Intelligent investment preference analysis and customer selection system based on large model

Abstract

The invention relates to the field of intelligent investment analysis, and discloses an intelligent investment preference analysis and customer selection system based on a large model, and the system comprises a user classification module, a behavior analysis module, a strategy optimization module and a customer screening module. Performing investment behavior feature clustering on the investment users to obtain classified investment users; performing real-time data updating on the multi-dimensional interaction data to obtain updated data, performing multi-modal space-time diagram convolution on the updated data to obtain a user behavior space-time tensor, and constructing a three-dimensional user portrait of the investment user; analyzing the investment adaptation degree of the classified investment users and the investment market, constructing an initial investment strategy of the classified investment users, and carrying out reinforcement learning on the initial investment strategy to obtain an optimized investment strategy; and performing user screening on the classified investment users to obtain a target investment user. According to the method, the comprehensiveness of intelligent investment preference analysis and the customer selection accuracy can be improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent investment analysis, and more particularly to an intelligent investment preference analysis and customer selection system based on a large model. Background Technology

[0002] Smart investing is an innovative investment approach that leverages artificial intelligence, big data, and other intelligent technologies to deeply analyze massive amounts of financial data and automatically generate investment strategies. Therefore, this technology can be used to analyze the investment preferences of investors and further help the financial market proactively select high-quality users or recommend different products based on different user investment preference types.

[0003] Currently, traditional data analysis techniques are widely used for investment preference analysis and customer selection. These techniques typically involve collecting basic customer information and transaction records, then using statistical analysis methods to process and mine the data, extracting features such as investment amount and type of investment product. Based on these features, a simple customer profile model is built to conduct investment preference analysis and customer selection. However, the financial market is highly complex and subject to real-time changes, making it difficult for traditional data analysis techniques to respond promptly and accurately to these rapid market shifts. This results in incomplete analysis of customer investment preferences and inaccurate customer selection results. Summary of the Invention

[0004] This invention provides an intelligent investment preference analysis and customer selection system based on a large model, the main purpose of which is to improve the comprehensiveness of intelligent investment preference analysis and the accuracy of customer selection.

[0005] To achieve the above objectives, the present invention provides an intelligent investment preference analysis and customer selection system based on a large model, comprising: a user classification module, a behavior analysis module, a strategy optimization module, and a customer screening module; The user classification module is used to collect multi-dimensional interaction data of investment users, perform cross-modal semantic fusion on the multi-dimensional interaction data to obtain customer dynamic behavior vectors, and perform investment behavior feature clustering on the investment users based on the customer dynamic behavior vectors to obtain classified investment users. The behavior analysis module is used to update the multidimensional interaction data in real time to obtain updated data, perform multimodal spatiotemporal graph convolution on the updated data to obtain a user behavior spatiotemporal tensor, query the investment market corresponding to the investment user, collect multi-source heterogeneous data of the investment market, use a trained knowledge augmentation model to construct a market situation embedding vector of the multi-source heterogeneous data, and use the user behavior spatiotemporal tensor and the market situation embedding vector to construct a three-dimensional user profile of the investment user. The strategy optimization module is used to analyze the investment fit between the categorized investment users and the investment market based on the three-dimensional user profile, construct an initial investment strategy for the categorized investment users based on the investment fit, query the historical investment feedback data of the categorized investment users, and perform reinforcement learning on the initial investment strategy based on the historical investment feedback data to obtain an optimized investment strategy. The customer screening module is used to update the customer dynamic behavior vector in real time to obtain an updated dynamic behavior vector. Based on the updated dynamic behavior vector, a trained graph neural network model is used to perform risk propagation analysis on the optimized investment strategy to obtain a risk score. The risk score is used to update the optimized investment strategy in real time to obtain a target strategy. The target strategy is used to screen the classified investment users to obtain target investment users.

[0006] Optionally, the step of performing cross-modal semantic fusion on the multi-dimensional interaction data to obtain a customer dynamic behavior vector includes: Extract transaction data, media data, and voice data from the multidimensional interactive data; Perform temporal convolution on the transaction data to obtain operation sequence features; Identify the semantic features of the media data; Speech emotion recognition is performed on the speech data to obtain intonation features; Multimodal alignment is performed on the operation sequence features, the semantic features, and the intonation features to obtain the customer dynamic behavior vector.

[0007] Optionally, the step of clustering investment behavior features of the investment users based on the customer dynamic behavior vector to obtain classified investment users includes: The customer dynamic behavior vectors are aggregated to obtain a behavior vector set; The investment users are categorized to obtain different user clusters; Based on the number of categories of users clustered together, the initial cluster centers of the investment users are selected from the set of behavior vectors; The vector distance from different data points in the behavior vector set to the initial cluster center is calculated using the following formula: ; in, Represents vector distance. This represents the i-th data point in the behavior vector set. Let represent the j-th initial cluster center, and m represent the vector dimension of the customer dynamic behavior vectors in the behavior vector set. express The value in the l-th dimension express The value in the l-th dimension; Based on the vector distance, the set of behavioral vectors is clustered and assigned to the initial cluster center to obtain the assigned clusters; Calculate the mean of the data in each cluster of the assigned clusters; Based on the data mean, the assigned clusters are updated to obtain the target clusters; Based on the target clustering, the investment users are classified to obtain classified investment users.

[0008] Optionally, performing multimodal spatiotemporal graph convolution on the updated data to obtain the user behavior spatiotemporal tensor includes: Extract the single-modal features of the updated data; The single-modal features are fused to obtain fused features; Perform a spatiotemporal graph convolution operation on the fused features to obtain preliminary spatiotemporal features; Perform multi-scale convolution operations on the preliminary spatiotemporal features to obtain multi-scale spatiotemporal features; The multi-scale spatiotemporal features are arranged in multiple dimensions to obtain the initial behavioral tensor. The initial behavior tensor is pooled to obtain a reduced-dimensional spatiotemporal tensor; The reduced-dimensional spatiotemporal tensor is regularized to obtain the user behavior spatiotemporal tensor.

[0009] Optionally, the step of using the trained knowledge-enhanced large model to construct the market situation embedding vector of the multi-source heterogeneous data includes: The multi-source heterogeneous data is preprocessed with modality alignment to obtain standardized data; Using the trained knowledge-enhanced large model, knowledge is injected into the standardized data to obtain an enhanced feature matrix; Cross-modal attention encoding is performed on the enhanced feature matrix to obtain a primary embedding vector; Perform a temporal causal convolution operation on the primary embedding vector to obtain the market situation embedding vector.

[0010] Optionally, using the user behavior spatiotemporal tensor and the market situation embedding vector, a three-dimensional user profile of the investment user is constructed, including: The user behavior spatiotemporal tensor and the market situation embedding vector are matrix-fused to obtain a joint representation matrix; The three-dimensional features of the joint representation matrix are decoupled to obtain the basic feature set; The basic feature set is dynamically normalized to obtain a normalized feature set; Using the normalized feature set, a three-dimensional user profile of the investment user is constructed.

[0011] Optionally, the step of analyzing the investment suitability between the categorized investment users and the investment market based on the three-dimensional user profile includes: The three-dimensional user profile is used to identify behavioral pattern segments of the categorized investment users; Using the aforementioned behavioral pattern fragments, construct a behavioral pattern dictionary for the categorized investment users; Query the market situation data of the investment market and perform hard rule quantification on the market situation data to obtain a set of situational status labels; The context state label set is used to match the behavior pattern dictionary to obtain suitable labels; Construct the adaptation degree matrix of the adaptation tag; By adjusting the parameters of the fitness matrix, the target fitness matrix is ​​obtained. The target fit matrix is ​​used to analyze the investment fit between the classified investment users and the investment market.

[0012] Optionally, the step of analyzing the investment fit between the categorized investment users and the investment market using the target fit matrix includes: The fit value between the categorized investment users and the investment market is calculated using the following formula: ; in, Indicates the fit value. This represents the dimension of the target fitness matrix. This represents the element in the i-th row and j-th column of the target fitness matrix; Based on the adaptation value, the investment market adaptation ranges of the classified investment users are divided to obtain multiple sets of investment adaptation ranges; The target suitable range is obtained by weighted filtering of the multiple sets of investment suitable ranges; Based on the target fit range, the investment fit between the classified investment users and the investment market is determined.

[0013] Optionally, the step of performing risk propagation analysis on the optimized investment strategy based on the updated dynamic behavior vector, using a trained graph neural network model to obtain a risk score, includes: Extract the key behavior nodes of the updated dynamic behavior vector to obtain the user risk feature set; The optimized investment strategy is deconstructed to obtain a portfolio relationship diagram; The user risk feature set is injected into the portfolio relationship graph to obtain a risk propagation graph; Using the trained graph neural network model, the propagation risk value of the risk propagation graph is calculated; The propagation risk value is numerically calibrated based on stress testing to obtain a risk score.

[0014] Optionally, the step of using the target strategy to screen the categorized investment users to obtain target investment users includes: The applicable conditions of the target strategy are analyzed to obtain the applicable rules; Using the applicable rules, construct the strategy-user matching rule table for the categorized investment users; The strategy-user matching rule table is scanned for user data to obtain a preliminary set of matched users; The preliminary matched user set is subjected to intelligent investment filtering to obtain effective candidate users; The valid candidate users are given priority scores to obtain the scored users; The rated users are then screened to obtain target investment users.

[0015] A method for intelligent investment preference analysis and customer selection based on a large model, characterized in that the method includes: Collect multidimensional interaction data of investment users, perform cross-modal semantic fusion on the multidimensional interaction data to obtain customer dynamic behavior vectors, and cluster the investment behavior features of the investment users based on the customer dynamic behavior vectors to obtain classified investment users; The multidimensional interactive data is updated in real time to obtain updated data. Multimodal spatiotemporal graph convolution is performed on the updated data to obtain user behavior spatiotemporal tensor. The investment market corresponding to the investment user is queried, and multi-source heterogeneous data of the investment market is collected. The trained knowledge augmentation model is used to construct the market situation embedding vector of the multi-source heterogeneous data. The user behavior spatiotemporal tensor and the market situation embedding vector are used to construct a three-dimensional user profile of the investment user. Based on the three-dimensional user profile, the investment fit between the categorized investment users and the investment market is analyzed. Based on the investment fit, an initial investment strategy for the categorized investment users is constructed. The historical investment feedback data of the categorized investment users is queried. Based on the historical investment feedback data, reinforcement learning is performed on the initial investment strategy to obtain an optimized investment strategy. The customer dynamic behavior vector is updated in real time to obtain an updated dynamic behavior vector. Based on the updated dynamic behavior vector, a trained graph neural network model is used to perform risk propagation analysis on the optimized investment strategy to obtain a risk score. The optimized investment strategy is updated in real time using the risk score to obtain a target strategy. The target strategy is used to filter the classified investment users to obtain target investment users.

[0016] Compared to existing technologies, this invention collects multi-dimensional interactive data and performs cross-modal semantic fusion on this data to uncover potential connections between different pieces of information, generating a comprehensive and dynamically reflecting customer dynamic behavior vector that makes user behavior characteristics more comprehensive and representative. Furthermore, this invention ensures data timeliness and accuracy by updating multi-dimensional interactive data in real time, and extracts more representative user behavior features by performing multi-modal spatiotemporal graph convolution on the updated data. Then, it utilizes a trained knowledge-enhanced large model to construct a market situation embedding vector, deeply mining market information, highlighting key features and trends, and using the user behavior spatiotemporal tensor and market situation embedding vector to construct a three-dimensional user profile, providing an intuitive and comprehensive view of investment users for investment analysis. This invention, by analyzing investment suitability based on three-dimensional user profiles, clarifies the positioning of different categories of investment users in the current investment market, understands their matching degree with the market and their strengths and weaknesses, and then constructs initial investment strategies based on investment suitability. Combining market investment opportunities and risk levels, it formulates suitable investment plans for different types of users. Furthermore, this invention updates customer dynamic behavior vectors in real time to ensure the model is based on the latest data. Then, based on the updated dynamic behavior vectors, it uses a trained graph neural network model to perform risk propagation analysis on optimized investment strategies, obtaining risk scores. This helps investment institutions understand the risk level of investment strategies in advance, and then filters categorized investment users based on the risk level of investment strategies, helping investment institutions accurately target customers. Therefore, this invention can improve the comprehensiveness of intelligent investment preference analysis and the accuracy of customer selection. Attached Figure Description

[0017] Figure 1 A functional module diagram of an intelligent investment preference analysis and customer selection system based on a large model, provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for intelligent investment preference analysis and customer selection based on a large model, provided as an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.

[0020] In practice, the server-side equipment deployed in a large-scale model-based intelligent investment preference analysis and customer selection system may consist of one or more devices. This system can be implemented as a business instance, a virtual machine, or hardware devices. For example, it can be implemented as a business instance deployed on one or more devices in a cloud node. Simply put, it can be understood as software deployed on a cloud node, providing large-scale model-based intelligent investment preference analysis and customer selection services to various users. Alternatively, it can be implemented as a virtual machine deployed on one or more devices in a cloud node, with application software installed to manage various users. Or, it can also be implemented as a server composed of numerous identical or different types of hardware devices, with one or more devices configured to provide large-scale model-based intelligent investment preference analysis and customer selection services to various users.

[0021] In terms of implementation, the intelligent investment preference analysis and customer selection system based on the large model and the user terminal are mutually compatible. That is, if the intelligent investment preference analysis and customer selection system based on the large model is implemented as an application installed on a cloud service platform, then the user terminal is implemented as a client that establishes a communication connection with the application; or if the intelligent investment preference analysis and customer selection system based on the large model is implemented as a website, then the user terminal is implemented as a webpage; or if the intelligent investment preference analysis and customer selection system based on the large model is implemented as a cloud service platform, then the user terminal is implemented as a mini-program in an instant messaging application.

[0022] Reference Figure 1 The diagram shown is a functional block diagram of an intelligent investment preference analysis and customer selection system based on a large model provided in an embodiment of the present invention.

[0023] The intelligent investment preference analysis and customer selection system 100 based on a large model described in this invention can be set up in a cloud server. In terms of implementation, it can be implemented as one or more service devices, or as an application installed in the cloud (e.g., a server, server cluster, etc., for intelligent investment preference analysis and customer selection based on a large model), or it can be developed as a website. Depending on the functions implemented, the intelligent investment preference analysis and customer selection system 100 based on a large model includes a user classification module 101, a behavior analysis module 102, a strategy optimization module 103, and a customer screening module 104.

[0024] In this embodiment of the invention, in a large-model-based intelligent investment preference analysis and customer selection tracking system, each of the above modules can be implemented independently and can call other modules. Here, "calling" can be understood as one module connecting to multiple modules of another type and providing corresponding services to those connected modules. In the large-model-based intelligent investment preference analysis and customer selection system provided by this embodiment of the invention, the applicability of the large-model-based intelligent investment preference analysis and customer selection architecture can be adjusted by adding modules and directly calling them without modifying the program code, achieving cluster-based horizontal expansion. This allows for quick and flexible expansion of the large-model-based intelligent investment preference analysis and customer selection system. In practical applications, the above modules can be set in the same device or different devices, or they can be set in virtual devices, such as service instances in a cloud server.

[0025] The following describes, with reference to specific embodiments, the various components and specific workflows of an intelligent investment preference analysis and customer selection system based on a large model.

[0026] The user classification module 101 is used to collect multi-dimensional interaction data of investment users, perform cross-modal semantic fusion on the multi-dimensional interaction data to obtain customer dynamic behavior vectors, and perform investment behavior feature clustering on the investment users based on the customer dynamic behavior vectors to obtain classified investment users.

[0027] The embodiments of the present invention can provide a comprehensive and rich data foundation for subsequent in-depth analysis of user investment characteristics by collecting multi-dimensional interaction data of investment users. Compared with single data, multi-dimensional data can more three-dimensionally and accurately depict the user's investment picture, avoiding the deviation in understanding the user due to information limitations.

[0028] Optionally, the multidimensional interactive data is obtained from the data repository of the financial trading platform, the user's social media database, and other channels.

[0029] The multidimensional interactive data refers to various types of data generated by investment users in various investment-related scenarios, such as transaction records, social media comments, investment consultation behavior, and device usage data.

[0030] Furthermore, the embodiments of the present invention, by performing cross-modal semantic fusion on the multi-dimensional interactive data to obtain customer dynamic behavior vectors, can transform data from different modalities into a unified information representation with semantic association through semantic-level fusion processing, thereby mining the potential connections between different information and generating more comprehensive and representative user behavior feature vectors.

[0031] The customer dynamic behavior vector refers to a mathematical representation that comprehensively and dynamically reflects the behavioral characteristics of investment users, obtained after processing multi-dimensional interactive data.

[0032] As an embodiment of the present invention, the step of performing cross-modal semantic fusion on the multidimensional interactive data to obtain a customer dynamic behavior vector includes: extracting transaction data, media data, and voice data from the multidimensional interactive data; performing temporal convolution on the transaction data to obtain operation sequence features; identifying the semantic features of the media data; performing voice emotion recognition on the voice data to obtain intonation features; and performing multimodal alignment on the operation sequence features, the semantic features, and the intonation features to obtain a customer dynamic behavior vector.

[0033] The operation sequence features refer to the feature vectors obtained by performing temporal convolution on transaction data; the semantic features refer to the semantic information such as the theme, emotion, viewpoint, and specific content related to investment expressed in the text; and the intonation features refer to the emotional tendency of users towards investment-related topics obtained through voice emotion recognition technology, such as positive, negative, or neutral.

[0034] Optionally, the extraction of transaction data, media data, and voice data from the multidimensional interactive data can be achieved using data acquisition tools. This includes extracting user transaction data from investment trading platform databases, covering fields such as transaction time, amount, and product; collecting media data posted and viewed by users from social media and news platforms using web crawling technology, including text and images; and collecting user voice data from investment consultation and other scenarios using a voice acquisition interface. The operation sequence features can be obtained by performing temporal convolution on the transaction data using a temporal convolutional neural network. The semantic features can be obtained by encoding the text in the media data using a pre-trained language model in natural language processing technology, such as BERT. The intonation features can be obtained by converting voice data into text using speech recognition technology, and then using a sentiment analysis model to analyze the sentiment tendency in the text and determine the user's attitude towards investment-related topics, such as positive, negative, or neutral. The customer dynamic behavior vector can utilize a cross-modal attention mechanism to calculate the correlation weights between different modal features, enabling operational sequence features, semantic features, and intonation features to be correlated within the same semantic space. Then, the weighted modal features are concatenated or fused to form a comprehensive customer dynamic behavior vector.

[0035] Furthermore, this embodiment of the invention uses the customer dynamic behavior vector to cluster the investment behavior characteristics of the investment users to obtain classified investment users. This allows for the reasonable classification of a large and complex group of investment users according to their behavior patterns, facilitating targeted analysis and management of different types of users.

[0036] As an embodiment of the present invention, the step of clustering investment users based on the customer dynamic behavior vector to obtain classified investment users includes: performing vector aggregation on the customer dynamic behavior vector to obtain a behavior vector set; classifying the investment users into different categories to obtain different category clusters; selecting initial cluster centers for the investment users from the behavior vector set based on the number of categories in the different category clusters; and calculating the vector distance from different data points in the behavior vector set to the initial cluster centers using the following formula: ; in, Represents vector distance. This represents the i-th data point in the behavior vector set. Let represent the j-th initial cluster center, and m represent the vector dimension of the customer dynamic behavior vectors in the behavior vector set. express The value in the l-th dimension express The value in the l-th dimension; Based on the vector distance, the set of behavioral vectors is clustered and assigned to the initial cluster center to obtain assigned clusters. The mean value of each cluster in the assigned clusters is calculated. Based on the mean value, the assigned clusters are updated to obtain target clusters. Based on the target clusters, the investment users are classified to obtain classified investment users.

[0037] The initial cluster centers refer to some representative points initially selected during the clustering analysis of investment behavior characteristics of investment users, which are used to guide the start of clustering.

[0038] Optionally, the behavior vector set is obtained by storing each customer's dynamic behavior vector in an orderly manner using data storage and management tools (such as a database system). The classification of investment users can be based on some basic attributes of investment users (such as age, investment experience, income level, etc.) or preliminary investment behavior characteristics (such as investment amount, investment product type preference, etc.) to initially classify all investment users, grouping users with similar attributes or characteristics into the same category, forming different user clusters. The assigned clustering compares the vector distance of each data point to each initial cluster center, assigning each data point to the cluster corresponding to the nearest initial cluster center. The target clustering calculates the average value of all data points (customer dynamic behavior vectors) in each assigned cluster across all dimensions, obtaining new cluster centers (data mean vectors). Then, based on the new cluster centers, the distances of data points in the behavior vector set to these new centers are recalculated, and clustering is performed again, repeating this process until the clustering results no longer change significantly. The classification of investment users can be achieved by classifying each investment user according to the cluster to which their corresponding customer dynamic behavior vector belongs, based on the final target clustering result.

[0039] It should be further explained that the above formula for calculating the vector distance is obtained by summing the squares of the differences between the data points in the behavior vector set and the initial cluster center in each dimension, and then taking the square root to obtain the Euclidean distance between them. This can measure the proximity of the data points to the initial cluster center in multidimensional space. The smaller the distance, the more similar the data point is to the cluster center, which provides a quantitative basis for subsequently assigning the data points to appropriate clusters.

[0040] The behavior analysis module 102 is used to update the multidimensional interactive data in real time to obtain updated data, perform multimodal spatiotemporal graph convolution on the updated data to obtain a user behavior spatiotemporal tensor, query the investment market corresponding to the investment user, collect multi-source heterogeneous data of the investment market, use a trained knowledge augmentation model to construct a market situation embedding vector of the multi-source heterogeneous data, and use the user behavior spatiotemporal tensor and the market situation embedding vector to construct a three-dimensional user profile of the investment user.

[0041] The embodiments of the present invention update the multidimensional interactive data in real time to ensure the timeliness and accuracy of the data, enabling the system to reflect the latest interactive behavior information of investment users in a timely manner.

[0042] Optionally, the updated data can be obtained by periodically polling or using a message queue, etc., to acquire the latest multidimensional interactive data, and then merge or replace it with the original data.

[0043] In this embodiment of the invention, by performing multimodal spatiotemporal graph convolution on the updated data, a user behavior spatiotemporal tensor is obtained, which can extract more representative user behavior features and transform them into a compact representation of the user behavior spatiotemporal tensor.

[0044] The user behavior spatiotemporal tensor refers to a multimodal data structure used to describe user behavior in the time and space dimensions.

[0045] As an embodiment of the present invention, the step of performing multimodal spatiotemporal graph convolution on the updated data to obtain a user behavior spatiotemporal tensor includes: extracting single-modal features from the updated data; performing feature fusion on the single-modal features to obtain fused features; performing spatiotemporal graph convolution on the fused features to obtain preliminary spatiotemporal features; performing multi-scale convolution on the preliminary spatiotemporal features to obtain multi-scale spatiotemporal features; arranging the multi-scale spatiotemporal features in multiple dimensions to obtain an initial behavior tensor; performing pooling on the initial behavior tensor to obtain a dimensionality-reduced spatiotemporal tensor; and performing regularization on the dimensionality-reduced spatiotemporal tensor to obtain the user behavior spatiotemporal tensor.

[0046] The initial spatiotemporal features refer to the feature representation obtained after performing spatiotemporal graph convolution on the fused features; the multi-scale spatiotemporal features refer to the features obtained after performing multi-scale convolution on the initial spatiotemporal features; and the initial behavioral tensor refers to the tensor structure obtained by arranging the multi-scale spatiotemporal features in multiple dimensions such as time, space, and modality.

[0047] Optionally, the single-modal features can be obtained by using corresponding feature extraction methods for different modalities in the updated data, such as transaction data, media data, and voice data. The fusion features can be obtained by concatenating the feature vectors of each modality in sequence according to the extracted single-modal features. The preliminary spatiotemporal features can be obtained by performing spatiotemporal graph convolution on the fusion features using a graph convolutional network (GCN). The multi-scale spatiotemporal features can be obtained by performing convolution operations on the preliminary spatiotemporal features using convolution kernels of different sizes (such as 3, 5, and 7, which can be set using a deep learning framework). The pooling operation on the initial behavior tensor to obtain the dimensionality-reduced spatiotemporal tensor can be implemented using the max pooling method. The user behavior spatiotemporal tensor can be obtained by regularizing the dimensionality-reduced spatiotemporal tensor using L2 regularization.

[0048] This invention, through querying the investment market corresponding to the investment user and collecting multi-source heterogeneous data from the investment market, can clarify the market environment in which the investment user is located, thereby providing rich background information for analyzing the investment user's behavior.

[0049] The investment market refers to the venues or fields where various investment activities are carried out, including the stock market, bond market, futures market, foreign exchange market, fund market, real estate market, and other financial derivatives markets. The multi-source heterogeneous data refers to data from multiple different data sources that differ in data structure, data format, and data semantics.

[0050] Optionally, the multi-source heterogeneous data can be obtained through methods such as web crawling to collect web page data, using API interfaces to obtain structured data from financial institutions or data platforms, and reading various types of data stored locally from the file system.

[0051] Furthermore, in this embodiment of the invention, by utilizing the trained knowledge-enhanced large model to construct the market situation embedding vector of the multi-source heterogeneous data, the collected multi-source heterogeneous data can be deeply mined and analyzed, transforming complex market information into a low-dimensional market situation embedding vector, highlighting the key features and trends in the market data.

[0052] The market situation embedding vector refers to a method that maps various situational information of the investment market, such as price trends, changes in trading volume, and market sentiment, into a vector representation in a low-dimensional vector space through specific algorithms and models.

[0053] It should be further explained that the trained knowledge-enhanced large model refers to a large model that, based on a traditional large model, is trained and optimized by introducing external knowledge sources or knowledge graphs, so that it can better understand and process natural language text and generate more accurate and targeted answers. It can be configured using a dataset containing a large amount of text data and knowledge graphs, and trained using a model based on the Transformer architecture, combined with knowledge embedding, multi-task learning and other techniques.

[0054] As an embodiment of the present invention, the step of constructing the market situation embedding vector of the multi-source heterogeneous data using a trained knowledge-enhanced big data model includes: performing modality alignment preprocessing on the multi-source heterogeneous data to obtain standardized data; using the trained knowledge-enhanced big data model to inject knowledge into the standardized data to obtain an enhanced feature matrix; performing cross-modal attention encoding on the enhanced feature matrix to obtain a primary embedding vector; and performing a temporal causal convolution operation on the primary embedding vector to obtain the market situation embedding vector.

[0055] The enhanced feature matrix refers to the data structure obtained after knowledge injection into standardized data by the knowledge-enhancing big model. It not only contains the feature information of the original multi-source heterogeneous data after preprocessing, but also integrates the relevant knowledge introduced by the knowledge-enhancing big model from external knowledge sources (such as knowledge graphs). The primary embedding vector refers to the vector representation obtained after cross-modal attention encoding of the enhanced feature matrix.

[0056] Optionally, the standardized data can first identify the features and formats of different modalities (such as text, numerical values, and images) in multi-source heterogeneous data, and then perform unified encoding and normalization processing on each modal data to obtain the data; the enhanced feature matrix can utilize knowledge graph technology to integrate entity and relationship information from the knowledge graph into a trained knowledge-enhanced large model. The large model will integrate relevant knowledge into the standardized data based on its learned knowledge system to achieve knowledge injection, and finally output the matrix; the primary embedding vector can utilize the attention mechanism in the Transformer architecture to construct a cross-modal attention module to encode the enhanced feature matrix; the market situation embedding vector can use a temporal convolutional network to arrange the primary embedding vectors in chronological order, and then use causal convolution kernels to perform convolution operations in the time dimension to obtain the data.

[0057] Furthermore, by utilizing the spatiotemporal tensor of user behavior and the market situation embedding vector, the embodiments of the present invention can construct a three-dimensional user profile of the investment user, providing an intuitive and comprehensive view of the investment user for investment analysis, which helps to better understand the needs, risk preferences and behavioral trends of the investment user.

[0058] The three-dimensional user profile refers to a model that provides a comprehensive and three-dimensional description of investment users. It is constructed based on multi-dimensional data such as user behavior spatiotemporal tensors and market situation embedding vectors, mapping various attributes and behavioral characteristics of investment users into three-dimensional space to form an intuitive and visual user image.

[0059] As an embodiment of the present invention, a three-dimensional user profile of the investment user is constructed using the user behavior spatiotemporal tensor and the market situation embedding vector, including: performing matrix-based fusion of the user behavior spatiotemporal tensor and the market situation embedding vector to obtain a joint representation matrix; decoupling the three-dimensional features of the joint representation matrix to obtain a basic feature group; performing dynamic normalization processing on the basic feature group to obtain a normalized feature group; and constructing the three-dimensional user profile of the investment user using the normalized feature group.

[0060] The joint representation matrix refers to the matrix obtained by matrix-fusing the user behavior spatiotemporal tensor and the market situation embedding vector, and the basic feature group refers to a set of features obtained by decoupling the joint representation matrix into three-dimensional features.

[0061] Optionally, the joint representation matrix can transform the user behavior spatiotemporal tensor and the market situation embedding vector into a matrix form. Based on the tensor and vector dimensions, it can be concatenated by rows or columns to integrate the information. For example, if the user behavior spatiotemporal tensor is a three-dimensional matrix and the market situation embedding vector is a one-dimensional vector, the vector can be expanded by rows and concatenated with the corresponding dimensions of the tensor. The basic feature group can be obtained by decomposing the user behavior, market situation, and other information represented by different dimensions in the joint representation matrix according to dimensions such as time, space, and feature attributes. The normalized feature group can determine dynamic normalization parameters based on the value range and variation patterns of different features in the basic feature group. Then, based on these parameters, the value of each feature is mapped to a standard range, such as [0, 1] or [-1, 1]. The three-dimensional user profile can utilize the features in the normalized feature group as the basis for describing the investment user profile in three dimensions (such as risk preference, investment activity, and market sensitivity). Each feature value is transformed into coordinates or attribute values ​​in the corresponding dimension to construct a user image in three-dimensional space.

[0062] The strategy optimization module 103 is used to analyze the investment fit between the categorized investment users and the investment market based on the three-dimensional user profile, construct an initial investment strategy for the categorized investment users based on the investment fit, query the historical investment feedback data of the categorized investment users, and perform reinforcement learning on the initial investment strategy based on the historical investment feedback data to obtain an optimized investment strategy.

[0063] This invention, through the analysis of the investment suitability of different categories of investment users with the investment market based on the three-dimensional user profile, can clarify the positioning of different categories of investment users in the current investment market, helping market analysts understand whether investment users are suitable for the current market environment, and in what aspects they have advantages or disadvantages.

[0064] The investment fit is a comprehensive indicator used to measure the degree of matching between investors and the investment market.

[0065] As an embodiment of the present invention, the step of analyzing the investment fit between the categorized investment users and the investment market based on the three-dimensional user profile includes: identifying behavioral pattern fragments of the categorized investment users using the three-dimensional user profile; constructing a behavioral pattern dictionary of the categorized investment users using the behavioral pattern fragments; querying market situation data of the investment market and performing hard rule quantification processing on the market situation data to obtain a contextual state label set; matching the contextual state label set with the behavioral pattern dictionary to obtain fit labels; constructing a fit degree matrix of the fit labels; adjusting the parameters of the fit degree matrix to obtain a target fit degree matrix; and analyzing the investment fit between the categorized investment users and the investment market using the target fit matrix.

[0066] The behavioral pattern fragments refer to relatively independent behavioral sequence units segmented from user behavior data contained in the 3D user profile according to specific rules or time windows. For example, a series of actions by a user who continuously purchases a certain type of fund product within a month can be considered a behavioral pattern fragment, reflecting a specific behavioral characteristic of the user during the investment process. The behavioral pattern dictionary refers to a collection of data that encodes and classifies behavioral pattern fragments and stores them in a dictionary structure. The market situation data refers to a data set reflecting the overall situation and changing trends of the investment market, including the rise and fall of market indices, the development dynamics of different industries, the impact of macroeconomic policies on the market, and fluctuations in interest rates and exchange rates. The contextual state label set refers to a set of labels obtained after the market situation data has undergone hard rule quantification. Through preset rules, the market situation data is divided into different categories or states, and each category or state is assigned a label. For example, based on the range of market index increases, market states are divided into "bull market," "oscillating market," and "bear market," and represented by corresponding labels. The fit matrix refers to a two-dimensional matrix constructed with fit tags as rows and columns, where the element values ​​in the matrix represent the degree of association or matching strength between two fit tags.

[0067] Optionally, the behavioral pattern fragments can be obtained using sequence segmentation algorithms, such as a sliding window method, to mine user behavior data contained in the 3D user profile, such as transaction time, transaction amount, and investment product selection. Based on specific time windows or behavioral logic, the user's behavioral sequence is segmented into multiple relatively independent fragments. The behavioral pattern dictionary can encode and classify the identified behavioral pattern fragments, assign a unique identifier to each unique behavioral pattern fragment, and store these identifiers and their corresponding behavioral pattern fragment information in a dictionary structure. The contextual state label set can obtain relevant data from market situation data, such as market indices, industry dynamics, and policy changes. Based on preset hard rules, such as setting different intervals of the market index corresponding to different states, conditional judgment statements are written using if-elif-else statements in Python to achieve hard rule quantification. The adaptation labels use string matching algorithms, such as regular expression matching or simple string comparison, to compare each label in the contextual state label set with the behavioral pattern fragments in the behavioral pattern dictionary, finding the matching behavioral pattern fragments. The fitness matrix can be constructed using a two-dimensional array as the initial matrix. Then, with the fitness labels as rows and columns, values ​​are assigned to each element in the matrix based on the correlation or matching strength between the labels. The target fitness matrix can be obtained by dynamically adjusting the parameters in the fitness matrix using an adaptive algorithm, such as the Kalman filter algorithm, based on factors such as real-time market changes and dynamic adjustments in user behavior.

[0068] Furthermore, as another optional embodiment of the present invention, the step of analyzing the investment fit between the categorized investment users and the investment market using the target fit matrix includes: calculating the fit value between the categorized investment users and the investment market using the following formula: ; in, Indicates the fit value. This represents the dimension of the target fitness matrix. This represents the element in the i-th row and j-th column of the target fitness matrix; Based on the fit value, the investment market fit intervals of the classified investment users are divided into multiple sets of investment fit intervals. The multiple sets of investment fit intervals are weighted and filtered to obtain the target fit interval. Based on the target fit interval, the investment fit degree between the classified investment users and the investment market is determined.

[0069] Optionally, the multiple investment matching intervals can be pre-defined according to business experience and data distribution characteristics, with rules for dividing the matching intervals set in advance. For example, if the matching value range is 0-1, matching values ​​less than 0.3 can be set as low matching intervals, 0.3-0.7 as medium matching intervals, and greater than 0.7 as high matching intervals. The calculated matching values ​​are then assigned to the corresponding intervals according to the rules. The target matching intervals can be weighted according to the importance of different investment matching intervals to investment decisions. For example, if the high matching interval is considered to have a greater impact on investment decisions, it can be assigned a higher weight, such as 0.6; the medium matching interval is weighted at 0.3, and the low matching interval at 0.1. Then, the product of the number of users in each interval and the corresponding weight is calculated, and the matching intervals are selected based on the size of the product. The investment fit can be determined in Python through simple conditional statements. For example, if the target fit range is a high fit range, the investment fit between the classified investment users and the investment market is determined to be high; if it is a medium fit range, the fit is medium; if it is a low fit range, the fit is low.

[0070] It should be further explained that the above formula for calculating the fit value assigns a weight related to the row and column position of each element in the target fit matrix, and then adds all the weighted elements together to obtain the fit value. This allows us to consider the different impacts of elements at different positions in the matrix on the fit. The further an element is from the top left corner of the matrix, the greater its weight and the greater its contribution to the fit value. This means that when analyzing investment fit, we pay more attention to elements at specific positions in the matrix. These elements represent more important investment characteristics or situations, thus providing a more comprehensive and detailed measurement of the fit between different types of investment users and the investment market.

[0071] Furthermore, in this embodiment of the invention, the initial investment strategy for the categorized investment users can be constructed based on the investment suitability analysis results, combined with market investment opportunities and risk levels, to formulate suitable investment plans for different categories of investment users.

[0072] The initial investment strategy refers to a preliminary investment plan or action plan formulated based on the investment suitability of different types of investors and the investment market, such as asset allocation, investment product selection, and risk management.

[0073] Optionally, the initial investment strategy can categorize users based on investment suitability, such as highly suitable aggressive users, moderately suitable stable users, and poorly suitable conservative users. Then, for different user types, the asset allocation ratio, investment product selection, investment timing, and risk management measures are determined. For example, aggressive users can allocate more to stocks, while conservative users can focus on low-risk assets such as bonds.

[0074] This invention allows for understanding the performance of investment users in actual investments and their responses to different investment strategies by querying the historical investment feedback data of the categorized investment users, thereby providing practical data support for subsequent strategy optimization.

[0075] The historical investment feedback data refers to various relevant information generated by categorized investment users during their past investment activities, such as transaction behavior, fund changes, and market feedback data.

[0076] Furthermore, in this embodiment of the invention, by performing reinforcement learning on the initial investment strategy based on the historical investment feedback data, an optimized investment strategy is obtained. This optimized investment strategy can adapt to dynamic changes in the market and the evolution of user investment behavior, thereby ensuring the adaptability and effectiveness of the investment strategy.

[0077] Optionally, the optimized investment strategy can first preprocess historical investment feedback data, extract features to construct state, action and reward functions, and then use reinforcement learning algorithms (such as DQN) to adjust the parameters of the initial investment strategy, and obtain the result when the reward function meets the required value.

[0078] The customer screening module 104 is used to update the customer dynamic behavior vector in real time to obtain an updated dynamic behavior vector. Based on the updated dynamic behavior vector, a trained graph neural network model is used to perform risk propagation analysis on the optimized investment strategy to obtain a risk score. The risk score is used to update the optimized investment strategy in real time to obtain a target strategy. The target strategy is used to screen the classified investment users to obtain target investment users.

[0079] The embodiments of the present invention update the customer dynamic behavior vector in real time to obtain the updated dynamic behavior vector, which can ensure that the data on which the model is based is up-to-date, accurately capture the dynamic changes in customer behavior, and provide accurate basic information for subsequent analysis and decision-making.

[0080] Optionally, the updated dynamic behavior vector can be obtained by collecting new customer behavior data in real time and correcting and supplementing the original customer dynamic behavior vector according to established feature extraction and vector transformation rules.

[0081] This invention, through the updated dynamic behavior vector, utilizes a trained graph neural network model to perform risk propagation analysis on the optimized investment strategy, obtaining a risk score. This risk score can help investment institutions understand the potential risk level of an investment strategy in advance, enabling them to take corresponding measures for risk control and management.

[0082] The trained graph neural network model refers to a graph neural network model that has been trained and optimized with a large amount of data and can be configured through a deep learning framework such as TensorFlow. The risk score refers to the quantitative assessment of the degree of risk faced by the optimized investment strategy under the current market environment and investor behavior.

[0083] As an embodiment of the present invention, the step of performing risk propagation analysis on the optimized investment strategy based on the updated dynamic behavior vector and using a trained graph neural network model to obtain a risk score includes: extracting key behavior nodes from the updated dynamic behavior vector to obtain a user risk feature set; deconstructing the optimized investment strategy to obtain a portfolio relationship graph; injecting the user risk feature set into the portfolio relationship graph to obtain a risk propagation graph; using the trained graph neural network model to calculate the propagation risk value of the risk propagation graph; and performing numerical calibration of the propagation risk value based on stress testing to obtain a risk score.

[0084] The user risk feature set refers to the set of key behavioral nodes extracted from the updated dynamic behavior vector. The portfolio relationship graph refers to the graph structure constructed by taking various investment products in the optimized investment strategy as nodes and the relationships between products such as allocation ratio, capital flow and mutual correlation as edges. The risk propagation graph refers to the graph obtained after injecting the user risk feature set into the portfolio relationship graph.

[0085] Optionally, the user risk feature set can utilize association rule algorithms in data mining techniques, such as the Apriori algorithm, to conduct in-depth analysis of the updated dynamic behavior vectors. Based on pre-defined importance indicators, nodes that significantly reflect user investment behavior risk are identified, and the data related to these key behavior nodes is stored. The portfolio relationship graph can use graph data structures and algorithms, such as NetworkX, to decompose information such as the various investment products in the optimized investment strategy and their allocation ratios and interrelationships. Then, a graph is constructed using investment products as nodes and relationships such as investment ratios and fund flows between products as edges. The risk propagation graph can utilize data fusion technology to accurately associate the user risk feature set data with corresponding nodes based on the constructed portfolio relationship graph by writing code. For example, data can be injected by matching the unique identifier of the node (such as the investment product code). The propagation risk value can be obtained by inputting the risk propagation graph into a trained graph neural network model, which calculates the nodes and edges in the graph according to the risk propagation pattern and rules, and outputs a value reflecting the degree of risk propagation in the portfolio. The risk score can utilize financial risk simulation techniques, such as Monte Carlo simulation, to generate risk data under various extreme market scenarios. Then, different levels of stress tests are applied to the calculated propagation risk value to simulate extreme market conditions. Based on the stress test results, the propagation risk value is adjusted and calibrated.

[0086] This invention, through the use of the risk score to update the optimized investment strategy in real time, obtains a target strategy that ensures the investment strategy can dynamically respond to market changes and changes in customer behavior, thereby improving the effectiveness and adaptability of the investment strategy and reducing risk.

[0087] Optionally, the target strategy can be obtained by adjusting key elements such as asset allocation and investment target selection in real time based on risk scores and pre-set risk and strategy adjustment rules, such as reducing the proportion of high-volatility assets when the risk is high.

[0088] This invention, through the use of the target strategy to screen the categorized investment users and obtain target investment users, can help investment institutions more accurately locate target customers and concentrate resources to provide these customers with investment services that better meet their needs.

[0089] As an embodiment of the present invention, the step of using the target strategy to screen the categorized investment users and obtain target investment users includes: parsing the applicable conditions of the target strategy to obtain applicable rules; using the applicable rules to construct a strategy-user matching rule table for the categorized investment users; scanning user data in the strategy-user matching rule table to obtain a preliminary matching user set; performing intelligent investment filtering on the preliminary matching user set to obtain valid candidate users; prioritizing the valid candidate users to obtain rated users; and screening the rated users to obtain target investment users.

[0090] Wherein, the applicable rules refer to the policy-user matching rule table, and the preliminary matched user set refers to...

[0091] Optionally, the applicable rules can be derived by using natural language processing technology combined with rule extraction algorithms to deeply analyze the composition and logic of the target strategy, and to identify the applicable conditions of the target strategy in terms of investment target type, market environment, and user risk preferences. These applicable conditions are then organized into quantifiable and executable rules, such as specifying the suitable range of investment products, market volatility range, and risk tolerance range. The strategy-user matching rule table can be designed to match user characteristics with the applicable conditions of the strategy for different categories of investment users based on the applicable rules, and then presented in tabular form. The rows of the table represent different categories of investment user groups, and the columns represent the applicable rules of the target strategy. The matching relationship between each user group and the rule is clearly stated in the corresponding cell, such as matching, not matching, or partially matching. The preliminary matched user set can be obtained by reading the strategy-user matching rule table, comparing the actual data of each category of investment users one by one according to the matching relationship set in the table, and then filtering out users whose feature data matches the applicable rules of the target strategy from the user database. The effective candidate users can be comprehensively evaluated using machine learning classification algorithms, such as random forests, based on factors such as the investment history, rationality of investment goals, and stability of financial status of users in the initial matching user set. Users with abnormal investment behavior, unrealistic investment goals, or potential financial risks are filtered out. The scoring users can consider factors such as the size of their investment, investment frequency, responsiveness to market changes, and loyalty. Each factor is assigned a corresponding weight, and the effective candidate users are quantitatively scored based on these factors. The scoring results are then correlated with user information. The target investment users can be selected based on business needs and objectives, with a threshold set to select the top 100 users with scores higher than the threshold. The specific threshold needs to be set according to actual application, or the top 100 users with the highest scores can be directly selected.

[0092] like Figure 2The diagram shown is a flowchart illustrating a large-scale model-based intelligent investment preference analysis and customer selection method according to an embodiment of the present invention. In this embodiment, the large-scale model-based intelligent investment preference analysis and customer selection method includes: Collect multidimensional interaction data of investment users, perform cross-modal semantic fusion on the multidimensional interaction data to obtain customer dynamic behavior vectors, and cluster the investment behavior features of the investment users based on the customer dynamic behavior vectors to obtain classified investment users; The multidimensional interactive data is updated in real time to obtain updated data. Multimodal spatiotemporal graph convolution is performed on the updated data to obtain user behavior spatiotemporal tensor. The investment market corresponding to the investment user is queried, and multi-source heterogeneous data of the investment market is collected. The trained knowledge augmentation model is used to construct the market situation embedding vector of the multi-source heterogeneous data. The user behavior spatiotemporal tensor and the market situation embedding vector are used to construct a three-dimensional user profile of the investment user. Based on the three-dimensional user profile, the investment fit between the categorized investment users and the investment market is analyzed. Based on the investment fit, an initial investment strategy for the categorized investment users is constructed. The historical investment feedback data of the categorized investment users is queried. Based on the historical investment feedback data, reinforcement learning is performed on the initial investment strategy to obtain an optimized investment strategy. The customer dynamic behavior vector is updated in real time to obtain an updated dynamic behavior vector. Based on the updated dynamic behavior vector, a trained graph neural network model is used to perform risk propagation analysis on the optimized investment strategy to obtain a risk score. The optimized investment strategy is updated in real time using the risk score to obtain a target strategy. The target strategy is used to filter the classified investment users to obtain target investment users.

[0093] In the several embodiments provided by this invention, it should be understood that the provided systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0094] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smart investment preference analysis and customer selection system based on a large model, characterized in that, The system, based on a large model for intelligent investment preference analysis and customer selection, includes: a user classification module, a behavior analysis module, a strategy optimization module, and a customer screening module. The user classification module is used to collect multi-dimensional interaction data of investment users, perform cross-modal semantic fusion on the multi-dimensional interaction data to obtain customer dynamic behavior vectors, and perform investment behavior feature clustering on the investment users based on the customer dynamic behavior vectors to obtain classified investment users. The behavior analysis module is used to update the multidimensional interaction data in real time to obtain updated data, perform multimodal spatiotemporal graph convolution on the updated data to obtain a user behavior spatiotemporal tensor, query the investment market corresponding to the investment user, collect multi-source heterogeneous data of the investment market, use a trained knowledge augmentation model to construct a market situation embedding vector of the multi-source heterogeneous data, and use the user behavior spatiotemporal tensor and the market situation embedding vector to construct a three-dimensional user profile of the investment user. The strategy optimization module is used to analyze the investment fit between the categorized investment users and the investment market based on the three-dimensional user profile, construct an initial investment strategy for the categorized investment users based on the investment fit, query the historical investment feedback data of the categorized investment users, and perform reinforcement learning on the initial investment strategy based on the historical investment feedback data to obtain an optimized investment strategy. The customer screening module is used to update the customer dynamic behavior vector in real time to obtain an updated dynamic behavior vector. Based on the updated dynamic behavior vector, a trained graph neural network model is used to perform risk propagation analysis on the optimized investment strategy to obtain a risk score. The risk score is used to update the optimized investment strategy in real time to obtain a target strategy. The target strategy is used to screen the classified investment users to obtain target investment users.

2. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, The process of performing cross-modal semantic fusion on the multidimensional interaction data to obtain a customer dynamic behavior vector includes: Extract transaction data, media data, and voice data from the multidimensional interactive data; Perform temporal convolution on the transaction data to obtain operation sequence features; Identify the semantic features of the media data; Speech emotion recognition is performed on the speech data to obtain intonation features; Multimodal alignment is performed on the operation sequence features, the semantic features, and the intonation features to obtain the customer dynamic behavior vector.

3. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, The process of clustering investment behavior features of investment users based on the customer dynamic behavior vector to obtain categorized investment users includes: The customer dynamic behavior vectors are aggregated to obtain a behavior vector set; The investment users are categorized to obtain different user clusters; Based on the number of categories of users clustered together, the initial cluster centers of the investment users are selected from the set of behavior vectors; The vector distance from different data points in the behavior vector set to the initial cluster center is calculated using the following formula: ； in, Represents vector distance. This represents the i-th data point in the behavior vector set. Let represent the j-th initial cluster center, and m represent the vector dimension of the customer dynamic behavior vectors in the behavior vector set. express The value in the l-th dimension express The value in the l-th dimension; Based on the vector distance, the set of behavioral vectors is clustered and assigned to the initial cluster centers to obtain the assigned clusters; Calculate the mean of the data in each cluster of the assigned clusters; Based on the data mean, the assigned clusters are updated to obtain the target clusters; Based on the target clustering, the investment users are classified to obtain classified investment users.

4. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, The step of performing multimodal spatiotemporal graph convolution on the updated data to obtain the user behavior spatiotemporal tensor includes: Extract the single-modal features of the updated data; The single-modal features are fused to obtain fused features; Perform a spatiotemporal graph convolution operation on the fused features to obtain preliminary spatiotemporal features; Perform multi-scale convolution operations on the preliminary spatiotemporal features to obtain multi-scale spatiotemporal features; The multi-scale spatiotemporal features are arranged in multiple dimensions to obtain the initial behavioral tensor. The initial behavior tensor is pooled to obtain a reduced-dimensional spatiotemporal tensor; The reduced-dimensional spatiotemporal tensor is regularized to obtain the user behavior spatiotemporal tensor.

5. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, The step of constructing a market situation embedding vector from the multi-source heterogeneous data using a trained knowledge-enhanced large model includes: The multi-source heterogeneous data is preprocessed with modality alignment to obtain standardized data; Using the trained knowledge-enhanced large model, knowledge is injected into the standardized data to obtain an enhanced feature matrix; Cross-modal attention encoding is performed on the enhanced feature matrix to obtain a primary embedding vector; Perform a temporal causal convolution operation on the primary embedding vector to obtain the market situation embedding vector.

6. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, Using the user behavior spatiotemporal tensor and the market situation embedding vector, a three-dimensional user profile of the investment user is constructed, including: The user behavior spatiotemporal tensor and the market situation embedding vector are matrix-fused to obtain a joint representation matrix; The three-dimensional features of the joint representation matrix are decoupled to obtain the basic feature set; The basic feature set is dynamically normalized to obtain a normalized feature set; Using the normalized feature set, a three-dimensional user profile of the investment user is constructed.

7. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, The analysis of the investment suitability between the categorized investment users and the investment market based on the three-dimensional user profile includes: The three-dimensional user profile is used to identify behavioral pattern segments of the categorized investment users; Using the aforementioned behavioral pattern fragments, construct a behavioral pattern dictionary for the categorized investment users; Query the market situation data of the investment market and perform hard rule quantification on the market situation data to obtain a set of situational status labels; The context state label set is used to match the behavior pattern dictionary to obtain suitable labels; Construct the adaptation degree matrix of the adaptation tag; By adjusting the parameters of the fitness matrix, the target fitness matrix is ​​obtained. The target fit matrix is ​​used to analyze the investment fit between the classified investment users and the investment market.

8. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 7, wherein analyzing the investment fit between the categorized investment users and the investment market using the target fit matrix includes: The fit value between the categorized investment users and the investment market is calculated using the following formula: ； in, Indicates the fit value. This represents the dimension of the target fitness matrix. This represents the element in the i-th row and j-th column of the target fitness matrix; Based on the adaptation value, the investment market adaptation ranges of the classified investment users are divided to obtain multiple sets of investment adaptation ranges; The target suitable range is obtained by weighted filtering of the multiple sets of investment suitable ranges; Based on the target fit range, the investment fit between the classified investment users and the investment market is determined.

9. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, The process involves using the updated dynamic behavior vector and a trained graph neural network model to perform risk propagation analysis on the optimized investment strategy, resulting in a risk score, including: Extract the key behavior nodes of the updated dynamic behavior vector to obtain the user risk feature set; The optimized investment strategy is deconstructed to obtain a portfolio relationship diagram; The user risk feature set is injected into the portfolio relationship graph to obtain a risk propagation graph; Using the trained graph neural network model, the propagation risk value of the risk propagation graph is calculated; The propagation risk value is numerically calibrated based on stress testing to obtain a risk score.

10. The intelligent investment preference analysis and customer selection system based on a large model as described in claim 1, characterized in that, The step of using the target strategy to screen the categorized investment users and obtain target investment users includes: The applicable conditions of the target strategy are analyzed to obtain the applicable rules; Using the applicable rules, construct the strategy-user matching rule table for the categorized investment users; The strategy-user matching rule table is scanned for user data to obtain a preliminary set of matched users; The preliminary matched user set is subjected to intelligent investment filtering to obtain effective candidate users; The valid candidate users are given priority scores to obtain the scored users; The rated users are then screened to obtain target investment users.

Images