A marketing management method and system based on big data

Abstract

The application discloses a marketing management method and system based on big data, and relates to the technical field of monitoring and analysis, which comprises the following steps: when an internal coordination score reaches a basic threshold, acquiring enterprise public dynamic information associated with a candidate anonymous group network segment and extracting an external intention vector; performing space vector similarity calculation on semantic trajectory feature data of the candidate anonymous group and the external intention vector to obtain an external corroboration score, and fusing the internal coordination score and the external corroboration score to obtain a comprehensive procurement intention score; when the comprehensive procurement intention score is lower than an action threshold, increasing a semantic task complementary weight in a hyperedge and re-dividing a dynamic co-occurrence hypergraph; when the comprehensive procurement intention score reaches the action threshold, aggregating behavior characteristics of the corresponding candidate anonymous group to generate a group virtual portrait, and outputting a marketing resource scheduling instruction matched with the corresponding network segment of the candidate anonymous group. The application has the effect of improving the efficiency of marketing management.

Description

Technical Field

[0001] This application relates to the field of monitoring and analysis technology, and in particular to a marketing management method and system based on big data. Background Technology

[0002] In a digital business environment, marketing management has become a crucial aspect for enterprises to acquire customer needs, allocate marketing resources, and improve business conversion efficiency. Meanwhile, the vast amount of session logs generated by users during online interactions such as searching, browsing, clicking, and dwelling records the attention potential customers have to products, services, and industry information, serving as an important data source for depicting changes in market demand. For marketing scenarios targeting enterprise clients, a single visitor identity often cannot fully represent the true purchasing entity. The combination of multiple anonymous sessions in terms of time, space, and business themes may more closely approximate the actual decision-making process involving different roles within the enterprise.

[0003] It's worth noting that a company's purchasing intent is usually not directly reflected in a single visit, but rather scattered across continuous access patterns across different time periods, device environments, and business pages. For example, technical personnel might focus on product compatibility information, managers might focus on cost and delivery information, and operations and maintenance personnel might focus on after-sales support information. These behaviors appear independent on the surface, but if their source network segments, active periods, and semantic themes are related, they may all point to the same purchasing need. Furthermore, publicly available information such as tenders, recruitment notices, and news from companies can also provide external corroboration of the business focus in anonymous sessions, requiring marketing management to shift from identifying isolated visits to analyzing the correlations between multi-source data.

[0004] However, existing marketing management methods still have significant shortcomings. Traditional solutions rely heavily on account logins, pageview counts, keyword matching, and static customer tags for judgment. In scenarios involving anonymous access, collaborative research by multiple people, and long enterprise procurement chains, this can easily lead to the same purchasing entity being split into multiple unrelated leads, and short-term browsing can be misjudged as high-intent demand. While some systems incorporate big data statistics, they mainly focus on traffic aggregation and user profiling, lacking verification of spatiotemporal co-occurrence relationships between conversations, semantic complementarity, and dynamic interactions with external enterprises. This results in lagging marketing resource allocation, inaccurate targeting, and difficulty in consistently identifying genuine purchasing intentions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a marketing management method and system based on big data.

[0006] Firstly, this application provides a marketing management method based on big data, comprising the following steps: Obtain anonymized session logs after desensitization, and extract spatiotemporal feature data and semantic trajectory feature data from the anonymized session logs; Spatiotemporal feature data and semantic trajectory feature data are mapped to anonymous session nodes in a dynamic co-occurrence hypergraph, and hyperedges connecting multiple anonymous session nodes are generated according to spatiotemporal proximity and semantic task complementarity. The dynamic co-occurrence hypergraph is segmented to obtain candidate anonymity groups, which are then entered into the intent detection loop. The intent detection loop includes: Calculate the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and fuse them to obtain the internal collaboration score; When the internal collaboration score is lower than the basic threshold, the corresponding candidate anonymization group is removed and the remaining candidate anonymization groups are detected. When the internal collaboration score reaches the basic threshold, obtain publicly available dynamic information of enterprises associated with candidate anonymous group network segments and extract external intent vectors. The semantic trajectory feature data of the candidate anonymized group is compared with the external intent vector to calculate the spatial vector similarity to obtain the external verification score. The internal collaboration score and the external verification score are then combined to obtain the comprehensive procurement intent score. When the overall procurement intention score is lower than the action threshold, increase the semantic task complementarity weights in the hyperedge and re-segment the dynamic co-occurrence hypergraph. When the overall procurement intent score reaches the action threshold, the behavioral characteristics of the corresponding candidate anonymous group are aggregated to generate a virtual profile of the group, and marketing resource scheduling instructions matching the network segment corresponding to the candidate anonymous group are output.

[0007] Secondly, this application provides a big data-based marketing management system, including: The data extraction module is used to obtain anonymized session logs and extract spatiotemporal feature data and semantic trajectory feature data from the anonymized session logs. The data mapping module is used to map spatiotemporal feature data and semantic trajectory feature data into anonymous session nodes in a dynamic co-occurrence hypergraph, and generate hyperedges connecting multiple anonymous session nodes according to spatiotemporal proximity and semantic task complementarity. The detection loop module is used to segment the dynamic co-occurrence hypergraph to obtain candidate anonymity groups, and then enter the intent detection loop. The intent detection loop includes: Calculate the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and fuse them to obtain the internal collaboration score; When the internal collaboration score is lower than the basic threshold, the corresponding candidate anonymization group is removed and the remaining candidate anonymization groups are detected. When the internal collaboration score reaches the basic threshold, obtain publicly available dynamic information of enterprises associated with candidate anonymous group network segments and extract external intent vectors. The semantic trajectory feature data of the candidate anonymized group is compared with the external intent vector to calculate the spatial vector similarity to obtain the external verification score. The internal collaboration score and the external verification score are then combined to obtain the comprehensive procurement intent score. When the overall procurement intention score is lower than the action threshold, increase the semantic task complementarity weights in the hyperedge and re-segment the dynamic co-occurrence hypergraph. The management output module is used to aggregate the behavioral characteristics of the corresponding candidate anonymous group to generate a virtual profile of the group when the comprehensive procurement intent score reaches the action threshold, and output marketing resource scheduling instructions that match the network segment corresponding to the candidate anonymous group.

[0008] In summary, this application includes at least one of the following beneficial technical effects: 1. This application provides a marketing management method based on big data. By acquiring anonymized session logs and simultaneously extracting spatiotemporal feature data and semantic trajectory feature data from them, the originally discrete and isolated access behaviors are transformed into multidimensional feature expressions that combine access location relationships, active time relationships, and business semantic relationships. The aforementioned spatiotemporal feature data and semantic trajectory feature data are mapped to anonymous session nodes in a dynamic co-occurrence hypergraph, and hyperedges that can connect multiple anonymous session nodes are generated according to spatiotemporal proximity relationships and semantic task complementarity relationships. This enables the system to no longer rely solely on single session, single IP, or single page access behavior for judgment, but to identify potential group intentions from the co-occurrence structure, collaborative time windows, and complementarity of business browsing paths among multiple anonymous sessions. By constructing a dynamic co-occurrence hypergraph, spatiotemporal features are used to constrain whether there is a real-world connection between anonymous session nodes, while semantic trajectory features are used to determine whether different anonymous sessions revolve around the same procurement task or related business links. The two work together to make the formation of candidate anonymous groups both spatially and temporally reasonable, as well as consistent and complementary in business semantics. This effectively avoids the misjudgment problem caused by traditional methods that rely solely on access counts, keyword hits, or single-point behavioral labels for clue identification. This application can organize dispersed anonymous access behaviors into a computable, divisible, and iteratively optimized group structure without directly identifying individual identities, significantly improving the stability, accuracy, and interpretability of enterprise-level procurement intent identification. 2. After obtaining candidate anonymous groups, an intent detection loop is introduced. By calculating the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymous groups, and fusing them to obtain an internal collaboration score, the system can initially screen candidate anonymous groups from two aspects: consistency of behavior within the group and complementarity of task division. When the internal collaboration score is lower than the basic threshold, low-confidence groups are promptly removed, and the remaining candidate anonymous groups are detected, thereby reducing interference from random co-occurrence, access to hot pages, or irrelevant browsing behavior. When the internal collaboration score reaches the basic threshold, this application further obtains publicly available dynamic information of enterprises associated with the network segments of the candidate anonymous groups and extracts external intent vectors. The semantic trajectory feature data of the candidate anonymous groups and the external intent vectors are used to calculate spatial vector similarity to obtain an external verification score. The internal collaboration score and the external verification score are then fused to form a comprehensive procurement intent score, so that the evidence of internal anonymous behavior and the external publicly available business dynamics form a mutual verification relationship, avoiding one-sided judgments based solely on on-site behavior. Especially when the overall procurement intent score is lower than the action threshold, this invention does not directly discard the group. Instead, it increases the semantic task complementarity weights in the hyperedges and re-segments the dynamic co-occurrence hypergraph. This allows for reverse adjustment of the graph structure based on the detection results, continuously optimizing the candidate group boundaries. When the overall procurement intent score reaches the action threshold, it aggregates the behavioral features of the corresponding candidate anonymous group to generate a virtual profile of the group and outputs marketing resource scheduling instructions that match the network segment corresponding to the candidate anonymous group. Through the above-mentioned cyclical mechanism, a closed-loop collaboration is formed between dynamic co-occurrence hypergraph segmentation, internal collaborative judgment, external intent verification, and marketing resource scheduling. This not only improves the accuracy of potential procurement group identification but also makes marketing resource allocation closer to the actual needs of enterprises, reducing ineffective outreach and resource waste, and improving the timeliness and targeting of lead discovery, customer nurturing, and marketing response. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a marketing management method based on big data, as described in this application.

[0011] Figure 2 This is a schematic diagram of a marketing management system based on big data, as described in this application embodiment. Detailed Implementation

[0012] The following description, in conjunction with the implementation of the present invention, is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention, and all such modifications and additions should fall within the protection scope of the present invention.

[0013] Example 1: This application transforms the originally scattered access behaviors in the anonymized anonymous session logs into a data structure with both spatiotemporal and semantic correlations, and uses a dynamic co-occurrence hypergraph to express the co-occurrence and complementary relationships between multiple anonymous session nodes. In enterprise procurement scenarios, the same procurement demand is often formed by different positions, different equipment environments, and different access times, making it difficult for a single anonymous session to directly reflect the true intent. This application establishes a processing link from anonymous session logs to candidate anonymous groups, and then to comprehensive procurement intent scores and marketing resource scheduling instructions, focusing on the data connection between spatiotemporal feature data, semantic trajectory feature data, and publicly available dynamic information of the enterprise, in order to reduce misjudgments and omissions in anonymous access scenarios.

[0014] This application discloses a marketing management method based on big data.

[0015] Reference Figure 1 A big data-based marketing management approach includes the following steps: Obtain anonymized session logs after desensitization, and extract spatiotemporal feature data and semantic trajectory feature data from the anonymized session logs; Furthermore, the anonymized session logs are obtained, and spatiotemporal feature data and semantic trajectory feature data are extracted from the anonymized session logs, including: Parse anonymous session logs and separate them into underlying network interaction records and front-end page browsing records; For the underlying network interaction records, extract IP network segments, border gateway protocol routing information, device environment hash data, and timestamp sequences; The IP network segment, border gateway protocol routing information, and device environment hash data are merged to generate a network segment environment identifier; Active time slices are segmented according to the timestamp sequence, and network segment environment identifiers are associated with active time slices to generate spatiotemporal feature data; Extract search keyword sequences, page business tag sequences, and page interaction sequences from the front-end page browsing history; Map search keyword sequences and page business tag sequences to industry knowledge graphs to obtain entity relationship paths; Using the page interaction sequence as the access weight of the entity relationship path, semantic trajectory feature data is generated.

[0016] Furthermore, the search keyword sequence and page business tag sequence are mapped to a pre-built industry knowledge graph to obtain entity relationship paths, including: The search keyword sequence is segmented and cleaned to obtain keyword units; Synonymize keyword units with page business tag sequences to generate business semantic units; Retrieve the business entity node corresponding to the business semantic unit in the industry knowledge graph; Extract the shortest business relationship chain between adjacent business entity nodes to generate entity relationship paths; Based on the business steps involved in the entity relationship path, the entity relationship path is marked as a technology adaptation path, a cost assessment path, an implementation and delivery path, and an operation and maintenance support path. The annotation results of each path are summarized into semantic task labels, and the semantic task labels and entity relationship paths together form semantic trajectory feature data.

[0017] In one specific implementation, the above method can be executed by a data processor in a marketing management platform. The data processor includes a log access program, a feature extraction program, a hypergraph construction program, a group segmentation program, an intent detection program, and a marketing resource scheduling program. All programs run in the same data processing environment. Anonymous session logs, dynamic co-occurrence hypergraphs, candidate anonymized groups, publicly available enterprise dynamic information, and group virtual profiles are all stored in the storage area in the form of structured data tables, graph data files, and vector index files. Taking an industrial software service provider's online marketing site as an example, the data processor reads the anonymized session logs once every preset collection period. The preset collection period can be set to fifteen minutes, thirty minutes, or one hour; this implementation uses thirty minutes. The anonymized session logs are derived from website access logs, site search logs, and page interaction logs. The logs do not retain direct identity fields such as personal names, mobile phone numbers, and email addresses. IP addresses are processed into IP network segments, and device fingerprints are processed into device environment hash data. A single record may include fields such as IP network segment, border gateway protocol routing information, device environment hash data, access time, search keywords, page business tags, click actions, dwell time, and scroll depth. Using anonymized session logs as input allows subsequent processing to focus on the relationships between data, rather than relying on personal identification.

[0018] After the log access program obtains the anonymized session logs, the feature extraction program extracts spatiotemporal feature data and semantic trajectory feature data from the anonymous session logs. The spatiotemporal feature data reflects the relationship between the anonymous session and its network source and time distribution, while the semantic trajectory feature data reflects the path relationship of the anonymous session in accessing business content. To ensure uninterrupted data flow in subsequent steps, the feature extraction program generates a unique anonymous session number for each anonymous session log. This anonymous session number is only used to associate the spatiotemporal feature data and semantic trajectory feature data within the same log entry and is not used to reconstruct the user's identity. For example, the anonymous session number... The corresponding spatiotemporal feature data includes the IP network segment 10.24.16.0 / 24, the border gateway protocol routing information AS number, the device environment hash data Hf37a, and the active time slice from 09:30 to 09:45 on May 18, 2025. The semantic trajectory feature data corresponding to this anonymous session number includes the search keyword sequence, the page business tag sequence, the page interaction sequence, and the entity relationship path formed by these data.

[0019] Specifically, the parsing of anonymous session logs has a significant impact on subsequent data quality. In this implementation, the log access program first parses the anonymous session logs, separating them into underlying network interaction records and front-end page browsing records. The underlying network interaction records mainly come from the network access layer, including IP network segments, border gateway protocol routing information, device environment hash data, and timestamp sequences. The front-end page browsing records mainly come from the page tracking layer, including search keyword sequences, page business tag sequences, and page interaction sequences. During separation, parsing rules can be established according to the source and type of log fields. For example, records containing IP network segments, AS numbers, routing prefixes, device environment hash data, and access times are written to the underlying network interaction records, while records containing site search terms, page categories, click types, dwell time, and scroll depth are written to the front-end page browsing records. Both types of records retain the same anonymous session number for subsequent merging.

[0020] For underlying network interaction records, the feature extraction program extracts IP network segments, border gateway protocol routing information, device environment hash data, and timestamp sequences. IP network segments can use CIDR format, such as 203.86.12.0 / 24; border gateway protocol routing information may include the AS number, routing prefix, and autonomous system name; device environment hash data can be generated by hashing de-identified fields such as browser type, operating system type, screen size, and language settings, with an example value of a9f3c21d; the timestamp sequence records the occurrence time of consecutive access actions within the anonymous session. The feature extraction program merges the IP network segment, border gateway protocol routing information, and device environment hash data to generate a network segment environment identifier. During merging, the IP network segment and border gateway protocol routing information are first concatenated into a network source string, then combined with the first few characters of the device environment hash data, generating a fixed-length network segment environment identifier through hash mapping. Using a network segment environment identifier can reduce the instability caused by the drift of a single IP network segment, while preserving consistent information about the enterprise network exit, routing affiliation, and device environment.

[0021] Timestamp sequences are used to segment active time slices. The feature extraction program compares the time interval between adjacent access actions with a preset interval threshold, which can be between ten and thirty minutes; in this implementation, fifteen minutes is used. When the interval between adjacent timestamps does not exceed fifteen minutes, the two access actions are grouped into the same active time slice; when the interval exceeds fifteen minutes, the latter access action is taken as the starting point of a new active time slice. Subsequently, the feature extraction program associates the network segment environment identifier with the active time slice to generate spatiotemporal feature data. The resulting spatiotemporal feature data includes both the consistency of the network source and the time window in which the access behavior occurs, and can be directly used when calculating the overlap of active time slices.

[0022] For front-end page browsing history, the feature extraction program extracts search keyword sequences, page business tag sequences, and page interaction sequences. Search keyword sequences come from the site search box and source search parameters, such as industrial data acquisition, private deployment, and interface authorization. Page business tag sequences come from pre-configured business categories on the page, such as data acquisition, system integration, price consultation, and after-sales service. Page interaction sequences include click actions, dwell time, scroll depth, and document download actions. The feature extraction program maps the search keyword sequences and page business tag sequences to an industry knowledge graph to obtain entity relationship paths. The industry knowledge graph can be pre-constructed from a product catalog, solution documents, industry terminology library, and business process library, where nodes are business entity nodes and edges are business relationships, such as relationships between product modules, deployment methods, pricing methods, delivery processes, and operation and maintenance support. The page interaction sequences are then used as access weights for entity relationship paths; for example, longer dwell time, higher scroll depth, and more explicit document download actions result in higher access weights for the corresponding entity relationship paths. Access weights can be calculated based on... Calculate, where, This indicates the access weight of the entity relationship path. This represents the normalized dwell time. This represents the normalized scroll depth. This represents the normalized score for actions such as downloading data and expanding forms. The resulting semantic trajectory feature data includes not only which business content was accessed, but also the intensity of those accesses.

[0023] in, , and Confirmation can be made separately under the same anonymous session ID, where, The value is obtained by comparing the page dwell time with the reference dwell time for the business topic to which the page belongs. The reference dwell time can be taken as the 75th percentile of the dwell time of similar pages in historical anonymous session logs. If the actual dwell time is... The recommended duration of stay is ,but This avoids amplifying the access weight due to a small number of abnormally long dwell times; The value is obtained by comparing the actual maximum scroll depth of the page with the total scrollable depth of the page. If the actual maximum scroll depth is... The total scrollable depth of the page is ,but This is used to characterize the extent to which the anonymous session covers the page content. Based on the action type and frequency in the page interaction sequence, first set basic action scores for downloading materials, expanding forms, clicking on price components, and clicking on case details. Then, sum the basic action scores within the same page and compare them with the preset action saturation value. If the sum of the basic action scores is... The preset action saturation value is ,but The preset action saturation value can be determined by the median of the sum of the base scores of similar page actions in historical effective conversion samples, so that... It can reflect a clear intention to interact without causing excessive bias due to repeated clicks.

[0024] In the process of mapping search keyword sequences and page business tag sequences to the industry knowledge graph, the feature extraction program first performs word segmentation and cleaning on the search keyword sequences to obtain keyword units. Word segmentation and cleaning includes removing stop words, unifying capitalization, eliminating invalid symbols, and merging number and model numbers. For example, the search keyword "private deployment cost" can be decomposed into two keyword units: "private deployment" and "cost". Then, the feature extraction program performs synonym merging with the page business tag sequences to generate business semantic units. Synonym merging can be done based on an industry terminology table; for example, "price", "fee", and "quotation" are merged into cost assessment-related business semantic units, while "installation", "deployment", and "launch" are merged into implementation and delivery-related business semantic units. Business semantic units reduce the differences in natural language expressions, making subsequent industry knowledge graph retrieval more stable.

[0025] After generating the business semantic units, the feature extraction program retrieves the corresponding business entity nodes in the industry knowledge graph. The retrieval can employ a combination of exact matching and edit distance matching. Business entity nodes with an edit distance similarity reaching a preset matching threshold are retained; this threshold can be set to 0.85. For adjacent business entity nodes, the feature extraction program extracts the shortest business relationship chain between them, generating an entity relationship path. The shortest business relationship chain can be obtained using a weighted shortest path algorithm, where the edge weight is represented by the reciprocal of the business relationship's closeness; the closer the business relationship, the smaller the edge weight. If the business entity node is privately deployed or uses an interface for authorization, the shortest business relationship chain may pass through nodes related to deployment methods, system integration, and interface management, forming an entity relationship path.

[0026] The preset matching threshold can be confirmed using historical mapping samples with manual verification. First, search keyword sequences and page business tag sequences are extracted from historical anonymous session logs to form an annotated sample set containing the correspondence between business semantic units and correct business entity nodes. Then, the edit distance similarity between each business semantic unit and candidate business entity nodes in the industry knowledge graph is calculated, and candidate thresholds are iterated within a range of 0.70 to 0.95 with a step size of 0.01 or 0.02. For each candidate threshold, the precision, recall, and F1 score of the retained results for business entity nodes are statistically analyzed. The candidate threshold with the highest F1 score and a precision not lower than the preset precision requirement is selected as the preset matching threshold. If the F1 scores of multiple candidate thresholds are close, the threshold with the higher value is selected to reduce the possibility of page business tag sequences being incorrectly mapped to irrelevant business entity nodes. In the actual system, when the annotated sample set shows a threshold of 0.85, it can balance synonym recall and mismatch suppression; therefore, 0.85 can be determined as the preset matching threshold and can be reviewed and adjusted monthly or quarterly using newly added annotated samples.

[0027] Subsequently, the feature extraction program labels the entity relationship paths according to the business stages they pass through: technology adaptation path, cost assessment path, implementation and delivery path, and operation and maintenance support path. If the entity relationship path passes through nodes such as interfaces, compatibility, and system integration, it is labeled as a technology adaptation path; if it passes through nodes such as pricing, authorization, and contracts, it is labeled as a cost assessment path; if it passes through nodes such as deployment, migration, and training, it is labeled as an implementation and delivery path; and if it passes through nodes such as work orders, maintenance, and upgrades, it is labeled as an operation and maintenance support path. The labeling results for each path are summarized into semantic task tags. These semantic task tags, together with the entity relationship paths, form semantic trajectory feature data, which is then written into subsequent anonymous session nodes along with the anonymous session number.

[0028] Spatiotemporal feature data and semantic trajectory feature data are mapped to anonymous session nodes in a dynamic co-occurrence hypergraph, and hyperedges connecting multiple anonymous session nodes are generated according to spatiotemporal proximity and semantic task complementarity. Furthermore, spatiotemporal feature data and semantic trajectory feature data are mapped to anonymous session nodes in a dynamic co-occurrence hypergraph. Hyperedges connecting multiple anonymous session nodes are generated according to spatiotemporal proximity and semantic task complementarity, including: Anonymous session nodes are established based on a single anonymous session as the basic unit, and spatiotemporal feature data is written into the spatiotemporal attribute field of the anonymous session node; Write semantic trajectory feature data into the semantic attribute field of the anonymous session node; Based on the network segment environment identifier and active time slice in the spatiotemporal attribute field, a spatiotemporal hyperedge connecting multiple anonymous session nodes is generated, and spatiotemporal connection weights are written to the spatiotemporal hyperedge. Based on the semantic task labels and entity relationship paths in the semantic attribute fields, generate semantic hyperedges that connect multiple anonymous session nodes, and write complementary weights of semantic tasks to the semantic hyperedges. By writing spatiotemporal hyperedges and semantic hyperedges into the same hypergraph structure, a dynamic co-occurrence hypergraph is obtained.

[0029] Furthermore, semantic hyperedges connecting multiple anonymous session nodes are generated, including: Read the semantic task tags of each anonymous session node; When the semantic task labels of multiple anonymous session nodes belong to the same business topic, and each semantic task label covers different business links, the multiple anonymous session nodes are merged into a candidate complementary node group. The percentage of duplicate entity relationship paths and the percentage of business process coverage within candidate complementary node groups are statistically analyzed. Based on the repetition ratio, homogeneous behavior deduction is performed on candidate complementary node groups, and semantic task complementary weights are generated based on the coverage ratio of business links. The semantic task complementary weights are written into the superedges corresponding to the candidate complementary node groups to obtain the semantic superedges.

[0030] In one specific implementation, after obtaining spatiotemporal feature data and semantic trajectory feature data, the hypergraph construction program maps them to anonymous session nodes in a dynamic co-occurrence hypergraph. Unlike ordinary graphs, a dynamic co-occurrence hypergraph allows a single hyperedge to connect more than two anonymous session nodes, making it suitable for representing situations where multiple anonymous sessions in an enterprise procurement scenario point to the same business theme. The hypergraph construction program establishes anonymous session nodes based on a single anonymous session as the basic unit, writing spatiotemporal feature data under the same anonymous session number into the spatiotemporal attribute field of the anonymous session node, and writing semantic trajectory feature data under the same anonymous session number into the semantic attribute field of the anonymous session node. Subsequently, the hypergraph construction program generates hyperedges connecting multiple anonymous session nodes according to spatiotemporal proximity and also generates hyperedges connecting multiple anonymous session nodes according to semantic task complementarity. The spatiotemporal proximity can be determined by the same IP network segment, similar border gateway protocol routing information, the same device environment hash data prefix, and overlapping active time slices; the semantic task complementarity is determined by the coverage of different business processes under the same business theme. For example, if three anonymous session nodes access the product deployment instructions, pricing plan page, and after-sales service page respectively, and their network segment environment identifiers are similar and their active time slices overlap, then these three anonymous session nodes can be connected by the same spatiotemporal hyperedge; if their semantic task tags point to the technology adaptation path, cost assessment path, and operation and maintenance support path respectively, then they can also be connected by the same semantic hyperedge.

[0031] Specifically, in the dynamic co-occurrence hypergraph construction phase, the hypergraph construction program establishes anonymous session nodes based on a single anonymous session as the basic unit. Anonymous session nodes may include node identifier fields, spatiotemporal attribute fields, semantic attribute fields, and weight cache fields. The spatiotemporal attribute fields contain spatiotemporal feature data, including network segment environment identifiers and active time slices; the semantic attribute fields contain semantic trajectory feature data, including entity relationship paths, semantic task tags, and access weights corresponding to page interaction sequences. Based on the network segment environment identifiers and active time slices in the spatiotemporal attribute fields, the hypergraph construction program generates spatiotemporal hyperedges connecting multiple anonymous session nodes and writes spatiotemporal connection weights to these hyperedges. The spatiotemporal connection weights can be configured according to... Calculate, where, Indicates the spatiotemporal connectivity weights. This represents the consistency score of the network segment environment identifier. This indicates the score for overlapping active time slices. This indicates the similarity score of the device environment hash data. , and These are weighting coefficients, and the sum of the three is 1. In implementation, 0.45 is acceptable. 0.35 is acceptable. A value of 0.20 can be used to ensure that the consistency of network source plays a major role in the spatiotemporal connection weight, while preserving time and device environment factors.

[0032] in, , and Confirmation can be made among multiple anonymous session nodes that intend to generate the same spatiotemporal hyperedge. The consistency score used to characterize the network segment environment identifier is determined by first reading the network segment environment identifier of each anonymous session node and then extracting the IP network segment, border gateway protocol routing information, and network source portion of the device environment hash data. If multiple anonymous session nodes have completely identical network segment environment identifiers, then... If the IP network segments are the same but the border gateway protocol routing information differs, then the calculation is based on the route prefix overlap ratio and AS number consistency. The higher the route prefix overlap ratio and the more consistent the AS number, the better. If the value is closer to 1, and the network segment is only adjacent or under the same border gateway protocol routing information, then a value between 0.4 and 0.7 is assigned. If there is no correlation between the two, then... Take 0. To characterize the active time-slice overlap score, it can perform pairwise comparisons of the active time-slices of multiple anonymous session nodes, calculate the ratio of the intersection duration to the union duration of the active time-slices for each node, and then take the average of all node pairs as the score. When there are multiple active time slices, active time slices within the same node can be merged first, and then participate in the overlap calculation. The resulting values ​​are limited to between 0 and 1. To characterize the similarity score of device environment hash data, the device environment hash data can be broken down into hash sub-items corresponding to anonymized source fields such as browser type, operating system type, terminal type, language settings, and screen size range. The similarity score is calculated as the ratio of the number of consistent hash sub-items to the total number of hash sub-items. If the system only retains the complete hash value, Hamming distance similarity can be used for calculation. Let the Hamming distance between two device environment hash data be... The hash length is The similarity is Multiple anonymous session nodes corresponding to Take the average similarity of all node pairs.

[0033] , and This can be confirmed through training with historical samples and business constraints. During implementation, first select a group of anonymous session nodes that have been confirmed by marketing personnel or historical conversion results to belong to the same enterprise's procurement research behavior as positive samples. Then select anonymous session node groups with random access, personal browsing, or no conversion records as negative samples. Calculate the performance of each sample group separately. , and ; then in And under the condition that all weight coefficients are not less than 0, for , and Perform a grid search, for example, traversing different combinations with a step size of 0.05, and use AUC, F1 score, or false-merging rate (the ability to distinguish between positive and negative samples) as evaluation metrics. If the evaluation results of multiple combinations are similar, prioritize the selected combination. higher and The next best combination aligns with the business characteristics of enterprise customers who typically share network exits, require relatively stable simultaneous collaborative access, and have varying equipment environments. Therefore, this can be used to... The value was determined to be 0.45. The value was determined to be 0.35. The value is set to 0.20. When access habits change in different industries or on different sites, the above weight coefficients can be recalculated quarterly using newly confirmed samples to ensure that the spatiotemporal connection weights are consistent with the actual co-occurrence patterns of anonymous sessions.

[0034] Based on the semantic task tags and entity relationship paths in the semantic attribute fields, the hypergraph builder generates semantic hyperedges connecting multiple anonymous session nodes and writes semantic task complementarity weights to these hyperedges. The semantic task complementarity weights do not simply pursue identical keywords, but rather focus on whether different business processes under the same business theme form a combination. For example, if one anonymous session node's semantic task tag is a technology adaptation path, another's is a cost assessment path, and a third's is an operation and maintenance support path, and all three entity relationship paths fall within the same product or solution theme, then the semantic task complementarity weight is higher than if all three anonymous session nodes repeatedly browse the same price page. Finally, the hypergraph builder writes the spatiotemporal hyperedges and semantic hyperedges into the same hypergraph structure, resulting in a dynamic co-occurrence hypergraph. In a dynamic co-occurrence hypergraph, the same anonymous session node can simultaneously belong to multiple spatiotemporal hyperedges and multiple semantic hyperedges, expressing the cross-relationships of enterprise procurement behavior in terms of time, network, and business content.

[0035] When generating semantic hyperedges, the hypergraph builder reads the semantic task labels of each anonymous session node. If the semantic task labels of multiple anonymous session nodes belong to the same business theme, and each semantic task label covers different business processes, the hypergraph builder merges the multiple anonymous session nodes into a candidate complementary node group. The business theme can be determined by the core business entity nodes in the entity relationship path. For example, if the core business entity nodes are a low-code platform, a data middle platform, or a work order system, then different business themes are formed. After the candidate complementary node group is formed, the hypergraph builder calculates the repetition rate of the entity relationship path and the business process coverage rate within the candidate complementary node group. (Repetition rate...) This can be expressed as the ratio of the number of duplicate entity relationship paths to the total number of entity relationship paths within the candidate complementary node group; business process coverage ratio. It can be expressed as the ratio of the number of covered business links to the total number of preset business links. In this embodiment, the total number of preset business links is four, corresponding to the technology adaptation path, cost assessment path, implementation and delivery path, and operation and maintenance support path.

[0036] Candidate complementary node groups are processed by deducting homogeneous behaviors based on the repetition ratio, and semantic task complementarity weights are generated based on the coverage ratio of business processes. Semantic task complementarity weights can be calculated as follows: Calculate, where, Indicates the complementary weights of semantic tasks. Indicates the coverage ratio of business processes. Indicates the repetition ratio. This represents the deduction coefficient for homogeneous behavior. The value can be between 0.5 and 2; in this implementation, it is set to 1. If the candidate complementary node group covers three of the four business links, The repetition rate is 0.75. If the value is 0.20, then the semantic task complementary weights are... The semantic task complementarity weight is 0.60; if the coverage ratio is the same but the repetition ratio is 0.60, the semantic task complementarity weight drops to 0.30. This calculation method ensures that anonymous session nodes that repeatedly browse the same content are not mistaken for a high-quality complementary group, while candidate complementary node groups with more complete business process coverage can obtain higher semantic task complementarity weights. Subsequently, the hypergraph construction program writes the semantic task complementarity weights into the hyperedges corresponding to the candidate complementary node groups, thus obtaining semantic hyperedges.

[0037] in, The degree of interference of homogeneous browsing behavior on the semantic task complementarity weights can be confirmed based on the historical anonymous session logs. During implementation, two types of samples are first collected from the historical anonymous session logs: one is a group of candidate complementary nodes with confirmed enterprise procurement collaborations, and the other is a group of homogeneous access nodes caused by hot pages, advertising landing pages, or document download pages. For each sample, the repetition rate of entity relationship paths and the coverage rate of business processes are statistically analyzed, and the samples are traversed in steps of 0.1 or 0.2 within the range of 0.5 to 2. Substitute Calculate the semantic task complement weights. For each candidate... By comparing the semantic task complementarity weight distributions of positive samples and homogeneous access node groups, priority is given to selecting those that significantly reduce the semantic task complementarity weights of homogeneous access node groups without significantly suppressing the values ​​of confirmed enterprise procurement collaboration samples. If there are many repeated page views in the business and the risk of accidental merging is high, then Take the value closest to 2, and enhance the deduction process for homogeneous behavior. If the business process is naturally less complex and the same entity relationship path is repeatedly viewed by multiple people, this is considered normal procurement research behavior. Take a value close to 0.5 to avoid excessive deduction. In this implementation, a value close to 0.5 can be used. Setting it to 1 indicates that the repetition ratio linearly reduces the complementary weights of the semantic task, which helps to achieve a balance between accuracy and stability.

[0038] The dynamic co-occurrence hypergraph is segmented to obtain candidate anonymity groups, which are then entered into the intent detection loop. The intent detection loop includes: Furthermore, the dynamic co-occurrence hypergraph is segmented to obtain candidate anonymity groups, including: Read anonymous session nodes, spatiotemporal hyperedges, semantic hyperedges, spatiotemporal connection weights, and semantic task complementary weights from the dynamic co-occurrence hypergraph; Anonymous session nodes are grouped according to spatiotemporal connectivity weights and semantic task complementary weights to obtain an initial anonymous group; The timestamp sequence in the anonymous session log is rearranged to generate background session data. Construct a background co-occurrence hypergraph based on background session data, and calculate the background population density in the background co-occurrence hypergraph; A background fluctuation threshold is generated based on the background population density and a preset fluctuation coefficient. Calculate the population density of the initial anonymized population, and identify the initial anonymized population whose population density reaches the background fluctuation threshold as the candidate anonymized population.

[0039] In one specific implementation, a group segmentation procedure divides the dynamic co-occurrence hypergraph to obtain candidate anonymity groups. During segmentation, the procedure reads the hyperedge connections between anonymous session nodes in the dynamic co-occurrence hypergraph and clusters anonymous session nodes with high connection strength in spatiotemporal proximity and semantic task complementarity into the same group. For a candidate anonymity group, if the anonymous session nodes within it are only accidentally close in time, and the business content is highly repetitive and lacks complementarity, the subsequent intent detection loop will reduce its retention probability; if the anonymous session nodes within it are complementary in multiple business processes, and the active time slices have stable overlap, it is more consistent with the characteristics of internal collaborative research within an enterprise. Therefore, the segmentation of the dynamic co-occurrence hypergraph is not simply based on the number of visits, but rather uses both spatiotemporal feature data and semantic trajectory feature data as the basis for group formation.

[0040] Specifically, after the dynamic co-occurrence hypergraph is formed, the group segmentation program reads the anonymous session nodes, spatiotemporal hyperedges, semantic hyperedges, spatiotemporal connectivity weights, and semantic task complementarity weights from the hypergraph. It then groups the anonymous session nodes according to these weights to obtain initial anonymous groups. Grouping can be optimized using a weighted modularity approach, treating each anonymous session node as a target for assignment, and combining the spatiotemporal connectivity weights and semantic task complementarity weights into a group attraction value between nodes. The group attraction value can be calculated as follows: Calculate, where, Indicates group attraction value. Indicates the spatiotemporal connectivity weights. Indicates the complementary weights of semantic tasks. This indicates the proportion of spatiotemporal connectivity weights. The value can be between 0.4 and 0.6. In this implementation, 0.5 is used to keep the spatiotemporal connectivity weight and the semantic task complementary weight balanced in the initial grouping. The group segmentation procedure iteratively adjusts the group affiliation of anonymous session nodes until the weighted modularity increment after two consecutive rounds of adjustment is less than the preset convergence threshold, which can be 0.001. At this point, the initial anonymous group is output.

[0041] To reduce misgrouping caused by random co-occurrence, the group segmentation program scrambles the timestamp sequence in the anonymous session logs to generate background session data. This scrambling only alters the time correspondence between different anonymous sessions, without changing the page business tag sequence and page interaction sequence within a single anonymous session. This results in background session data that retains the overall access frequency but breaks down the actual time-series correlations. Based on this background session data, the group segmentation program constructs a background co-occurrence hypergraph and calculates the background group density within it. Group density can be calculated by... Calculate, where, Indicates population density. This represents the sum of spatiotemporal connectivity weights and semantic task complementarity weights within the group. This represents the number of anonymous session nodes within the group. Background group density represents the average connection strength that may form under random temporal relationships.

[0042] Subsequently, the group segmentation program generates a background fluctuation threshold based on the background group density and a preset fluctuation coefficient. The background fluctuation threshold can be set according to... Calculate, where, Indicates the background fluctuation threshold. Indicates background population density. This indicates the preset volatility coefficient. The value can be between 0.15 and 0.50; in this embodiment, it is 0.30. The group segmentation procedure calculates the group density of the initial anonymized group and identifies the initial anonymized group whose group density reaches the background fluctuation threshold as candidate anonymized groups. This screening makes the candidate anonymized groups have a more obvious internal correlation strength relative to the random background.

[0043] in, The background population density fluctuation range in the background co-occurrence hypergraph can be used to confirm the data. During implementation, the population segmentation program repeatedly rearranges the timestamp sequence in the anonymous session log to generate multiple sets of background session data, and constructs background co-occurrence hypergraphs for each set, calculating the background population density in each hypergraph. Subsequently, the mean and standard deviation of these background population densities are calculated, or their quantile distribution is analyzed, incorporating the natural fluctuations caused by random co-occurrence into the threshold setting. If the mean and standard deviation method is used, it can be set... Related to the ratio of standard deviation to mean, for example ,in, This represents the standard deviation of the background population density over multiple iterations. This represents the mean of the background population density. Represents the confidence coefficient. A value of 1 to 2 can be used; if the quantile method is used, the relative difference between the 90th or 95th percentile of the background population density and the mean can be taken as the mean. The reference value obtained in this way. It can reflect the normal fluctuations in population density under random backgrounds, avoiding the retention of accidental co-occurrences due to excessively low background fluctuation thresholds, and also avoiding the filtering of genuine candidate anonymized groups due to excessively high background fluctuation thresholds; in sites with stable traffic, A value of 0.30 is usually acceptable.

[0044] Calculate the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and fuse them to obtain the internal collaboration score; Furthermore, the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group are calculated and fused to obtain the internal collaboration score, including: Extract the active time slices of each anonymous session node within the candidate anonymous group; The overlap of active time slices is statistically analyzed to obtain the degree of overlap of active time for the candidate anonymized group; Extract the entity relationship path and semantic task label of each anonymous session node within the candidate anonymous group; Calculate the coverage distribution of entity relationship paths in different business processes to obtain task chain coverage data; Based on the repetition ratio of task chain coverage data and semantic task labels, a semantic complementarity index is generated; The internal collaboration score is obtained by weighting the overlap of active time and the semantic complementarity index.

[0045] In one specific implementation, after candidate anonymity groups are generated, the intent detection program enters an intent detection loop. For each candidate anonymity group, the intent detection program first calculates the active time overlap and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and then fuses them to obtain an internal collaboration score. The active time overlap is used to measure the degree of overlap of anonymous session nodes within the candidate anonymity group in active time slices, and the semantic complementarity index is used to measure the degree of complementarity of anonymous session nodes within the candidate anonymity group in terms of technology adaptation paths, cost assessment paths, implementation and delivery paths, and operation and maintenance support paths. The internal collaboration score can be calculated using the formula... Calculate, where, Indicates the internal collaboration score. Indicates the degree of overlap in active times. Indicates the semantic complementarity index. This indicates the weight corresponding to the overlap of active times. This represents the weight corresponding to the semantic complementarity index, and In actual deployment, A value of 0.35 to 0.55 is acceptable. A value of 0.45 to 0.65 is acceptable; in industries where enterprise customers have long procurement chains, It can be set to 0.6 to make the semantic complementarity index contribute more to the internal collaboration score. All terms in the above formula are normalized to between 0 and 1 to avoid bias caused by different original units such as access duration and number of paths.

[0046] Specifically, within the intent detection loop, the calculation of the internal collaboration score can employ a more detailed process. The intent detection program extracts the active time slices of each anonymous session node within the candidate anonymity group and performs overlap statistics on these active time slices to obtain the active time overlap degree of the candidate anonymity group. If a candidate anonymity group contains n anonymous session nodes, the intent detection program can calculate the ratio of the intersection duration to the union duration of the active time slices of any two anonymous session nodes, and then take the average of all node pairs as the active time overlap degree. The formula can be written as follows: ,in, Indicates the degree of overlap in active times. Indicates the first The anonymous session node and the first The duration of the intersection of the active time slices of anonymous session nodes. This indicates the duration of the union of the two active time slices. This indicates the number of node pairs. The closer this value is to 1, the more likely the anonymous session nodes within the candidate anonymous group are to be active together in similar time periods.

[0047] Simultaneously, the intent detection program extracts the entity relationship paths and semantic task tags of each anonymous session node within the candidate anonymity group, calculates the coverage distribution of entity relationship paths in different business stages, and obtains task chain coverage data. Task chain coverage data can be represented as a four-dimensional vector; for example, [1,1,0,1] represents that the candidate anonymity group covers the technology adaptation path, cost assessment path, and operation and maintenance support path, but not the implementation and delivery path. The intent detection program generates a semantic complementarity index based on the repetition ratio of the task chain coverage data and semantic task tags. The semantic complementarity index can be calculated according to... Calculate, where, Indicates the semantic complementarity index. This indicates the number of business processes that have been covered. This indicates the total number of preset business processes. This indicates the repetition rate of semantic task tags. Next, the intent detection program weights the active time overlap and semantic complementarity index to obtain an internal collaboration score. Since the internal collaboration score includes both time overlap and semantic complementarity, the two types of data are cross-validated within the same candidate anonymization group, reducing the bias of judging intent solely based on the number of views.

[0048] When the internal collaboration score is lower than the basic threshold, the corresponding candidate anonymization group is removed and the remaining candidate anonymization groups are detected. When the internal collaboration score reaches the basic threshold, obtain publicly available dynamic information of enterprises associated with candidate anonymous group network segments and extract external intent vectors. The semantic trajectory feature data of the candidate anonymized group is compared with the external intent vector to calculate the spatial vector similarity to obtain the external verification score. The internal collaboration score and the external verification score are then combined to obtain the comprehensive procurement intent score. Furthermore, obtain publicly available dynamic information about enterprises associated with candidate anonymized group network segments, and extract external intent vectors, including: Read the network segment environment identifiers corresponding to the candidate anonymity groups; Based on the network segment environment identifier, retrieve the network segment belonging organization data from the public network registration data to obtain the organization candidate set; Extract bidding announcements, job postings, and company news data corresponding to the candidate organization data from publicly available commercial data sources to form publicly available dynamic information about the companies. Text cleaning is performed on publicly available corporate information to extract business keywords and behavioral action words; By mapping business subject terms and behavioral action terms to the industry knowledge graph, external business entity nodes are obtained. Generate external intent vectors based on the distribution of external business entity nodes in the industry knowledge graph.

[0049] In one specific implementation, the base threshold is determined jointly by historical confirmed business opportunity data and historical invalid access data. The intent detection program can sort the internal collaboration scores corresponding to the historical confirmed business opportunity data from low to high, take the 20th percentile as the reference lower limit of the base threshold, and then correct it by combining the 80th percentile of the internal collaboration scores in the historical invalid access data. For example, if the 20th percentile of the historical confirmed business opportunity data is 0.62 and the 80th percentile of the historical invalid access data is 0.54, then the base threshold can be set to 0.60. If the internal collaboration score of the current candidate anonymized group is lower than the base threshold, the intent detection program removes the corresponding candidate anonymized group and continues to detect the remaining candidate anonymized groups. The removed candidate anonymized groups are only retained as statistical records and are not included in the retrieval process of the company's publicly available dynamic information. This processing ensures that the retrieval of the company's publicly available dynamic information focuses on data objects with internal collaboration relationships that have reached a certain strength, reducing invalid searches and subsequent misallocation of marketing resources.

[0050] When the internal collaboration score reaches a basic threshold, the intent detection program acquires publicly available dynamic information about enterprises associated with the candidate anonymity group's network segments and extracts external intent vectors. The candidate anonymity group's network segments originate from the network environment identifiers of anonymous session nodes within the candidate anonymity group. Publicly available dynamic information about enterprises can be derived from public network registration data, bidding announcements, job postings, and company news. The intent detection program does not directly use this publicly available dynamic information as a conclusion; instead, it cleans and converts it into external intent vectors, which are then used to calculate spatial vector similarity with the semantic trajectory feature data of the candidate anonymity group. Spatial vector similarity can be calculated using cosine similarity, as shown in the formula: ,in, Indicates the external verification score. This represents a group semantic vector formed by the semantic trajectory feature data of the candidate anonymized group. Represents the external intent vector. This represents the sum of the products of the corresponding dimensions of two vectors. and These represent the magnitudes of the two vectors, respectively. Cosine similarity takes topic orientation as the primary comparison object and is not directly affected by text length and access frequency, making it suitable for comparing the business consistency between anonymous access trajectories and publicly available corporate information.

[0051] Subsequently, the intent detection program integrates internal collaboration scores and external corroboration scores to obtain a comprehensive procurement intent score. The comprehensive procurement intent score can be calculated using the formula... Calculate, where, Indicates the overall procurement intention. Indicates the internal collaboration score. Indicates the external verification score. This indicates the weight corresponding to the internal collaboration score. This indicates the weight corresponding to the external corroboration score, and In one embodiment, Take 0.55, A value of 0.45 is used to allow both the collaborative relationships between internal anonymous session nodes and the corroboration relationships of publicly available corporate information to participate in the judgment. The action threshold can be determined based on the effective conversion samples after historical marketing resource investment. For example, the 30th percentile of the comprehensive purchase intent score corresponding to the historical effective conversion samples can be set as the action threshold. If the value is 0.72, then the candidate anonymous group with a comprehensive purchase intent score of 0.72 will enter the group virtual profile generation process.

[0052] Specifically, regarding the introduction of external data, the intent detection program first reads the network segment environment identifier corresponding to the candidate anonymous group. The IP network segment and border gateway protocol routing information in the network segment environment identifier are used to retrieve the network segment's affiliated organization data from public network registration data, obtaining an organizational candidate set. Public network registration data may include Internet Number Resource Registration Information, Autonomous System Registration Information, and public WHOIS information. Since the same network segment may correspond to multiple entities such as group companies, branches, and cloud service exits, the organizational candidate set can retain multiple candidate organizations, and an attribution confidence score is written for each candidate organization. The attribution confidence score can be calculated jointly by the IP network segment matching accuracy, the consistency of border gateway protocol routing information, and the frequency of occurrence of the organization name.

[0053] The intent detection program extracts bidding announcement data, job posting data, and corporate news data corresponding to the candidate set of organizations from publicly available commercial data sources, forming publicly available dynamic information about the enterprises. Bidding announcement data may include project name, procurement category, release time, and bidding keywords; job posting data may include job title, job responsibilities, and skill requirements; corporate news data may include news headlines, text summaries, and release time. To avoid interference from outdated information, the intent detection program can set a valid time window, such as collecting only content from publicly available commercial data sources within the last 90 days or 180 days. Subsequently, the intent detection program performs text cleaning on the publicly available dynamic information of the enterprises, extracting business keywords and behavioral action words. Text cleaning includes removing announcement template words, commonly used words in company registration, and conjunctions without business meaning; business keywords reflect the content of the object, such as data governance, customer management, and interface platform; behavioral action words reflect the action tendency, such as procurement, construction, upgrade, recruitment, and expansion. Business keywords and behavioral action words are mapped to an industry knowledge graph to obtain external business entity nodes.

[0054] Finally, the intent detection program generates external intent vectors based on the distribution of external business entity nodes in the industry knowledge graph. Each dimension of the external intent vector corresponds to a business entity node or business process, and the dimension value can be obtained by weighting the frequency of occurrence, the publication time decay coefficient, and the confidence level of the organization affiliation. The publication time decay coefficient can be calculated by... Calculate, where, This represents the decay coefficient over time. This indicates the number of days between the current date and the date the information was published. The time decay parameter τ can range from 30 to 90 days; in this implementation, 60 days is used. This gives recently released information a higher weight in the external intent vector.

[0055] When the overall procurement intention score is lower than the action threshold, increase the semantic task complementarity weights in the hyperedge and re-segment the dynamic co-occurrence hypergraph.

[0056] When the overall procurement intent score reaches the action threshold, the behavioral characteristics of the corresponding candidate anonymous group are aggregated to generate a virtual profile of the group, and marketing resource scheduling instructions matching the network segment corresponding to the candidate anonymous group are output.

[0057] Furthermore, when the overall procurement intention score is lower than the action threshold, the semantic task complementarity weights in the hyperedge are increased and the dynamic co-occurrence hypergraph is re-segmented, including: Read the internal collaboration score and external corroboration score of anonymous candidate groups whose overall procurement intention score is below the action threshold; When the internal collaboration score reaches the basic threshold and the external verification score is lower than the verification threshold, semantic weight correction data is generated. Based on semantic weight correction data, the complementary weight of semantic tasks in semantic hyperedges is increased, and the connection weight of network segment environment identifiers in spatiotemporal hyperedges is decreased. Reconstruct the dynamic co-occurrence hypergraph based on the corrected semantic hyperedges and spatiotemporal hyperedges; The reconstructed dynamic co-occurrence hypergraph is segmented again to generate new candidate anonymity groups, which are then fed into the intent detection loop.

[0058] Furthermore, the system aggregates the behavioral characteristics of the corresponding candidate anonymous groups to generate a virtual profile of the group, and outputs marketing resource scheduling instructions that match the network segments corresponding to the candidate anonymous groups, including: Extract the number of anonymous session nodes, network segment environment identifiers, semantic task tags, page interaction sequences, and comprehensive procurement intent scores from candidate anonymous groups that have reached the action threshold; Group size data is generated based on the overlap between the number of anonymous session nodes and active time slices. Generate topic attention distribution data based on semantic task tags and page interaction sequences; Procurement stage data is generated based on the order of coverage of entity relationship paths in different business processes. By combining group size data, topic focus distribution data, procurement stage data, and comprehensive procurement intent, a virtual profile of the group is generated. Based on the network segment environment identifiers, topic attention distribution data, and procurement stage data in the group's virtual profile, marketing resource scheduling instructions are generated.

[0059] In one specific implementation, when the overall procurement intent score is lower than the action threshold, the intent detection program does not immediately output a marketing resource scheduling instruction. Instead, it increases the semantic task complementarity weights in the hyperedges and re-segments the dynamic co-occurrence hypergraph. This step addresses a common scenario: the semantic task complementarity within the candidate anonymity group is significant, but due to unstable network segment affiliation, changes in proxy exits, or delayed updates to public information, the external verification score is insufficient, causing the overall procurement intent score to temporarily fall short of the action threshold. In this case, the hypergraph construction program adjusts the hyperedge weights in the dynamic co-occurrence hypergraph based on the correction results output by the intent detection program, making the re-segmentation more focused on semantic task complementarity relationships rather than over-relying on the spatiotemporal proximity of a single network segment. If the overall procurement intent score reaches the action threshold, the marketing resource scheduling program aggregates the behavioral characteristics of the corresponding candidate anonymity group to generate a virtual profile of the group and outputs a marketing resource scheduling instruction matching the network segment corresponding to the candidate anonymity group. The marketing resource scheduling instruction may include fields such as network segment environment identifier, topic attention distribution data, procurement stage data, recommended content type, and resource priority. Its output target is the marketing resource queue within the platform, rather than performing control operations on the user terminal.

[0060] Specifically, when the overall procurement intent score is lower than the action threshold, the intent detection program reads the internal collaboration score and external corroboration score of the candidate anonymized group whose overall procurement intent score is lower than the action threshold. If the internal collaboration score reaches the base threshold and the external corroboration score is lower than the corroboration threshold, the intent detection program generates semantic weight correction data. The corroboration threshold can be determined based on the 25th percentile of the external corroboration score in historical confirmed business opportunity data, for example, 0.50. The semantic weight correction data may include the semantic task complementarity weight increment, the spatiotemporal connectivity weight deduction, and the reconstruction marker. The semantic task complementarity weight increment can be calculated according to... Calculate, where, This represents the incremental weights for complementary semantic tasks. This represents the correction factor. A value of 0.1 to 0.3 is acceptable. Indicates the internal collaboration score. This indicates the external verification score; the deduction of spatiotemporal connectivity weights can be calculated according to... Calculate, where, This represents the reduction in spatiotemporal connectivity weights. This represents the deduction factor. A value of 0.05 to 0.2 is acceptable. This correction logic expresses the following data processing implication: when there is obvious internal collaboration within the candidate anonymity group but insufficient external corroboration, the re-segmentation should appropriately increase the importance of semantic complementarity and reduce the rigid binding brought about by network segment environment identification.

[0061] in, and Confirmation can be achieved through calibration experiments using historically confirmed samples: First, anonymous groups that have been manually confirmed or verified by subsequent conversion links are considered positive samples, while randomly co-occurring sessions on the same network segment but without business association, or sessions triggered by a single external clue, are considered negative samples. Their internal coordination scores are then calculated. External verification score And the group identification results before and after correction; for The value can be iterated within the range of 0.1 to 0.3 with a step size of 0.02 or 0.05 to observe the impact of the correction term on the final group confidence. Priority should be given to values ​​that improve positive sample recall without significantly increasing the false positive rate of negative samples. If the external corroborating data source is stable and highly reliable, then... A value close to 0.3 can be taken. However, if the external verification data is sparse or noisy, then... Take a value close to 0.1 to avoid external factors excessively influencing internal collaborative judgment; for It can be traversed within the range of 0.05 to 0.2, and will Substitute the spatiotemporal connectivity weight reduction process and compare the differences. The following network segments have low false aggregation rate, true collaborative group retention rate, and overall F1 score. If there are many mis-aggregations in the sample, then improve... To enhance deductions, if the actual enterprise survey sample has insufficient external corroboration, the deduction will be reduced. To avoid excessively weakening the spatiotemporal connectivity weights, the final set selected should maximize the AUC or F1 score on the validation set while maintaining a false merging rate below the business tolerance limit. and And the newly confirmed samples are reviewed and adjusted monthly or quarterly.

[0062] The hypergraph construction program adjusts the data based on semantic weights, increasing the semantic task complementarity weights in semantic hyperedges and decreasing the connection weights of network segment environment identifiers in spatiotemporal hyperedges. The adjusted semantic task complementarity weights do not exceed a preset upper limit, which can be 1; the adjusted spatiotemporal connection weights do not fall below a preset lower limit, which can be 0. Subsequently, the hypergraph construction program reconstructs a dynamic co-occurrence hypergraph according to the adjusted semantic and spatiotemporal hyperedges. The group segmentation program further segments the reconstructed dynamic co-occurrence hypergraph, generating new candidate anonymity groups, and sends these new candidate anonymity groups into the intent detection loop. Loop termination conditions may include the new candidate anonymity group reaching an action threshold, the candidate anonymity group being removed, or the number of reconstructions reaching a preset upper limit. The preset upper limit can be two to three times to avoid infinite loops due to insufficient external data. This processing allows the same batch of anonymous session nodes to still have the opportunity to be reorganized according to semantic task complementarity relationships when network segment relationships are unstable, thus better reflecting the actual process of multi-person collaborative research in enterprises.

[0063] When the overall purchase intent score of a candidate anonymous group reaches the action threshold, the marketing resource scheduling program begins to aggregate the behavioral characteristics of the corresponding candidate anonymous group to generate a virtual profile of the group. The marketing resource scheduling program extracts the number of anonymous session nodes, network segment environment identifiers, semantic task tags, page interaction sequences, and overall purchase intent score from the candidate anonymous group that has reached the action threshold. The overlap between the number of anonymous session nodes and active time slices is used to generate group size data, which may include fields such as the number of anonymous session nodes, the peak number of simultaneously active nodes, and the average active time slice length. For example, if a candidate anonymous group contains eight anonymous session nodes, and five of these nodes overlap during the same morning period, then the group size data indicates that this candidate anonymous group is not a single, occasional visitor.

[0064] The marketing resource scheduling program generates topic attention distribution data based on semantic task tags and page interaction sequences. This data can be weighted by business topic dimension. For example, the page interaction sequence of each anonymous session node is converted into access weights for entity relationship paths, then aggregated by semantic task tags, resulting in a distribution of 0.38 for the technology adaptation path, 0.27 for the cost assessment path, 0.18 for the implementation and delivery path, and 0.17 for the operation and maintenance support path. If a data download occurs on a page corresponding to the technology adaptation path, the access weight of that path is increased accordingly; if it's just a brief visit, the weight is lower. The marketing resource scheduling program also generates procurement stage data based on the coverage order of entity relationship paths in different business stages. If the anonymous candidate group first visits the technology adaptation path, and then the cost assessment and implementation and delivery paths, the procurement stage data can be labeled as the solution comparison stage; if the cost assessment path and contract-related entity relationship paths have higher weights, the procurement stage data can be labeled as the budget confirmation stage. This labeling is determined by the order and access weight of entity relationship paths within active time slices, and does not rely on subjective manual input.

[0065] Finally, the marketing resource scheduling program merges group size data, topic attention distribution data, procurement stage data, and comprehensive procurement intent score to generate a virtual group profile. This virtual profile can be stored in a structured record format, including network segment environment identifiers, the number of anonymous session nodes, topic attention distribution data, procurement stage data, comprehensive procurement intent score, and update time. Based on the network segment environment identifiers, topic attention distribution data, and procurement stage data in the virtual group profile, the marketing resource scheduling program generates marketing resource scheduling instructions. If the technical adaptation path has the highest proportion in the topic attention distribution data and the procurement stage data is in the solution comparison phase, the marketing resource scheduling instruction can target the content queues of technical white papers, compatibility specifications, and industry case studies. If the cost assessment path has an increased proportion and the procurement stage data is in the budget confirmation phase, the marketing resource scheduling instruction can target the content queues of quotation scheme descriptions, authorization model descriptions, and implementation cycle descriptions. The marketing resource scheduling instruction carries the corresponding network segment of the candidate anonymous group, recommended content type, resource priority, and effective time window. This information is read by the resource scheduling program within the marketing management platform and used to arrange subsequent content delivery, sales lead ranking, and enterprise customer follow-up plans. Throughout the process, the spatiotemporal feature data and semantic trajectory feature data of anonymous session logs are dynamically co-occurred to form candidate anonymous groups through a hypergraph. The company’s publicly available dynamic information is used to externally verify the candidate anonymous groups. The virtual profile of the groups then receives comprehensive procurement intent and outputs marketing resource scheduling instructions. Each data object is continuously used in the preceding and following steps, avoiding the judgment gaps caused by isolated single-point data.

[0066] Example 2: This application also discloses a marketing management system based on big data.

[0067] Reference Figure 2 A big data-based marketing management system includes: The data extraction module is used to obtain anonymized session logs and extract spatiotemporal feature data and semantic trajectory feature data from the anonymized session logs. The data mapping module is used to map spatiotemporal feature data and semantic trajectory feature data into anonymous session nodes in a dynamic co-occurrence hypergraph, and generate hyperedges connecting multiple anonymous session nodes according to spatiotemporal proximity and semantic task complementarity. The detection loop module is used to segment the dynamic co-occurrence hypergraph to obtain candidate anonymity groups, and then enter the intent detection loop. The intent detection loop includes: Calculate the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and fuse them to obtain the internal collaboration score; When the internal collaboration score is lower than the basic threshold, the corresponding candidate anonymization group is removed and the remaining candidate anonymization groups are detected. When the internal collaboration score reaches the basic threshold, obtain publicly available dynamic information of enterprises associated with candidate anonymous group network segments and extract external intent vectors. The semantic trajectory feature data of the candidate anonymized group is compared with the external intent vector to calculate the spatial vector similarity to obtain the external verification score. The internal collaboration score and the external verification score are then combined to obtain the comprehensive procurement intent score. When the overall procurement intention score is lower than the action threshold, increase the semantic task complementarity weights in the hyperedge and re-segment the dynamic co-occurrence hypergraph. The management output module is used to aggregate the behavioral characteristics of the corresponding candidate anonymous group to generate a virtual profile of the group when the comprehensive procurement intent score reaches the action threshold, and output marketing resource scheduling instructions that match the network segment corresponding to the candidate anonymous group.

[0068] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0069] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0070] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A marketing management method based on big data, characterized in that, Includes the following steps: Obtain anonymized session logs after desensitization, and extract spatiotemporal feature data and semantic trajectory feature data from the anonymized session logs; Spatiotemporal feature data and semantic trajectory feature data are mapped to anonymous session nodes in a dynamic co-occurrence hypergraph, and hyperedges connecting multiple anonymous session nodes are generated according to spatiotemporal proximity and semantic task complementarity. The dynamic co-occurrence hypergraph is segmented to obtain candidate anonymity groups, which are then entered into the intent detection loop. The intent detection loop includes: Calculate the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and fuse them to obtain the internal collaboration score; When the internal collaboration score is lower than the basic threshold, the corresponding candidate anonymization group is removed and the remaining candidate anonymization groups are detected. When the internal collaboration score reaches the basic threshold, obtain publicly available dynamic information of enterprises associated with candidate anonymous group network segments and extract external intent vectors. The semantic trajectory feature data of the candidate anonymized group is compared with the external intent vector to calculate the spatial vector similarity to obtain the external verification score. The internal collaboration score and the external verification score are then combined to obtain the comprehensive procurement intent score. When the overall procurement intention score is lower than the action threshold, increase the semantic task complementarity weights in the hyperedge and re-segment the dynamic co-occurrence hypergraph. When the overall procurement intent score reaches the action threshold, the behavioral characteristics of the corresponding candidate anonymous group are aggregated to generate a virtual profile of the group, and marketing resource scheduling instructions matching the network segment corresponding to the candidate anonymous group are output.

2. The marketing management method based on big data according to claim 1, characterized in that, Obtain the anonymized session logs and extract spatiotemporal feature data and semantic trajectory feature data from them, including: Parse anonymous session logs and separate them into underlying network interaction records and front-end page browsing records; For the underlying network interaction records, extract IP network segments, border gateway protocol routing information, device environment hash data, and timestamp sequences; The IP network segment, border gateway protocol routing information, and device environment hash data are merged to generate a network segment environment identifier; Active time slices are segmented according to the timestamp sequence, and network segment environment identifiers are associated with active time slices to generate spatiotemporal feature data; Extract search keyword sequences, page business tag sequences, and page interaction sequences from the front-end page browsing history; Map search keyword sequences and page business tag sequences to industry knowledge graphs to obtain entity relationship paths; Using the page interaction sequence as the access weight of the entity relationship path, semantic trajectory feature data is generated.

3. The marketing management method based on big data according to claim 2, characterized in that, By mapping search keyword sequences and page business tag sequences to a pre-built industry knowledge graph, entity relationship paths are obtained, including: The search keyword sequence is segmented and cleaned to obtain keyword units; Synonymize keyword units with page business tag sequences to generate business semantic units; Retrieve the business entity node corresponding to the business semantic unit in the industry knowledge graph; Extract the shortest business relationship chain between adjacent business entity nodes to generate entity relationship paths; Based on the business steps involved in the entity relationship path, the entity relationship path is marked as a technology adaptation path, a cost assessment path, an implementation and delivery path, and an operation and maintenance support path. The annotation results of each path are summarized into semantic task labels, and the semantic task labels and entity relationship paths together form semantic trajectory feature data.

4. The marketing management method based on big data according to claim 3, characterized in that, Spatiotemporal feature data and semantic trajectory feature data are mapped to anonymous session nodes in a dynamic co-occurrence hypergraph. Hyperedges connecting multiple anonymous session nodes are generated according to spatiotemporal proximity and semantic task complementarity, including: Anonymous session nodes are established based on a single anonymous session as the basic unit, and spatiotemporal feature data is written into the spatiotemporal attribute field of the anonymous session node; Write semantic trajectory feature data into the semantic attribute field of the anonymous session node; Based on the network segment environment identifier and active time slice in the spatiotemporal attribute field, a spatiotemporal hyperedge connecting multiple anonymous session nodes is generated, and spatiotemporal connection weights are written to the spatiotemporal hyperedge. Based on the semantic task labels and entity relationship paths in the semantic attribute fields, generate semantic hyperedges that connect multiple anonymous session nodes, and write complementary weights of semantic tasks to the semantic hyperedges. By writing spatiotemporal hyperedges and semantic hyperedges into the same hypergraph structure, a dynamic co-occurrence hypergraph is obtained.

5. A marketing management method based on big data according to claim 4, characterized in that, Generate semantic hyperedges connecting multiple anonymous session nodes, including: Read the semantic task tags of each anonymous session node; When the semantic task labels of multiple anonymous session nodes belong to the same business topic, and each semantic task label covers different business links, the multiple anonymous session nodes are merged into a candidate complementary node group. The percentage of duplicate entity relationship paths and the percentage of business process coverage within candidate complementary node groups are statistically analyzed. Based on the repetition ratio, homogeneous behavior deduction is performed on candidate complementary node groups, and semantic task complementary weights are generated based on the coverage ratio of business links. The semantic task complementary weights are written into the superedges corresponding to the candidate complementary node groups to obtain the semantic superedges.

6. The marketing management method based on big data according to claim 5, characterized in that, Segmenting the dynamic co-occurrence hypergraph yields candidate anonymity groups, including: Read anonymous session nodes, spatiotemporal hyperedges, semantic hyperedges, spatiotemporal connection weights, and semantic task complementary weights from the dynamic co-occurrence hypergraph; Anonymous session nodes are grouped according to spatiotemporal connectivity weights and semantic task complementary weights to obtain an initial anonymous group; The timestamp sequence in the anonymous session log is rearranged to generate background session data. Construct a background co-occurrence hypergraph based on background session data, and calculate the background population density in the background co-occurrence hypergraph; A background fluctuation threshold is generated based on the background population density and a preset fluctuation coefficient. Calculate the population density of the initial anonymized population, and identify the initial anonymized population whose population density reaches the background fluctuation threshold as the candidate anonymized population.

7. A marketing management method based on big data according to claim 3, characterized in that, Calculate the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and fuse them to obtain the internal collaboration score, including: Extract the active time slices of each anonymous session node within the candidate anonymous group; The overlap of active time slices is statistically analyzed to obtain the degree of overlap of active time for the candidate anonymized group; Extract the entity relationship path and semantic task label of each anonymous session node within the candidate anonymous group; Calculate the coverage distribution of entity relationship paths in different business processes to obtain task chain coverage data; Based on the repetition ratio of task chain coverage data and semantic task labels, a semantic complementarity index is generated; The internal collaboration score is obtained by weighting the overlap of active time and the semantic complementarity index.

8. A marketing management method based on big data according to claim 2, characterized in that, Obtain publicly available dynamic information about enterprises associated with candidate anonymized group network segments, and extract external intent vectors, including: Read the network segment environment identifiers corresponding to the candidate anonymity groups; Based on the network segment environment identifier, retrieve the network segment belonging organization data from the public network registration data to obtain the organization candidate set; Extract bidding announcements, job postings, and company news data corresponding to the candidate organization data from publicly available commercial data sources to form publicly available dynamic information about the companies. Text cleaning is performed on publicly available corporate information to extract business keywords and behavioral action words; By mapping business subject terms and behavioral action terms to the industry knowledge graph, external business entity nodes are obtained. Generate external intent vectors based on the distribution of external business entity nodes in the industry knowledge graph.

9. A marketing management method based on big data according to claim 6, characterized in that, When the overall procurement intention score is lower than the action threshold, the semantic task complementarity weights in the hyperedges are increased and the dynamic co-occurrence hypergraph is re-segmented, including: Read the internal collaboration score and external corroboration score of anonymous candidate groups whose overall procurement intention score is below the action threshold; When the internal collaboration score reaches the basic threshold and the external verification score is lower than the verification threshold, semantic weight correction data is generated. Based on semantic weight correction data, the complementary weight of semantic tasks in semantic hyperedges is increased, and the connection weight of network segment environment identifiers in spatiotemporal hyperedges is decreased. Reconstruct the dynamic co-occurrence hypergraph based on the corrected semantic hyperedges and spatiotemporal hyperedges; The reconstructed dynamic co-occurrence hypergraph is segmented again to generate new candidate anonymity groups, which are then fed into the intent detection loop.

10. A marketing management method based on big data according to claim 7, characterized in that, Aggregate the behavioral characteristics of the corresponding candidate anonymous groups to generate a virtual profile of the group, and output marketing resource scheduling instructions that match the network segments corresponding to the candidate anonymous groups, including: Extract the number of anonymous session nodes, network segment environment identifiers, semantic task tags, page interaction sequences, and comprehensive procurement intent scores from candidate anonymous groups that have reached the action threshold; Group size data is generated based on the overlap between the number of anonymous session nodes and active time slices. Generate topic attention distribution data based on semantic task tags and page interaction sequences; Procurement stage data is generated based on the order of coverage of entity relationship paths in different business processes. By combining group size data, topic focus distribution data, procurement stage data, and comprehensive procurement intent, a virtual profile of the group is generated. Based on the network segment environment identifiers, topic attention distribution data, and procurement stage data in the group's virtual profile, marketing resource scheduling instructions are generated.

11. A big data-based marketing management method, applied to the big data-based marketing management system described in any one of claims 1-10, characterized in that, include: The data extraction module is used to obtain anonymized session logs and extract spatiotemporal feature data and semantic trajectory feature data from the anonymized session logs. The data mapping module is used to map spatiotemporal feature data and semantic trajectory feature data into anonymous session nodes in a dynamic co-occurrence hypergraph, and generate hyperedges connecting multiple anonymous session nodes according to spatiotemporal proximity and semantic task complementarity. The detection loop module is used to segment the dynamic co-occurrence hypergraph to obtain candidate anonymity groups, and then enter the intent detection loop. The intent detection loop includes: Calculate the overlap of active time and semantic complementarity index of anonymous session nodes within the candidate anonymity group, and fuse them to obtain the internal collaboration score; When the internal collaboration score is lower than the basic threshold, the corresponding candidate anonymization group is removed and the remaining candidate anonymization groups are detected. When the internal collaboration score reaches the basic threshold, obtain publicly available dynamic information of enterprises associated with candidate anonymous group network segments and extract external intent vectors. The semantic trajectory feature data of the candidate anonymized group is compared with the external intent vector to calculate the spatial vector similarity to obtain the external verification score. The internal collaboration score and the external verification score are then combined to obtain the comprehensive procurement intent score. When the overall procurement intention score is lower than the action threshold, increase the semantic task complementarity weights in the hyperedge and re-segment the dynamic co-occurrence hypergraph. The management output module is used to aggregate the behavioral characteristics of the corresponding candidate anonymous group to generate a virtual profile of the group when the comprehensive procurement intent score reaches the action threshold, and output marketing resource scheduling instructions that match the network segment corresponding to the candidate anonymous group.

Images