User maintenance method and device based on simulated user agent, medium and equipment

Abstract

The application discloses a user maintenance method and device based on simulated user agents, a medium and equipment, relates to the technical field of artificial intelligence, and can be applied to the financial technology scene, and mainly solves the problem of low timeliness in user maintenance. Mainly includes constructing an agent social network; in response to the update of external environment information, real-time external environment information is obtained, a multi-dimensional transaction environment state is simulated according to the real-time external environment information, and the multi-dimensional transaction environment state is injected into the agent social network, so that each simulated user agent makes multi-round transaction decisions according to the multi-dimensional transaction environment state and the decision influence relationship; the risk categories of each simulated user agent are identified; according to the decision preference type of each simulated user agent and the corresponding risk category of each simulated user agent, the target maintenance strategy of each simulated user agent is determined, and the target maintenance strategy is used to execute maintenance operations on the users of the corresponding decision preference type. It is mainly used for maintaining users with loss risk.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology and can be applied to financial technology scenarios. In particular, it relates to a user maintenance method, apparatus, medium, and equipment based on a simulated user intelligent agent. Background Technology

[0002] User maintenance refers to a series of strategies and actions that involve the continuous operation and management of acquired users based on user transaction records, behavioral patterns, and other data, in order to extend user lifecycle, enhance user value, and reduce churn risk. User maintenance is a crucial bridge from transaction monitoring to value continuation and an essential means to reduce churn and accumulate data assets.

[0003] Existing user maintenance methods often analyze investors' time-series behaviors such as login frequency, transaction frequency, and fund flows, using models like LSTM, GRU, and Transformer for sequence modeling. However, these methods typically treat investors as isolated individuals, ignoring group effects such as social influence and emotional transmission among investors. They struggle to capture the chain reaction of user churn triggered by sudden market events or public opinion crises, resulting in an inability to accurately and promptly identify user churn events and thus hinder timely user maintenance, leading to low timeliness. Summary of the Invention

[0004] In view of this, the present invention provides a user maintenance method, apparatus, medium, and equipment based on a simulated user intelligent agent, the main purpose of which is to address the problem of low timeliness in existing user maintenance.

[0005] According to one aspect of the present invention, a user maintenance method based on a simulated user agent is provided, comprising: Based on the multidimensional business-related data and user historical behavior data of the target product, an intelligent agent social network is constructed, wherein the intelligent agent social network includes simulated user intelligent agents matching different decision preference types and the decision influence relationship between each simulated user intelligent agent; In response to updates to external environment information, real-time external environment information is acquired, a multi-dimensional transaction environment state is simulated based on the real-time external environment information, and the multi-dimensional transaction environment state is injected into the intelligent agent social network, so that each simulated user intelligent agent can make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship. Based on the final transaction decisions of each simulated user agent, identify the churn risk category of each simulated user agent; For each simulated user agent's decision preference type and corresponding churn risk category, a target maintenance strategy is determined for each simulated user agent, and maintenance operations are performed on users with the corresponding decision preference type according to the target maintenance strategy.

[0006] According to another aspect of the present invention, a user maintenance device based on a simulated user intelligent agent is provided, comprising: The construction module is used to construct an intelligent agent social network based on the multidimensional business-related data and user historical behavior data of the target product. The intelligent agent social network includes simulated user intelligent agents that match different decision preference types and the decision influence relationships between each simulated user intelligent agent. The decision module is used to respond to updates of external environment information, acquire real-time external environment information, simulate a multi-dimensional transaction environment state based on the real-time external environment information, and inject the multi-dimensional transaction environment state into the intelligent agent social network, so that each simulated user intelligent agent can make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship. The risk identification module is used to identify the churn risk category of each simulated user agent based on the final transaction decision of each simulated user agent; The strategy generation module is used to determine the target maintenance strategy for each of the simulated user agents based on their decision preference type and corresponding churn risk category, and to perform maintenance operations on users with the corresponding decision preference type according to the target maintenance strategy.

[0007] According to another aspect of the present invention, a medium is provided, the medium storing at least one executable instruction that causes a processor to perform operations corresponding to the user maintenance method based on simulated user agents described above.

[0008] According to another aspect of the present invention, a device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the user maintenance method based on the simulated user agent described above.

[0009] By employing the above-described technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages: This invention provides a user maintenance method, apparatus, medium, and device based on simulated user agents. First, an agent social network is constructed based on multi-dimensional business-related data of the target product and historical user behavior data. Responding to updates in external environment information, real-time external environment information is acquired, and a multi-dimensional transaction environment state is simulated based on this information. This multi-dimensional transaction environment state is then injected into the agent social network, enabling each simulated user agent to make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and decision-making influence relationships. Based on the final transaction decisions of each simulated user agent, the churn risk category of each simulated user agent is identified. For each simulated user agent's decision preference type and corresponding churn risk category, a target maintenance strategy is determined, and maintenance operations are performed on users with the corresponding decision preference type according to the target maintenance strategy. Compared with existing technologies, this invention, by constructing an agent social network and simulating users with different decision preferences making decisions based on each agent within the network, as well as the propagation of influence between user decisions, introduces social influence, emotional transmission, and other group effects into the user churn risk process. This allows for timely detection of user churn risks and user maintenance, thereby improving the timeliness of user maintenance.

[0010] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0011] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 The following is a flowchart of a user maintenance method based on a simulated user agent provided by an embodiment of the present invention; Figure 2 This invention provides a flowchart of another user maintenance method based on a simulated user agent, according to an embodiment of the present invention. Figure 3 This invention provides a flowchart of another user maintenance method based on a simulated user agent, according to an embodiment of the present invention. Figure 4 This diagram illustrates a block diagram of a user maintenance device based on a simulated user agent, according to an embodiment of the present invention. Figure 5 A schematic diagram of the structure of a device provided in an embodiment of the present invention is shown. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] This invention provides a user maintenance method based on simulated user agents. This method constructs agents to simulate users with different decision-making preferences for a target product. Based on the propagation and influence of decisions made by users of different preference types, a social network of agents is built to simulate the scenario where users influence each other's decisions and ultimately affect the overall transaction environment. Decision-making operations in this social network are triggered by updates to external environmental information. When the external environmental information affecting user decisions changes, a multi-dimensional transaction environment state is simulated using the current real-time external environmental information, and within this virtual world, the simulated user agents simulate the thinking and decision-making methods of the corresponding user groups, making feedback decisions on the multi-dimensional transaction environment state and tracking the propagation of these decisions within the social network. This yields the final decision for each simulated user agent. Based on the final decision, the churn risk and maintenance strategy for each simulated user agent's corresponding user group are determined. The execution entity of this method is the server of the subsystem used to identify and retain users within a financial business system; this server can be a cloud server or a local server. The server connects to interfaces of multiple information sources to monitor external environment information and connects to the sub-business systems of the target product to obtain user behavior data and business data of the product.

[0014] like Figure 1 As shown, the method includes steps 101-104: 101. Construct an intelligent agent social network based on the multidimensional business-related data and user historical behavior data of the target product.

[0015] In this embodiment of the invention, the target product is a financial product that requires monitoring of customer churn risk and customer retention efforts. The target product can be any financial product requiring monitoring and customer maintenance, such as fund products, wealth management products, loan products, or insurance products. To construct an intelligent agent social network capable of simulating its users and their social networks for the current target product, the current executing entity obtains business-related data of different dimensions from interfaces of multiple information sources, such as market data, product change information, news and public opinion data, and policy documents, thus obtaining multi-dimensional business-related data. User historical behavior data includes user historical transaction data and customer service dialogue records.

[0016] It's important to note that the intelligent agent social network includes simulated user agents matching different decision-making preference types and the decision-influence relationships between these simulated user agents. Simulated user agents are agents constructed based on the characteristics of users with corresponding decision-making preference types to simulate the decision-making process of that category of users; each decision-making preference type corresponds to one or more simulated agents. Decision-influence relationships characterize the mutual impact of decisions made by users; for example, the loss of some investors may lead to the departure of a large number of related investors. By constructing an intelligent agent social network, it's possible to simulate not only changes in user decisions based on changes in the external environment, but also the impact of different users on each other's next decisions after a user has made a decision. This allows for proactive identification of customer decisions and the ability to detect situations where a product's declining performance triggers a chain reaction of customer churn.

[0017] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2 As shown, based on the multidimensional business-related data and user historical behavior data of the target product, an intelligent agent social network is constructed, including: 1011. Perform entity recognition and entity relationship extraction on the multidimensional business-related data and the user historical behavior data, and construct a domain knowledge graph with entities as nodes and entity relationships as edges.

[0018] 1012. Based on the user's historical behavior data, perform cluster analysis to identify all decision preference types, and construct a corresponding simulated user agent for each decision preference type.

[0019] 1013. Extract the decision influence relationships between the simulated user agents based on the domain knowledge graph, and construct an agent social network with the simulated user agents as network nodes and the decision influence relationships as edge attributes.

[0020] In this embodiment of the invention, entities include user entities, market event entities, emotional states, and intervention strategies. Relationships between entities include operational relationships, decision-influence relationships, attention relationships, transmission relationships, and association relationships. Taking a three-year closed-end wealth management product as an example, the knowledge graph construction process requires integrating all relevant business data and user historical behavior to identify key entities such as user entities, market event entities, emotional state entities, and intervention strategy entities. Using these entities as nodes and the relationships between them (such as operational relationships, decision-influence relationships, attention relationships, transmission relationships, and association relationships) as edges, a knowledge graph specific to this product is constructed. Furthermore, cluster analysis is performed based on user historical behavior data to identify all decision preference types, and a corresponding simulated user agent is constructed for each decision preference type. Finally, the decision-influence relationships between simulated user agents are extracted based on the domain knowledge graph. Using simulated user agents as network nodes and decision-influence relationships as edge attributes, an agent social network is constructed, enabling more realistic prediction of group behavior and the development of retention strategies.

[0021] It's important to note that a single simulated user agent can only represent the static decision-making logic of a particular type of investor. In contrast, investors in real financial markets do not make decisions in isolation; they observe and imitate each other, experience contagious emotions, or are influenced by group behavior. Domain knowledge graphs precisely record the various relationships between user entities and between users and events—such as attention relationships, transmission relationships, and operational relationships. These relationships inherently imply the propagation paths of decision-making influence. By extracting these relationships from the graph, isolated agents can be connected into an interactive social network, making churn risk prediction and retention strategy formulation more closely reflect real market dynamics.

[0022] In one embodiment of the present invention, for further explanation and limitation, the process of constructing a simulated user agent of any decision preference type includes: Extract the target historical behavior data of users with the aforementioned decision preference type; The target historical behavior data is arranged in chronological order to construct a long-term memory sequence containing multiple memories, and the long-term memory sequence is stored in a vector database to obtain a long-term memory bank; Extract the user's factual behavioral attributes from the long-term memory, and construct a basic profile based on the factual behavioral attributes; The loss aversion coefficient is obtained by fitting user feedback behavior after different transaction results in the long memory sequence, and the conformity tendency coefficient is obtained by fitting user feedback behavior under different market hotness. The loss aversion coefficient and the conformity tendency coefficient are stored as psychological features. Using the multidimensional transaction environment state extracted from the target historical behavior data as training samples, the basic profile and the psychological features as state feature inputs, and the event context as context inputs, a reinforcement learning strategy model is trained using the behavior cloning method to obtain a decision model whose decision output satisfies the actual decision distribution of the decision preference type.

[0023] In this embodiment of the invention, the simulated user agent includes a long-term memory bank, a basic profile, psychological characteristics, and a decision-making model. User historical behavior data is divided according to the user's decision preference type. Therefore, when training an agent for any decision preference type, user historical behavior data belonging only to the current decision preference type is extracted, i.e., the target user's historical behavior data. Furthermore, a long-term memory sequence containing multiple memories is constructed chronologically and stored in a vector database to form the long-term memory bank. This long-term memory sequence differs from common short-term conversational memory; each memory in the sequence includes event context, a timestamp, and user feedback behavior, thus preserving the user's complete decision-making trajectory on a transaction-by-transaction basis. For example, the market state before and after a purchase, changes in holdings, and subsequent feedback—a coherent memory sequence can extract the user's transaction and decision-making characteristics.

[0024] Then, the system extracts factual behavioral attributes generated from actual user transactions, such as transaction frequency, holding period, and product preference—statistically objective indicators—from this memory to construct a basic profile. Next, it fits the user feedback behavior after different transaction outcomes in the long-term memory sequence to obtain a loss aversion coefficient. This coefficient quantifies how much a user is more sensitive to losses than to gains when faced with equal profits or losses. The loss aversion coefficient can be fitted by statistically analyzing the probability differences in different feedback behaviors triggered by users after each transaction settlement, such as selling a fund, under the same magnitude of profit and loss—e.g., redemption, additional investment, or switching products. For example, the ratio of the redemption rate after average loss to the redemption rate after average profit can be used as the loss aversion coefficient.

[0025] Meanwhile, a conformity coefficient was obtained by fitting user feedback behavior under different market conditions. This coefficient quantifies the degree to which users tend to follow the mainstream trend rather than stick to independent judgment when the overall market is rising or falling. The conformity coefficient is specifically obtained by statistically analyzing the proportion of users who take the same actions as the mainstream during periods of significant market rise or fall (e.g., buying when the market is rising and selling when it is falling), and subtracting the proportion of users taking the same actions during periods of market stability. The larger the positive value of this coefficient, the stronger the conformity tendency.

[0026] The two coefficients mentioned above together serve as psychological features. Finally, multi-dimensional trading environment states extracted from the target's historical behavioral data, such as price volatility, market trading volume, and sector popularity, are used as training samples. Basic profiles and psychological characteristics are used as state feature inputs describing the user's intrinsic attributes, while external information such as financial report releases and policy changes is used as contextual input. A behavioral cloning approach is employed to mimic the user's historical decision-making trajectory to train the reinforcement learning strategy model, resulting in the current simulated user agent's decision-making model. Through this training method, the model's decision output in a given state can approximate the actual decision distribution of that decision preference type.

[0027] 102. In response to the update of external environment information, obtain real-time external environment information, simulate a multi-dimensional transaction environment state based on the real-time external environment information, and inject the multi-dimensional transaction environment state into the intelligent agent social network, so that each simulated user intelligent agent can make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship.

[0028] In this embodiment of the invention, updates to external environmental information are continuously monitored. Once changes are detected in real-time information such as market data, public opinion news, policy announcements, or sudden events, this information is immediately collected and transformed into a standardized multi-dimensional trading environment state vector through multi-dimensional mapping and numerical encoding. This vector includes state representations such as volatility, sentiment scores, and policy factors. This state vector is then used as input to the constructed intelligent agent social network. Each simulated user agent in the intelligent agent social network, upon receiving the current multi-dimensional trading environment state, combines its own basic profile and psychological characteristics with reference to the historical decisions or current tendencies of neighboring agents defined by decision-making influence relationships within the social network. This results in actions such as buying, selling, or holding. In this way, the evolutionary process of group decision-making formed by investors observing, imitating, or contagious emotions in a dynamically changing environment in the real market can be simulated, thus providing a dynamic and predictable basis for predicting churn risk and developing targeted retention strategies.

[0029] In one embodiment of the present invention, for further explanation and limitation, a multi-dimensional trading environment state is simulated based on the real-time external environment information, including: Based on the semantic feature extraction results of the real-time external environment information, the real-time external environment information is classified to obtain at least one type of external environment information from market data, news data, policy documents, and product announcements. According to the feature extraction method of matching information type, environmental state features are extracted from the external environment information to obtain an environmental state vector; Based on the environmental state vector, a multi-dimensional transaction environment state is identified from the domain knowledge graph through a graph retrieval-based enhanced generation method.

[0030] In this embodiment of the invention, since real-time external environment information can be multiple pieces of information from multiple data sources and covers different information types, after obtaining the real-time external environment information, it is first necessary to determine the information type covered by this external information through semantic feature recognition. This category can be one or more of market data, news data, policy documents, and product announcements. After determining the information type, feature extraction is performed on the information according to the feature extraction method corresponding to each information type: quantitative indicators are generated for market data, such as calculating price volatility and changes in trading volume; sentiment analysis and keyword extraction are performed for news data to quantify market sentiment and hot topics; rule parsing and impact assessment are performed for policy documents to extract policy direction and intensity factors; and product information is extracted for product announcements, such as net asset value adjustments or fee changes, ultimately obtaining the environmental state vectors corresponding to different types of information.

[0031] The process involves obtaining environmental state vectors. Based on these vectors, a graph-based retrieval-enhanced generative approach is used to retrieve relevant entities and relationships from the domain knowledge graph. This generates a complete representation that integrates external environmental information and structured knowledge from the graph, identifying the current multidimensional trading environment state and providing structured environmental input for subsequent simulations of user agent decision-making. Specifically, during the retrieval process, the environmental state vectors serve as the query index for graph retrieval. Vectors such as market sentiment, intensity factors, and quantitative indicators form a set of feature values. These feature values ​​are matched with market event entities in the graph. For example, the currently calculated high volatility feature can be matched with nodes of drastic fluctuations in the graph, and negative news sentiment scores can be matched with nodes of bearish public opinion. This yields event entities, relationship paths, and associated subgraphs that match the current environmental state vector. Then, the retrieved structured knowledge, such as the transmission path of user behavior in historically similar market environments or the impact chain of specific events on product entities, is semantically fused and reordered with the original environmental state vectors to obtain fused features. Finally, a generative model outputs an enhanced representation that integrates the real-time environmental state vectors and prior graph association knowledge—the multidimensional trading environment state. The graph retrieval-enhanced generation method utilizes the structured knowledge of the graph to supplement and enrich the original environmental information. The final identified multidimensional transaction environment state is a product of the fusion of the original environmental information and the prior knowledge of the graph.

[0032] In one embodiment of the present invention, for further explanation and limitation, each of the simulated user intelligent agents performs multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship, including: Each of the simulated user agents makes a first-round transaction decision based on the multi-dimensional transaction environment state; After completing the i-th round of transaction decisions for each of the simulated user agents, a decision update is performed for each simulated user agent to complete the (i+1)-th round of transaction decisions for each simulated user agent, where i∈[1,N] and N represents the preset iteration round: Specifically, the decision update includes: Based on the decision influence relationship, the associated simulated user agent is determined, and the next round of transaction decision is made for the simulated user agent according to the i-th round transaction decision of the associated simulated user agent, so as to obtain the i+1-th round transaction decision of each simulated user agent. The market environment state is updated based on the (i+1)th round of trading decisions of each simulated user agent. The multidimensional trading environment state is then updated based on the updated market environment state, and the updated multidimensional trading environment state is re-injected into the agent's social network as the environment perception input for the next round of trading decisions. After completing the i-th round of transaction decisions for each of the simulated user agents, the step of updating the decision for each simulated user agent is repeated until the number of transaction decision rounds equals the preset number of iteration rounds.

[0033] In this embodiment of the invention, all simulated user agents simultaneously complete the first round of trading decisions based on the current multidimensional trading environment state. After completing each round of trading decisions, for each simulated user agent, other simulated user agents associated with it are identified based on the decision influence relationships in the agent social network. The actual decision results of these associated agents in the current round (round i) guide the current agent to make the next round of trading decisions, i.e., decision update, thus completing the next round (round i+1). This process is repeated cyclically, with i increasing from 1 to a preset iteration round N. Finally, all trading actions generated by all agents in round i+1 are summarized to update the market environment state. For example, price changes or trading volume changes are simulated, and the multidimensional trading environment state is refreshed based on the updated market environment state. The refreshed multidimensional trading environment state is then reinjected into the agent social network to drive the next round of iterative decisions. This process is repeated until all N iterations are completed, thereby simulating the dynamic decision-making evolution process of a group of investors under the combined influence of mutual influence and environmental feedback.

[0034] In one embodiment of the present invention, for further explanation and limitation, the process by which any simulated user intelligent agent makes decisions based on the multidimensional transaction environment state includes: The simulated user agent's perception of transaction results is updated based on the market environment status, the transaction trust level of the simulated user agent is updated based on the public opinion environment status, the transaction motivation of the simulated user agent is updated based on the policy environment status, and the long-term memory is queried based on the event environment status to obtain experience feedback behavior. The updated transaction result perception, transaction trust level, and transaction motivation, along with the retrieved experience feedback behavior, are concatenated into a multi-dimensional transaction environment state vector. The multidimensional trading environment state vector, the basic profile and psychological characteristics of the simulated user agent are used as inputs to the decision model deployed in the simulated user agent. The decision model is then used to make decisions, resulting in the trading decisions of the simulated user agent.

[0035] The trading environment state includes market environment state, public opinion environment state, policy environment state, and event environment state. In this embodiment of the invention, the internal state of the simulated user agent is updated according to each dimension of the multi-dimensional trading environment state: the perception of trading results is updated according to the market environment state, that is, the agent's subjective feelings about the current profit or loss of holdings or the gains and losses of historical transactions, reflecting its emotions and satisfaction after profit or loss; the trading trust level is updated according to the public opinion environment state, that is, the agent's confidence in the reliability of current market information and the correctness of its own decisions, affecting whether it is willing to continue to execute the original strategy; the trading motivation is updated according to the policy environment state, that is, the strength of the agent's internal driving force to trade, reflecting its willingness to buy, sell, or wait and see due to changes in external rules; and the experience feedback behavior under similar historical events is retrieved from the long-term memory bank according to the event environment state. Ideally, the updated transaction result perception, transaction trust level, transaction motivation, and queried experience feedback behavior should be concatenated into a multi-dimensional transaction environment state vector. This vector, along with the basic profile and psychological characteristics of the simulated user agent, should then be input into the decision model within the agent. The decision model should then output the agent's transaction decisions, thereby achieving a refined simulation of user psychological changes and behavioral responses under different external environments.

[0036] 103. Based on the final transaction decisions of each simulated user agent, identify the churn risk category of each simulated user agent.

[0037] In this embodiment of the invention, based on the final trading decisions output by each simulated user agent after all N iterations (e.g., liquidation, continued holding, partial redemption, or additional investment), and combined with the decision preference type corresponding to the agent, the churn risk category of each simulated user agent is identified through a preset churn risk mapping rule or classification model. The churn risk categories include mild warning, moderate warning, severe warning, and imminent churn. For example, the criteria for a mild warning might be: decreased trading frequency, reduced inquiries, and a 30% decrease in trading or inquiry frequency compared to the historical average; the criteria for a moderate warning might be: continuous redemption or frequent portfolio adjustments, with redemption operations occurring for three consecutive rounds; the criteria for a severe warning might be: continuous capital outflow, with a capital outflow ratio >50%; and the criteria for imminent churn might be: applying for cancellation or full redemption.

[0038] 104. For each of the simulated user agents’ decision preference types and their corresponding churn risk categories, determine the target maintenance strategy for each of the simulated user agents, and perform maintenance operations on users with the corresponding decision preference types according to the target maintenance strategy.

[0039] In this embodiment of the invention, after determining the churn risk categories for different decision-making preference types, it is equivalent to predicting the churn risk categories of users with corresponding decision-making preference types. It is then necessary to determine whether intervention is required based on the risk category, and what user maintenance strategy to apply for intervention. Mild and moderate warnings can be monitored and their causes analyzed, but specific target maintenance strategies are not required for intervention. Moderate and severe warnings require intervention, i.e., a target maintenance strategy needs to be selected. Since users with different decision-making preference types are more receptive to different intervention strategies, when determining the target maintenance strategy, it is also necessary to combine the decision-making preference type and propose targeted maintenance strategies to improve the success rate of user maintenance.

[0040] In one embodiment of the present invention, for further illustration and limitation, such as Figure 3 As shown, the process of determining the target maintenance strategy for any simulated user agent includes: 1041. Based on the churn risk category and decision preference type of the simulated user agent, multiple candidate maintenance strategies are matched from the maintenance strategy set.

[0041] 1042. The candidate maintenance strategies are tested in parallel on the simulated user agent to obtain the multi-dimensional execution effect parameters of each candidate maintenance strategy.

[0042] 1043. Perform a weighted summation on the execution effect parameters of each dimension, and determine the candidate maintenance strategy with the largest weighted summation value as the target maintenance strategy.

[0043] In this embodiment of the invention, candidate maintenance strategies may include dedicated customer service support, providing users with revenue attribution reports, product conversion suggestions, fee discounts, redemption recovery reminders, and displaying social proof. The maintenance strategy set includes the mapping relationship between different risk categories and different decision preference types and maintenance strategies. For example, for conservative users at high risk of churn, multiple maintenance strategies such as fee discounts, dedicated customer service support, or product conversion suggestions can be considered as candidates. These candidate maintenance strategies are tested in parallel on a simulated user agent, i.e., each strategy is simulated and executed in the current simulation environment, and multi-dimensional execution effect parameters are collected after each strategy is executed, such as the decrease in churn probability, expected revenue contribution, strategy execution cost, and user satisfaction score. Finally, the execution effect parameters of each dimension are weighted and summed. The weights of each dimension can be preset and adjusted according to business objectives, such as prioritizing reducing churn rate or controlling costs.

[0044] This invention provides a user maintenance method based on simulated user agents. First, an agent social network is constructed based on multi-dimensional business-related data of the target product and historical user behavior data. Responding to updates in external environment information, real-time external environment information is acquired, and a multi-dimensional transaction environment state is simulated based on this information. This multi-dimensional transaction environment state is then injected into the agent social network, enabling each simulated user agent to make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and decision-making influence relationships. Based on the final transaction decisions of each simulated user agent, the churn risk category of each simulated user agent is identified. For each simulated user agent's decision preference type and corresponding churn risk category, a target maintenance strategy is determined, and maintenance operations are performed on users with the corresponding decision preference type according to the target maintenance strategy. Compared with existing technologies, this invention, by constructing an agent social network and simulating users with different decision preferences making decisions based on each agent within the network, as well as the propagation of influence between user decisions, introduces social influence, emotional transmission, and other group effects into the user churn risk process. This allows for timely detection of user churn risks and user maintenance, thereby improving the timeliness of user maintenance.

[0045] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this embodiment of the invention provides a user maintenance device based on a simulated user intelligent agent, such as... Figure 4 As shown, the device includes: Module 31 is used to construct an intelligent agent social network based on the multidimensional business association data and user historical behavior data of the target product. The intelligent agent social network includes simulated user intelligent agents matching different decision preference types and the decision influence relationship between each simulated user intelligent agent. Decision module 32 is used to respond to updates of external environment information, obtain real-time external environment information, simulate a multi-dimensional transaction environment state based on the real-time external environment information, and inject the multi-dimensional transaction environment state into the intelligent agent social network, so that each simulated user intelligent agent can make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship. The risk identification module 33 is used to identify the churn risk category of each simulated user agent based on the final transaction decision of each simulated user agent; The strategy generation module 34 is used to determine the target maintenance strategy for each of the simulated user agents based on their decision preference type and corresponding churn risk category, and to perform maintenance operations on users with the corresponding decision preference type according to the target maintenance strategy.

[0046] Furthermore, the decision module 32 includes: The first decision-making unit is used for each of the simulated user intelligent agents to make a first round of transaction decisions based on the multi-dimensional transaction environment state. The second decision-making unit, after completing the i-th round of transaction decisions for each of the simulated user agents, executes the following process for each simulated user agent to complete the (i+1)-th round of transaction decisions for each simulated user agent, where i∈[1,N] and N represents the preset iteration round: Based on the decision influence relationship, the associated simulated user agent is determined, and the simulated user agent is used to make the next round of transaction decisions based on the i-th round of transaction decisions of the associated simulated user agent, so as to complete the i+1-th round of transaction decisions for each simulated user agent. The environment update unit is used to update the market environment state based on the (i+1)th round of transaction decisions of each simulated user agent, update the multidimensional transaction environment state according to the updated market environment state, and re-inject the updated multidimensional transaction environment state into the agent social network.

[0047] Furthermore, the multi-dimensional transaction environment state includes market environment state, public opinion environment state, policy environment state, and event environment state; the process by which any simulated user agent in the first or second decision-making unit makes a decision based on the multi-dimensional transaction environment state includes: The simulated user agent's perception of transaction results is updated based on the market environment status, the transaction trust level of the simulated user agent is updated based on the public opinion environment status, the transaction motivation of the simulated user agent is updated based on the policy environment status, and the long-term memory is queried based on the event environment status to obtain experience feedback behavior. The updated transaction result perception, transaction trust level, and transaction motivation, along with the retrieved experience feedback behavior, are concatenated into a multi-dimensional transaction environment state vector. The multidimensional trading environment state vector, the basic profile and psychological characteristics of the simulated user agent are used as inputs to the decision model deployed in the simulated user agent. The decision model is then used to make decisions, resulting in the trading decisions of the simulated user agent.

[0048] Furthermore, module 31 includes: The relationship extraction unit is used to perform entity recognition and entity relationship extraction on the multidimensional business association data and the user historical behavior data, and to construct a domain knowledge graph with entities as nodes and entity relationships as edges. The entities include user entities, market event entities, emotional states and intervention strategies, and the entity relationships include business operation relationships, decision influence relationships, attention relationships, transmission relationships and association relationships. The identification unit is used to perform cluster analysis based on the user's historical behavior data, identify all decision preference types, and construct a corresponding simulated user agent for each decision preference type; The construction unit extracts the decision influence relationships between the simulated user agents based on the domain knowledge graph, and constructs an agent social network with the simulated user agents as network nodes and the decision influence relationships as edge attributes.

[0049] Furthermore, the simulated user agent includes a long-term memory bank, a basic profile, psychological characteristics, and a decision-making model; in specific application scenarios, the construction unit is specifically used in the construction process of a simulated user agent of any decision preference type, including: Extract the target historical behavior data of users with the aforementioned decision preference type; The target historical behavior data is constructed into a long-term memory sequence containing multiple memories in chronological order, and the long-term memory sequence is stored in a vector database to obtain a long-term memory bank, wherein each memory includes event context, timestamp and user feedback behavior; Extract the user's factual behavioral attributes from the long-term memory, and construct a basic profile based on the factual behavioral attributes; The loss aversion coefficient is obtained by fitting user feedback behavior after different transaction results in the long memory sequence, and the conformity tendency coefficient is obtained by fitting user feedback behavior under different market hotness. The loss aversion coefficient and the conformity tendency coefficient are stored as psychological features. Using the multidimensional transaction environment state extracted from the target historical behavior data as training samples, the basic profile and the psychological features as state feature inputs, and the event context as context inputs, a reinforcement learning strategy model is trained using the behavior cloning method to obtain a decision model whose decision output satisfies the actual decision distribution of the decision preference type.

[0050] Furthermore, the decision module 32 also includes: The classification unit is used to classify the real-time external environment information based on the semantic feature extraction results of the real-time external environment information, and obtain at least one type of external environment information among market data, news data, policy documents, and product announcements. An environment state extraction unit is used to extract environment state features from the external environment information according to the feature extraction method of the matching information type, thereby obtaining an environment state vector. The feature extraction method of the matching information type includes: generating quantitative indicators from the market data; performing sentiment analysis and keyword extraction from the news data; performing rule parsing and impact assessment from the policy documents; and extracting product information from the product announcements. The environment identification unit is used to identify the multi-dimensional transaction environment state from the domain knowledge graph based on the environment state vector and through a graph retrieval-based enhanced generation method.

[0051] Furthermore, the strategy generation module 34 includes: The matching unit is used to match multiple candidate maintenance strategies from the maintenance strategy set based on the churn risk category and decision preference type of the simulated user agent; The testing unit is used to perform parallel sandbox tests on the simulated user agent to obtain multi-dimensional execution effect parameters of each candidate maintenance strategy. The determination unit is used to perform a weighted summation of the execution effect parameters of each dimension, and to determine the candidate maintenance strategy with the largest weighted summation value as the target maintenance strategy.

[0052] This invention provides a user maintenance device based on simulated user agents. First, an agent social network is constructed based on multi-dimensional business-related data of the target product and historical user behavior data. Responding to updates in external environment information, real-time external environment information is acquired, and a multi-dimensional transaction environment state is simulated based on this information. This multi-dimensional transaction environment state is then injected into the agent social network, enabling each simulated user agent to make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and decision-making influence relationships. Based on the final transaction decisions of each simulated user agent, the churn risk category of each simulated user agent is identified. For each simulated user agent's decision preference type and corresponding churn risk category, a target maintenance strategy is determined, and maintenance operations are performed on users with the corresponding decision preference type according to the target maintenance strategy. Compared with existing technologies, this invention, by constructing an agent social network and simulating users with different decision preferences making decisions based on each agent within the network, as well as the propagation of influence between user decisions, introduces social influence, emotional transmission, and other group effects into the user churn risk process. This allows for timely detection of user churn risks and user maintenance, thereby improving the timeliness of user maintenance.

[0053] According to one embodiment of the present invention, a medium is provided having at least one executable instruction that can execute the user maintenance method based on simulated user agent in any of the above method embodiments.

[0054] Figure 5 The diagram shows a structural schematic of a device according to an embodiment of the present invention. The specific implementation of the device is not limited by the specific embodiments of the present invention.

[0055] like Figure 5 As shown, the device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0056] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0057] Communication interface 404 is used to communicate with other network elements such as clients or other servers.

[0058] The processor 402 is used to execute program 410, specifically to execute the relevant steps in the above-described user maintenance method embodiment based on simulated user intelligent agents.

[0059] Specifically, program 410 may include program code that includes computer operation instructions.

[0060] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The device may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0061] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0062] Specifically, program 410 can be used to cause processor 402 to perform the following operations: Based on the multidimensional business-related data and user historical behavior data of the target product, an intelligent agent social network is constructed, wherein the intelligent agent social network includes simulated user intelligent agents matching different decision preference types and the decision influence relationship between each simulated user intelligent agent; In response to updates to external environment information, real-time external environment information is acquired, a multi-dimensional transaction environment state is simulated based on the real-time external environment information, and the multi-dimensional transaction environment state is injected into the intelligent agent social network, so that each simulated user intelligent agent can make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship. Based on the final transaction decisions of each simulated user agent, identify the churn risk category of each simulated user agent; For each simulated user agent's decision preference type and corresponding churn risk category, a target maintenance strategy is determined for each simulated user agent, and maintenance operations are performed on users with the corresponding decision preference type according to the target maintenance strategy.

[0063] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0064] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A user maintenance method based on simulated user intelligent agents, characterized in that, include: Based on the multidimensional business-related data and user historical behavior data of the target product, an intelligent agent social network is constructed, wherein the intelligent agent social network includes simulated user intelligent agents matching different decision preference types and the decision influence relationship between each simulated user intelligent agent; In response to updates to external environment information, real-time external environment information is acquired, a multi-dimensional transaction environment state is simulated based on the real-time external environment information, and the multi-dimensional transaction environment state is injected into the intelligent agent social network, so that each simulated user intelligent agent can make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship. Based on the final transaction decisions of each simulated user agent, identify the churn risk category of each simulated user agent; For each simulated user agent's decision preference type and corresponding churn risk category, a target maintenance strategy is determined for each simulated user agent, and maintenance operations are performed on users with the corresponding decision preference type according to the target maintenance strategy.

2. The method according to claim 1, characterized in that, Each of the simulated user agents makes multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision-making influence relationship, including: Each of the simulated user agents makes a first-round transaction decision based on the multi-dimensional transaction environment state; After completing the i-th round of transaction decisions for each of the simulated user agents, a decision update is performed for each of the simulated user agents to complete the (i+1)-th round of transaction decisions for each simulated user agent, where i∈[1,N] and N represents the preset iteration round; Specifically, the decision update includes: Based on the decision influence relationship, the associated simulated user agent is determined, and the next round of transaction decision is made for the simulated user agent according to the i-th round transaction decision of the associated simulated user agent, so as to obtain the i+1-th round transaction decision of each simulated user agent. The market environment state is updated based on the (i+1)th round of trading decisions of each simulated user agent. The multidimensional trading environment state is then updated based on the updated market environment state, and the updated multidimensional trading environment state is re-injected into the agent's social network as the environment perception input for the next round of trading decisions. After completing the i-th round of transaction decisions for each of the simulated user agents, the step of updating the decision for each simulated user agent is repeated until the number of transaction decision rounds equals the preset number of iteration rounds.

3. The method according to claim 2, characterized in that, The multidimensional trading environment status includes market environment status, public opinion environment status, policy environment status, and event environment status; The process by which any simulated user agent makes decisions based on the state of the multidimensional transaction environment includes: The simulated user agent's perception of transaction results is updated based on the market environment status, the transaction trust level of the simulated user agent is updated based on the public opinion environment status, the transaction motivation of the simulated user agent is updated based on the policy environment status, and the long-term memory is queried based on the event environment status to obtain experience feedback behavior. The updated transaction result perception, transaction trust level, and transaction motivation, along with the retrieved experience feedback behavior, are concatenated into a multi-dimensional transaction environment state vector. The multidimensional trading environment state vector, the basic profile and psychological characteristics of the simulated user agent are used as inputs to the decision model deployed in the simulated user agent. The decision model is then used to make decisions, resulting in the trading decisions of the simulated user agent.

4. The method according to claim 1, characterized in that, Based on the multidimensional business-related data and user historical behavior data of the target product, an intelligent agent social network is constructed, including: Entity identification and entity relationship extraction are performed on the multidimensional business-related data and the user historical behavior data. A domain knowledge graph is constructed with entities as nodes and entity relationships as edges. The entities include user entities, market event entities, emotional states and intervention strategies. The entity relationships include business operation relationships, decision-making influence relationships, attention relationships, transmission relationships and association relationships. Cluster analysis is performed based on the user's historical behavior data to identify all decision preference types, and a corresponding simulated user agent is constructed for each decision preference type; The decision-making influence relationships between the simulated user agents are extracted based on the domain knowledge graph, and a social network of agents is constructed with the simulated user agents as network nodes and the decision-making influence relationships as edge attributes.

5. The method according to claim 4, characterized in that, The simulated user agent includes a long-term memory bank, a basic profile, psychological characteristics, and a decision-making model; The process of constructing a simulated user agent for any decision preference type includes: Extract the target historical behavior data of users with the aforementioned decision preference type; The target historical behavior data is constructed into a long-term memory sequence containing multiple memories in chronological order, and the long-term memory sequence is stored in a vector database to obtain a long-term memory bank, wherein each memory includes event context, timestamp and user feedback behavior; Extract the user's factual behavioral attributes from the long-term memory, and construct a basic profile based on the factual behavioral attributes; The loss aversion coefficient is obtained by fitting user feedback behavior after different transaction results in the long memory sequence, and the conformity tendency coefficient is obtained by fitting user feedback behavior under different market hotness. The loss aversion coefficient and the conformity tendency coefficient are stored as psychological features. Using the multidimensional transaction environment state extracted from the target historical behavior data as training samples, the basic profile and the psychological features as state feature inputs, and the event context as context inputs, a reinforcement learning strategy model is trained using the behavior cloning method to obtain a decision model whose decision output satisfies the actual decision distribution of the decision preference type.

6. The method according to claim 1, characterized in that, Based on the aforementioned real-time external environment information, a multi-dimensional trading environment state is simulated, including: Based on the semantic feature extraction results of the real-time external environment information, the real-time external environment information is classified to obtain at least one type of external environment information from market data, news data, policy documents, and product announcements. According to the feature extraction method of the matching information type, environmental state features are extracted from the external environment information to obtain an environmental state vector. The feature extraction method of the matching information type includes: generating quantitative indicators from the market data; performing sentiment analysis and keyword extraction from the news data; performing rule parsing and impact assessment from the policy documents; and extracting product information from the product announcements. Based on the environmental state vector, a multi-dimensional transaction environment state is identified from the domain knowledge graph through a graph retrieval-based enhanced generation method.

7. The method according to claim 1, characterized in that, The process of determining the target maintenance strategy for any simulated user agent includes: Based on the churn risk category and decision preference type of the simulated user agent, multiple candidate maintenance strategies are matched from the maintenance strategy set; The candidate maintenance strategies are tested in parallel sandbox on the simulated user agent to obtain the multi-dimensional execution effect parameters of each candidate maintenance strategy. The execution effect parameters of each dimension are weighted and summed, and the candidate maintenance strategy with the largest weighted sum is determined as the target maintenance strategy.

8. A user maintenance device based on a simulated user intelligent agent, characterized in that, include: The construction module is used to construct an intelligent agent social network based on the multidimensional business-related data and user historical behavior data of the target product. The intelligent agent social network includes simulated user intelligent agents that match different decision preference types and the decision influence relationships between each simulated user intelligent agent. The decision module is used to respond to updates of external environment information, acquire real-time external environment information, simulate a multi-dimensional transaction environment state based on the real-time external environment information, and inject the multi-dimensional transaction environment state into the intelligent agent social network, so that each simulated user intelligent agent can make multiple rounds of transaction decisions based on the multi-dimensional transaction environment state and the decision influence relationship. The risk identification module is used to identify the churn risk category of each simulated user agent based on the final transaction decision of each simulated user agent; The strategy generation module is used to determine the target maintenance strategy for each of the simulated user agents based on their decision preference type and corresponding churn risk category, and to perform maintenance operations on users with the corresponding decision preference type according to the target maintenance strategy.

9. A medium, characterized in that, The medium stores at least one executable instruction that causes the processor to perform the operation corresponding to the user maintenance method based on simulated user agents as described in any one of claims 1-7.

10. A device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the user maintenance method based on simulated user agents as described in any one of claims 1-7.

Images