Financial product determination method and device, program product and electronic equipment

Abstract

The invention discloses a financial product determination method and device, a program product and electronic equipment, and relates to the field of financial science and technology, and the method comprises the steps: obtaining user information of a target user after the authorization of the target user is obtained, carrying out the feature extraction of the user information through employing a singular value decomposition algorithm, and obtaining a feature extraction result; processing the extracted user hidden feature vectors and financial product hidden feature vectors to obtain a first prediction score corresponding to each financial product, and processing the user information, the user hidden feature vectors and the financial product hidden feature vectors by adopting a preset determination structure to obtain a second prediction score corresponding to each financial product, and based on the first prediction score and the second prediction score corresponding to each financial product, determining a target score corresponding to each financial product by adopting a weight calculation structure, and determining the financial product indicated by the maximum target score as a target financial product. According to the method and the device, the technical problem of relatively low accuracy of determining the financial product in related technologies is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of financial technology, in particular to a financial product determination method and device, program product and electronic device. BACKGROUND

[0002] In the current financial market, there are various types of investment products, including funds, bonds, financial products, insurance and stocks, etc. Different products have great differences in yield, risk level and investment cycle, etc. For non-financial practitioners, due to the lack of professional financial knowledge and risk identification ability, it is difficult to make a reasonable choice among the numerous investment products, resulting in low investment decision-making efficiency, strong blindness and unstable income.

[0003] In related technologies, collaborative filtering or content-based determination algorithms are generally used to determine financial products for users. Although this method can achieve personalized determination of financial products to a certain extent, it has the following problems: (1) the user feature and product feature data are high-dimensional and sparse, resulting in insufficient model generalization ability; (2) the traditional algorithm is difficult to capture the nonlinear relationship between user preferences and product attributes; (3) poor adaptability to the "cold start" problem of new users or new products, resulting in insufficient determination accuracy.

[0004] At present, there is no effective solution to the above problems. SUMMARY

[0005] The embodiments of the present application provide a financial product determination method and device, program product and electronic device to at least solve the technical problem of low accuracy in determining financial products in related technologies.

[0006] According to one aspect of the present invention, a method for determining a financial product is provided, comprising: obtaining user information of the target user with authorization from the target user, wherein the user information includes at least: categorical data and numerical data, wherein categorical data refers to data with different categories and numerical data refers to data with specific values; extracting features from the user information using a singular value decomposition algorithm to obtain a user latent feature vector and a financial product latent feature vector, and processing the user latent feature vector and the financial product latent feature vector using a singular value decomposition algorithm to obtain a first predicted score corresponding to each financial product; processing the user information, the user latent feature vector, and the financial product latent feature vector using a preset determination structure to obtain a second predicted score corresponding to each financial product, wherein the preset determination structure is a structure in the scoring model; and determining a target score corresponding to each financial product based on the first predicted score and the second predicted score corresponding to each financial product using a weight calculation structure, and determining the financial product indicated by the maximum target score as the target financial product, wherein the weight calculation structure is a structure in the scoring model, and the target financial product is used to be displayed on the target user's terminal client.

[0007] Furthermore, the predefined structure includes at least an embedding layer, a factorization machine structure, a neural network structure, and an output layer. The step of processing user information, user latent feature vectors, and financial product latent feature vectors using this predefined structure to obtain a second predicted score for each financial product includes: extracting features from user information to obtain categorical and numerical features; processing each categorical feature using the embedding layer to obtain an embedding vector; processing all embedding vectors, user latent feature vectors, and financial product latent feature vectors using the factorization machine structure to obtain a first feature; processing all embedding vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors using the neural network structure to obtain a second feature; and determining the second predicted score for each financial product based on the first and second features using the output layer.

[0008] Furthermore, the step of processing all embedded vectors, user latent feature vectors, and financial product latent feature vectors using a factorization machine structure to obtain the first feature includes: combining all embedded vectors, user latent feature vectors, and financial product latent feature vectors in pairs to obtain multiple feature vector combinations; determining the dot product of each feature vector combination; and connecting all dot products based on the factorization machine weights and biases of the factorization machine structure to obtain the first feature.

[0009] Furthermore, the neural network structure includes: multiple fully connected layers, each with an activation function, a fully connected weight matrix, and a fully connected bias. The step of processing all embedding vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors using the neural network structure to obtain the second feature includes: concatenating all embedding vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors to obtain connected feature vectors; processing the connected feature vectors using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the first fully connected layer to obtain initial features; processing the initial features using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the next fully connected layer, until the feature output by the last fully connected layer is obtained, and this feature is represented as the second feature.

[0010] Furthermore, the step of determining the second predicted score for each financial product using an output layer based on the first feature and the second feature includes: processing the first feature using an output layer to obtain a first initial score; processing the second feature using an output layer to obtain a second initial score; and determining the second predicted score based on the first initial score and the second initial score.

[0011] Furthermore, the step of determining the target score for each financial product based on the first predicted score and the second predicted score corresponding to each financial product using a weighted calculation structure includes: processing the first predicted score using an activation function based on the first preset weight and the first preset bias of the weighted calculation structure to obtain the first weight; processing the second predicted score using an activation function based on the second preset weight and the second preset bias of the weighted calculation structure to obtain the second weight; and weighting the first weight, the second weight, the first predicted score, and the second predicted score to obtain the target score.

[0012] Furthermore, before processing user information, user latent feature vectors, and financial product latent feature vectors using a pre-defined structure to obtain the second predicted score for each financial product, the process further includes: constructing an initial scoring model, wherein the model structure of the initial scoring model includes: a pre-defined structure and a weight calculation structure; collecting multiple historical user information sets and processing the historical user information using a singular value decomposition algorithm to obtain the historical user latent feature vector and historical financial product latent feature vector corresponding to each historical user information set; for each financial product, labeling the financial product based on each historical user information set to obtain labeling information, wherein the labeling information includes at least: real... The scoring process involves training an initial scoring model using labeled information, all historical user information, the latent feature vector of each historical user, and the latent feature vector of historical financial products. The training continues until the loss value determined by the loss function is less than a preset loss threshold, yielding the target model parameters. The loss value is determined by the loss function based on the predicted score output by the initial scoring model and the actual score. The target model parameters include at least: factorization machine weights, factorization machine biases, fully connected weight matrix, fully connected biases, first preset weights, first preset biases, second preset weights, and second preset biases. The initial scoring model is then adjusted based on the target model parameters to obtain the final scoring model.

[0013] According to another aspect of the present invention, a device for determining financial products is also provided, comprising: an acquisition unit, configured to acquire user information of a target user upon obtaining authorization from the target user, wherein the user information includes at least: categorical data and numerical data, wherein categorical data refers to data with different categories and numerical data refers to data with specific values; a first processing unit, configured to extract features from the user information using a singular value decomposition algorithm to obtain a user latent feature vector and a financial product latent feature vector, and to process the user latent feature vector and the financial product latent feature vector using a singular value decomposition algorithm to obtain a first predicted score corresponding to each financial product; a second processing unit, configured to process the user information, the user latent feature vector, and the financial product latent feature vector using a preset determination structure to obtain a second predicted score corresponding to each financial product, wherein the preset determination structure is a structure in a scoring model; and a determination unit, configured to determine a target score corresponding to each financial product based on the first predicted score and the second predicted score corresponding to each financial product using a weighted calculation structure, and to determine the financial product indicated by the maximum target score as the target financial product, wherein the weighted calculation structure is a structure in a scoring model, and the target financial product is used to be displayed on the target user's terminal client.

[0014] Furthermore, the pre-defined structure includes at least an embedding layer, a factorization machine structure, a neural network structure, and an output layer. The second processing unit includes: a first extraction module for extracting features from user information to obtain categorical features and numerical features; a first processing module for processing each categorical feature using the embedding layer to obtain an embedding vector; a second processing module for processing all embedding vectors, user latent feature vectors, and financial product latent feature vectors using the factorization machine structure to obtain a first feature; a third processing module for processing all embedding vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors using the neural network structure to obtain a second feature; and a first determination module for determining a second predicted score for each financial product based on the first and second features using the output layer.

[0015] Furthermore, the second processing module includes: a first combination submodule, used to combine all embedded vectors, user latent feature vectors and financial product latent feature vectors in pairs to obtain multiple feature vector combinations; a first determination submodule, used to determine the dot product of each feature vector combination; and a first connection submodule, used to connect all dot products based on the factorization machine weights and factorization machine biases of the factorization machine structure to obtain the first feature.

[0016] Furthermore, the neural network structure includes: multiple fully connected layers, each with an activation function, a fully connected weight matrix, and a fully connected bias; the third processing module includes: a second connection submodule, used to connect all embedded vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors to obtain connected feature vectors; a first processing submodule, used to process the connected feature vectors using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the first fully connected layer to obtain initial features; and a second processing submodule, used to process the initial features using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the next fully connected layer, until the features output by the last fully connected layer are obtained, and these features are represented as second features.

[0017] Furthermore, the first determining module includes: processing the first feature using the output layer to obtain a first initial score; processing the second feature using the output layer to obtain a second initial score; and determining a second predicted score based on the first initial score and the second initial score.

[0018] Further, the determining unit includes: a fourth processing module, used to calculate a first preset weight and a first preset bias based on the weighting structure, and process the first predicted score using an activation function to obtain a first weight; a fifth processing module, used to calculate a second preset weight and a second preset bias based on the weighting structure, and process the second predicted score using an activation function to obtain a second weight; and a sixth processing module, used to perform weighted processing on the first weight, the second weight, the first predicted score, and the second predicted score to obtain a target score.

[0019] Furthermore, the determining device also includes: a first construction module, used to construct an initial scoring model before processing user information, user latent feature vectors, and financial product latent feature vectors using a preset determining structure to obtain a second predicted score corresponding to each financial product, wherein the model structure of the initial scoring model includes: a preset determining structure and a weight calculation structure; a seventh processing module, used to collect multiple historical user information and process the historical user information using a singular value decomposition algorithm to obtain a historical user latent feature vector and a historical financial product latent feature vector corresponding to each historical user information; and a first annotation module, used to annotate each financial product based on each historical user information to obtain annotation information, wherein the annotation information is up to... The system includes at least: the actual rating; a first training module, used to train the initial rating model using labeled information, all historical user information, the historical user latent feature vector corresponding to each historical user information, and the historical financial product latent feature vector, until the loss value determined by the loss function is less than a preset loss threshold, thus obtaining the target model parameters, wherein the loss value is determined by the loss function based on the predicted rating output by the initial rating model and the actual rating; the target model parameters include at least: factorization machine weights, factorization machine biases, fully connected weight matrix, fully connected biases, a first preset weight, a first preset bias, a second preset weight, and a second preset bias; and a first adjustment module, used to adjust the initial rating model based on the target model parameters to obtain the rating model.

[0020] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the method for determining any of the above-mentioned financial products.

[0021] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for determining any of the above-described financial products.

[0022] In this invention, with the authorization of the target user, user information of the target user is obtained. A singular value decomposition (SVD) algorithm is used to extract features from the user information, resulting in latent feature vectors for the user and financial products. SVD is then used to process these latent feature vectors to obtain a first predicted score for each financial product. A preset determination structure is then used to process the user information, user latent feature vectors, and financial product latent feature vectors to obtain a second predicted score for each financial product. Based on the first and second predicted scores for each financial product, a weighted calculation structure is used to determine the target score for each financial product. The financial product indicated by the highest target score is then identified as the target financial product. This solves the technical problem of low accuracy in determining financial products in related technologies.

[0023] This invention employs a hybrid deterministic algorithm, utilizing data acquisition and feature processing to improve accuracy and personalization. This enhances the robustness and dynamic adaptability of the deterministic system, addressing the shortcomings of traditional methods in handling high-dimensional sparse data and cold-start problems. Specifically, after obtaining authorization from the target user, user information, including categorical and numerical data, is acquired. Features are extracted using Singular Value Decomposition (SVD) to obtain latent feature vectors for the user and financial products, and a first predicted score is calculated accordingly. Then, the user information and latent feature vectors are processed using a pre-defined deterministic structure in the scoring model (such as DeepFM (Deep Factorization Machines)) to obtain a second predicted score. Finally, the weighted calculation structure in the scoring model combines the first and second predicted scores to generate a target score. The financial product with the highest target score is selected as the final determination result, enabling personalized display on the user's terminal. This not only improves the efficiency and accuracy of the deterministic system but also effectively alleviates the cold-start problem faced by new users or new products. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0025] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for determining financial products is shown.

[0026] Figure 2 This is a flowchart of a method for determining financial products according to Embodiment 1 of the present invention;

[0027] Figure 3 This is a schematic diagram of an optional DeepFM-based financial product determination system according to an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of an optional DeepFM-based financial product identification process according to an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram of an optional financial product determination device according to an embodiment of the present invention;

[0030] Figure 6 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] It should be noted that all relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected and involved in this invention are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data comply with the relevant laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and it does not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface. The relevant information is obtained only after receiving consent from the aforementioned user or organization. If the user chooses to refuse, the process proceeds to an expert decision-making process.

[0034] This invention proposes a financial product identification method based on DeepFM, which improves the accuracy, personalization, and robustness of product identification. In this invention, the DeepFM model effectively learns high-order feature interaction relationships, and when combined with SVD latent preference representation, it significantly improves the prediction accuracy of investment product identification. Furthermore, through joint modeling of explicit and implicit semantic features, it achieves precise matching for users with different risk preferences, investment periods, and return targets, making the identification results more personalized. In addition, DeepFM reduces its reliance on large amounts of historical interaction data through feature cross-learning, allowing for reasonable identification of new users and new products, thus solving the cold start and sparsity problems. Simultaneously, through online learning and feedback mechanisms, the model can automatically adjust the identification weights based on market conditions and user behavior, achieving continuous optimization.

[0035] The present invention will now be described in detail with reference to various embodiments.

[0036] Example 1

[0037] According to an embodiment of this application, an embodiment of a method for determining financial products is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0038] The method embodiment provided in Embodiment 1 of this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing a method for determining financial products is shown. Figure 1 As shown, computer terminal 10 (or mobile device) may include one or more ( Figure 1The processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions may also be included. In addition, it may include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera, wherein the network interface can be connected to wired and / or wireless networks. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0039] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0040] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the financial product determination method in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned financial product determination method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0041] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0042] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0043] Under the aforementioned operating environment, this application provides the following: Figure 2 The method for determining the financial products shown. Figure 2 This is a flowchart of a method for determining financial products according to Embodiment 1 of the present invention, as shown below. Figure 2 As shown, the method includes the following steps:

[0044] Step S201: With the authorization of the target user, obtain the user information of the target user, wherein the user information includes at least: categorical data and numerical data. Categorical data refers to data with different categories, and numerical data refers to data with specific values.

[0045] In this embodiment of the invention, while adhering to privacy protection principles, explicit authorization from the target user is obtained to acquire user characteristics and financial product characteristic data (user information). This user information includes categorical data and numerical data. Categorical data refers to data with fixed categories, such as user ID, product ID, risk preference, etc.; numerical data refers to data covering specific values, such as click count, historical investment amount, etc. This data comprehensively reflects the user's personal attributes and investment behavior characteristics.

[0046] Step S202: Use the singular value decomposition algorithm to extract features from user information to obtain latent feature vectors of users and financial products. Then, use the singular value decomposition algorithm to process the latent feature vectors of users and financial products to obtain the first predicted score for each financial product.

[0047] In this embodiment of the invention, the SVD algorithm is used to extract features from user information, forming latent feature vectors (user_factors) and latent feature vectors (item_factors) for financial products. SVD projects high-dimensional data into a low-dimensional space through matrix factorization, capturing the implicit relationships between users and products—those connections that are not easily identifiable from the original data, such as the sensitivity of user risk preferences and the volatility of product returns. The SVD process helps reduce data dimensionality and computational complexity while preserving key user preferences and product attribute information. Then, the latent feature vectors generated by SVD are used to predict ratings for all financial products that the user may be interested in, obtaining a first predicted rating for each financial product.

[0048] Step S203: The user information, user latent feature vector, and financial product latent feature vector are processed using a preset determined structure to obtain the second predicted score for each financial product. The preset determined structure is the structure in the scoring model.

[0049] In this embodiment of the invention, a Deep Factorization Machine (DeepFM) model (i.e., a pre-defined structure) is used to process user information, user latent feature vectors, and financial product latent feature vectors to obtain a second predicted score for each financial product.

[0050] Here, the predefined deterministic structure, as a deep learning framework, is the structure in the scoring model, capable of automatically identifying and learning complex relationships between features. The DeepFM model takes user information, user latent feature vectors, and financial product latent feature vectors as input, and through a series of calculations, outputs a second predicted score for each financial product. By fusing Factorization Machines (FM) and Deep Neural Networks (DNN), DeepFM takes into account both low-order and high-order interactions of features, improving the accuracy and personalization level of deterministic determination.

[0051] Step S204: Based on the first predicted score and the second predicted score corresponding to each financial product, a weighted calculation structure is used to determine the target score corresponding to each financial product, and the financial product indicated by the maximum target score is determined as the target financial product. Here, the weighted calculation structure is the structure in the scoring model, and the target financial product is used to be displayed on the target user's terminal client.

[0052] In this embodiment of the invention, based on the first and second predicted scores, the scores are further fused through a weighted calculation structure in the scoring model to obtain the target score corresponding to each financial product. Here, the weighted calculation structure is an additional neural network layer that dynamically adjusts the contribution of the two scoring sources through learning, thereby optimizing the final determination result. The final generated target score integrates the user's historical behavior, implicit preferences, and product characteristics, enabling a more comprehensive assessment of the user's potential interest in the financial product.

[0053] Then, based on the target score for each financial product, the product with the highest score is selected as the target financial product. This process automatically selects the investment option that best meets the user's needs by comparing the scores of different products, demonstrating the intelligence and personalization of the algorithm. Afterwards, the selected target financial product will be displayed on the user's terminal client, including key information such as product name, expected rate of return, and risk level, facilitating quick understanding and decision-making for the user. Simultaneously, it supports collecting user feedback to continuously optimize the scoring model.

[0054] In summary, the hybrid deterministic algorithm, through data acquisition and feature processing, improves the accuracy and personalization of the determination process, thereby enhancing the robustness and dynamic adaptability of the deterministic system. This addresses the shortcomings of traditional deterministic methods when dealing with high-dimensional sparse data and the cold-start problem. Specifically, after obtaining authorization from the target user, user information containing both categorical and numerical data is acquired. Features are extracted using Singular Value Decomposition (SVD) to obtain latent feature vectors for the user and financial products, and a first predicted score is calculated accordingly. Then, the user information and latent feature vectors are processed using a pre-defined deterministic structure (such as DeepFM) in the scoring model to obtain a second predicted score. Finally, the weighted calculation structure in the scoring model combines the first and second predicted scores to generate a target score. The financial product with the highest target score is selected as the final determination result, enabling personalized display on the user's terminal. This approach not only improves the efficiency and accuracy of the deterministic system but also effectively alleviates the cold-start challenge faced by new users or new products.

[0055] Optionally, the preset determination structure includes at least an embedding layer, a factorization machine structure, a neural network structure, and an output layer. To improve the accuracy of determining the second predicted score corresponding to each financial product, in the financial product determination method provided in Embodiment 1 of this application, feature extraction is performed on user information to obtain categorical features and numerical features; the embedding layer processes each categorical feature to obtain an embedding vector; the factorization machine structure processes all embedding vectors, user latent feature vectors, and financial product latent feature vectors to obtain a first feature; the neural network structure processes all embedding vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors to obtain a second feature; based on the first feature and the second feature, the output layer determines the second predicted score corresponding to each financial product.

[0056] In this embodiment of the invention, input data can be fed into a preset structure to output a second predicted score for each financial product. The input data includes three parts: The first part is classification features: including userId, productId, and RiskPreference. These are all integers representing the category attributes of the user or financial product. userId represents the user identifier, productId represents the identifier of the investment product (such as a fund, bond, or stock), and RiskPreference represents the user's risk preference level (e.g., low risk, medium risk, high risk). The second part is numerical features: including clicks and InvestmentAmount. Clicks represent the number of times the user clicks on the product, and InvestmentAmount represents the user's historical investment amount (in yuan). The third part is the SVD latent feature vector: including user_factors and item_factors, representing the user latent feature vector (capturing user investment preferences, such as sensitivity to returns or risks) and the financial product latent feature vector (capturing product attributes, such as yield fluctuations or market correlations) extracted from SVD decomposition, respectively.

[0057] Here, the predefined structure includes at least: an embedding layer, a factorization machine structure, a neural network structure, and an output layer. The specific processing procedure of the predefined structure for the input data is as follows:

[0058] First, feature extraction can be performed on the collected user information to obtain categorical and numerical features. Categorical features include user ID, financial product ID, and risk preference, reflecting the user's attributes and behavioral categories; while numerical features involve specific numerical indicators such as historical investment amounts, providing a measure of user behavior.

[0059] For each categorical feature, an embedding layer is applied to transform the categorical feature from a high-dimensional discrete form into a low-dimensional continuous vector representation, i.e., an embedding vector. Embedding vectors can more compactly express feature information, helping the model learn the potential relationships between features and reducing computational complexity. This transformation process is accomplished by learning the correspondence between each category and its position in the vector space. For example, a risk preference level feature might be mapped to different points in the vector space, with each point representing the semantic position of different risk preference levels in the model's understanding.

[0060] For example, for all categorical features, the model maps high-dimensional discrete features to low-dimensional dense vectors through an embedding layer. This mapping captures the semantic information of the features and reduces the computational complexity caused by dimensionality. As shown in Equation (1), for each categorical feature... Transformed into a low-dimensional vector through the embedding layer. The embedded vector is flattened into a one-dimensional vector by the Flatten layer so that it can be subsequently input into other parts of the model.

[0061] (1);

[0062] Then, the embedding vectors of all categorical features, user latent feature vectors, and financial product latent feature vectors are input into the factorization machine structure. By calculating the pairwise interactions between features, the first feature is generated. Here, the factorization machine is used to handle sparse and high-dimensional data, and can efficiently capture low-order interactions between features, providing valuable predictive information even with limited data.

[0063] Next, all embedded vectors, numerical features, user latent feature vectors, and financial product latent feature vectors are concatenated and used as input to a neural network structure. This process is then processed through a multi-layer fully connected network to generate a second feature. This process leverages the capabilities of deep learning to automatically learn higher-order nonlinear relationships between features, capturing more complex investment behavior patterns and product attribute combinations, thereby improving the personalization and accuracy of the determination.

[0064] Finally, the first feature generated by the factorization machine is combined with the second feature generated by the neural network, and a second predicted score for each financial product is generated through the output layer (usually a fully connected layer). The output layer integrates all the processed information and outputs a final score prediction. This score prediction combines the information from the interaction of low-order and high-order features, providing a comprehensive evaluation basis for determining the algorithm.

[0065] In this embodiment, by combining the high-order feature learning capabilities of deep learning with the low-order feature interaction capture capabilities of factorization machines, the accuracy and personalization of financial product determination are improved. Even in scenarios with sparse data or new users, reasonable and personalized determinations can be provided through the combined use of embedding layers and SVD latent feature vectors. This comprehensive scoring prediction method not only enhances the model's generalization ability but also effectively alleviates the cold start problem, ensuring that the determination system remains efficient and accurate in the ever-changing financial market.

[0066] To improve the accuracy of determining the first feature, in the method for determining financial products provided in Embodiment 1 of this application, all embedded vectors, user latent feature vectors, and financial product latent feature vectors are combined in pairs to obtain multiple feature vector combinations; the dot product of each feature vector combination is determined; based on the factorization machine weights and factorization machine biases of the factorization machine structure, all dot products are connected to obtain the first feature.

[0067] In this embodiment of the invention, the FM structure (Factor Machine structure) is used to capture low-order interactions between features. For all embedded vectors... (Assuming there are m categorical features) and SVD latent feature vectors, calculate the pairwise dot product between each pair of features. Concatenate all dot product results into a vector, and then pass it through a fully connected layer to generate the output of the FM structure. (First characteristic). As shown in formula (2).

[0068] (2);

[0069] in, This represents the weight vector (factorization machine weights). The vector representing the i-th feature, i = 1, 2, ..., m. Represents the dot product between elements. and These represent the latent feature vectors of users and financial products, respectively. This indicates the bias term (factorization machine bias).

[0070] Specifically, all the classification feature vectors (i.e., embedding vectors) processed by the embedding layer, as well as the user latent feature vectors and financial product latent feature vectors extracted by the SVD algorithm, are combined in pairs to generate multiple combined vectors in order to explore the potential interaction relationships between these features. In this way, even in high-dimensional space, these interactions that are not easy to be intuitively identified can be captured.

[0071] For each combination of feature vectors, their dot product (i.e., inner product) is calculated to obtain a series of dot product results. Dot product calculation can reveal the similarity between two vectors in the feature space. For financial product identification, it can quantify the degree of matching between user features and product features to capture low-order interaction relationships between features.

[0072] Then, all the calculated dot product results are concatenated to form a set of feature vectors, which serve as the input to the factorization machine (FM) structure. The FM structure includes weight vectors and bias terms. Through the FM calculation process, the weight vectors are used to perform a weighted sum of the dot product results, and the bias term is added to generate the first feature. In this way, the FM structure captures the pairwise interaction effects between features, effectively learning the association between users and financial products even in the case of sparse data, thus obtaining preliminary score predictions.

[0073] In this embodiment, categorical and numerical features can be effectively extracted from user information, and the complex low-order interaction relationships between users and financial products can be captured through pairwise combinations and dot product calculations of feature vectors. The introduction of the factorization machine structure not only optimizes the score prediction of feature combinations, but is also particularly suitable for processing high-dimensional sparse data, ensuring the deterministic performance of the model in cold-start scenarios.

[0074] Optionally, the neural network structure includes multiple fully connected layers, each with an activation function, a fully connected weight matrix, and a fully connected bias. To improve the accuracy of determining the second feature, in the method for determining financial products provided in Embodiment 1 of this application, all embedded vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors are concatenated to obtain a connected feature vector. Based on the fully connected weight matrix and fully connected bias corresponding to the first fully connected layer, an activation function is used to process the connected feature vector to obtain an initial feature. Based on the fully connected weight matrix and fully connected bias corresponding to the next fully connected layer, an activation function is used to process the initial feature until the feature output by the last fully connected layer is obtained, and the feature is represented as the second feature.

[0075] In this embodiment of the invention, the neural network structure includes multiple fully connected layers, each corresponding to an activation function, a fully connected weight matrix, and a fully connected bias. The neural network structure is used to capture high-order interactions between features. All embedded vectors (from categorical features), SVD latent feature vectors, and numerical features are concatenated into a single vector. Processed step by step through multiple fully connected layers Each layer includes a linear transformation, an activation function, and a bias, as shown in Equation (3).

[0076] (3);

[0077] in, Indicates the second characteristic, This represents the concatenated input vector. This represents the weight matrix for each layer. This indicates the offset of each layer. The ReLU (Rectified Linear Unit) activation function is shown in Equation (4).

[0078] (4);

[0079] Where x represents the feature vector input to each fully connected layer.

[0080] Specifically, all embedded vectors, numerical features, and latent feature vectors of users and financial products are concatenated to form an integrated feature vector containing all key information, denoted as the concatenated feature vector. This concatenated feature vector is then input into the first fully connected layer of a Deep Neural Network (DNN), where it undergoes linear transformation using weight matrices and bias vectors, followed by activation functions (e.g., ReLU) to generate initial features. These initial features are then fed into the next layer and subsequent layers of the DNN, repeating the linear transformation and activation function process until the final fully connected layer. Each layer is equipped with independent weight matrices and bias terms to capture feature combinations at different levels. Multi-layer processing progressively refines and constructs higher-order feature representations, thereby deepening the understanding of the intrinsic relationship between user preferences and financial product attributes and improving the model's prediction accuracy. The final output feature obtained after continuous fully connected layer processing is considered the second feature. This feature integrates all information extracted from the original data, and through multi-level abstraction and non-linear combination in deep learning, forms a highly condensed understanding of user financial product preferences.

[0081] In this embodiment, deep neural networks (DNNs) are used to process user features and financial product features in a multi-level, non-linear manner, significantly enhancing the personalization and accuracy of the determination system. Specifically, DNNs can automatically learn the complex relationships between users and products, providing high-quality rating predictions even for high-dimensional, sparse datasets. This not only fully unlocks the potential value of the data but also effectively avoids overfitting, improves the model's generalization ability, and ensures that the determination results meet user needs in different scenarios. It also enhances the ability to handle the cold start problem for new users or new products. Through deep feature learning, the model possesses stronger dynamic adaptability and optimization potential, providing users with more accurate investment product determination, thus promoting the efficiency and stability of investment decisions.

[0082] To improve the accuracy of determining the second predicted score for each financial product, in the method for determining financial products provided in Embodiment 1 of this application, the output layer processes the first feature to obtain the first initial score; the output layer processes the second feature to obtain the second initial score; and the second predicted score is determined based on the first initial score and the second initial score.

[0083] In this embodiment of the invention, the DeepFM output layer adds the output of the FM structure to the output of the DNN structure, combining the interaction of low-order and high-order features to generate a preliminary prediction result. This combination allows DeepFM to simultaneously utilize the efficiency of FM, the deep modeling capability of DNN, and the user-item implicit relationship extracted by SVD, as shown in Equation (5).

[0084] (5);

[0085] in, This indicates the second predicted score.

[0086] Specifically, by connecting the output (first feature) of the FM structure to the output layer, a final linear transformation is performed using the weights and biases of the output layer to obtain a predicted score associated with the investment product, i.e., the first initial score. The output layer is typically a single-node layer that integrates the feature information processed by all previous layers to provide a specific numerical score prediction. This score represents the model's estimate of the match between a specific user and a particular financial product. The second feature (i.e., the output of the DNN structure) is also passed to the output layer and undergoes the same processing logic as the first initial score, but with different weights and bias parameters to generate the second initial score. This process emphasizes that the model can assess the user's potential interest and match with the financial product from different perspectives, providing diverse inputs for subsequent score fusion. Then, a weighted average of the two score predictions is calculated, with the weights depending on the evaluation and adjustment of the accuracy of each branch's score prediction during model training. That is, the two score predictions are weighted and summed through a weighted mixing layer or an additional linear layer to obtain the second predicted score.

[0087] In this embodiment, by integrating the outputs of DeepFM and DNN, the predictive power and personalization level of the financial product determination system are significantly enhanced. The first and second initial scores quantify the relationship between users and financial products from the perspectives of low-level interaction and high-level abstraction, respectively, while the final second predicted score is a comprehensive reflection of both, balancing the direct correlation and complex interaction of features, and providing a more comprehensive and accurate score prediction.

[0088] To improve the accuracy of determining the target score for each financial product, in the method for determining financial products provided in Embodiment 1 of this application, based on the first preset weight and the first preset bias of the weight calculation structure, an activation function is used to process the first predicted score to obtain the first weight; based on the second preset weight and the second preset bias of the weight calculation structure, an activation function is used to process the second predicted score to obtain the second weight; the first weight, the second weight, the first predicted score, and the second predicted score are weighted to obtain the target score.

[0089] In this embodiment of the invention, it is not directly used Instead of using a weighted mixing mechanism, the prediction results of DeepFM and SVD are combined as the final output. The DeepFM output is weighted through a fully connected layer (with sigmoid activation function) as shown in Equation (6).

[0090] (6);

[0091] in, This indicates the second preset weight. This indicates the second preset bias. This represents the activation function. This indicates the second weight.

[0092] The weights of the SVD prediction score svd_pred (first prediction score) are calculated through another fully connected layer (with sigmoid activation function), as shown in Equation (7).

[0093] (7);

[0094] in, Indicates the first preset weight. Indicates the first preset bias. This represents the activation function. This indicates the first weight.

[0095] Final prediction score (Target score) is the weighted sum of the two, as shown in formula (8).

[0096] (8);

[0097] This weighted hybrid approach dynamically adjusts the contributions of DeepFM and SVD predictions using learned weights. SVD is a classic collaborative filtering method based on user-item rating matrix decomposition, while DeepFM utilizes rich feature information. The weighting mechanism allows the model to adaptively combine the advantages of both.

[0098] In this embodiment, by calculating the weights of the two rating predictions, the adaptive adjustment of the model's trust and contribution to the outputs of different determination algorithms is emphasized. The generation of weights for DeepFM rating prediction and SVD rating prediction, as well as their dynamic fusion when generating the target rating, enable the determination results to more accurately reflect the user's true preferences and the true attributes of financial products, while enhancing the model's ability to handle data sparsity and cold start problems.

[0099] To improve the accuracy of the scoring model, in the method for determining financial products provided in Embodiment 1 of this application, before processing user information, user latent feature vectors, and financial product latent feature vectors using a preset determination structure to obtain the second predicted score corresponding to each financial product, an initial scoring model is constructed. The initial scoring model's structure includes: a preset determination structure and a weight calculation structure; collecting multiple historical user information sets and processing the historical user information using a singular value decomposition algorithm to obtain the historical user latent feature vector and historical financial product latent feature vector corresponding to each historical user information set; for each financial product, labeling the financial product based on each historical user information set to obtain labeling information, wherein... The annotation information includes at least: the actual rating; the initial rating model is trained using the annotation information, all historical user information, the historical user latent feature vector corresponding to each historical user information, and the historical financial product latent feature vector until the loss value determined by the loss function is less than a preset loss threshold, thus obtaining the target model parameters. The loss value is determined by the loss function based on the predicted rating output by the initial rating model and the actual rating; the target model parameters include at least: factorization machine weights, factorization machine biases, fully connected weight matrix, fully connected biases, first preset weights, first preset biases, second preset weights, and second preset biases; the initial rating model is adjusted based on the target model parameters to obtain the rating model.

[0100] In this embodiment of the invention, the scoring model is trained using the Adam (Adaptive Moment Estimation) optimizer, and the loss function is the Mean Squared Error (MSE). MSE can effectively measure the difference between the predicted value and the true value, as shown in Equation (9).

[0101] (9);

[0102] in, These are real ratings. This is the predicted score, where n represents the number of samples.

[0103] Specifically, an initial scoring model can be constructed first, comprising a predefined structure and a weight calculation structure. The predefined structure mainly involves the architectural design of the feature embedding layer, the FM part, and the DNN part, including the type, number, and parameter configuration of the layers. The weight calculation structure is set up for generating score prediction weights later, and contains two fully connected layers (used to handle DeepFM and SVD score predictions respectively). The model parameters (such as the weight matrix and bias vector) are initialized to random values ​​or samples from a predefined distribution to prepare for the subsequent training process.

[0104] Historical user information is collected, including transaction records, user attributes, and product preferences, and then processed using the Singular Value Decomposition (SVD) algorithm. SVD decomposes the user and product rating matrix into user latent feature matrices and product latent feature matrices, yielding historical user latent feature vectors and historical financial product latent feature vectors. These vectors can capture the potential correlations and preference patterns between users and products.

[0105] For each financial product, annotation information, i.e., the true rating, is added based on users' ratings, clicks, purchase records, and other historical user information. This annotation information is a key input during model training, allowing the model to learn and optimize its parameters by comparing predicted results with the true ratings.

[0106] Then, the labeled information, all historical user information, historical user latent feature vectors, and historical financial product latent feature vectors are used as inputs. The model parameters are adjusted through an iterative training process until the loss value determined by the loss function is lower than a preset loss threshold. The loss function (such as mean squared error, MSE) measures the difference between the model's predicted score and the actual score and is the optimization objective during model training.

[0107] After the model training converges, the model parameters at this point are recorded, including the factorization machine weights, factorization machine biases, fully connected weight matrix, fully connected bias, first preset weights, first preset bias, second preset weights, and second preset bias, which constitute the target model parameters. These parameters have been thoroughly learned from the training data and can effectively predict users' ratings of financial products.

[0108] Next, the target model parameters obtained during training are applied to the initial scoring model, replacing the original randomly initialized parameters, thereby optimizing the model. The adjusted scoring model possesses the learned feature interaction capabilities and user-product preference matching capabilities, enabling it to provide users with more accurate investment product identification.

[0109] In this embodiment, a scoring model based on DeepFM and SVD is constructed. From data preprocessing and feature extraction to model training and parameter optimization, each step can improve the predictive accuracy and personalization level of the system. The model fully utilizes the latent feature vectors of users and products extracted by SVD, combined with the feature cross-validation capability of DeepFM, to effectively handle high-dimensional sparse data and the cold-start problem, providing stable and accurate determinations. By minimizing the loss function, the model continuously adjusts its parameters during training to ensure that the predicted score is as close as possible to the true score, thereby providing users with financial product determinations that better match their needs and preferences in practical applications.

[0110] Figure 3 This is a schematic diagram of an optional DeepFM-based financial product determination system according to an embodiment of the present invention, such as... Figure 3 As shown, it includes: a data acquisition module, a feature processing module, a model determination module, a model training and optimization module, and a model output module. The data acquisition module acquires raw data; the feature processing module preprocesses the raw data and generates input features; the model determination module generates a definitive score based on DeepFM and SVD; the model training and optimization module continuously optimizes model performance; and finally, the model output module presents the definitive result to the user.

[0111] Specifically, (1) Data Acquisition Module: This module is used to collect user characteristic and financial product characteristic data. This includes: categorical features (such as user ID, product ID, risk preference), numerical features (such as click count, historical investment amount), and latent feature vectors (user latent features and product latent features) generated through Singular Value Decomposition (SVD). This module obtains real-time user behavior data and product information from the financial platform through a database interface or API (Application Programming Interface) to ensure the integrity and real-time nature of the input data.

[0112] (2) Feature processing module: responsible for preprocessing and feature engineering of input data. Specifically, it includes: classification feature embedding, numerical feature normalization and SVD feature generation.

[0113] (3) Deterministic model: A weighted hybrid deterministic model of DeepFM and SVD is adopted.

[0114] (4) Model Training and Optimization Module: Responsible for model training and parameter optimization. Gradient descent training is performed using the Adam optimizer, with mean squared error (MSE) as the loss function to minimize the difference between the predicted and actual scores. Online learning functionality is supported, allowing for dynamic updates of model parameters based on real-time user feedback.

[0115] (5) Output Module: Based on the scores generated by the model, the Top-N (first N) results are selected from the financial product database and displayed to the user through the user interface. The results include information such as product name, expected rate of return, and risk level, and support user interaction and feedback collection.

[0116] Figure 4 This is a schematic diagram of an optional DeepFM-based financial product identification process according to an embodiment of the present invention, such as... Figure 4 As shown, the process includes the following:

[0117] (1) Extract features from the collected data to obtain input features, including: classification features, numerical features and SVD latent feature vectors.

[0118] (2) The high-dimensional discrete features are mapped to low-dimensional dense vectors through the Embedding layer. This mapping can capture the semantic information of the features and reduce the computational complexity caused by the dimensionality.

[0119] (3) The FM part in the DeepFM model is used to capture low-order interaction relationships between features, including: FM interaction and FM output.

[0120] (4) The DNN part in the DeepFM model is used to capture the high-order interaction relationship between features, including: DNN input, DNN layer, and DNN output.

[0121] (5) The DeepFM output in the DeepFM model adds the output of the FM part to the output of the DNN part, integrates the interaction of low-order and high-order features, and generates a preliminary prediction result (including DeepFM prediction score).

[0122] (6) Perform DeepFM weight calculation (to obtain the weight of DeepFM prediction score) and SVD weight calculation (to obtain the weight of SVD prediction score) through the weight calculation and weighted fusion module, and then obtain the target score through weighted fusion.

[0123] The financial product determination method provided in this application adopts a hybrid determination algorithm. Through data acquisition and feature processing, it improves the accuracy and personalization of determination, thereby enhancing the robustness and dynamic adaptability of the determination system. This solves the shortcomings of traditional determination methods when facing high-dimensional sparse data and cold-start problems. Specifically, after obtaining authorization from the target user, user information containing categorical and numerical data is acquired. Features are extracted using the Singular Value Decomposition (SVD) algorithm to obtain latent feature vectors for the user and financial products, and a first predicted score is calculated accordingly. Then, the user information and latent feature vectors are processed through a preset determination structure (such as DeepFM) in the scoring model to obtain a second predicted score. Subsequently, the weight calculation structure in the scoring model is used to combine the first and second predicted scores to generate a target score. The financial product with the highest target score is selected as the determination result, achieving personalized display on the user's terminal. In this way, not only is the efficiency and accuracy of the determination system improved, but the cold-start problem faced by new users or new products is also effectively alleviated.

[0124] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0125] Example 2

[0126] This application also provides a device for determining financial products. It should be noted that this device can be used to execute the method for determining financial products provided in this application. The following describes the device for determining financial products provided in this application.

[0127] According to an embodiment of this application, an apparatus for implementing the above-described method for determining financial products is also provided. Figure 5 This is a schematic diagram of an optional financial product determination device according to an embodiment of the present invention, such as... Figure 5 As shown, the determining device may include: an acquisition unit 50, a first processing unit 51, a second processing unit 52, and a determining unit 53.

[0128] The acquisition unit 50 is used to acquire user information of the target user when the target user authorizes it. The user information includes at least: categorical data and numerical data. Categorical data refers to data with different categories, and numerical data refers to data with specific values.

[0129] The first processing unit 51 is used to extract features from user information using the singular value decomposition algorithm to obtain user latent feature vectors and financial product latent feature vectors, and to process the user latent feature vectors and financial product latent feature vectors using the singular value decomposition algorithm to obtain the first predicted score corresponding to each financial product.

[0130] The second processing unit 52 is used to process user information, user latent feature vector and financial product latent feature vector using a preset determined structure to obtain a second predicted score for each financial product, wherein the preset determined structure is the structure in the scoring model.

[0131] The determining unit 53 is used to determine the target score corresponding to each financial product based on the first predicted score and the second predicted score corresponding to each financial product, using a weight calculation structure, and to determine the financial product indicated by the maximum target score as the target financial product. The weight calculation structure is a structure in the scoring model, and the target financial product is used to be displayed on the target user's terminal client.

[0132] The financial product determination apparatus provided in this application adopts a hybrid determination algorithm. Through data acquisition and feature processing, it improves the accuracy and personalization of determination, thereby enhancing the robustness and dynamic adaptability of the determination system. This solves the shortcomings of traditional determination methods when facing high-dimensional sparse data and cold-start problems. Specifically, after obtaining authorization from the target user, user information containing categorical and numerical data is acquired. Features are extracted using the Singular Value Decomposition (SVD) algorithm to obtain the user's latent feature vector and the financial product's latent feature vector, and a first predicted score is calculated accordingly. Then, the user information and latent feature vector are processed through a preset determination structure (such as DeepFM) in the scoring model to obtain a second predicted score. Subsequently, the weight calculation structure in the scoring model is used to combine the first and second predicted scores to generate a target score. The financial product with the highest target score is selected as the determination result, realizing personalized display on the user terminal. In this way, not only is the efficiency and accuracy of the determination system improved, but the cold-start problem faced by new users or new products is also effectively alleviated.

[0133] Optionally, the predefined structure includes at least an embedding layer, a factorization machine structure, a neural network structure, and an output layer. The second processing unit includes: a first extraction module for extracting features from user information to obtain categorical features and numerical features; a first processing module for processing each categorical feature using the embedding layer to obtain an embedding vector; a second processing module for processing all embedding vectors, user latent feature vectors, and financial product latent feature vectors using the factorization machine structure to obtain a first feature; a third processing module for processing all embedding vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors using the neural network structure to obtain a second feature; and a first determination module for determining a second predicted score for each financial product based on the first and second features using the output layer.

[0134] Optionally, the second processing module includes: a first combination submodule, used to combine all embedded vectors, user latent feature vectors and financial product latent feature vectors in pairs to obtain multiple feature vector combinations; a first determination submodule, used to determine the dot product of each feature vector combination; and a first connection submodule, used to connect all dot products based on the factorization machine weights and factorization machine biases of the factorization machine structure to obtain the first feature.

[0135] Optionally, the neural network structure includes: multiple fully connected layers, each fully connected layer corresponding to an activation function, a fully connected weight matrix, and a fully connected bias; the third processing module includes: a second connection submodule, used to connect all embedded vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors to obtain connected feature vectors; a first processing submodule, used to process the connected feature vectors using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the first fully connected layer to obtain initial features; and a second processing submodule, used to process the initial features using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the next fully connected layer, until the features output by the last fully connected layer are obtained, and the features are represented as second features.

[0136] Optionally, the first determining module includes: processing the first feature using the output layer to obtain a first initial score; processing the second feature using the output layer to obtain a second initial score; and determining a second predicted score based on the first initial score and the second initial score.

[0137] Optionally, the determining unit includes: a fourth processing module, used to calculate a first preset weight and a first preset bias based on the weighting structure, and process the first predicted score using an activation function to obtain a first weight; a fifth processing module, used to calculate a second preset weight and a second preset bias based on the weighting structure, and process the second predicted score using an activation function to obtain a second weight; and a sixth processing module, used to perform weighted processing on the first weight, the second weight, the first predicted score, and the second predicted score to obtain a target score.

[0138] Optionally, the determining device further includes: a first construction module, used to construct an initial scoring model before processing user information, user latent feature vectors, and financial product latent feature vectors using a preset determining structure to obtain a second predicted score corresponding to each financial product, wherein the model structure of the initial scoring model includes: a preset determining structure and a weight calculation structure; a seventh processing module, used to collect multiple historical user information and process the historical user information using a singular value decomposition algorithm to obtain a historical user latent feature vector and a historical financial product latent feature vector corresponding to each historical user information; and a first annotation module, used to annotate each financial product based on each historical user information to obtain annotation information, wherein the annotation information is up to... The system includes at least: the actual rating; a first training module, used to train the initial rating model using labeled information, all historical user information, the historical user latent feature vector corresponding to each historical user information, and the historical financial product latent feature vector, until the loss value determined by the loss function is less than a preset loss threshold, thus obtaining the target model parameters, wherein the loss value is determined by the loss function based on the predicted rating output by the initial rating model and the actual rating; the target model parameters include at least: factorization machine weights, factorization machine biases, fully connected weight matrix, fully connected biases, a first preset weight, a first preset bias, a second preset weight, and a second preset bias; and a first adjustment module, used to adjust the initial rating model based on the target model parameters to obtain the rating model.

[0139] The aforementioned determining device may further include a processor and a memory. The aforementioned acquisition unit 50, first processing unit 51, second processing unit 52, determining unit 53, etc., are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0140] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, a weighted calculation structure is used to determine the target score for each financial product based on the first and second predicted scores corresponding to each financial product. The financial product indicated by the highest target score is then identified as the target financial product.

[0141] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0142] It should be noted that the acquisition unit 50, the first processing unit 51, the second processing unit 52, and the determination unit 53 mentioned above correspond to steps S201 to S204 in Embodiment 1. The instances and application scenarios implemented by the above units and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above units can be hardware or software components stored in memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The above units can also be part of a device and can run in the computer terminal 10 provided in Embodiment 1.

[0143] Example 3

[0144] Embodiments of this application may provide an electronic device. Figure 6 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Figure 6 As shown, the electronic device may include: one or more ( Figure 6 (Only one is shown) Processor 602, memory 604, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0145] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the financial product determination method and apparatus in this application embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned financial product determination method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0146] The processor can access information and applications stored in memory via a transmission device to perform the following steps: With authorization from the target user, obtain the target user's user information, which includes at least: categorical data and numerical data. Categorical data refers to data with different categories, and numerical data refers to data with specific values. Use a singular value decomposition (SVD) algorithm to extract features from the user information, obtaining user latent feature vectors and financial product latent feature vectors. Process these latent feature vectors using the SVD algorithm to obtain a first predicted score for each financial product. Process the user information, user latent feature vectors, and financial product latent feature vectors using a pre-defined structure to obtain a second predicted score for each financial product. The pre-defined structure is a structure within the scoring model. Based on the first and second predicted scores for each financial product, use a weighted calculation structure to determine the target score for each financial product, and identify the financial product with the highest target score as the target financial product. The weighted calculation structure is a structure within the scoring model, and the target financial product is used for display on the target user's terminal client.

[0147] The processor can access information and applications stored in memory via a transmission device to perform the following steps: extracting features from user information to obtain categorical and numerical features; processing each categorical feature using an embedding layer to obtain an embedding vector; processing all embedding vectors, user latent feature vectors, and financial product latent feature vectors using a factorization machine structure to obtain a first feature; processing all embedding vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors using a neural network structure to obtain a second feature; and determining a second predicted score for each financial product using an output layer based on the first and second features.

[0148] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: combine all embedded vectors, user latent feature vectors, and financial product latent feature vectors in pairs to obtain multiple feature vector combinations; determine the dot product of each feature vector combination; and connect all dot products based on the factorization machine weights and biases of the factorization machine structure to obtain the first feature.

[0149] The processor can access information and applications stored in the memory via a transmission device to perform the following steps: concatenate all embedded vectors, all numerical features, user latent feature vectors, and financial product latent feature vectors to obtain concatenated feature vectors; process the concatenated feature vectors using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the first fully connected layer to obtain initial features; process the initial features using an activation function based on the fully connected weight matrix and fully connected bias corresponding to the next fully connected layer until the features output by the last fully connected layer are obtained, and these features are represented as second features.

[0150] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: process the first feature using the output layer to obtain a first initial score; process the second feature using the output layer to obtain a second initial score; and determine a second predicted score based on the first initial score and the second initial score.

[0151] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: based on the first preset weight and the first preset bias of the weight calculation structure, process the first predicted score using an activation function to obtain the first weight; based on the second preset weight and the second preset bias of the weight calculation structure, process the second predicted score using an activation function to obtain the second weight; and perform weighted processing on the first weight, the second weight, the first predicted score, and the second predicted score to obtain the target score.

[0152] The processor can access information and applications stored in memory via a transmission device to execute the following steps: Constructing an initial scoring model, wherein the model structure of the initial scoring model includes: a preset determination structure and a weight calculation structure; collecting multiple historical user information sets and processing the historical user information using a singular value decomposition algorithm to obtain the historical user latent feature vector and the historical financial product latent feature vector corresponding to each historical user information set; for each financial product, labeling the financial product based on each historical user information set to obtain labeling information, wherein the labeling information includes at least: the actual score; training the initial scoring model using the labeling information, all historical user information sets, the historical user latent feature vector corresponding to each historical user information set, and the historical financial product latent feature vector, until the loss value determined by the loss function is less than a preset loss threshold, obtaining the target model parameters, wherein the loss value is determined by the loss function based on the predicted score output by the initial scoring model and the actual score; the target model parameters include at least: factorization machine weights, factorization machine biases, fully connected weight matrix, fully connected biases, a first preset weight, a first preset bias, a second preset weight, and a second preset bias; adjusting the initial scoring model based on the target model parameters to obtain the scoring model.

[0153] This application provides a scheme for determining financial products. By employing a hybrid determination algorithm, through data acquisition and feature processing, it improves the accuracy and personalization of the determination process, thereby enhancing the robustness and dynamic adaptability of the determination system. This addresses the shortcomings of traditional determination methods when facing high-dimensional sparse data and cold-start problems. Specifically, after obtaining authorization from the target user, user information containing both categorical and numerical data is acquired. Features are extracted using the Singular Value Decomposition (SVD) algorithm to obtain latent feature vectors for the user and the financial product, and a first predicted score is calculated accordingly. Then, the user information and latent feature vectors are processed using a preset determination structure (such as DeepFM) in the scoring model to obtain a second predicted score. Finally, using the weight calculation structure in the scoring model, the first and second predicted scores are combined to generate a target score. The financial product with the highest target score is selected as the determination result, enabling personalized display on the user's terminal. This not only improves the efficiency and accuracy of the determination system but also effectively alleviates the cold-start problem faced by new users or new products.

[0154] Those skilled in the art will understand that Figure 6 The structure shown is for illustrative purposes only. Electronic devices can also be terminal devices such as smartphones, tablets, PDAs, and mobile internet devices (MIDs). Figure 6 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 6 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 6 The different configurations shown.

[0155] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0156] Example 4

[0157] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the method for determining financial products provided in Embodiment 1.

[0158] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0159] This application also provides a computer program product that, when executed on a data processing device, is suitable for performing the steps of a method for determining a financial product.

[0160] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0161] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0162] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0163] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0164] Furthermore, the functional units in the various embodiments of this application 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 as a software functional unit.

[0165] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0166] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for determining a financial product, characterized in that, include: With the authorization of the target user, the user information of the target user is obtained, wherein the user information includes at least: categorical data and numerical data, wherein the categorical data refers to data with different categories, and the numerical data refers to data with specific values; The user information is used to extract features using the singular value decomposition algorithm to obtain the user latent feature vector and the financial product latent feature vector. The user latent feature vector and the financial product latent feature vector are then processed using the singular value decomposition algorithm to obtain the first predicted score for each financial product. The user information, the user latent feature vector, and the financial product latent feature vector are processed using a preset determined structure to obtain a second predicted score for each financial product, wherein the preset determined structure is the structure in the scoring model; Based on the first predicted score and the second predicted score corresponding to each financial product, a weighted calculation structure is used to determine the target score corresponding to each financial product, and the financial product indicated by the maximum target score is determined as the target financial product. The weighted calculation structure is a structure in the scoring model, and the target financial product is used to be displayed on the target user's terminal client.

2. The determination method according to claim 1, characterized in that, The preset structure includes at least an embedding layer, a factorization machine structure, a neural network structure, and an output layer. The step of processing the user information, the user latent feature vector, and the financial product latent feature vector using the preset structure to obtain a second predicted score for each financial product includes: Feature extraction is performed on the user information to obtain classification features and numerical features; The embedding layer is used to process each of the classification features to obtain an embedding vector; The factorization machine structure is used to process all the embedding vectors, the user latent feature vectors, and the financial product latent feature vectors to obtain the first feature; The neural network structure is used to process all the embedding vectors, all the numerical features, the user latent feature vector, and the financial product latent feature vector to obtain the second feature. Based on the first feature and the second feature, the output layer is used to determine the second predicted score corresponding to each of the financial products.

3. The determination method according to claim 2, characterized in that, The step of processing all the embedding vectors, the user latent feature vectors, and the financial product latent feature vectors using the aforementioned factorization machine structure to obtain the first feature includes: By combining all the embedding vectors, the user latent feature vectors, and the financial product latent feature vectors in pairs, multiple feature vector combinations are obtained. Determine the dot product of each of the said eigenvector combinations; Based on the factorization machine weights and biases of the aforementioned factorization machine structure, all the dot products are connected to obtain the first feature.

4. The determination method according to claim 2, characterized in that, The neural network structure includes multiple fully connected layers, each of which corresponds to an activation function, a fully connected weight matrix, and a fully connected bias. The step of processing all the embedding vectors, all the numerical features, the user latent feature vector, and the financial product latent feature vector using the neural network structure to obtain the second feature includes: By concatenating all the embedded vectors, all the numerical features, the user latent feature vector, and the financial product latent feature vector, a concatenated feature vector is obtained. Based on the fully connected weight matrix and fully connected bias corresponding to the first fully connected layer, the activation function is used to process the connected feature vector to obtain initial features; Based on the fully connected weight matrix and fully connected bias corresponding to the next fully connected layer, the initial features are processed using the activation function until the features output by the last fully connected layer are obtained, and the features are represented as the second feature.

5. The determination method according to claim 2, characterized in that, Based on the first feature and the second feature, the step of determining the second predicted score corresponding to each financial product using the output layer includes: The first feature is processed by the output layer to obtain a first initial score; The output layer is used to process the second feature to obtain a second initial score; The second predicted score is determined based on the first initial score and the second initial score.

6. The determination method according to claim 1, characterized in that, The step of determining the target score for each financial product based on the first predicted score and the second predicted score corresponding to each financial product, using a weighted calculation structure, includes: Based on the first preset weight and the first preset bias of the weight calculation structure, the first predicted score is processed by an activation function to obtain the first weight; Based on the second preset weight and the second preset bias of the weight calculation structure, the second predicted score is processed by an activation function to obtain the second weight; The first weight, the second weight, the first predicted score, and the second predicted score are weighted to obtain the target score.

7. The determination method according to claim 1, characterized in that, Before processing the user information, the user latent feature vector, and the financial product latent feature vector using a preset structure to obtain the second predicted score corresponding to each financial product, the process further includes: Construct an initial scoring model, wherein the model structure of the initial scoring model includes: the preset determination structure and the weight calculation structure; Collect multiple historical user information and process the historical user information using the singular value decomposition algorithm to obtain the historical user latent feature vector and the historical financial product latent feature vector corresponding to each historical user information. For each of the financial products, the financial products are labeled based on each of the historical user information to obtain labeling information, wherein the labeling information includes at least: real ratings; The initial scoring model is trained using the labeled information, all historical user information, the historical user latent feature vector corresponding to each historical user information, and the historical financial product latent feature vector until the loss value determined by the loss function is less than a preset loss threshold, thereby obtaining the target model parameters. The loss value is determined by the loss function based on the predicted score output by the initial scoring model and the actual score. The target model parameters include at least: factorization machine weights, factorization machine biases, fully connected weight matrix, fully connected biases, a first preset weight, a first preset bias, a second preset weight, and a second preset bias. The initial scoring model is adjusted based on the target model parameters to obtain the scoring model.

8. A device for determining a financial product, characterized in that, include: The acquisition unit is used to acquire user information of the target user when the target user authorizes the acquisition, wherein the user information includes at least: categorical data and numerical data, wherein the categorical data refers to data with different categories and the numerical data refers to data with specific values; The first processing unit is used to extract features from the user information using the singular value decomposition algorithm to obtain user latent feature vectors and financial product latent feature vectors, and to process the user latent feature vectors and financial product latent feature vectors using the singular value decomposition algorithm to obtain a first predicted score for each financial product. The second processing unit is used to process the user information, the user latent feature vector, and the financial product latent feature vector using a preset determined structure to obtain a second predicted score for each financial product, wherein the preset determined structure is a structure in the scoring model; The determining unit is configured to determine the target score corresponding to each financial product based on the first predicted score and the second predicted score corresponding to each financial product, using a weighted calculation structure, and determine the financial product indicated by the maximum target score as the target financial product. The weighted calculation structure is a structure in the scoring model, and the target financial product is used to be displayed on the target user's terminal client.

9. A computer program product, characterized in that, The method includes a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for determining a financial product as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for determining a financial product as described in any one of claims 1 to 7.

Images