Scene self-adaptive scoring method and system based on double-agent collaboration

Abstract

The invention discloses a scene self-adaptive scoring method and system based on double-agent collaboration, and belongs to the technical field of artificial intelligence and data processing. The method comprises the following steps: identifying an intelligent agent through a user intention, and receiving and identifying a customer demand containing a target marketing scene; the user intention recognition agent transmits recognized target marketing scene information to a scoring modeling agent; the score modeling agent adaptively screens and obtains personalized user data related to the target marketing scene from a database containing operator data according to the target marketing scene; and the score modeling agent fuses the personalized user data and the universal user data, and performs score modeling on the users based on the fused data to obtain scores of the users in the target marketing scene. According to the invention, automation, precision and high efficiency of the scoring system are realized, and the efficiency and accuracy of marketing scoring are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and data processing, in particular to a scene adaptive scoring method and system based on double-agent cooperation. BACKGROUND

[0002] With the rapid development of big data and artificial intelligence technology, data-driven fine operation has become the core competitiveness of various industries. In the field of marketing, traditional extensive user delivery mode is gradually being eliminated, and instead, accurate insight into user needs and personalized touch are taking place. User scoring system as a key tool to achieve precision marketing aims to quantify the potential value of each user in a specific scenario through the analysis of user data.

[0003] Currently, the existing user scoring technology system mainly has the following technical defects: Single data source and data island problem: user behavior data is scattered in different platforms and data owners. Due to data security and privacy protection considerations, these data are difficult to be effectively integrated. This leads most scoring models to rely on single source data, resulting in the lack of key features and the lack of accuracy and fairness of the generated scoring results.

[0004] Poor scene adaptability and large resource consumption: marketing scenarios are complex and varied, covering finance, education, travel and many other fields. Traditional scoring models usually need to be independently designed, developed and trained for each specific marketing scenario, lacking a general scoring system that can adapt to different scenarios. This mode not only limits the accuracy of scoring, but also causes a large amount of repeated consumption of human and computing resources.

[0005] Fragmented process and low efficiency: in the traditional marketing process, customer demand identification, scoring modeling and user touch are usually three independent stages completed by different teams or even different service providers. This leads to high communication costs between stages, untimely information synchronization, security risks in data transfer, and low overall operational efficiency.

[0006] Limited data processing capacity: traditional scoring models mostly rely on structured numerical or categorical data, and have weak processing capacity for unstructured data such as text and dialogue records, making it difficult to deeply mine the user intent and preferences contained therein.

[0007] Therefore, how to integrate multi-source heterogeneous data, especially to utilize the unique data resources of operators, and to build an efficient scoring system that can automatically adapt to different marketing scenarios and realize full-process integration, is a technical problem to be solved by those skilled in the art. SUMMARY

[0008] The main purpose of the present application is to provide a scene adaptive scoring system, method, device and medium based on double-agent cooperation, to solve the problem of single scoring model data source, poor scene adaptability, fragmented process and low efficiency in the prior art.

[0009] To achieve the above-mentioned purpose, one aspect of the present application provides a scene adaptive scoring method based on double-agent cooperation, which comprises: The user intent recognition agent receives and identifies the customer demand containing the target marketing scene; The user intent recognition agent passes the identified target marketing scene information to the scoring modeling agent; The scoring modeling agent adaptively filters and obtains personalized user data related to the target marketing scene from a database containing operator data according to the target marketing scene; The scoring modeling agent fuses the personalized user data with general user data, and performs scoring modeling on the user based on the fused data to obtain the score of each user in the target marketing scene, and According to the score, information is reached to the target user whose score meets the preset reach condition; feedback data of the target user is collected, and the feedback data is used for model optimization of the user intent recognition agent or the scoring modeling agent.

[0010] Further, the user intent recognition agent receives and identifies the customer demand containing the target marketing scene, comprising: evaluating the intent clarity of the customer demand through a first large language model; if the intent clarity meets the preset clarity condition, identifying the customer demand through a rule-based decision reasoning engine to obtain the target marketing scene; if the intent clarity does not meet the preset clarity condition, identifying the customer demand through a second large language model to obtain the target marketing scene; wherein the parameter quantity of the second large language model is greater than that of the first large language model.

[0011] Further, the scoring modeling agent adaptively filters and obtains personalized user data related to the target marketing scene according to the target marketing scene, comprising: the scoring modeling agent analyzes the target marketing scene information based on a pre-trained language model to generate a database query statement for querying the personalized user data; and executes the database query statement to obtain the personalized user data from the database.

[0012] Further, the operator data comprises at least one of B-domain data, O-domain data and M-domain data of a personal user; the general user data comprises at least one of user basic attribute information and communication behavior information; and the personalized user data comprises user online behavior data, interest preference data or travel trajectory data corresponding to the target marketing scene.

[0013] Further, before the scoring modeling agent acquires the personalized user data, the method further comprises: pre-processing data in the database, the pre-processing comprising: missing value filling based on a random forest algorithm, feature binning based on a chi-square binning method, and evidence weight WOE binning.

[0014] Further, the scoring modeling agent performs scoring modeling on users based on the fused data, comprising: taking the fused data as an input of a multi-layer perception model; and obtaining scores of each user in the target marketing scene through training of the multi-layer perception model.

[0015] Another aspect of the present application also provides a scene adaptive scoring device based on double-agent cooperation, comprising: a user intention recognition agent, configured to receive and recognize customer demand containing a target marketing scene, and deliver the recognized target marketing scene information to a scoring modeling agent; and the scoring modeling agent, configured to adaptively filter and acquire personalized user data related to the target marketing scene from a database containing operator data according to the target marketing scene; and fuse the personalized user data with general user data, and perform scoring modeling on users based on the fused data to obtain scores of each user in the target marketing scene.

[0016] Further, the customer demand containing a target marketing scene is received and recognized, comprising: a first large language model is used to evaluate the intention clarity of the customer demand; if the intention clarity meets a preset clarity condition, a rule-based decision reasoning engine is used to recognize the customer demand to obtain the target marketing scene; if the intention clarity does not meet the preset clarity condition, a second large language model is used to recognize the customer demand to obtain the target marketing scene; wherein the parameter quantity of the second large language model is greater than that of the first large language model.

[0017] Further, the personalized user data related to the target marketing scene is adaptively filtered and acquired from the database containing operator data according to the target marketing scene, comprising: The scoring modeling agent parses the target marketing scenario information based on a pre-trained language model to generate a database query statement for querying the personalized user data. The database query statement is executed to obtain the personalized user data from the database.

[0018] Further, the scoring modeling based on the fused data obtains the scores of each user in the target marketing scenario, including: The fused data is taken as an input of a multi-layer perception model; The multi-layer perception model is trained to obtain the scores of each user in the target marketing scenario.

[0019] Compared with the prior art, the technical solution provided by the present application has at least the following beneficial effects: The accuracy and fairness of the scoring are improved: by innovatively fusing the B-domain, O-domain and M-domain data unique to the operator, the data dimension of the scoring model is greatly enriched, the information bias caused by a single data source is overcome, and the user scoring is more comprehensive, objective and accurate.

[0020] High scene adaptability is achieved: the dual-agent collaborative architecture is introduced, the powerful natural language understanding capability of the large language model is used to accurately identify the real intention of the customer and the marketing scenario, and the related data is adaptively matched for modeling, without the need to develop a model for each scene, greatly improving the flexibility and universality of the system.

[0021] The overall process automation efficiency and security are improved: the customer demand identification, scoring modeling, user touch and effect feedback are integrated in a closed-loop system, realizing end-to-end automation processing, reducing manual intervention, avoiding security risks and communication costs caused by multi-link data transfer, and significantly improving the overall operation efficiency.

[0022] Resource consumption is reduced: through scene-adaptive data matching and modeling, the use of full data for analysis in each scene is avoided, the data storage and computing resources are saved to the greatest extent on the premise of ensuring the accuracy of the scoring. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be introduced below.

[0024] Figure 1 is the overall framework diagram of the scene-adaptive scoring system based on dual-agent collaboration according to an embodiment of the present application; Figure 2is a user intention recognition flowchart according to an embodiment of the present application; Figure 3 is a flowchart of a scenario adaptive scoring method based on double-agent cooperation according to an embodiment of the present application; Figure 4 is a scenario adaptive scoring system architecture diagram based on double-agent cooperation according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0026] The present application provides a scenario adaptive scoring system based on double-agent cooperation. The system aims to solve the personalized scoring of customers to users in different demand scenarios and accurately reach the target demand through an automated and integrated process. The system will be combined with Figures 1-3 The complete process of the scenario adaptive scoring method based on double-agent cooperation of the present application will be described in detail.

[0027] Step S101: receiving and identifying the customer demand containing the target marketing scenario through the user intention recognition agent This step is performed by the user intention recognition agent, which aims to accurately identify the marketing intention of the customer (such as a marketing personnel) through dialogue. In a preferred embodiment, the recognition process of the user intention recognition agent can adopt a hybrid architecture. First, the user intention recognition agent evaluates the clarity of the received customer demand through a first large language model (for example, a lightweight R1-1.5b model). If the evaluation result shows that the clarity meets the preset clarity condition (for example, the question pattern is fixed and the expression method is standardized), a rule-based decision reasoning engine is started to identify the customer demand, so as to efficiently and accurately obtain the target marketing scenario. If the evaluation result shows that the clarity does not meet the preset clarity condition (for example, the question is complex or the expression method is diverse), a second large language model (for example, an R1-32b model with more parameters than the first large language model) is started to deeply identify the customer demand, so as to more accurately understand the real intention and finally obtain the target marketing scenario. The specific steps are described as follows.

[0028] In this embodiment, the data resources are systematically integrated under the premise of complying with national laws, regulations and regulatory policies, ensuring information security and user privacy.

[0029] In one aspect, the operator data is integrated, and the B domain (business domain, such as user basic attributes, billing information), O domain (operation domain, such as call behavior, SMS behavior), and M domain (mobile domain, such as Internet behavior, terminal mobile data, travel trajectory data, and interest preference data) data of the individual user data in the Unicom data are synchronized to the data operation and modeling platform in a timely manner through an ETL (extraction, transformation, and loading) job. This provides a rich, unique, and multi-dimensional data basis for subsequent modeling, and solves the problems of difficult information collection and missing key information in the prior art.

[0030] On the other hand, training and fine-tuning data are prepared for the subsequent user intent recognition agent. Based on user historical conversation records, user feedback data, and other methods, a data set containing user intent and corresponding classification labels is widely collected. Data annotation experts accurately annotate these data to ensure that the data set covers a variety of user intents and provides high-quality corpora for model fine-tuning.

[0031] To improve data quality, the data used for user intent recognition and user scoring modeling needs to be preliminarily processed and screened. For the data used to fine-tune the user intent recognition agent, data annotation experts will accurately annotate the user historical records and user feedback data collected in step one in this step to ensure the intent classification label of each sample and provide high-quality data for model training. Finally, the data is cleaned, and noise and duplicate data are removed. At the same time, text data is segmented, and stop words are removed to make the data easier for the model to understand and process.

[0032] Three user intent and tagging result examples are given in Table 1. The tagging result mainly identifies information such as scene, gender, age, and region. The scene is a required field, and the age of the marketing user needs to be at least eighteen years old. In addition, the travel scene involves the origin and destination, so the travel city field is added to identify the scene.

[0033] Table 1. User intent and tagging result examples

[0034] As shown in Figure 2 This step is performed by the user intent recognition agent, which aims to accurately identify the marketing intent of the customer through the conversation. The user intent recognition agent in this embodiment adopts a hybrid architecture to balance efficiency and accuracy.

[0035] First, the agent receives the text or voice information input by the customer through the interface. For clear and patterned requests, the system will prefer to use the rule-based decision engine. This engine quickly extracts the intent through keyword list and regular expression matching. For matched keywords, the intent score is calculated according to the pre-set rules, and the highest scoring intent is selected.

[0036] The rule-based decision engine module usually uses hierarchical rule matching and template-driven methods to handle tasks with fixed patterns and relatively standardized expressions identified by the Deepseek large language model based on R1-1.5b, and provides fast and accurate responses. First, segmentation and entity recognition are needed to perform basic processing such as word segmentation, part-of-speech tagging, named entity recognition (NER) on the input text, extract key language units, and unify synonyms or near-synonymous expressions and filter meaningless symbols or stop words; then, the extracted key language units (query) are matched with the pre-defined keyword list (word-stock) by regular expressions to extract intent trigger words;

[0037] Match = REG(query, word-stock) (2) Tasks that are not successfully matched will be passed to the complex intent recognition module for complex task intent recognition; for tasks that match keywords, complex templates will be matched by combining dependency syntax or slot filling. For example:

[0038] A[input text] --> B{template library matching} B --> | is | C[extract intent and slot] B --> | no | D[enter complex intent recognition module] Then, based on the matching results, the intent category is assigned. If a high-precision template (such as "outbound industry ticket scenario") is hit, the user intent is directly determined. Otherwise, the highest scoring intent is selected by calculating the scores of multiple rule matching results. The specific formula is as follows:

[0039] (3) where wj is the rule weight, and Match(Rulej) is the rule matching score (0 or 1). The final user score in each intent is obtained, and the highest scoring intent is selected. If multiple intents have the same score, the task will be passed to the complex intent recognition module for further intent recognition to obtain the accurate user intent.

[0040] The complex intent recognition module is used to identify complex intent tasks that cannot be recognized by the rule engine. With the deep understanding of natural language of the DeepSeek R1-32B large language model, it can better understand the user's diversified request expression, and flexibly and accurately process the user's request. At the same time, compared with other models, DeepSeek R1-32B has the characteristics of high efficiency and accuracy, and can be called and used in various ways, suitable for different scenes and needs.

[0041] After the deployment of the model is completed, the data labeled by the data labeling expert in step S102 is used to fine-tune the model. First, the DeepSeek R1 API is used to call the model, which has better stability and response speed than other similar products. Then, in order to prevent the collected data from being too different from the general field and causing the model to not get good results, we fine-tune the DeepSeek R1-32B model on the newly collected specific task. Finally, the user's input is sent to the fine-tuned model for processing, and the model is used for intent recognition and classification to obtain the user's target intent scene and the circled user province information and age related information, providing effective data for the target scene adaptive user scoring modeling.

[0042] To verify the effectiveness of the module, 80% of the labeled data is used to fine-tune the model, and the remaining 20% is used as the test set. Experiments show that the user intent recognition rate of this module reaches 92%, indicating that it can effectively capture customer demand, and the rationality of this scheme is verified.

[0043] In this step, the agent based on the large language model interacts with the user to accurately identify the user's intent and lay a solid foundation for the subsequent downstream modeling tasks. Through this hybrid architecture, the intent recognition system can more accurately and quickly understand the user's intent, while effectively reducing the overall demand for hardware and reducing the hardware burden.

[0044] Step S102: The user intent recognition agent passes the target marketing scene information identified to the scoring modeling agent; This step needs to be based on the user intent obtained in step S101 to preliminarily circumscribe the user's customer group and match the scene related data adaptively. The user intent recognition agent passes the user intent and demand identified to the target scene adaptive user scoring modeling agent.

[0045] For the data used for user scoring modeling, a series of processing operations are performed: (1) Missing value filling: Random Forest algorithm is used for filling. For missing values of discrete features, a random forest classifier is used for prediction filling; for missing values of continuous features, a random forest regressor is used for prediction filling.

[0046] (2) Feature binning: the chi-square binning method is used to discretize continuous variables. This method assesses the similarity of adjacent intervals by calculating the chi-square statistic, and merges intervals with high similarity until the preset conditions are met, thereby simplifying the data structure and improving model stability.

[0047] (3) WOE (Weight of Evidence) binning: for each bin or discrete variable, the WOE value is calculated to quantify the degree of influence of each value on the target variable (such as "good" / "bad" samples). The calculation formula of WOE value is as follows:

[0048] (1) where Bad total represents the total number of bad samples, Bad X =X i represents the number of bad samples and X is the i-th group, Good X =X i represents the number of good samples and X is the i-th group; Good total represents the total number of good samples. By calculating the WOE value, the value of the bin or discrete variable can be converted into a numerical value, which can intuitively reflect the influence direction and degree of the bin or value on the target variable. A positive WOE value indicates that the bin or value is strongly associated with "good" samples, and a negative WOE value indicates that it is strongly associated with "bad" samples. In this way, in subsequent analysis and modeling, these quantitative information can be more easily utilized to improve the accuracy and interpretability of the model. In summary, through a series of data processing operations such as missing value filling, feature binning and WOE binning, the quality and usability of the data can be effectively improved, and the potential value of the data can be mined, providing strong support for subsequent analysis and modeling work.

[0049] Step S103: The scoring modeling agent adaptively filters and obtains personalized user data related to the target marketing scenario from a database containing operator data according to the target marketing scenario; The scoring modeling agent preliminarily circumscribes the target user group based on the user group information (including age, region, etc.) mentioned in the obtained user intention and demand, and if the customer does not explicitly mention the relevant information, data modeling is performed for all users. At the same time, the user demand obtained is transmitted to the DeepSeek R1-32B large language model, and the powerful natural language processing capability thereof is used for analysis, and a corresponding database query statement is generated according to the analyzed target demand scene to filter out the matched personalized user related data. For example, if the user's demand scene is children programming, the model needs to identify the user's online behavior data related to children programming in education, and needs to include the user's access times, duration, traffic and other related information of the related APP and website.

[0050] Then, the agent obtains the scene related personalized user data from the database through the query statement generated by the database interface. In addition, in order to more fully mine the value of the operator data and obtain more accurate user scores, the application still fully utilizes the user's basic attribute information (age, gender, region, etc.) and communication behavior (call, message and bill, etc.) information as the user feature data and scene personalized data of each demand scene as the data source to jointly perform multi-dimensional modeling to obtain more accurate user scores.

[0051] Step S104: The scoring modeling agent fuses the personalized user data and the general user data, and performs scoring modeling on the users based on the fused data to obtain the scores of the users in the target marketing scene Based on the above obtained scene personalized information and general scene information of each user, the data is merged, and then a multi-layer perception machine in a deep learning algorithm is used to realize user scoring. First, the embedding representation of the user is initialized based on the feature embedding of the user, specifically, for the n features of the user, each feature is initialized by a one-hot code vector. Then, a corresponding trainable feature embedding matrix is defined x ,

[0052] (4) wherein represents the one-hot code vector of the xth feature of the user, and the latent feature embedding of the user can be obtained by splicing all the user features, wherein and are the embedding representation of the user and the splicing operation, respectively.

[0053] (5) Then, the user embedding is sent into the multi-layer perception machine to calculate the user score, wherein represents a predicted user rating value.

[0054] (6) Finally, the accurate user rating can be obtained by minimizing the mean square error (MSE) loss, where R represents the number of all users, represents the real user rating. By optimizing the parameters of the iterative MLP, the accurate user rating can be obtained.

[0055] Step S105: According to the score, the information of the target user whose score meets the preset reach condition is reached, the feedback data of the target user is collected, and the feedback data is used for model optimization of the user intention recognition agent or the score modeling agent.

[0056] Based on the generated score result, high-score users are screened out. This step uses various ways such as mail, short message, voice, mailbox, 5G message number and enterprise business card to reach the user, establishes a feedback mechanism that can collect user feedback data on the reached information. With the unique advantages of the operator, under the premise of meeting the user privacy security agreement, the reach channel and method are selected according to the score result through each line, and personalized information is pushed to the user, and the user feedback data such as whether to click or purchase is collected and fed back to the data warehouse for model optimization adjustment. This step improves the whole process service ability and efficiency through various reach methods, realizes accurate push of marketing information, and continuously optimizes the model through feedback data to improve the accuracy of marketing score. It is the final implementation step of the whole system, and the feedback data is used for model optimization of the user intention recognition agent or the score modeling agent.

[0057] In order to verify the effectiveness of the scoring system proposed in the present application, in this embodiment, ticket recommendation in the travel industry is taken as an example to demonstrate, and the interaction content between the customer and the system is as follows: Dialogue robot: "Do you have any needs now? Do you need to promote for which scene" Customer: "We want to do marketing service for Beijing-Shanghai flight tickets, please help me find users who have demand for the flight tickets of this route" Dialogue robot: "OK, do you want to promote for which specific route" Customer: "This time, we want to promote Shenzhen-Chongqing, Beijing-Chengdu, and Chengdu-Guangzhou routes" Dialogue robot: "OK, do you have any restrictions on the age and region of the user group for promotion" Customer: "18-55 years old is required".

[0058] Through the above interaction information, the intent recognition agent performs user intent recognition, and delivers the recognized scene result to a target scene adaptive user score modeling module to splice the user general features and personalized features in the scene, so as to obtain a score of each user in the scene. Then, users who score 70 (the score threshold is different in different scenes) are screened out and reached through short messages.

[0059] To evaluate the performance, the Click-Through Rate (CTR) and User Conversion Rate (UCR) of the user are used to evaluate the target scene adaptive user score modeling agent. Table 2 shows the user CRT and UCR results in different routes in the ticket recommendation scene of the travel industry. It can be found from the experimental results that the method proposed in the application shows excellent user click rate and conversion rate in most routes, effectively verifying the effectiveness of the intent recognition module and the adaptive scene user score module in the application.

[0060] Table 2. User results in the travel-ticket recommendation scene

[0061] The scene adaptive scoring system based on double-agent cooperation provides a complete system for customers, which provides reliable scoring reference for customer user scoring needs in different scenes under the premise of ensuring data security, and accurately obtains user intent and adaptive scene matching with the help of large language model multi-agent technology. At the same time, thanks to the model that can adapt to different scoring scenes, it also maximizes the reduction of energy consumption. Finally, the user is reached through various touch channels such as email, short message, voice, mailbox, 5G message number and enterprise business card, and user feedback is collected based on user touch to optimize the data model.

[0062] The embodiment of the application also provides a scene adaptive scoring system 400 based on double-agent cooperation, comprising: A user intent recognition agent 401 configured to perform the function of user intent recognition (S103) in the foregoing method embodiment. Specifically, it is responsible for interacting with customers, and identifying the target marketing scene and related constraints through an internal hybrid model architecture (lightweight LLM + rule engine, or heavy LLM).

[0063] A score modeling agent 402 configured to perform the function of target scene adaptive user score modeling (S104) in the foregoing method embodiment. Specifically, it receives scene information from the user intent recognition agent, automatically generates a data query, obtains and fuses personalized and general data, and calculates the user score using an internal scoring model.

[0064] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. The present application can have various changes and modifications for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A scene adaptive scoring method based on dual-agent collaboration, characterized in that, include: The intelligent agent that recognizes user intent receives and identifies customer needs that include the target marketing scenario. The user intent recognition agent transmits the identified target marketing scenario information to the scoring modeling agent; The scoring modeling agent adaptively filters and retrieves personalized user data related to the target marketing scenario from a database containing operator data, based on the target marketing scenario. The rating modeling agent integrates the personalized user data with general user data, and performs rating modeling on the integrated data to obtain the rating of each user in the target marketing scenario. Based on the rating, information is sent to target users whose ratings meet preset reach conditions, feedback data from the target users is collected, and the feedback data is used to optimize the model of the user intent recognition agent or the rating modeling agent.

2. The method according to claim 1, characterized in that, The process of receiving and identifying customer needs containing target marketing scenarios through a user intent recognition intelligent agent includes: The clarity of intent regarding the customer's needs is assessed using a primary language model. If the clarity of intent meets the preset clarity conditions, then the customer needs are identified through a rule-based decision reasoning engine to obtain the target marketing scenario; If the clarity of intent does not meet the preset clarity condition, the customer needs are identified through a second major language model to obtain the target marketing scenario; wherein the number of parameters of the second major language model is greater than the number of parameters of the first major language model.

3. The method according to claim 1, characterized in that, The scoring modeling agent adaptively filters and acquires personalized user data related to the target marketing scenario based on the target marketing scenario, including: The scoring modeling agent, based on a pre-trained language model, parses the target marketing scenario information to generate a database query statement for retrieving the personalized user data; and Execute the database query statement to retrieve the personalized user data from the database.

4. The method according to claim 1, characterized in that, The operator data includes at least one of the following: B-domain data, O-domain data, and M-domain data of individual users; The general user data includes at least one of: basic user attribute information and communication behavior information; The personalized user data includes: user online behavior data, interest preference data, or travel trajectory data corresponding to the target marketing scenario.

5. The method according to claim 1, characterized in that, Before the rating modeling agent acquires the personalized user data, the method further includes: The data in the database is preprocessed, including: missing value imputation based on the random forest algorithm, feature binning based on the chi-square binning method, and weight of evidence (WOE) binning.

6. The method according to claim 1, characterized in that, The rating modeling agent performs user rating modeling based on the fused data, including: The fused data is used as input to a multilayer perceptron model; The multilayer perceptron model is used to train the system and obtain the scores of each user in the target marketing scenario.

7. A scene adaptive scoring system based on dual-agent collaboration, characterized in that, include: The user intent recognition agent is used to receive and identify customer needs that include target marketing scenarios, and to pass the identified target marketing scenario information to the scoring modeling agent; The scoring modeling agent is used to adaptively filter and obtain personalized user data related to the target marketing scenario from a database containing operator data, based on the target marketing scenario. The personalized user data is then integrated with the general user data, and a user rating model is performed based on the integrated data to obtain the rating of each user in the target marketing scenario.

8. The scene adaptive scoring system based on dual-agent collaboration according to claim 7, characterized in that, The method for receiving and identifying customer needs that include target marketing scenarios includes: The clarity of intent regarding the customer's needs is assessed using a primary language model. If the clarity of intent meets the preset clarity conditions, then the customer needs are identified through a rule-based decision reasoning engine to obtain the target marketing scenario; If the clarity of intent does not meet the preset clarity condition, the customer needs are identified through a second major language model to obtain the target marketing scenario; wherein the number of parameters of the second major language model is greater than the number of parameters of the first major language model.

9. The scene adaptive scoring system based on dual-agent collaboration according to claim 7, characterized in that, The method for adaptively filtering and obtaining personalized user data related to the target marketing scenario from a database containing operator data, based on the target marketing scenario, includes: The scoring modeling agent, based on a pre-trained language model, parses the target marketing scenario information to generate a database query statement for retrieving the personalized user data; and Execute the database query statement to retrieve the personalized user data from the database.

10. The scene adaptive scoring system based on dual-agent collaboration according to claim 7, characterized in that, The user rating modeling based on the fused data yields a rating for each user within the target marketing scenario, including: The fused data is used as input to a multilayer perceptron model; The multilayer perceptron model is used to train the system and obtain the scores of each user in the target marketing scenario.

Images