Search

CN-121981826-A - User portrait construction method and system based on financial consultation behaviors

CN121981826ACN 121981826 ACN121981826 ACN 121981826ACN-121981826-A

Abstract

The invention discloses a user portrait construction method and a system based on financial consultation behaviors, wherein the user portrait construction method comprises the steps of obtaining a user consultation log and generating a multidimensional feature vector comprising a semantic feature vector and a behavior feature vector; the method comprises the steps of carrying out cluster analysis on multidimensional feature vectors to obtain a plurality of user prototype clusters, constructing a probability distribution model of a financial user prototype based on the number of samples in the user prototype clusters, sampling the probability distribution model to obtain a target user prototype, generating a primary user portrait according to the target user prototype by using a large language model, checking the primary user portrait according to preset rules, and correcting the primary user portrait according to check results to obtain the target user portrait. According to the invention, the semantic meaning and the behavior characteristic depth of the user consultation log are fused, the potential preference of the user is captured, the user portrait is corrected through the preset rule, the model illusion is effectively restrained, and the restoration degree and the credibility of the user behavior simulation are improved.

Inventors

  • LI JUAN
  • LONG YUN

Assignees

  • 北京工业大学

Dates

Publication Date
20260505
Application Date
20260120

Claims (10)

  1. 1. A user portrait construction method based on financial consultation behavior is characterized by comprising the following steps: acquiring a user consultation log, and generating a multidimensional feature vector based on the user consultation log, wherein the multidimensional feature vector comprises a semantic feature vector and a behavior feature vector; performing cluster analysis on the multidimensional feature vectors to obtain a plurality of user prototype clusters, and constructing a probability distribution model of a financial user prototype based on the number of samples in the user prototype clusters; Sampling the probability distribution model to obtain a target user prototype, and generating a primary user portrait according to the target user prototype by using a large language model, wherein the primary user portrait comprises a user attribute tag, an asset annual fluctuation rate and an investment target; And verifying the primary user portrait according to a preset rule, and correcting the primary user portrait according to a verification result to obtain a target user portrait, wherein the preset rule comprises a risk matching rule and a cognitive matching rule.
  2. 2. The user portrayal construction method of claim 1, wherein the generating a multi-dimensional feature vector based on the user consultation log comprises: Preprocessing the user consultation log, wherein the preprocessing comprises the steps of removing an invalid log, unifying text formats and removing redundant information; carrying out semantic coding on the preprocessed user consultation log by utilizing a pre-trained financial language model to generate the semantic feature vector, wherein the semantic feature vector is used for reflecting a financial topic focused by a user; performing behavior feature extraction on the preprocessed user consultation log based on the behavior statistical feature indexes to generate behavior feature vectors, wherein the behavior feature vectors are used for representing the user operation style and the user cognition level; And splicing the semantic feature vector and the behavior feature vector to obtain the multidimensional feature vector.
  3. 3. The user portrayal construction method according to claim 2, characterized in that the behavior statistical characteristic index comprises a professional degree, a rigor degree and an emotion value; the behavior feature extraction is performed on the preprocessed user consultation log based on the behavior statistical feature index to generate a behavior feature vector, and the behavior feature vector comprises: Extracting a question text of the preprocessed user consultation log; Calculating the specialty based on the occurrence frequency or occurrence density of the financial entity codes and the professional terms in the questioning text; Calculating the stringency by analyzing the misprinted word rate, punctuation standardability and sentence integrity in the question text; Carrying out emotion analysis on the questioning text through an emotion analysis tool so as to calculate the emotion value; and generating the behavior characteristic vector according to the specialty, the strictness and the emotion value.
  4. 4. The user portrait construction method according to claim 1, wherein the performing cluster analysis on the multidimensional feature vector to obtain a plurality of user prototype clusters includes: clustering the multidimensional feature vectors by adopting a K-means++ algorithm; Evaluating the clustering effects corresponding to the different clustering clusters through an elbow rule or a contour coefficient, and determining the number of clusters and the user prototype clusters according to the clustering effects; User attribute tags are determined based on the center vector of the user prototype cluster, the user attribute tags including risk preference type, field of interest category, expertise level, and operational style characteristics.
  5. 5. The user representation construction method of claim 4, wherein constructing a probability distribution model of a financial user prototype based on the number of samples in the user prototype cluster comprises: acquiring the number of user consultation logs contained in each user prototype cluster as the sample number; Calculating the duty ratio weight of the sample number in the total user consultation log number; and establishing the probability distribution model based on the duty ratio weight, wherein the probability distribution model is used for representing the distribution proportion of the user attribute tags.
  6. 6. The user representation construction method of claim 1, wherein said generating a primary user representation from said target user prototype using a large language model comprises: generating natural language prompt words according to the feature description of the target user prototype; and inputting the prompt word into the large language model, inquiring user holding information corresponding to the target user prototype from a financial database by utilizing the large language model based on a retrieval enhancement generation strategy, and generating the primary user portrait based on the target user prototype and the user holding information.
  7. 7. The user portrayal construction method of claim 6, wherein the querying the user's cabin holding information corresponding to the target user prototype from a financial database based on a search enhancement generation policy using the large language model comprises: Generating a structured query instruction according to the attention field category and the risk preference type of the target user prototype; retrieving product information in the financial database by utilizing the query instruction, wherein the product information comprises product entity data, historical market data and user holding data; And generating the user holding information based on the product information.
  8. 8. The user representation construction method according to claim 1, wherein the verifying the primary user representation according to a preset rule comprises: acquiring risk preference attributes and investment experiences of the primary user portraits, wherein the risk preference attributes comprise conservation and aggressive types, and the investment experiences comprise experience sufficiency and investment novices; If the risk preference attribute is conservative, generating a verification result of successful risk verification under the condition that the asset annual fluctuation rate is smaller than or equal to a first preset threshold value, and generating a verification result of failed risk verification under the condition that the asset annual fluctuation rate is larger than the first preset threshold value; If the risk preference attribute is aggressive, generating a verification result of successful risk verification under the condition that the investment target contains a preset high-risk product, and generating a verification result of failed risk verification under the condition that the investment target does not contain the preset high-risk product; If the investment experience is an investment novice, generating a verification result of successful cognitive verification under the condition that the professional degree of the behavior feature vector is smaller than or equal to a second preset threshold value, and generating a verification result of failed cognitive verification under the condition that the professional degree of the behavior feature vector is larger than the second preset threshold value; If the investment experience is sufficient, generating a verification result of the success of the cognitive verification under the condition that the investment target contains the preset complex financial product, and generating a verification result of the failure of the cognitive verification under the condition that the investment target does not contain the preset complex financial product.
  9. 9. The user representation construction method according to claim 8, wherein correcting the primary user representation based on the verification result to obtain a target user representation comprises: judging evidence sources of attribute information causing conflict under the condition that the verification result is that risk verification fails or cognition verification fails; If the evidence source is the purchasing behavior evidence in the user consultation log, correcting the risk preference label; and if the evidence source is the information generated by the large language model, correcting the investment target.
  10. 10. A user portrayal construction system based on financial consulting actions for executing a user portrayal construction method based on financial consulting actions as claimed in any one of claims 1-9, comprising: the vector generation module is configured to acquire a user consultation log and generate a multidimensional feature vector based on the user consultation log, wherein the multidimensional feature vector comprises a semantic feature vector and a behavior feature vector; the cluster analysis module is configured to perform cluster analysis on the multidimensional feature vectors to obtain a plurality of user prototype clusters, and construct a probability distribution model of a financial user prototype based on the number of samples in the user prototype clusters; The portrait generation module is configured to sample the probability distribution model to obtain a target user prototype, and generate a primary user portrait according to the target user prototype by using a large language model, wherein the primary user portrait comprises a user attribute tag, an asset annual fluctuation rate and an investment target; and the portrait modification module is configured to verify the primary user portrait according to preset rules and modify the primary user portrait according to a verification result to obtain a target user portrait, wherein the preset rules comprise risk matching rules and cognitive matching rules.

Description

User portrait construction method and system based on financial consultation behaviors Technical Field The invention relates to the technical field of financial user portraits, in particular to a user portrayal construction method and system based on financial consultation behaviors. Background In the field of financial information consultation, the construction of high-fidelity and multidimensional user portraits is a core foundation for realizing intelligent casting and consultation accurate service, effective deduction of market emotion and personalized financial service landing. With the rapid development of artificial intelligence technology, particularly, large Language Models (LLM) show excellent capability in the field of natural language understanding and generation, the intelligent agent based on LLM is widely applied to social science and user behavior simulation research, and a user portrait is taken as a core reference basis of intelligent agent behavior logic, and a construction method of the intelligent agent based on LLM forms related technical schemes in the fields of general simulation and verticality. In the prior art, text portraits containing information such as names, professions, character features and the like are set for an agent through prompt word engineering, simulation of macroscopic social behaviors of human beings is realized by means of social common knowledge in a LLM pre-training knowledge base, a scene is simulated in a vertical field such as a recommendation system, a structural portraits module containing demographic information and preferences in a specific field is designed by a related scheme, and the agent is initialized in a real data alignment or regular sampling mode so as to simulate specific business behaviors such as browsing, clicking and the like. However, the above method tends to output samples of "typical users" conforming to the statistical average rule, while the real financial market user distribution has significant long-tail characteristics, including a large number of extreme risk preferences, specific investment beliefs and other "edge users", and it is difficult to fully cover such long-tail samples critical to stress test and comprehensive evaluation only by means of LLM probability prediction, so that there is a problem that the diversity of the generated samples is insufficient. In addition, the method focuses on semantic analysis of the user consultation log, namely only focuses on what the user's questions, but ignores relevant characteristics of how to question which are implied by the professional degree and the operation style of the user in the financial consultation scene, such as insufficient mining of behavior statistical characteristics of code mention density, professional term abbreviation use habit, emotional misspelling and the like, so that the reduction of the investment style of the user is not stereoscopic enough, and the reality of behavior simulation is affected. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a user portrait construction method and a system based on financial consultation behaviors, which aim to improve the authenticity and the credibility of the simulation of the financial consultation behaviors and generate financial user portraits with high credibility. The invention discloses a user portrait construction method based on financial consultation behaviors, which comprises the following steps: acquiring a user consultation log, and generating a multidimensional feature vector based on the user consultation log, wherein the multidimensional feature vector comprises a semantic feature vector and a behavior feature vector; Performing cluster analysis on the multidimensional feature vectors to obtain a plurality of user prototype clusters, and constructing a probability distribution model of the financial user prototype based on the number of samples in the user prototype clusters; sampling the probability distribution model to obtain a target user prototype, and generating a primary user portrait according to the target user prototype by using the large language model, wherein the primary user portrait comprises a user attribute tag, an asset annual fluctuation rate and an investment target; and verifying the primary user portrait according to a preset rule, and correcting the primary user portrait according to a verification result to obtain a target user portrait, wherein the preset rule comprises a risk matching rule and a cognitive matching rule. Preferably, generating the multidimensional feature vector based on the user consultation log includes: preprocessing the user consultation log, wherein the preprocessing comprises the steps of removing invalid logs, unifying text formats and removing redundant information; Semantic coding is carried out on the preprocessed user consultation log by utilizing a pre-trained financial language model so as to generate a s