Financial product recommendation method, device and equipment

Abstract

The embodiment of the invention provides a financial product recommendation method, device and equipment, and relates to the technical field of financial science and technology. The financial product recommendation method comprises the steps that user feature data and financial product feature data are input into N neural collaborative filtering networks of a recommendation strategy network, N initial preference scores of users for different financial products are generated, and N is an integer larger than or equal to 2; performing weighted fusion on the N initial preference scores through a routing network of the recommendation strategy network to generate comprehensive preference scores of the user for different financial products; and determining a sorting result of the recommended financial products according to the comprehensive preference score. According to the method, the accuracy of financial product recommendation can be improved.

Description

Technical Field

[0001] This application relates to the field of financial technology, and in particular to a method, apparatus and equipment for recommending financial products. Background Technology

[0002] In financial settings, personalized financial product recommendations have become a crucial tool for improving user satisfaction and conversion rates. Financial institutions analyze multi-dimensional data such as users' historical transaction records, risk preferences, and financial status to provide customized recommendations for financial products such as loans, investments, and insurance. However, the high degree of privacy and sparse nature of financial institutions' multi-dimensional data pose challenges to traditional financial product recommendation methods. Summary of the Invention

[0003] This application provides a method, apparatus, and device for recommending financial products, in order to improve the accuracy of financial product recommendations.

[0004] In a first aspect, this application provides a financial product recommendation method, comprising: inputting user feature data and financial product feature data into N neural collaborative filtering networks of a recommendation strategy network to generate N initial preference scores for different financial products, where N is an integer greater than or equal to 2; weighting and fusing the N initial preference scores through the routing network of the recommendation strategy network to generate a comprehensive preference score for different financial products; and determining the ranking result of the recommended financial products based on the comprehensive preference score.

[0005] In one possible embodiment, the neural collaborative filtering network includes parallel generalized matrix factorization branches and a first multilayer perceptron. The process of inputting user feature data and financial product feature data into N neural collaborative filtering networks of the recommendation strategy network to generate N initial preference scores for different financial products includes: for the i-th neural collaborative filtering network (i∈[1,N]): performing a linear transformation on the embedding vectors of the user feature data and the financial product feature data through the generalized matrix factorization branch to obtain a first feature vector; concatenating the embedding vectors of the user feature data and the financial product feature data through the first multilayer perceptron, and performing a nonlinear transformation on the concatenated embedding vector to obtain a second feature vector; performing feature fusion processing based on the first and second feature vectors to obtain a fused feature vector; and calculating the i-th initial preference score based on the fused feature vector.

[0006] In one possible embodiment, the input data of the neural collaborative filtering network further includes: external environment data, the embedding vector of which is used to fuse with the embedding vector of the user feature data.

[0007] In one possible embodiment, the routing network includes at least one second multilayer perceptron. The routing network of the recommendation strategy network weights and fuses N initial preference scores to generate a comprehensive preference score for the user on different financial products. This includes: performing a nonlinear transformation on a multidimensional direction vector using at least one second multilayer perceptron to obtain weighting coefficients. The multidimensional direction vector represents the weights of a multi-objective task, which includes multiple of the following for the user's financial products: click-through rate, purchase behavior, satisfaction rating, and retention time; and then linearly weighting the N initial preference scores based on the weighting coefficients to generate the comprehensive preference score.

[0008] In one possible embodiment, after obtaining the weighting coefficients, the financial product recommendation method further includes: adjusting the dimensional weights of the multi-dimensional directional vector and updating the weighting coefficients based on the user's real-time feedback interaction data.

[0009] In one possible embodiment, the financial product recommendation method is applied to an electronic device. After obtaining the weighted coefficients, the financial product recommendation method further includes: displaying the correlation between the weighted coefficients and the embedding vectors of user feature data and financial product feature data on a first interface of the electronic device.

[0010] In one possible embodiment, the policy network is trained as follows: based on a reinforcement learning algorithm, the ranking results of the recommended financial products predicted by the initial recommendation policy network are associated with the user's historical interaction data and stored in an experience replay buffer; experience data is randomly selected from the experience replay buffer as training samples; the initial recommendation policy network is updated through the training samples until the loss function of the initial recommendation policy network converges, and the corresponding initial recommendation policy network is determined as the recommendation policy network.

[0011] In one possible embodiment, the reinforcement learning algorithm includes a proximal policy optimization algorithm.

[0012] Secondly, this application provides a financial product recommendation device, comprising: an initial preference score generation module, used to input user feature data and financial product feature data into N neural collaborative filtering networks of a recommendation strategy network to generate N initial preference scores for different financial products, where N is an integer greater than or equal to 2; a comprehensive preference score generation module, used to perform weighted fusion of the N initial preference scores through the routing network of the recommendation strategy network to generate a comprehensive preference score for different financial products; and a recommendation module, used to determine the ranking result of the recommended financial products based on the comprehensive preference score.

[0013] Thirdly, this application provides an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method as described in any of the first aspects.

[0014] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any of the first aspects.

[0015] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method of any one of the first aspects.

[0016] In this embodiment, user feature data and financial product feature data are input into N neural collaborative filtering networks (NCFs) of the recommendation strategy network. Each NCF independently models the interaction relationship between the user and the financial product using a deep learning network and independently generates an initial preference score for the user on the financial product. This avoids the local optima problem that can easily occur when a single NCF models the interaction relationship between the user and the financial product. A routing network weights and fuses the N initial preference scores output by the N N NCFs to generate a comprehensive preference score that reflects the user's global preference for different financial products. Ranking the financial products based on the comprehensive preference score accurately determines the ranking of the recommended financial products.

[0017] Furthermore, by assigning weights to N initial preference scores, the routing network enables N neural collaborative filtering networks to explore the multi-dimensional interaction between users and financial products in parallel. For example, in scenarios with sparse user feature data, the routing network prioritizes activating neural collaborative filtering networks that capture static user attributes (such as risk preferences), while in the cold start phase of new financial products, it activates neural collaborative filtering networks that focus on dynamic features of financial products (such as market popularity). This mechanism, through the parallel exploration of the multi-dimensional interaction between users and financial products by N neural collaborative filtering networks, significantly improves the adaptability of the recommendation strategy network to sparse data. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] Figure 1 This is a schematic diagram illustrating the application of the financial product recommendation method of this application to a financial institution;

[0020] Figure 2 This is a flowchart of a financial product recommendation method according to an embodiment of this application;

[0021] Figure 3 A flowchart illustrating a financial product recommendation method according to yet another embodiment of this application;

[0022] Figure 4This is a schematic diagram of a financial product recommendation method according to another embodiment of this application;

[0023] Figure 5 This is a schematic diagram of the structure of a financial product recommendation device provided in an embodiment of this application;

[0024] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application;

[0025] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0027] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all 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 the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0028] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0029] It should be noted that the financial product recommendation method, apparatus and equipment provided in this application can be applied to the field of financial technology, or to any field other than financial technology. This application does not limit the application field of the financial product recommendation method, apparatus and equipment.

[0030] Figure 1 This is a schematic diagram illustrating the application of the financial product recommendation method of this application to a financial institution. For example... Figure 1 As shown, when a user initiates a financial product recommendation request through the financial application 2 of terminal device 1, terminal device 1 sends the financial product recommendation request to the financial institution's server 3. Server 3 responds to the financial product recommendation request and recommends financial products to the user by executing the financial product recommendation method of this embodiment.

[0031] In related technology 1, a collaborative filtering (CF) algorithm is used for financial product recommendation. The collaborative filtering algorithm recommends financial products by calculating user similarity or financial product similarity.

[0032] Specifically, user-based collaborative filtering (User-CF) predicts the similarity of users to related financial products by analyzing the behavior of similar users, and then recommends financial products with high similarity to users. Item-based collaborative filtering (Item-CF) recommends similar financial products to users by analyzing the co-occurrence relationships of items.

[0033] However, financial product recommendations based on collaborative filtering algorithms rely on a large amount of historical data, which leads to the cold start problem. That is, when user or item data is sparse (such as new users or new financial products), there is a lack of sufficient historical data to support accurate similarity calculations, resulting in a significant decrease in the accuracy of financial product recommendations.

[0034] In addition, the interaction data between users and financial products in financial institutions is often sparse, and the relationship between users and financial products may not be close enough. This makes it impossible to effectively calculate the accurate similarity when using collaborative filtering algorithms to recommend financial products, thus affecting the recommendation effect of financial products.

[0035] Furthermore, using collaborative filtering algorithms for financial product recommendations may lead to over-recommendation, which means over-recommending popular financial products while ignoring some less popular but potentially suitable financial products, thus limiting the diversity of financial product recommendations.

[0036] In related technology 2, the financial product recommendation process is modeled as a reinforcement learning problem, and the financial product recommendation strategy is dynamically adjusted through the interaction between the agent and the environment.

[0037] Specifically, the intelligent agent performs the action of recommending financial products, using user feedback such as clicks and purchases as reward signals. Through interaction with the environment, the intelligent agent continuously optimizes its financial product recommendation strategy, learning how to make optimal financial product recommendations based on user feedback.

[0038] Reinforcement learning-based financial product recommendation methods can adapt to users' long-term behavioral patterns and have strong personalization and adaptability. However, they also face at least one of the following problems:

[0039] High computational resource consumption: Reinforcement learning requires a large number of training cycles, and policy updates depend on real-time interactive data, resulting in high training costs.

[0040] The challenge lies in striking a balance between exploration and utilization: overexploration may recommend irrelevant financial products, reducing user experience; overutilization, on the other hand, leads to simplistic recommendations that fail to meet diverse user needs.

[0041] Poor interpretability: The black-box nature of deep learning models makes recommendation decisions lack transparency, making it difficult for users to understand the recommendation logic and affecting trust.

[0042] The financial product recommendation method, apparatus, and equipment provided in this application are intended to at least solve one of the above-mentioned technical problems in the related technologies.

[0043] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0044] Figure 2 This is a flowchart illustrating a financial product recommendation method according to an embodiment of this application. The financial product recommendation method of this application embodiment can be executed by an electronic device, such as a server, and specifically includes steps S110 to S130.

[0045] S110. Input user feature data and financial product feature data into N neural collaborative filtering networks of the recommendation strategy network to generate N initial preference scores for different financial products.

[0046] User characteristic data includes, for example, risk preference data and transaction records. Financial product characteristic data includes, for example, yield data and risk level.

[0047] In one possible embodiment, user feature data is associated with, for example, a user identifier, and financial product feature data is associated with, for example, a financial product identifier. The server may input the user identifier and financial product identifier into N neural collaborative filtering networks of a recommendation strategy network. These N neural collaborative filtering networks search for associated user feature data and financial product feature data based on the user identifier and financial product identifier, and generate N initial preference scores for different financial products based on these scores.

[0048] Neural Collaborative Filtering (NCF) uses deep learning networks to model the interaction between users and financial products. N is an integer greater than or equal to 2. N ...

[0049] S120. The N initial preference scores are weighted and fused through the routing network of the recommendation strategy network to generate a comprehensive preference score for users on different financial products.

[0050] A routing network refers to a network that coordinates N initial preference scores output by N neural collaborative filtering networks through a learnable weight allocation mechanism.

[0051] To illustrate with a specific example: If the recommendation strategy network includes NCF1 and NCF2, NCF1 has an initial preference score of 0.8 for financial product P1, NCF2 has an initial preference score of 0.7 for financial product P1, and the weight of the routing network is [0.6, 0.4], then the overall preference score for financial product P1 is 0.76.

[0052] S130. Based on the comprehensive preference score, determine the ranking of recommended financial products.

[0053] For example, sorting by comprehensive preference score from highest to lowest, the top k financial products out of the entire pool are used as the ranking result for recommended financial products, where k is an integer greater than or equal to 1.

[0054] For example, the server can push financial products to users based on the ranking results of recommended financial products determined in the embodiments of this application, thereby improving the accuracy of financial product recommendations and increasing the conversion rate of financial products.

[0055] In this embodiment, user feature data and financial product feature data are input into N neural collaborative filtering networks (NCFs) of the recommendation strategy network. Each NCF independently models the interaction relationship between the user and the financial product using a deep learning network and independently generates an initial preference score for the user on the financial product. This avoids the local optima problem that can easily occur when a single NCF models the interaction relationship between the user and the financial product. A routing network weights and fuses the N initial preference scores output by the N N NCFs to generate a comprehensive preference score that reflects the user's global preference for different financial products. Ranking the financial products based on the comprehensive preference score accurately determines the ranking of the recommended financial products.

[0056] Furthermore, by assigning weights to N initial preference scores, the routing network enables N neural collaborative filtering networks to explore the multi-dimensional interaction between users and financial products in parallel. For example, in scenarios with sparse user feature data, the routing network prioritizes activating neural collaborative filtering networks that capture static user attributes (such as risk preferences), while in the cold start phase of new financial products, it activates neural collaborative filtering networks that focus on dynamic features of financial products (such as market popularity). This mechanism, through the parallel exploration of the multi-dimensional interaction between users and financial products by N neural collaborative filtering networks, significantly improves the adaptability of the recommendation strategy network to sparse data.

[0057] In one possible embodiment, the neural collaborative filtering network includes parallel generalized matrix factorization branches and a first multilayer perceptron.

[0058] Generalized Matrix Factorization (GMF) is used to model the linear interaction between users and financial products. A Multilayer Perceptron (MLP) is a neural network composed of multiple fully connected layers that undergo nonlinear transformations through activation functions. The first-layer MLP is used to model the nonlinear interaction between users and financial products.

[0059] Figure 3 This is a flowchart of a financial product recommendation method according to another embodiment of this application.

[0060] like Figure 3 As shown, step S110, which inputs user feature data and financial product feature data into N neural collaborative filtering networks of the recommendation strategy network to generate N initial preference scores for different financial products, includes: for the i-th neural collaborative filtering network in the N neural collaborative filtering networks, performing the following steps S111 to S114.

[0061] S111. The embedding vectors of user feature data and financial product feature data are linearly transformed by the generalized matrix decomposition branch to obtain the first feature vector.

[0062] Linear transformation processing includes dot product processing. For example, element-wise dot product processing is performed on the first and second sets of embedded vectors through generalized matrix factorization to model the linear interaction relationships between different users and different financial products. The first set of embedded vectors includes at least one embedded vector of user feature data indexed by a user identifier, and the second set of embedded vectors includes at least one embedded vector of financial product feature data indexed by a financial product identifier.

[0063] S112. The embedding vectors of user feature data and financial product feature data are concatenated using the first multilayer perceptron, and the concatenated embedding vectors are subjected to nonlinear transformation to obtain the second feature vector.

[0064] For example, the following formula (1) represents the first multilayer perceptron processing the concatenated embedding vector through a fully connected layer. A nonlinear transformation is performed to obtain the second eigenvector. .

[0065] (1)

[0066] In formula (1), This represents the activation function. Indicates the weights of the fully connected layer. This represents the offset of the fully connected layer.

[0067] It should be noted that the first multilayer perceptron can also perform multiple nonlinear transformations similar to formula (1) on the spliced ​​embedded vector through multiple fully connected layers to obtain the second feature vector.

[0068] S113. Perform feature fusion processing based on the first feature vector and the second feature vector to obtain the fused feature vector.

[0069] For example, feature fusion processing includes: splicing operations or addition operations.

[0070] S114. Calculate the i-th initial preference score based on the fused feature vector.

[0071] For example, the following formula (2) represents the server processing the fused feature vector through the output layer of the first multilayer perceptron to calculate the i-th initial preference score. .

[0072] (2)

[0073] In formula (2), This represents the activation function. This represents the weights of the output layer of the first multilayer perceptron. This represents the offset of the output layer of the first multilayer perceptron. This represents the fused feature vector.

[0074] In this embodiment, the neural collaborative filtering network includes parallel generalized matrix factorization (GMMF) branches and a first multilayer perceptron. The GMMF branches capture the linear relationship between users and financial products. Based on this, the first multilayer perceptron concatenates the embedding vectors of user feature data and financial product feature data, and then performs a nonlinear transformation on the concatenated embedding vector to capture the nonlinear relationship between users and financial products. This feature fusion processing overcomes the limitations of single linear or nonlinear networks, allowing the fused feature vector to comprehensively cover the features between users and financial products. Consequently, the initial preference score calculated based on the fused feature vector is more accurate.

[0075] In one possible embodiment, the input data of the neural collaborative filtering network further includes: external environment data, the embedding vector of which is used to fuse with the embedding vector of the user feature data.

[0076] External environmental data refers to external data that is unrelated to users and financial products but influences the recommendation of financial products. Examples of external environmental data include market interest rates and economic cycle data.

[0077] Based on the above example, step S111, which involves performing a linear transformation on the embedding vectors of user feature data and financial product feature data through a generalized matrix factorization branch to obtain the first feature vector, includes: performing feature fusion processing on the embedding vectors of user feature data and financial product feature data through a generalized matrix factorization branch to obtain the first fusion vector; and performing a linear transformation on the first fusion vector and the embedding vectors of financial product feature data through a generalized matrix factorization branch to obtain the first feature vector.

[0078] Step S112, which involves concatenating the embedding vectors of user feature data and financial product feature data using a first multilayer perceptron and performing a nonlinear transformation on the concatenated embedding vector to obtain a second feature vector, includes: performing feature fusion processing on the embedding vectors of user feature data and external environment data to obtain a second fusion vector; and concatenating the second fusion vector with the embedding vectors of financial product feature data using a first multilayer perceptron and performing a nonlinear transformation on the concatenated embedding vector to obtain a second feature vector.

[0079] In this embodiment, by introducing external environmental data, the recommendation strategy network can dynamically respond to the external environment. For example, when the external environmental data includes market interest rates and market interest rates fluctuate, the recommendation strategy network can optimize the output initial preference score by incorporating the fluctuating market interest rates, thereby improving the accuracy of the initial preference score.

[0080] like Figure 3 As shown, in one possible embodiment, the routing network includes at least one second multilayer perceptron. Step S130, which involves weighted fusion of N initial preference scores through the routing network of the recommendation strategy network to generate a comprehensive preference score for the user on different financial products, includes steps S131 to S132.

[0081] S131. The multidimensional direction vector is processed by nonlinear transformation through at least one second multilayer perceptron to obtain weighting coefficients.

[0082] The multidimensional directional vector represents the weights of a multi-objective task, which includes multiple of the following for users' financial products: click-through rate, purchase behavior, satisfaction rating, and retention time.

[0083] In the case of a routing network including a second multilayer perceptron, the second multilayer perceptron maps multidimensional direction vectors into high-dimensional features through nonlinear transformation processing to extract patterns of multi-objective tasks. It also compresses and activates the high-dimensional features through nonlinear transformation processing to generate weighted coefficients.

[0084] In the case where the routing network includes multiple second multilayer perceptrons MLP21 and MLP22 connected in sequence, some of the second multilayer perceptrons MLP21 are used to map multidimensional direction vectors into high-dimensional features to extract patterns of multi-objective tasks, and some of the second multilayer perceptrons MLP22 are used to compress and activate the high-dimensional features to generate weighted coefficients.

[0085] S132. Based on the weighting coefficients, linearly weight the N initial preference scores to generate a comprehensive preference score.

[0086] It should be noted that the routing network has the ability to assign weights through a learnable weight allocation mechanism. Different input multidimensional direction vectors (i.e., different task priorities) will cause the routing network to output different weighting coefficients, thereby enabling the recommendation policy network to match the task priorities indicated by the multidimensional direction vectors.

[0087] In this embodiment, each dimension of the multidimensional direction vector corresponds to the weight of a target task. The weighting coefficients generated by the multidimensional direction vector linearly weight the N initial preference scores, enabling the routing network to flexibly adjust the contribution of the N initial preference scores output by the N neural collaborative filtering networks based on the weights of each task. This avoids the recommendation lag caused by the static weight allocation of the N initial preference scores, allowing the recommendation strategy network to adapt to the weight allocation requirements of different tasks. By comprehensively evaluating preference scores, it accurately assesses user preferences and thus accurately recommends financial products.

[0088] In the above embodiments, after the multi-dimensional direction vector is input into the routing network, the routing network outputs weighted coefficients. In one possible embodiment, user feature data and financial product feature data can be input into the routing network. The routing network can perform feature fusion processing on the user feature data, financial product features, and multi-dimensional direction vector to obtain a third fused vector, and then perform feature extraction and calculate weighted coefficients on the third fused vector. Thus, the routing network can combine user features and financial product features to generate personalized weighted coefficients for users and financial products.

[0089] In one possible implementation, the multidimensional direction vector can be adjusted to adapt to flexible and dynamic multi-objective tasks. For example, users who successfully recommend financial products in the first category exhibit a high click-through rate, so the weight of click-through rate in the multidimensional direction vector can be increased when recommending financial products in the first category. Users who successfully recommend financial products in the second category exhibit a long retention time, so the weight of retention time in the multidimensional direction vector can be increased when recommending financial products in the second category.

[0090] In one possible embodiment, after obtaining the weighting coefficients in step S131, the product recommendation method may further include: adjusting the dimensional weights of the multidimensional directional vector and updating the weighting coefficients based on the user's real-time feedback interaction data.

[0091] Real-time user feedback interaction data refers to users' immediate behavioral data regarding financial products, such as clicks, purchases, and satisfaction ratings. For example, when a user clicks on a financial product repeatedly, the server can record the user's real-time feedback interaction data (such as clicks) and adjust the multi-dimensional directional vector to increase the weight of the click-through rate dimension. As another example, when a user's survey results indicate a preference for high-risk financial products, the server can record the user's real-time feedback interaction data (such as satisfaction ratings) and adjust the multi-dimensional directional vector to give higher weight to the initial preference rating output by a neural collaborative filtering network recommending high-risk products.

[0092] For example, the multidimensional direction vector is updated in real time using an online learning algorithm such as incremental gradient descent.

[0093] In this embodiment, the dimensional weights of the multi-dimensional directional vector are adjusted and the weighting coefficients are updated based on real-time user feedback and interaction data. This allows the recommendation strategy network to adapt to dynamic changes in user interest in financial products. For example, when a user shifts from a conservative to an aggressive investment approach, the server can adjust the multi-dimensional directional vector to instantly prioritize high-yield financial products, thereby significantly reducing recommendation lag and improving user satisfaction and timeliness.

[0094] In addition, in this embodiment, the dynamic adjustment of the multidimensional direction vector enables the recommendation strategy network to adjust the weights of N initial preference scores in real time based on the user's real-time feedback interaction data, avoiding the limitations of the initial preference scores output by a single neural collaborative filtering network, thereby generating more accurate initial preference scores in cold start scenarios.

[0095] Figure 4 A schematic diagram illustrating a financial product recommendation method according to yet another embodiment of this application is shown.

[0096] like Figure 4 As shown, the recommendation strategy network includes a routing network and four neural collaborative filtering networks (i.e., neural collaborative filtering networks 1 to 4, where N is 4). User feature data and financial product feature data are input into each neural filtering network of the recommendation strategy network to obtain four initial preference scores for different financial products. A multidimensional directional vector (along with user feature data and financial product feature data) is input into the routing network to obtain weighting coefficients. The four initial preference scores for different financial products are weighted and fused based on these weighting coefficients to obtain a fourth fused vector. The embedding layer of the recommendation strategy network maps the fourth fused vector to a high-dimensional embedding space. Then, a third multilayer perceptron performs a nonlinear transformation on the vector processed by the embedding layer. The nonlinearly transformed vector is then converted into a probability distribution of the user's preferences for different financial products through a softmax layer (this probability distribution represents the user's preferences for different financial products). This generates a comprehensive preference score for the user's preferences for different financial products, and the ranking of recommended financial products can be determined based on this comprehensive preference score.

[0097] In one possible embodiment, the method is applied to an electronic device. After obtaining the weighting coefficients in step S131, the financial product recommendation method further includes: displaying the correlation between the weighting coefficients and the embedding vectors of user feature data and financial product feature data on a first interface of the electronic device.

[0098] In this embodiment, after generating the weighted coefficients of the multi-dimensional direction vector, the routing network visualizes and associates these weighted coefficients with the embedding vectors of user feature data and financial product feature data. This allows users to intuitively understand the priority of each task in the multi-objective task of the recommendation strategy network (such as the weight allocation between click-through rate and purchase behavior), improving the transparency of the recommendation strategy network and addressing its black-box nature. For example, users can observe through the first interface why a certain high-yield product is prioritized for recommendation, thereby enhancing their trust in financial institutions.

[0099] like Figure 3 As shown, in one possible embodiment, the recommendation policy network is trained in the following manner:

[0100] S101. Based on the reinforcement learning algorithm, the ranking results of the recommended financial products predicted by the initial recommendation strategy network are associated with the user's historical interaction data and stored in the experience replay buffer.

[0101] Reinforcement learning algorithms include proximal policy optimization (PPO), which will be used as an example below. PPO limits the magnitude of network parameter updates in the initial recommendation policy network through a pruning function, ensuring the stability of training the initial recommendation policy network.

[0102] Historical interaction data is a specific action in the action space, which includes any financial product in the ranking results of the predicted recommended financial products or any financial product in the ranking results of the recommended financial products without selecting the predicted ones.

[0103] The initial recommendation policy network refers to the recommendation policy network before training. The initial recommendation policy network has the same network structure as the policy recommendation network but different network parameters.

[0104] The experience replay buffer is used to store user and historical interaction data with relationships. This user and historical interaction data with relationships can be represented, for example, as quadruples. Quadruple Including status ,action ,award and the next state .

[0105] The state includes multidimensional direction vectors, user feature data, financial product feature data, and the ranking results of recommended financial products predicted by the initial recommendation strategy network.

[0106] In one possible implementation, the reward is calculated using a reward function. This reward function can be a weighted sum of click-through rate (CTR), purchase behavior, satisfaction rating, and retention time. The weights of CTR, purchase behavior, satisfaction rating, and retention time can be set based on expert experience or the specific application scenario.

[0107] Specifically, for click-through rate (CTR), if a user clicks on a financial product in the ranking results predicted by the initial recommendation strategy network, the reward value increases by 1; otherwise, it decreases by 1. For purchase behavior, if a user purchases a financial product in the ranking results predicted by the initial recommendation strategy network, the reward value increases by 10; otherwise, it decreases by 1. For user satisfaction ratings, rewards can be set based on user satisfaction ratings for the financial products in the ranking results predicted by the initial recommendation strategy network, for example, a rating range of 1 to 5. For user retention time, rewards can be set based on the time users spend on the pages of the financial products in the ranking results predicted by the initial recommendation strategy network. If the retention time exceeds a preset threshold (e.g., 15 seconds), the reward value increases by 1; otherwise, it decreases by 1.

[0108] S102. Randomly select empirical data from the empirical replay buffer as training samples.

[0109] Randomly selecting empirical data as training samples can reduce the correlation between training samples and alleviate the overfitting of the initial recommendation strategy network.

[0110] S103. Update the initial recommendation policy network using training samples until the loss function of the initial recommendation policy network converges, and determine the corresponding initial recommendation policy network as the recommendation policy network.

[0111] During the process of updating the initial recommendation policy network using training samples, the training samples are input into the initial recommendation policy network to obtain the ranking results of the recommended financial products predicted by the initial recommendation policy network. The loss function based on the PPO algorithm calculates the loss value between the ranking results of the recommended financial products predicted by the initial recommendation policy network and the ranking results labels of the recommended financial products from the training samples. If the loss value is greater than a threshold, the next round of training iterations continues; if the loss value is less than the threshold, training is complete.

[0112] It should be noted that during the training of the initial recommendation policy network, the routing network of the initial recommendation policy network can randomly select a multi-dimensional direction vector from multiple multi-dimensional direction vectors to predict the weighting coefficients.

[0113] For example, the following formula (3) represents the loss function of the PPO algorithm. .

[0114] (3)

[0115] In formula (3), Loss from pruning strategy Value function loss Entropy regularization term Related. This represents the expectation of the training samples. This represents the network parameters of the recommendation policy network, which are the parameters to be optimized by the near-end policy optimization algorithm. and These represent weighting coefficients used in the loss function of the equilibrium value. With entropy regularization term The importance of entropy regularization. The larger the value of , the stronger the randomness of the initial recommendation strategy network's output ranking of financial products; the entropy regularization term. This is used to encourage the initial recommendation policy network to explore the network state space.

[0116] The following formula (4) represents the loss of the pruning strategy. .

[0117] (4)

[0118] In formula (4), This represents the update ratio, which is the ratio of the current ranking of financial products output by the initial recommendation strategy network to the previous ranking of financial products output by the initial recommendation strategy network in terms of action probability. This is an estimate of the advantage function, used to measure the performance of a given state. Take action below How much better are the returns than average? The cutting factor (usually ranging from 0.1 to 0.2) is the factor used for cropping. limit The range is in [ ]between. This represents the clipping function.

[0119] The following formula (5) represents the update ratio.

[0120] (5)

[0121] In formula (5), This indicates the current state of the initial recommendation policy network. Take action below The probability of. This indicates that the old initial recommendation policy network (the initial recommendation policy network before the update) is in state Take action below The probability of this change. The update ratio reflects the degree of change in the current initial recommendation policy network compared to the old initial recommendation policy network.

[0122] In this embodiment, the experience replay buffer storage mechanism enables the server to fully utilize users' historical interaction data as training samples, thereby increasing the diversity of training samples and reducing training costs. Randomly sampled training samples are used to train the initial recommendation policy network, resulting in a recommendation policy network with stronger generalization ability and reduced computational resource consumption.

[0123] On the other hand, training the initial recommendation policy network using a proximal policy optimization algorithm can improve training stability.

[0124] Figure 5 This is a schematic diagram of a financial product recommendation device provided in an embodiment of this application. Figure 5 As shown in the embodiment of this application, the financial product recommendation device includes: an initial preference score generation module 210, a comprehensive preference score generation module 220, and a recommendation module 230.

[0125] The initial preference rating generation module 210 is used to input user feature data and financial product feature data into N neural collaborative filtering networks of the recommendation strategy network to generate N initial preference ratings for different financial products, where N is an integer greater than or equal to 2.

[0126] The comprehensive preference score generation module 220 is used to weight and fuse N initial preference scores through the routing network of the recommendation strategy network to generate a comprehensive preference score for users on different financial products.

[0127] The recommendation module 230 is used to determine the ranking of recommended financial products based on the comprehensive preference score.

[0128] In one possible embodiment, the neural collaborative filtering network includes parallel generalized matrix factorization branches and a first multilayer perceptron; the initial preference score generation module includes: a linear transformation processing submodule, used to perform linear transformation processing on the embedding vectors of user feature data and financial product feature data through the generalized matrix factorization branches to obtain a first feature vector; a nonlinear transformation processing submodule, used to concatenate the embedding vectors of user feature data and financial product feature data through the first multilayer perceptron, and perform nonlinear transformation processing on the concatenated embedding vector to obtain a second feature vector; a fusion processing submodule, used to perform feature fusion processing based on the first feature vector and the second feature vector to obtain a fused feature vector; and a calculation submodule, used to calculate the i-th initial preference score based on the fused feature vector.

[0129] In one possible embodiment, the input data of the neural collaborative filtering network further includes: external environment data, the embedding vector of which is used to fuse with the embedding vector of the user feature data.

[0130] In one possible embodiment, the routing network includes at least one second multilayer perceptron, and the comprehensive preference score generation module includes: a weighting coefficient determination submodule, used to perform nonlinear transformation processing on a multidimensional direction vector through at least one second multilayer perceptron to obtain weighting coefficients, wherein the multidimensional direction vector represents the weights of a multi-objective task, and the multi-objective task includes multiple of the following for a user's financial product: click-through rate, purchase behavior, satisfaction rating, and retention time; and a weighting submodule, used to linearly weight N initial preference scores based on the weighting coefficients to generate a comprehensive preference score.

[0131] In one possible embodiment, the financial product recommendation device further includes a weighted coefficient update module, used to adjust the dimensional weights of the multi-dimensional directional vector and update the weighted coefficients based on the user's real-time feedback interaction data.

[0132] In one possible embodiment, the financial product recommendation device is applied to an electronic device. After obtaining the weighted coefficients, the financial product recommendation device further includes a display module for displaying the correlation between the weighted coefficients and the embedding vectors of user feature data and financial product feature data on a first interface of the electronic device.

[0133] In one possible embodiment, the policy network is trained as follows: based on a reinforcement learning algorithm, the ranking results of the recommended financial products predicted by the initial recommendation policy network are associated with the user's historical interaction data and stored in an experience replay buffer; experience data is randomly selected from the experience replay buffer as training samples; the initial recommendation policy network is updated through the training samples until the loss function of the initial recommendation policy network converges, and the corresponding initial recommendation policy network is determined as the recommendation policy network.

[0134] In one possible embodiment, the reinforcement learning algorithm includes a proximal policy optimization algorithm.

[0135] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, this application embodiment provides an electronic device including a processor 301 and a memory 302. Optionally, the device further includes a communication component 303. The processor 301, memory 302, and communication component 303 are connected via a bus 304.

[0136] In the specific implementation process, the memory 302 stores code, and the processor 301 runs the code stored in the memory 302 to execute the method of the above method embodiment.

[0137] The specific implementation process of processor 301 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0138] In the above Figure 6 In the illustrated embodiments, it should be understood that the processor 301 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0139] The memory 302 may include high-speed RAM memory, and may also include non-volatile memory (NVM), such as at least one disk storage.

[0140] Bus 304 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 304 can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus 304 in the accompanying drawings of this application is not limited to only one bus or one type of bus.

[0141] This application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods described in the above-described method embodiments.

[0142] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0143] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0144] This application provides a computer program product, including a computer program that, when executed by a processor, implements the methods provided in any of the embodiments described above.

[0145] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0146] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0147] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0148] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0149] When integrated units / modules are implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0150] If the integrated unit / module is implemented as a software program module and sold or used as an independent financial product, it can be stored in a computer-readable storage device (CMD). 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 financial product. This computer software financial product is stored in a memory 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 of the various embodiments of this application. The aforementioned memory 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.

[0151] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0152] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0153] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method of recommending a financial product, characterized by, The method comprises the following steps: inputting user feature data and financial product feature data into N neural collaborative filtering networks of a recommendation strategy network to generate N initial preference scores of a user for different financial products, N being an integer greater than or equal to 2; performing weighted fusion on the N initial preference scores through a routing network of the recommendation strategy network to generate a comprehensive preference score of the user for different financial products; determining a ranking result of recommended financial products according to the comprehensive preference score.

2. The method of claim 1, wherein, The neural collaborative filtering network comprises a parallel general matrix factorization branch and a first multilayer perceptron; the step of inputting user feature data and financial product feature data into N neural collaborative filtering networks of a recommendation strategy network to generate N initial preference scores of a user for different financial products comprises the following steps for an i-th neural collaborative filtering network in the N neural collaborative filtering networks, i∈[1, N]: performing linear transformation processing on an embedding vector of the user feature data and an embedding vector of the financial product feature data through the general matrix factorization branch to obtain a first feature vector; performing splicing processing on the embedding vector of the user feature data and the embedding vector of the financial product feature data through the first multilayer perceptron, and performing nonlinear transformation processing on the spliced embedding vector to obtain a second feature vector; performing feature fusion processing on the first feature vector and the second feature vector to obtain a fusion feature vector; calculating an i-th initial preference score according to the fusion feature vector.

3. The method of claim 2, wherein, The input data of the neural collaborative filtering network further comprises external environment data, and an embedding vector of the external environment data is used for fusion with an embedding vector of the user feature data.

4. The method of claim 2, wherein, The routing network comprises at least one second multilayer perceptron, and the step of performing weighted fusion on the N initial preference scores through a routing network of the recommendation strategy network to generate a comprehensive preference score of the user for different financial products comprises the following steps: performing nonlinear transformation processing on a multidimensional direction vector through the at least one second multilayer perceptron to obtain a weighting coefficient, the multidimensional direction vector representing the weights of a multi-objective task, the multi-objective task including multiple ones of the following for the user on the financial product: click rate, purchase behavior, satisfaction score, retention time; performing linear weighting on the N initial preference scores based on the weighting coefficient to generate the comprehensive preference score.

5. The method of claim 4, wherein, After the weighting coefficient is obtained, the method further comprises the following steps: adjusting the dimension weight of the multidimensional direction vector and updating the weighting coefficient according to real-time feedback interaction data of the user.

6. The method of claim 4, wherein, The method is applied to an electronic device, and after the weighting coefficient is obtained, the method further comprises the following steps: displaying, on a first interface of the electronic device, an association relationship between the weighting coefficient and the embedding vector of the user feature data and the embedding vector of the financial product feature data.

7. The method according to any one of claims 1 to 6, characterized in that, The strategy network is trained in the following manner: based on a reinforcement learning algorithm, associating a ranking result of a recommended financial product predicted by an initial recommendation strategy network with historical interaction data of a user, and storing the ranking result in an experience replay buffer; randomly selecting experience data from the experience replay buffer as a training sample; The initial recommendation strategy network is updated through the training sample until a loss function of the initial recommendation strategy network converges, and a corresponding initial recommendation strategy network is determined as the recommendation strategy network.

8. The method of claim 7, wherein, The reinforcement learning algorithm includes a proximal policy optimization algorithm.

9. A financial product recommendation device characterized by comprising: The method comprises the steps of: An initial preference score generation module is configured to input user feature data and financial product feature data into N neural collaborative filtering networks of a recommendation strategy network, and generate N initial preference scores of a user for different financial products, where N is an integer greater than or equal to 2. An integrated preference score generation module is configured to weight and integrate the N initial preference scores through a routing network of the recommendation strategy network, and generate an integrated preference score of the user for different financial products. A recommendation module is configured to determine a ranking result of recommended financial products according to the integrated preference score.

10. An electronic device, comprising: The method comprises the steps of: A processor and a memory connected in communication with the processor; The memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory to implement the method according to any one of claims 1 to 8.

Images