Training of resource object interaction prediction model and resource object information processing method

Abstract

The application relates to a resource object interaction prediction model training method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: obtaining a resource object representation set for a target resource object; the resource object representation set is used for training a client user model deployed on a client of each user account to obtain model training gradient information corresponding to each client user model; aggregating the model training gradient information to obtain aggregated gradient information; training a resource object model and each client user model according to the aggregated gradient information; determining a resource object interaction prediction model according to the resource object model and each client user model; and the resource object interaction prediction model is used for determining prediction probability information of each user account initiating interaction with a resource object to be predicted. The method can improve the training efficiency of the resource object interaction prediction model.

Description

Technical Field

[0001] This application relates to the fields of computer technology and financial technology, and in particular to a method, apparatus, computer device, storage medium and computer program product for training a resource object interaction prediction model and processing resource object information. Background Technology

[0002] With the development of computer technology, more and more users are operating various virtual resources on resource transfer platforms. For example, purchasing wealth management products on banking product platforms.

[0003] Currently, various business units often analyze user interaction records on resource transfer platforms to build models that can determine users' preferences for various virtual resources, thereby accurately recommending virtual resources to different users. However, when sending local user data from each client to the server for model training, it often leads to the leakage of user information, and the measures taken to protect user information result in low efficiency in model training.

[0004] Therefore, traditional technologies suffer from low model training efficiency for predicting resource object interactions. Summary of the Invention

[0005] Based on this, it is necessary to provide a training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for a resource object interaction prediction model that can improve the model training efficiency of the resource object interaction prediction model, in order to address the above-mentioned technical problems.

[0006] A method for training a resource object interaction prediction model, characterized in that the method includes:

[0007] Obtain the resource object representation set for the target resource object; the target resource object is the resource object that each user account has interacted with in the historical time period; the resource object representation set is used to characterize the resource object interaction features of the user account; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model;

[0008] Aggregate the training gradient information of each model to obtain aggregated gradient information;

[0009] Based on the aggregated gradient information, the resource object model and each client user model are trained; the resource object model is used to output the resource object representation corresponding to the input resource object;

[0010] Based on the resource object model and each client user model, a resource object interaction prediction model is determined; the resource object interaction prediction model is used to determine the prediction probability information of each user account initiating an interaction with the resource object to be predicted.

[0011] In one embodiment, obtaining the resource object representation set for the target resource object includes:

[0012] Retrieve resource object data for the target resource object;

[0013] Input the resource object data of the target resource object into the initial resource object model to obtain the resource object representation set.

[0014] In one embodiment, the method further includes:

[0015] Input the resource object representation set into each client user model to obtain the client user representation corresponding to each client;

[0016] Obtain the candidate resource object representation for each client. Based on the client user representation and the candidate resource object representation, determine the prediction probability information of the candidate resource object on the client. The prediction probability information is used to determine the model training gradient information.

[0017] In one embodiment, the resource object model and each client user model are trained based on aggregated gradient information, including:

[0018] Based on the aggregated gradient information, the initial resource object model is updated using backpropagation training to obtain the updated resource object model.

[0019] The updated resource object representation set is output from the updated resource object model;

[0020] The updated resource object representation set is sent to each client to perform a new round of model training on the updated client user model on each client, and to obtain the aggregated gradient information corresponding to the new round of model training; the updated client user model is obtained by updating the client user model using the aggregated gradient information;

[0021] The process involves updating the initial resource object model using backpropagation training based on the aggregated gradient information, obtaining the updated resource object model, and continuing until the resource object model and each client user model meet the training termination conditions.

[0022] In one embodiment, the training gradient information of each client model is aggregated to obtain aggregated gradient information, including:

[0023] The model training gradient information from each client is aggregated to obtain at least one local aggregated gradient information; the local aggregated gradient information is obtained by aggregating a preset portion of the model training gradient information according to a preset aggregation method.

[0024] The aggregated gradient information is obtained by globally aggregating the local aggregated gradient information.

[0025] In one embodiment, the method further includes:

[0026] Sending an interactive prediction command to the target client; the interactive prediction command is used to instruct the target client to input the resource object to be predicted into the resource object interactive prediction model to obtain the prediction probability information corresponding to the resource object to be predicted.

[0027] In one embodiment, the target client is configured to determine the prediction probability information of the user account in the target client initiating an interaction with the resource object to be predicted, based on the resource object representation corresponding to the resource object to be predicted and the user behavior representation of the target client.

[0028] A method for processing resource object information, characterized in that the method includes:

[0029] Obtain the resource object to be predicted;

[0030] The resource object to be predicted is input into a pre-trained resource object interaction prediction model to obtain the prediction probability information corresponding to the resource object to be predicted; the prediction probability information is used to determine the probability that the user account in the target client will initiate an interaction with the resource object to be predicted.

[0031] Among them, the pre-trained resource object interaction prediction model is determined based on the resource object model and each client user model; the resource object model and each client user model are obtained based on aggregated gradient information; the aggregated gradient information is obtained by aggregating the training gradient information of each model; the training gradient information of each model is obtained by training the client user model deployed on the client where each user account is located using the resource object representation set; the resource object representation set is used to characterize the resource object interaction features corresponding to the interaction between the user account and the target resource object in a historical time period; the target resource object is the resource object that each user account has interacted with in a historical time period.

[0032] A training device for a resource object interaction prediction model, characterized in that the device comprises:

[0033] The acquisition module is used to acquire the resource object representation set for the target resource object; the target resource object is the resource object that each user account has interacted with in the historical time period; the resource object representation set is used to characterize the resource object interaction features of the user account; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model;

[0034] The aggregation module is used to aggregate the training gradient information of each model to obtain aggregated gradient information.

[0035] The training module is used to train the resource object model and each client user model based on the aggregated gradient information; the resource object model is used to output the resource object representation corresponding to the input resource object.

[0036] The determination module is used to determine the resource object interaction prediction model based on the resource object model and each client user model; the resource object interaction prediction model is used to determine the prediction probability information of each user account initiating interaction with the resource object to be predicted.

[0037] A resource object information processing device, characterized in that the device comprises:

[0038] The acquisition module is used to acquire the resource objects to be predicted.

[0039] The input module is used to input the resource object to be predicted into the pre-trained resource object interaction prediction model to obtain the prediction probability information corresponding to the resource object to be predicted; the prediction probability information is used to determine the probability that the user account in the target client will initiate an interaction with the resource object to be predicted.

[0040] Among them, the pre-trained resource object interaction prediction model is determined based on the resource object model and each client user model; the resource object model and each client user model are obtained based on aggregated gradient information; the aggregated gradient information is obtained by aggregating the training gradient information of each model; the training gradient information of each model is obtained by training the client user model deployed on the client where each user account is located using the resource object representation set; the resource object representation set is used to characterize the resource object interaction features corresponding to the interaction between the user account and the target resource object in a historical time period; the target resource object is the resource object that each user account has interacted with in a historical time period.

[0041] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the method described above.

[0042] A computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the above-described method.

[0043] A computer program product includes a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the above-described method.

[0044] The training method, apparatus, computer equipment, storage medium, and computer program product of the aforementioned resource object interaction prediction model involve: acquiring a resource object representation set for a target resource object; the target resource object being the resource objects interacted with by each user account over a historical time period; the resource object representation set representing the resource object interaction characteristics of user accounts; training client-side user models deployed on the clients of each user account using the resource object representation set to obtain model training gradient information corresponding to each client-side user model; aggregating the training gradient information of each model to obtain aggregated gradient information; training the resource object model and each client-side user model based on the aggregated gradient information; the resource object model outputting a resource object representation corresponding to the input resource object; and determining the resource object interaction prediction model based on the resource object model and each client-side user model. The prediction model is used to determine the probability of each user account initiating an interaction with the resource object to be predicted. This allows the resource object interaction prediction model to be divided into a resource object model on a central server and user models on each client. The more robust resource object model is trained on the central server, while the lighter client-side user models are trained on each client, reducing the computational overhead on each client during model training. Furthermore, by aggregating the model training gradient information from each client, the model training gradient information is encrypted. Simultaneously, since each client uses the resource object representation set distributed by the central server to train its user model, compared to directly deploying the resource object model on each client, the communication overhead between the central server and each client during model training is reduced, thereby improving the training efficiency of the resource object interaction prediction model. Attached Figure Description

[0045] Figure 1 This is an application environment diagram of the training method for the resource object interaction prediction model in one embodiment;

[0046] Figure 2 This is a flowchart illustrating the training method of a resource object interaction prediction model in one embodiment;

[0047] Figure 3 This is an architecture diagram of a resource object interaction prediction model in one embodiment;

[0048] Figure 4This is a schematic diagram of a parameter aggregation mechanism in one embodiment;

[0049] Figure 5 This is a flowchart illustrating the training method of the resource object interaction prediction model in another embodiment;

[0050] Figure 6 This is a flowchart illustrating a resource object interaction prediction method in one embodiment;

[0051] Figure 7 This is a structural block diagram of a training device for a resource object interaction prediction model in one embodiment;

[0052] Figure 8 This is a structural block diagram of a resource object information processing device in one embodiment;

[0053] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] It should be noted that the present application discloses a training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for a resource object interaction prediction model, which can be applied to the field of financial technology.

[0056] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0057] The training method for the resource object interaction prediction model provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, the central server 102 communicates with the client 104 via a network. The central server 102 acquires a resource object representation set for the target resource object; the target resource object is the resource object that each user account has interacted with in a historical time period; the resource object representation set is used to characterize the resource object interaction features of user accounts; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model; the central server 102 aggregates the training gradient information of each model to obtain aggregated gradient information; the central server 102 trains the resource object model and each client user model based on the aggregated gradient information; the resource object model is used to output the resource object representation corresponding to the input resource object; the central server 102 determines the resource object interaction prediction model based on the resource object model on the central server 102 and the client user models on each client 104; the resource object interaction prediction model is used to determine the prediction probability information of each user account initiating interaction with the resource object to be predicted. The central server 102 can be implemented using a standalone server or a server cluster composed of multiple servers. Client 104 can be, but is not limited to, various personal computers, laptops, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc.

[0058] In one embodiment, such as Figure 2 As shown, a training method for a resource object interaction prediction model is provided, which is then applied to... Figure 1 Taking the central server 102 as an example, the following steps are included:

[0059] Step S202: Obtain the resource object representation set for the target resource object; the target resource object is the resource object that each user account has interacted with in the historical time period; the resource object representation set is used to characterize the resource object interaction features of the user account; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model.

[0060] In this context, the resource object can refer to the resource to be acquired. In practical applications, the resource object can refer to financial products, wealth management products, etc. recommended by the bank.

[0061] In this context, the resource object representation can be the vector representation corresponding to the resource object. In practical applications, the resource object representation can be the financial product representation. When a financial product n is given, the financial product representation n can be learned from the content of the financial product using a financial product model.

[0062] The resource object representation set can refer to a dataset composed of resource object representations.

[0063] In this context, a user account can refer to an account used for resource transfer on a resource transfer platform. In practical applications, a resource transfer platform can refer to a bank transaction system, and a user account can refer to a customer account used for transactions within that system.

[0064] The historical time period can refer to any time period before the current point in time.

[0065] Among them, the resource object interaction characteristics can refer to the interaction characteristics of the resource object when the user account initiates an interaction with the resource object. In practical applications, the resource object interaction characteristics can refer to the financial product interaction characteristics. The financial product interaction characteristics represent the interaction characteristic information of the financial product when the user account clicks on the financial product.

[0066] In this context, a client-side user model can be a model that understands user representation from data generated when a user account interacts with resource objects over a historical period. User representation can refer to the characteristics of a user's interactive behavior. In practical applications, a client-side user model can be a model that models user click behavior characteristics from data generated when a user account clicks on financial products over a historical period.

[0067] In practical applications, for user representation, the client-side user model learns the user representation from the user's historical clicks on financial products. The financial product representations [n1, n2, ..., nM] representing the user's historical clicks on financial products are used as input to the user model. The user representation u is then calculated through the user model, and u is the output of the user model. For financial product representation, the financial product representation n is learned from the financial product content using the financial product model.

[0068] The aforementioned user model can be implemented using various model structures. When using Efficient-FedRec (an efficient privacy-preserving framework for news recommendation) as the framework for the resource-object interaction prediction model, the NRMS (multi-head attention neural network recommendation algorithm) user model can be adopted, which combines multi-head self-attention networks and additional attention networks. Similarly, the aforementioned financial product model can be implemented using various model structures. A shallow NLP model (a natural language processing model) can be used, and when using Efficient-FedRec as the framework for the resource-object interaction prediction model, the PLM-NR model can be employed.

[0069] In practice, the central server obtains the dataset consisting of the resource object representations for the target resource object.

[0070] In practical applications, the central server obtains a dataset consisting of financial product representations for the target financial product. The target financial product is the financial product that the user of the target client has interacted with, and the target client is the client that participates in model training.

[0071] Step S204: Aggregate the training gradient information of each model to obtain aggregated gradient information.

[0072] Aggregated gradient information can refer to aggregated gradient information obtained by aggregating model training gradient information according to a secure aggregation protocol.

[0073] In practice, the central server aggregates the model training gradient information corresponding to the client user models of each client to obtain aggregated gradient information.

[0074] Step S206: Based on the aggregated gradient information, train the resource object model and each client user model; the resource object model is used to output the resource object representation corresponding to the input resource object.

[0075] In this context, the resource object model can refer to a model used to model the interaction characteristics of resource objects, that is, the resource object model can output a representation of resource objects. In practical applications, the resource object model can refer to a financial product model, which can be a model used to model the interaction characteristics of financial products, that is, the financial product model can output a representation of financial products.

[0076] In practice, the central server trains the resource object model and each client user model based on the aggregated gradient information. In each training process, the updated resource object model can output a new resource object representation set based on the resource object data.

[0077] In practical applications, during each training process, the central server updates the financial product model and the user model of each client based on the aggregated gradient information. The updated financial product model can output a new financial product representation set based on the financial product data. The central server can then distribute the updated user model and the new financial product representation set to each client to conduct a new round of model training.

[0078] Step S208: Based on the resource object model and each client user model, determine the resource object interaction prediction model; the resource object interaction prediction model is used to determine the prediction probability information of each user account initiating an interaction with the resource object to be predicted.

[0079] Among them, the resource object interaction prediction model can be a prediction model used to predict the probability of a user interacting with a resource object. In practical applications, the resource object interaction prediction model can be a prediction model used to predict the probability of a user clicking on a financial product.

[0080] Among them, the predicted probability information can refer to the probability of a user interacting with a resource object. In practical applications, the predicted probability information can refer to the probability of a user clicking on a financial product.

[0081] In practice, the central server determines the resource object interaction prediction model based on the resource object model and the user models of each client.

[0082] In practical applications, the central server determines the financial product recommendation model based on the financial product model and the user models of each client. The financial product model is relatively heavy and is deployed on the central server, while the user model is a lightweight model and is deployed on each client. The financial product recommendation model is a model composed of the financial product model on the central server and the user models of each client. It can train the lightweight user model on each client and train the relatively heavy financial product model on the central server, thereby reducing the communication and computing overhead of the client.

[0083] For example, instead of requesting the entire model from a central server, users might request only vector representations of their clicks, browsing, and purchases of financial products within their user behavior and user model. However, directly requesting the financial product representations within user behavior would leak users' private information. To address this, a randomly selected group of clients exchanges representations of a set of joint financial products involved in a set of user behavior through a secure aggregation protocol. Representing the client group as Us = {u1, u2, ..., us}, the joint financial product set is calculated as n = ∪ui ∈ UsNi. Therefore, the central server only knows the financial products accessed by one set of clients. Finally, users store a local copy of the user model Θtu and the financial product representations of the joint financial product set Θtes. Since the financial product representations of the joint financial product set are much smaller than the financial product models, the propagation cost is reduced.

[0084] The above describes the decomposition of the financial product recommendation model into a financial product model with parameter set Θn and a user model with parameter set Θu. A central server maintains the financial product model and generates financial product representations using parameter set Θe, while retaining a global user encoder. (The global user encoder refers to a system where users actively display their interests by selecting certain tags, and these tags are relatively fixed and do not change over a long period, reflecting the user's long-term, stable interests.) The global user encoder learns feature vectors from the answer behavior sequence and user tags to form the final user vector representation u. Its goal is to collaboratively train an accurate financial product recommendation model without disclosing user privacy (the user's behavior Bu is stored on the local device).

[0085] For the ease of understanding of those skilled in the art, Figure 3 An exemplary architecture diagram of a resource object interaction prediction model is provided. First, a central server distributes user models and resource object representation sets to various clients (client 1, ..., client n). The training gradient information of the models trained by each client is securely aggregated and then uploaded to the central server. The central server updates the user models on each client based on the aggregated gradient information corresponding to each user model. Simultaneously, the central server calculates the gradient information of the resource object models based on the aggregated gradient information of the resource object representations, thereby updating the resource object models in the central server based on the gradient information of the resource object models. The updated resource object models then output a new resource object representation set for redistribution to each client. Specifically, in the client, the user model outputs a user representation u based on the resource object representation set {n1, n2, ..., nM}. When interaction prediction is needed for candidate resource objects nc, the interaction prediction score can be determined based on the user representation u and the candidate resource object representation nc, thereby further determining the loss information to calculate the current model training gradient information.

[0086] The training method, apparatus, computer equipment, storage medium, and computer program product of the aforementioned resource object interaction prediction model, through obtaining a resource object representation set for the target resource object; the target resource object is the resource object that each user account has interacted with in a historical time period; the resource object representation set is used to characterize the resource object interaction features of user accounts; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model; the training gradient information of each model is aggregated to obtain aggregated gradient information; based on the aggregated gradient information, the resource object model and each client user model are trained; the resource object model is used to output the resource object representation corresponding to the input resource object; based on the resource object model and each client user model, the resource object interaction prediction model is determined; the resource object interaction prediction model is then determined. The test model is used to determine the prediction probability information of each user account initiating interaction with the resource object to be predicted. In this way, the resource object interaction prediction model can be divided into a resource object model on the central server and a user model on each client. The more intensive resource object model is trained on the central server, and the lighter client user model is trained on each client. This reduces the computational overhead of each client during model training. Furthermore, by aggregating the model training gradient information of each client, the model training gradient information is encrypted. At the same time, since each client uses the resource object representation set distributed by the central server to train the user model, compared with directly deploying the resource object model on each client, the communication overhead between the central server and each client during model training is reduced, thereby improving the training efficiency of the resource object interaction prediction model.

[0087] In another embodiment, obtaining a resource object representation set for a target resource object includes: obtaining resource object data for the target resource object; and inputting the resource object data of the target resource object into an initial resource object model to obtain a resource object representation set.

[0088] Resource object data can refer to the initial interaction data when a user interacts with a resource object. In practical applications, resource object data can refer to financial product data, which can be the initial interaction data when a user clicks on a financial product.

[0089] The initial resource object model can refer to an untrained resource object model. In practical applications, the initial resource object model can refer to an untrained financial product model.

[0090] In practice, the central server obtains the resource object data for the target resource object, and then inputs the resource object data of the target resource object into the initial resource object model to obtain the resource object representation set.

[0091] In practical applications, the central server inputs the financial product data from each client into the initial financial product model to obtain the initial financial product representation set. Subsequently, the central server distributes the initial financial product representation set and the initial user model to each client.

[0092] The technical solution of this embodiment obtains resource object data for the target resource object; inputs the resource object data of the target resource object into the initial resource object model to obtain a resource object representation set; in this way, a resource object representation set shared by all clients can be obtained, and user data on each client can be better protected to prevent information leakage.

[0093] In another embodiment, the method further includes: inputting a resource object representation set into each client user model to obtain a client user representation corresponding to each client; obtaining a candidate resource object representation corresponding to each client; and determining the prediction probability information of the candidate resource object on the client based on the client user representation and the candidate resource object representation; the prediction probability information is used to determine the model training gradient information.

[0094] Here, the client user representation can refer to the vector representation of the user interaction behavior characteristics of the client.

[0095] Among them, the candidate resource object representation can be a definite resource object representation that requires prediction probability information.

[0096] In the specific implementation, the central server inputs the resource object representation set into each client user model to obtain the client user representation corresponding to each client. The central server obtains the candidate resource object representation corresponding to each client. Based on the client user representation and the candidate resource object representation, the central server determines the prediction probability information of the candidate resource object on the client. Based on the prediction probability information, the central server determines the model training gradient information.

[0097] For example, the central server inputs the financial product representation set into the user model of each client to obtain the user representation corresponding to each client. The central server obtains the candidate financial product representation corresponding to each client. Based on the user representation and the candidate financial product object representation, the central server determines the click score of the candidate financial product on the client. Based on the click score, the central server determines the model training gradient information.

[0098] In practical applications, given a client u, the user representation u is calculated using the historical clicks of u on financial products through a local user encoder; for candidate financial products n... c n are represented by candidate financial products. c The user represents u, which calculates the click score s using a click predictor, where the click predictor can be a dot product.

[0099] When using classification cross-entropy loss for model training, for each clicked candidate financial product, K unclicked financial products are sampled in the same iteration, and the user and the degree of user preference for the financial products are abstracted into y. i and s i This allows for the prediction of similarity, thereby understanding customer preferences for candidate financial products. The loss calculation formula for the training samples is as follows:

[0100]

[0101] Where i represents the label of the i-th financial product, and user u is y i The predicted score is s i The average loss for all training samples is:

[0102]

[0103] The local gradients of the global user encoder and the local gradients of the financial product representation are as follows:

[0104]

[0105]

[0106] The technical solution of this embodiment obtains the client user representation corresponding to each client by inputting the resource object representation set into each client user model; obtains the candidate resource object representation corresponding to each client; and determines the prediction probability information of the candidate resource object on the client based on the client user representation and the candidate resource object representation. In this way, the prediction probability information of the candidate resource object on the client can be accurately determined, which is beneficial to determining the model training gradient information of the client user model.

[0107] In another embodiment, training the resource object model and each client user model based on the aggregated gradient information includes: updating the initial resource object model using backpropagation training based on the aggregated gradient information to obtain the updated resource object model; outputting the updated resource object representation set through the updated resource object model; sending the updated resource object representation set to each client to perform a new round of model training on the updated client user model on each client to obtain the aggregated gradient information corresponding to the new round of model training; the updated client user model is obtained by updating the client user model using the aggregated gradient information; returning to the step of updating the initial resource object model using backpropagation training based on the aggregated gradient information to obtain the updated resource object model, until the resource object model and each client user model that meet the training termination condition are obtained.

[0108] The initial resource object model can be the initial model of the resource object model.

[0109] The updated resource object model can be the model obtained by updating the resource object model.

[0110] The updated resource object representation set can be the resource object representation set obtained by inputting resource object data into the updated resource object model.

[0111] The updated client user model can be obtained by updating the client user model.

[0112] In the specific implementation, the central server updates the initial resource object model using backpropagation training based on the aggregated gradient information, obtaining the updated resource object model. The central server then outputs the updated resource object representation set, which is sent to each client so that each client can use the updated resource object representation set to train the updated client user model for a new round of model training, obtaining the corresponding aggregated gradient information after the new round of model training. The central server then trains the model again based on the aggregated gradient information after the new round of model training, until the model converges. The resource object model and each client user model at the time of model convergence are then used as the final model.

[0113] In practical applications, the global user model is directly updated using the gradient of the user model through the FedAdam algorithm (a stochastic optimization algorithm), as follows:

[0114]

[0115]

[0116]

[0117] Where η is the learning rate, and β1, β2, and τ are parameters of FedAdam. The financial product model is updated through a backpropagation training process. For each financial product ni in the joint financial product set, the central server has its content and the gradient of its financial product representation. Using the financial product content as input, its financial product representation ni is calculated through the financial product model.

[0118] Financial product model g t n The gradient is:

[0119]

[0120] Where Θtn represents the parameters of the financial product model in round t. The calculation results when updating the new model using the Adam optimizer are as follows:

[0121]

[0122]

[0123]

[0124] Where η is the learning rate, and β1, β2, and τ are Adam's hyperparameters. The updated financial product model is used to infer the financial product representation. Finally, the updated financial product representation and the user encoder are distributed to all clients.

[0125] The technical solution of this embodiment updates the initial resource object model using backpropagation training based on aggregated gradient information to obtain an updated resource object model; outputs an updated resource object representation set from the updated resource object model; sends the updated resource object representation set to each client to perform a new round of model training on the updated client user model on each client, obtaining the corresponding aggregated gradient information after the new round of model training; the updated client user model is obtained by updating the client user model using aggregated gradient information; returns to the step of updating the initial resource object model using backpropagation training based on aggregated gradient information to obtain an updated resource object model, until a resource object model and each client user model that meet the training termination condition are obtained; thus, a final model that meets the model convergence condition can be obtained, enabling the training of the resource object model and each client user model.

[0126] In another embodiment, aggregating the model training gradient information of each client to obtain aggregated gradient information includes: aggregating the model training gradient information of each client to obtain at least one local aggregated gradient information; the local aggregated gradient information is obtained by aggregating a preset portion of the model training gradient information according to a preset aggregation method; and aggregating the local aggregated gradient information globally to obtain aggregated gradient information.

[0127] In the specific implementation, the central server aggregates the model training gradient information of a preset part according to the preset aggregation method to obtain at least one local aggregated gradient information. The central server then performs global aggregation of the local aggregated gradient information to obtain the aggregated gradient information.

[0128] In practical applications, the central server needs to calculate the weighted sum of the gradients of the user model, the financial product representation, and the number of samples from the randomly selected user group Us. Since local gradients contain private information, secure aggregation is used for the summation. The aggregated gradients of the user model and the financial product representation are expressed as follows:

[0129]

[0130]

[0131] Since each user only sends the gradient of the financial product representation in the joint financial product set n, which is much smaller than the financial product model, the communication overhead is reduced.

[0132] In gradient aggregation, a two-stage federated learning architecture based on SMPC (Secure Multi-party Computation) is adopted, which focuses on protecting the local parameters w generated by the participants. i The participants will w iDecompose into n meaningless values: the first n-1 values ​​are random numbers, and the nth value is determined by the formula:

[0133]

[0134] The calculation shows that Q is the modulus, which needs to be pre-selected, and V is the correct parameter value to be generated. By secretly selecting n-1 values ​​that are not equal to Q, the V value that needs to be protected is calculated.

[0135] In this method, participants secretly exchange shares with each other, with each participant holding a portion of the parameter vector. Participants then perform local aggregation of these secret shares, followed by global aggregation to obtain w*. This two-stage exchange and aggregation of secret shares eliminates the randomness of share splitting. The parameter aggregation mechanism of this method is as follows: Figure 4 As shown, the two-stage federated learning architecture described above uses a voting process to generate a participating committee. Committee members exchange and aggregate secret shares to generate aggregated parameters, which can solve the problem of excessive communication costs caused by mutual exchange of secret shares.

[0136] The technical solution of this embodiment aggregates the model training gradient information of each client to obtain at least one local aggregated gradient information. The local aggregated gradient information is obtained by aggregating a preset part of the model training gradient information according to a preset aggregation method. The local aggregated gradient information is then globally aggregated to obtain aggregated gradient information. In this way, the communication overhead of the aggregation process can be reduced through two-stage information aggregation.

[0137] In another embodiment, the method further includes: sending an interactive prediction instruction to the target client; the interactive prediction instruction is used to instruct the target client to input the resource object to be predicted into the resource object interactive prediction model to obtain the prediction probability information corresponding to the resource object to be predicted.

[0138] The resource object to be predicted can refer to the resource object for which probability information needs to be predicted. In practical applications, the resource object to be predicted can refer to financial products for which the probability information of user clicks needs to be predicted.

[0139] In practice, the central server sends an interactive prediction instruction to the target client, instructing the target client to input the resource object to be predicted into the resource object interactive prediction model to obtain the prediction probability information corresponding to the resource object to be predicted.

[0140] The technical solution of this embodiment, by sending an interactive prediction instruction to the target client, can instruct the client to input the resource object to be predicted into the trained resource object interactive prediction model, thereby accurately outputting the prediction probability information of the resource object to be predicted for the target client.

[0141] In another embodiment, the target client is used to determine the prediction probability information of the user account in the target client initiating an interaction with the resource object to be predicted, based on the resource object representation corresponding to the resource object to be predicted and the user behavior representation of the target client.

[0142] In the specific implementation, the target client determines the prediction probability information of the local user initiating an interaction with the resource object to be predicted based on the resource object representation corresponding to the resource object to be predicted and the local user behavior representation.

[0143] In this embodiment, the target client determines the prediction probability information of the user account in the target client initiating interaction with the resource object to be predicted based on the resource object representation corresponding to the resource object to be predicted and the user behavior representation of the target client. This can accurately determine the prediction probability information of the user account in the target client initiating interaction with the resource object to be predicted.

[0144] In another embodiment, such as Figure 5 As shown, a training method for a resource object interaction prediction model is provided, which is then applied to... Figure 1 Taking the central server 102 as an example, the following steps are included:

[0145] Step S502: Obtain the resource object representation set for the target resource object; the target resource object is the resource object that each user account has interacted with in the historical time period; the resource object representation set is used to characterize the resource object interaction features of the user account; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model.

[0146] Step S504: Aggregate the training gradient information of each model to obtain aggregated gradient information.

[0147] Step S506: Based on the aggregated gradient information, train the resource object model and each client user model; the resource object model is used to output the resource object representation corresponding to the input resource object.

[0148] Step S508: Based on the resource object model and each client user model, determine the resource object interaction prediction model; the resource object interaction prediction model is used to determine the prediction probability information of each user account initiating an interaction with the resource object to be predicted.

[0149] Step S510: Send an interactive prediction instruction to the target client; the interactive prediction instruction is used to instruct the target client to input the resource object to be predicted into the resource object interactive prediction model to obtain the prediction probability information corresponding to the resource object to be predicted.

[0150] It should be noted that the specific limitations of the above steps can be found in the specific limitations of the training method for a resource object interaction prediction model described above.

[0151] In one embodiment, such as Figure 6 As shown, a resource object interaction prediction method is provided, which is applied to... Figure 1 Taking client 104 as an example, the explanation includes the following steps:

[0152] Step S602: Obtain the resource object to be predicted.

[0153] The resource object to be predicted can refer to the resource object in the client that needs to have its probability information predicted.

[0154] In the actual implementation, the client obtains the resource object to be predicted.

[0155] Step S604: Input the resource object to be predicted into the pre-trained resource object interaction prediction model to obtain the prediction probability information corresponding to the resource object to be predicted; the prediction probability information is used to determine the probability that a user account in the target client will initiate an interaction with the resource object to be predicted; wherein, the pre-trained resource object interaction prediction model is determined based on the resource object model and each client user model; the resource object model and each client user model are obtained based on aggregated gradient information; the aggregated gradient information is obtained by aggregating the training gradient information of each model; the training gradient information of each model is obtained by training the client user model deployed on the client where each user account is located using the resource object representation set; the resource object representation set is used to characterize the resource object interaction features corresponding to the interaction between the user account and the target resource object in historical time periods; the target resource object is the resource object that each user account has interacted with in historical time periods.

[0156] In practice, the client inputs the resource object to be predicted into a pre-trained resource object interaction prediction model to obtain the prediction probability information corresponding to the resource object to be predicted, which is the prediction probability information of the client initiating an interaction with the resource object to be predicted.

[0157] The above-mentioned resource object interaction prediction method obtains the resource object to be predicted, inputs the resource object to be predicted into a pre-trained resource object interaction prediction model, and obtains the prediction probability information corresponding to the resource object to be predicted. In this way, the probability information of the client initiating an interaction with the resource object to be predicted can be accurately determined.

[0158] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0159] Based on the same inventive concept, this application also provides a training apparatus for a resource object interaction prediction model, which is used to implement the training method for the resource object interaction prediction model described above. The solution provided by this apparatus is similar to the implementation described in the above method. Therefore, the specific limitations of one or more resource object interaction prediction model training apparatus embodiments provided below can be found in the limitations of the resource object interaction prediction model training method described above, and will not be repeated here.

[0160] In one embodiment, such as Figure 7 As shown, a training device for a resource object interaction prediction model is provided, comprising:

[0161] The acquisition module 702 is used to acquire the resource object representation set for the target resource object; the target resource object is the resource object that each user account has interacted with in the historical time period; the resource object representation set is used to characterize the resource object interaction features of the user account; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model;

[0162] The aggregation module 704 is used to aggregate the training gradient information of each model to obtain aggregated gradient information.

[0163] Training module 706 is used to train the resource object model and each client user model based on the aggregated gradient information; the resource object model is used to output the resource object representation corresponding to the input resource object;

[0164] The determination module 708 is used to determine the resource object interaction prediction model based on the resource object model and each client user model; the resource object interaction prediction model is used to determine the prediction probability information of each user account initiating interaction with the resource object to be predicted.

[0165] In one embodiment, the acquisition module 702 is specifically used to acquire resource object data for the target resource object; input the resource object data of the target resource object into the initial resource object model to obtain a resource object representation set.

[0166] In one embodiment, the apparatus further includes: an input module, specifically configured to input a resource object representation set into each client user model to obtain a client user representation corresponding to each client; obtain a candidate resource object representation corresponding to each client; and determine the prediction probability information of the candidate resource object on the client based on the client user representation and the candidate resource object representation; the prediction probability information is used to determine the model training gradient information.

[0167] In one embodiment, the training module 706 is specifically used to update the initial resource object model using backpropagation training based on the aggregated gradient information to obtain the updated resource object model; output the updated resource object representation set through the updated resource object model; send the updated resource object representation set to each client to perform a new round of model training on the updated client user model on each client to obtain the aggregated gradient information corresponding to the new round of model training; the updated client user model is obtained by updating the client user model using the aggregated gradient information; return to the step of updating the initial resource object model using backpropagation training based on the aggregated gradient information to obtain the updated resource object model, until the resource object model and each client user model that meet the training termination condition are obtained.

[0168] In one embodiment, the aggregation module 704 is specifically used to aggregate the model training gradient information of each client to obtain at least one local aggregated gradient information; the local aggregated gradient information is obtained by aggregating a preset part of the model training gradient information according to a preset aggregation method; and the local aggregated gradient information is globally aggregated to obtain aggregated gradient information.

[0169] In one embodiment, the apparatus further includes: a sending module, specifically used to send an interactive prediction instruction to a target client; the interactive prediction instruction instructs the target client to input the resource object to be predicted into the resource object interactive prediction model to obtain the prediction probability information corresponding to the resource object to be predicted.

[0170] In one embodiment, the prediction module, the target client is used to determine the prediction probability information of the user account in the target client initiating an interaction with the resource object to be predicted, based on the resource object representation corresponding to the resource object to be predicted and the user behavior representation of the target client.

[0171] In one embodiment, such as Figure 8As shown, a resource object information processing device is provided, comprising:

[0172] The acquisition module 802 is used to acquire the resource object to be predicted;

[0173] The input module 804 is used to input the resource object to be predicted into the pre-trained resource object interaction prediction model to obtain the prediction probability information corresponding to the resource object to be predicted; the prediction probability information is used to determine the probability that the user account in the target client will initiate an interaction with the resource object to be predicted.

[0174] Among them, the pre-trained resource object interaction prediction model is determined based on the resource object model and each client user model; the resource object model and each client user model are obtained based on aggregated gradient information; the aggregated gradient information is obtained by aggregating the training gradient information of each model; the training gradient information of each model is obtained by training the client user model deployed on the client where each user account is located using the resource object representation set; the resource object representation set is used to characterize the resource object interaction features corresponding to the interaction between the user account and the target resource object in a historical time period; the target resource object is the resource object that each user account has interacted with in a historical time period.

[0175] The modules in the training device and resource object information processing device of the aforementioned resource object interaction prediction model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0176] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores training data for a resource object interaction prediction model and resource object information processing data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a method for training a resource object interaction prediction model and processing resource object information.

[0177] As will be understood by those skilled in the art, Figure 9The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0178] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the training method and resource object information processing method for the resource object interaction prediction model described above. Here, the steps of the training method and resource object information processing method for the resource object interaction prediction model can be any of the steps in the training method and resource object information processing method for the resource object interaction prediction model described in the various embodiments above.

[0179] In one embodiment, a computer-readable storage medium is provided, storing a computer program that, when executed by a processor, causes the processor to perform the steps of the above-described method for training a resource object interaction prediction model and method for processing resource object information. Here, the steps of the method for training a resource object interaction prediction model and method for processing resource object information can be steps from the methods for training a resource object interaction prediction model and methods for processing resource object information in the various embodiments described above.

[0180] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, causes the processor to perform the steps of the above-described method for training a resource object interaction prediction model and method for processing resource object information. Here, the steps of the method for training a resource object interaction prediction model and method for processing resource object information can be steps from the methods for training a resource object interaction prediction model and methods for processing resource object information in the various embodiments described above.

[0181] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0182] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0183] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0184] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A training method for a resource object interaction prediction model, characterized in that, Applied to a central server, the method includes: Obtain resource object data for the target resource object; input the resource object data of the target resource object into the initial resource object model to obtain the resource object representation set; the target resource object is the resource object that each user account has interacted with in a historical time period; the resource object representation set is used to characterize the resource object interaction features of the user account; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model; Aggregate the training gradient information of each model to obtain aggregated gradient information; Based on the aggregated gradient information, the resource object model and each of the client user models are trained; the resource object model is used to output the resource object representation corresponding to the input resource object; Based on the resource object model and each of the client user models, a resource object interaction prediction model is determined; the resource object interaction prediction model is used to determine the prediction probability information of each of the user accounts initiating an interaction with the resource object to be predicted; wherein, the resource object interaction prediction model is composed of the resource object model deployed on the central server and the client user model deployed on each of the clients; The step of training the resource object model and each of the client user models based on the aggregated gradient information includes: Based on the aggregated gradient information, the initial resource object model is updated using backpropagation training to obtain the updated resource object model; and the client user model deployed on each of the clients is synchronously updated based on the aggregated gradient information. The updated resource object representation set is output through the updated resource object model; The updated resource object representation set is sent again to each of the clients to perform a new round of model training on the updated client user model on each of the clients, and to obtain the aggregated gradient information corresponding to the new round of model training; the updated client user model is obtained by updating the client user model using the aggregated gradient information; The process returns to the step of updating the initial resource object model using backpropagation training based on the aggregated gradient information to obtain the updated resource object model, until the resource object model and each client user model that meet the training termination condition are obtained.

2. The method according to claim 1, characterized in that, The method further includes: The resource object representation set is input into each of the client user models to obtain the client user representation corresponding to each client. Obtain the candidate resource object representation corresponding to each client, and determine the prediction probability information of the candidate resource object on the client based on the client user representation and the candidate resource object representation; the prediction probability information is used to determine the model training gradient information.

3. The method according to claim 1, characterized in that, The aggregation of the training gradient information of each client model to obtain aggregated gradient information includes: The model training gradient information of each client is aggregated to obtain at least one local aggregated gradient information; the local aggregated gradient information is obtained by aggregating a preset portion of the model training gradient information according to a preset aggregation method. The aggregated gradient information is obtained by globally aggregating the local aggregated gradient information.

4. The method according to claim 1, characterized in that, The method further includes: Sending an interactive prediction instruction to the target client; the interactive prediction instruction is used to instruct the target client to input the resource object to be predicted into the resource object interactive prediction model to obtain the prediction probability information corresponding to the resource object to be predicted.

5. The method according to claim 4, characterized in that, The target client is used to determine the prediction probability information of a user account in the target client initiating an interaction with the resource object to be predicted, based on the resource object representation corresponding to the resource object to be predicted and the user behavior representation of the target client.

6. A method for processing resource object information, characterized in that, The method includes: Obtain the resource object to be predicted; The resource object to be predicted is input into a pre-trained resource object interaction prediction model to obtain the prediction probability information corresponding to the resource object to be predicted; the prediction probability information is used to determine the probability that a user account in the target client will initiate an interaction with the resource object to be predicted; the resource object interaction prediction model is trained by the training method of the resource object interaction prediction model according to any one of claims 1-5. The pre-trained resource object interaction prediction model is determined by the central server based on the resource object model and each client user model; the resource object model and each client user model are obtained based on aggregated gradient information; the aggregated gradient information is obtained by aggregating the training gradient information of each model; the training gradient information of each model is obtained by training the client user model deployed on the client of each user account using a resource object representation set; the resource object representation set is used to characterize the resource object interaction features corresponding to the interaction between the user account and the target resource object in a historical time period; the target resource object is the resource object that each user account has interacted with in the historical time period.

7. A training device for a resource object interaction prediction model, located in a central server, characterized in that, The device includes: The acquisition module is used to acquire resource object data for a target resource object; input the resource object data of the target resource object into an initial resource object model to obtain the resource object representation set; the target resource object is the resource object that each user account has interacted with in a historical time period; the resource object representation set is used to characterize the resource object interaction features of the user account; the resource object representation set is used to train the client user model deployed on the client of each user account to obtain the model training gradient information corresponding to each client user model; The aggregation module is used to aggregate the training gradient information of each model to obtain aggregated gradient information; The training module is used to train the resource object model and each of the client user models based on the aggregated gradient information; the resource object model is used to output the resource object representation corresponding to the input resource object; The determination module is used to determine a resource object interaction prediction model based on the resource object model and each of the client user models; the resource object interaction prediction model is used to determine the prediction probability information of each of the user accounts initiating an interaction with the resource object to be predicted; wherein, the resource object interaction prediction model is composed of the resource object model deployed on the central server and the client user model deployed on each of the clients; The training module is specifically used to update the initial resource object model using backpropagation training based on the aggregated gradient information to obtain an updated resource object model; and to synchronously update the client user models deployed on each of the clients based on the aggregated gradient information; output an updated resource object representation set through the updated resource object model; and send the updated resource object representation set back to each of the clients to perform a new round of model training on the updated client user models on each of the clients to obtain the aggregated gradient information corresponding to the new round of model training; the updated client user model is obtained by updating the client user model using the aggregated gradient information; and return to the step of updating the initial resource object model using backpropagation training based on the aggregated gradient information to obtain an updated resource object model, until the resource object model and each of the client user models that meet the training termination conditions are obtained.

8. A resource object information processing device, characterized in that, The device includes: The acquisition module is used to acquire the resource objects to be predicted. An input module is used to input the resource object to be predicted into a pre-trained resource object interaction prediction model to obtain prediction probability information corresponding to the resource object to be predicted; the prediction probability information is used to determine the probability that a user account in the target client will initiate an interaction with the resource object to be predicted; the resource object interaction prediction model is trained by the training method of the resource object interaction prediction model according to any one of claims 1-5. The pre-trained resource object interaction prediction model is determined by the central server based on the resource object model and each client user model; the resource object model and each client user model are obtained based on aggregated gradient information; the aggregated gradient information is obtained by aggregating the training gradient information of each model; the training gradient information of each model is obtained by training the client user model deployed on the client of each user account using a resource object representation set; the resource object representation set is used to characterize the resource object interaction features corresponding to the interaction between the user account and the target resource object in a historical time period; the target resource object is the resource object that each user account has interacted with in the historical time period.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images