Search

CN-121998769-A - Short-risk dynamic pricing method based on AI large model

CN121998769ACN 121998769 ACN121998769 ACN 121998769ACN-121998769-A

Abstract

The application discloses a short risk dynamic pricing method based on an AI large model, which comprises the steps of S1, constructing a multi-mode risk dynamic sensing framework, outputting a uniform dynamic risk vector with a fixed length, S2, establishing a real-time risk pricing self-adaptive mechanism based on the uniform dynamic risk vector with the fixed length output in S1, realizing real-time pricing decisions of cross risk, S3, designing an intelligent strategy generation and conflict resolution system, optimizing the real-time pricing decisions of the cross risk in S2, ensuring that the pricing decisions realize the optimal balance of market competitiveness and profit targets on the premise of meeting precision calculation constraint and supervision requirements, S4, constructing a complete decision tracing mechanism by adopting a multi-mode interpretable artificial intelligence technology, generating attribution reports based on S1-S3 output data, and realizing the conversion of pricing decisions from black boxes to white boxes. The technology provides a new generation of intelligent pricing solution for short risk service through innovative application of the AI large model.

Inventors

  • XUAN WUJING
  • LIU TIANYI
  • YANG LIU
  • ZHENG MINGMING

Assignees

  • 中国人寿保险股份有限公司新疆维吾尔自治区分公司

Dates

Publication Date
20260508
Application Date
20251211

Claims (10)

  1. 1. A short-risk dynamic pricing method based on an AI large model is characterized by comprising the following steps: S1, constructing a multi-mode risk dynamic sensing framework, and outputting a uniform dynamic risk vector with a fixed length; S2, based on the uniform dynamic risk vector with fixed length output in the S1, a real-time risk pricing self-adaptive mechanism is established, and real-time pricing decision crossing dangerous seeds is realized; S3, designing an intelligent strategy generation and conflict resolution system, optimizing the real-time pricing decision of the cross-risk in the S2, and ensuring that the pricing decision achieves the optimal balance of market competitiveness and profitability targets on the premise of meeting the requirements of fine calculation constraint and supervision; S4, constructing a complete decision tracing mechanism by adopting a multi-mode interpretable artificial intelligence technology, and generating an attribution report based on the S1-S3 output data to realize the conversion of a pricing decision from a black box to a white box.
  2. 2. The AI-large-model-based short risk dynamic pricing method according to claim 1, wherein the constructing the multi-modal risk dynamic perception architecture in S1 comprises: S1.1, constructing a multi-mode data access and preprocessing module, which specifically comprises an unstructured data processing sub-module and a structured data processing sub-module, wherein the unstructured data processing sub-module is used for processing medical texts, vehicle damage images and voice records, text data is subjected to automatic cleaning, denoising and standardization processing, image data is subjected to normalization, compression and data enhancement processing, voice data is converted into text through an automatic voice recognition technology, and the structured data processing sub-module is used for processing policy information, historical claim records, credit scores and real-time sensor data, and carrying out data verification, cleaning and statistical feature generation; S1.2, constructing an AI large model feature coding module, which specifically comprises a text feature coder, a visual feature extractor and a structured data embedder; the text feature encoder adopts a large language model after corpus training in the insurance field, performs fine-grained entity recognition, emotion analysis and medical knowledge graph linking, extracts accident responsibility ambiguity, loss degree implications, medical diagnosis relevance and other risk dimensions from text data, integrates a visual feature extractor in the large visual model after vehicle injury and medical image data set training, recognizes an impaired part through a target detection network, utilizes an image segmentation and regression model to quantify the injury degree, and generates a visual injury severity index and an image risk score; S1.3, a dynamic risk feature fusion module is constructed, and the dynamic risk feature fusion module specifically comprises a cross-modal feature alignment sub-module and a depth nonlinear fusion sub-module, wherein the cross-modal feature alignment sub-module adopts a cross attention mechanism to realize bidirectional information interaction among multi-modal features, alignment and complementation of feature semantic layers are realized by calculating correlation weights among different modal features, when consistency among different modal features is detected, an inconsistent risk sign is automatically generated when a text description and an image feature have obvious differences, the depth nonlinear fusion sub-module comprises a multi-layer fully connected neural network, nonlinear transformation and information compression are carried out on the aligned high-dimensional feature vectors, and finally, uniform dynamic risk vectors with fixed lengths are output through complex interaction among different features learned by a multi-layer perceptron.
  3. 3. The AI-large-model-based short risk dynamic pricing method according to claim 2, wherein the constructing the multi-modal risk dynamic perception architecture in S1 further comprises: s1.4, a stream updating and real-time reasoning module is constructed, which specifically comprises an event driving processing sub-module and an increment reasoning sub-module, wherein the event driving processing sub-module monitors a data input port, when new driving behavior data, image uploading or voice recording is received, a feature recalculation flow is automatically triggered, the increment reasoning module adopts a lightweight updating strategy to only run a corresponding feature encoder for the new data, and the new feature is quickly combined with a cached historical fusion vector through an increment fusion network to generate an updated dynamic risk vector.
  4. 4. The AI-large-model-based short-risk dynamic pricing method of claim 3, wherein the step S1.4 further comprises a version management sub-module for recording and tracking the feature vector update process to ensure traceability of the evolution of the risk portraits when constructing the streaming update and real-time reasoning module.
  5. 5. The AI-large-model-based short risk dynamic pricing method according to claim 2, wherein the establishing of the real-time risk pricing adaptive mechanism in S2 comprises: S2.1, constructing a multi-source flow data acquisition module, and receiving user behavior data of various dangerous types in real time through a distributed message middleware, wherein the vehicle risk field acquires driving behavior sequences transmitted by OBD equipment, the health risk field acquires physiological index time sequence data monitored by wearable equipment, the unexpected risk field acquires user movement track and activity mode data, the life risk field receives periodic physical examination and life style monitoring data, the data are preprocessed and input into a dynamic characteristic extraction engine based on a large language model, the engine utilizes a pre-trained language model to perform unified characterization learning on the multi-mode time sequence data, a long-term dependence is captured through an attention mechanism, and an online fine tuning technology is adopted to enable the model to continuously adapt to the change of data distribution; S2.2, constructing a dynamic risk scoring model, adopting a large language model as a core framework of sequence modeling, and processing streaming data by utilizing strong context understanding capability, wherein the method specifically comprises the following steps: S2.21, performing feature extraction and risk pattern recognition on the streaming data of various dangerous types by using a time sequence encoder based on a transducer; s2.22, converting a real-time risk scoring task into a text generation problem of a language model by adopting a prompt learning technology, and realizing risk assessment by dynamically constructing a prompt template; S2.23, utilizing a thinking chain technology to enable the model to show an reasoning process from the original data to risk scoring; s2.24, applying a parameter efficient fine adjustment method, such as LoRA, to realize quick online adaptation to a basic large model; S2.3, constructing a multi-task risk assessment framework based on a large language model, maintaining a dynamically updated risk knowledge base by adopting a memory-enhanced large model framework, storing a recent risk assessment mode and emergency risk information, simultaneously, realizing deep analysis of complex risk scenes and assessment of risk infection effects of risk crossing through reasoning capacity of the large model, and finally realizing real-time pricing decision of risk crossing.
  6. 6. The AI-large-model-based short-risk dynamic pricing method of claim 2, wherein the constructing the dynamic risk scoring model in S2.2 further comprises: S2.25, designing an event-triggered model updating mechanism, and automatically starting an incremental learning process when an abnormal mode is detected, so that updating of the risk assessment model is ensured to be completed in second-level time.
  7. 7. The AI-large-model-based short-risk dynamic pricing method of claim 5, wherein designing the intelligent policy generation and conflict resolution system in S3 comprises: S3.1, a large language model subjected to fine adjustment of knowledge in the fine calculation field is adopted as a decision core to construct a strategy generation module, a multi-head attention mechanism based on a Transformer framework is constructed aiming at four short risk services of vehicle risk, health risk, accident risk and life risk, and specific risk feature vectors and pricing factors of various risks are respectively processed; s3.2, constructing a conflict resolution module by adopting a multi-objective constraint optimization framework based on a large model, wherein the method specifically comprises the following steps of: s3.21, converting the accurate constraint condition into a computable vector representation by utilizing the semantic understanding capability of the large language model; S3.22, analyzing and adapting to the change of the constraint condition in real time through the context capacity of the large model; s3.23, designing a reward shaping mechanism based on a large model, and automatically adjusting the weight of a reward function of each dangerous pricing strategy by analyzing market feedback and profit effects of historical pricing decisions; S3.24, aiming at strategy conflict among multiple dangerous seeds, adopting course learning strategies guided by a large model, and gradually optimizing pricing strategy cooperation of various dangerous seeds in a sequence-to-sequence modeling mode; S3.3, a compliance checking module is built by the real-time supervision compliance checking system based on the large language model, the deep understanding capability of the large language model to supervision files and company policies is utilized, consistency of the generated strategies and supervision requirements is automatically detected through text implication and semantic similarity calculation, the method specifically comprises the steps of using few-shot learning capability of the large language model to quickly adapt to supervision policy updating, displaying a logic process of compliance judgment through a chain of thought technology, and generating an interpretable compliance report.
  8. 8. The AI-large-model-based short-risk dynamic pricing method of claim 7, wherein the step S3.3 of constructing the compliance verification module further comprises a large-model-based multi-granularity feedback mechanism, wherein real-time optimization suggestions are provided in a full process of policy generation, conflict resolution and compliance checking, and the pricing policy is ensured to achieve the optimal balance of market competitiveness and profitability targets on the premise of meeting fine-calculation constraints and regulatory requirements.
  9. 9. The AI-large-model-based short-risk dynamic pricing method of claim 7, wherein constructing the complete decision-traceback mechanism in S4 comprises: S4.1, establishing a decision logic analysis module based on a thinking chain technology, and carrying out semantic reconstruction on the whole pricing decision process by utilizing a large language model subjected to fine adjustment of knowledge in the field of fine calculation, wherein the method specifically comprises the following steps: S4.11, identifying key risk factors with the largest influence on final pricing in input features by adopting an attention weight visualization technology; s4.12, converting a complex numerical calculation process into a decision logic chain of natural language description by utilizing the sequence-to-sequence generation capability of the large model; S4.13, constructing a domain-specific interpretation template by prompting engineering technology, and ensuring that the generated attribution report accords with the accurate calculation professional specification; s4.2, constructing a multi-granularity interpretable output architecture of an attribution report generation layer, which specifically comprises the following steps: s4.21, calculating the relative importance of each risk factor in pricing decision through a hierarchical attention mechanism based on an attribution algorithm of a transducer; s4.22, automatically generating a complete report containing risk factor weight distribution, decision logic paths and accurate calculation basis by utilizing text generation capacity of the large model; s4.23, adopting a knowledge graph enhancement technology to perform association analysis on the pricing decision, the insurance clauses, the supervision requirements and the fine calculation principle, and providing authoritative decision basis reference; S4.3, constructing a multi-version report generation mechanism based on a large model so as to achieve the dual purpose of supervision compliance and customer communication, and enabling the large model to generate interpretation reports of corresponding detail degree and expression mode according to different audiences through a field self-adaptive fine tuning technology, wherein the supervision-oriented report stands out the compliance and statistical significance, and the customer-oriented report stands out popular and easy to understand.
  10. 10. The AI-large-model-based short-risk dynamic pricing method of claim 9, wherein the step S4.3 of constructing a large-model-based multi-version report generation mechanism further comprises a real-time compliance checking function, and compliance verification is performed on the generated interpretation content by using the large model to ensure that all expressions meet regulatory requirements.

Description

Short-risk dynamic pricing method based on AI large model Technical Field The application relates to the technical field of intersection of artificial intelligence and insurance pricing, in particular to a short-risk dynamic pricing method based on an AI large model. Background Short-risk dynamic pricing technology is undergoing spanning development from experience-driven static determination to AI-energized real-time accurate pricing, and the core evolution path encloses The three main technical lines of risk quantitative modeling, intelligent algorithm optimization and multi-source data fusion are developed, and the technical iterative logic and the existing bottleneck are as follows: 1. Static statistical modeling phase, namely risk quantification based on experience rules. Early short risk pricing takes a fine calculation rule and a basic statistical model as cores, and the core aims to realize preliminary classification pricing of risks. In 2010, smith J.A multiple linear regression pricing model is provided, and the basic rate is determined by limited static factors such as age, occupation and the like, but the model cannot capture the dynamic change and differentiation characteristics of individual risks. In 2013, zhang Wei adopts a Generalized Linear Model (GLM) to optimize risk factor weight, so that pricing stability is improved, but the model is limited by a preset statistical distribution assumption, and unstructured data such as texts, behavior tracks and the like are difficult to adapt. In 2015, li Ming fuses historical claim settlement data by means of a Bayesian network, so that the logic of risk quantification is enhanced, but the characteristic engineering of manual leading is relied on, and the updating period of a pricing scheme is as long as half a year to one year, so that market fluctuation and burst risk cannot be responded. In 2017, wang Hong proposed an interval group pricing model, and reduced the calculation cost by simplifying risk classification, but the pricing fairness was insufficient due to neglect of individual risk heterogeneity, and it was difficult to meet the personalized demand. 2. And a machine learning optimization stage, namely multi-factor accurate risk identification. Along with the expansion of data dimension, a machine learning algorithm becomes a core tool for breaking through the limitation of the traditional model, and the pricing accuracy and factor mining capability are improved in a focusing mode. In 2018, yu Jiang integrates customer base information and claim settlement data by adopting a random forest algorithm, so that the pricing error is reduced by 8.3% compared with a traditional model, but feature redundancy is easy to occur in a high-dimensional sparse data scene, and the model efficiency is influenced. In 2020, chen Yi strengthens feature interaction relation capture through a gradient lifting tree (XGBoost), improves the fineness of risk identification, but the performance of Chen Yi is highly dependent on high-quality labeling data and is sensitive to data loss and noise. In 2021, the Robert team combines a Support Vector Machine (SVM) with a rule engine to achieve fast dynamic adjustment of short risk pricing, but suffers from insufficient interpretability due to the "black box" nature of the model, facing regulatory compliance pressures. 2022, huang Bangju and the like propose LightGBM multi-factor pricing models, the pricing efficiency of small short risks is obviously improved, but data sources are limited to an enterprise intranet, the real-time adaptation capability to sudden scenes such as natural disasters, sudden market changes and the like is lacking, 2024, wei Yongjiang, xia Anyu discloses an insurance product pricing and risk assessment method, device and storage medium based on lightGBM and a neural network aiming at specific tasks, the patent mainly relies on structured data and preset rules to carry out single risk assessment data, unstructured data (such as medical texts, vehicle damage images, voice records and the like) cannot be understood deeply and semantically, the model update period is usually several weeks or even months, and real dynamic pricing is difficult to realize. 3. And AI large model fusion stage, namely multi-source data dynamic risk perception. In recent years, the AI large model becomes a core technical direction of short-risk dynamic pricing by virtue of strong heterogeneous data processing and feature learning capabilities. In 2023, the public security team integrates the wearable device health data based on the open source large model, and pushes out a healthy short risk pricing scheme with steps deducting the premium, but the model suitability is limited to a single scene, and the generalization capability of cross-risk (such as aviation risk and accident risk) is insufficient. In 2024, the safe risk team analyzes driving behavior data collected by vehicle-mounted equipment by using a Large Language Mod