Search

CN-121981806-A - Self-adaptive intelligent matching recommendation method and system for financial products

CN121981806ACN 121981806 ACN121981806 ACN 121981806ACN-121981806-A

Abstract

The invention relates to a self-adaptive intelligent matching recommendation method and system for financial products, which relate to the technical field of financial product matching, and are characterized in that multi-source user data are obtained, real-time streaming processing is carried out on the multi-source user data, dynamic micro-behavior characteristics and static attribute characteristics of a user are extracted, the dynamic micro-behavior characteristics, the static attribute characteristics and external macro-environment factors are input into a pre-built multi-target game recommendation model, the multi-target game recommendation model is used for building a non-cooperative game framework among a user agent, a mechanism agent and a supervision agent based on a game theory, and initial recommendation combination considering three-party interest requirements is generated by solving Nash equilibrium.

Inventors

  • HUANG XIAOBO

Assignees

  • 南通久哲科技发展有限公司

Dates

Publication Date
20260505
Application Date
20260408

Claims (10)

  1. 1. The self-adaptive intelligent matching recommendation method for the financial products is characterized by comprising the steps of acquiring multi-source user data, performing real-time streaming on the multi-source user data, and extracting dynamic micro-behavior characteristics and static attribute characteristics of users; inputting the dynamic micro-behavioral characteristics, the static attribute characteristics and the external macro environment factors into a pre-constructed multi-target game recommendation model; The multi-target game recommendation model builds a non-cooperative game framework among a user agent, an organization agent and a supervision agent based on a game theory, and generates an initial recommendation combination considering three-party interest appeal by solving Nash equilibrium; inputting the initial recommendation combination into an interpretability engine constructed based on causal inference, and generating a counterfactual interpretation text for the recommendation combination; pushing the recommendation combination and the corresponding counterfactual interpretation text to a user side, collecting user feedback, and updating parameters of the multi-target game recommendation model in real time by using an online learning mechanism.
  2. 2. The method of claim 1, wherein the dynamic micro-behavior features include one or more of a user's mouse-over time, a page scroll speed, a repeated comparison path, and an input frame hesitation time of the interactive interface, and the dynamic micro-behavior features are extracted in real time through a sliding time window based on a stream computing framework.
  3. 3. The method of claim 2, wherein the external macro-environmental factors comprise one or more of real-time interest rate curves, consumer price indices, and industry policy event vectorized representations, and the multi-objective game recommendation model dynamically adjusts revenue functions of the user agent and the institution agent by using the macro-environmental factors as uncontrollable environmental variables.
  4. 4. The method for adaptively matching and recommending financial products according to claim 3, wherein said generating a counterfactual interpretation text for said recommended combination comprises: Positioning an association path between the current portrait of the user and the recommended product based on the financial knowledge graph; carrying out anti-facts hypothesis modification on key feature variables in the user portrait through a causal intervention technology; The multi-target game recommendation model is driven to recalculate based on the modified characteristic variable, and a comparison recommendation result is generated; And naturally linguistic the comparison recommendation result into a fixed sentence pattern by using a large language model to output.
  5. 5. The intelligent matching recommendation method for self-adaptive financial products of claim 3, wherein after an initial recommendation combination is generated, the initial recommendation combination is input into a compliance verification module based on a generated countermeasure network, the compliance verification module comprises a discriminator for judging whether the recommendation combination accords with a preset supervision rule and an in-institution control strategy, and the discrimination result is fed back to a multi-objective game recommendation model as a part of a loss function to perform countermeasure training.
  6. 6. An adaptive intelligent matching recommendation system for financial products, which is applied to the adaptive intelligent matching recommendation method for financial products according to any one of claims 1-4, is characterized by comprising a data acquisition and processing module, a data processing module and a data processing module, wherein the data acquisition and processing module is used for acquiring multi-source user data and performing real-time stream processing to extract dynamic micro-behavior characteristics and static attribute characteristics of users; The multi-target game recommendation engine is internally provided with a user agent, an organization agent and a supervision agent which are constructed based on a game theory and are used for receiving the characteristics and external macroscopic environment factors and generating initial recommendation combinations by solving Nash equilibrium; the interpretive engine is connected with the multi-target game recommendation engine, and is internally provided with a financial knowledge graph and a causal inference model, and is used for generating a counterfactual interpretation text for the recommendation combination; and the online learning feedback module is used for collecting user feedback and driving the multi-target game recommendation engine to update parameters online.
  7. 7. The intelligent matching recommendation system for adaptive financial products of claim 6, further comprising a privacy computation module disposed at a boundary of the data provider for performing encrypted parameter exchange and gradient update using a longitudinal federal learning protocol to ensure that the original user data does not go out of domain when training the multi-objective game recommendation engine in conjunction with multiple data sources.
  8. 8. The intelligent adaptive financial product matching recommendation system of claim 6 wherein said multi-objective gaming recommendation engine further comprises a lifecycle adapter coupled to an external payment system for obtaining future cash flow forecast data for a user and performing a spatiotemporal matching optimization of a expiration date or an opening date of a product with an expected funds outflow time of the user as a pair of gaming variables when solving for a Nash equilibrium.
  9. 9. The intelligent matching recommendation system for self-adaptive financial products of claim 8, wherein a cognitive deviation correction factor is introduced into a utility function of a user agent constructed in the multi-objective game recommendation engine, and the cognitive deviation correction factor is used for dynamically reducing recommendation weight of a high-risk high-fluidity product when detecting that the historical behavior of the user has a tendency of rising, falling or excessive trading so as to guide the user behavior to return to a long-term average value.
  10. 10. The intelligent matching recommendation system for self-adaptive financial products of claim 9, wherein the data acquisition and processing module comprises a feature self-adaptive weighting unit, the feature self-adaptive weighting unit adopts an online learning algorithm, and calculates contribution degree of each dynamic micro-behavior feature to recommendation accuracy in real time according to user feedback collected by the online learning feedback module, and dynamically adjusts input weight of the feature in the next round of recommendation according to the contribution degree.

Description

Self-adaptive intelligent matching recommendation method and system for financial products Technical Field The invention relates to the technical field of financial product matching, in particular to a self-adaptive intelligent matching recommendation method and system for financial products. Background Financial products are one type of tool that financial institutions design to conduct transactions in the marketplace. It is essentially a contract with legal effectiveness, the core of which is to connect both the supply and demand of funds. With the rapid development of the financial market, the variety of financial products is increasingly abundant, and the selection difficulty facing investors is continuously increased. The traditional financial product recommendation method is mainly based on static matching of user portraits and product labels or based on traditional recommendation algorithms such as collaborative filtering. However, these methods have technical drawbacks. The traditional method generally adopts an offline batch processing mode to update the model, and can not capture microscopic behavior changes of the user in the interaction process in real time, so that the recommendation result is delayed from the current real requirement of the user. Aiming at the problems, the invention provides a self-adaptive intelligent matching recommendation method and system for financial products, which realize compliance and intelligent recommendation. Disclosure of Invention In order to realize intelligent recommendation of financial products, the invention provides a self-adaptive intelligent matching recommendation method and system for financial products. In a first aspect, the invention provides a self-adaptive intelligent matching recommendation method for financial products, which adopts the following technical scheme: an adaptive financial product intelligent matching recommendation method comprises the following steps: acquiring multi-source user data, carrying out real-time streaming processing on the multi-source user data, and extracting dynamic micro-behavior characteristics and static attribute characteristics of a user; inputting the dynamic micro-behavioral characteristics, the static attribute characteristics and the external macro environment factors into a pre-constructed multi-target game recommendation model; The multi-target game recommendation model builds a non-cooperative game framework among a user agent, an organization agent and a supervision agent based on a game theory, and generates an initial recommendation combination considering three-party interest appeal by solving Nash equilibrium; inputting the initial recommendation combination into an interpretability engine constructed based on causal inference, and generating a counterfactual interpretation text for the recommendation combination; pushing the recommendation combination and the corresponding counterfactual interpretation text to a user side, collecting user feedback, and updating parameters of the multi-target game recommendation model in real time by using an online learning mechanism. Optionally, the dynamic micro-behavior feature includes one or more of a mouse-over time length, a page scrolling speed, a repeated comparison path and an input frame hesitation time length of the user on the interactive interface, and the real-time extraction is performed through a sliding time window based on the stream computing frame. Optionally, the external macro-environmental factors comprise one or more of real-time interest rate curves, consumer price indexes and industry policy event vectorized representations, and the multi-target game recommendation model dynamically adjusts the profit functions of the user agent and the institution agent by taking the macro-environmental factors as uncontrollable environmental variables. Optionally, the generating the counterfactual interpretation text for the recommendation combination specifically includes: Positioning an association path between the current portrait of the user and the recommended product based on the financial knowledge graph; a certain key characteristic variable (such as risk bearing capacity) in the user portrait is modified by a causal intervention technology; The multi-target game recommendation model is driven to recalculate based on the modified characteristic variable, and a comparison recommendation result is generated; The comparison recommendation is naturally linguistic into a sentence pattern of "if you's [ feature variable ] is [ anti-fact value ], we will recommend [ comparison product ] for you, and it expects [ benefit/risk ] to output a change [ difference ]". Optionally, after the initial recommended combination is generated, the initial recommended combination is input into a compliance verification module based on a generated countermeasure network, wherein the compliance verification module comprises a discriminator for judging whether the reco