Information push method and system based on user behavior sequence

Abstract

The present application provides an information push method and system based on user behavior sequence. First, by acquiring user interest learning tasks generated by a basic user behavior mining network, a progressive training process is used to perform user interest learning on the basic user behavior mining network, which can gradually improve the performance and stability of the network and ensure that the generated target user behavior mining network has higher quality; secondly, by monitoring the training results of the first progressive training process in a second progressive training node and performing user interest learning of the second progressive training process when the network parameters are stable, dynamic adjustment and optimization of the training process can be achieved, thereby improving training efficiency and effect; finally, by mining a user interest point map of a target user behavior sequence based on the target user behavior mining network, user interests can be grasped more accurately, and information push services that better meet users' needs can be provided to improve user experience.

Description

Technical Field

[0001] The present application relates to the technical field of digital financial services, and more specifically, to a method and system for pushing information based on user behavior sequences. Background Art

[0002] With the rapid development and popularization of the Internet, users will generate a large amount of user behavior data when using digital financial pages, such as online banking, securities trading platforms, insurance service websites, etc. These data include a series of operations such as user clicks, browsing, purchases, and evaluations, which contain rich user interest information and consumption habits. How to effectively mine and utilize these user behavior data to provide more accurate and personalized information push services is an important challenge facing the current financial technology field. Summary of the invention

[0003] In view of this, the purpose of the present application is to provide an information push method and system based on user behavior sequence.

[0004] According to a first aspect of the present application, there is provided an information push method based on a user behavior sequence, the method comprising:

[0005] Obtaining a user interest learning task generated for a basic user behavior mining network, the user interest learning task is used to instruct the basic user behavior mining network to perform multiple progressive training processes of user interest learning until a target user behavior mining network is generated, and after mining a user interest point map of a target user behavior sequence of an input target user based on the target user behavior mining network, information is pushed to the target user according to the user interest point map, wherein the target user behavior mining network is a progressive training network model generated by performing user interest learning based on template user behavior sequences of multiple digital financial pages;

[0006] According to the user interest learning task, performing user interest learning of a first progressive training process on the basic user behavior mining network in a first progressive training node, wherein a plurality of the progressive training processes include the first progressive training process, and the first progressive training node is a progressive training node of the first progressive training process;

[0007] Monitoring the training result of the first progressive training process in a second progressive training node, where the second progressive training node is a progressive training node of the second progressive training process, a plurality of the progressive training processes include the second progressive training process, and the first progressive training process is a forward progressive training node of the second progressive training node;

[0008] When it is monitored that the training result indicates that a network parameter is in a stable state under the first progressive training process, user interest learning of the second progressive training process is performed in the second progressive training node.

[0009] In a possible implementation manner of the first aspect, after monitoring the training result of the first progressive training process in the second progressive training node, the method further includes:

[0010] When it is monitored that the training result indicates that the first progressive training process does not show a stable state of network parameters, monitoring the training progress state of the first progressive training process;

[0011] When it is monitored that the training progress state meets the training validity requirement, sampling the training result in the set monitoring node until the training result indicates that the network parameter is in a stable state under the first progressive training process;

[0012] When it is monitored that the training progress status does not meet the training effectiveness requirement, the first progressive training process is removed, and user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

[0013] In a possible implementation of the first aspect, monitoring the training progress status of the first progressive training process includes:

[0014] Obtaining training results of the first progressive training process in X monitoring nodes;

[0015] Determine a statistic of a target monitoring node among the X monitoring nodes, wherein the first progressive training process is in an unstable state of network parameters in the target monitoring node;

[0016] When it is monitored that the statistic is less than Y, it is determined that the training progress status meets the training effectiveness requirement, and when it is monitored that the statistic is not less than Y, it is determined that the training progress status does not meet the training effectiveness requirement.

[0017] In a possible implementation manner of the first aspect, after sampling the training result in the setting monitoring node until the training result indicates that a network parameter is in a stable state under the first progressive training process, the method further includes:

[0018] When it is monitored that the training results in the set monitoring node indicate that the first progressive training process has not reached a stable state of network parameters, the first progressive training process is removed, and user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

[0019] In a possible implementation manner of the first aspect, performing user interest learning of the second progressive training process in the second progressive training node includes:

[0020] Performing user interest learning of the second progressive training process on the user behavior mining network generated by the first progressive training process in the second progressive training node;

[0021] Alternatively, user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

[0022] In a possible implementation manner of the first aspect, performing user interest learning of the second progressive training process on the user behavior mining network generated by the first progressive training process in the second progressive training node includes:

[0023] Determine a first user behavior mining network generated by the first progressive training process, where the first user behavior mining network is a progressive training network model generated by performing user interest learning based on a plurality of template user behavior sequences S1 of the first progressive training process;

[0024] Acquire a plurality of template user behavior sequences S2 of the second progressive training process, wherein the template user behavior sequence S2 is different from the template user behavior sequence S1;

[0025] The first user behavior mining network is used to learn user interests based on the plurality of template user behavior sequences S2, and parameter information of the first user behavior mining network is progressively optimized during the user interest learning process until a second user behavior mining network that meets the training termination requirements is generated.

[0026] In a possible implementation of the first aspect, before the step of performing user interest learning of a first progressive training process on the basic user behavior mining network in the first progressive training node, the method further includes:

[0027] Acquire multiple template user behavior sequences of a first digital financial page S3 and multiple template user behavior sequences of a second digital financial page S4, wherein the multiple digital financial pages include the first digital financial page and the second digital financial page;

[0028] Integrate the features of the plurality of template user behavior sequences S3 and the plurality of template user behavior sequences S4 to generate a plurality of template user behavior sequences;

[0029] The performing of a first progressive training process of user interest learning on the basic user behavior mining network in the first progressive training node includes:

[0030] In the first progressive training node, user interest learning of the first progressive training process is performed on the basic user behavior mining network according to the plurality of template user behavior sequences.

[0031] In a possible implementation of the first aspect, before the step of acquiring multiple template user behavior sequences S3 of the first digital financial page and multiple template user behavior sequences S4 of the second digital financial page, the method further includes:

[0032] Performing a first confusion perturbation strategy processing on the plurality of template user behavior sequences S3 of the first digital financial page, and loading the plurality of template user behavior sequences S3 processed by the first confusion perturbation strategy into the second digital financial page;

[0033] Acquire a global template user behavior sequence fed back by the second digital financial page and a plurality of template user behavior sequences S4 processed by the second confusion perturbation strategy, wherein the global template user behavior sequence is a template user behavior sequence generated after the plurality of template user behavior sequences S3 processed by the first confusion perturbation strategy and the plurality of template user behavior sequences S4 of the second digital financial page are processed by the second confusion perturbation strategy;

[0034] Performing decoding conversion of the first confusion perturbation strategy on the global template user behavior sequence and the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, generating multiple template user behavior sequences S3 processed by the second confusion perturbation strategy, and multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, wherein the template user behavior sequence S3 covers the multiple template user behavior sequences S3 processed by the second confusion perturbation strategy, and the template user behavior sequence S4 covers the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy;

[0035] The step of integrating the features of the plurality of template user behavior sequences S3 and the plurality of template user behavior sequences S4 to generate a plurality of template user behavior sequences includes:

[0036] Determine a shared template user behavior sequence of the multiple template user behavior sequences S3 processed by the second confusion perturbation strategy and the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, and output the shared template user behavior sequence as the template user behavior sequence.

[0037] In a possible implementation manner of the first aspect, after performing user interest learning of the second progressive training process in the second progressive training node, the method further includes:

[0038] Obtain the user behavior mining network that completes user interest learning;

[0039] Performing network performance verification on the user behavior mining network that has completed user interest learning according to multiple verification template user behavior sequences, and generating network performance indicators corresponding to the user behavior mining network that has completed user interest learning;

[0040] When it is monitored that the network performance indicator is less than a threshold indicator parameter, activating the next progressive training process of the first progressive training process until the target user behavior mining network is generated;

[0041] When it is monitored that the network performance index is not less than the threshold index parameter, the user behavior mining network that completes the user interest learning is output as the target user behavior mining network;

[0042] In a possible implementation manner of the first aspect, activating a next progressive training process of the first progressive training process until the target user behavior mining network is generated includes:

[0043] Monitoring, within a current round of progressive training nodes, past training results of a forward progressive training node of the current round of progressive training nodes, the current round of progressive training nodes being the next progressive training process of the first progressive training process, and the current round of progressive training nodes being the progressive training nodes of the current round of progressive training processes;

[0044] When it is monitored that the past training results indicate that the forward progressive training node of the current round of progressive training nodes does not have a stable network parameter state, outputting the next progressive training process of the current round of progressive training process as the current round of progressive training process;

[0045] When it is monitored that the past training results indicate that a network parameter stable state occurs under the forward progressive training node of the current progressive training node, user interest learning of the current progressive training process is performed in the current progressive training node to generate a user behavior mining network that completes user interest learning in this round;

[0046] Verifying the user behavior mining network that has completed user interest learning in this round according to the plurality of verification template user behavior sequences, and generating a network performance indicator for this round corresponding to the user behavior mining network that has completed user interest learning in this round;

[0047] When it is monitored that the network performance indicator of the current round is less than the threshold indicator parameter, outputting the next progressive training process of the current round of progressive training process as the progressive training process of the current round;

[0048] When it is monitored that the network performance index of this round is not less than the threshold index parameter, the user behavior mining network that completes user interest learning in this round is output as the target user behavior mining network.

[0049] According to the second aspect of the present application, there is provided an information push system based on user behavior sequence, the information push system based on user behavior sequence comprising a machine-readable storage medium and a processor, the machine-readable storage medium storing machine-executable instructions, when the processor executes the machine-executable instructions, the information push system based on user behavior sequence implements the aforementioned information push method based on user behavior sequence.

[0050] According to a third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the aforementioned information push method based on user behavior sequence is implemented.

[0051] According to any one of the above aspects, the technical effect of the present application is:

[0052] The embodiment of the present application firstly obtains the user interest learning task generated by the basic user behavior mining network, which can realize the deep mining and learning of user interests, thereby improving the accuracy and personalization of information push; secondly, the user interest learning of the basic user behavior mining network is carried out by using a progressive training process, which can gradually improve the performance and stability of the network, and ensure that the generated target user behavior mining network has higher quality; furthermore, by monitoring the training results of the first progressive training process in the second progressive training node, and carrying out the user interest learning of the second progressive training process under the stable state of network parameters, the dynamic adjustment and optimization of the training process can be realized, and the training efficiency and effect can be improved; finally, based on the user interest point map of the target user behavior sequence mined by the target user behavior mining network, the user interest can be grasped more accurately, and the user can be provided with an information push service that better meets their needs, thereby improving the user experience. Thus, the accuracy and personalization of information push can be effectively improved, and the user experience can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0054] Figure 1 One of the flow diagrams of the information push method based on user behavior sequence provided in the embodiment of the present application;

[0055] Figure 2 Flow chart 2 of the information push method based on user behavior sequence provided in the embodiment of the present application;

[0056] Figure 3 A schematic diagram of the component structure of an information push system based on a user behavior sequence for implementing the above-mentioned information push method based on a user behavior sequence provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below according to the drawings in the embodiment of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description, and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the embodiment of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add multiple other operations to the flowchart under the guidance of the content of the present application, and can also destroy multiple operations from the flowchart.

[0058] Figure 1 The flowchart of the information push method and system based on user behavior sequence provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the information push method based on user behavior sequence of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The detailed steps of the information push method based on user behavior sequence include:

[0059] Step S110, obtaining a user interest learning task generated for a basic user behavior mining network, wherein the user interest learning task is used to instruct the basic user behavior mining network to perform multiple progressive training processes of user interest learning until a target user behavior mining network is generated, and after mining a user interest point map of a target user behavior sequence of an input target user based on the target user behavior mining network, information is pushed to the target user according to the user interest point map, wherein the target user behavior mining network is a progressive training network model generated by performing user interest learning based on template user behavior sequences of multiple digital financial pages.

[0060] In this embodiment, in the financial field, the information push system based on the user behavior sequence acts as a server and receives a specific user interest learning task. This user interest learning task can be designed by the business team based on market demand and user feedback, and aims to improve the accuracy and effectiveness of personalized information push by deeply mining the user's behavior data when using financial applications. The user interest learning task details the structure, data source, training objectives, and evaluation criteria of the basic user behavior mining network used.

[0061] For example, suppose a financial technology company has developed a comprehensive financial application that covers multiple services such as stock trading, fund investment, and insurance purchase. In order to improve user experience and increase user stickiness, the information push function can be optimized through the machine learning system on the server. The user interest learning task received by the server requires it to use the user's browsing, clicking, trading and other behavioral data in the application, through a progressive training process, to gradually optimize an existing basic user behavior mining network, so that it can more accurately identify the user's financial interests, and push personalized financial information, product recommendations and other information accordingly.

[0062] For example, in one example, suppose a digital financial platform system wants to improve the accuracy of its personalized recommendation system to better meet user needs and improve user experience. The platform collects a large amount of user behavior data, including browsing records, click records, purchase records, etc. In order to achieve in-depth mining and learning of user interests, the platform adopts the above-mentioned information push method based on user behavior sequence.

[0063] First, the platform obtains the user interest learning task generated by the basic user behavior mining network. This task is to learn user interests based on the template user behavior sequences of multiple digital financial pages to generate the target user behavior mining network. The template user behavior sequence can be some typical user behavior patterns preset by the platform, such as "browse financial products → view product details → purchase products".

[0064] Then, based on the user interest learning task, the platform conducts user interest learning in the first progressive training process on the basic user behavior mining network in the first progressive training node. In this process, the network will learn how to predict the user's interest points based on the user behavior sequence. For example, when a user browses a financial product, the network may predict that the user is interested in the product's yield, risk, etc.

[0065] Then, in the second progressive training node, the platform monitors the training results of the first progressive training process. If the network parameters are monitored to be stable, that is, the network performance is no longer significantly improved, the platform will perform user interest learning in the second progressive training process in the second progressive training node. In this process, the network will further learn how to predict the user's interest points based on more complex user behavior sequences. For example, when a user browses multiple financial products at the same time and compares them, the network may predict that the user is interested in the advantages and disadvantages of different products, applicable scenarios, etc.

[0066] Finally, after multiple progressive training processes, a target user behavior mining network is generated. The network can mine the user's interest point map based on the input target user's target user behavior sequence (such as a user's browsing and clicking records in the past week). This map can include information such as the user's interest in different types of financial products and preferences for different risk levels. Based on this map, the platform can push information that is more in line with the user's needs and interests, such as recommending similar or related financial products and providing personalized financial advice. As a result, the digital financial platform system can grasp the user's interests and needs more accurately and provide users with information push services that are more in line with their expectations, thereby improving the user experience and the competitiveness of the platform.

[0067] Step S120, based on the user interest learning task, performing user interest learning of the first progressive training process on the basic user behavior mining network in a first progressive training node, multiple progressive training processes including the first progressive training process, and the first progressive training node is a progressive training node of the first progressive training process.

[0068] In this embodiment, the server starts to train the basic user behavior mining network at the first progressive training node according to the requirements of the user interest learning task. This first progressive training node is the starting stage of the entire training process. Its main purpose is to perform preliminary optimization and adjustment on the basic user behavior mining network to lay a good foundation for subsequent training.

[0069] At the first progressive training node, the server first loads the basic user behavior mining network and prepares a large amount of user behavior data as training samples. These user behavior data include various user behavior records in financial applications, such as the stock pages browsed, the types of funds concerned, and the insurance products purchased. The server uses these user behavior data to train the basic user behavior mining network, and gradually adjusts the parameters such as weights and biases of the basic user behavior mining network through the back propagation algorithm, so that it can better capture the characteristics and patterns of user interests. At the same time, the server can also use various regularization techniques and optimization algorithms to improve the stability and efficiency of training.

[0070] Step S130, monitoring the training results of the first progressive training process in a second progressive training node, wherein the second progressive training node is a progressive training node of the second progressive training process, multiple progressive training processes include the second progressive training process, and the first progressive training process is a forward progressive training node of the second progressive training node.

[0071] In this embodiment, after completing the training of the first progressive training node, the server enters the second progressive training node and begins to monitor and evaluate the training results of the first stage. This step is to ensure the effectiveness of the training, discover possible problems in a timely manner, and provide guidance for the next stage of training.

[0072] For example, in the second progressive training node, the server used an independent validation data set to evaluate the training effect of the first stage. In detail, the server calculated the accuracy, recall, F1 score and other indicators of the basic user behavior mining network on the validation set, and plotted the corresponding learning curve and confusion matrix. Through these analyses, the server found that the basic user behavior mining network performed well in user interest recognition in certain categories, while there was a lot of room for improvement in other categories. In addition, the server also monitored the loss function value and gradient changes during the network training process to ensure that the training did not fall into local optimality or overfitting.

[0073] Step S140: when it is monitored that the training result indicates that a network parameter is in a stable state under the first progressive training process, user interest learning of the second progressive training process is performed in the second progressive training node.

[0074] In this embodiment, when the server monitors that the network parameters under the first progressive training process have reached a stable state and the performance meets the preset requirements, the next stage of user interest learning can be started at the second progressive training node. The goal of this stage is to further improve the performance of the basic user behavior mining network so that it can more accurately identify and mine user interests.

[0075] For example, in the second progressive training node, the server adjusted the training strategy based on the monitoring results of the first phase. For example, the server increased the depth and width of the basic user behavior mining network and introduced more complex nonlinear transformations and feature interactions to enhance the basic user behavior mining network's ability to express and generalize user behavior. At the same time, the server also uses more advanced optimization algorithms and learning rate scheduling strategies to accelerate the training process and prevent overfitting. Through these improvements, the server aims to enable the target user behavior mining network to more accurately identify and push financial information that meets the personalized needs of users.

[0076] Based on the above steps, the embodiment of the present application firstly obtains the user interest learning task generated by the basic user behavior mining network, which can realize the deep mining and learning of user interests, thereby improving the accuracy and personalization of information push; secondly, the user interest learning of the basic user behavior mining network is carried out by using a progressive training process, which can gradually improve the performance and stability of the network, and ensure that the generated target user behavior mining network has higher quality; furthermore, by monitoring the training results of the first progressive training process in the second progressive training node, and carrying out the user interest learning of the second progressive training process under the stable state of network parameters, the dynamic adjustment and optimization of the training process can be realized, and the training efficiency and effect can be improved; finally, based on the user interest point map of the target user behavior sequence mined by the target user behavior mining network, the user interest can be grasped more accurately, and the user can be provided with an information push service that better meets their needs, thereby improving the user experience. Thus, the accuracy and personalization of information push can be effectively improved, and the user experience can be improved.

[0077] In a possible implementation, after step S130, the method further includes:

[0078] Step A110: When it is detected that the training result indicates that the first progressive training process does not reach a stable state of network parameters, the training progress state of the first progressive training process is monitored.

[0079] Step A120, when it is monitored that the training progress status meets the training validity requirement, the training result is sampled in the set monitoring node until the training result indicates that the network parameters are in a stable state under the first progressive training process.

[0080] Step A130, when it is monitored that the training progress status does not meet the training effectiveness requirement, the first progressive training process is removed, and user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

[0081] In this embodiment, the server completes monitoring of the training results of the first progressive training process in the second progressive training node. At this time, according to different monitoring results, the server will adopt different processing strategies to ensure the effectiveness and efficiency of the training.

[0082] For example, suppose the server finds that the training results of the first progressive training process have not reached a stable state of network parameters, which means that the network needs further training and adjustment. In order to better understand the training situation, the server begins to monitor the training progress of the first progressive training process, including monitoring the change trends of indicators such as loss function value, accuracy, recall rate, etc. during the training process, as well as the gradient update and learning rate changes.

[0083] If the server detects that the training progress meets the effective training requirements (such as the loss function continues to decrease, the accuracy rate steadily increases, etc.), it will continue to sample the training results in the set monitoring nodes to observe the changes in network parameters in more detail. The server will regularly save model snapshots during the training process so that it can be rolled back or compared and analyzed when necessary. This process will continue until the training results indicate that the network parameters are stable under the first progressive training process.

[0084] However, if the server detects that the training progress of the first progressive training process does not meet the effective training requirements (for example, the loss function fluctuates greatly, the accuracy rate does not improve significantly or even decreases), it will determine that the current training strategy may have problems or is no longer effective. In this case, in order to avoid wasting computing resources and time, the server will decide to remove the first progressive training process.

[0085] After removing the first progressive training process, the server will directly perform user interest learning of the second progressive training process on the basic user behavior mining network in the second progressive training node. This may be because the design of the first progressive training process is flawed, or the data and algorithms used are no longer applicable to the current task and environment. By switching to the second progressive training process, the server hopes to use different strategies and methods to optimize network performance and better mine user interests.

[0086] In the second progressive training process, the server can use different network structures, learning rate scheduling strategies, regularization methods, etc. to improve the training effect. At the same time, it will continue to monitor the training results and progress status to ensure the effectiveness and efficiency of the training. If the second progressive training process still cannot achieve the expected results, the server can further adjust the training strategy or seek other solutions.

[0087] In a possible implementation, step A110 may include:

[0088] Step A111: Obtain training results of the first progressive training process in X monitoring nodes.

[0089] Step A112, determining the statistics of a target monitoring node among the X monitoring nodes, wherein the first progressive training process is in an unstable state of network parameters in the target monitoring node.

[0090] Step A113, when it is monitored that the statistic is less than Y, it is determined that the training progress status meets the training effectiveness requirement, and when it is monitored that the statistic is not less than Y, it is determined that the training progress status does not meet the training effectiveness requirement.

[0091] In this embodiment, during the training process of the server, monitoring the training progress status is a key step to ensure the effectiveness of the training and timely adjust the strategy. In particular, for the first progressive training process, since it may involve multiple monitoring nodes and complex network parameter changes, a detailed monitoring mechanism is required.

[0092] For example, the server will first set up X monitoring nodes, which are distributed at different stages of the first progressive training process to collect key data during the training process. Each monitoring node will record training indicators such as loss function value, accuracy, recall rate, and changes in network parameters. The server will regularly obtain training results from these monitoring nodes to conduct a comprehensive analysis of the training progress.

[0093] After obtaining the training results of X monitoring nodes, the server will further analyze the data, especially focusing on those monitoring nodes whose network parameters are in an unstable state, that is, the target monitoring nodes. For these target monitoring nodes, the server will calculate specific statistics, such as the mean and standard deviation of the loss function, in order to more accurately evaluate the training progress.

[0094] The server sets a threshold Y to determine whether the training progress status meets the training validity requirements. When the statistics of the target monitoring node are less than Y, the server considers that the training progress status is good and meets the training validity requirements, which means that the network parameters are gradually stabilizing, the training strategy is effective, and training can continue.

[0095] However, if the monitored statistic is not less than Y (i.e. greater than or equal to Y), the server will determine that the training progress is not good and does not meet the effective training requirements. This may be due to unreasonable network structure, improper learning rate setting, data quality problems, etc. In this case, the server can consider adjusting the training strategy, optimizing the network structure, or reprocessing the data to improve the training effect.

[0096] Through such a detailed monitoring mechanism, the server can promptly discover problems in the training process and take corresponding measures to adjust and optimize, thereby ensuring the effective completion of user interest learning tasks.

[0097] In a possible implementation, after step A120, when it is monitored that the training results in the set monitoring node indicate that the first progressive training process has not reached a stable state of network parameters, the first progressive training process is removed, and user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

[0098] If the training results sampled in the set monitoring node indicate that the first progressive training process fails to stabilize the network parameters, the server will decide to remove the first progressive training process and perform user interest learning of the second progressive training process on the basic user behavior mining network in the second progressive training node.

[0099] For example, suppose that after a period of sampling and analysis, the server finds that the network parameters are still in a state of large fluctuations and there is no sign of convergence. This may mean that there are some problems with the first progressive training process, such as improper learning rate setting, unreasonable network structure, etc. In order to avoid further wasting computing resources, the server decides to stop the current training process and switch to the second progressive training process.

[0100] In the second progressive training process, the server can adjust the network structure, use different optimization algorithms, or adjust the learning rate to try to improve the training effect. At the same time, the server will continue to monitor the changes in network parameters under the new training process to ensure that the training can develop in the right direction and eventually reach a stable state.

[0101] In a possible implementation, step S140 may include: performing user interest learning of the second progressive training process on the user behavior mining network generated by the first progressive training process in the second progressive training node. Or, performing user interest learning of the second progressive training process on the basic user behavior mining network in the second progressive training node.

[0102] In this embodiment, when the server continues training in the second progressive training node, it can choose to further optimize the user behavior mining network that has been generated by the first progressive training process, which means that the server will use the network structure and parameters obtained in the previous training process as a starting point, and on this basis, perform user interest learning in the second progressive training process.

[0103] For example, suppose that in the first progressive training process, the server has built a preliminary user behavior mining network that can predict the user's interest preferences based on the user's historical behavior data. However, due to some reasons (such as training not converging, performance not meeting the standard, etc.), the server decides to remove the first progressive training process and switch to the second progressive training process.

[0104] In the second progressive training node, the server will load the user behavior mining network generated by the first progressive training process, including its network structure and parameters. Then, the server will use new training data, optimization algorithms or learning strategies to further train the network. This process aims to optimize the performance of the model by adjusting the network parameters and structure, so that it can better capture the user's interest preferences.

[0105] Through the second progressive training process, the server hopes to improve the accuracy and generalization ability of the user behavior mining network, thereby providing users with more accurate and personalized recommendation services.

[0106] Another case is that the server chooses to perform user interest learning of the second progressive training process on the basic user behavior mining network in the second progressive training node, which means that the server will build a new network from scratch instead of optimizing based on the results of the first progressive training process.

[0107] For example, suppose the server encounters serious problems in the first progressive training process, such as poor data quality, unreasonable network structure design, etc., resulting in the training results failing to meet the requirements. In this case, the server may decide to abandon the network generated by the first progressive training process and directly perform the second progressive training process on the basic user behavior mining network.

[0108] In the second progressive training node, the server will reinitialize a basic user behavior mining network, which may adopt a completely new network structure, initialization parameters and training strategy. Then, the server will train the network with new training data, aiming to learn the user's interest preferences from scratch.

[0109] Through this method, the server can get rid of the problems and limitations that may exist in the first progressive training process and learn user interests in a more flexible and efficient way, which helps to improve the performance and accuracy of the model and provide users with a better recommendation experience.

[0110] In a possible implementation, the performing of user interest learning of the second progressive training process on the user behavior mining network generated by the first progressive training process in the second progressive training node includes:

[0111] Step S141, determining a first user behavior mining network generated by the first progressive training process, wherein the first user behavior mining network is a progressive training network model generated by performing user interest learning based on a plurality of template user behavior sequences S1 of the first progressive training process.

[0112] After the server completes the first progressive training process, a user behavior mining network will be generated, which we call the first user behavior mining network. This network is a progressive training network model generated after user interest learning based on multiple template user behavior sequences S1 in the first progressive training process.

[0113] For example, assume that the server has completed the first progressive training process of user shopping behavior on the e-commerce platform. In this process, the server uses the template user behavior sequence S1, which may contain user browsing, adding to cart, purchasing, and evaluation behavior data. Based on this data, the server constructs a first user behavior mining network that can predict user shopping interests through a machine learning algorithm.

[0114] Step S142: Acquire a plurality of template user behavior sequences S2 of the second progressive training process, wherein the template user behavior sequence S2 is different from the template user behavior sequence S1.

[0115] When the server is ready to perform the second progressive training process, it needs new user behavior data as input. Therefore, the server obtains multiple template user behavior sequences S2, which are different from the template user behavior sequences S1 used in the first progressive training process.

[0116] For example, in order to further improve the accuracy of user interest prediction, the server decided to introduce more types of user behavior data. Therefore, in addition to the previous shopping behavior data, the server now also obtains the user's sharing, liking, commenting and other behavior data on the social platform as the template user behavior sequence S2. These data provide a more comprehensive perspective for user interest learning.

[0117] Step S143, performing user interest learning on the first user behavior mining network according to the plurality of template user behavior sequences S2, and progressively optimizing parameter information of the first user behavior mining network during the user interest learning process until a second user behavior mining network that meets the training termination requirements is generated.

[0118] With the new template user behavior sequence S2, the server will use this data to further train and optimize the first user behavior mining network. This process aims to improve the performance of the model by adjusting the network parameters and structure so that it can more accurately capture the user's interest preferences.

[0119] For example, the server combines the newly acquired social behavior data with the original shopping behavior data and uses them as input to train the first user behavior mining network. During the training process, the server adjusts the network's weights, biases and other parameter information based on the difference between the predicted results and the actual user behavior. In this way, the model can gradually learn more about the knowledge and patterns of user interests.

[0120] In the process of learning user interests, the server will progressively optimize the parameter information of the first user behavior mining network, which means that the server will continuously adjust the structure and parameters of the network to improve its performance until a second user behavior mining network that meets the training termination requirements is generated.

[0121] For example, after multiple rounds of training and optimization, the server finds that the performance of the model has been significantly improved. Now it can more accurately predict user interest preferences and has better generalization capabilities for new user behavior data. Therefore, the server decides to stop training and save the current network state as the second user behavior mining network. This new network will be used in subsequent user interest prediction and recommendation tasks.

[0122] In a possible implementation, before step S120, the method further includes:

[0123] Step B110, obtaining a plurality of template user behavior sequences S3 of a first digital financial page and a plurality of template user behavior sequences S4 of a second digital financial page, wherein the plurality of digital financial pages include the first digital financial page and the second digital financial page.

[0124] In this embodiment, the server first collects multiple template user behavior sequences S3 from the first digital financial page, which may include user behavior data such as clicks, browsing, and transactions on the page. Similarly, the server also collects multiple template user behavior sequences S4 from the second digital financial page, which reflect different behavior patterns of users on the page.

[0125] For example, suppose the first digital financial page is an online banking platform, and the second digital financial page is a stock trading platform. The server extracts the user's behavior data over a period of time from these two pages through log analysis, user tracking and other technical means, forming template user behavior sequences S3 and S4, which record in detail how the user interacts with the page, which financial products are viewed, and which trading operations are performed.

[0126] Step B120 , integrating features of the plurality of template user behavior sequences S3 and the plurality of template user behavior sequences S4 to generate a plurality of template user behavior sequences.

[0127] The server then performs feature integration on the template user behavior sequences S3 and S4 collected from two different digital financial pages. This process aims to fuse the behavior data from different data sources together to form a more comprehensive and unified user behavior representation. Through feature integration, the server can capture the comprehensive behavior patterns of users on multiple financial pages.

[0128] For example, the server uses feature engineering technology to process user behavior data collected from online banking platforms and stock trading platforms. For example, features such as the number of user logins, transaction amounts, and browsing time can be extracted from the two data sources and preprocessed by normalization and deduplication. The server then fuses these features together to form a new, more comprehensive set of template user behavior sequences, each of which contains comprehensive behavior information of users on two financial pages.

[0129] Step S120 may include: in the first progressive training node, performing user interest learning of the first progressive training process on the basic user behavior mining network according to a plurality of the template user behavior sequences.

[0130] In the first progressive training node, the server will use multiple template user behavior sequences after feature integration as input data to train the basic user behavior mining network. This process aims to enable the network to automatically learn the user's interest preferences and behavior patterns through machine learning algorithms. By continuously adjusting the network parameters and structure, the server hopes to improve the prediction accuracy and generalization ability of the model.

[0131] For example, the server inputs the integrated template user behavior sequence into the basic user behavior mining network, which can be a deep neural network, convolutional neural network, or other types of machine learning models. During the training process, the server updates the network parameters such as weights and biases based on the difference between the prediction results of the basic user behavior mining network and the actual user behavior. Through multiple rounds of training and optimization, the basic user behavior mining network gradually learns the user's interest preferences and behavior patterns, so that it can more accurately predict the user's future behavior. Finally, when the basic user behavior mining network reaches certain performance requirements, the server will stop training and save the current network state as the result of the first progressive training process.

[0132] In a possible implementation manner, before step B110, the method further includes:

[0133] Step B101, performing a first confusion perturbation strategy processing on multiple template user behavior sequences S3 of the first digital financial page, and loading the multiple template user behavior sequences S3 processed by the first confusion perturbation strategy into the second digital financial page.

[0134] Before the server starts collecting user behavior data, the first obfuscation perturbation strategy is first processed on multiple template user behavior sequences S3 on the first digital financial page. This processing is to increase the diversity and complexity of the data while protecting user privacy. The processed sequences are loaded onto the second digital financial page to observe the user's reaction to these obfuscated behavior patterns.

[0135] Assume that the server has user behavior data on a first digital financial page (such as an online banking platform). In order to protect user privacy and increase data diversity, the server uses a first obfuscation perturbation strategy to process this data, which may include randomizing the occurrence time of certain behaviors, replacing or deleting certain sensitive information, adding noise, etc. The processed behavior sequence is loaded onto a second digital financial page (such as an investment consulting platform) and used as test data.

[0136] Step B102, obtaining a global template user behavior sequence fed back by the second digital financial page, and multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, wherein the global template user behavior sequence is a template user behavior sequence generated after the multiple template user behavior sequences S3 processed by the first confusion perturbation strategy and the multiple template user behavior sequences S4 of the second digital financial page are processed by the second confusion perturbation strategy.

[0137] The server collects feedback data from the second digital financial page, including global template user behavior sequences, which are generated after further processing (second obfuscation perturbation strategy) of all obfuscated behavior sequences loaded on the page. At the same time, the server also obtains multiple template user behavior sequences S4 processed by the second obfuscation perturbation strategy, which reflect the actual behavior patterns of users on the second digital financial page.

[0138] For example, on the second digital financial page, the server collects data on the user's interaction with the obfuscated behavior sequence. After the data is processed by the second obfuscation perturbation strategy, a global template user behavior sequence and multiple template user behavior sequences S4 are generated. The global template user behavior sequence is a sequence that integrates all user behavior patterns, while S4 is a collection of sequences that reflect the behavior patterns of a single user or user group.

[0139] Step B103, performing decoding conversion of the first confusion perturbation strategy on the global template user behavior sequence and the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, generating multiple template user behavior sequences S3 processed by the second confusion perturbation strategy, and multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, wherein the template user behavior sequence S3 covers the multiple template user behavior sequences S3 processed by the second confusion perturbation strategy, and the template user behavior sequence S4 covers the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy.

[0140] In order to match the feedback data with the original user behavior data, the server needs to perform decoding conversion of the first confusion perturbation strategy on the global template user behavior sequence and multiple template user behavior sequences S4 processed by the second confusion perturbation strategy. This process aims to restore the original state of the data for subsequent analysis and processing.

[0141] For example, the server uses the decoding algorithm corresponding to the first confusion perturbation strategy to process the global template user behavior sequence and S4. The decoding process may include steps such as removing noise, restoring replaced or deleted information, and adjusting the time sequence. The processed data will be closer to the original user behavior data, which is convenient for the server to perform subsequent feature integration and model training.

[0142] Step B120 may include: determining a shared template user behavior sequence of the multiple template user behavior sequences S3 processed by the second confusion perturbation strategy and the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, and outputting the shared template user behavior sequence as the template user behavior sequence.

[0143] After completing the data collection and processing, the server begins to integrate the features of the template user behavior sequences from different digital financial pages. It first determines the shared template user behavior sequences between the multiple template user behavior sequences S3 and S4 processed by the second confusion perturbation strategy. These shared sequences reflect the common behavior patterns of users on two different pages. Then, the server outputs these shared sequences as new template user behavior sequences for subsequent user interest learning and model training.

[0144] For example, by analyzing the processed S3 and S4 sequence sets, the server discovered some behavioral patterns that frequently appeared on both pages (such as regularly checking account balances, paying attention to specific financial products, etc.). These shared behavioral patterns were extracted and integrated into a new template user behavior sequence set. Each sequence in this set represents the common behavioral characteristics of users on two digital financial pages, providing strong support for subsequent user interest learning and personalized recommendations.

[0145] In a possible implementation, after step S140, see Figure 2 , the method further comprises:

[0146] Step S150, obtaining a user behavior mining network that completes user interest learning.

[0147] Step S160, performing network performance verification on the user behavior mining network that has completed user interest learning according to a plurality of verification template user behavior sequences, and generating network performance indicators corresponding to the user behavior mining network that has completed user interest learning.

[0148] Step S170: when it is monitored that the network performance indicator is less than a threshold indicator parameter, activating the next progressive training process of the first progressive training process until the target user behavior mining network is generated.

[0149] Step S180: when it is monitored that the network performance index is not less than the threshold index parameter, the user behavior mining network that completes user interest learning is output as the target user behavior mining network.

[0150] In this embodiment, after the server completes the user interest learning of the second progressive training process in the second progressive training node, it will obtain the current user behavior mining network, which has been learned and optimized to have the ability to more accurately predict user interests and behaviors. For example, suppose that the server is conducting a user behavior mining project on financial products. In the second progressive training node, the server uses a large amount of user behavior data to train the network so that it can better understand the user's interest preferences and purchasing habits. After the training is completed, the server obtains this optimized user behavior mining network.

[0151] To evaluate the performance of the user behavior mining network, the server tests the network using a set of validation template user behavior sequences, which are pre-defined and represent typical user behavior patterns. The server compares the network's predictions with these sequences and calculates network performance metrics such as accuracy, recall, etc.

[0152] For example, the server selects a set of representative user behavior sequences from the validation data set as validation templates. Then, the user behavior mining network is tested using these templates to see whether the network can accurately predict user behaviors in these sequences. After the test is completed, the server calculates the network's performance indicators based on the difference between the predicted results and the actual sequences.

[0153] If the monitoring results of the server show that the network performance index is lower than the preset threshold index parameter, it means that the current network performance is not enough to meet the requirements. In order to further improve the performance, the server will activate the next progressive training process of the first progressive training process to further train and optimize the network. This process will continue until a target user behavior mining network that meets the performance requirements is generated.

[0154] For example, after analyzing the network performance indicators, the server finds that some key indicators are lower than the preset threshold. In order to improve network performance, the server decides to start the next progressive training process. In this new training process, the server can use a more complex network structure, introduce more features, or use a more advanced training algorithm to optimize the network. This process will be repeated until the network performance reaches the preset target.

[0155] If the monitoring results of the server show that the network performance index reaches or exceeds the preset threshold index parameters, it means that the current network has sufficient performance to meet the demand. In this case, the server will output the user behavior mining network that has completed user interest learning as the target user behavior mining network for subsequent application and deployment.

[0156] For example, after multiple rounds of training and verification, the server finds that the performance of the user behavior mining network has stabilized and reached the preset goal. At this time, the server decides to save the current network as the target user behavior mining network and output it to the relevant application system for deployment and use. In this way, the application system can use this network to more accurately predict the user's interests and behaviors, thereby providing better personalized services and recommendations.

[0157] In a possible implementation, step S170 may include:

[0158] Step S171, monitoring the past training results of the forward progressive training node of the current round of progressive training node within the current round of progressive training node, the current round of progressive training node is the next progressive training process of the first progressive training process, and the current round of progressive training node is the progressive training node of the current round of progressive training process.

[0159] The server uses a progressive training process when training the user behavior mining network. In the progressive training node of the current round (called "this round"), the server will first review and monitor the training results of the previous round (that is, the "forward progressive training node"), which contains various data and information in the previous round of training, such as changes in network parameters, convergence of training errors, etc.

[0160] For example, suppose the server is conducting the third round of user behavior mining network training. Before this round begins, the server will first check the result records of the second round of training, which may include the adjustment of network weights during training, the downward trend graph of the loss function, etc., to evaluate the effect and stability of the second round of training.

[0161] Step S172, when it is monitored that the past training results indicate that the forward progressive training nodes of the current round of progressive training nodes do not have a stable network parameter state, the next progressive training process of the current round of progressive training process is output as the current round of progressive training process.

[0162] If the server finds that the network parameters have not reached a stable state during the training process (such as the loss function continues to fluctuate, the weight parameters change dramatically, etc.) when monitoring the past training results of the forward progressive training node, then the server will determine that the previous round of training has not been completed or has not achieved the expected results.

[0163] For example, when analyzing the results of the second round of training, the server found that the loss function did not converge to a lower level, but fluctuated repeatedly within a certain range, which indicated that the network parameters had not reached a stable state and the second round of training might require more iterations or adjustment of the training strategy.

[0164] When the current round of training fails to reach a stable state, the server will take the next round (i.e. "the next progressive training process of this round of progressive training process") as a new training starting point and continue to optimize the training. This usually means adjusting the learning rate, increasing the amount of training data, or introducing new training techniques while maintaining the current network structure.

[0165] For example, because the second round of training did not reach a stable state, the server decided to start the third round of training. In the third round of training, the server can reduce the learning rate to reduce parameter fluctuations during training, or add more user behavior data to enhance the generalization ability of the network.

[0166] Step S173, when it is monitored that the past training results indicate that a network parameter stable state appears under the forward progressive training node of the current progressive training node, user interest learning of the current progressive training process is performed within the current progressive training node to generate a user behavior mining network that completes user interest learning in this round.

[0167] If the server finds that the network parameters have stabilized (such as a steady decrease in the loss function and small changes in the weight parameters) when monitoring the past training results of the forward progressive training nodes, the server will determine that the previous round of training has been completed and achieved the expected results.

[0168] For example, when analyzing the results of the second round of training, the server found that the loss function had stabilized at a lower level and the adjustment of the weight parameters had also become gentler, which indicated that the network parameters had reached a stable state and the second round of training could be considered successful.

[0169] In the current round of progressive training nodes, the server will use the optimized user behavior mining network to learn user interests. This process usually involves using a large amount of user behavior data to train the network so that it can better understand and predict user interests and behaviors. After the training is completed, the server will get an updated user behavior mining network.

[0170] In the third round of training, the server uses the optimized network structure and parameter settings, as well as more user behavior data for training. Through this round of training, the server expects the network to better capture the user's interest characteristics and behavior patterns. After the training is completed, the server obtains a user behavior mining network with stronger performance.

[0171] Step S174, verifying the user behavior mining network that has completed user interest learning in this round according to a plurality of the verification template user behavior sequences, and generating a network performance index for this round corresponding to the user behavior mining network that has completed user interest learning in this round.

[0172] In order to evaluate the performance of the user behavior mining network obtained in the current round of training, the server will test the network using a set of predefined verification template user behavior sequences. These sequences represent typical user behavior patterns and can be used to verify whether the network can accurately predict the user's actual behavior. After the test is completed, the server will calculate the network's performance indicators based on the difference between the predicted results and the actual sequence.

[0173] For example, the server selected a set of representative user behavior sequences from the validation data set as validation templates. It then used these templates to test the user behavior mining network obtained from the third round of training. The test results showed that the network's accuracy in predicting user behavior has been significantly improved.

[0174] Step S175, when it is monitored that the network performance indicator of this round is less than the threshold indicator parameter, the next progressive training process of the progressive training process of this round is output as the progressive training process of this round.

[0175] If the server finds that the performance index of the user behavior mining network obtained by the current round of training is lower than the preset threshold value (such as insufficient accuracy, too low recall rate, etc.) when evaluating the performance of the user behavior mining network, then the server will determine that the current network still needs further optimization.

[0176] For example, although the performance of the network improved after the third round of training, the evaluation results of the server showed that some key performance indicators of the network still did not reach the preset threshold values, which indicated that the network still needed further optimization and training to meet actual needs.

[0177] When the previous round of training fails to achieve the expected performance, the server will use the next round as a new training starting point and continue to optimize the training. This usually means adjusting the training strategy, introducing new features, or improving the network structure while maintaining the current network structure.

[0178] For example, since the network performance after the third round of training still does not meet the expected target, the server decides to start the fourth round of training. In the fourth round of training, the server can try to introduce more user feature information, adjust the network structure, or use more advanced training algorithms to further improve the performance of the network.

[0179] Step S176, when it is monitored that the network performance index of this round is not less than the threshold index parameter, the user behavior mining network that completes the user interest learning in this round is output as the target user behavior mining network.

[0180] If the server finds that the performance index of the network has reached or exceeded the preset threshold value when evaluating the performance of the user behavior mining network obtained in the current round of training, the server will determine that the current network has sufficient performance to meet actual needs.

[0181] For example, after the fourth round of training, the evaluation results of the server showed that the performance indicators of the network have reached the preset threshold value or even higher, which indicates that the network has sufficient performance to meet the needs of user behavior mining.

[0182] When the evaluation results of the server show that the user behavior mining network obtained from the current round of training has sufficient performance, the server will output this network as the final target user behavior mining network, which will be used in subsequent user behavior prediction, personalized recommendation and other tasks.

[0183] For example, since the network performance after the fourth round of training has met the actual needs, the server decides to save this network as the target user behavior mining network. Subsequent application systems can use this network to more accurately predict user interests and behaviors, thereby providing better personalized services and recommendations.

[0184] Figure 3The information push system 100 based on user behavior sequence shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the information push system 100 based on user behavior sequence may also include a transceiver 1004, which may be used for data interaction between the server and other servers, such as data transmission and / or data reception. It should be noted that the transceiver 1004 is not limited to one in actual scheduling, and the structure of the information push system 100 based on user behavior sequence does not constitute a limitation on the embodiments of the present application.

[0185] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0186] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0187] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to carry or store program code and can be read by a computer, without limitation herein.

[0188] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.

[0189] An embodiment of the present application provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0190] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the implementation order of these steps is not limited to the order indicated by the arrows. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders based on demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to the actual implementation scenario, and some or all of these sub-steps or stages may be executed at the same time, and each sub-step or stage in these sub-steps or stages may also be executed at different times respectively. In different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0191] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A method for pushing information based on user behavior sequence, characterized in that: The method comprises: Obtaining a user interest learning task generated for a basic user behavior mining network, the user interest learning task is used to instruct the basic user behavior mining network to perform multiple progressive training processes of user interest learning until a target user behavior mining network is generated, and after mining a user interest point map of a target user behavior sequence of an input target user based on the target user behavior mining network, information is pushed to the target user according to the user interest point map, wherein the target user behavior mining network is a progressive training network model generated by performing user interest learning based on template user behavior sequences of multiple digital financial pages; According to the user interest learning task, performing user interest learning of a first progressive training process on the basic user behavior mining network in a first progressive training node, wherein a plurality of the progressive training processes include the first progressive training process, and the first progressive training node is a progressive training node of the first progressive training process; Monitoring the training result of the first progressive training process in a second progressive training node, where the second progressive training node is a progressive training node of the second progressive training process, a plurality of the progressive training processes include the second progressive training process, and the first progressive training process is a forward progressive training node of the second progressive training node; When it is monitored that the training result indicates that a network parameter is in a stable state under the first progressive training process, performing user interest learning of the second progressive training process in the second progressive training node; Before the step of performing user interest learning of a first progressive training process on the basic user behavior mining network in the first progressive training node, the method further includes: Acquire multiple template user behavior sequences of a first digital financial page S3 and multiple template user behavior sequences of a second digital financial page S4, wherein the multiple digital financial pages include the first digital financial page and the second digital financial page; Integrate the features of the plurality of template user behavior sequences S3 and the plurality of template user behavior sequences S4 to generate a plurality of template user behavior sequences; The performing of a first progressive training process of user interest learning on the basic user behavior mining network in the first progressive training node includes: In the first progressive training node, the basic user behavior mining network is subjected to the user interest learning of the first progressive training process according to the plurality of template user behavior sequences; Before the step of acquiring multiple template user behavior sequences S3 of the first digital financial page and multiple template user behavior sequences S4 of the second digital financial page, the method further includes: Performing a first confusion perturbation strategy processing on the plurality of template user behavior sequences S3 of the first digital financial page, and loading the plurality of template user behavior sequences S3 processed by the first confusion perturbation strategy into the second digital financial page; Acquire a global template user behavior sequence fed back by the second digital financial page and a plurality of template user behavior sequences S4 processed by the second confusion perturbation strategy, wherein the global template user behavior sequence is a template user behavior sequence generated after the plurality of template user behavior sequences S3 processed by the first confusion perturbation strategy and the plurality of template user behavior sequences S4 of the second digital financial page are processed by the second confusion perturbation strategy; Performing decoding conversion of the first confusion perturbation strategy on the global template user behavior sequence and the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, generating multiple template user behavior sequences S3 processed by the second confusion perturbation strategy, and multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, wherein the template user behavior sequence S3 covers the multiple template user behavior sequences S3 processed by the second confusion perturbation strategy, and the template user behavior sequence S4 covers the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy; The step of integrating the features of the plurality of template user behavior sequences S3 and the plurality of template user behavior sequences S4 to generate a plurality of template user behavior sequences includes: Determine a shared template user behavior sequence of the multiple template user behavior sequences S3 processed by the second confusion perturbation strategy and the multiple template user behavior sequences S4 processed by the second confusion perturbation strategy, and output the shared template user behavior sequence as the template user behavior sequence.

2. The information push method based on user behavior sequence according to claim 1 is characterized in that: After monitoring the training result of the first progressive training process in the second progressive training node, the method further includes: When it is monitored that the training result indicates that the first progressive training process does not show a stable state of network parameters, monitoring the training progress state of the first progressive training process; When it is monitored that the training progress state meets the training validity requirement, sampling the training result in the set monitoring node until the training result indicates that the network parameter is in a stable state under the first progressive training process; When it is monitored that the training progress status does not meet the training effectiveness requirement, the first progressive training process is removed, and user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

3. The information push method based on user behavior sequence according to claim 2 is characterized in that: The monitoring of the training progress status of the first progressive training process includes: Obtaining training results of the first progressive training process in X monitoring nodes; Determine a statistic of a target monitoring node among the X monitoring nodes, wherein the first progressive training process is in an unstable state of network parameters in the target monitoring node; When it is monitored that the statistic is less than Y, it is determined that the training progress status meets the training effectiveness requirement, and when it is monitored that the statistic is not less than Y, it is determined that the training progress status does not meet the training effectiveness requirement.

4. The information push method based on user behavior sequence according to claim 2 is characterized in that: The training result is sampled in the setting monitoring node until the training result indicates that a network parameter is in a stable state under the first progressive training process, and the method further includes: When it is monitored that the training results in the set monitoring node indicate that the first progressive training process has not reached a stable state of network parameters, the first progressive training process is removed, and user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

5. The information push method based on user behavior sequence according to claim 1, characterized in that: The performing the user interest learning of the second progressive training process in the second progressive training node includes: Performing user interest learning of the second progressive training process on the user behavior mining network generated by the first progressive training process in the second progressive training node; Alternatively, user interest learning of the second progressive training process is performed on the basic user behavior mining network in the second progressive training node.

6. The information push method based on user behavior sequence according to claim 5, characterized in that: The performing the user interest learning of the second progressive training process on the user behavior mining network generated by the first progressive training process in the second progressive training node includes: Determine a first user behavior mining network generated by the first progressive training process, where the first user behavior mining network is a progressive training network model generated by performing user interest learning based on a plurality of template user behavior sequences S1 of the first progressive training process; Acquire a plurality of template user behavior sequences S2 of the second progressive training process, wherein the template user behavior sequence S2 is different from the template user behavior sequence S1; The first user behavior mining network is used to learn user interests based on the plurality of template user behavior sequences S2, and parameter information of the first user behavior mining network is progressively optimized during the user interest learning process until a second user behavior mining network that meets the training termination requirements is generated.

7. The information push method based on user behavior sequence according to any one of claims 1 to 6, characterized in that: After performing user interest learning of the second progressive training process in the second progressive training node, the method further includes: Obtain the user behavior mining network that completes user interest learning; Performing network performance verification on the user behavior mining network that has completed user interest learning according to multiple verification template user behavior sequences, and generating network performance indicators corresponding to the user behavior mining network that has completed user interest learning; When it is monitored that the network performance indicator is less than a threshold indicator parameter, activating the next progressive training process of the first progressive training process until the target user behavior mining network is generated; When it is monitored that the network performance index is not less than the threshold index parameter, the user behavior mining network that completes the user interest learning is output as the target user behavior mining network; The step of activating the next progressive training process of the first progressive training process until the target user behavior mining network is generated includes: Monitoring, within a current round of progressive training nodes, past training results of a forward progressive training node of the current round of progressive training nodes, the current round of progressive training nodes being the next progressive training process of the first progressive training process, and the current round of progressive training nodes being the progressive training nodes of the current round of progressive training processes; When it is monitored that the past training results indicate that the forward progressive training node of the current round of progressive training nodes does not have a stable network parameter state, outputting the next progressive training process of the current round of progressive training process as the current round of progressive training process; When it is monitored that the past training results indicate that a network parameter stable state occurs under the forward progressive training node of the current progressive training node, user interest learning of the current progressive training process is performed in the current progressive training node to generate a user behavior mining network that completes user interest learning in this round; Verifying the user behavior mining network that has completed user interest learning in this round according to the plurality of verification template user behavior sequences, and generating a network performance indicator for this round corresponding to the user behavior mining network that has completed user interest learning in this round; When it is monitored that the network performance indicator of the current round is less than the threshold indicator parameter, outputting the next progressive training process of the current round of progressive training process as the progressive training process of the current round; When it is monitored that the network performance index of this round is not less than the threshold index parameter, the user behavior mining network that completes user interest learning in this round is output as the target user behavior mining network.

8. An information push system based on user behavior sequence, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the information push method based on user behavior sequence described in any one of claims 1 to 7 is implemented.

Images