A multi-modal data-driven customer portrait intelligent analysis method

Abstract

The application discloses a kind of multi-modal data-driven customer portrait intelligent analysis methods.The method obtains the multi-modal user interaction data in the specified time window of internet advertising service platform, it is parsed into discrete interaction node and converges to generate interaction node universe;The metric value of each interaction node pair in interaction node universe is calculated, and the non-euclidean metric matrix of modal semantic and space-time decay coupling is constructed;In interaction node universe, landmark point set is extracted, and based on non-euclidean metric matrix, witness complex filtering flow evolving with distance threshold is constructed;The rank data of each order homology group of witness complex filtering flow is calculated, and sampling adjustment instruction is generated according to homology group rank data, to feedback control the sampling frequency of multi-modal user interaction data;The birth and death state of topological structure is recorded, and continuous bar chart is generated, and the structured customer portrait is output by distance matching.The application can realize the adaptive adjustment of data acquisition frequency, and improve the precision of internet advertising service platform customer portrait and the click rate and conversion rate of advertisement targeted delivery.

Description

Technical Field

[0001] This invention relates to the field of internet advertising service technology, specifically to a multimodal data-driven intelligent analysis method for customer profiling, used for user intent recognition and targeted advertising. Background Technology

[0002] Intelligent customer profiling analysis is a crucial component of e-commerce platforms and digital marketing. By collecting and mining data on user interactions, the analysis system can build feature models that reflect user preferences, providing support for product recommendations and platform operations.

[0003] In existing technologies, internet advertising service platforms typically collect multimodal user interaction data, including text search records, product image interaction coordinates, and page dwell time. After acquiring this data, the analysis system uses feature extraction algorithms to convert the operation records into corresponding feature vectors. Subsequently, through feature concatenation or weighted fusion, the data is input into the analysis model for intent classification, thereby outputting customer profiles. In actual e-commerce operation scenarios, the accuracy of customer profiles directly determines the conversion rate of product recommendations and the platform's operational efficiency. In actual internet advertising service operation scenarios, advertisers' budgets and user attention resources are very limited. The accuracy of ad targeting directly determines the click-through rate (CTR), conversion rate (CVR), and return on investment (ROI). Existing technologies cannot effectively depict the evolution trend of user intent in browsing, searching, and scrolling sequences, resulting in a discrepancy between the output user profile and the actual advertising audience's intent. This leads to overexposure or underexposure of ads, wasting a large amount of advertising budget on ineffective exposure, affecting the traffic monetization efficiency of internet advertising platforms and the advertising experience for advertisers.

[0004] However, existing customer profiling analysis methods suffer from a weakness in eliminating feature alignment errors when processing multimodal interaction data. User interaction behavior involves multiple modalities, including textual, visual, and temporal modalities, and semantic differences and temporal misalignments exist between these modal data. Existing analysis systems typically map features from different modalities to a unified real number space before calculating distances, resulting in insufficient quantification granularity for semantic differences between modalities and a lack of explicit quantification methods for the temporal decay effect introduced by the time difference of event occurrences, thus limiting the accuracy of heterogeneous data fusion. The aforementioned data processing methods are insufficient to effectively reduce feature alignment biases between heterogeneous data and lack the ability to characterize the spatiotemporal evolution of user behavior, leading to biased customer profiles in the output. Summary of the Invention

[0005] To address the problems mentioned in the background, this invention provides a multimodal data-driven intelligent customer profiling analysis method, comprising the following steps:

[0006] Step S100: Obtain multimodal user interaction data within a specified time window from the Internet advertising service platform, parse the multimodal user interaction data into discrete interaction nodes, and aggregate them to generate a complete set of interaction nodes;

[0007] Step S200: Calculate the metric values ​​of each pair of interactive nodes in the complete set of interactive nodes, and construct a non-Euclidean metric matrix that couples modal semantics with spatiotemporal decay.

[0008] Step S300: Extract a set of landmark points from the complete set of interactive nodes, and construct a witness complex filter flow that evolves with the distance threshold based on the non-Euclidean metric matrix;

[0009] Step S400: Calculate the homology group rank data of each order of the witness complex filter stream, and generate a sampling adjustment command based on the homology group rank data to control the sampling frequency of the multimodal user interaction data.

[0010] Step S500: Record the birth and death states of the topology, generate a continuous bar chart, and output a structured customer profile through distance matching;

[0011] Preferably, the method further includes, after step S500: sending the output structured customer profile to the advertising targeting module of the Internet advertising service platform, wherein the advertising targeting module calls a preset advertising content matching engine based on the intent classification tags bound in the customer profile, retrieves advertising creative content with a matching degree higher than a preset threshold from the candidate advertising material library, generates personalized advertising content, and delivers it to the advertising display interface of the user terminal device.

[0012] Preferably, the multimodal user interaction data comes from the user terminal devices connected to the Internet advertising service platform, including the user's text search history on the advertising landing page, the click and swipe interaction coordinates of the advertising material images, and the user's dwell time on the advertising page;

[0013] Preferably, the process of acquiring multimodal user interaction data and generating a complete set of interaction nodes includes: listening to interaction events of user terminal devices, capturing operation behaviors including text search records, product image interaction coordinates, and page dwell time, and attaching event timestamps to the operation behaviors; performing vector space mapping to segment continuous operation behavior data into discrete interaction nodes containing single-dimensional modal data; for discrete interaction nodes with different modal data, respectively calling word embedding networks, position encoding networks, and normalization functions to convert discrete interaction nodes into corresponding text feature vectors, visual feature vectors, and time feature scalars; uniformly mapping the extracted feature vectors and time feature scalars to a set real number vector space to generate initial feature vectors with consistent dimensions, and binding them with event timestamps and modal attribute identifiers to generate a complete set of interaction nodes.

[0014] Preferably, the process of constructing the non-Euclidean metric matrix includes: extracting the initial feature vectors of the interactive nodes and calculating their squared Euclidean distance in the real vector space; reading the cross-modal transition relation matrix and determining the corresponding modal semantic transition penalty coefficient based on the similarity and difference of the modal attribute identifiers of the interactive nodes; calculating the absolute time difference of the event occurrence timestamp identifiers of the interactive nodes, and generating an exponential decay factor with the natural constant as the base and the product of the time decay constant and the absolute time difference as the exponent, in conjunction with the set time decay constant; multiplying the calculated squared Euclidean distance, modal semantic transition penalty coefficient, and exponential decay factor to obtain the metric value, and summing the metric values ​​of all interactive node pairs and filling them into the corresponding row and column coordinate positions in the two-dimensional square matrix to generate the non-Euclidean metric matrix.

[0015] Preferably, the process of extracting the landmark set and constructing the witness complex filter flow includes: determining the target landmark number limit by combining available memory and computational latency threshold; selecting the discrete interaction node with the largest L2 norm of the feature vector as the first landmark; performing an iterative extraction operation, traversing and querying the metric values ​​from unselected nodes to known landmarks, extracting the minimum metric value of each node, locking the node corresponding to the maximum value among all minimum metric values ​​as the new landmark, until the number limit is reached, and assigning the remaining nodes to the witness set; based on the non-Euclidean metric matrix, extracting the minimum metric value from the witness point to the landmark set as the baseline distance; controlling the topology filtering parameters to continuously increase, determining whether the simplex formed by the landmarks satisfies the set relaxed witness conditions, including the simplex that meets the conditions into the witness complex, recording the order of simplex generation to generate a witness complex filter flow containing multiple simplexes.

[0016] Preferably, the process of controlling the sampling frequency of multimodal user interaction data based on the feedback of homology group rank data includes: constructing an edge matrix based on the generation order of the simplex and vertex dimension information in the witness complex filtering flow; performing matrix canonical form elimination to extract homology group rank data of each order; locking the zero-dimensional Betti number representing the number of connected components and the one-dimensional Betti number representing the number of one-dimensional holes, and packaging them into a topological invariant signal; performing a weighted average operation on the historical Betti number and the current Betti number within a specified sliding time window, comparing them with the preset connectivity determination threshold and hole determination threshold, and outputting a trigger state bit; calling the sampling frequency dynamic update equation based on the activation status of the trigger state bit, and calculating the target sampling frequency of each mode in the next time window by combining the modal polarity control factor and sensitivity gain coefficient; and encapsulating the generated target sampling frequency into a sampling configuration update parameter package for distribution, thereby adjusting the event listening time interval of the data acquisition layer for each modal operation behavior.

[0017] Preferably, the process of outputting a structured customer profile includes: extracting the topological feature occurrence distance parameter and disappearance distance parameter corresponding to the witness complex filter flow boundary evolution data, and pairing them to form a lifecycle interval; classifying the lifecycle interval according to the homology group dimension attribute to which the topological feature belongs, and aggregating to generate structured persistent bar chart data in the form of a multidimensional array; calling the standard intent homology class template library that is internally persistently stored and bound with intent classification labels, and establishing a bijective matching relationship between the triplet data structure contained in the target feature set and the reference feature set; calculating the infimum of the sum of absolute distances between the endpoints of the triplets in the target feature set and the endpoints of the matched triplets in the reference feature set under all established matching relationships, and setting it as the Wasserstein distance value between the two; locking the intent classification label of the template corresponding to the first smallest Wasserstein distance value after ascending sorting, and associating it with the user terminal device identifier to generate a structured macro customer profile.

[0018] The present invention, by adopting the above technical solution, can bring the following beneficial effects:

[0019] 1. This invention extracts the feature vectors, timestamps, and modal attribute identifiers of interactive nodes, determines the modal semantic penalty coefficient by combining the cross-modal transformation matrix, and introduces a time decay constant to calculate the exponential decay factor. Then, based on the squared Euclidean distance, the above factors are multiplied to construct a non-Euclidean metric matrix. This realizes the unified mapping of multimodal interactive data with semantic differences and temporal misalignment to a non-Euclidean space, thereby improving the accuracy of internet advertising service platforms in recognizing users' real-time intentions, providing highly reliable audience preference input for advertising targeting systems, reducing invalid ad exposures caused by profile bias, and improving ad click-through rates.

[0020] 2. This invention extracts the homology group rank data of the witness complex filter flow, compares the smoothed zero-dimensional and one-dimensional Betty numbers with the judgment threshold to output the trigger state bit, and then calls the update equation to calculate the target sampling frequency of each modality in the next time window and converts it into a hardware interrupt adjustment instruction to be issued. This realizes the dynamic adjustment of the time interval of data collection of each modality according to the evolution trend of the topological features of interactive behavior. In this way, while ensuring the capture density of core business data (such as text search logs), it suppresses the interference of high-noise visual data on downstream advertising targeting models, and improves the cost-effectiveness of data collection and advertising response and delivery efficiency of Internet advertising platforms.

[0021] 3. This invention determines the appearance and disappearance distance parameters of topological features by extracting the boundary evolution data of the homology class generator, aggregates the life cycle intervals according to dimensional attributes to generate continuous bar chart data in the form of a multidimensional array, and then establishes a bijective matching relationship between the target feature set and the reference feature set. The infimum of the sum of the absolute distances of the endpoints is calculated as the distance metric, thereby realizing the locking of the minimum matching distance template with intent classification labels based on the cost of topological deformation. This improves the output stability of structured customer profiles in Internet advertising scenarios, enabling the advertising targeting module to perform accurate advertising content matching based on highly stable intent labels, thereby improving advertising conversion rate (CVR) and advertising ROI. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of the multimodal data-driven intelligent customer profiling analysis method of the present invention.

[0023] Figure 2 This is the overall architecture diagram of the multimodal data-driven intelligent customer profiling analysis system of the present invention;

[0024] Figure 3 This is a detailed flowchart of the steps for generating the complete set of data acquisition and interaction nodes in this invention;

[0025] Figure 4 This is a detailed flowchart of the non-Euclidean metric matrix construction steps of the present invention;

[0026] Figure 5 This is a detailed flowchart of the steps for extracting the landmark set and dividing the witness points in this invention;

[0027] Figure 6 This is a detailed flowchart of the steps for calculating the homology group rank data and controlling the sampling frequency in this invention.

[0028] Figure 7 This is a detailed flowchart of the customer profile output steps of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example 1

[0031] As an optional implementation method, this embodiment combines the appendix. Figure 1This paper provides an overall description of the multimodal data-driven intelligent analysis method for customer profiling from a methodological and process perspective. It is understood that there is a corresponding execution relationship between the method steps described in this embodiment and the system modules described in Embodiment 2 below. Those skilled in the art can make adaptive adjustments to the execution order of each step according to the actual business scenario.

[0032] Please refer to the appendix. Figure 1 This invention provides a multimodal data-driven intelligent analysis method for customer profiling, comprising the following steps:

[0033] Step S100: Obtain multimodal user interaction data within a specified time window from the Internet advertising service platform, parse the multimodal user interaction data into discrete interaction nodes, and aggregate them to generate a complete set of interaction nodes;

[0034] Multimodal user interaction data includes text search records, product image interaction coordinates, and page lifecycle timestamps. The data acquisition layer 10 continuously monitors the operation events of the user terminal device according to the preset initial sampling frequency. The data acquisition layer 10 performs field separation and data cleaning on the received operation events. The tensor computing server 20 calls the pre-trained word embedding network and image feature extraction network to map discrete interaction nodes into initial feature vectors of fixed dimensions. The tensor computing server 20 synchronously records the occurrence timestamp and modal attribute identifier corresponding to each interaction node.

[0035] Step S200: Calculate the metric values ​​of each pair of interactive nodes in the complete set of interactive nodes, and construct a non-Euclidean metric matrix that couples modal semantics with spatiotemporal decay.

[0036] Tensor operation module 22 traverses the entire set of interaction nodes, extracts the initial feature vectors of any two interaction nodes, searches for the modality semantic transition penalty coefficient in the preset knowledge graph based on the modality attribute identifiers of the two interaction nodes, calculates the absolute time difference of the timestamps of the two interaction nodes, and calculates the exponential decay factor based on the preset time decay constant. Tensor operation module 22 multiplies the Euclidean distance of the initial feature vectors, the modality semantic transition penalty coefficient, and the exponential decay factor to obtain the metric value in the non-Euclidean space. Tensor operation module 22 fills the metric values ​​of all interaction node pairs into a two-dimensional matrix and outputs the non-Euclidean metric matrix.

[0037] Step S300: Extract the set of landmarks from the entire set of interactive nodes, and construct a witness complex filter flow that evolves with the distance threshold based on the non-Euclidean metric matrix;

[0038] The topology homology analysis cluster 30 runs the maximum-minimum distance algorithm to downsample the entire set of interaction nodes. The topology homology analysis cluster 30 calculates the L2 norm of the eigenvectors of each discrete interaction node in the entire set of interaction nodes, selects the discrete interaction node with the largest L2 norm as the initial landmark, and iteratively calculates the minimum metric value from the remaining nodes to the selected node set. The topology homology analysis cluster 30 selects the node that maximizes the minimum metric value as the new landmark, until a preset limit is reached to form a landmark set. The topology homology analysis cluster 30 sets the remaining nodes in the entire set of interaction nodes as the witness set. The topology homology analysis cluster 30 sets a continuously increasing distance threshold parameter. The topology homology analysis cluster 30 determines whether the metric values ​​of the landmark and the witness in the non-Euclidean metric matrix satisfy the relaxation judgment condition including the distance threshold parameter. The topology homology analysis cluster 30 connects the landmarks that satisfy the judgment condition to generate a witness complex filter flow containing multiple simplexes.

[0039] Step S400: Calculate the homology group rank data of each order of the witness complex filter flow, and generate sampling adjustment instructions based on the homology group rank data to control the sampling frequency of multimodal user interaction data.

[0040] The topology coherence analysis cluster 30 performs matrix standard form elimination on the simplex structure contained in the witness complex filter flow. The topology coherence analysis cluster 30 extracts the topology invariant signal characterizing the system state. The topology invariant signal contains 0-dimensional Betti numbers and 1-dimensional Betti numbers. The sampling controller 40 receives the 0-dimensional Betti numbers and 1-dimensional Betti numbers. The sampling controller 40 multiplies the 0-dimensional Betti numbers and 1-dimensional Betti numbers by a set sensitivity adjustment gain coefficient and calculates the frequency update variables corresponding to different modal modules by combining them with a preset modal polarity control factor. The sampling controller 40 adds the frequency update variables corresponding to different modal modules to the initial sampling frequency to generate the target sampling frequency. The sampling controller 40 encapsulates the target sampling frequency into a sampling configuration update parameter package and sends it to the data acquisition layer 10. The data acquisition layer 10 parses the parameter package and adjusts the event listening time interval of each front-end data acquisition interface according to the target sampling period value.

[0041] Step S500: Record the birth and death states of the topology, generate a continuous bar chart, and output a structured customer profile through distance matching.

[0042] The topology coherence analysis cluster 30 records the occurrence distance parameters and disappearance distance parameters of specific topology features during the distance threshold evolution process. The topology coherence analysis cluster 30 maps the occurrence distance parameters and disappearance distance parameters into persistent bar chart structure data. The profile database 50 calls the pre-set standard intent coherence class template library. The profile database 50 calculates the Wasserstein distance value between the persistent bar chart structure data of the current session and the data of each template library. The profile database 50 locks the target template according to the minimum matching distance principle. The profile database 50 extracts the intent classification label associated with the target template, persists the intent classification label as the final customer profile data, and distributes it to the business end.

[0043] As an optional implementation, after the structured customer profile is output in step S500, the profile database 50 writes the customer profile data into the advertising business distribution queue. The downstream advertising targeting module reads the customer profile in the advertising business distribution queue, calls the corresponding advertising content matching template according to the intent classification tag, retrieves advertising creative content with a matching degree higher than a preset threshold from the candidate advertising material library, generates personalized advertising content, and delivers it to the advertising display interface of the user's terminal device.

[0044] Example 2

[0045] Based on the above embodiment one, as an optional implementation method, the above multimodal data-driven intelligent customer profiling analysis method is implemented based on a multimodal data-driven intelligent customer profiling analysis system. This embodiment is combined with the appendix. Figure 2 This document provides an exemplary description of the overall architecture and interactive relationships of the various functional modules of a multimodal data-driven intelligent customer profiling analysis system, so that those skilled in the art can more clearly understand the underlying system support logic of each method step in Embodiment 1.

[0046] See attached document Figure 2 , Figure 2 This is an overall architecture diagram of a multimodal data-driven intelligent customer profiling analysis system according to an embodiment of the present invention. The multimodal data-driven intelligent customer profiling analysis system specifically includes a data acquisition layer 10, a tensor calculation server 20, a topology coherence analysis cluster 30, a sampling controller 40, and a profiling database 50. The above components interact with each other through a communication bus or network interface to achieve the coordination of front-end multimodal data stream acquisition and back-end topology calculation.

[0047] The data acquisition layer 10 establishes a communication connection with external user terminal devices through a network interface to acquire multimodal user interaction data during the operation of the e-commerce platform. The subordinate feature implementation methods under the jurisdiction of the data acquisition layer 10 include a log collection and embedding program deployed on the front-end page, a screen touch sensor status reading module of the mobile terminal, and a page rendering timer. Using the above programs and modules, the data acquisition layer 10 directly acquires the text search string, product image interaction coordinates, and start and end timestamps of the page lifecycle of the client within a specified time window.

[0048] Tensor computing server 20 receives raw interactive logs transmitted by data acquisition layer 10. Tensor computing server 20 is configured with vectorization encoding module 21 and tensor operation module 22. Vectorization encoding module 21 performs data cleaning and local feature extraction on raw logs of different modalities, mapping the raw logs into high-dimensional real number vectors of a unified dimension. Tensor operation module 22 calculates and generates a non-Euclidean metric matrix that couples modal semantics and spatiotemporal decay based on the extracted feature vectors, the timestamps of the events, and the modal attribute identifiers of the records.

[0049] The topological homology analysis cluster 30 is used to receive the above metric matrix and perform algebraic topological structure calculations. The topological homology analysis cluster 30 is logically divided into a downsampling execution sub-node 31 and a simple complex construction sub-node 32. The downsampling execution sub-node 31 runs the maximum-minimum distance algorithm to partition the data point set. The simple complex construction sub-node 32 uses the partitioned data point set to perform homology algebra operations to generate a multi-dimensional simple complex structure and runs the matrix canonical form elimination algorithm to output the rank data of the continuous homology group.

[0050] The sampling controller 40 establishes a closed-loop communication connection with the topology coherence analysis cluster 30 and the data acquisition layer 10. The sampling controller 40 receives the rank data output by the topology coherence analysis cluster 30 and converts the rank data into sampling adjustment instructions according to the internally integrated threshold judgment logic. The sampling controller 40 sends the sampling configuration update parameter package to the sampling interval configuration item of the corresponding modal data acquisition interface in the data acquisition layer 10, changing the event listening time interval or triggering condition of the front-end data acquisition interface.

[0051] The profile database 50 stores pre-labeled standard intent coherence class template data. The profile database 50 receives the current session topology feature data generated by the topology coherence analysis cluster 30, and uses the built-in metric space distance matching module to output the final classification label and persist it.

[0052] After the profile database 50 associates the persistently stored customer profiles with the current user terminal device identifier, it provides a profile query interface to the advertising targeting engine of the Internet advertising service platform. The advertising targeting engine retrieves advertising creative content with a matching degree higher than a preset threshold from the candidate advertising material library based on the intent classification tags in the customer profile, generates personalized advertising content, and renders it to the advertising display interface of the user terminal device.

[0053] Example 3

[0054] Based on the above embodiments one and two, this embodiment, as a preferred implementation, combines the appendix... Figure 3 To be continued Figure 5 The detailed execution process of steps S100 to S300 in Embodiment 1 will be further explained below. It is understood that steps S110 to S350 below are further refinements of steps S100 to S300. Those skilled in the art can use existing software engineering methods to modularly encapsulate the following steps.

[0055] As a preferred embodiment, the following steps S110 to S150 are a further refinement of step S100 in Embodiment 1, which describes "acquiring multimodal user interaction data within a specified time window of the e-commerce platform, parsing the multimodal user interaction data into discrete interaction nodes, and aggregating them to generate a complete set of interaction nodes." These steps detail the process of collecting, cleaning, mapping vector space, and generating a complete set of interaction nodes for multimodal heterogeneous interaction data, specifically including the following steps:

[0056] In step S110, the data acquisition layer 10 listens for interactive events of the user terminal device according to the set trigger conditions. The data acquisition layer 10 loads the log collection script in the front-end running environment of the user terminal device. The log collection script captures the operation behavior that occurs on the client interface. The specific data types of the operation behavior are divided into text search records, product image interaction coordinates and page dwell time. The data acquisition layer 10 reads the system clock of the device operating system and adds an event occurrence timestamp to each captured operation behavior.

[0057] In step S120, the data acquisition layer 10 encapsulates the captured operation behavior data and event timestamp identifier into a data packet according to the set reporting cycle. The data acquisition layer 10 sends the data packet to the tensor calculation server 20 through the network communication interface. The tensor calculation server 20 performs unpacking and field parsing operations on the data packet. The tensor calculation server 20 verifies the integrity of the parsed fields and discards abnormal data packets containing null values ​​or illegal characters.

[0058] In step S130, the vectorization encoding module 21 receives the verified operation behavior data and performs vector space mapping according to the data type identifier. The vectorization encoding module 21 divides the continuous operation behavior data into independent discrete interaction nodes. Each discrete interaction node contains single-dimensional modal data. The modal attribute identifiers of the discrete interaction nodes are divided into text modality, visual modality and temporal modality.

[0059] In step S140, the vectorization encoding module 21 calls the set feature encoder to extract local features from the discrete interaction nodes. For discrete interaction nodes containing text search records, the vectorization encoding module 21 calls the word embedding network to convert the text search records into text feature vectors. For discrete interaction nodes containing product image interaction coordinates, the vectorization encoding module 21 calls the position encoding network to output visual feature vectors. For discrete interaction nodes containing time modalities such as page dwell time, the vectorization encoding module 21 calls the normalization function to convert the page dwell time into a fixed-dimensional time feature scalar.

[0060] In step S150, the vectorization encoding module 21 maps the extracted text feature vector, visual feature vector and time feature scalar to a set real number vector space. The vectorization encoding module 21 performs an alignment and truncation operation on the mapped vector dimensions to generate an initial feature vector with consistent dimensions. The vectorization encoding module 21 binds the initial feature vector, the event occurrence timestamp identifier and the modal attribute identifier to output a complete set of interactive nodes containing multi-source heterogeneous information.

[0061] As a preferred embodiment, steps S210 to S260 below are a further refinement of step S200 in Embodiment 1, which describes "calculating the metric values ​​of each pair of interactive nodes in the complete set of interactive nodes and constructing a non-Euclidean metric matrix coupled with modal semantics and spatiotemporal decay." These steps detail the factorial calculation process for each element in the non-Euclidean metric matrix, and the specific implementation process includes the following steps:

[0062] In step S210, the tensor operation module 22 traverses the entire set of interaction nodes, extracts the initial feature vector, event occurrence timestamp identifier, and modal attribute identifier corresponding to any two interaction nodes, and loads the extracted initial feature vector, event occurrence timestamp identifier, and modal attribute identifier into the memory work area.

[0063] Step S220: Tensor operation module 22 calculates the squared Euclidean distance between the initial feature vectors of the two interactive nodes in the real vector space. The squared Euclidean distance is used to characterize the degree of difference between the two interactive nodes in the local semantic space.

[0064] In step S230, the tensor operation module 22 determines the modality semantic transition penalty coefficient based on the extracted modality attribute identifier. The tensor operation module 22 reads the preset cross-modality transition relationship matrix. The cross-modality transition relationship matrix stores the cross-cross cost values ​​of different modality data pairs in business logic. The cross-cross cost values ​​in the cross-modality transition relationship matrix are pre-calibrated by statistically analyzing the frequency of different modal transitions in the historical global log. The lower the transition frequency, the larger the corresponding cross-cross cost value. When the modality attribute identifiers of two interactive nodes are the same, the tensor operation module 22 assigns the modality semantic transition penalty coefficient to the set benchmark value. When the modality attribute identifiers of two interactive nodes are different, the tensor operation module 22 extracts the corresponding cross-cross cost value greater than the benchmark value based on the cross-modality transition relationship matrix and assigns it to the modality semantic transition penalty coefficient.

[0065] In step S240, the tensor operation module 22 calculates the absolute time difference between the event timestamps of the two interactive nodes. The tensor operation module 22 calls the system's preset time decay constant, which is a real number greater than zero. The tensor operation module 22 uses the natural constant as the base and the product of the time decay constant and the absolute time difference as the exponent to calculate and generate the exponential decay factor.

[0066] In step S250, the tensor operation module 22 multiplies the squared Euclidean distance, the modality semantic transformation penalty coefficient, and the exponential decay factor to obtain the metric value in non-Euclidean space. The tensor operation module 22 then calculates the metric value using a predefined metric function, the expression of which is:

[0067]

[0068] In the formula, and Represents any two discrete interaction nodes in the universal set of interaction nodes; and Representing discrete interaction nodes and Corresponding modal attribute identifier;

[0069] Represents discrete interaction nodes and The values ​​between them; This represents the modality semantic transition penalty coefficient; , Representing discrete interaction nodes and The corresponding initial feature vector; Represents the initial eigenvector and The square Euclidean distance between them; and Representing discrete interaction nodes and The corresponding event timestamp identifier; This represents the set time decay constant, and ;

[0070] It should be noted that when When it approaches zero, the modality semantic transition penalty coefficient Preserve the preset modal difference baseline value to ensure that the measurement values ​​between different modal data nodes are not lower than the baseline value, so as to maintain the distinguishability of modal heterogeneity;

[0071] In step S260, the tensor operation module 22 summarizes the metric values ​​of all interactive node pairs in the complete set of interactive nodes. The tensor operation module 22 constructs a two-dimensional square matrix with a dimension equal to the total number of interactive nodes. The tensor operation module 22 fills the summarized metric values ​​into the corresponding row and column coordinate positions in the two-dimensional square matrix to generate a non-Euclidean metric matrix. The tensor operation module 22 outputs the non-Euclidean metric matrix to the topology homology analysis cluster 30.

[0072] As a preferred embodiment, the following steps S310 to S350 are a further refinement of step S300 in Embodiment 1, which describes "extracting a set of landmark points from the entire set of interactive nodes and constructing a witness complex filtering flow that evolves with a distance threshold based on a non-Euclidean metric matrix." These steps detail the process of extracting downsampled landmark points and partitioning the witness point set based on the maximum-minimum distance algorithm. The specific implementation process includes the following steps:

[0073] In step S310, the downsampling execution sub-node 31 receives the non-Euclidean metric matrix and the complete set of interactive nodes output by the tensor operation module 22. The downsampling execution sub-node 31 reads the current available memory balance and the preset maximum computational latency threshold of the topology homology analysis cluster 30. The downsampling execution sub-node 31 calculates the downsampling ratio parameter based on the available memory balance and the maximum computational latency threshold. The downsampling execution sub-node 31 calculates the target landmark number limit by combining the total number of nodes in the complete set of interactive nodes with the downsampling ratio parameter.

[0074] Step S320: The downsampling execution sub-node 31 selects an initial landmark point in the entire set of interaction nodes. The downsampling execution sub-node 31 calculates the L2 norm of the feature vector of each discrete interaction node in the entire set of interaction nodes. The downsampling execution sub-node 31 selects the discrete interaction node with the largest L2 norm of the feature vector as the first landmark point and adds the first landmark point to the initialized landmark point set.

[0075] Step S330: Downsampling execution sub-node 31 performs iterative extraction. In each iteration, downsampling execution sub-node 31 traverses the remaining nodes in the entire set of interaction nodes that were not selected into the landmark set. Based on the non-Euclidean metric matrix, downsampling execution sub-node 31 queries the metric values ​​of each remaining node to all known landmarks in the current landmark set. Downsampling execution sub-node 31 extracts the minimum metric value corresponding to each remaining node. Downsampling execution sub-node 31 locks the maximum value among all extracted minimum metric values ​​and identifies the remaining node corresponding to the locked maximum value as the new landmark. Downsampling execution sub-node 31 uses an iterative update function to determine the new landmark. The calculation expression of the iterative update function is:

[0076]

[0077] In the formula, The variable representing the current iteration round being executed, and , Indicates the limit on the number of target landmarks; Represents the complete set of interactive nodes; Indicates the preceding The set of known landmarks obtained from the next iteration; This indicates that all interactive nodes are excluded from the set. The remaining nodes that were not selected; Represents the set of known landmarks Landmarks in the middle; Indicates the remaining nodes that were not selected. Landmarks The metric value in a non-Euclidean metric matrix; Indicates the first The new landmark point determined by the next iteration operation;

[0078] In step S340, the downsampling execution sub-node 31 adds the determined new landmarks to the landmark set. The downsampling execution sub-node 31 determines whether the number of elements in the landmark set has reached the target landmark number limit. If the target landmark number limit has not been reached, the control flow of the downsampling execution sub-node 31 returns to execute the iterative extraction operation. If the target landmark number limit has been reached, the downsampling execution sub-node 31 stops the iteration.

[0079] In step S350, the downsampling execution sub-node 31 outputs the final generated set of landmark points. The downsampling execution sub-node 31 divides the remaining nodes in the entire set of interactive nodes, excluding the set of landmark points, into the set of witness points. The downsampling execution sub-node 31 transmits the set of landmark points and the set of witness points to the simple complex construction sub-node 32.

[0080] Example 4

[0081] Based on the above embodiments one to three, as a preferred implementation method, this embodiment combines the appendix... Figure 6 To be continued Figure 7 The following further explains the detailed execution process of steps S400 to S500 in Embodiment 1 and the topology homology analysis cluster 30, sampling controller 40 and portrait database 50 in Embodiment 2. It can be understood that steps S410 to S850 below are further refinements of steps S400 to S500.

[0082] As a preferred embodiment, the following steps S410 to S450 are a further refinement of the "calculation of homology group rank data of each order of the witness complex filter flow" described in step S400 of Embodiment 1, used to explain in detail the witness complex construction and Betti number calculation process, specifically including the following steps:

[0083] In step S410, the simple complex construction sub-node 32 receives the set of landmark points and the set of witness points. The simple complex construction sub-node 32 traverses each witness point in the set of witness points. Based on the non-Euclidean metric matrix, the simple complex construction sub-node 32 extracts the metric values ​​from each witness point to all landmark points in the set of landmark points. The simple complex construction sub-node 32 extracts the minimum value among all extracted metric values ​​and defines the minimum value as the baseline distance of the corresponding witness point.

[0084] In step S420, the simplex construction sub-node 32 sets the initial values ​​of the topology filtering parameters to zero, a step size constant, and a maximum evolution threshold. The simplex construction sub-node 32 controls the topology filtering parameters to continuously increase according to the step size constant until the maximum evolution threshold is reached. During the increase of the topology filtering parameters, for any simplex composed of landmarks from the landmark point set, the simplex construction sub-node 32 determines whether the corresponding simplex satisfies the relaxation witness condition. The formula for determining the relaxation witness condition is:

[0085] For simplex Each vertex ,in ,satisfy ;

[0086] In the formula, The index variable represents the vertex of the simplex, and ; The simplex contains the first... Each landmark vertex; Represents any witness point in the set of witness points; Represents the vertex of the landmark. With Witness Point The metric value in a non-Euclidean metric matrix; Indicates witness point The corresponding baseline distance; Indicates the topology filtering parameters;

[0087] Step S430: When there is at least one witness point that makes the simplex satisfy the relaxed witness condition, the simplex construction sub-node 32 will include the simplex that satisfies the condition into the witness complex corresponding to the current topology filtering parameter. As the topology filtering parameter increases continuously, the simplex construction sub-node 32 records the order of simplex generation and generates a witness complex filtering flow containing multiple simplexes.

[0088] In step S440, the simplex construction sub-node 32 constructs an edge matrix according to the generation order of simplexes in the witness simplex filter flow. The simplex construction sub-node 32 extracts the vertex dimension information of each simplex in the witness simplex filter flow. The simplex construction sub-node 32 uses the q-dimensional simplex as the column vector of the edge matrix and the (q-1)-dimensional simplex as the row vector of the edge matrix. If the (q-1)-dimensional simplex is a face of the q-dimensional simplex, the simplex construction sub-node 32 records the corresponding element value in the edge matrix as 1; otherwise, it records the corresponding element value as 0. The simplex construction sub-node 32 performs matrix normal form elimination operation on the edge matrix.

[0089] In step S450, the simple complex construction sub-node 32 extracts the rank data of each order homology group based on the canonical form matrix after elimination. The simple complex construction sub-node 32 locks the 0-dimensional Betti number corresponding to the 0-order homology group and the 1-dimensional Betti number corresponding to the 1-order homology group. The 0-dimensional Betti number is used to characterize the number of connected components in the non-Euclidean space of the total set of interactive nodes, and the 1-dimensional Betti number is used to characterize the number of one-dimensional holes in the non-Euclidean space of the total set of interactive nodes. The simple complex construction sub-node 32 packages the 0-dimensional Betti number and the 1-dimensional Betti number into a topological invariant signal and sends the topological invariant signal to the sampling controller 40.

[0090] As a preferred embodiment, the following steps S510 to S550 are a further refinement of the "generating sampling adjustment instructions based on homology group rank data" described in step S400 of Embodiment 1, used to explain in detail the process of generating topological feature feedback signals, determining thresholds, and dynamically updating sampling frequencies, specifically including the following steps:

[0091] In step S510, the sampling controller 40 receives the topological invariant signal sent by the topological coherence analysis cluster 30. The topological invariant signal includes the 0-dimensional Betty number and the 1-dimensional Betty number within the current time window. The sampling controller 40 maps the 0-dimensional Betty number to the interest point dispersion index of the current set of interactive nodes and maps the 1-dimensional Betty number to the topological complexity index of the interaction behavior pattern. Based on the combined change trend of the interest point dispersion index and the topological complexity index, a trigger state bit representing the client's browsing behavior tending to be dispersed or hesitant is generated.

[0092] In step S520, the sampling controller 40 reads the topological invariant signal within the historical time window cached in the buffer area, the sampling controller 40 extracts multiple historical 0-dimensional Betti numbers and historical 1-dimensional Betti numbers within the set sliding time window, the sampling controller 40 performs a weighted average operation on the extracted historical values ​​and the current values, and outputs the smoothed 0-dimensional Betti numbers and smoothed 1-dimensional Betti numbers.

[0093] In step S530, the sampling controller 40 extracts a preset set of judgment thresholds, which includes a connectivity judgment threshold and a hole judgment threshold. The sampling controller 40 compares the smoothed 0-dimensional Betty number with the connectivity judgment threshold and the smoothed 1-dimensional Betty number with the hole judgment threshold. The sampling controller 40 outputs discrete trigger state bits based on the comparison results. The trigger state bits include a dispersed state bit representing the divergence of browsing behavior and a hesitant state bit representing the hesitation of browsing behavior.

[0094] In step S540, the sampling controller 40 generates a control feedback signal based on the activation status of the trigger state bit. The sampling controller 40 calls the preset sampling frequency dynamic update equation to calculate the target sampling frequency of the target modal sensor in the next time window. The calculation expression of the sampling frequency dynamic update equation is as follows:

[0095]

[0096] In the formula, Representing modes In the next time window The target sampling frequency within; Representing modes In the current time window The initial sampling frequency within; Indicates modality The set modal polarity control factor; This represents the smoothed 1-dimensional Betty number within the current time window; This represents the smoothed 0-Vibetti number within the current time window; This represents the sensitivity gain coefficient set for a 1-dimensional Betty number; This represents the sensitivity gain coefficient set for the 0 Vibetti number; Indicates modality The preset minimum sampling frequency threshold, and ;

[0097] in, and The dimensions are consistent with the sampling frequency, ensuring that the numerical calculation results of the equation dynamically updated by the sampling frequency are consistent with... They have the same physical dimensions;

[0098] For the text search modality, the modality polarity control factor is set to a positive number; for the product image sliding modality, the modality polarity control factor is set to a negative number. When the smoothed 1-dimensional Betty number is greater than the hole determination threshold, triggering the hesitation state, the sampling controller 40 uses the above equation to calculate the increased text modality target sampling frequency and the decreased visual modality target sampling frequency, and generates a first feedback signal. The first feedback signal instructs the data acquisition layer 10 to shorten the text search log acquisition time interval and simultaneously extend the product image sliding acquisition time interval. When the smoothed 0-dimensional Betty number is greater than the connectivity determination threshold, triggering the dispersion state, the sampling controller 40 uses the above equation to calculate the decreased text modality target sampling frequency and the increased visual modality target sampling frequency, and generates a second feedback signal. The second feedback signal instructs the data acquisition layer 10 to extend the text search log acquisition time interval and simultaneously shorten the product image sliding acquisition time interval.

[0099] In step S550, the sampling controller 40 encapsulates the generated first feedback signal or second feedback signal into a sampling configuration update parameter package. The sampling controller 40 sends the sampling configuration update parameter package to the data acquisition layer 10. The data acquisition layer 10 parses the parameter package and extracts the target modality identifier and the target sampling period value. Based on the target sampling period value, it adjusts the event listening time interval of each front-end data acquisition interface. The closed-loop adjustment process of the sampling period reduces the amount of meaningless visual data acquisition in a high-redundancy environment and improves the capture density of strong logic text data.

[0100] As a preferred embodiment, the following steps S610 to S640 are a further refinement of the "feedback control of the sampling frequency of multimodal user interaction data" described in step S400 of Embodiment 1, used to explain in detail the process by which the data acquisition layer receives and executes the sampling adjustment command and dynamically changes the sampling frequency of the underlying modal sensor, specifically including the following steps:

[0101] In step S610, the communication bus interface of the data acquisition layer 10 receives the sampling configuration update parameter package sent by the sampling controller 40. The data acquisition layer 10 loads the sampling configuration update parameter package into the parameter buffer of the data acquisition layer 10. The data acquisition layer 10 calls the built-in parameter parsing engine to parse the sampling configuration update parameter package. The data acquisition layer 10 extracts the terminal device identifier, the target mode identifier, and the corresponding target sampling period value from the parsed parameter data frame.

[0102] In step S620, the data acquisition layer 10 establishes a communication addressing connection with a specific user terminal device based on the extracted terminal device identifier. The data acquisition layer 10 locks the associated front-end data acquisition interface in the specific user terminal device after the communication addressing connection is established based on the extracted target modality identifier. When the target modality identifier points to the text search modality, the data acquisition layer 10 addresses and locks the virtual keyboard input listening interface in the front-end page of the specific user terminal device. When the target modality identifier points to the product image sliding modality, the data acquisition layer 10 addresses and locks the screen touch coordinate reporting interface of the specific user terminal device.

[0103] In step S630, the data acquisition layer 10 writes the target sampling period value into the sampling interval configuration item of the corresponding interface and executes the change of the front-end sampling logic. For the text search mode, the data acquisition layer 10 reduces the event debounce delay parameter of the virtual keyboard input listening interface according to the shortened target sampling period value. By reducing the event debounce delay parameter, the data acquisition layer 10 increases the number of keyboard input events captured per unit time. For the product image sliding mode, the data acquisition layer 10 increases the event throttling time interval of the screen touch coordinate reporting interface according to the extended target sampling period value. By increasing the event throttling time interval, the data acquisition layer 10 reduces the frequency of visual interaction coordinate data being transmitted upward.

[0104] In step S640, after the data acquisition layer 10 completes the numerical update of each modal sampling interval configuration item, it executes the global clock synchronization logic. The data acquisition layer 10 resets the timestamp allocation step size of the log acquisition script, so that the subsequently captured operation behavior data has an event occurrence timestamp identifier that is strictly aligned with the new sampling period. The data acquisition layer 10 clears the interaction nodes remaining in memory in the old sampling period. The data acquisition layer 10 repackages the multimodal heterogeneous interaction data according to the updated sampling configuration and restores the data packet reporting task to the tensor computing server 20 through the network communication interface.

[0105] As a preferred embodiment, the following steps S710 to S740 are a further refinement of the "recording the birth and death states of the topology and generating a persistent bar chart" described in step S500 of Embodiment 1, used to explain in detail the generation process of the topology persistent bar chart data, specifically including the following steps:

[0106] In step S710, the topological homology analysis cluster 30 reads the matrix canonical form elimination record of the witness complex filtering flow during the evolution process. The topological homology analysis cluster 30 extracts the boundary evolution data corresponding to each homology class generator. Based on the boundary evolution data, the topological homology analysis cluster 30 determines the occurrence distance parameter and the disappearance distance parameter of each topological feature. The occurrence distance parameter represents the value of the topological filtering parameter when the simplex closes to form a specific topological feature. The disappearance distance parameter represents the value of the topological filtering parameter when a higher-dimensional simplex fills the corresponding topological feature.

[0107] In step S720, the topology coherence analysis cluster 30 performs a boundary verification operation on the extracted vanishing distance parameter. For topological features that are not filled within the maximum evolution threshold, their corresponding boundary evolution data are in an open state. The topology coherence analysis cluster 30 forcibly assigns the vanishing distance parameter of the topological features in the open state to the preset maximum evolution threshold. The topology coherence analysis cluster 30 pairs the occurrence distance parameter of the same topological feature with the corresponding vanishing distance parameter to form a numerically closed life cycle interval.

[0108] In step S730, the topology homology analysis cluster 30 reads the dimension attribute of the homology group to which each topology feature belongs. The topology homology analysis cluster 30 classifies all lifecycle intervals according to the dimension attribute of the homology group to which each topology feature belongs. The topology homology analysis cluster 30 aggregates lifecycle intervals with 0-dimensional attribute labels into 0-dimensional feature subsets and aggregates lifecycle intervals with 1-dimensional attribute labels into 1-dimensional feature subsets.

[0109] In step S740, the topological homology analysis cluster 30 maps the 0-dimensional feature subset and the 1-dimensional feature subset into a multi-dimensional array. The topological homology analysis cluster 30 defines each element in the multi-dimensional array as a triplet data structure containing the topological dimension, the occurrence distance parameter, and the disappearance distance parameter. The topological homology analysis cluster 30 extracts all the lifetime interval endpoint values ​​in the array and generates structured continuous bar chart data based on the triplet data structure set.

[0110] As a preferred embodiment, the following steps S810 to S850 are a further refinement of the "outputting a structured customer profile through distance matching" described in step S500 of Embodiment 1, used to explain in detail the standard intent coherence class template matching process based on Wasserstein distance, specifically including the following steps:

[0111] In step S810, the profile database 50 reads the structured persistent bar chart data in the memory interaction area. The profile database 50 calls the standard intent coherence class template library stored internally. The standard intent coherence class template library contains multiple pre-generated template bar chart data, and each template bar chart data is bound to an intent classification label.

[0112] In step S820, the portrait database 50 uses the structured persistent bar chart data as the target feature set and the called template bar chart data as the reference feature set. The portrait database 50 establishes a bijective matching relationship between the triplet data structures contained in the target feature set and the triplet data structures contained in the reference feature set. For triplet data structures in the target feature set or the reference feature set that fail to establish a bijective matching relationship, the portrait database 50 directly projects the unmatched triplet data structures onto the diagonal line where the appearance distance parameter and the disappearance distance parameter are equal. The portrait database 50 calculates the sum of the absolute distances between the endpoints of the triplets in the target feature set and the endpoints of the matched triplets in the reference feature set under all the above-established matching relationships.

[0113] In step S830, the portrait database 50 extracts the infimum of all calculated absolute distance sums. The portrait database 50 sets the infimum of the extracted absolute distance sums as the Wasserstein distance value between the target feature set and the reference feature set. The Wasserstein distance value is used to characterize the geometric cost value of the topology of the current session being deformed and converted into the topology represented by template bar chart data.

[0114] In step S840, the profile database 50 traverses all template bar chart data in the standard intent coherence class template library, calculates the Wasserstein distance value between the structured persistent bar chart data and each template bar chart data, sorts all Wasserstein distance values ​​in ascending order, locks the smallest Wasserstein distance value at the top of the sort, and extracts the intent classification label bound to the template bar chart data corresponding to the locked smallest Wasserstein distance value.

[0115] In step S850, the profile database 50 associates the extracted intent classification tags with the current user terminal device identifier to generate a structured macro profile of the customer. The profile database 50 writes the structured macro profile of the customer into the advertising business distribution queue and provides it to the downstream advertising targeting module or advertising push reach module.

Claims

1. A multimodal data-driven intelligent analysis method for customer profiling, characterized in that, Includes the following steps: Step S100: Obtain multimodal user interaction data within a specified time window from the Internet advertising service platform, parse the multimodal user interaction data into discrete interaction nodes, and aggregate them to generate a complete set of interaction nodes; Step S200: Calculate the metric values ​​of each pair of interactive nodes in the complete set of interactive nodes, and construct a non-Euclidean metric matrix that couples modal semantics with spatiotemporal decay. Step S300: Extract a set of landmark points from the complete set of interactive nodes, and construct a witness complex filter flow that evolves with the distance threshold based on the non-Euclidean metric matrix; Step S400: Calculate the homology group rank data of each order of the witness complex filter stream, and generate a sampling adjustment command based on the homology group rank data to control the sampling frequency of the multimodal user interaction data. Step S500: Record the birth and death states of the topology, generate a continuous bar chart, and output a structured customer profile through distance matching; The method further includes: sending the structured customer profile to the advertising targeting module of the Internet advertising service platform, wherein the advertising targeting module retrieves advertising creative content with a matching degree higher than a preset threshold from the candidate advertising material library based on intent classification tags, generates personalized advertising content, and delivers it to the user's terminal device.

2. The multimodal data-driven intelligent customer profiling analysis method according to claim 1, characterized in that, The process of acquiring multimodal user interaction data from user terminal devices connected to the internet advertising service platform within a specified time window, parsing the multimodal user interaction data into discrete interaction nodes, and aggregating them to generate a complete set of interaction nodes includes: Based on the set trigger conditions, listen for the interaction events of the user terminal device, capture the operation behavior that occurs on the client interface, and the specific data types of the operation behavior are divided into text search records, product image interaction coordinates and page dwell time, and attach an event occurrence timestamp to each captured operation behavior; Based on the data type identifier, a vector space mapping is performed to divide the continuous operation behavior data into independent discrete interaction nodes. Each discrete interaction node contains single-dimensional modal data, and the modal attribute identifiers of the discrete interaction nodes are divided into text modality, visual modality, and temporal modality. For discrete interactive nodes containing text search records, a word embedding network is invoked to convert the text search records into text feature vectors; for discrete interactive nodes containing product image interaction coordinates, a location encoding network is invoked to output visual feature vectors; for discrete interactive nodes containing time modalities such as page dwell time, a normalization function is invoked to convert the page dwell time into a fixed-dimensional time feature scalar. The extracted text feature vectors, visual feature vectors, and time feature scalars are uniformly mapped to a set real number vector space. An alignment and truncation operation is performed to generate an initial feature vector with consistent dimensions. The initial feature vector, the event timestamp identifier, and the modal attribute identifier are bound together, and a complete set of interactive nodes containing multi-source heterogeneous information is output.

3. The multimodal data-driven intelligent customer profiling analysis method according to claim 1, characterized in that, The calculation of the metric values ​​for each pair of interaction nodes in the complete set of interaction nodes, and the construction of a non-Euclidean metric matrix coupling modal semantics and spatiotemporal decay, includes: Extract the initial feature vector, event timestamp identifier, and modal attribute identifier corresponding to any two interactive nodes, and calculate the squared Euclidean distance of the initial feature vector in the real vector space; Read the preset cross-modal transition relationship matrix. When the modal attribute identifiers of two interactive nodes are the same, assign the modal semantic transition penalty coefficient to the set benchmark value. When the modal attribute identifiers of two interactive nodes are different, extract the corresponding cross-cost value greater than the benchmark value according to the cross-modal transition relationship matrix and assign it to the modal semantic transition penalty coefficient. Calculate the absolute time difference between the event timestamps of two interacting nodes. Use a system-preset time decay constant greater than zero, with the natural constant as the base and the product of the time decay constant and the absolute time difference as the exponent, to calculate an exponential decay factor. Multiply the squared Euclidean distance, the modality semantic transition penalty coefficient, and the exponential decay factor to obtain the metric value in non-Euclidean space. Summarize the metric values ​​of all interacting node pairs in the complete set of interacting nodes and fill them into the corresponding row and column coordinates in a two-dimensional matrix to generate a non-Euclidean metric matrix. The expression for calculating the metric value is: In the formula, and Represents any two discrete interaction nodes in the universal set of interaction nodes; and Representing discrete interaction nodes and Corresponding modal attribute identifier; Represents discrete interaction nodes and The values ​​between them; This represents the modality semantic transition penalty coefficient; , Representing discrete interaction nodes and The corresponding initial feature vector; Represents the initial eigenvector and The square Euclidean distance between them; and Representing discrete interaction nodes and The corresponding event timestamp identifier; This represents the set time decay constant, and .

4. The multimodal data-driven intelligent customer profiling analysis method according to claim 1, characterized in that, The step of extracting a set of landmarks from the complete set of interactive nodes includes: Read the current available memory balance and the preset maximum computational latency threshold, calculate the downsampling ratio parameter based on the available memory balance and the maximum computational latency threshold, and calculate the target landmark number limit by combining the total number of nodes in the entire set of interactive nodes and the downsampling ratio parameter. Calculate the L2 norm of the feature vectors of each discrete interaction node in the global set of interaction nodes, and select the discrete interaction node with the largest L2 norm of the feature vector as the first landmark point to add to the initialized landmark point set. Perform iterative extraction operation. In each iteration, traverse the remaining nodes in the entire set of interactive nodes that have not been selected into the set of landmarks. Based on the non-Euclidean metric matrix, query the metric values ​​of each remaining node to all known landmarks in the current set of landmarks. Extract the minimum metric value corresponding to each remaining node, lock the maximum value among all extracted minimum metric values, and confirm the remaining node corresponding to the locked maximum value as a new landmark. New landmarks are determined using an iterative update function, the expression for which is calculated is: In the formula, The variable representing the current iteration round being executed, and , Indicates the limit on the number of target landmarks; Represents the complete set of interactive nodes; Indicates the preceding The set of known landmarks obtained from the next iteration; This indicates that all interactive nodes are excluded from the set. The remaining nodes that were not selected; Represents the set of known landmarks Landmarks in the middle; Indicates the remaining nodes that were not selected. Landmarks The metric value in a non-Euclidean metric matrix; Indicates the first The new landmark point determined by the next iteration operation; The newly identified landmarks are added to the landmark set until the number of elements in the landmark set reaches the target landmark number limit, at which point the iteration stops. The remaining nodes in the entire set of interactive nodes, excluding the landmark set, are then assigned to the witness set.

5. The multimodal data-driven intelligent customer profiling analysis method according to claim 4, characterized in that, The construction of the witness complex filtering flow that evolves with the distance threshold based on the non-Euclidean metric matrix includes: Traverse each witness point in the witness point set, and extract the metric values ​​from each witness point to all landmark points in the landmark point set according to the non-Euclidean metric matrix. The minimum value among all extracted metric values ​​is defined as the base distance of the corresponding witness point. The initial values ​​of the topology filtering parameters are set to zero, a step size constant, and a maximum evolution threshold. The topology filtering parameters are controlled to continuously increase according to the step size constant until the maximum evolution threshold is reached. During the process of increasing the topology filtering parameters, for any simplex formed by landmarks in the landmark point set, it is determined whether the corresponding simplex satisfies the relaxation witness condition. The formula for determining the relaxation witness condition is as follows: For simplex Each vertex ,in ,satisfy ; In the formula, The dimension variable representing the simplex; The index variable representing the vertex of the simplex; The simplex contains the first... Each landmark vertex; Represents any witness point in the set of witness points; Represents the vertex of the landmark. With Witness Point The metric value in a non-Euclidean metric matrix; Indicates witness point The corresponding baseline distance; Indicates the topology filtering parameters; When there is at least one witness point that makes the simplex satisfy the relaxed witness condition, the simplex that satisfies the condition is included in the witness complex corresponding to the current topology filtering parameter, and the order of simplex generation is recorded to generate a witness complex filtering flow containing multiple simplexes.

6. The multimodal data-driven intelligent customer profiling analysis method according to claim 1, characterized in that, The calculation of the homology group rank data of each order of the witness complex filter flow includes: Based on the generation order of simplexes in the witness complex filter flow, an edge matrix is ​​constructed, and vertex dimension information of each simplex in the witness complex filter flow is extracted; Use the q-dimensional simplex as the column vector of the edge matrix and the (q-1)-dimensional simplex as the row vector of the edge matrix. If the (q-1)-dimensional simplex is a face of the q-dimensional simplex, record the corresponding element value in the edge matrix as 1; otherwise, record the corresponding element value as 0. Perform matrix canonical form elimination on the marginal matrix, and extract the rank data of each homology group based on the canonical form matrix after elimination; Lock the 0-dimensional Betty number corresponding to the 0th order homology group and the 1-dimensional Betty number corresponding to the 1st order homology group. The 0-dimensional Betty number is used to characterize the number of connected components in the universal set of interactive nodes in non-Euclidean space, and the 1-dimensional Betty number is used to characterize the number of one-dimensional holes in the universal set of interactive nodes in non-Euclidean space. Pack the 0-dimensional Betti numbers and the 1-dimensional Betti numbers into a topological invariant signal.

7. The multimodal data-driven intelligent customer profiling analysis method according to claim 6, characterized in that, The step of generating sampling adjustment instructions based on homology group rank data to control the sampling frequency of multimodal user interaction data includes: Read the topological invariant signal within the historical time window, extract multiple historical 0-dimensional Betty numbers and historical 1-dimensional Betty numbers within the set sliding time window, perform a weighted average operation on the extracted historical values ​​and the current 0-dimensional Betty numbers and 1-dimensional Betty numbers, and output the smoothed 0-dimensional Betty numbers and smoothed 1-dimensional Betty numbers. The smoothed 0-dimensional Betty number is numerically compared with the connectivity threshold, and the smoothed 1-dimensional Betty number is numerically compared with the hole threshold. Based on the comparison results, discrete trigger state bits are output. The trigger state bits include a dispersed state bit that represents the convergence of browsing behavior and a hesitant state bit that represents the divergence of browsing behavior. Based on the activation status of the trigger state bit, the preset sampling frequency dynamic update equation is invoked to calculate the target sampling frequency of the target modal sensor in the next time window. The calculation expression of the sampling frequency dynamic update equation is as follows: In the formula, Representing modes In the next time window The target sampling frequency within; Representing modes In the current time window The initial sampling frequency within; Indicates modality The set modal polarity control factor; This represents the smoothed 1-dimensional Betty number within the current time window; This represents the smoothed 0-Vibetti number within the current time window; This represents the sensitivity gain coefficient set for a 1-dimensional Betty number; This represents the sensitivity gain coefficient set for a Vibbetti number of 0. Indicates modality The preset minimum sampling frequency threshold, and .

8. The multimodal data-driven intelligent customer profiling analysis method according to claim 7, characterized in that, The step of generating sampling adjustment instructions based on homology group rank data to control the sampling frequency of multimodal user interaction data further includes: For the text search modality, the modality polarity control factor is set to a positive number; for the product image sliding modality, the modality polarity control factor is set to a negative number. When the smoothed 1D Betty number is greater than the hole detection threshold, triggering the hesitation state, the increased text modality target sampling frequency and the decreased visual modality target sampling frequency are calculated, and the first feedback signal is generated to indicate that the collection time interval of the text search log is shortened and the collection time interval of the product image sliding is extended simultaneously. When the smoothed 0-dimensional Betty number is greater than the connectivity determination threshold, triggering the scatter state bit, the reduced text modality target sampling frequency and the increased visual modality target sampling frequency are calculated, and a second feedback signal is generated to indicate the extension of the text search log collection time interval and the synchronous shortening of the product image sliding collection time interval. The generated first or second feedback signal is encapsulated into a sampling configuration update parameter package and sent out. The terminal device identifier, target modality identifier and corresponding target sampling period value are extracted from the parsed parameter data frame. A communication addressing connection is established with a specific user terminal device based on the extracted terminal device identifier. The front-end data acquisition interface associated with the specific user terminal device is locked based on the extracted target modality identifier. The target sampling period value is written into the sampling interval configuration item of the corresponding interface to adjust the event listening time interval of the data acquisition layer for each modality operation behavior.

9. The multimodal data-driven intelligent customer profiling analysis method according to claim 1, characterized in that, The process of recording the birth and death states of the topological structure and generating a persistent bar chart includes: Read the matrix canonical elimination records of the witness complex filter flow during its evolution process, and extract the boundary evolution data corresponding to each homology class generator; determine the occurrence distance parameter and disappearance distance parameter of each topological feature based on the boundary evolution data. The occurrence distance parameter represents the value of the topological filter parameter when the simplex closes to form a specific topological feature, and the disappearance distance parameter represents the value of the topological filter parameter when a higher-dimensional simplex fills the corresponding topological feature. For topological features that are not filled within the maximum evolution threshold, the disappearance distance parameter of the topological features in the open state is forcibly assigned to the preset maximum evolution threshold; the appearance distance parameter of the same topological feature is paired with the corresponding disappearance distance parameter to form a numerically closed life cycle interval. Read the homology group dimension attribute of each topological feature, aggregate the lifecycle intervals with 0-dimensional attribute labels into 0-dimensional feature subsets, and aggregate the lifecycle intervals with 1-dimensional attribute labels into 1-dimensional feature subsets. The 0-dimensional feature subset and the 1-dimensional feature subset are mapped to a multi-dimensional array. Each element in the multi-dimensional array is defined as a triplet data structure containing a topological dimension, an appearance distance parameter, and a disappearance distance parameter. All life cycle interval endpoint values ​​in the array are extracted to generate structured continuous bar chart data based on the triplet data structure set.

10. The multimodal data-driven intelligent customer profiling analysis method according to claim 9, characterized in that, The structured customer profile output through distance matching includes: The system calls the internally persistent standard intent coherence class template library, which contains multiple pre-generated template bar chart data, each of which is bound to an intent category label. Using structured persistent bar chart data as the target feature set and the called template bar chart data as the reference feature set, a bijective matching relationship is established between the triplet data structure contained in the target feature set and the triplet data structure contained in the reference feature set. For triplet data structures in the target feature set or reference feature set that fail to establish a bijective matching relationship, the unmatched triplet data structures are directly projected and matched onto the diagonal line where the appearance distance parameter and the disappearance distance parameter are equal; the sum of absolute distances between the endpoints of the triplet in the target feature set and the endpoints of the matched triplet in the reference feature set are calculated under all the above-established matching relationships. Extract the infimum from the sum of all calculated absolute distances, and set the infimum of the extracted sum of absolute distances as the Wasserstein distance value between the target feature set and the reference feature set. Calculate the Wasserstein distance values ​​between the structured persistent bar chart data and each template bar chart data and sort them in ascending order. Lock the smallest Wasserstein distance value at the top of the sorted list and extract the intent classification label bound to the template bar chart data corresponding to the locked smallest Wasserstein distance value. The extracted intent classification tags are associated with the current user terminal device identifier to generate a structured macro-profile of customers.

Images