device operation records to generate a migration correlation parameter representing the consistency of the user's cross-device behavior;

[0085] Optionally, step 103 may specifically include the following steps:

[0086] 1031. In the edge computing node, using the hardware acceleration unit, compare the high-frequency click area numbers on the device interfaces in the cross-device operation records of any two terminal devices, and extract the matching position distribution features of consecutive matches in the operation records of the two devices;

[0087] 1032. Based on the matching position distribution characteristics, calculate, by the hardware acceleration unit, a coverage density index of the operation traces between the two devices, where the coverage density index is formed by multiplying the ratio of the sum of the matching segment lengths to the total length of the longer operation record by the dispersion of the matching segment positions.

[0088] 1033. Apply a time decay factor to the coverage density index through the hardware acceleration unit to generate a cross-device trajectory similarity parameter, where the time decay factor is inversely proportional to the time interval between the operation records of the two devices.

[0089] 1034. After traversing all device combinations, select device pairs whose cross-device trajectory similarity parameter values ​​exceed a preset threshold, and generate migration association parameters that characterize the consistency of the user's cross-device behavior.

[0090] In the above steps, edge computing nodes refer to distributed servers deployed near user devices, which are used to execute computing tasks near the data source to reduce transmission delays; hardware acceleration units refer to dedicated computing chips integrated into edge nodes, such as GPUs, FPGAs, or ASICs, which are used to improve the execution efficiency of intensive computing tasks; trajectory similarity calculation refers to measuring the consistency of cross-device behavior by quantifying the degree of overlap of user operation trajectories between different devices; migration association parameters refer to device association strength indicators generated after similarity screening, which are used to reflect the habitual patterns of user migration operations between multiple devices; coverage density index is a composite indicator that combines the operation trajectory matching length and location distribution concentration, which is used to characterize the similarity of behavior patterns between two devices; time decay factor refers to the coefficient that dynamically adjusts the similarity weight based on the time interval, and its value decays as the time interval between the operation records of the two devices increases.

[0091] In an embodiment of the present application, first, in step 1031, a hardware acceleration unit is used in the edge computing node to compare the high-frequency click area numbers of the cross-device operation records. Specifically, the interfaces of different terminal devices are divided into multiple areas according to a preset grid, and each area is assigned a unique number. For example, the mobile phone screen is divided into a 10×10 grid and numbered from 1 to 100. Through the parallel computing module in the hardware acceleration unit, the high-frequency click area numbers of the operation records of the two devices are extracted, and a hash algorithm is used to quickly generate a high-frequency area number sequence. Subsequently, continuous matching segments are detected in the number sequence. For example, the number sequence of device A contains 5-8-3, and the number sequence of device B contains 5-8-3-2. The continuous matching segment 5-8-3 is extracted as the matching position distribution feature, and its starting position, length and distribution range are recorded.

[0092] Secondly, the coverage density index of the operation trajectory between the two devices is calculated based on the matching position distribution characteristics through step 1032. Specifically, the ratio of the sum of the lengths of the matching segments to the total length of the longer operation records of the two devices is calculated. For example, the total length of the trajectory of device A is 100 steps, the total length of the trajectory of device B is 80 steps, and the total length of the matching segment is 60 steps, then the ratio is 60 / 100=0.6. At the same time, the discreteness of the matching segment position distribution is calculated, and the standard deviation is used to measure the distribution concentration of the matching segment in the original operation sequence. For example, the standard deviation of the matching segment interval distance is 1.2. Finally, the coverage density index is obtained by multiplying the ratio by the inverse of the discreteness, that is, 0.6×(1 / 1.2)=0.5. The higher the index, the stronger the spatial overlap and continuity of the operation trajectories of the two devices.

[0093] Next, in step 1033, a time decay factor is applied to the coverage density index to generate a cross-device trajectory similarity parameter. The time decay factor is calculated using an exponential function. For example, if the decay coefficient is set to 0.1 and the time interval between the operation records of two devices is 2 hours, the decay factor is 1 / (1+0.1×2)=0.83. Multiplying the coverage density index of 0.5 by the decay factor of 0.83 yields a similarity parameter of 0.415. The decay calculations for all device pairs are batch processed by the floating-point arithmetic module in the hardware acceleration unit to ensure that the similarity parameter decreases significantly with increasing time intervals.

[0094] Finally, through step 1034, all device combinations are traversed and high-similarity device pairs are screened to generate migration association parameters. Utilizing the parallel computing capability of the hardware acceleration unit, the similarity parameters of all device combinations are calculated simultaneously, for example, C(100,2)=4950 calculations are performed on 100 devices. A similarity threshold is set, for example, 0.4, and device pairs whose similarity parameters exceed this value are screened out, such as 0.415 for device AB and 0.42 for device CD. The device identifiers and their similarity values ​​in the screening results are stored in the migration association parameter set, which is used in subsequent recommendation models to identify consistent patterns of user cross-device behavior. For example, when the user's shopping cart operations on mobile phones and tablets are highly correlated, cross-device advertising recommendations are automatically triggered.

[0095] In practical applications, for example, in the deployment of a cross-terminal recommendation system on an e-commerce platform, edge computing nodes process user multi-device behavior data through hardware acceleration units. When a user browses a sports camera product page on a smartphone, the edge node captures the coordinates of six consecutive clicks in the product parameter area in real time. The click hotspot is concentrated in the range of 200-220 pixels on the X axis and 350-380 pixels on the Y axis. Simultaneously, the user switches to a tablet device 15 seconds later, generating similar click trajectories when comparing similar products on the same page. The hotspot is distributed in the range of 195-215 pixels on the X axis and 345-375 pixels on the Y axis. The hardware acceleration unit maps and compares the click area numbers of the two devices, extracting 12 consecutive matching location feature points. The calculated matching segment accounts for 82% of the total length, with a position dispersion coefficient of 0.83 and an initial coverage density index of 0.68. Taking into account the 15-second interval between the two device operations and applying a time decay factor of 0.92, the final trajectory similarity parameter is corrected to 0.63. When the user subsequently accessed the product review video via their smart TV at an 18-second interval, the edge node recalculated the trajectory similarity parameter between the tablet and TV to be 0.58, and the migration association parameters formed between the three devices all exceeded the 0.55 threshold. Based on this, when the user browses outdoor equipment using an in-vehicle device, the platform, taking into account the consistency of their cross-device behavior, pushes a waterproof accessory package compatible with the action camera. Actual operational data shows that the hardware acceleration unit shortens the time required for trajectory similarity calculation, improves the accuracy of migration association parameters, drives up the conversion rate of cross-device associated products, and improves the completeness of behavioral link recognition for more than three devices.

[0096] In the overall solution for step 103 above, leveraging the hardware acceleration capabilities of edge computing nodes, real-time trajectory comparisons are performed on the frequently clicked area numbers of multiple devices, extracting location matching features for cross-device operations. The coverage density index is calculated by multiplying the length of the matching segments by the dispersion of the location distribution. The time decay factor is dynamically adjusted based on the time interval between operations to quantify the trajectory similarity between devices. Ultimately, highly similar device combinations are screened, and migration correlation parameters representing the consistency of user behavior are generated. This effectively identifies the correlation strength of cross-device operation patterns, providing a precise basis for cross-device behavior continuity analysis for ad recommendations and improving the contextual relevance of ad reach in multi-device scenarios.

[0097] 104. Input the spatiotemporal relationship graph into a behavior analysis model to extract the user's consistent behavior characteristics across different terminals, combine the migration association parameters, and fuse the device's geographic location change data and interface element distribution characteristics to generate a multi-dimensional scene portrait;

[0098] Optionally, step 104 may specifically include the following steps:

[0099] 1041. Add an attribute vector to each node in the spatiotemporal relationship graph, including the physical position change of the corresponding terminal, the screen size, and the shape characteristics of the interface click area;

[0100] 1042. Through the node propagation mechanism in the behavior analysis model, the attribute vectors of adjacent nodes are aggregated along the hierarchical edge direction to generate the behavior code of each node;

[0101] 1043. Convert the migration association parameter into a dynamic scaling coefficient that is inversely correlated with the edge weight in the spatiotemporal relationship graph, and adjust the edge weight in real time according to the dynamic scaling coefficient when the nodes are aggregated;

[0102] Among them, step 1043 may specifically include the following processes: converting the migration association parameter of each device pair into a dynamic scaling coefficient, wherein the dynamic scaling coefficient is inversely proportional to the value of the migration association parameter, and the larger the value of the migration association parameter, the smaller the corresponding dynamic scaling coefficient; according to the generation order and path direction of each hyperedge in the spatiotemporal relationship graph, extracting the migration association parameter of the corresponding device pair of the hyperedge, multiplying the dynamic scaling coefficient with the path length of the associated device in the hyperedge, and generating a hyperedge weight adjustment factor; in the node aggregation process, when the attribute vector is transmitted along the hyperedge direction, the original hierarchical weight of the hyperedge is scaled in real time based on the hyperedge weight adjustment factor, so that the hyperedge hierarchical weight between the device pairs with high values ​​of the migration association parameter is suppressed, and the hyperedge hierarchical weight of the device pairs with low values ​​of the migration association parameter is enhanced.

[0103] 1044. Extract the spatial offset related to the device position in the behavior code, and calculate the cross-device behavior continuity index based on the shape similarity of the click area on the interface;

[0104] 1045. Nonlinearly fuse the cross-device behavior continuity indicator with the adjusted edge weight to generate a multi-dimensional scenario portrait that characterizes the consistency of the user's multi-device behavior.

[0105] In the above steps, the attribute vector refers to the multidimensional feature vector attached to the node of the spatiotemporal relationship graph, which includes the physical position offset distance of the terminal device, the screen resolution size, and the aspect ratio or contour features of the interface click area; the node propagation mechanism refers to the information transmission algorithm based on the graph neural network, which updates the current node representation by aggregating the features of adjacent nodes; the dynamic scaling factor is a weight adjustment factor inversely related to the migration association parameter, which is used to control the contribution intensity of the hyperedge in feature propagation; the spatial offset refers to the longitude and latitude difference or distance measurement of the geographical location change when the user operates between different devices; the shape similarity of the interface click area is an indicator of the morphological consistency of the cross-device click area calculated by the geometric contour matching algorithm; nonlinear fusion refers to the operation of weighted combination of multi-source features using a fully connected neural network layer or an attention mechanism; the multi-dimensional scene portrait is a comprehensive feature expression of the user's cross-device behavior at the spatial, temporal, and interactive morphological levels.

[0106] In the embodiment of the present application, first, an attribute vector is added to each node in the spatiotemporal relationship graph through step 1041. Specifically, the geographical location information of the terminal is extracted from the device metadata, such as GPS coordinates or IP positioning, and the movement distance of the device during continuous operation is calculated as the physical position change; the device screen size parameters, such as 1920×1080 pixels, are obtained and converted into standardized values; the shape of the high-frequency click area is analyzed through the image contour detection algorithm, and the aspect ratio, circularity or convex hull area is extracted as the shape feature. The above three types of data are spliced ​​into a multidimensional attribute vector and bound to the corresponding device node. For example, the vector of node A is a position change of 0.5km, a screen size of 0.8, and a shape similarity of 0.7.

[0107] Next, the behavior analysis model in step 1042 executes the node propagation mechanism to generate the behavior code. Specifically, a graph convolutional network algorithm is used to propagate node attributes along the hyperedge direction in the spatiotemporal graph. For device node B, features are aggregated from its upstream node A and downstream node C: node A's attribute vector is scaled by the hyperedge weights and concatenated with node B's original vector. This is then input into the weight matrix of the graph convolution layer for a linear transformation, and then activated using the ReLU function to generate the updated behavior code for node B. This process is iterated layer by layer until the behavior codes of all nodes converge, forming a feature representation that integrates neighborhood information.

[0108] Next, step 1043 converts the migration association parameter into a dynamic scaling factor and adjusts the hyperedge weight. For example, the migration association parameter of device pair AB is 0.8, and the inverse function is used to convert it into a dynamic scaling factor of 1 / 0.8=1.25. According to the physical connection path length of the devices in the hyperedge, such as A to B spanning two intermediate devices, the scaling factor is multiplied by the path length to generate a weight adjustment factor of 1.25×2=2.5. In the node aggregation stage, the original hyperedge level weight is 0.6, which is scaled by the adjustment factor to 0.6×2.5=1.5. This makes the device pairs with lower migration association parameters, such as parameter 0.2 corresponding to the adjustment factor 5, occupy a higher weight in feature propagation, suppressing the influence of highly associated device pairs.

[0109] Then, step 1044 is used to extract the cross-device behavior continuity index from the behavior code. The spatial offset component related to the geographic location is separated from the behavior code, and the Euclidean distance of the position change between device nodes is calculated. At the same time, the shape features of the interface click areas of adjacent device nodes are compared using the cosine similarity algorithm. If the aspect ratio difference is less than 0.1, it is considered similar. After normalizing the spatial offset to the range of 0-1, it is weighted and summed with the shape similarity value, such as weights of 0.7 and 0.3 respectively, to generate a cross-device behavior continuity index, which is used to quantify the coherence of user operations in space and interface form.

[0110] Finally, a multi-dimensional scene portrait is generated through nonlinear fusion in step 1045. The cross-device behavior continuity index and the adjusted hyperedge weights are input into the bidirectional attention mechanism layer: first, the two types of features are mapped to the same dimensional space through a fully connected network, and the attention score is calculated to determine the contribution ratio of each feature. Then, a gated fusion unit is used to splice and compress the weighted features, outputting a multi-dimensional scene portrait that includes temporal continuity, spatial movement patterns, and interface interaction inertia. For example, in a work scenario, a user frequently uses a mobile phone and a laptop for cross-device operations, and the interface click area morphology is highly consistent. The final portrait generates a labeled feature "Office Collaboration Scene - Cross-Device Coherence Index 0.85".

[0111] In practical applications, for example, in a cross-terminal recommendation practice on an e-commerce platform, the platform processes user multi-device behavior data through a behavioral analysis model. When a user browses drone products on a 6.5-inch smartphone, the device's GPS location indicates it is in a commercial area, with the interface's click hotspot distributed vertically in a strip-like pattern spanning 280-300 pixels on the right side of the screen's X-axis. 14 minutes later, when the user switches to a 10.4-inch tablet, the device's location shifts to a residential area, and the click hotspot shifts to a horizontal block-like pattern spanning 800-820 pixels on the bottom side of the screen's Y-axis. The model assigns attribute vectors to each of the two device nodes, including a location offset of 1.2 kilometers, a screen size difference index of 0.37, and a click area shape difference coefficient of 0.68. When aggregated through a node propagation mechanism, the detected migration correlation parameter between the two devices is 0.62, which is converted to a dynamic scaling factor of 1.61, which is then multiplied by the original hyperedge level weight of 0.75 to a value of 0.47. When the user subsequently watched the aerial video on a 55-inch smart TV, the model calculated a shape similarity of 0.82 between the touch area on the tablet's interface. Combined with the spatial continuity index of 0.71 formed by the position offsets between the three devices, this was fused with the adjusted hyperedge weights using a hyperbolic tangent function to generate a multidimensional scene profile with a device-scene compatibility of 0.65 and a behavioral coherence of 0.78. In practice, when a user browsed on a car device, the platform recommended wind-resistant drone accessories based on the outdoor scene preference value of 0.83 in the profile. When the user switched to an office desktop, a portable folding drone was recommended based on the office scene relevance of 0.72. Data showed that this approach improved cross-device recommendation accuracy, reducing the average time it takes for users to compare tablet parameters and place an order on their mobile phone to 5 minutes and 12 seconds. The dynamic scaling mechanism reduced the mismatch rate for highly mobile device pairs, improving scene recognition accuracy within the multidimensional profile.

[0112] In the overall solution of step 104 above, a multi-dimensional scene portrait generation mechanism is constructed through the collaborative calculation of the spatiotemporal relationship graph and the migration association parameters. Specifically, the attribute of the device nodes in the spatiotemporal relationship graph is expanded through the behavior analysis model, and feature vectors such as the geographic location offset and the interface click area morphology are added. The node propagation mechanism is used to aggregate the behavior codes of adjacent devices along the hierarchical edge direction. By calculating the spatial offset in the node behavior code and the similarity of the interface click area morphology, and combining the weighted hierarchical edges for nonlinear feature fusion, a multi-dimensional scene portrait is finally generated that integrates spatiotemporal continuity, interface interaction consistency, and device migration association strength. This significantly improves the context-awareness capability of cross-device behavior prediction and provides fine-grained user intent representation for advertising recommendations in dynamic scenarios.

[0113] 105. Based on the fusion results of the coherent behavioral features and the multi-dimensional scenario portraits, identify the interest migration pattern in the user's cross-device behavior, filter the advertising content that matches the user's current cross-device scenario from the advertising library according to the interest migration pattern, and dynamically adjust the advertising push strategy.

[0114] Optionally, step 105 may specifically include the following steps:

[0115] 1051. Extracting a cross-device interaction hotspot area from the multi-dimensional scene portrait based on a fusion result of the coherent behavior feature and the multi-dimensional scene portrait;

[0116] 1052. Perform spatiotemporal overlay analysis on the device switching frequency and operation sequence repetition period in the continuous behavior features and the interaction hotspot areas to generate a probability matrix representing the user's cross-device behavior interest migration pattern. The probability matrix is ​​constructed by analyzing the hotspot area transfer paths and time distribution density of operation sequences between different devices.

[0117] 1053. Identify the core migration path and associated device combination of the interest migration pattern in the user's cross-device behavior based on the transition probability peak of the continuous device switching path in the probability matrix;

[0118] 1054. Based on the core migration path, filter advertising content that matches the usage scenario tag of the associated device combination from the advertising library, and calculate the matching weight between the advertising scenario feature vector and the corresponding migration path in the probability matrix;

[0119] 1055. Dynamically construct an advertisement push priority queue, wherein the sorting weight of the advertisement push priority queue is adjusted in real time by the superposition result of the matching weight and the activity of the high-frequency click area of ​​the current device interface, and the advertisement push strategy is adjusted based on the advertisement push priority queue.

[0120] In the above steps, the interest migration pattern refers to the changing pattern of user interest preferences when transferring operations between multiple devices, which is quantified through the spatiotemporal correlation between device switching paths and interaction hot spots; the probability matrix is ​​a two-dimensional matrix constructed by counting the transfer frequency and time density of operation sequences between different devices, which is used to describe the possible distribution of user cross-device behavior paths; the core migration path refers to the device jump sequence corresponding to the transfer probability peak in the probability matrix, which represents the user's most frequent or most stable cross-device operation habits; the matching weight is the scenario fit score between the advertising content and the user migration path, which is calculated based on the feature vector similarity; the advertising push priority queue is a push list dynamically sorted according to the advertising matching degree and real-time user behavior, and the weight integrates long-term interest and immediate activity.

[0121] In the embodiment of the present application, first, step 1051 is used to extract cross-device interaction hotspots from the multi-dimensional scene portrait. Specifically, a cluster analysis is performed on the shape features of the interface click areas in the scene portrait, and the density peak of the high-frequency click area is detected using the density-based clustering algorithm DBSCAN. Areas with similar spatial positions and click frequencies exceeding a threshold are marked as interaction hotspots. For example, the shopping cart icon area of ​​a mobile device and the payment button area of ​​a tablet are marked as a cross-device hotspot pair due to their high click frequencies and adjacent locations, forming a hotspot distribution map.

[0122] Secondly, a probability matrix representing the interest migration pattern is generated through step 1052. The device switching frequency and the operation sequence repetition period in the coherent behavior characteristics are normalized, and the device switching events are mapped to the time axis of the interactive hotspot area through spatiotemporal superposition analysis technology. For each device switching event, the number of hot zone transfers from device A to device B is counted, and its proportion of the total number of device A switches is calculated as the transfer probability value, which is filled into the corresponding row and column positions of the matrix. For example, the hot zone transfer probability from device A to device B is 0.7, and that from device B to device C is 0.3, forming a probability distribution matrix for cross-device behavior.

[0123] Next, step 1053 identifies the core migration path. A peak detection algorithm is used within the probability matrix to traverse all possible device switching paths and calculate the product of the transition probabilities for each path. If the product of the transition probabilities for paths A, B, and C is 0.7 multiplied by 0.6 (equal to 0.42), and a threshold of 0.4 is set, then the path is determined to be the core migration path. The associated device combination A, B, and C is extracted as the core link for the user's cross-device interest migration.

[0124] Then, in step 1054, matching ad content is filtered and a matching weight is calculated. Ads matching the usage scenario tags associated with the device combination associated with the core migration path are retrieved from the ad library. The scenario feature vector is extracted from the ad metadata. For example, if the ad category is electronics and the device type is mobile phone or tablet, the cosine similarity algorithm is used to calculate the similarity between the ad feature vector and the feature vector of the corresponding migration path in the probability matrix to obtain the matching weight. For example, if the feature vector of ad X has a similarity of 0.85 with the device type sequence of paths A, B, and C, its matching weight is 0.85.

[0125] Finally, step 1055 is used to dynamically construct an advertising push priority queue. The activity of the high-frequency click area of ​​the current device interface is obtained in real time. For example, the number of clicks on the shopping cart area of ​​the current mobile phone interface accounts for 30% of the total operations of the current session. The matching weight and the activity are weighted and summed according to a preset ratio. For example, the matching weight accounts for 0.6 and the activity accounts for 0.4 to generate an advertising ranking score. The score of advertisement X is 0.85 times 0.6 plus 0.3 times 0.4, which is equal to 0.63. The score of advertisement Y is 0.72 times 0.6 plus 0.5 times 0.4, which is equal to 0.632. Then advertisement Y is displayed first. A priority queue is generated based on the real-time updated ranking score, and the push strategy is adjusted to refresh the queue every 5 minutes, giving priority to triggering high-weight advertisements when the user switches to the next device in the core migration path.

[0126] In practical applications, for example, an e-commerce platform's cross-device ad recommendation practice uses dynamic recommendation optimization by analyzing user behavior across multiple devices. When a user browsed for wireless earphones on a smartphone, the platform detected seven consecutive clicks on the noise cancellation feature parameter area. The hotspot was concentrated in the lower right corner of the screen, between 250 and 280 pixels on the X axis and 600 and 630 pixels on the Y axis, with an average interval of 2.3 seconds between clicks. Eight minutes and 15 seconds later, the user switched to a tablet device and generated five clicks on the same coordinate area on the product comparison page, with an 85% overlap of hotspots. Edge computing nodes used spatiotemporal overlay analysis to generate a probability matrix representing interest migration, identifying the core migration path from smartphone feature research to tablet comparisons. This path had a peak probability of 0.81 during evening hours. Based on this information, the ad selection engine matched an ad for active noise cancellation earphones, calculating a match weight of 0.73 for the current device combination scenario. Combined with real-time heart rate data from the smartwatch, the platform prioritized ads related to exercise scenarios. When the user played music on their in-car device the next day, the system detected that the interval between consecutive device switching had shortened to an average of 4 minutes and 50 seconds. Using the vehicle's GPS location to reveal the characteristics of their commuting route, the system dynamically adjusted the ad to a set of in-car noise-cancelling earbuds, increasing the matching weight based on the commuting time scenario. Actual application data shows that recommendations based on interest migration patterns have increased cross-device ad conversion rates, boosted click-through rates for ads matching the core migration path, shortened the user's decision cycle from comparing tablet specifications to purchasing a smartwatch, reduced ineffective ad exposure, and maintained a high level of accuracy in multi-device scenario tag recognition.

[0127] In the overall solution of step 105 above, a dynamic advertising push decision-making mechanism is constructed by integrating cross-device coherent behavioral characteristics and multi-dimensional scene portraits. Specifically, based on the spatiotemporal superposition analysis of user behavior interaction hot spots and device switching frequencies, a probability matrix representing the interest migration path is generated, and the core migration path and associated device combinations are identified through the transfer probability peaks in the matrix. The system calculates the matching degree between the advertising content and the scene labels corresponding to the migration path, and adjusts the push priority in real time based on the advertising scene feature vector and the current device interface click activity, forming an advertising sorting queue that adapts to cross-end scene changes. This mechanism can capture the continuity of user intentions in the collaborative use of multiple devices, ensure the deep fit between advertising content and cross-device behavioral scenarios, achieve accurate synchronization between advertising exposure timing and user interest migration nodes, and enhance the contextual awareness and user experience consistency of cross-end advertising recommendations.

[0128] The following is a complete embodiment of steps 101 to 105:

[0129] like Figure 2 As shown, after a smart home appliance brand deployed a user behavior migration analysis module in its multi-terminal recommendation system, it recorded that a user triggered the fitness data synchronization function through a smartwatch at 8:15 in the morning. Within 3 minutes and 42 seconds after the synchronization, the mobile health app generated 6 calorie counter usage records. When the user switched to the office laptop to visit the shopping website at 9:07, the system captured the device switching interval of 52 minutes. Using the FPGA chip of the edge node, it completed cross-terminal trajectory matching within 0.6 milliseconds and found that the mouse wheel's rest position on the coffee machine product page overlapped with 83% of the page focus areas in the mobile browsing history. Combined with smart home gateway data, the system identified that the IoT device in the user's home kitchen triggered the use of the steam oven 19 times in the past 7 days, and each operation was preceded by a mobile recipe search behavior 45 minutes before.

[0130] The system leveraged a 256-core distributed computing cluster to construct a cross-device spatiotemporal graph in 1.5 seconds. Analysis revealed that when users accessed home appliance forums via tablets between 8:00 PM and 10:00 PM, 67% of their touches were concentrated around keywords like "quiet design" and "energy-saving mode." When a cooking program played on a smart TV at 9:18 PM, the recommendation engine correlated the location data of three offline experience stores the user had searched for that day via their car's screen at 6:30 PM. Combined with the average weekly cooking frequency of 5.2 times recorded by the kitchen's smoke sensor, the recommendation engine generated a recommendation for a "quiet kitchen appliance package." The next morning at 7:50 AM, an ad for a brand of silent blender pushed to the user's phone lock screen during their commute achieved a 12.3% click-through rate. Meanwhile, an ad for the same brand of silent microwave oven displayed in the browser sidebar on their office PC generated a 4.7% cross-device conversion rate.

[0131] Actual operational data shows that the system has reduced cross-device recommendation response time to 0.8 seconds, increased the ad exposure device synergy index to 8.6, and increased the combined purchase rate of target products on smart TVs and mobile phones by 9.8% compared to single-device recommendation models. When a user views the ingredient list on their smart refrigerator's touchscreen at 9:05 PM, the system dynamically loads discount information on six small kitchen appliances that match the current inventory based on a comparison of 23 air fryers viewed on a tablet at 2:27 PM that day, creating a closed-loop precision marketing campaign across multiple devices.

[0132] Optionally, the method also includes: in the trajectory similarity calculation, adversarial learning is performed on the device operation records and interface click distribution data through a multimodal adversarial training network to optimize and generate migration association parameters with time alignment function; based on the optimized migration association parameters, an edge weight dynamic adjustment model is constructed, the node connection strength in the spatiotemporal relationship graph is updated through the adversarial learning loss value, and the updated spatiotemporal relationship graph is input into the behavior analysis model to optimize the extraction of coherent behavior features; a joint learning framework is established between edge computing nodes, and the weight parameters output by the edge weight dynamic adjustment model are adjusted based on the weight parameters, and the generation rules of the multi-dimensional scene portrait are iteratively optimized; the advertising push strategy is input into the behavior inference engine, and an adaptive feedback strategy is generated in combination with the noise characteristics output by the adversarial learning, and the parameter update mechanism of the model is adjusted based on the adaptive feedback strategy.

[0133] In the above steps, the multimodal adversarial training network refers to a deep learning framework that simultaneously processes the time series data of device operation records and the spatial data of interface click distribution, and improves the feature alignment capability through the adversarial game between the generator and the discriminator; the temporal alignment function refers to a mechanism that automatically matches the operation trajectories of different devices with key behavior fragments in the time dimension through adversarial learning; the edge weight dynamic adjustment model refers to an adaptive network module that adjusts the hyperedge weights in the spatiotemporal relationship graph in real time based on the adversarial learning loss value; the joint learning framework refers to a distributed training mechanism in which multiple edge computing nodes share model weight parameters under the premise of data privacy protection to collaboratively optimize the portrait generation rules; the adaptive feedback strategy refers to a closed-loop optimization mechanism that generates reverse signals in real time according to the effect of advertising push, and adjusts the model parameter updates through noise feature injection.

[0134] In the embodiment of the present application, the trajectory similarity calculation process is first optimized through a multimodal adversarial training network. Specifically, a dual network structure of a generator and a discriminator is constructed: the generator uses a temporal convolutional network to process the time series of device operation records and extracts operation interval and duration features; at the same time, another branch of the generator models the spatial correlation of interface click distribution through a graph attention network, and outputs an operation trajectory embedding vector that integrates spatiotemporal features. The discriminator receives the real user behavior sequence and the pseudo sequence synthesized by the generator, and forces the migration association parameters generated by the generator to have the ability to align timing through the adversarial loss function. For example, the generator needs to make the click timestamp difference between the mobile phone and the tablet in the shopping cart operation stage less than 100 milliseconds, and the discriminator identifies the misaligned segments and feeds back the loss value. After iterative optimization, migration association parameters with precise timing alignment are generated.

[0135] Secondly, a dynamic edge weight adjustment model is constructed based on the optimized migration association parameters. The discriminator loss value in adversarial learning is used as a signal input, and a weight adjustment factor is designed: when the adversarial loss value of a specific device pair is higher than the threshold, indicating that its temporal alignment is low, the hierarchical weight of the corresponding hyperedge in the spatiotemporal relationship graph is reduced. Specifically, through a recurrent neural network with a gating mechanism, the historical weight value is fused with the current adversarial loss value to generate a dynamic scaling coefficient. For example, if the adversarial loss value of device AB exceeds 0.5 three times in a row, its hyperedge weight will gradually decay from 0.8 to 0.3, thereby weakening the influence of low-quality association paths in the behavior analysis model.

[0136] Next, a joint learning framework is established between edge computing nodes to optimize the scene portrait generation rules. After each edge node locally trains the edge weight dynamic adjustment model, it only uploads the model weight gradient to the central server, and generates a global model through a secure aggregation algorithm. For example, the weight gradient of node 1 is ΔW1, and that of node 2 is ΔW2. After aggregation, the global gradient is (ΔW1+ΔW2) / 2, which is then sent to each node to update the local model. In each round of iteration, the multi-dimensional scene portrait is regenerated based on the updated weight parameters. For example, the weight of the geographic location feature is adjusted from 30% to 45%, so that the portrait focuses more on the user's cross-device movement pattern.

[0137] Finally, the ad push strategy is fed into the behavioral inference engine for adaptive feedback. During adversarial learning, the noise features output by the discriminator are extracted as adversarial samples and fed into the feedback strategy generator. This generator uses a reinforcement learning mechanism to dynamically adjust the model parameter update rules based on changes in ad click-through rates. If the click-through rate of a particular ad type drops by more than 20% due to noise interference, the learning rate of the corresponding feature channel is reduced. Conversely, if the click-through rate of an anti-interference ad remains stable, its ranking weight is increased. For example, when adversarial noise causes fluctuations in the performance of electronic product ads, the feedback strategy automatically increases the feature regularization strength of these ads to mitigate the risk of overfitting.

[0138] In practical applications, for example, in the implementation of a cross-device recommendation system for smart homes, edge computing nodes process user behavior data using a multimodal adversarial training network. When a user triggers air quality monitoring with a smartwatch, the adversarial network generator collects device acceleration data at a rate of 12 samples per second and modally aligns it with six consecutive click events (each less than 1.2 seconds apart) in the x-axis region of 180-210 pixels on the tablet. The initial timing offset of 23 seconds was reduced to 4.3 seconds after three rounds of training. After the system detected that a user had spent 127 seconds browsing fresh produce on the smart refrigerator's touchscreen, the edge weight dynamic adjustment model increased the device connection strength from an initial 0.68 to 0.82 based on an adversarial loss of 0.34. A planar attenuation factor of 1.2 kilometers between the delivery address and the refrigerator's coordinates was also applied. The behavioral inference engine injects spatial noise of ±15 pixels along the x-axis into the smart TV's viewing behavior. When a deviation of 0.47 in the user's focus area is detected, a backpropagation mechanism is triggered, adjusting the convolution kernel learning rate to 0.0021. The federated learning framework synchronizes the updated weight matrix to three regional servers within 47 milliseconds, enabling iterative cross-node optimization of kitchen scenario parameters. In actual deployments, the timing synchronization coefficient remains stable above 0.89, device connection strength dynamically adjusts from 0.45 to 0.93, model update iteration cycles are shortened to within 12 training rounds, and user cross-device decision paths are compressed to an average of 4 minutes and 17 seconds, with system response latency maintained at milliseconds.

[0139] Through the above steps, a multimodal adversarial training mechanism is introduced to enhance cross-device behavior modeling capabilities. In trajectory similarity calculation, an adversarial learning framework is used to jointly optimize device operation records and interface click distributions. An adversarial game between the generator and the discriminator achieves cross-modal temporal alignment, enabling transfer correlation parameters to capture implicit behavioral correlations between devices. An edge weight adjustment model dynamically updates the spatiotemporal graph based on an adversarial loss. The noise characteristics output by the generator are used to reversely correct node connection strengths, enhancing the stability of the graph structure under the influence of abnormal behavior. A joint learning framework is deployed between edge computing nodes to achieve cross-node collaborative optimization of model parameters based on a dynamic edge weight parameter distribution, ensuring the adaptability of multi-dimensional scene profiling rules to localized behavior patterns. The advertising push strategy integrates the noise characteristics generated by adversarial learning through a behavioral inference engine, constructing a feedback-driven adaptive adjustment mechanism. This real-time detection of policy execution deviations and reverse-iterative updates of model parameters form a closed-loop optimization chain from behavior analysis to ad reach, significantly improving the dynamic adaptability and anti-interference capabilities of the recommendation system in cross-device scenarios.

[0140] Optionally, the method also includes: embedding an incremental graph convolution module in the spatiotemporal relationship graph, adjusting the aggregation range of adjacent nodes based on the device switching frequency and the calibration value output by adversarial learning; constructing a dual-path memory enhancement unit, the first path storing the optimized spatiotemporal features of the high-frequency click area and synchronizing them to the generation rules of the multi-dimensional scene portrait, and the second path recording abnormal events in the cross-device behavior path and triggering topology reconstruction based on weight parameters; integrating a timing prediction module in the hardware acceleration unit, predicting the estimated value of the scene portrait in the future time period based on the migration association parameters generated by adversarial learning; when the deviation between the estimated value of the scene portrait and the actual behavior of the user exceeds the set value for three consecutive times, triggering topology reconstruction, and resetting the generation rules of the multi-dimensional scene portrait according to the reconstructed spatiotemporal relationship graph.

[0141] In the above steps, the incremental graph convolution module refers to a graph neural network component based on dynamically adjusting the aggregation range of neighbor nodes, and adaptively controlling the feature propagation radius through the calibration value of the device switching frequency and the adversarial learning output; the dual-path memory enhancement unit refers to a memory network that includes independent storage and abnormal event triggering. The first path is used to solidify the spatiotemporal features of high-frequency click areas, and the second path is used to detect cross-device behavior anomalies and drive graph structure updates; the timing prediction module refers to the time series prediction model integrated in the hardware acceleration unit, which predicts the evolution trend of future scene portraits based on migration correlation parameters; topology reconstruction refers to the structural reorganization of the spatiotemporal relationship graph triggered by user behavior deviations, including the redistribution of node connection weights and the dynamic resetting of multi-dimensional scene portrait generation rules.

[0142] In an embodiment of the present application, an incremental graph convolution module is first embedded in the spatiotemporal relationship graph. Specifically, the connection density index between nodes is calculated based on the device switching frequency. For example, the number of switches from device A to device B per hour is 5 times, and the density index is 5 / maximum number of switches 10=0.5. At the same time, the migration association parameter calibration value output by the adversarial learning, such as the calibration value of device AB is 0.8, is normalized to an aggregation radius coefficient of 0-1 through the Sigmoid function. The density index is multiplied by the calibration coefficient to obtain the final aggregation range adjustment factor of 0.5×0.8=0.4. In the incremental graph convolution, only the adjacent node features whose adjustment factors exceed the threshold of 0.3 are aggregated. For example, the neighbor node B of device A is included in the aggregation range due to the adjustment factor of 0.4, while node C is excluded due to the adjustment factor of 0.2, thereby realizing dynamic neighbor selection.

[0143] Secondly, a dual-path memory enhancement unit is constructed. The first path uses a key-value memory network to store the spatiotemporal features of high-frequency click areas: the screen coordinates, click timestamps, and device identifiers of high-frequency areas are encoded into vectors. The attention mechanism is used to retrieve features that match the current scene profile generation rules. For example, when the generation rules focus on shopping scenarios, coordinate features related to clicks on product detail pages are preferentially extracted. The second path uses a long short-term memory network (LSTM) to detect cross-device behavior path anomalies: the interval time and path transition probability of users continuously switching devices are recorded. If the switching time interval suddenly increases by 3 times and the transition probability decreases by 50%, it is marked as an abnormal event. When an abnormal event triggers topology reconstruction, the node connection strength is recalculated based on the current edge weight parameters. For example, the weight of the abnormal path A to B is reduced from 0.6 to 0.2, and a new high-probability path B to D is added with a weight of 0.7.

[0144] Next, a time series prediction module is integrated into the hardware acceleration unit. A temporal convolutional network (TCN) is used, with migration-related parameters as input features, to predict the estimated scene profile values ​​for the next three time units. For example, if the current migration-related parameter sequence is [0.7, 0.8, 0.6], the TCN output prediction value is 0.75. When the deviation between actual user behavior and the predicted value exceeds the set threshold three times in a row—for example, the actual values ​​are 0.5, 0.4, and 0.3, while the predicted values ​​are 0.6, 0.55, and 0.5, with deviations of 16.7%, 27.3%, and 40%, respectively—it is considered a continuous deviation. At this point, topology reconstruction is triggered, and the spatiotemporal relationship graph is rebuilt based on the latest behavior data. The spatiotemporal feature weights in the multi-dimensional scene profile generation rules are reset, for example, adjusting the time attenuation factor weight from 0.6 to 0.4 and increasing the spatial offset weight from 0.3 to 0.5.

[0145] Finally, a dynamic closed-loop mechanism optimizes recommendation effectiveness. After topology reconstruction, the incremental graph convolution module adjusts the aggregation range based on the newly generated spatiotemporal relationship graph. The dual-path memory unit simultaneously updates high-frequency features and anomaly detection logic, and the time series prediction module is retrained to adapt to the new profile generation rules. For example, the reconstructed scene profile emphasizes spatial consistency across devices, and the ad push strategy prioritizes matching ads with devices in close proximity. If the prediction deviation exceeds the standard again, the system automatically switches to the backup profile generation rule version to ensure recommendation stability.

[0146] In practical applications, in the implementation of a smart home cross-device recommendation system, the incremental graph convolution module, based on a touch trajectory sampling rate of 8 frames per second, detects seven consecutive clicks within the 200-230 pixel area on a 10.1-inch tablet, with a device switching interval of 3 minutes and 12 seconds. Combining a switching frequency threshold of 5 times per hour with a 0.68 calibration value output by adversarial learning, it expands the node aggregation radius from two layers to four layers, covering the kitchen and living room device nodes. The first channel of the dual-channel memory unit synchronizes the 2.8N touch pressure characteristics of the tablet's 200-230 pixel hot zone on the X axis to the scene profiling engine every 20 seconds. The second channel captures 15 abnormal clicks on the mobile delivery page, triggering topology reconstruction. This reduces the weight of the smartwatch-related edge from 0.75 to 0.53 and adds a connection edge to the central control screen. The timing prediction module generates a 25-minute scene estimate in the hardware acceleration unit based on a 0.79 migration parameter. When the user's actual behavior is detected to deviate from the prediction of the 250-280 pixel focus area on the VR device's X-axis three times in a row, the full reconstruction process is initiated: the number of graph convolution layers is reduced from 4 to 2, the time decay factor is reset to 0.58, and new profiling rules are established. After reconstruction, the prediction error is reduced from 41 seconds to 9 seconds, and the aggregation radius is dynamically adjusted to 3 layers to adapt to the new behavior pattern.

[0147] The above steps dynamically adjust the node aggregation range through the incremental graph convolution module, and optimize the local behavior feature extraction capability by combining the device switching frequency and the adversarial learning calibration value. The dual-path memory units operate in parallel, one path solidifies the spatiotemporal distribution characteristics of the high-frequency click area to iterate the portrait generation rules, and the other path monitors the abnormal cross-device behavior path and triggers the weight-based topology reconstruction to achieve adaptive adjustment of the graph structure. The timing prediction module embedded in the hardware acceleration unit relies on the migration correlation parameters to predict the evolution trend of the scene portrait. When the predicted value and the actual behavior trajectory exceed the limit for three consecutive times, the system automatically starts the topology reconstruction process and simultaneously resets the portrait generation rules. Through the closed-loop control of prediction execution and feedback correction, the system can perceive the sudden changes in user behavior patterns in real time, balance the modeling requirements of long-term interest solidification and short-term behavior drift, and ensure the accuracy and robustness of multi-terminal advertising recommendations in dynamic scenarios.

[0148] Figure 3 The present invention provides a structural diagram of an advertising recommendation system integrating user dynamic behavior modeling. Figure 3 As shown, the system includes:

[0149] Acquisition module 31 records the user's cross-device operations on multiple terminal devices, synchronously records the unique identification code of each device, captures the user's high-frequency click area data on the device interface, and marks the time when each operation occurs;

[0150] The calculation module 32 calculates the time interval between user switching between different terminal devices based on the unique identification code and the occurrence time, combines and analyzes the switching time interval with the high-frequency click area data to generate a feature set reflecting the correlation between the user's cross-device behavior, and constructs a spatiotemporal relationship graph containing device nodes and behavior paths based on the feature set;

[0151] Generating module 33, in the edge computing node, performing trajectory similarity calculation on the cross-device operation records through a hardware acceleration unit to generate migration correlation parameters representing the consistency of the user's cross-device behavior;

[0152] Analysis module 34 inputs the spatiotemporal relationship graph into a behavior analysis model, extracts consistent behavior characteristics of users across different terminals, combines the migration association parameters, and integrates device geographic location change data and interface element distribution characteristics to generate a multi-dimensional scene portrait;

[0153] Adjustment module 35, the fusion result of the coherent behavioral characteristics and the multi-dimensional scene portrait, identifies the interest migration pattern in the user's cross-device behavior, filters the advertising content that matches the user's current cross-device scene from the advertising library according to the interest migration pattern, and dynamically adjusts the advertising push strategy.

[0154] Figure 3 The advertising recommendation system integrating user dynamic behavior modeling can be executed Figure 1 The implementation principles and technical effects of the advertising recommendation method that integrates user dynamic behavior modeling described in the illustrated embodiment will not be elaborated on here. The specific manner in which each module and unit performs operations in the advertising recommendation system that integrates user dynamic behavior modeling in the above embodiment has been described in detail in the relevant embodiments of the method and will not be elaborated on here.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An advertising recommendation method integrating user dynamic behavior modeling, characterized in that: include: Obtain user operation records across multiple devices, synchronously record the unique identification code of each device, capture the user's high-frequency click area data on the device interface, and mark the time when each operation occurs; Calculating the time interval between user switching between different terminal devices based on the unique identification code and the occurrence time, combining and analyzing the switching time interval with the high-frequency click area data to generate a feature set reflecting the correlation between the user's cross-device behavior, and constructing a spatiotemporal relationship graph containing device nodes and behavior paths based on the feature set; In the edge computing node, a hardware acceleration unit is used to calculate the trajectory similarity of the cross-device operation records to generate a migration correlation parameter representing the consistency of the user's cross-device behavior; Input the spatiotemporal relationship graph into a behavior analysis model to extract the user's consistent behavior characteristics across different terminals, combine the migration association parameters, and fuse the device's geographic location change data and interface element distribution characteristics to generate a multi-dimensional scene portrait; Based on the fusion results of the coherent behavioral features and the multi-dimensional scene portrait, the interest migration pattern in the user's cross-device behavior is identified, and advertising content that matches the user's current cross-device scene is filtered from the advertising library according to the interest migration pattern, and the advertising push strategy is dynamically adjusted; It is characterized by further comprising: In trajectory similarity calculation, a multimodal adversarial training network is used to perform adversarial learning on device operation records and interface click distribution data, optimizing and generating migration correlation parameters with time-series alignment capabilities. Based on the optimized migration association parameters, a dynamic adjustment model for edge weights is constructed, node connection strengths in the spatiotemporal relationship graph are updated by adversarial learning loss values, and the updated spatiotemporal relationship graph is input into the behavior analysis model to optimize the extraction of coherent behavior features; Establishing a joint learning framework between edge computing nodes, dynamically adjusting the weight parameters of the model output based on the edge weights, and iteratively optimizing the generation rules of the multi-dimensional scene portrait; The advertising push strategy is input into the behavioral inference engine, and an adaptive feedback strategy is generated based on the noise characteristics output by adversarial learning. Based on the adaptive feedback strategy, the parameter update mechanism of the model is adjusted.

2. The method according to claim 1, characterized in that Also includes: Embedding an incremental graph convolution module in the spatiotemporal relationship graph to adjust the adjacent node aggregation range based on the device switching frequency and the calibration value output by adversarial learning; Construct a dual-path memory enhancement unit. The first path stores the optimized spatiotemporal features of high-frequency click areas and synchronizes them with the generation rules of the multi-dimensional scene portrait. The second path records abnormal events in cross-device behavior paths and triggers topology reconstruction based on weight parameters. Integrating a time series prediction module into the hardware acceleration unit to predict the estimated value of the scene portrait in the future time period based on the migration correlation parameters generated by adversarial learning; When the deviation between the estimated value of the scene portrait and the actual behavior of the user exceeds the set value three times in a row, topology reconstruction is triggered, and the generation rules of the multi-dimensional scene portrait are reset according to the reconstructed spatiotemporal relationship graph.

3. The method according to claim 1, characterized in that Based on the unique identification code and the occurrence time, the user's switching time interval between different terminal devices is calculated, the switching time interval is combined with the high-frequency click area data for analysis, and a feature set reflecting the correlation between the user's cross-device behavior is generated. Based on the feature set, a spatiotemporal relationship graph containing device nodes and behavior paths is constructed, including: Matching the user's operation record sequences on multiple terminals according to the unique identification code, and sorting the operation records of each device according to the occurrence time; Traverse the operation record sequence of adjacent terminals and calculate the interval from the last operation time of the user on the current terminal to the first operation time of the next terminal as the switching time interval; Extracting the interface click density distribution corresponding to each terminal in the high-frequency click area data, and counting the high-frequency click area coordinates of the previous and next terminals when each switching time interval occurs; The switching time interval is combined with the coordinates of the high-frequency click area to generate a cross-device behavior correlation feature set. Based on the correlation feature set, multiple device operations in which the user's continuous switching time interval is less than a set value are combined into a hyperedge to construct a spatiotemporal relationship graph with hierarchical edges.

4. The method according to claim 1, wherein The spatiotemporal relationship graph is input into a behavior analysis model to extract the user's consistent behavior characteristics across different terminals. The model combines the migration association parameters with the device's geographic location change data and interface element distribution characteristics to generate a multi-dimensional scene portrait, including: Adding an attribute vector to each node in the spatiotemporal relationship graph, including the physical position change of the corresponding terminal, the screen size, and the shape characteristics of the interface click area; Through the node propagation mechanism in the behavior analysis model, the attribute vectors of adjacent nodes are aggregated along the hierarchical edge direction to generate the behavior code of each node; Converting the migration association parameter into a dynamic scaling coefficient that is inversely associated with the edge weight in the spatiotemporal relationship graph, and adjusting the edge weight in real time according to the dynamic scaling coefficient when nodes are aggregated; Extracting the spatial offset related to the device position in the behavior code, and calculating the cross-device behavior continuity index based on the shape similarity of the click area on the interface; The cross-device behavior continuity indicator is nonlinearly fused with the adjusted edge weights to generate a multi-dimensional scenario portrait that characterizes the consistency of the user's multi-device behavior.

5. The method according to claim 1, characterized in that In the edge computing node, a hardware acceleration unit is used to perform trajectory similarity calculation on the cross-device operation records to generate migration correlation parameters that characterize the consistency of the user's cross-device behavior, including: In the edge computing node, the hardware acceleration unit is used to compare the high-frequency click area numbers on the device interface in the cross-device operation records of any two terminal devices, and the matching position distribution characteristics of continuous matches in the operation records of the two devices are extracted; Based on the matching position distribution characteristics, the hardware acceleration unit calculates a coverage density index of the operation traces between the two devices, where the coverage density index is formed by multiplying the ratio of the sum of the matching segment lengths to the total length of the longer operation record by the dispersion of the matching segment positions; Applying a time decay factor to the coverage density index by the hardware acceleration unit to generate a cross-device trajectory similarity parameter, wherein the time decay factor is inversely proportional to the time interval between the operation records of the two devices; After traversing all device combinations, device pairs whose values ​​of the cross-device trajectory similarity parameters exceed a preset threshold are selected to generate migration association parameters that characterize the consistency of the user's cross-device behavior.

6. The method according to claim 1, characterized in that Based on the fusion results of the coherent behavioral features and the multi-dimensional scenario portrait, the interest migration pattern in the user's cross-device behavior is identified, and advertising content that matches the user's current cross-device scenario is filtered from the advertising library according to the interest migration pattern, and the advertising push strategy is dynamically adjusted, including: Extracting cross-device interaction hotspots from the multi-dimensional scene portrait based on a fusion result of the coherent behavior features and the multi-dimensional scene portrait; Performing a spatiotemporal overlay analysis of the device switching frequency and operation sequence repetition period in the coherent behavior features with the interaction hotspot areas to generate a probability matrix representing the user's cross-device behavior interest migration pattern. The probability matrix is ​​constructed by analyzing the hotspot area transfer paths and time distribution density of operation sequences between different devices; Identifying the core migration path and associated device combinations of the interest migration pattern in the user's cross-device behavior based on the transition probability peaks of the continuous device switching paths in the probability matrix; Based on the core migration path, screening advertising content that matches the usage scenario tag of the associated device combination from the advertising library, and calculating the matching weight between the advertising scenario feature vector and the corresponding migration path in the probability matrix; Dynamically construct an advertising push priority queue, and the sorting weight of the advertising push priority queue is adjusted in real time by the superposition result of the matching weight and the activity of the high-frequency click area of ​​the current device interface. Based on the advertising push priority queue, adjust the advertising push strategy.

7. The method according to claim 3, characterized in that The switching time interval and the coordinates of the high-frequency click area are combined to generate a cross-device behavior correlation feature set. Based on the correlation feature set, multiple device operations in which the user switches continuously for a time interval less than a set value are combined into a hyperedge, and a spatiotemporal relationship graph with hierarchical edges is constructed, including: Combine the switching time interval with the coordinates of the high-frequency click areas of the terminal before and after the switching, calculate the inverse of the switching time interval as the temporal correlation strength value, and calculate the plane distance between the coordinates of the high-frequency click areas of the terminal before and after the switching as the spatial correlation difference value; The temporal correlation strength value and the spatial correlation difference value are fused into a multidimensional feature vector according to a preset ratio, wherein the proportion of the temporal correlation strength value increases as the switching time interval shortens, thereby generating a cross-device behavior correlation feature set; Based on the associated feature set, extract a multi-device operation combination with a time interval less than a set value in the user's continuous device switching events; if the number of associated terminal devices is greater than or equal to three and the continuous switching time intervals are all less than the set threshold, then connect the multi-device operation combination through a hyperedge; Assign a hierarchical weight to each hyperedge, and construct a spatiotemporal relationship graph with hierarchical edges based on the hierarchical weight and connection direction of the hyperedge.

8. The method according to claim 4, characterized in that Converting the migration association parameter into a dynamic scaling coefficient that is inversely associated with the edge weight in the spatiotemporal relationship graph, and adjusting the edge weight in real time according to the dynamic scaling coefficient when nodes are aggregated, including: Converting the migration-associated parameter of each device pair into a dynamic scaling coefficient, wherein the dynamic scaling coefficient is inversely proportional to the value of the migration-associated parameter, and a larger value of the migration-associated parameter corresponds to a smaller dynamic scaling coefficient; Extracting the migration association parameter of the device pair corresponding to the hyperedge according to the generation order and path direction of each hyperedge in the spatiotemporal relationship graph, and performing a product operation on the dynamic scaling coefficient and the path length of the device associated with the hyperedge to generate a hyperedge weight adjustment factor; During the node aggregation process, when the attribute vector is transmitted along the hyperedge direction, the original hierarchical weight of the hyperedge is scaled in real time based on the hyperedge weight adjustment factor, so that the hyperedge hierarchical weight between the device pairs with high values ​​of the migration association parameter is suppressed, and the hyperedge hierarchical weight with low values ​​of the migration association parameter is enhanced.

9. An advertising recommendation system integrating user dynamic behavior modeling, characterized in that: include: The acquisition module obtains the user's cross-device operation records on multiple terminal devices, synchronously records the unique identification code of each device, captures the user's high-frequency click area data on the device interface, and marks the time when each operation occurs; a calculation module, which calculates the time interval between user switching between different terminal devices based on the unique identification code and the occurrence time, combines and analyzes the switching time interval with the high-frequency click area data to generate a feature set reflecting the correlation between the user's cross-device behavior, and constructs a spatiotemporal relationship graph containing device nodes and behavior paths based on the feature set; a generation module, which calculates the trajectory similarity of the cross-device operation records through a hardware acceleration unit in an edge computing node to generate a migration correlation parameter representing the consistency of the user's cross-device behavior; An analysis module inputs the spatiotemporal relationship graph into a behavior analysis model, extracts the user's consistent behavior characteristics across different terminals, combines the migration association parameters, and integrates the device's geographic location change data and interface element distribution characteristics to generate a multi-dimensional scene portrait; An adjustment module, based on the fusion result of the coherent behavioral features and the multi-dimensional scene portrait, identifies the interest migration pattern in the user's cross-device behavior, selects advertising content that matches the user's current cross-device scene from the advertising library according to the interest migration pattern, and dynamically adjusts the advertising push strategy; It is characterized by further comprising: In trajectory similarity calculation, a multimodal adversarial training network is used to perform adversarial learning on device operation records and interface click distribution data, optimizing and generating migration correlation parameters with time-series alignment capabilities. Based on the optimized migration association parameters, a dynamic adjustment model for edge weights is constructed, node connection strengths in the spatiotemporal relationship graph are updated by adversarial learning loss values, and the updated spatiotemporal relationship graph is input into the behavior analysis model to optimize the extraction of coherent behavior features; Establishing a joint learning framework between edge computing nodes, dynamically adjusting the weight parameters of the model output based on the edge weights, and iteratively optimizing the generation rules of the multi-dimensional scene portrait; The advertising push strategy is input into the behavioral inference engine, and an adaptive feedback strategy is generated based on the noise characteristics output by adversarial learning. Based on the adaptive feedback strategy, the parameter update mechanism of the model is adjusted.

Images