A Method for Constructing Consumer Finance Sales Lead Networks Based on Large Language Models

Abstract

This invention provides a method for constructing a consumer finance sales lead network based on a large language model, belonging to the field of data processing technology. Specifically, it includes: a data access module that reads related data to obtain unstructured text data; a feature enhancement module that uses a semantic network construction strategy and a large language model to perform in-depth analysis of the unstructured text data, performing intent recognition, sentiment analysis, and topic extraction to generate semantic feature tags; a network dynamic construction module responsible for forming a dynamic, multi-dimensional relationship network; an inference module that generates natural language-described, actionable sales leads; and a reach strategy generation module that automatically generates or recommends personalized communication scripts based on the generated sales leads. Furthermore, based on the semantic network construction strategy and the recommendation matching of different marketing activities, it determines the user profile groups that require optimization of the semantic network construction strategy, thereby improving the matching degree of the recommendation processing.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a method for constructing a consumer finance sales lead network based on a large language model. Background Technology

[0002] In the consumer finance industry, accurately and efficiently identifying and converting sales leads is crucial for driving business growth. Currently, the industry's commonly used lead generation methods primarily rely on rule engines or traditional machine learning models. These methods typically have the following limitations:

[0003] First, the problem of data silos and fragmentation is serious. Customer data is scattered across multiple internal business systems (such as application, transaction, and repayment) and external partners (such as scenario platforms and data service providers), making it difficult to form a unified, real-time, and complete panoramic view of customers, resulting in a one-sided understanding of customers.

[0004] Secondly, the clue identification dimension is limited. Existing methods mostly rely on limited structured data (such as credit scores and income levels), making it difficult to effectively process and mine the deep intentions and potential demand signals contained in unstructured data (such as APP browsing paths and customer service dialogue texts).

[0005] Finally, the value of customer networks is not being fully utilized. Customers are not isolated individuals; their social relationships, device connections, and debt networks—these "network attributes"—are valuable information for assessing risk and identifying cross-selling opportunities. Traditional methods lack the ability to effectively construct and analyze such complex relationship networks.

[0006] Therefore, there is an urgent need for a technical solution that can integrate multi-source data, deeply understand customer intent, dynamically build relationship networks, and intelligently generate high-value sales leads, namely, a consumer finance sales lead network construction method based on a large language model. Summary of the Invention

[0007] To achieve the objectives of this invention, the following technical solution is adopted:

[0008] A consumer finance sales lead network construction system based on a large language model, specifically including:

[0009] Data access module: used to integrate related data from users and read related data to obtain unstructured text data;

[0010] Feature enhancement module: Based on the correlation between the user's user profile group and the user tags of marketing campaigns with different profile recommendation bias types, determine the user's semantic network construction strategy, use the semantic network construction strategy and large language model to perform in-depth analysis of unstructured text data, perform intent recognition, sentiment analysis and topic extraction, and generate semantic feature tags;

[0011] Network dynamic construction module: Define entities as network nodes, and use the reasoning ability of large models to construct explicit and implicit relationship edges, thereby forming a dynamic, multi-dimensional relationship network;

[0012] Reasoning module: Places the user in the relational network for contextual analysis, and generates actionable sales leads in natural language through multi-step reasoning;

[0013] Outreach strategy generation module: Based on the generated sales leads, automatically generate or recommend personalized communication scripts, and determine the user profile groups that need to be optimized based on the semantic network construction strategy and the recommendation matching of different marketing activities.

[0014] It should be noted that the associated data includes data from internal business systems, data from third-party credit reporting agencies, and user authorization behavior data.

[0015] Specifically, implicit relationship edges include "potential interest-related edges", "social influence edges", and "risk transmission edges".

[0016] Specifically, this application provides a method for constructing a consumer finance sales lead network based on a large language model, applied to the aforementioned consumer finance sales lead network construction system based on a large language model, specifically including:

[0017] S1 uses user profiles as a basis to divide users into different user profile groups. Based on the user data and semantic feature tags of the user profile groups and the correlation with marketing activities, it determines the real-time construction of profile groups within the user profile groups.

[0018] S2 determines the type of profile recommendation deviation for the marketing campaign based on the association between the marketing campaign and the semantic feature tags of different real-time profile groups;

[0019] S3 When the user profile group does not belong to the real-time profile building group, determine the semantic network construction strategy of the user profile group based on the association between the user profile group and the user tags of marketing activities with different profile recommendation deviation types. Based on the semantic network construction strategy and the recommendation matching of different marketing activities, determine the user profile group that needs to be optimized by the semantic network construction strategy.

[0020] Furthermore, the user profile is determined based on the parsing results of the user's associated data, including various types of semantic feature tags.

[0021] Furthermore, users are segmented into different user profile groups, specifically including:

[0022] Users with the same semantic feature labels are grouped into the same user profile group.

[0023] Furthermore, the user data of the user profile group includes the number of users in the user profile group.

[0024] Furthermore, the association between the semantic feature tags and the marketing campaign is determined based on the association between the semantic feature tags and the user tags of the marketing campaign.

[0025] Furthermore, the method for determining the real-time constructed user profile group within the user profile group is as follows:

[0026] The number of users in the user profile group is determined based on the user data of the user profile group;

[0027] Based on the association between the semantic feature tags of the user profile group and the marketing activities, determine the semantic feature tags associated with different marketing activities;

[0028] Based on the number of users in the user profile group and the semantic feature tags associated with different marketing campaigns, the real-time profile building group in the user profile group is determined.

[0029] The beneficial effects of this invention are as follows:

[0030] By breaking down data barriers and integrating multi-source heterogeneous data, we can build a "digital twin" that deeply understands customer intentions and risks. This goes beyond the analysis of a single individual, proactively building and utilizing complex customer relationship networks to discover group characteristics, centers of influence, and hidden cross-selling chains. It generates highly accurate and relevant sales leads and supporting strategies, directly empowering frontline marketing and sales teams, improving conversion rates, and reducing customer acquisition costs.

[0031] Through a multi-step screening process, the selection of groups is evaluated sequentially across four dimensions: group size, breadth of marketing connections, scarcity of marketing connections, and overall strategic weight. This ensures that the selected groups are not only substantial in size but also closely, uniquely, or with high value in relation to the marketing campaign. The core logic lies in using a tiered filtering mechanism to accurately identify high-priority groups from a massive user profile pool—those whose members, upon exhibiting new behavior, warrant immediate initiation of complex network construction and AI inference processes.

[0032] This system proactively decides whether to invest high-cost resources in building a real-time semantic network for each user profile by meticulously assessing the "value loss risk" and "feedback vacuum risk" caused by push delays between each user group and its associated marketing campaigns. Its core logic is to prioritize limited real-time computing resources to groups most likely to lack feedback data and suffer the greatest subsequent business losses due to slow push efficiency. Through proactive, deep network insights, it generates more accurate and context-rich leads to improve the potential conversion rate of a single push, thereby partially offsetting the negative impact of insufficient push volume and accumulating higher-quality early data for evaluating the reliability of LLM analysis. Essentially, under the constraint of push efficiency, this achieves a strategic substitution and compensation for "insight accuracy" in terms of "reach breadth."

[0033] When a marketing campaign with a low conversion rate (indicating a risk of model bias) is identified, manual verification depends on the campaign accumulating sufficient new user data. However, if the target user group of the campaign is largely under a semantic network construction strategy with low update efficiency, data feedback is slow, and manual verification will be delayed indefinitely. Therefore, this invention designs a decision-making process to accurately identify the group among the user groups associated with inefficient campaigns that has become a bottleneck for data feedback due to the inefficiency of the current strategy. By upgrading their strategy to a higher priority (e.g., setting them as a group for real-time profile building), the overall data update and push efficiency of the problematic campaign is systematically improved, thereby accelerating the attainment of the trigger condition for manual verification and enabling rapid model diagnosis and repair.

[0034] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0036] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0037] Figure 1 This is a framework diagram of a consumer finance sales lead network construction system based on a large language model;

[0038] Figure 2 This is a flowchart of a method for constructing a consumer finance sales lead network based on a large language model;

[0039] Figure 3 This is a flowchart illustrating the method for determining the real-time user profile group.

[0040] Figure 4 This is a flowchart illustrating the method for determining the semantic network construction strategy for user profile groups. Detailed Implementation

[0041] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0042] Example 1

[0043] like Figure 1 As shown, a consumer finance sales lead network construction system based on a large language model specifically includes:

[0044] Data access module: used to integrate related data from users and read related data to obtain unstructured text data;

[0045] Feature enhancement module: Based on the correlation between the user's user profile group and the user tags of marketing campaigns with different profile recommendation bias types, determine the user's semantic network construction strategy, use the semantic network construction strategy and large language model to perform in-depth analysis of unstructured text data, perform intent recognition, sentiment analysis and topic extraction, and generate semantic feature tags;

[0046] Network dynamic construction module: Define entities as network nodes, and use the reasoning ability of large models to construct relational edges and implicit relational edges, thereby forming a dynamic, multi-dimensional relational network;

[0047] Reasoning module: Places the user in the relational network for contextual analysis, and generates actionable sales leads in natural language through multi-step reasoning;

[0048] Outreach strategy generation module: Based on the generated sales leads, automatically generate or recommend personalized communication scripts, and determine the user profile groups that need to be optimized based on the semantic network construction strategy and the recommendation matching of different marketing activities.

[0049] It should be noted that the associated data includes data from internal business systems, data from third-party credit reporting agencies, and user authorization behavior data.

[0050] Specifically, implicit relationship edges include "potential interest-related edges", "social influence edges", and "risk transmission edges".

[0051] Specifically, the explicit relationship edge refers to the intuitive and clear relationship directly related to the user, such as transactions and applications. In one possible embodiment, explicit relationship edges are constructed between the customer and the channel and the product through accessed customer credit application, transaction and other data, that is, the correspondence between the customer and the product and the channel is constructed.

[0052] Specifically, the implicit relationship edges are based on LLM analysis of indirect, explicit relationships such as behavioral similarity ("browsing similar products"), textual relevance ("consulting similar questions"), and device or network association. For example, if user A and user B browse similar products within the same period, there is a potential interest between the users and the product, thus a "potential interest relationship edge" is constructed between user A and user B.

[0053] If User A and User B share the same contacts or are each other's emergency contacts, a social influence relationship exists between User A and User B, thus constructing a "social influence edge".

[0054] If user A and user B have a customer profile similarity that reaches a set threshold, their risk profiles are quite similar. A "risk transmission edge" will be established between user A and user B. This edge is used to provide timely warnings to user B when credit risk arises for user A, thereby reducing marketing dumping and preventing financial risks.

[0055] In one possible embodiment, the data access layer accesses customer transaction data from the credit card center, online loan application data, third-party credit reports, user app behavior event streams, and customer service system dialogue recordings converted to text data in real time or in batches through API interfaces, database connections, etc. The collection of the above data requires user authorization and consent.

[0056] The feature enhancement module is the core. First, it analyzes the customer service dialogue text, prompting users to "Analyze the following customer dialogue, extract their core financial needs, concerns about product features, and emotional tendencies, and output them as tags." For example, regarding the dialogue "Do you have any installment interest rate promotions this month? I want to buy a laptop for my daughter on installment," LLM can output tags such as: "Need: Education / Electronics Installment," "Concerns: Rates / Promotions," and "Emotion: Clear Intent."

[0057] Next, the network dynamics module defines nodes: each customer, each financial product, and each marketing channel is an independent node. An explicit relationship edge is constructed based on the statement "Customer A applied for Product B". Simultaneously, LLM analysis reveals that "Customer C and Customer D viewed the detailed product page of the same car product within three days," thus constructing a "potential interest association edge," with weights calculated based on behavioral similarity.

[0058] In the intelligent reasoning and application layer, the system selects customer E, who recently searched for the keyword "home renovation," as the target. The reasoning module searches customer E's network and finds three strongly connected contacts (related in "potential interest association edge" and "social influence edge," and according to the risk transmission edge, there is no risk of delinquency) who have all applied for "Category A loan" products in the past six months. After receiving this network context, LLM generates a clue: "Target customer E has an intention to spend on home renovation, and their social circle has a high acceptance of 'Category A loan' products, creating a demonstration effect. It is recommended to combine the 'Home Furnishing Festival' event and push an upgraded version of the 'Category A loan' product introduction through the customer's exclusive APP, emphasizing 'fast approval and favorable rates.'" The outreach strategy module further generates a personalized push message.

[0059] After marketers reach out, regardless of whether customer E applies, the result will be recorded in the feedback and learning module. If the conversion is successful, the system will strengthen the correlation weight between "browsing renovation information" and "Type A loan" products, and between "social network success stories" and "conversion," to optimize future lead generation.

[0060] Example 2

[0061] like Figure 2 As shown, this application provides a method for constructing a consumer finance sales lead network based on a large language model, applied to the aforementioned consumer finance sales lead network construction system based on a large language model, specifically including:

[0062] S1 uses user profiles as a basis to divide users into different user profile groups. Based on the user data and semantic feature tags of the user profile groups and the correlation with marketing activities, it determines the real-time construction of profile groups within the user profile groups.

[0063] S2 determines the type of profile recommendation deviation for the marketing campaign based on the association between the marketing campaign and the semantic feature tags of different real-time profile groups;

[0064] S3 When the user profile group does not belong to the real-time profile building group, determine the semantic network construction strategy of the user profile group based on the association between the user profile group and the user tags of marketing activities with different profile recommendation deviation types. Based on the semantic network construction strategy and the recommendation matching of different marketing activities, determine the user profile group that needs to be optimized by the semantic network construction strategy.

[0065] Furthermore, the user profile is determined based on the parsing results of the user's associated data, including various types of semantic feature tags.

[0066] Furthermore, users are segmented into different user profile groups, specifically including:

[0067] Users with the same semantic feature labels are grouped into the same user profile group.

[0068] Furthermore, the user data of the user profile group includes the number of users in the user profile group.

[0069] Furthermore, the association between the semantic feature tags and the marketing campaign is determined based on the association between the semantic feature tags and the user tags of the marketing campaign.

[0070] Specifically, such as Figure 3 As shown, the method for determining the real-time user profile group is as follows:

[0071] This embodiment describes a decision-making method for dynamically determining whether to include user profile groups in real-time profile building based on their characteristics and size. The core decision-making objective of this method is to prioritize real-time monitoring and lead generation for customer groups with the greatest marketing potential and strategic value, given limited real-time computing resources, thereby achieving optimal allocation of marketing resources and maximizing response efficiency. This method employs a multi-step screening logic, evaluating groups sequentially from four dimensions: group size, breadth of marketing associations, scarcity of marketing associations, and comprehensive strategic weight. This ensures that selected groups are not only substantial in size but also closely and uniquely associated with marketing activities, or possess high value. Its core logic lies in using a hierarchical filtering mechanism to accurately identify high-priority groups from massive user profile groups—those whose members exhibit new behavior warranting immediate initiation of complex network construction and AI inference processes.

[0072] S11 uses the user data of the user profile group to determine the number of users in the user profile group;

[0073] The "user data" of a user profile group mainly refers to the collection of individual users included in the group, and the "number of users" is the basic indicator for measuring the size of the group.

[0074] This step is designed to follow the fundamental principle of scale in marketing campaign impact. Only when a user group has a certain size will real-time analysis and marketing interventions have economies of scale and commercial value. If the number of users in the group is too small, even if the individual users are highly valuable, the marginal benefit of investing real-time computing resources in dynamic network building and lead generation may be low. From a resource input-output perspective, batch or periodic processing models may be more suitable. This step, as the first layer of filtering, aims to quickly eliminate groups that are too small to justify high-cost real-time processing, allowing for more focused and targeted evaluations and improving overall system efficiency.

[0075] S12 determines the semantic feature tags associated with different marketing activities based on the association between the semantic feature tags of the user profile group and the marketing activities;

[0076] "Semantic feature tags" are intent, interest, or attribute identifiers extracted by LLM from user behavior data (such as "study abroad intention" or "high-end car owner"). "Association status" refers to whether a semantic tag is predefined as a target customer characteristic for a specific marketing campaign (such as "overseas education loan" or "vehicle loan").

[0077] This step aims to assess the breadth of association between user profiles and the existing marketing system. If a group's semantic tags are associated with multiple marketing campaigns, it indicates that the group has diverse potential product needs or cross-selling opportunities. Real-time monitoring of this group can trigger various marketing actions, thereby improving the efficiency of marketing resource utilization and potential returns. This avoids over-investing resources simply because the group is large but has a singular need, ensuring that the selected group possesses a certain degree of diversity and flexibility in terms of business opportunities. It connects static user characteristics with dynamic marketing assets and is a key step in determining the commercial value of a group.

[0078] S13 determines the real-time profile building group within the user profile group based on the number of users in the user profile group and the semantic feature tags associated with different marketing activities.

[0079] It should be noted that, based on the number of users in the user profile group and the semantic feature tags associated with different marketing campaigns, the real-time constructed profile group within the user profile group is determined, specifically including:

[0080] S131 Based on the number of users in the user profile group, determine the proportion of users in the user profile group among all users, and use it as the proportion of the number of users. Determine whether the proportion of the number of users is greater than a preset threshold for the proportion of the number of users. If yes, proceed to the next step. If no, determine that the user profile group does not belong to the real-time profile building group.

[0081] To determine if the percentage of users exceeds a preset threshold, for example: Suppose the system has 1 million users, and a group with the "recently purchased a home" tag has 100,000 users, accounting for 10%. If the preset threshold is 5%, then it passes.

[0082] This is a quantitative refinement of step S11. While absolute numbers may be influenced by the platform's overall user base, "percentage" better reflects the relative importance and representativeness of this group within the overall user base. By setting a percentage threshold, we ensure that the selected group is not only large in absolute numbers but also represents a strategically important market segment that the platform needs to focus on. This helps to allocate system resources towards core user segments.

[0083] S132 Based on the semantic feature tags of the user profile group, determine the semantic feature tags associated with different marketing activities, take the marketing activities with associated semantic feature tags as associated marketing activities, and determine whether the number of associated marketing activities is greater than the preset marketing activity number threshold. If yes, determine that the user profile group belongs to the real-time profile building group; otherwise, proceed to the next step.

[0084] In the above steps, it is determined whether the number of associated marketing activities exceeds a preset threshold. For example, continuing from the previous example, the "recent homebuyer" group may be associated with multiple marketing activities such as "home renovation installment payments," "home furnishing loans," and "property insurance." If the activity number threshold is 2, and the group is associated with 3 activities, then it passes.

[0085] This step directly quantifies the "breadth of associations" assessed in step S12. A large number of associated activities indicates that this group is a "high-value interface," and changes in its behavior may simultaneously impact multiple business lines. Real-time monitoring of this group allows for the capture of cross-business sales opportunities or risk signals, enabling coordinated responses. This promotes cross-departmental marketing collaboration and in-depth exploration of customer lifetime value.

[0086] S133 uses user profile groups with semantic feature tags that are related to different related marketing activities as related user profile groups, and determines whether there are related marketing activities in which the number of related user profile groups is less than a preset group number threshold. If yes, proceed to the next step; otherwise, determine that the user profile group does not belong to the real-time constructed profile group.

[0087] "Associated user profile groups" refers to all user groups that possess semantic tags associated with a particular marketing campaign. "Scarcity" is reflected in the fact that a marketing campaign is associated with only a very small number of user groups. For example, suppose there is a tag "yacht enthusiasts" that is only associated with the marketing campaign "high-end marine insurance," and there is only one user group (i.e., this group) containing this tag on the entire platform. If the preset threshold for the number of groups is 3, then this campaign meets the "scarcity" condition.

[0088] This step incorporates considerations of strategic scarcity and unique opportunities. Some marketing campaigns (typically targeting high-net-worth individuals, niche markets, or emerging businesses) have very specific and narrow target audiences. The user base serving these campaigns may be small, but they represent almost the entirety of the business's target customer base, making them extremely valuable. Any behavioral changes within this group must be captured immediately, or valuable and irreplaceable sales opportunities or risk warning windows may be missed. This setup ensures the system's real-time coverage of niche but critical markets.

[0089] S134 determines the weight of different related marketing activities based on the number of related user profile groups in different related marketing activities. When the sum of the weights of different related marketing activities is greater than the preset weight threshold, it is determined that the user profile group belongs to the real-time profile group.

[0090] In the above steps, the weighting rule is: the more user groups associated with a marketing campaign, the smaller the "weight" contribution of that campaign to that specific group. This is because the uniqueness or importance of that group in the campaign is reduced.

[0091] Example: Continuing from example S132, this group is associated with activities A, B, and C. Assume activity A is associated with 50 groups (low weight, i.e., falling within the range of 40 or more groups, with a weight of 0.1), activity B is associated with 5 groups (medium weight, i.e., falling within the range of 5 to 40 groups, with a weight of 0.5), and activity C is associated only with its own group and one other group (high weight, i.e., falling within the range of 0 to 5 groups, with a weight of 0.9). Then the total weight = 0.1 + 0.5 + 0.9 = 1.5. If the preset weight threshold is 1.2, then it passes.

[0092] This step involves a comprehensive, weighted assessment of group value. It balances "breadth of associations" (number of activities) and "depth of associations / uniqueness" (weight of each activity). A group may have a high overall strategic value even if it has a moderate number of associated activities, but if most of these activities are scarce (i.e., have high weight). Conversely, a group with many associated activities, but which are mostly mass-market activities (low weight), may have lower necessity. This weighting mechanism allows for more refined decision-making, enabling the identification of "strategic hub" user groups that are irreplaceable in specific areas or contribute to multiple highly unique businesses.

[0093] Ultimately, real-time processing is triggered: The system will initiate high-priority monitoring for its members only when a user profile group is identified as a "real-time profile building group." Once any user in this group generates new unstructured data (such as a customer service inquiry or an app browsing record), the system will immediately execute the following chain reaction:

[0094] LLM Real-Time Parsing: Invokes LLM to extract updated semantic feature labels from new data.

[0095] Dynamic network construction: Focusing on the user and the updated tags, quickly associate nodes such as products, channels, and activities, and recalculate or construct their "potential interest association edges" and "social influence edges".

[0096] Network reasoning and clue generation: The user is placed in a refreshed relational network context, and LLM performs multi-step reasoning to generate natural language clues such as "User X just consulted about electric vehicle charging issues. In his 'environmental advocate' group, 30% of the members have purchased new energy vehicle insurance in the past six months. It is recommended to push 'green car insurance' packages and charging pile cooperation discount information."

[0097] Marketing Recommendations: Trigger or recommend relevant marketing campaigns based on generated leads.

[0098] It is understood that the weight of the related marketing campaign is determined based on the number of related user profile groups of the related marketing campaign, and the larger the number of related user profile groups of the related marketing campaign, the smaller the weight.

[0099] It should be noted that if the user profile group belongs to the real-time profile building group, then once new unstructured data exists in the user profile group, the semantic feature tags of the user are extracted, and the entities of users, products, channels, and marketing activities are defined as network nodes. Implicit relationship edges such as "potential interest association edges", "social influence edges", and "risk transmission edges" are constructed to form a dynamic, multi-dimensional relationship network. The user is placed in the above-constructed relationship network for contextual analysis. Using LLM as the inference engine, the customer and its network association information are input, and through multi-step inference, natural language-described, actionable sales leads are generated for marketing activity recommendation processing.

[0100] Specifically, the method for determining the type of profile recommendation bias in the marketing campaign is as follows:

[0101] This embodiment describes an intelligent diagnostic method for evaluating and diagnosing whether and to what extent marketing campaigns target specific user profiles when making recommendations. The core decision-making objective of this method is to dynamically monitor and quantify the matching degree and coverage between each marketing campaign and the high-value customer group currently being monitored by the system (i.e., the "real-time profile-building group"). This allows the method to determine whether the target customer definition of the marketing campaign is accurate, up-to-date, and effective in pushing the campaign. The core logic is to consider the "real-time profile-building group" as the most behaviorally active, data-fresh, and strategically valuable "golden customer circle." By analyzing the closeness of the association between marketing campaigns and these core circles, the efficiency of user push processing for marketing campaigns is determined, thus reflecting the effectiveness and reliability of the LLM model's multi-network-based marketing campaign push processing. This is a data-driven, reverse-verification quality control mechanism for the effectiveness of marketing targeting.

[0102] S21 determines the real-time constructed profile group with semantic feature tags that are associated with the marketing activity based on the association between the marketing activity and the semantic feature tags of different real-time constructed profile groups, that is, the real-time associated profile group of the marketing activity.

[0103] "Real-time related profile group" refers to the user group that belongs to the "real-time profile building group" (i.e., the high-priority group selected through multiple steps in Example 1) and whose semantic feature tags match the target customer characteristics (i.e. marketing tags) preset by the marketing campaign to be evaluated.

[0104] This step aims to bridge the gap between marketing strategies and the most active customer data in the real world. "Real-time profiling" involves the system dynamically identifying value clusters based on users' latest behavior, representing the true pulse of the market. By identifying which such groups are relevant to a marketing campaign, this step places the static marketing plan (with pre-defined labels) against a dynamically changing customer landscape, providing the data foundation for subsequent deviation diagnosis. It ensures that the evaluation is based on the latest and most relevant market slices, rather than outdated or generalized customer segmentation.

[0105] S22 determines the profile recommendation deviation type of the marketing campaign based on the real-time associated profile group data of the marketing campaign.

[0106] It is understandable that if the marketing campaign does not have a real-time associated user profile, then the user profile recommendation deviation type of the marketing campaign is determined to be a severe deviation type.

[0107] Scenario 1: Severe Deviation Type (When the marketing campaign does not have any real-time relevant user profiles):

[0108] Example: A campaign targeting "Gen Z tech enthusiasts" offered "latest headphones with interest-free installments," using pre-defined tags such as "age 18-25," "enthusiastic about new tech products," and "frequent music app users." However, the system found that none of the groups currently tagged as "real-time profile building groups" simultaneously met all these tag combinations. It's possible that Gen Z is currently more focused on "job training" or "short trips," with very little real-time relevance to trendy tech products.

[0109] This determination means that the target customer positioning of the marketing campaign is completely out of sync with the high-value, high-activity customer segments currently identified by the system. This is a dangerous signal, indicating that the recommendation processing efficiency of the marketing campaign is low within the existing user recommendation processing scheme. In related marketing campaigns, the semantic analysis results of the LLM model may be difficult to effectively verify, leading to slower model update and iteration efficiency. Therefore, the profile recommendation bias type of the marketing campaign is determined to be a severe bias type.

[0110] Additionally, it can be understood that if the marketing campaign has a real-time associated profile group, the number of the real-time associated profile group of the marketing campaign is obtained, and it is determined whether the number of the real-time associated profile group of the marketing campaign is greater than a preset real-time group number threshold. If so, the profile recommendation deviation type of the marketing campaign is determined to be the reliable recommendation type; otherwise, the profile recommendation deviation type of the marketing campaign is determined to be the general deviation type.

[0111] Recommended reliable type (when the number of real-time associated profile groups exceeds a preset threshold):

[0112] Example: A "car owner loan" campaign has preset tags such as "car owners" and "frequent drivers". The system found that in the current real-time profile building, more than 5 groups were associated with highly relevant tags such as "weekend road trip enthusiasts" and "commuting distance of more than 20 kilometers".

[0113] This assessment indicates that the marketing campaign's target positioning resonates strongly with multiple currently active core customer segments. The large number of associated groups signifies that the campaign covers a broad range of high-value customers in real time, and its pre-defined tags effectively identify a large number of target individuals from the dynamic market, thereby improving the campaign's delivery efficiency. This demonstrates the accuracy and timeliness of the marketing strategy; the label "reliable recommendation" is positive feedback, indicating high recommendation efficiency. Furthermore, the semantic analysis results of the LLM model can be effectively validated in related marketing activities, leading to rapid model updates and iterations.

[0114] General deviation type (when there is a real-time associated profile group, but the number does not exceed the threshold):

[0115] Example: A "high-end study abroad loan" campaign has preset tags such as "planning to study abroad" and "family annual income of over 500,000 yuan". The system found that only one real-time profile-building group (e.g., "international school parent community") is associated with it, and the size of this group is moderate and does not meet the reliable standard of "large number".

[0116] This assessment reveals a narrow or insufficiently targeted approach. While the marketing campaign did reach some high-value customers in real time, its coverage was limited, resulting in generally low recommendation efficiency.

[0117] Specifically, such as Figure 4 As shown, the method for determining the semantic network construction strategy for the user profile group is as follows:

[0118] This embodiment aims to directly address the core contradiction of inefficient marketing campaign execution, which leads to the inability to collect timely and effective user feedback, making it difficult to assess the reliability of LLM based on multi-network correlation analysis. Its decision-making objective is to proactively allocate computing resources (real-time semantic network construction and LLM inference) to user groups most likely to experience "data black boxes" and "business value loss" due to push delays through a forward-looking, risk-assessment-based resource scheduling system. The logic is as follows: 1) Identify which groups' marketing outreach has failed or become inefficient due to efficiency issues (bias type); 2) Assess the magnitude of business losses caused by the failure of these groups (correlation impact factors); 3) Based on the above two points, decide whether to initiate high-cost real-time analysis to generate more accurate leads, attempting to compensate for the lack of "reach quantity" with "quality" and accumulating key data for assessing LLM reliability. Essentially, this is a strategy to maximize learning efficiency and business defense capabilities by optimizing the allocation of insight resources under limited push capabilities.

[0119] S31 determines the type of recommendation bias in the marketing activities associated with the user profile group based on the association between the user tags of the user profile group and marketing activities with different types of profile recommendation bias.

[0120] "Profile Recommendation Deviation Type" is the diagnostic result of calling the previous implementation example for each related marketing campaign. It is determined based on the "number of users pushed to this user profile group in the recent period" of the campaign, and is divided into "Severe Deviation" (extremely low push volume, feedback vacuum), "General Deviation" (insufficient push volume, sparse and biased feedback), and "Reliable Recommendation" (sufficient push volume, stable feedback).

[0121] This step is the problem diagnosis and risk assessment layer. It directly quantifies the specific consequences of "slow push efficiency" on this group. "Severe bias" means that this group is almost excluded from the current marketing cycle, and the LLM analysis of them is in a completely unverifiable "black box" state, making reliability assessment impossible, which is the highest level of risk. "Moderate bias" means that there is a small amount of insufficient feedback, the assessment confidence is low, and there is a risk of misjudgment. The significance of this step is that it accurately locates the macro-level push efficiency problem to each group and transforms it into a comparable risk level, providing the most urgent and direct decision-making basis for "whether it is worthwhile to invest resources to break down data silos".

[0122] S32 determines the association influence factor between the user profile group and different related marketing activities based on the ranking results of the user profile group in the related profile groups of different related marketing activities;

[0123] "Ranking Results" refers to the ranking of all target user profile groups in a marketing campaign associated with this group, from highest to lowest based on the number of users included. "Association Influence Factor" is a value from 0 to 1. The higher the ranking (the larger the user base), the larger the factor value, representing the strategic scale and core importance of this group in this marketing campaign.

[0124] For a given user profile group G and its associated marketing campaign C, the association influence factor IG,C of this group in this campaign is calculated as follows:

[0125]

[0126] in:

[0127] I G,C : The correlation influence factor of user profile group G in marketing campaign C, with a value range of [0,1].

[0128] Rank G,C The ranking result of group G among all related profile groups of activity C (arranged from highest to lowest number of users), starting from 1 (the group with the most users is ranked 1st).

[0129] N C Total number of related user profiles for Marketing Campaign C.

[0130] This step is the value and loss assessment layer. In the context of low push efficiency, assessing the importance of a group involves not only looking at its potential conversion rate, but also the overall opportunity loss that could result from the continued failure or stagnation of marketing strategies targeting that group due to a lack of effective reach and feedback. For groups with high ranking and large influence factors, their failure will directly impact the core of the campaign. The significance of this step is that it quantifies the expected business loss caused by "marketing failure to that group due to slow push efficiency." This ensures that when allocating scarce real-time computing resources, the system will prioritize those core groups with the "most severe consequences after failure," providing clear strategic defense and loss mitigation value.

[0131] S33 determines the semantic network construction strategy for the user profile group based on the profile recommendation deviation type of the related marketing activities and the association influence factors of different related marketing activities.

[0132] The "semantic network construction strategy" is the final output action instruction, which is divided into the "pre-set construction strategy" (highest priority, aimed at urgently breaking the data black box), the "second pre-set construction strategy" (medium priority, aimed at optimizing decision-making under sparse feedback), and the "third pre-set construction strategy" (lowest priority, aimed at maintaining a virtuous cycle). The decision-making process is a refined arbitration process containing sub-steps S331-S334.

[0133] This step is the resource optimization and action generation layer. It transforms the two-dimensional "risk-value" assessment into differentiated and executable resource scheduling solutions through a set of hierarchical and progressive decision-making rules. Its core significance lies in achieving optimal investment decisions under constraints, ensuring that valuable real-time computing resources are used to address the most pressing and highest-return "data-business" dual crises.

[0134] It should be noted that the ranking results of the user profile groups are obtained by sorting the number of users in the related marketing campaign profile groups from high to low.

[0135] Specifically, the correlation influence factor is determined in conjunction with the ranking result. The higher the ranking result of the user profile group in the correlation marketing activity, the larger the correlation influence factor, and its value ranges from 0 to 1.

[0136] It is understandable that if there are severely biased related marketing activities in the related marketing activities of the user profile group, then the semantic network construction strategy of the user profile group is determined to be the preset construction strategy.

[0137] Furthermore, if there are no severely biased related marketing campaigns in the related marketing activities of the user profile group, the following are also included:

[0138] S331 determines whether there is a general deviation type of related marketing activity in the related marketing activities. If yes, proceed to the next step. If no, determine that the semantic network construction strategy of the user profile group is the third preset construction strategy.

[0139] Determine if there are any cross-marketing campaigns exhibiting a "general deviation" type. If not, determine the strategy as the third preset construction strategy.

[0140] This step is the first level of triage. If all associated activities of a group are of the "reliable recommendation" type, it means that the group is currently in a good state of marketing service and data feedback, and there is no urgent risk of a "data black box". Therefore, it is given the lowest priority (third strategy) to avoid unnecessary intervention in its health status and save resources.

[0141] S332 obtains the number of associated marketing activities of the general deviation type, and determines whether the number of associated marketing activities of the general deviation type is greater than the preset deviation activity number threshold. If yes, the semantic network construction strategy of the user profile group is determined to be the preset construction strategy. If no, proceed to the next step.

[0142] This step assesses the breadth of the problem. If multiple marketing campaigns associated with a group are under-targeting them (a high number of "general deviation" campaigns), it indicates that the group faces systemic, cross-campaign reach barriers, rather than a problem with a single campaign. This widespread "push anomaly" severely hinders a comprehensive understanding of their interest graph and value. Therefore, it should be prioritized at the highest level (pre-defined strategy), aiming to address or mitigate the common problems across multiple campaigns in one go through deep network insights for maximum efficiency.

[0143] S333 determines whether there are any related marketing activities with a correlation impact factor greater than a preset impact factor threshold based on the correlation impact factor of related marketing activities with different general deviation types. If yes, proceed to the next step; otherwise, determine the semantic network construction strategy of the user profile group as the second preset construction strategy.

[0144] This step assesses the depth and criticality of the problem. When the number of problematic activities is small, it's necessary to examine whether these problems occur in the most critical business areas. If all "general deviation" activities are not core business for this group (low impact factor), then even optimization will yield limited overall benefits. In this case, assigning them medium priority (secondary strategy) is appropriate. This prevents resources from being wasted on optimizing "peripheral issues" with minimal impact on the business.

[0145] S334 determines a comprehensive influence factor by summing the influence factors of related marketing activities with different general deviation types, and determines a semantic network construction strategy for the user profile group based on the comprehensive influence factor.

[0146] This step is the final comprehensive value assessment. When a group is associated with a few but very important "general deviation" activities, it is necessary to assess whether its overall strategic value is high enough to warrant initiating a highest-priority response. If the sum of the comprehensive impact factors exceeds a preset value, it means that the combined impact of these high-value activities is significant enough to require urgent attention to prevent damage to core business. Conversely, it remains classified as a medium priority. This achieves a convergent judgment from "individual activity importance" to "overall strategic value of the group," making the decision more robust.

[0147] Assume the user profile group "G-Cross-border E-commerce Sellers" (80,000 users) is associated with three marketing campaigns:

[0148] Activity P (Cross-border E-commerce Business Loan): Group G ranks 1st in terms of user numbers among its target groups (correlation influence factor = 0.95). In the past 7 days, 400 users (0.5% of the group) were reached with the activity, which was diagnosed as "severe bias".

[0149] Activity Q (Cross-border payment fee discount): Ranked 3rd in Group G (impact factor = 0.75). Number of push users = 2000 (accounting for 2.5%), diagnosed as "general deviation".

[0150] Activity R (Corporate Credit Card): Ranked 5th in the G group (Influence Factor = 0.60). Number of users pushed to the campaign = 16,000 (20%), diagnosed as "Reliable Recommendation". Execution process:

[0151] Step S31: Determine the type of associated activity bias in group G: activity P (severe bias), activity Q (moderate bias), activity R (reliable recommendation).

[0152] Step S32: Determine the associated influencing factors: Activity P (0.95), Activity Q (0.75), Activity R (0.60).

[0153] Step S33 Decision:

[0154] Because of the "severe deviation" activity P, according to the rules, the semantic network construction strategy of group G is directly determined to be the "preset construction strategy" (highest priority). There is no need to proceed to the judgment in S331-S334.

[0155] Trigger and Action (based on preset policy rules): Condition check: The number of push users (400) of activity P (serious deviation) in the most recent period is much less than the preset push user number threshold (assumed to be 8000).

[0156] Decision: Therefore, in the next preset time period (the following hour), the system will initiate real-time semantic network construction processing for the G group.

[0157] Real-time processing and value realization:

[0158] During the monitoring period, seller "User A" in group G clicked "Inventory Preparation Funding Loan" in the financial software. The system immediately captured the unstructured text and used LLM parsing to extract semantic feature tags: "Strong demand for financing for peak season inventory preparation" and "Focus on capital turnover efficiency".

[0159] Dynamically update the network centered on user A: strongly connect user A to the "Activity P (Business Loan)" node through "potential interest association edges"; discover that there are 3 related sellers in user A's social network who have just successfully applied for similar loans, and construct "social influence edges".

[0160] LLM uses this enhanced network context to infer and generate natural language cues: "User A and their 'peak season inventory preparation' sub-circle have a clear and urgent need for business loans. Traditional tags failed to capture this real-time intent. Recommendations: 1. Immediately push Activity P to User A and their sub-circle with high priority, and match it with the phrase 'peak season exclusive, fast approval'; 2. Optimize the tagging system of Activity P, and add scenario-based tags such as 'peak season financing' and 'inventory turnover'."

[0161] Results: This highly accurate and contextualized lead was adopted by the push system, achieving precise reach to user A and some users within their network. If this push achieves a high click-through / inquiry rate, it will provide high-quality early data points for the first time to evaluate the reliability of LLM analysis in "severely biased" groups, breaking the data black box.

[0162] It should be noted that if the comprehensive impact factor is greater than the preset value of the impact factor, the semantic network construction strategy of the user profile group is determined to be the preset construction strategy; if the comprehensive impact factor is not greater than the preset value of the impact factor, the semantic network construction strategy of the user profile group is determined to be the second preset construction strategy.

[0163] It should be noted that if the semantic network construction strategy for the user profile group is a preset construction strategy, then as long as the number of users pushed to the related marketing activities of the user profile group within the most recent preset time period is less than the preset threshold of the number of users pushed, the semantic network construction processing of the user profile group will be carried out in the future preset time period. Once new unstructured data exists in the user of the user profile group, the semantic feature tags of the user will be extracted, and the entities of users, products, channels, and marketing activities will be defined as network nodes. Implicit relationship edges such as "potential interest association edge", "social influence edge", and "risk transmission edge" will be constructed to form a dynamic, multi-dimensional relationship network. The user will be placed in the above-constructed relationship network for contextual analysis. Using LLM as the inference engine, the customer and its network association information will be input, and through multi-step inference, natural language-described, actionable sales leads will be generated for marketing activity recommendation processing.

[0164] Additionally, it can be understood that if the semantic network construction strategy for the user profile group is the second preset construction strategy, then as long as the sum of the association influence factors of the related marketing activities of the user profile group in the most recent preset time period, where the number of users pushed to the activity is less than the preset threshold for the number of users pushed to the activity, is greater than the preset factor threshold, then the semantic network construction processing of the user profile group will be carried out in the future preset time period. Once new unstructured data exists in the user profile group, the semantic feature tags of the user will be extracted, and the entities of users, products, channels, and marketing activities will be defined as network nodes. Implicit relationship edges such as "potential interest association edges", "social influence edges", and "risk transmission edges" will be constructed to form a dynamic, multi-dimensional relationship network. The user will be placed in the above-constructed relationship network for contextual analysis. Using LLM as the inference engine, the customer and their network association information will be input, and through multi-step inference, natural language-described, actionable sales leads will be generated for marketing activity recommendation processing.

[0165] Additionally, it can be understood that if the semantic network construction strategy for the user profile group is the third preset construction strategy, then as long as the sum of the association influence factors of the related marketing activities of the user profile group in the most recent preset time period, where the number of users pushed to the activity is less than the preset threshold for the number of users pushed to the activity, is greater than the second preset factor threshold, then the semantic network construction processing of the user profile group will be carried out in the future preset time period. Once new unstructured data exists in the user profile group, the semantic feature tags of the user will be extracted, and the entities of users, products, channels, and marketing activities will be defined as network nodes. Implicit relationship edges such as "potential interest association edges", "social influence edges", and "risk transmission edges" will be constructed to form a dynamic, multi-dimensional relationship network. The user will be placed in the above-constructed relationship network for contextual analysis. Using LLM as the inference engine, the customer and their network association information will be input, and through multi-step inference, natural language-described, actionable sales leads will be generated for marketing activity recommendation processing.

[0166] It should be noted that the preset factor threshold is less than the second preset factor threshold.

[0167] Specifically, the method for determining the user profile groups that need to be optimized in the semantic network construction strategy is as follows:

[0168] The fundamental goal of this embodiment is to address the problem of significant delays in manually reviewing large-scale models of inefficient marketing campaigns due to uneven "data update efficiency" among different user groups. Its core logical chain is as follows:

[0169] Problem identified: Marketing campaigns with low conversion rates ("conversion bias") may have misunderstandings in their overall models and require urgent manual review.

[0170] Diagnostic Bottleneck: The triggering of manual verification depends on the accumulation of sufficient new push user data (reaching the "first preset quantity range"). However, if the majority of the target user group of the activity is currently assigned a semantic network construction strategy with low update efficiency (such as the "third preset strategy"), then the new behavioral data of these users is difficult to be captured, parsed and used for push notifications in a timely manner, resulting in slow growth in the "number of new push users" and indefinite delays in manual verification.

[0171] Optimization Implementation: Therefore, the decision-making objective of this embodiment is to accurately identify those groups among the user groups associated with "conversion deviation marketing campaigns" that are being dragged down by inefficient strategies and becoming "bottlenecks" in data feedback. By optimizing and upgrading the construction strategies for these groups (e.g., upgrading to "real-time group profiling"), the efficiency of data and status updates in associated campaigns can be significantly improved, thereby accelerating the accumulation of new user data pushed to the problematic campaigns and prompting them to reach the conditions for triggering manual verification more quickly, enabling rapid diagnosis and repair.

[0172] S41 Based on the aforementioned semantic network construction strategy, determine the semantic network construction strategy for the associated profile groups of different marketing activities;

[0173] "Semantic network construction strategy" refers to the dynamic analysis priority assigned to each user profile group in the early stage of the system (such as in Example 3), which is divided into "preset construction strategy" (highest priority, updated in real time), "second preset construction strategy" (medium priority, updated under conditions), and "third preset construction strategy" (low priority, updated periodically or in batches). "Associated profile group" refers to the set of target customer groups preset for a certain marketing campaign.

[0174] This step is set up to build a "strategy map" of the current system resource allocation, clarifying the current data insight status of each user group on which each marketing campaign depends. This is the foundation for assessing "data update efficiency bottlenecks." Only by clearly understanding the correspondence between campaign, group, and strategy can we analyze which campaigns may face the risk of slow data feedback due to the strategy distribution of their associated groups. For example, if most of a campaign's target group is under the "third preset construction strategy," then the overall speed at which the campaign acquires fresh user behavior data will be very slow, which is precisely the scenario that subsequent optimization needs to focus on.

[0175] S42 determines the conversion data of users targeted by different marketing campaigns based on the recommendation matching results.

[0176] "Recommendation matching" refers to whether the actual users pushed to the marketing campaign are those recommended and matched by the system's AI model. "Conversion data of pushed users" specifically refers to "conversion rate," calculated as: (Number of users showing interest in the push / Total number of pushed users) * 100%. Here, "interest behavior" can be key conversion nodes such as clicks, inquiries, and applications, depending on the business definition.

[0177] This step aims to obtain the ultimate business metric for measuring the effectiveness of marketing campaigns and to identify "problem campaigns" that need to be prioritized. Low conversion rates are the most direct signal that the model may have misunderstandings or that the strategy is failing. By continuously monitoring this metric, the system can quickly filter out those campaigns that most urgently require in-depth diagnosis and optimization from among many others, providing clear targets for subsequent bottleneck analysis and resource reallocation. This ensures that the entire optimization process is always oriented towards solving actual business problems.

[0178] S43, based on the semantic network construction strategy of the associated profile groups of the marketing campaign and the conversion data of the pushed users, determines the user profile groups that need to be optimized in the semantic network construction strategy.

[0179] The "user profile group requiring optimization" is the core output object of this embodiment. It specifically refers to those groups currently associated with marketing campaigns with low conversion rates ('conversion deviation marketing campaigns') and whose semantic network construction strategies are inefficient, thus becoming a bottleneck in the speed of data feedback and diagnosis for the entire problematic campaign. The specific action for "optimizing" them is usually to upgrade them to "real-time profile building group", that is, to adopt the highest priority "preset construction strategy".

[0180] This step is the core decision-making layer for diagnosing bottlenecks and generating optimization instructions. It systematically answers the question, "To accelerate the manual verification of a problematic activity, which user group(s) should we prioritize upgrading their strategies?" through a set of screening logics from macro to micro, from phenomena to root causes (S431-S435). Its significance lies in achieving precision and purposefulness in resource allocation: instead of indiscriminately upgrading all inefficient activity-related groups, it identifies the "key lever points" that most effectively accelerate the overall data flow by analyzing the correlation between strategy distribution and conversion effects, thereby achieving maximum diagnostic acceleration with minimal resource intervention.

[0181] Specifically, the conversion data includes the conversion rate of users who were pushed to the marketing campaign, which is determined based on the proportion of users who were interested in the marketing campaign.

[0182] Furthermore, based on the semantic network construction strategy for the associated user profiles of the marketing campaign and the conversion data of the pushed users, the user profile groups that need to be optimized in the semantic network construction strategy are identified, specifically including:

[0183] S431 uses the conversion data of users pushed by different marketing activities to determine whether there are marketing activities where the conversion rate of users pushed is less than the preset conversion rate threshold. If so, proceed to the next step. If not, there is no need to manually review the identification deviation of the marketing activity model. Therefore, it is determined that all user profile groups do not belong to the user profile groups that need to be optimized by the semantic network construction strategy.

[0184] The "preset conversion rate threshold" is a pass / fail line set based on historical data and business objectives. It's used to determine the health of a marketing campaign's performance. This step acts as the master switch and filter for the entire optimization process. Its existence ensures that optimization is "problem-driven" rather than "blindly implemented." Only when the system detects that a marketing campaign's performance is below expectations (conversion rate below the threshold) does it indicate a potential risk of model bias requiring manual verification. Only then is subsequent bottleneck analysis and strategy upgrades necessary. If all campaigns meet the targets, the system is healthy and no additional optimization process is needed. This reflects the system's economic efficiency and stability, avoiding resource consumption and excessive intervention in problem-free scenarios.

[0185] S432 identifies marketing activities with a conversion rate of less than a preset conversion rate threshold as conversion deviation marketing activities. It then determines whether there are conversion deviation marketing activities in the associated marketing activities of the user profile group. If yes, it proceeds to the next step. If no, since there are no conversion deviation marketing activities in the associated marketing activities of the user profile group, there is no need for manual review of the model identification deviation of the marketing activities. Therefore, the user profile group does not belong to the user profile group that needs to be optimized by the semantic network construction strategy.

[0186] "Conversion Deviation Marketing Campaigns" refer to the marketing campaigns identified in the previous step that failed to meet performance targets. "Related Marketing Campaigns" refer to those marketing campaigns that were pre-targeted within a specific user profile group; this step involves attribution and initial screening. It establishes the connection between the user profile group and the problematic campaigns. Whether a group needs optimization depends first on whether it is directly related to the "problem." If all marketing campaigns associated with a group perform well, then regardless of the efficiency of its own strategy, it is not part of the current "diagnosis acceleration" bottleneck that needs to be addressed, and therefore should be excluded from the scope of this optimization. This step quickly narrows the optimization scope from the "all user groups" to "groups directly related to the problematic campaigns," improving the efficiency and targeting of subsequent steps.

[0187] S433 determines whether the proportion of conversion deviation marketing activities in the associated marketing activities of the user profile group is greater than the preset threshold for the proportion of deviation marketing activities. If yes, it is determined that the user profile group belongs to the user profile group that needs to be optimized by the semantic network construction strategy. If no, proceed to the next step.

[0188] "Quantity percentage" refers to the number of "conversion deviation marketing campaigns" associated with this user profile group divided by the total number of marketing campaigns associated with it. "Preset deviation marketing campaign quantity percentage threshold" is a percentage threshold (e.g., 50%).

[0189] This step aims to assess the severity and prevalence of the problem for this user profile group. If a majority (exceeding a threshold) of the activities associated with a group are ineffective, it indicates that the group is likely in a systemic, multifaceted environment of service failure or cognitive bias. Prioritizing optimization for such groups is of high value and urgency because improving their data update efficiency can positively impact the diagnosis of multiple problematic activities simultaneously, reflecting the strategic priority principle in optimizing resource allocation.

[0190] S434 determines whether there is a real-time profile group in the conversion deviation marketing activities of the user profile group. If yes, proceed to the next step; otherwise, determine that the user profile group belongs to the user profile group that needs to be optimized by the semantic network construction strategy.

[0191] "Real-time profile building groups" refers to those groups that have already implemented the highest priority "pre-built strategy", and their data is being monitored and updated in real time or near real time.

[0192] This step is crucial for identifying "clear bottlenecks" and is the core of the logic in this embodiment. If a group is associated with a problem activity but is not yet a "real-time profiled group," it means it is in a state of low data update efficiency. It is a direct and clear bottleneck causing slow growth in new user data for its associated problem activities, thus delaying manual verification. Therefore, once such a group is identified, it should be immediately determined as an object requiring optimization without further complex judgment. Upgrading its strategy is the most direct and effective measure to streamline the data feedback chain and accelerate diagnosis.

[0193] S435 classifies marketing campaigns that do not have conversion deviations in real-time user profile building as other marketing campaigns, and determines whether the user profile group belongs to the user profile group whose semantic network building strategy needs to be optimized based on the semantic network building strategy of the related marketing campaigns of the user profile group.

[0194] "Other marketing campaigns" here specifically refers to those conversion-biased marketing campaigns in S434 that do not have "real-time profile building for target groups". "Push-biased marketing campaigns" are a type of campaign identified by analyzing strategy distribution patterns. They are defined as follows: Under this campaign, the total number of users covered by the associated groups in the low-to-medium efficiency strategies (second and third presets) is very low, and the number of users in the "second preset strategy" is also very small. This indicates that the strategy configuration of this campaign may be ineffective overall, resulting in generally slow data updates.

[0195] This step addresses more complex or hidden bottleneck scenarios. When a problem campaign has enabled live builds for some of its core groups (S434 condition is true), but the overall campaign performance remains poor and data growth is slow, the bottleneck may not lie with the core groups, but with other groups that constitute the majority of the campaign's user base and are under inefficient strategies. S435 identifies this "strategy pattern failure" by determining whether the campaign is a "push-biased marketing campaign." If identified successfully, it means the optimization scope needs to be expanded to include other related groups under inefficient strategies within the campaign, systematically improving the overall data output efficiency of the campaign and thus accelerating diagnosis.

[0196] Specifically, based on the semantic network construction strategy of the associated user profile group in other marketing activities, push-biased marketing activities in the other marketing activities are identified. When there are multiple push-biased marketing activities in the associated marketing activities of the user profile group, the user profile group is determined to be a user profile group that needs to be optimized by the semantic network construction strategy.

[0197] It is understood that the push deviation marketing campaign refers to other marketing campaigns where the sum of the proportion of users in all related profile groups is less than a preset user proportion threshold, and the number of related profile groups in the second preset construction strategy is less than a preset value of the number of groups.

[0198] In one possible embodiment:

[0199] Suppose the system identifies the marketing campaign "High-end Study Abroad Loan Funding Plan" as having a conversion rate consistently below a preset threshold of 1.5%, and classifies it as a conversion deviation marketing campaign (A1). This campaign is associated with three target user profile groups:

[0200] Group G1: Families of international school high school students (Current strategy: Pre-defined construction strategy, which builds profiles of the group in real time with high update efficiency)

[0201] Group G2: Outstanding undergraduate students from key universities (Current strategy: Second pre-set construction strategy, efficiency being updated)

[0202] Group G3: Young white-collar workers with overseas experience (Current strategy: Third preset construction strategy, low update efficiency)

[0203] Manual verification trigger conditions: Activity A1 needs to accumulate 5,000 new push users (first preset number range) and the conversion rate is less than 1.5% before manual verification can be triggered.

[0204] Bottleneck analysis: Currently, only G1 can contribute data efficiently. Due to the low efficiency of their strategies, G2 and G3 users' new study abroad intentions are difficult to capture and push in a timely manner, resulting in slow growth of the "number of new push users" for Activity A1, and manual verification is a long way off.

[0205] Execute the S43 optimization process: S431: The conversion rate of activity A1 is below the threshold, continue the process. S432: Groups G1, G2, and G3 are all associated with activity A1, continue the process.

[0206] S433: Assuming that activity A1 is not the only or primary activity associated with these three groups (the percentage of deviation activities is less than the threshold, such as 20%), all three groups proceed to the next step.

[0207] S434 Decision: Check activity A1's "Real-time Construction of Profile Groups": Includes G1.

[0208] Therefore, for groups G2 and G3: they are associated with A1, but they are not real-time profile building groups of A1 → According to rule S434, since there are real-time profile building groups, it is determined that G2 and G3 do not belong to the "user profile groups that need optimization processing".

[0209] Assuming G1 also does not belong to the real-time optimized profile group and belongs to the third preset construction strategy:

[0210] At this point, proceed to step S435, where the second preset construction strategy is 1, namely G2, which is less than 2. When the number of users of G1 / G2 / G3 accounts for less than 5% of the total number of users, it is determined that the activity A1 belongs to the push deviation marketing activity.

[0211] If G2 and G3 have two marketing campaigns with different push notifications, then G2 and G3 will be considered as "user profile groups that need optimization".

[0212] Optimized Results: Data update efficiency for G2 and G3 has been significantly improved. A large number of previously delayed potential study abroad intentions (such as school inquiries and participation in study abroad fairs) can now be analyzed in real-time by LLM and converted into targeted push notifications. This has accelerated the growth of "new push users" for Activity A1, significantly reducing the time required to reach 5,000 new users and trigger manual verification. Experts can intervene earlier, potentially discovering that the model incorrectly interprets the needs of "young white-collar workers" as "children's education" rather than "self-improvement," and can promptly correct the model's logic.

[0213] It should be noted that if the user profile group belongs to the user profile group that needs to be optimized by the semantic network construction strategy, then it will be regarded as the user profile group to be constructed in real time.

[0214] Furthermore, when the conversion rate of the marketing campaign is less than a preset conversion rate threshold, and whenever the number of new users pushed to the campaign is within a first preset number range, a manual review of the model's identification deviation is performed to determine whether there is a deviation in the model's understanding.

[0215] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0216] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0217] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for constructing a consumer finance sales lead network based on a large language model, characterized in that, Specifically, it includes: Based on user profiles, users are divided into different user profile groups; based on the number of users in the user profile group and the marketing activities associated with the semantic feature tags of the user profile group, the real-time constructed profile group in the user profile group is determined by the number of users, the number of associated marketing activities, and the number of all user profile groups corresponding to the semantic feature tags of the associated marketing activities. Based on the association between marketing campaigns and the semantic feature tags of different real-time profile groups, real-time profile groups with semantic feature tags that are associated with the marketing campaigns are identified as the real-time associated profile groups of the marketing campaigns. Based on whether the marketing campaigns have real-time associated profile groups and the number of such groups, the profile recommendation deviation type of the marketing campaigns is determined. When a user profile group does not belong to a real-time profile group, the association between the user profile group and the semantic feature tags of different marketing activities is used to determine the profile recommendation deviation type of the user profile group in the related marketing activities. Based on the ranking result of the user profile group obtained by sorting the user profile group by the number of users in the related profile groups of the related marketing activities, the association influence factor between the user profile group and different related marketing activities is determined. Based on the profile recommendation deviation type of the user profile group in the related marketing activities and the association influence factor of different related marketing activities, the semantic network construction strategy of the user profile group is determined. The semantic network construction strategy includes: the highest priority preset construction strategy, the medium priority second preset construction strategy, and the lowest priority third preset construction strategy. Based on the semantic network construction strategy of the user profile groups, the semantic network construction strategy of the associated profile groups of different marketing activities is determined. Based on the recommendation matching of different marketing activities, the conversion data of the users pushed by the marketing activities is determined. Based on the semantic network construction strategy of the associated profile groups of the marketing activities and the conversion data of the users pushed, the user profile groups that need to be optimized by the semantic network construction strategy are determined.

2. The method for constructing a consumer finance sales lead network based on a large language model as described in claim 1, characterized in that, The user profile is determined based on the parsing results of the user's associated data, including various types of semantic feature tags.

3. The method for constructing a consumer finance sales lead network based on a large language model as described in claim 1, characterized in that, Users are segmented into different user profile groups, specifically including: Users with the same semantic feature labels are grouped into the same user profile group.

4. The method for constructing a consumer finance sales lead network based on a large language model as described in claim 1, characterized in that, If the user profile group belongs to the real-time profile building group, then once new unstructured data is available for users in the user profile group, the semantic feature tags of the users will be extracted and marketing activity recommendation processing will be carried out.

5. The method for constructing a consumer finance sales lead network based on a large language model as described in claim 1, characterized in that, The ranking of the user profile groups is obtained by sorting the number of users from high to low based on the related profile groups of the related marketing activities.

6. A consumer finance sales lead network construction system based on a large language model, comprising the method for constructing a consumer finance sales lead network based on a large language model as described in any one of claims 1-5, characterized in that, Specifically, it includes: Data access module: used to integrate related data from users and read related data to obtain unstructured text data; Feature enhancement module: Based on the semantic network construction strategy of the user's user profile group, it uses the semantic network construction strategy and large language model to perform in-depth analysis of unstructured text data, perform intent recognition, sentiment analysis and topic extraction, and generate semantic feature labels; Network dynamic construction module: Define entities as network nodes, and use the reasoning ability of large models to construct explicit and implicit relationship edges, thereby forming a dynamic, multi-dimensional relationship network; Reasoning module: Places the user in the relational network for contextual analysis, and generates actionable sales leads in natural language through multi-step reasoning; Outreach strategy generation module: Based on the generated sales leads, automatically generate or recommend personalized communication scripts, and determine the user profile groups that need to be optimized based on the semantic network construction strategy and the recommendation matching of different marketing activities.

7. The consumer finance sales lead network construction system based on a large language model as described in claim 6, characterized in that, The associated data includes data from internal business systems, data from third-party credit reporting agencies, and user authorization behavior data.

Images