User screening method and device based on graph contrast learning, equipment and storage medium
By constructing a bipartite graph network and utilizing graph contrastive learning methods, high-value users are automatically screened, solving the problems of high manpower consumption and strong subjectivity in traditional manual screening, and achieving efficient and accurate user screening and maximizing bank profits.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BANK OF CHINA
- Filing Date
- 2022-08-18
- Publication Date
- 2026-07-24
AI Technical Summary
Traditional methods of manually screening high-value users are labor-intensive and highly subjective, making it difficult to accurately identify truly high-value users.
By constructing a bipartite graph network using a graph contrastive learning approach, and using the GraphSAGE graph network model to encode user and financial product behavioral data, supervised and unsupervised learning is performed. The graph network model is trained by combining cross-entropy loss and InfoNCE loss functions to automatically filter high-value users.
It enables efficient and intelligent identification of high-value users, reduces labor costs, improves the accuracy of user screening, and allows for the development of effective strategies for high-value users to maximize bank profits.
Smart Images
Figure CN115329159B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of user screening technology, and in particular to a user screening method, apparatus, device, and storage medium based on graph contrastive learning. Background Technology
[0002] In banking operations, it is often necessary to identify high-value users in order to provide targeted services to them. The traditional method is to identify high-value users through manual screening, which wastes a lot of human resources. Moreover, manual screening is too subjective and not conducive to identifying truly high-value users. Summary of the Invention
[0003] To address the shortcomings of the prior art, this invention provides a user screening method, apparatus, device, and storage medium based on graph comparison learning, in order to improve the accuracy of user screening.
[0004] The first aspect of this application provides a user screening method based on graph contrastive learning, including:
[0005] Multiple tuples are generated based on user behavior records; wherein, each tuple includes a user, a financial product, and a user-product weight; the user-product weight is determined based on the user's behavior toward the financial product.
[0006] A bipartite graph network is constructed using the plurality of binary pairs; wherein the bipartite graph network includes a first node representing a user, a second node representing a financial product, and an edge connecting the first node and the second node; the edge has a weight corresponding to the user product of the first node and the second node it connects;
[0007] Using the bipartite graph network as input, a graph network model is trained based on supervised learning and unsupervised contrastive learning methods to obtain a trained graph network model.
[0008] The trained graph network model is used to process the bipartite graph network to obtain the screening index for each of the first nodes; the screening index serves as the basis for screening the users represented by each of the first nodes.
[0009] Optionally, generating multiple tuples based on user behavior records includes:
[0010] For each user-product combination, determine the user product weight corresponding to the user-product combination, and form a binary tuple corresponding to the user-product combination by combining the users and financial products in the user-product combination and the user product weight corresponding to the user-product combination.
[0011] The process of determining the user-product weights corresponding to the user-product combination includes:
[0012] Determine the types and intensity of users' historical behavior towards financial products within the user-product portfolio;
[0013] If the behavior type is click, and the behavior intensity is less than the first dwell threshold, the user product weight is determined to be the first weight; the behavior intensity corresponding to the click is the dwell time.
[0014] If the behavior type is click, and the behavior intensity is greater than or equal to the first dwell threshold and less than the second dwell threshold, the user product weight is determined to be the second weight; the second weight is greater than the first weight.
[0015] If the behavior type is click, and the behavior intensity is greater than or equal to the second dwell threshold, the user product weight is determined to be the third weight; the third weight is greater than the second weight.
[0016] If the behavior type is purchase, the user product weight is determined to be a fourth weight that is greater than the third weight, based on the corresponding behavior intensity; the fourth weight is positively correlated with the behavior intensity corresponding to purchase.
[0017] Optionally, the step of using the bipartite graph network as input to train the graph network model based on supervised learning and unsupervised contrastive learning methods to obtain the trained graph network model includes:
[0018] The bipartite graph network is subjected to two heuristic data augmentations to obtain the first view and the second view corresponding to the bipartite graph network.
[0019] The graph network model is used to encode the first view and the second view to obtain the first representation vector of each node in the first view and the second representation vector of each node in the second view.
[0020] The contrastive loss value of the graph network model is calculated using multiple samples composed of the first representation vector and the second representation vector; wherein, each sample includes a first representation vector and a second representation vector, samples whose first and second representation vectors correspond to the same node are positive samples, and samples whose first and second representation vectors correspond to different nodes are negative samples.
[0021] The first node in the bipartite graph network is encoded using a graph network model to obtain the third representation vector of the first node;
[0022] The cross-entropy loss value of the graph network model is determined based on the third representation vector and the real labels of the user represented by the pre-labeled first node;
[0023] The overall loss value of the graph network model is determined based on the contrast loss value and the cross-entropy loss value.
[0024] The parameters of the graph network model are updated based on the comprehensive loss value of the graph network model until the comprehensive loss value meets the preset comprehensive convergence condition, thereby obtaining the trained graph network model.
[0025] Optionally, the step of calculating the contrastive loss value of the graph network model using multiple samples composed of the first representation vector and the second representation vector includes:
[0026] For each sample, if the sample is a positive sample, the negative of the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample; if the sample is a negative sample, the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample.
[0027] The sample loss of each sample is summed to obtain the contrastive loss of the graph network model.
[0028] A second aspect of this application provides a user screening device based on graph contrastive learning, comprising:
[0029] A generation unit is used to generate multiple tuples based on user behavior records; wherein, the tuples include a user, a financial product, and a user product weight; the user product weight is determined based on the user's behavior toward the financial product;
[0030] A construction unit is used to construct a bipartite graph network using the plurality of binary pairs; wherein the bipartite graph network includes a first node representing a user, a second node representing a financial product, and an edge connecting the first node and the second node; the edge has a weight corresponding to the user product of the connected first node and second node;
[0031] The training unit is used to train the graph network model based on supervised learning and unsupervised contrastive learning methods, using the bipartite graph network as input, to obtain the trained graph network model.
[0032] The processing unit is used to process the bipartite graph network using the trained graph network model to obtain the screening index for each of the first nodes; the screening index serves as the basis for screening the users represented by each of the first nodes.
[0033] Optionally, when the generation unit generates multiple tuples based on user behavior records, it is specifically used for:
[0034] For each user-product combination, determine the user product weight corresponding to the user-product combination, and form a binary tuple corresponding to the user-product combination by combining the users and financial products in the user-product combination and the user product weight corresponding to the user-product combination.
[0035] The process of determining the user-product weights corresponding to the user-product combination includes:
[0036] Determine the types and intensity of users' historical behavior towards financial products within the user-product portfolio;
[0037] If the behavior type is click, and the behavior intensity is less than the first dwell threshold, the user product weight is determined to be the first weight; the behavior intensity corresponding to the click is the dwell time.
[0038] If the behavior type is click, and the behavior intensity is greater than or equal to the first dwell threshold and less than the second dwell threshold, the user product weight is determined to be the second weight; the second weight is greater than the first weight.
[0039] If the behavior type is click, and the behavior intensity is greater than or equal to the second dwell threshold, the user product weight is determined to be the third weight; the third weight is greater than the second weight.
[0040] If the behavior type is purchase, the user product weight is determined to be a fourth weight that is greater than the third weight, based on the corresponding behavior intensity; the fourth weight is positively correlated with the behavior intensity corresponding to purchase.
[0041] Optionally, the training unit uses the bipartite graph network as input and trains the graph network model based on supervised learning and unsupervised contrastive learning methods. When obtaining the trained graph network model, it is specifically used for:
[0042] The bipartite graph network is subjected to two heuristic data augmentations to obtain the first view and the second view corresponding to the bipartite graph network.
[0043] The graph network model is used to encode the first view and the second view to obtain the first representation vector of each node in the first view and the second representation vector of each node in the second view.
[0044] The contrastive loss value of the graph network model is calculated using multiple samples composed of the first representation vector and the second representation vector; wherein, each sample includes a first representation vector and a second representation vector, samples whose first and second representation vectors correspond to the same node are positive samples, and samples whose first and second representation vectors correspond to different nodes are negative samples.
[0045] The first node in the bipartite graph network is encoded using a graph network model to obtain the third representation vector of the first node;
[0046] The cross-entropy loss value of the graph network model is determined based on the third representation vector and the real labels of the user represented by the pre-labeled first node;
[0047] The overall loss value of the graph network model is determined based on the contrast loss value and the cross-entropy loss value.
[0048] The parameters of the graph network model are updated based on the comprehensive loss value of the graph network model until the comprehensive loss value meets the preset comprehensive convergence condition, thereby obtaining the trained graph network model.
[0049] Optionally, when the training unit calculates the contrastive loss value of the graph network model using multiple samples composed of the first representation vector and the second representation vector, it is specifically used for:
[0050] For each sample, if the sample is a positive sample, the negative of the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample; if the sample is a negative sample, the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample.
[0051] The sample loss of each sample is summed to obtain the contrastive loss of the graph network model.
[0052] A third aspect of this application provides an electronic device, including a memory and a processor;
[0053] The memory is used to store computer programs;
[0054] The processor is used to execute the computer program, which, when executed, is specifically used to implement the user screening method based on graph contrastive learning provided in any of the first aspects of this application.
[0055] The fourth aspect of this application provides a computer storage medium for storing a computer program, which, when executed, is specifically used to implement the user screening method based on graph contrastive learning provided in any one of the first aspects of this application.
[0056] This application provides a user screening method, apparatus, device, and storage medium based on graph contrastive learning, applicable to the fields of artificial intelligence or finance. The method includes: generating multiple tuples based on user behavior records; wherein each tuple includes a user, a financial product, and a user-product weight; the user-product weight is determined based on the user's behavior towards the financial product; constructing a bipartite graph network using the multiple tuples; wherein the bipartite graph network includes a first node representing a user, a second node representing a financial product, and edges connecting the first and second nodes; each edge has a user-product weight corresponding to the connected first and second nodes; training a graph network model using the bipartite graph network as input based on supervised learning and unsupervised contrastive learning methods to obtain a trained graph network model; processing the bipartite graph network using the trained graph network model to obtain a screening index for each first node; the screening index serves as the basis for screening users represented by each first node. This solution automatically determines the screening index for each user by constructing a bipartite graph network and using the graph network model for analysis, thereby enabling the automatic screening of high-value users based on the screening index. Attached Figure Description
[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0058] Figure 1 A flowchart illustrating the user selection method based on graph contrastive learning provided in this application embodiment;
[0059] Figure 2 A schematic diagram of a user screening device based on graph contrast learning provided in an embodiment of this application;
[0060] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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 are within the scope of protection of the present invention.
[0062] Graph contrastive learning: In the field of graph learning, due to the lack of labels or difficulty in labeling graph data, researchers have applied contrastive learning techniques to graph representation learning tasks and achieved very good results. This series of algorithm studies is called graph contrastive learning.
[0063] In banking operations, it's often necessary to identify high-value customers to provide targeted services. Traditionally, this involves manual screening, which wastes significant human resources and is heavily influenced by subjectivity, hindering the identification of truly high-value customers. This invention uses graph-based comparative learning to intelligently and quickly identify high-value customers. This allows for more dedicated human resources to maintain relationships with these customers, ultimately maximizing bank profits.
[0064] To address the aforementioned issues, this application establishes a high-value user acquisition method based on graph contrastive learning. This method constructs a bipartite graph network using user and financial product behavioral data. The network nodes include user nodes and financial product nodes, with edges connecting them. Edge weights are calculated based on behavior type and intensity to measure user preference for different financial products. Then, the GraphSAGE graph network model is used to encode the representation vectors of the graph nodes, which are compared with the user's actual labels to calculate the cross-entropy loss function. Based on contrastive learning, two heuristic data augmentations are performed on the bipartite graph, such as randomly adding or removing edges and randomly deleting nodes, resulting in two views of the graph network. The graph network model is then used again to encode the representation vectors of the nodes in each view, and the contrastive learning InfoNCE loss function is calculated. Finally, the cross-entropy loss function and InfoNCE loss function values are weighted and fused to obtain the final loss function value, which is then used to update the graph network model parameters. The trained graph network model is then used to predict user value, intelligently and quickly identifying high-value users, and developing corresponding strategies to maximize bank profits for these high-value users.
[0065] The GraphSAGE graph network model is an existing graph neural network algorithm (or graph neural network model). The specific working principle of this algorithm can be found in relevant existing technical literature, and will not be elaborated here.
[0066] The InfoNCE loss function is an existing loss function in the field of contrastive learning. Its specific definition and expression can be found in relevant existing technical literature, and will not be elaborated here.
[0067] This solution mainly includes the following modules:
[0068] Log collection and processing module: Based on the bank client's behavior logs, it obtains customer behavior data on financial products, including wealth management, funds, and insurance, and processes it into a binary tuple "user-financial product-weight". The weight is calculated based on the behavior type and intensity. For example, a click without purchasing but a dwell time of less than 5 seconds has a weight of 0.1, a click without purchasing but a dwell time of less than 1 minute has a weight of 1, a click without purchasing but a dwell time of more than 1 minute has a weight of 2, and a click and purchase has a weight of 3. The weight can be appropriately increased based on the user's purchase amount.
[0069] User-Product Bipartite Graph Construction Module: Based on the preprocessed user-product tuples, a bipartite graph network is established, where the network nodes include user nodes and financial product nodes. The edges of the network connect user nodes and financial product nodes, and the edge weights are the weight values in the tuples, used to measure the degree of user preference for different financial products.
[0070] Graph node representation module: The representation vector of the graph node is obtained by encoding using the Graphsage graph network model. The model uses the average pooling aggregation function and two aggregation layers. The cross-entropy loss function value is calculated by comparing with the user's real label.
[0071] Unsupervised comparison module for graph nodes: Based on comparison learning, two heuristic data augmentations are performed on the bipartite graph, such as randomly adding or removing edges and randomly deleting nodes, to obtain two views of the graph network. The Graphsage graph network model is then used to encode the representation vectors of the nodes in the two views respectively. There are a total of 2N view node representation vectors in each training batch. Samples from the same graph node are defined as positive samples, and otherwise as negative samples. The InfoNCE loss function value is calculated based on the positive and negative samples.
[0072] High-value user identification module: The cross-entropy loss function value and the InfoNCE loss function value are weighted and fused to obtain the final loss function value, which is then used to update the graph network model parameters. The trained graph network model is used to predict the value of users, intelligently and quickly screen and identify high-value users, and formulate corresponding strategies for these high-value users to maximize bank profits.
[0073] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0074] Please see Figure 1 This is a schematic diagram of a user screening method based on graph contrastive learning provided in an embodiment of this application. The method may include the following steps.
[0075] S101, generate multiple tuples based on user behavior records.
[0076] The binary pair includes the user, the financial product, and the user product weight; the user product weight is determined based on the user's behavior towards the financial product.
[0077] In step S101, multiple users registered in the banking system can be collected first. Then, for each user, if a product link corresponding to a financial product has been pushed to that user, the user and the financial product are identified as a user-product combination.
[0078] For example, if a user A registered in the banking system has had product links for financial products A and B pushed to their client in the past week, then a user-product combination of (user A, financial product A) and a user-product combination of (user A, financial product B) can be identified.
[0079] Based on the user-product combination defined above, multiple tuples are generated according to user behavior records, which may include:
[0080] For each user-product combination, determine the user-product weight corresponding to the user-product combination, and form a binary pair corresponding to the user-product combination by combining the users and financial products in the user-product combination and the user-product weight corresponding to the user-product combination.
[0081] The process of determining the user-product weights corresponding to the user-product combination includes:
[0082] Determine the types and intensity of users' historical behavior towards financial products within the user-product portfolio;
[0083] For a user-product combination, the user may not take any action to interact with the product links of the financial products within it. In this case, the combination is deleted, and no corresponding binary tuple is generated. If the user interacts with the product links of the financial products within it, the type of interaction may be either a click or a transaction.
[0084] If the behavior type is click and the behavior intensity is less than the first dwell threshold, the user's product weight is determined as the first weight; the behavior intensity corresponding to the click is the dwell time.
[0085] For a user-product combination, if the behavior type is "click," it means that the user clicked the product link of the financial product. In this case, the client will display the details page of the financial product. The above-mentioned dwell time is equivalent to the duration of the details page of the financial product displayed on the client.
[0086] The dwell time can be timed starting when the user clicks the product link and stopping when the user actively exits the details page.
[0087] If the behavior type is click, and the behavior intensity is greater than or equal to the first dwell threshold and less than the second dwell threshold, the user's product weight is determined to be the second weight; the second weight is greater than the first weight.
[0088] If the behavior type is click, and the behavior intensity is greater than or equal to the second dwell threshold, the user's product weight is determined to be the third weight; the third weight is greater than the second weight.
[0089] If the behavior type is purchase, the user's product weight is determined to be a fourth weight that is greater than the third weight, based on the intensity of the corresponding behavior; the fourth weight is positively correlated with the intensity of the purchase behavior.
[0090] It should be noted that "purchase" refers to clicking on a product link and making a purchase. In other words, once a user clicks on a product link, regardless of how much time has passed, as long as the user ultimately purchases the financial product, the behavior type of this interaction is defined as a purchase.
[0091] The intensity of the purchase behavior can be defined as the amount of financial products a user buys, such as how much money a user buys in financial products.
[0092] The above thresholds and weights can be preset as needed. For example, the first dwell time threshold can be 5 seconds, and the first weight can be 0.1; the second dwell time can be 1 minute, and the second weight can be 1; the third weight can be 2; the fourth weight is determined according to the purchase amount. For example, the initial fourth weight can be 3. If the purchase amount is greater than 10,000, the fourth weight can be 4. If the purchase amount is greater than 100,000, the fourth weight can be 5, and so on.
[0093] S102, constructs a bipartite graph network using multiple pairs.
[0094] The bipartite graph network includes a first node representing a user, a second node representing a financial product, and edges connecting the first and second nodes; the edges have user product weights corresponding to the first and second nodes they connect.
[0095] Specifically, in step S102, first nodes corresponding to each user in multiple pairs can be generated first, then second nodes corresponding to each financial product can be generated. Finally, each pair is traversed. For each pair, an edge is added between the first node corresponding to the user in the pair and the second node corresponding to the financial product in the pair. The weight of the user product in the pair is assigned to this edge. After traversing all pairs, the construction of the bipartite graph network is completed.
[0096] S103 uses a bipartite graph network as input and trains the graph network model based on supervised learning and unsupervised contrastive learning methods to obtain a trained graph network model.
[0097] Optionally, the specific process of training a graph network model using a bipartite graph network as input, based on supervised learning methods and unsupervised contrastive learning methods, to obtain the trained graph network model may include:
[0098] B1. Perform two heuristic data augmentations on the bipartite graph network to obtain the first and second views corresponding to the bipartite graph network.
[0099] Heuristic data augmentation can specifically involve randomly adding or subtracting nodes and edges in a bipartite graph network, i.e., randomly adding edges or nodes, and randomly deleting existing edges or nodes. The weight of the newly added edge can be the lowest weight mentioned above.
[0100] B2 uses a graph network model to encode the first view and the second view, obtaining the first representation vector of each node in the first view and the second representation vector of each node in the second view.
[0101] The graph network model here can specifically be the aforementioned Graphsage graph network model. For the specific architecture of the Graphsage graph network model, as well as the specific methods for encoding the first and second views of the bipartite graph network using the Graphsage graph network model, please refer to the relevant existing technical literature, which will not be elaborated here.
[0102] Optionally, the aggregation function used in this graph network model can be the average pooling aggregation function, and the model can have two aggregation layers.
[0103] It should be noted that in step B2, both the first node and the second node will be encoded to obtain the corresponding first representation vector and second representation vector.
[0104] B3 calculates the contrastive loss value of the graph network model using multiple samples composed of the first and second representation vectors.
[0105] Each sample includes a first representation vector and a second representation vector. Samples whose first and second representation vectors correspond to the same node are positive samples, and samples whose first and second representation vectors correspond to different nodes are negative samples.
[0106] The aforementioned comparative loss value can specifically be a function value calculated using the aforementioned InfoNCE loss function.
[0107] Optionally, the contrastive loss value of the graph network model can be calculated using multiple samples composed of the first and second representation vectors, which may include:
[0108] For each sample, if the sample is a positive sample, the negative of the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss; if the sample is a negative sample, the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss.
[0109] The sample loss of each sample is summed to obtain the contrastive loss of the graph network model.
[0110] The negative of similarity, which is the value obtained by adding a negative sign before the similarity score, is equivalent to negative similarity. For example, if the similarity is 20%, then the negative of the similarity score is -20%.
[0111] In other words, for a positive sample, the closer the first representation vector and the second representation vector contained therein are, the smaller the sample loss of the positive sample; for a negative sample, the less similar the first representation vector and the second representation vector contained therein are, the smaller the sample loss of the negative sample.
[0112] B4 uses a graph network model to encode the first node in a bipartite graph network, obtaining the third representation vector of the first node.
[0113] As mentioned earlier, the graph network model here can specifically be the aforementioned Graphsage graph network model. For the specific architecture of the Graphsage graph network model, as well as the specific methods for encoding the first and second views of the bipartite graph network using the Graphsage graph network model, please refer to the relevant existing technical literature, which will not be elaborated here.
[0114] B5 determines the cross-entropy loss value of the graph network model based on the third representation vector and the user's real label represented by the pre-labeled first node.
[0115] A user's true identity can be determined by analyzing each user's information in advance, including recent income statements (e.g., income over the past year), asset information, the number and type of financial products currently held, the holding period, and personal information such as age and occupation.
[0116] In step B5, the cross-entropy loss value can be determined based on the difference between each user's third representation vector and the true label. The greater the difference between the third representation vector and the true label, the greater the cross-entropy loss value, and vice versa. The specific calculation method for the cross-entropy loss value can be found in relevant existing technical literature and will not be elaborated further.
[0117] B6. Determine the overall loss value of the graph network model based on the contrast loss value and the cross-entropy loss value.
[0118] Optionally, the comparative loss value and the cross-entropy loss value can be added together to obtain the comprehensive loss value, or the two can be added together with certain weights to obtain the comprehensive loss value.
[0119] B7. Update the parameters of the graph network model based on the comprehensive loss value of the graph network model until the comprehensive loss value meets the preset comprehensive convergence condition, and obtain the trained graph network model.
[0120] In step B7, the overall convergence condition can be that the overall loss value is less than or equal to a preset convergence threshold. If the overall convergence condition is not met, the parameters of the graph network model can be updated according to the overall loss value, and then the process can be repeated from step B1 to B7 until the overall loss value meets the overall convergence condition.
[0121] S104 uses the trained graph network model to process the bipartite graph network and obtains the screening index for each first node.
[0122] Screening metrics serve as the basis for selecting users represented by each first node. Specifically, the screening metrics can be the user value of each user.
[0123] After obtaining the screening criteria for each user, in S104, users whose screening criteria are higher than a certain threshold can be identified as high-value users, or N users can be selected as high-value users according to the screening criteria from high to low. This embodiment does not limit the specific screening method.
[0124] This application provides a user screening method based on graph contrastive learning, applicable to the fields of artificial intelligence or finance. The method includes: generating multiple tuples based on user behavior records; wherein each tuple includes a user, a financial product, and a user-product weight; the user-product weight is determined based on the user's behavior towards the financial product; constructing a bipartite graph network using the multiple tuples; wherein the bipartite graph network includes a first node representing a user, a second node representing a financial product, and edges connecting the first and second nodes; each edge has a user-product weight corresponding to the connected first and second nodes; using the bipartite graph network as input, training a graph network model based on supervised learning and unsupervised contrastive learning methods to obtain a trained graph network model; processing the bipartite graph network using the trained graph network model to obtain a screening index for each first node; the screening index serves as the basis for screening users represented by each first node. This solution automatically determines the screening index for each user by constructing a bipartite graph network and using the graph network model for analysis, thereby automatically screening high-value users based on the screening index.
[0125] According to the user screening method based on graph contrast learning provided in the embodiments of this application, the embodiments of this application also provide a user screening device based on graph contrast learning. Please refer to [link to relevant documentation]. Figure 2The device may include the following units.
[0126] The generation unit 201 is used to generate multiple tuples based on user behavior records. Each tuple includes a user, a financial product, and a user-product weight; the user-product weight is determined based on the user's behavior towards the financial product.
[0127] Construction unit 202 is used to construct a bipartite graph network using multiple pairs of tuples.
[0128] The bipartite graph network includes a first node representing a user, a second node representing a financial product, and edges connecting the first and second nodes; the edges have user product weights corresponding to the first and second nodes they connect.
[0129] Training unit 203 is used to train a graph network model based on supervised learning and unsupervised contrastive learning methods, using a bipartite graph network as input, to obtain a trained graph network model.
[0130] The processing unit 204 is used to process the bipartite graph network using the trained graph network model to obtain the screening index of each first node; the screening index serves as the basis for screening the users represented by each first node.
[0131] Optionally, when the generation unit 201 generates multiple tuples based on user behavior records, it is specifically used for:
[0132] For each user-product combination, determine the user-product weight corresponding to the user-product combination, and form a binary pair corresponding to the user-product combination by combining the users and financial products in the user-product combination and the user-product weight corresponding to the user-product combination.
[0133] The process of determining the user-product weights corresponding to the user-product combination includes:
[0134] Determine the types and intensity of users' historical behavior towards financial products within the user-product portfolio;
[0135] If the behavior type is click and the behavior intensity is less than the first dwell threshold, the user's product weight is determined as the first weight; the behavior intensity corresponding to the click is the dwell time.
[0136] If the behavior type is click, and the behavior intensity is greater than or equal to the first dwell threshold and less than the second dwell threshold, the user's product weight is determined to be the second weight; the second weight is greater than the first weight.
[0137] If the behavior type is click, and the behavior intensity is greater than or equal to the second dwell threshold, the user's product weight is determined to be the third weight; the third weight is greater than the second weight.
[0138] If the behavior type is purchase, the user's product weight is determined to be a fourth weight that is greater than the third weight, based on the intensity of the corresponding behavior; the fourth weight is positively correlated with the intensity of the purchase behavior.
[0139] Optionally, training unit 203 uses a bipartite graph network as input and trains the graph network model based on supervised learning and unsupervised contrastive learning methods. When obtaining the trained graph network model, it is specifically used for:
[0140] Two heuristic data augmentations were performed on the bipartite graph network to obtain the first and second views corresponding to the bipartite graph network.
[0141] The graph network model is used to encode the first view and the second view to obtain the first representation vector of each node in the first view and the second representation vector of each node in the second view.
[0142] The contrastive loss value of the graph network model is calculated using multiple samples composed of the first representation vector and the second representation vector. Each sample includes a first representation vector and a second representation vector. Samples whose first and second representation vectors correspond to the same node are positive samples, and samples whose first and second representation vectors correspond to different nodes are negative samples.
[0143] The first node in the bipartite graph network is encoded using a graph network model to obtain the third representation vector of the first node;
[0144] The cross-entropy loss value of the graph network model is determined based on the third representation vector and the user's real label represented by the pre-labeled first node;
[0145] The overall loss value of the graph network model is determined based on the comparative loss value and the cross-entropy loss value.
[0146] The parameters of the graph network model are updated based on the overall loss value until the overall loss value meets the preset overall convergence condition, thus obtaining the trained graph network model.
[0147] Optionally, when training unit 203 calculates the contrastive loss value of the graph network model using multiple samples composed of the first and second representation vectors, it is specifically used for:
[0148] For each sample, if the sample is a positive sample, the negative of the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss; if the sample is a negative sample, the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss.
[0149] The sample loss of each sample is summed to obtain the contrastive loss of the graph network model.
[0150] The user screening device based on graph contrast learning provided in this embodiment can be referred to in the relevant steps of the user screening method based on graph contrast learning provided in any embodiment of this application for its specific working principle, which will not be repeated here.
[0151] This application provides a user screening device based on graph contrastive learning, applicable to the fields of artificial intelligence or finance. The device includes: a generation unit 201 that generates multiple pairs of data based on user behavior records; wherein each pair includes a user, a financial product, and a user-product weight; the user-product weight is determined based on the user's behavior towards the financial product; a construction unit 202 that constructs a bipartite graph network using the multiple pairs of data; wherein the bipartite graph network includes a first node representing a user, a second node representing a financial product, and edges connecting the first and second nodes; each edge has a user-product weight corresponding to the first and second nodes connected; a training unit 203 that uses the bipartite graph network as input and trains a graph network model based on supervised learning and unsupervised contrastive learning methods to obtain a trained graph network model; and a processing unit 204 that processes the bipartite graph network using the trained graph network model to obtain a screening index for each first node; the screening index serves as the basis for screening users represented by each first node. This solution automatically determines the screening index for each user by constructing a bipartite graph network and using the graph network model for analysis, thereby enabling the automatic screening of high-value users based on the screening index.
[0152] This application also provides an electronic device; please refer to [link to relevant documentation]. Figure 3 It includes a memory 301 and a processor 302.
[0153] The memory 301 is used to store computer programs.
[0154] The processor 302 is used to execute a computer program, which, when executed, is specifically used to implement the user screening method based on graph contrast learning provided in any embodiment of this application.
[0155] This application also provides a computer storage medium for storing a computer program. When the computer program is executed, it is specifically used to implement the user screening method based on graph contrast learning provided in any embodiment of this application.
[0156] It should be noted that the user screening method, apparatus, device, and storage medium based on graph contrastive learning provided by this invention can be used in the fields of artificial intelligence or finance. The above are merely examples and do not limit the application areas of the user screening method, apparatus, device, and storage medium based on graph contrastive learning provided by this invention.
[0157] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0158] It should be noted that the concepts of "first" and "second" mentioned in this invention are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.
[0159] Those skilled in the art will be able to implement or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A user screening method based on graph contrastive learning, characterized in that, include: Multiple tuples are generated based on user behavior records; wherein, each tuple includes a user, a financial product, and a user-product weight; the user-product weight is determined based on the user's behavior toward the financial product. A bipartite graph network is constructed using the plurality of binary pairs; wherein the bipartite graph network includes a first node representing a user, a second node representing a financial product, and an edge connecting the first node and the second node; the edge has a weight corresponding to the user product of the first node and the second node it connects; Using the bipartite graph network as input, a graph network model is trained based on supervised learning and unsupervised contrastive learning methods to obtain a trained graph network model. The trained graph network model is used to process the bipartite graph network to obtain the screening index for each first node; the screening index serves as the basis for screening the users represented by each first node. The step of using the bipartite graph network as input to train a graph network model based on supervised learning and unsupervised contrastive learning methods to obtain a trained graph network model includes: The bipartite graph network is subjected to two heuristic data augmentations to obtain the first view and the second view corresponding to the bipartite graph network. The graph network model is used to encode the first view and the second view to obtain the first representation vector of each node in the first view and the second representation vector of each node in the second view. The contrastive loss value of the graph network model is calculated using multiple samples composed of the first representation vector and the second representation vector; wherein, each sample includes a first representation vector and a second representation vector, samples whose first and second representation vectors correspond to the same node are positive samples, and samples whose first and second representation vectors correspond to different nodes are negative samples. The first node in the bipartite graph network is encoded using a graph network model to obtain the third representation vector of the first node; The cross-entropy loss value of the graph network model is determined based on the third representation vector and the real labels of the user represented by the pre-labeled first node; The overall loss value of the graph network model is determined based on the contrast loss value and the cross-entropy loss value. The parameters of the graph network model are updated based on the comprehensive loss value of the graph network model until the comprehensive loss value meets the preset comprehensive convergence condition, thereby obtaining the trained graph network model.
2. The method according to claim 1, characterized in that, The generation of multiple tuples based on user behavior records includes: For each user-product combination, determine the user product weight corresponding to the user-product combination, and form a binary tuple corresponding to the user-product combination by combining the users and financial products in the user-product combination and the user product weight corresponding to the user-product combination. The process of determining the user-product weights corresponding to the user-product combination includes: Determine the types and intensity of users' historical behavior towards financial products within the user-product portfolio; If the behavior type is click, and the behavior intensity is less than the first dwell threshold, the user product weight is determined to be the first weight; the behavior intensity corresponding to the click is the dwell time. If the behavior type is click, and the behavior intensity is greater than or equal to the first dwell threshold and less than the second dwell threshold, the user product weight is determined to be the second weight; the second weight is greater than the first weight. If the behavior type is click, and the behavior intensity is greater than or equal to the second dwell threshold, the user product weight is determined to be the third weight; the third weight is greater than the second weight. If the behavior type is purchase, the user product weight is determined to be a fourth weight that is greater than the third weight, based on the corresponding behavior intensity; the fourth weight is positively correlated with the behavior intensity corresponding to purchase.
3. The method according to claim 1, characterized in that, The step of calculating the contrastive loss value of the graph network model using multiple samples composed of the first representation vector and the second representation vector includes: For each sample, if the sample is a positive sample, the negative of the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample; if the sample is a negative sample, the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample. The sample loss of each sample is summed to obtain the contrastive loss of the graph network model.
4. A user screening device based on graph contrastive learning, characterized in that, include: A generation unit is used to generate multiple tuples based on user behavior records; wherein, the tuples include a user, a financial product, and a user product weight; the user product weight is determined based on the user's behavior toward the financial product; A construction unit is used to construct a bipartite graph network using the plurality of binary pairs; wherein the bipartite graph network includes a first node representing a user, a second node representing a financial product, and an edge connecting the first node and the second node; the edge has a weight corresponding to the user product of the connected first node and second node; The training unit is used to train the graph network model based on supervised learning and unsupervised contrastive learning methods, using the bipartite graph network as input, to obtain the trained graph network model. The processing unit is used to process the bipartite graph network using the trained graph network model to obtain the screening index for each of the first nodes; the screening index serves as the basis for screening the users represented by each of the first nodes. The training unit, using the bipartite graph network as input, trains the graph network model based on supervised learning and unsupervised contrastive learning methods. When obtaining the trained graph network model, it is specifically used for: The bipartite graph network is subjected to two heuristic data augmentations to obtain the first view and the second view corresponding to the bipartite graph network. The graph network model is used to encode the first view and the second view to obtain the first representation vector of each node in the first view and the second representation vector of each node in the second view. The contrastive loss value of the graph network model is calculated using multiple samples composed of the first representation vector and the second representation vector; wherein, each sample includes a first representation vector and a second representation vector, samples whose first and second representation vectors correspond to the same node are positive samples, and samples whose first and second representation vectors correspond to different nodes are negative samples. The first node in the bipartite graph network is encoded using a graph network model to obtain the third representation vector of the first node; The cross-entropy loss value of the graph network model is determined based on the third representation vector and the real labels of the user represented by the pre-labeled first node; The overall loss value of the graph network model is determined based on the contrast loss value and the cross-entropy loss value. The parameters of the graph network model are updated based on the comprehensive loss value of the graph network model until the comprehensive loss value meets the preset comprehensive convergence condition, thereby obtaining the trained graph network model.
5. The apparatus according to claim 4, characterized in that, When the generation unit generates multiple tuples based on user behavior records, it is specifically used for: For each user-product combination, determine the user product weight corresponding to the user-product combination, and form a binary tuple corresponding to the user-product combination by combining the users and financial products in the user-product combination and the user product weight corresponding to the user-product combination. The process of determining the user-product weights corresponding to the user-product combination includes: Determine the types and intensity of users' historical behavior towards financial products within the user-product portfolio; If the behavior type is click, and the behavior intensity is less than the first dwell threshold, the user product weight is determined to be the first weight; the behavior intensity corresponding to the click is the dwell time. If the behavior type is click, and the behavior intensity is greater than or equal to the first dwell threshold and less than the second dwell threshold, the user product weight is determined to be the second weight; the second weight is greater than the first weight. If the behavior type is click, and the behavior intensity is greater than or equal to the second dwell threshold, the user product weight is determined to be the third weight; the third weight is greater than the second weight. If the behavior type is purchase, the user product weight is determined to be a fourth weight that is greater than the third weight, based on the corresponding behavior intensity; the fourth weight is positively correlated with the behavior intensity corresponding to purchase.
6. The apparatus according to claim 4, characterized in that, When the training unit calculates the contrastive loss value of the graph network model using multiple samples composed of the first representation vector and the second representation vector, it is specifically used for: For each sample, if the sample is a positive sample, the negative of the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample; if the sample is a negative sample, the similarity between the first representation vector and the second representation vector contained in the sample is determined as the sample loss of the sample. The sample loss of each sample is summed to obtain the contrastive loss of the graph network model.
7. A computer storage medium, characterized in that, Used to store computer programs, which, when executed, are specifically used to implement the user screening method based on graph contrastive learning as described in any one of claims 1 to 3.
8. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is used to execute the computer program, specifically to implement the user screening method based on graph contrastive learning as described in any one of claims 1 to 3.