Financial product recommendation method, system and device based on large language model assistance

Abstract

The invention discloses a financial product recommendation method, system and device based on large language model assistance, and the method comprises the steps: obtaining user-related data and financial product-related data, and carrying out the feature extraction, and obtaining user-related features and financial product-related features; constructing an interpretation generation module, wherein the interpretation generation module comprises a large language model, an auto-encoder network and a gating network; combining the weighted explanation features, the user features, the financial product related features, the user financing preference change features, the context features and the recommendation reason features to form a feature matrix, and training a traditional recommendation pre-training model through the feature matrix to obtain an improved recommendation model; and processing the to-be-recommended data based on the improved recommendation model to obtain a financial product recommendation result. The large language model and the traditional recommendation model are deeply fused, so that the interpretability of the recommendation system is remarkably improved, and the recommendation accuracy of the model is further enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system, and apparatus for recommending financial products based on a large language model. Background Technology

[0002] In the era of digital finance, financial product recommendation systems have become a key tool for financial institutions to improve service quality and optimize user experience. These systems deeply integrate cutting-edge technologies such as big data analytics and machine learning, conducting in-depth analysis from multiple dimensions, including user behavior patterns, risk preferences, and financial status, to provide users with highly customized financial product recommendation services. Currently, mainstream recommendation systems primarily adopt algorithmic architectures such as collaborative filtering, content recommendation, and hybrid recommendation, and are based on user profiling and customer segmentation analysis to achieve precise matching between products and user needs.

[0003] However, current financial product recommendation systems generally encounter the problem of being "black box models." While these systems can predict users' potential product preferences based on their past purchasing behavior, they fall short in clearly explaining the logic behind recommendation decisions. In an era of information overload, the need for explainable recommendation systems has become increasingly urgent: users crave to understand the specific reasons for product recommendations to improve transparency and trust in their decisions; financial institutions need insights into the core drivers of product popularity to optimize product design and marketing strategies; and platform operators urgently need to understand the preference characteristics of specific user groups to enhance the precision of marketing and user experience. This demand has spurred the rapid development of research in the field of explainable recommendation systems.

[0004] Current financial product recommendation systems generally face two core problems: First, the systems lack transparency: While these systems can use users' past behavioral data to predict their preferences for financial products, the models they rely on are often "black boxes," making it difficult to clearly articulate the logic behind the recommendation decisions. This lack of transparency not only makes it difficult for users to grasp the basis of the recommendations, reducing the clarity and trust in the decision-making process, but also hinders financial institutions from gaining a deeper understanding of user preferences, limiting the further development of product improvement and precision marketing strategies.

[0005] Secondly, there is the challenge of balancing interpretability and model performance: Currently, improving the interpretability of recommender systems often requires adding extra constraints, which to some extent limits the model's flexibility and expressiveness, potentially leading to a decrease in recommendation accuracy. Therefore, many systems still rely on black-box models. Since enterprises typically prioritize performance, they are unwilling to sacrifice recommendation accuracy for interpretability, making the conflict between interpretability and model performance a long-standing challenge in the financial industry.

[0006] To address the aforementioned problems, this invention proposes a technical solution that innovatively integrates the powerful text understanding and reasoning capabilities of large language models with traditional recommendation models. This design not only provides clear and reasonable explanations for the model's recommendation results, enhancing users' understanding and trust in the recommendation decisions, but also leverages the explanations generated by the large language model to further optimize the recommendation strategy. Therefore, while improving model interpretability, it also increases recommendation accuracy, achieving a dual breakthrough in interpretability and performance. Summary of the Invention

[0007] This invention addresses the shortcomings of existing technologies by providing a method, system, and apparatus for recommending financial products based on a large language model.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A financial product recommendation method based on a large language model includes the following steps: Acquire user-related data and financial product-related data and extract features to obtain user-related features and financial product-related features. The user-related features include at least user features and user historical purchase sequences, and the financial product-related features include candidate product features and contextual features. An explanation generation module is constructed, comprising a large language model, an autoencoder network, and a gating network. Based on the large language model, user-related features and financial product-related features are analyzed to obtain changes in user investment preferences. Candidate financial products are evaluated to obtain recommendation reasons. The autoencoder network receives user investment preference changes and recommendation reasons, performs feature compression and information extraction, and then reconstructs the data to obtain user investment preference change features and recommendation reason features. The gating network concatenates user features, candidate product features, and contextual features, applies a linear rectifier function, and normalizes using a flexible maximum value to obtain weighted explanation features. The recommendation reasons include matching degree results and recommendation explanations. The weighted explanatory features, user features, financial product-related features, user financial preference change features, context features, and recommendation reason features are combined to form a feature matrix. The traditional recommendation pre-training model is then trained using the feature matrix to obtain an improved recommendation model. The improved recommendation model is used to process the data to be recommended, and the results of financial product recommendations are obtained.

[0009] As one possible implementation method, the user characteristics include demographic characteristics, user device characteristics, and user activity characteristics; The user's historical purchase sequence includes the user's historical purchases of financial products and corresponding product descriptions; The characteristics of candidate products include at least risk characteristics, wherein the candidate products are those that have passed the initial screening by the recall model; Contextual features include continuous features and labels; Among them, demographic characteristics include at least age, gender, and place of origin; user device characteristics include at least device operating system and device operator; and user activity characteristics include at least platform activity, scenario activity, and user account level.

[0010] As one possible implementation, the large language model includes a user profile unit and a recommendation reason unit; The user profiling unit receives relevant user features and extracts key information; it analyzes the key information to obtain changes in user financial management preferences, and then obtains the trend of changes in user financial management preferences over time. The recommendation reason unit analyzes the relevant characteristics of financial products to obtain various attribute characteristics of candidate products; it then performs a matching degree analysis between the various attribute characteristics of candidate products and changes in users' financial purchase preferences to obtain matching degree results, wherein each matching degree result includes a detailed explanation of the recommendation reason.

[0011] As one possible implementation, the autoencoder network is a multilayer perceptron network, including an encoder and a decoder; The encoder maps the received user financial preference changes and recommendation reasons to a low-dimensional space to achieve feature compression and information extraction, thereby obtaining low-dimensional user financial preference change features and low-dimensional recommendation reason features. The decoder reconstructs the low-dimensional user financial preference change features and low-dimensional recommendation reason features to restore the semantic information of the input data and obtain the user financial preference change features and recommendation reason features. The recommendation reason features include risk matching features, return potential matching features and other explanatory features. The autoencoder network is trained using the backpropagation algorithm. A cross-entropy loss function is constructed, and the reconstruction error between the input data and the reconstructed data is used as the optimization objective until the reconstruction error reaches its minimum value, at which point the training is complete.

[0012] As one possible implementation, the encoder includes a first input unit, a first hiding unit, and a first output unit, wherein the first hiding unit includes at least a first hiding layer and a second hiding layer; The decoder includes a second input unit, a second hidden unit, and a second output unit. The second hidden unit includes at least decoding a first hidden layer and decoding a second hidden layer. The first hidden layer consists of H1 neurons with the ReLU activation function, the second hidden layer consists of H2 neurons with the ReLU activation function, and so on until the last hidden layer outputs low-dimensional user financial preference change features and low-dimensional recommendation reason features; The second input unit receives low-dimensional user financial management preference change characteristics and low-dimensional recommendation reason characteristics; The first hidden layer is decoded by H3 neurons with ReLU activation function; the second hidden layer is decoded by H4 neurons with ReLU activation function; and so on until the output of the last hidden layer is obtained. The second output unit consists of D neurons with Sigmoid activation function.

[0013] As one possible implementation, the gating network module includes an input unit and a gating network; The input unit receives user features, candidate product features, and context features; The gated network concatenates user features, candidate product features, and context features and calculates weighted explanatory features using normalized weights. Specifically, it concatenates user features, candidate product features, and context features to obtain combined features; performs a nonlinear transformation on the combined features using the ReLU activation function to obtain a hidden layer representation; normalizes the hidden layer representation using the SoftMax function to obtain dynamic weights; and processes the dynamic weights and explanatory features to obtain the weighted explanatory features.

[0014] As one possible implementation, the feature matrix is ​​represented as: I 特征矩阵 =[I 用户理财偏好变化特征 ;I 加权解释特征 ;I 用户特征 ;I 产品特征 ;I 上下文特征 ].

[0015] A financial product recommendation system based on a large language model includes: The data acquisition module acquires user-related data and financial product-related data and performs feature extraction to obtain user-related features and financial product-related features. The user-related features include at least user features and user historical purchase sequences, and the financial product-related features include candidate product features and contextual features. A model building module is used to construct an explanation generation module. This module includes a large language model, an autoencoder network, and a gating network. Based on the large language model, it analyzes user-related features and financial product-related features to obtain changes in user investment preferences and evaluates candidate financial products to obtain recommendation reasons. The autoencoder network receives user investment preference changes and recommendation reasons, performs feature compression and information extraction, and then reconstructs the data to obtain user investment preference change features and recommendation reason features. The gating network concatenates user features, candidate product features, and contextual features, applies a linear rectifier function, and normalizes using a flexible maximum value to obtain weighted explanation features. The recommendation reasons include matching degree results and recommendation explanations. An improved training module is used to merge weighted explanatory features, user features, financial product-related features, user financial preference change features, context features, and recommendation reason features to form a feature matrix. The traditional recommendation pre-training model is then trained using the feature matrix to obtain the improved recommendation model. The result reasoning module processes the data to be recommended based on the improved recommendation model to obtain the financial product recommendation results.

[0016] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0017] A financial product recommendation device based on a large language model includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the method described above.

[0018] This invention, by adopting the above technical solutions, has significant technical effects: By leveraging a large language model to generate user-oriented candidate product recommendation reasons, the correlation between user needs and product features is fully explored. Through a dynamic weight allocation mechanism, the generated explanatory information is input as a key feature into the traditional recommendation model, achieving transparency in the recommendation logic. By deeply integrating the large language model with the traditional recommendation model, the interpretability of the recommendation system is significantly improved, while the recommendation accuracy of the model is further enhanced. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a schematic diagram of the overall structure of the system of the present invention; Figure 3 This is a schematic diagram of the overall process of a specific embodiment of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.

[0022] Example 1: A financial product recommendation method based on a large language model, such as... Figure 1 As shown, it includes the following steps: S100. Obtain user-related data and financial product-related data and perform feature extraction to obtain user-related features and financial product-related features. The user-related features include at least user features and user historical purchase sequences, and the financial product-related features include candidate product features and contextual features. S200. Construct an explanation generation module, which includes a large language model, an autoencoder network, and a gating network. Based on the large language model, it analyzes user-related features and financial product-related features to obtain changes in user financial preferences and evaluates candidate financial products to obtain recommendation reasons. The autoencoder network receives user financial preference changes and recommendation reasons, performs feature compression and information extraction, and then reconstructs the data to obtain user financial preference change features and recommendation reason features. The gating network concatenates user features, candidate product features, and context features, applies a linear rectifier function, and normalizes using a flexible maximum value to obtain weighted explanation features. The recommendation reasons include matching degree results and recommendation explanations. S300: The weighted explanatory features, user features, financial product-related features, user financial preference change features, context features, and recommendation reason features are merged to form a feature matrix. The traditional recommendation pre-training model is then trained using the feature matrix to obtain an improved recommendation model. S400. Based on the improved recommendation model, the data to be recommended is processed to obtain the financial product recommendation results.

[0023] The core innovation of this invention lies in the integration of a large language model and recommendation rationale within the framework of traditional financial product recommendation systems. By leveraging the text understanding and reasoning capabilities of the large language model, clear and reasonable explanations are generated for recommended products, significantly improving the interpretability of traditional recommendation systems. Through a dynamic weight allocation mechanism, the generated explanatory information is input as additional features into the recommendation model, further optimizing the recommendation strategy and simultaneously improving recommendation accuracy. This design not only solves the black-box problem of traditional recommendation systems but also achieves a dual optimization of interpretability and recommendation performance.

[0024] The purpose of the explanation generation module is to construct dynamic features of user financial preferences for each input sample and generate multi-dimensional explanatory vectors (embeddings) for candidate products. These explanatory vectors consist of two parts: first, structured features based on financial domain knowledge; and second, supplementary semantic explanations generated by a large language model. Specifically, this module employs a two-stage processing mechanism: in the first stage, leveraging the text generation capabilities of the large language model, recommendations based on user profiles, historical behavior, and market dynamics are generated that conform to financial semantics; in the second stage, a pre-trained autoencoder network transforms the recommendations generated by the large language model into high-dimensional vector representations. Ultimately, these explanatory embeddings, along with user features, product features, and contextual features, serve as input to a deep neural network recommendation model, achieving semantic enhancement and improved decision interpretability of the recommendation system.

[0025] In one embodiment, the user characteristics include demographic characteristics, user device characteristics, and user activity characteristics; The user's historical purchase sequence includes the user's historical purchases of financial products and corresponding product descriptions; The characteristics of candidate products include at least risk characteristics, wherein the candidate products are those that have passed the initial screening by the recall model; Contextual features include continuous features and labels; Among them, demographic characteristics include at least age, gender, and place of origin; user device characteristics include at least device operating system and device operator; and user activity characteristics include at least platform activity, scenario activity, and user account level.

[0026] Financial product characteristics include risk level, historical maximum drawdown, Sharpe ratio, investment period, etc.; contextual characteristics include market volatility index, regulatory policy labels, etc.; candidate products are the set of candidate products after initial screening by the recall model, and the candidate product characteristics are the corresponding features.

[0027] The final feature matrix is ​​represented as: I 特征矩阵 =[I 用户理财偏好变化特征 ;I 加权解释特征 ;I 用户特征 ;I 产品特征 ;I 上下文特征 ].

[0028] In one embodiment, the large language model includes a user profile unit and a recommendation reason unit; The user profiling unit receives relevant user features and extracts key information; it analyzes the key information to obtain changes in user financial management preferences, and then obtains the trend of changes in user financial management preferences over time. The recommendation reason unit analyzes the relevant characteristics of financial products to obtain various attribute characteristics of candidate products; it then performs a matching degree analysis between the various attribute characteristics of candidate products and changes in users' financial purchase preferences to obtain matching degree results, wherein each matching degree result includes a detailed explanation of the recommendation reason.

[0029] The prompt words used by the large language model are as follows: Role: You are a professional financial analyst who can accurately analyze and generate information on changes in a user's financial product purchase history based on a given list of such purchases. You will also provide detailed reasons for recommending or not recommending candidate financial products, taking into account the user's basic information, candidate product information, and the current market atmosphere. These reasons must cover risk, return, maximum drawdown matching, and other reasonable explanations.

[0030] Skills: Analyzing changes in financial preferences; 1. Carefully examine the given list of users' historical financial product purchases and extract key information, such as the risk level, return performance, and maximum drawdown of products purchased at different times.

[0031] 2. By comparing data from different stages, we can summarize the changing trends of users' financial purchase preferences over time, such as whether risk appetite is rising or falling, and whether expectations for returns have changed.

[0032] Skills: Evaluating candidate products; 1. Conduct in-depth analysis of the risk characteristics, expected return levels, and maximum drawdown of candidate products.

[0033] 2. Perform a matching degree analysis between the various indicators of the candidate products and changes in users' financial purchase preferences.

[0034] Skills: Provide reasons for recommendation; 1. Based on the matching results, explain in detail the reasons for recommending or not recommending, and be sure to comprehensively cover the matching situation of risks, returns, and maximum drawdown. You can also provide reasonable explanations from other aspects such as product characteristics and market environment.

[0035] Output format Changes in user financial management preferences: [Detailed explanation] Reasons for recommendation: Risk matching situation: [Detailed explanation] Benefit matching: [Detailed explanation] Maximum drawdown matching scenario: [Detailed explanation] Other explanations: [If any, please provide details] This also includes some limitations: Answers must be based on the given list of the user's historical purchases of financial products and candidate product information, and irrelevant information must not be included.

[0036] The reasons for recommending or not recommending should be clear, comprehensive, and logical.

[0037] The generated content should be concise and clear, highlighting the key points and avoiding lengthy and complex descriptions.

[0038] All generated content must strictly comply with relevant laws, regulations, and regulatory requirements of the financial industry to ensure compliance. Prohibited terms include: guaranteed principal, sure profit, and no risk.

[0039] In one embodiment, the autoencoder network is a multilayer perceptron network, including an encoder and a decoder; The encoder maps the received user financial preference changes and recommendation reasons to a low-dimensional space to achieve feature compression and information extraction, thereby obtaining low-dimensional user financial preference change features and low-dimensional recommendation reason features. The decoder reconstructs the low-dimensional user financial preference change features and low-dimensional recommendation reason features to restore the semantic information of the input data and obtain the user financial preference change features and recommendation reason features. The recommendation reason features include risk matching features, return potential matching features and other explanatory features. The autoencoder network is trained using the backpropagation algorithm. A cross-entropy loss function is constructed, and the reconstruction error between the input data and the reconstructed data is used as the optimization objective until the reconstruction error reaches its minimum value, at which point the training is complete. The cross-entropy loss function is expressed as follows:

[0040] in, Indicates the location of The probability of an accurate prediction This indicates the reconstruction error.

[0041] Specifically, the encoder is a multi-layer hidden layer encoder, including a first input unit, a first hidden unit and a first output unit, wherein the first hidden unit includes at least a first hidden layer and a second hidden layer; The decoder is a multi-layer hidden layer decoder, including a second input unit, a second hidden unit and a second output unit. The second hidden unit includes at least decoding the first hidden layer and decoding the second hidden layer. The first hidden layer consists of H1 neurons, with the ReLU activation function. The output of the first hidden layer is expressed as: h1 = f1(W1x + b1). The second hidden layer consists of H2 neurons, and the activation function is the ReLU function. The output of the second hidden layer is expressed as: h2 = f2(W2h1 + b2) Among them, W1 ∈ RH1×D represents the first encoding weight matrix, b1 ∈ RH1 represents the first encoding bias vector, f1(z) = max(0, z) represents the ReLU function, x represents the input feature, that is, the change in the user's financial management preference and the recommendation reason, W2 ∈ RH2×H1 represents the second encoding weight matrix, b2 ∈ RH2 represents the second encoding bias vector, f2(z) = max(0, z) represents the ReLU function, and h1 represents the output of the first hidden layer; Until the output of the last hidden layer represents the low-dimensional user financial management preference change feature and the low-dimensional recommendation reason feature, expressed as z ∈ Rd, where d represents the dimension of the low-dimensional feature, and d < D, and z represents the low-dimensional user financial management preference change feature and the low-dimensional recommendation reason feature; The second input unit receives the low-dimensional user financial management preference change feature and the low-dimensional recommendation reason feature; The first decoded hidden layer consists of H3 neurons, and the activation function is the ReLU function. The output of the first decoded hidden layer is expressed as: h1′ = f1′(W1′z + b1′); The second decoded hidden layer consists of H4 neurons, and the activation function is the ReLU function. The output of the second decoded hidden layer is expressed as: h2′ = f2′(W2′h1′ + b2′); Among them, W1′ ∈ RH1×d represents the first decoding weight matrix, b1′ ∈ RH3 represents the first decoding bias vector, W2′ ∈ RH4×H3 represents the second decoding weight matrix, and b2′ ∈ RH4 represents the second decoding bias vector; Until the output of the last hidden layer is obtained, the second output unit consists of D neurons, and the activation function is the Sigmoid function. Then the output of the decoder is expressed as:

[0042] Among them, WD ∈ RD×Hlast′ represents the output weight matrix, and bD ∈ RD represents the output bias vector.

[0043] The encoder maps the input text features to a low-dimensional latent space to achieve feature compression and information extraction; the decoder reconstructs based on the compressed low-dimensional features to restore the semantic information of the input data. The network is trained by the backpropagation algorithm, and the optimization objective is to minimize the reconstruction error between the input data and the reconstructed data, so as to ensure that the learned latent representation can fully capture the essential features of the input data.

[0044] In one specific embodiment, the gated network module includes an input unit and a gated network; the input unit receives user features, candidate product features, and context features; the gated network concatenates the user features, candidate product features, and context features and calculates weighted explanatory features through normalized weights, specifically: concatenating user features, candidate product features, and context features to obtain combined features; performing a nonlinear transformation on the combined features using the ReLU activation function to obtain a hidden layer representation; normalizing the hidden layer representation based on the SoftMax function to obtain dynamic weights; and processing the dynamic weights and explanatory features to obtain weighted explanatory features; The combined features are represented as follows:

[0045] The hidden layer is represented as follows:

[0046] The dynamic weights are represented as follows:

[0047] The weighted interpretation features are represented as follows:

[0048] in, These represent the weight matrices, These represent the bias matrices, These represent user characteristics, candidate product characteristics, and contextual characteristics, respectively. Indicates dynamic weights. Represents combined features, This indicates the weighted interpretation feature. This represents the hidden layer representation. This represents the bias matrix.

[0049] By introducing a gating network to implement a dynamic weight adjustment mechanism, the system can adaptively learn the contribution weights of different explanatory features to the current prediction based on the user's product context. The input unit contains four types of features: user features, candidate product features, contextual features, and explanatory features. The gating network adopts a lightweight design, built on a single-layer multilayer perceptron (MLP), and achieves dynamic adjustment through normalized weight calculation.

[0050] Through the above mechanism, the model can flexibly capture the importance of explanatory features in different user product scenarios, thereby improving the accuracy and interpretability of predictions.

[0051] The training process of this invention can be divided into two stages: The first stage involves training an autoencoder network. The recommendation reasons generated by the large language model are semantically divided into four dimensions: risk matching, benefit matching, maximum drawdown matching, and other supplementary explanations. These structured features are fed into the autoencoder network as input, where the encoder performs feature compression and latent representation learning. The decoder then reconstructs the original input.

[0052] The training objective is to minimize the difference between the input and the reconstructed output to ensure that the autoencoder network can effectively capture the core semantic information of the interpreted text. The network parameters are optimized using the cross-entropy loss function through backpropagation until the model achieves stable performance on the validation set.

[0053] The second stage involves joint training of the recommendation model and the gating network. After the autoencoder network is trained, the parameters of the large language model and the autoencoder network are fixed, and the training focuses on the traditional recommendation model and the gating network. The gating network adopts a lightweight single-layer MLP structure and adaptively learns the contribution weights of different explanatory features to the current user-item interaction scenario through a dynamic weight adjustment mechanism. During training, the AdamOptimizer is used as the optimizer, and its adaptive learning rate characteristic can effectively adjust the parameter update pace, improving training efficiency and stability. Based on the loss value calculated by forward propagation, the parameters of the recommendation model and the gating network are updated through backpropagation.

[0054] The detailed training process is as follows: Overall input: dataset Pre-trained large model LLM, learning rate η, number of training epochs E, batch size B; Recommendation reason generation: For each sample implement; User profile enhancement: Using LLM to enrich user profiles with historical purchase sequences and candidate product features; Prepare prompts: Connect the reasons for recommendation and changes in financial preferences to generate prompts and link them with the user's historical purchase sequence and candidate products; Generating Recommendation Reasons and Changes in Investment Preferences: Using LLM to generate recommendation reasons and changes in investment preferences across four dimensions: ( ), (Matching profit potential) (Maximum drawdown matching) (Other possible reasons) (Changes in investment preferences); Calculate the loss: calculate separately , , , , ; Loop ends Fine-tuning the autoencoder: The autoencoder AE(·) is trained by minimizing the loss function as follows: ; Deep neural network recommendation models: Initialization: Trainable parameters: Gated network parameters Traditional recommendation algorithm parameters ; Execute for each round from epoch=1 to E: Shuffle the dataset D; For each batch from b=1 to implement: Select batch ; For each sample implement: Obtain the recommendation reasons and financial preference feature vectors: , , , , ; Gated networks obtain the feature vector weights of the recommendation reasons: ; We obtain the weighted interpretation features: + + + ; Obtaining user characteristics Product features ), context features ( ); The above feature vectors are concatenated to form the input layer of the recommendation model: ; Calculate the predicted values ​​of the recommendation model: ; Loop ends Calculate batch loss: ; Update all trainable parameters: ,in ; Batch cycle ends End of cycle return .

[0055] The training is complete when the optimal parameters of the model are obtained.

[0056] Example 2: A financial product recommendation system based on a large language model, such as... Figure 2As shown, it includes: The data acquisition module 100 acquires user-related data and financial product-related data and performs feature extraction to obtain user-related features and financial product-related features. The user-related features include at least user features and user historical purchase sequences, and the financial product-related features include candidate product features and contextual features. A model construction module 200 is used to construct an explanation generation module. This module includes a large language model, an autoencoder network, and a gating network. Based on the large language model, it analyzes user-related features and financial product-related features to obtain changes in user investment preferences and evaluates candidate financial products to obtain recommendation reasons. The autoencoder network receives user investment preference changes and recommendation reasons, performs feature compression and information extraction, and then reconstructs the data to obtain user investment preference change features and recommendation reason features. The gating network concatenates user features, candidate product features, and contextual features, applies a linear rectifier function, and normalizes using a flexible maximum value to obtain weighted explanation features. The recommendation reasons include matching degree results and recommendation explanations. The improved training module 300 is used to merge weighted explanatory features, user features, financial product-related features, user financial preference change features, context features, and recommendation reason features to form a feature matrix. The traditional recommendation pre-training model is then trained using the feature matrix to obtain the improved recommendation model. The result reasoning module 400 processes the data to be recommended based on the improved recommendation model to obtain the financial product recommendation results.

[0057] Various changes and modifications made without departing from the spirit and scope of this invention, and all equivalent technical solutions, also fall within the scope of this invention.

[0058] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] It should be noted that: The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0064] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.

Claims

1. A financial product recommendation method based on a large language model assistance, characterized in that, The method comprises the following steps: obtaining user-related data and financial product-related data and performing feature extraction to obtain user-related features and financial product-related features, wherein the user-related features at least include user features and user historical purchase sequences, and the financial product-related features include candidate product features and context features; constructing an explanation generation module, wherein the explanation generation module includes a large language model, an autoencoder network and a gating network, analyzing the user-related features and the financial product-related features based on the large language model to obtain user financial preference changes, and evaluating the candidate financial products to obtain recommendation reasons; the autoencoder network receives the user financial preference changes and the recommendation reasons, performs feature compression and information extraction, and then reconstructs to obtain user financial preference change features and recommendation reason features; the gating network splices the user features, the candidate product features and the context features, performs a linear rectifier function and flexible maximum value normalization to obtain weighted explanation features, wherein the recommendation reasons include matching degree results and recommendation explanations; merging the weighted explanation features, the user features, the financial product-related features, the user financial preference change features, the context features and the recommendation reason features to form a feature matrix, and training a traditional recommendation pre-training model through the feature matrix to obtain an improved recommendation model; processing to-be-recommended data based on the improved recommendation model to obtain a financial product recommendation result.

2. The method of claim 1, wherein, The user features include demographic attribute features, user device features and user activity features; the user historical purchase sequences include user historical purchase financial products and corresponding product introductions; the candidate product features at least include risk features, wherein the candidate products are candidate products after preliminary screening by a recall model; the context features include continuous features and labels; wherein the demographic attribute features at least include age, gender and place of origin; the user device features at least include device operating systems and device operators; and the user activity features at least include platform activity, scene activity and user account levels. 3.The method of claim 1, wherein, The large language model includes a user portrait unit and a recommendation reason unit; the user portrait unit receives user-related features and extracts key information; analyzes the key information to obtain user financial preference changes, and further obtains trends of user financial preference changes over time; the recommendation reason unit analyzes the financial product-related features to obtain various attribute features of the candidate products; performing matching degree analysis on the various attribute features of the candidate products and the user financial purchase preference changes to obtain matching degree results, wherein each matching degree result includes detailed elaboration of the recommendation reasons.

4. The method of claim 1, wherein, The autoencoder network is a multi-layer perceptron network, including an encoder and a decoder; the encoder maps the received user financial preference changes and the recommendation reasons to a low-dimensional space to realize feature compression and information extraction, and obtains low-dimensional user financial preference change features and low-dimensional recommendation reason features; and the decoder reconstructs the low-dimensional user financial preference change features and the low-dimensional recommendation reason features to obtain user financial preference change features and recommendation reason features. The decoder reconstructs the low-dimensional user financial preference change feature and the low-dimensional recommendation reason feature to restore the semantic information of the input data, to obtain the user financial preference change feature and the recommendation reason feature, and the recommendation reason feature includes a risk matching feature, a yield potential matching feature and other explanation features; The autoencoder network is trained through a back propagation algorithm, a cross-entropy loss function is constructed, and a reconstruction error between the input data and the reconstructed data is taken as an optimization target until the reconstruction error reaches a minimum value, and then the training is completed.

5. The method of claim 4, wherein, The encoder includes a first input unit, a first hidden unit and a first output unit, and the first hidden unit includes at least a first hidden layer and a second hidden layer; The decoder includes a second input unit, a second hidden unit and a second output unit, and the second hidden unit includes at least a decoding first hidden layer and a decoding second hidden layer; The first layer hidden layer includes H1 neurons, and the activation function is a ReLU function; the second layer hidden layer includes H2 neurons, and the activation function is a ReLU function, and until the last layer hidden layer outputs the low-dimensional user financial preference change feature and the low-dimensional recommendation reason feature; The second input unit receives the low-dimensional user financial preference change feature and the low-dimensional recommendation reason feature; The decoding first hidden layer includes H3 neurons, and the activation function is a ReLU function; the decoding second hidden layer includes H4 neurons, and the activation function is a ReLU function; and until the output of the last layer hidden layer, the second output unit includes D neurons, and the activation function is a Sigmoid function.

6. The large language model-based auxiliary financial product recommendation method according to claim 1, characterized in that, The gating network module includes an input unit and a gating network; The input unit receives user features, candidate product features and context features; The gating network splices the user features, the candidate product features and the context features and calculates through normalized weights to obtain weighted explanation features, specifically: splicing the user features, the candidate product features and the context features to obtain combined features; performing nonlinear transformation on the combined features through a ReLU function activation function to obtain a hidden layer representation; performing normalization processing on the hidden layer representation based on a SoftMax function to obtain dynamic weights; and processing the dynamic weights and the explanation features to obtain the weighted explanation features.

7. The method of claim 1, wherein, The feature matrix is represented as: I 特征矩阵 = [I 用户理财偏好变化特征 ;I 加权解释特征 ;I 用户特征 ;I 产品特征 ;I 上下文特征 ].

8. A financial product recommendation system based on large language model assistance, characterized in that, It includes: A data acquisition module acquires user-related data and financial product-related data and performs feature extraction to obtain user-related features and financial product-related features, and the user-related features at least include user features and user historical purchase sequences, and the financial product-related features include candidate product features and context features; A model construction module is used to construct an explanation generation module, and the explanation generation module includes a large language model, an autoencoder network and a gating network, analyzes the user-related features and the financial product-related features based on the large language model to obtain user financial preference changes, and evaluates candidate financial products to obtain recommendation reasons; The autoencoder network receives user financial preference changes and recommendation reasons, performs feature compression and information extraction, and then reconstructs to obtain user financial preference change features and recommendation reason features; The gating network splices the user features, candidate product features and context features, linear rectifier functions and flexible maximum value normalization to obtain weighted explanation features, wherein the recommendation reasons include matching degree results and recommendation explanations; The improved training module is used for merging the weighted explanation features, user features, financial product related features, user financial preference change features, context features and recommendation reason features to form a feature matrix, and training a traditional recommendation pre-training model through the feature matrix to obtain an improved recommendation model; The result reasoning module processes the to-be-recommended data based on the improved recommendation model to obtain a financial product recommendation result.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the method of any one of claims 1 to 7.

10. A financial product recommendation device based on a large language model assistance, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1 to 7.

Images