Push model training, message pushing method and device, storage medium and equipment

Abstract

The application relates to a push model training method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: inputting a target sample and a reference sample into an initial push difference prediction model for interaction possibility difference prediction, obtaining an interaction difference degree corresponding to the target sample and an interaction difference degree corresponding to the reference sample, and calculating the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample to obtain a training interaction difference. Loss calculation is performed based on the training interaction difference and a sample interaction difference label to obtain training loss information. The initial push difference prediction model is updated based on the training loss information and is iterated until a training completion condition is reached, a target push difference prediction model is obtained, and the target push difference prediction model is used for predicting a target interaction difference degree and performing message pushing. The method can improve the message pushing accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a push model training, message push method, apparatus, computer device, storage medium, and computer program product. Background Technology

[0002] With the development of computer network technology, push notification technology has emerged. Push notifications refer to users pre-setting their desired message subscriptions, which the publisher then pushes to subscribers. Currently, publishers push multiple messages to subscribers in chronological or interest-based order. However, when pushing messages in interest-based order, the publisher needs to consume server resources to calculate the subscriber's interest level for each message. When pushing messages to a large number of subscribers in interest-based order, there is a risk that subscribers may not interact with the messages, leading to wasted server resources. Summary of the Invention

[0003] Therefore, it is necessary to provide a push model training, push method, device, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of message push and thus save resources, in order to address the above-mentioned technical problems.

[0004] Firstly, this application provides a method for training a push notification model. The method includes:

[0005] Obtain training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when pushing historical messages in the target order, and the reference samples are obtained when pushing historical messages in the reference order.

[0006] The target sample is input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the target sample. The reference sample is also input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0007] The difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample is calculated to obtain the training interaction difference. Based on the training interaction difference and the sample, the loss is calculated on the interaction difference label to obtain the training loss information.

[0008] The initial push difference prediction model is updated based on the training loss information, and the steps of obtaining training sample pairs and corresponding sample pair interaction difference labels are iteratively executed until the training completion condition is met, and the target push difference prediction model is obtained. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference.

[0009] Secondly, this application also provides a push model training device. The device includes:

[0010] The sample acquisition module is used to acquire training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when historical messages are pushed in the target order, and the reference samples are obtained when historical messages are pushed in the reference order.

[0011] The prediction module is used to input the target sample into the initial push difference prediction model to predict the interaction probability difference and obtain the interaction difference degree corresponding to the target sample. It also inputs the reference sample into the initial push difference prediction model to predict the interaction probability difference and obtain the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0012] The loss calculation module is used to calculate the difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample, to obtain the training interaction difference, and to calculate the loss based on the training interaction difference and the interaction difference label of the sample to obtain the training loss information.

[0013] The iteration module is used to update the initial push difference prediction model based on the training loss information and return the steps of obtaining training sample pairs and corresponding sample pair interaction difference labels to be executed iteratively until the training completion condition is met, and the target push difference prediction model is obtained. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference.

[0014] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0015] Obtain training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when pushing historical messages in the target order, and the reference samples are obtained when pushing historical messages in the reference order.

[0016] The target sample is input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the target sample. The reference sample is also input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0017] The difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample is calculated to obtain the training interaction difference. Based on the training interaction difference and the sample, the loss is calculated on the interaction difference label to obtain the training loss information.

[0018] The initial push difference prediction model is updated based on the training loss information, and the steps of obtaining training sample pairs and corresponding sample pair interaction difference labels are iteratively executed until the training completion condition is met, and the target push difference prediction model is obtained. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference.

[0019] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0020] Obtain training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when pushing historical messages in the target order, and the reference samples are obtained when pushing historical messages in the reference order.

[0021] The target sample is input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the target sample. The reference sample is also input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0022] The difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample is calculated to obtain the training interaction difference. Based on the training interaction difference and the sample, the loss is calculated on the interaction difference label to obtain the training loss information.

[0023] The initial push difference prediction model is updated based on the training loss information, and the steps of obtaining training sample pairs and corresponding sample pair interaction difference labels are iteratively executed until the training completion condition is met, and the target push difference prediction model is obtained. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference.

[0024] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0025] Obtain training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when pushing historical messages in the target order, and the reference samples are obtained when pushing historical messages in the reference order.

[0026] The target sample is input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the target sample. The reference sample is also input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0027] The difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample is calculated to obtain the training interaction difference. Based on the training interaction difference and the sample, the loss is calculated on the interaction difference label to obtain the training loss information.

[0028] The initial push difference prediction model is updated based on the training loss information, and the steps of obtaining training sample pairs and corresponding sample pair interaction difference labels are iteratively executed until the training completion condition is met, and the target push difference prediction model is obtained. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference.

[0029] The aforementioned push model training method, apparatus, computer equipment, storage medium, and computer program product train an initial push difference prediction model using target samples and reference samples to obtain training interaction differences. Then, loss is calculated using the training interaction differences and sample-based interaction difference labels to obtain training loss information. This training loss information is used to update the initial push difference prediction model, and this process is iterated continuously until the training completion condition is met, resulting in a target push difference prediction model. By calculating the loss using training interaction differences and sample-based interaction difference labels and updating the initial push difference prediction model, the accuracy of the obtained target push difference prediction model is improved. The target push difference prediction model is then used to predict the degree of target interaction difference when pushing messages according to the target order versus the reference order, thereby improving the accuracy of the target interaction difference degree. Pushing messages according to the target interaction difference degree further improves message push accuracy and avoids the problem of wasted server resources.

[0030] Firstly, this application provides a message push method. The method includes:

[0031] Get the message push request, which carries the object to be pushed to;

[0032] The target interaction difference degree is obtained based on the target object to be pushed. This degree is achieved by inputting the feature data of the target object into a target push difference prediction model to predict interaction probability differences. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels. The distributions of target and reference samples in the training sample pairs are input into an initial push difference prediction model to predict interaction probability differences, resulting in the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample. The difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample is calculated to obtain the training interaction difference. Loss is calculated based on the training interaction difference and the sample pair interaction difference labels to obtain training loss information. The initial push difference prediction model is then trained based on this training loss information. The interaction difference degree characterizes the difference between the target interaction probability and the reference interaction probability. The target interaction probability characterizes the interaction probability obtained when messages are pushed according to the target order; the reference interaction probability characterizes the interaction probability obtained when messages are pushed according to the reference order.

[0033] When the degree of difference in target interaction meets the preset target order push conditions, respond to the message push request, obtain each message to be pushed corresponding to the object to be pushed and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the object to be pushed according to the target push order.

[0034] Secondly, this application also provides a message push device. The device includes:

[0035] The request retrieval module is used to retrieve message push requests, which carry the object to be pushed.

[0036] The difference level acquisition module is used to obtain the target interaction difference level based on the object to be pushed. The target interaction difference level is obtained by inputting the feature data of the object to be pushed into the target push difference prediction model to predict the interaction probability difference. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels, inputs the target sample and reference sample distribution in the training sample pairs into the initial push difference prediction model to predict the interaction probability difference, obtains the interaction difference level corresponding to the target sample and the interaction difference level corresponding to the reference sample, and calculates the difference between the interaction difference level corresponding to the target sample and the interaction difference level corresponding to the reference sample to obtain the training interaction difference. Based on the training interaction difference and the sample pair interaction difference labels, loss is calculated to obtain training loss information. The initial push difference prediction model is trained based on the training loss information. The interaction difference level is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0037] The request and response module is used to respond to message push requests when the degree of difference in target interaction meets the preset target order push conditions, obtain each message to be pushed corresponding to the target object and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the target object according to the target push order.

[0038] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0039] Get the message push request, which carries the object to be pushed to;

[0040] The target interaction difference degree is obtained based on the target object to be pushed. This degree is achieved by inputting the feature data of the target object into a target push difference prediction model to predict interaction probability differences. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels. The distributions of target and reference samples in the training sample pairs are input into an initial push difference prediction model to predict interaction probability differences, resulting in the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample. The difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample is calculated to obtain the training interaction difference. Loss is calculated based on the training interaction difference and the sample pair interaction difference labels to obtain training loss information. The initial push difference prediction model is then trained based on this training loss information. The interaction difference degree characterizes the difference between the target interaction probability and the reference interaction probability. The target interaction probability characterizes the interaction probability obtained when messages are pushed according to the target order; the reference interaction probability characterizes the interaction probability obtained when messages are pushed according to the reference order.

[0041] When the degree of difference in target interaction meets the preset target order push conditions, respond to the message push request, obtain each message to be pushed corresponding to the object to be pushed and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the object to be pushed according to the target push order.

[0042] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0043] Get the message push request, which carries the object to be pushed to;

[0044] The target interaction difference degree is obtained based on the target object to be pushed. This degree is achieved by inputting the feature data of the target object into a target push difference prediction model to predict interaction probability differences. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels. The distributions of target and reference samples in the training sample pairs are input into an initial push difference prediction model to predict interaction probability differences, resulting in the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample. The difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample is calculated to obtain the training interaction difference. Loss is calculated based on the training interaction difference and the sample pair interaction difference labels to obtain training loss information. The initial push difference prediction model is then trained based on this training loss information. The interaction difference degree characterizes the difference between the target interaction probability and the reference interaction probability. The target interaction probability characterizes the interaction probability obtained when messages are pushed according to the target order; the reference interaction probability characterizes the interaction probability obtained when messages are pushed according to the reference order.

[0045] When the degree of difference in target interaction meets the preset target order push conditions, respond to the message push request, obtain each message to be pushed corresponding to the object to be pushed and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the object to be pushed according to the target push order.

[0046] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0047] Get the message push request, which carries the object to be pushed to;

[0048] The target interaction difference degree is obtained based on the target object to be pushed. This degree is achieved by inputting the feature data of the target object into a target push difference prediction model to predict interaction probability differences. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels. The distributions of target and reference samples in the training sample pairs are input into an initial push difference prediction model to predict interaction probability differences, resulting in the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample. The difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample is calculated to obtain the training interaction difference. Loss is calculated based on the training interaction difference and the sample pair interaction difference labels to obtain training loss information. The initial push difference prediction model is then trained based on this training loss information. The interaction difference degree characterizes the difference between the target interaction probability and the reference interaction probability. The target interaction probability characterizes the interaction probability obtained when messages are pushed according to the target order; the reference interaction probability characterizes the interaction probability obtained when messages are pushed according to the reference order.

[0049] When the degree of difference in target interaction meets the preset target order push conditions, respond to the message push request, obtain each message to be pushed corresponding to the object to be pushed and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the object to be pushed according to the target push order.

[0050] The aforementioned message push method, apparatus, computer device, storage medium, and computer program product acquire a message push request carrying a target object, and obtains the corresponding target interaction difference degree based on the target object. When the target interaction difference degree meets the preset target order push conditions, the system responds to the message push request, acquires each message to be pushed corresponding to the target object and the target push order corresponding to each message, and pushes each message to be pushed to the terminal corresponding to the target object in the target push order. Since the target interaction difference degree is obtained using a target push difference prediction model, the accuracy of the target interaction difference degree is improved. Furthermore, only when the target interaction difference degree meets the preset target order push conditions are the messages to be pushed to the terminal corresponding to the target object pushed in the target push order, thus avoiding server resource waste and saving server resources. Attached Figure Description

[0051] Figure 1 This is a diagram illustrating the application environment of the push model training method in one embodiment;

[0052] Figure 2 This is a flowchart illustrating a push model training method in one embodiment;

[0053] Figure 3 This is a schematic diagram of the process of pushing the difference prediction model prediction in one embodiment;

[0054] Figure 4 This is a schematic diagram of the architecture of a push difference prediction model in a specific embodiment;

[0055] Figure 5 This is a schematic diagram of the process for obtaining training sample pairs in one embodiment;

[0056] Figure 6 This is a schematic diagram of the process for obtaining a sample set in one embodiment;

[0057] Figure 7 This is a schematic diagram of the causal relationship in a specific embodiment;

[0058] Figure 8 This is a flowchart illustrating a message push method in one embodiment;

[0059] Figure 9 This is a flowchart illustrating the message push method in another embodiment;

[0060] Figure 10 This is a comparative diagram of server resources in a specific embodiment;

[0061] Figure 11 This is a flowchart illustrating a specific embodiment of a push model training method;

[0062] Figure 12 This is a structural block diagram of a push-model training device in one embodiment;

[0063] Figure 13 This is a structural block diagram of a message push device in one embodiment;

[0064] Figure 14 This is an internal structural diagram of a computer device in one embodiment;

[0065] Figure 15 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0066] 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.

[0067] The push model training method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other server. Server 104 receives a model training instruction from terminal 102. Server 104 obtains training sample pairs and corresponding interaction difference labels based on the instruction. The training sample pairs include target samples and reference samples. Target samples are obtained when historical messages are pushed according to the target order, and reference samples are obtained when historical messages are pushed according to the reference order. Server 104 inputs the target samples into an initial push difference prediction model to predict the interaction probability difference, obtaining the interaction difference degree corresponding to the target samples. Similarly, server 104 inputs the reference samples into the initial push difference prediction model to predict the interaction probability difference, obtaining the interaction difference degree corresponding to the reference samples. The interaction difference degree characterizes the difference between the target interaction probability and the reference interaction probability. The target interaction probability characterizes the message push according to the target order. The server 104 calculates the difference between the interaction probability obtained when pushing messages in the target order and the interaction probability obtained when pushing messages in the reference order. The server 104 calculates the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample, obtaining the training interaction difference. Based on the training interaction difference and the interaction difference labels of the samples, the server 104 calculates the loss, obtaining the training loss information. The server 104 updates the initial push difference prediction model based on the training loss information and iteratively executes the steps of obtaining training sample pairs and corresponding sample pair interaction difference labels until the training completion condition is met, thus obtaining the target push difference prediction model. The target push difference prediction model is used to predict the target interaction difference degree when pushing messages in the target order relative to pushing messages in the reference order, and pushes messages based on the target interaction difference degree. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0068] In one embodiment, such as Figure 2 As shown, a push model training method is provided, which is then applied to... Figure 1 Taking a server as an example, it can be understood that this method can also be applied to servers, and also to systems including terminals and servers, and implemented through the interaction between the terminal and the server. This embodiment includes the following steps:

[0069] Step 202: Obtain training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when historical messages are pushed in the target order, and the reference samples are obtained when historical messages are pushed in the reference order.

[0070] The training sample pair includes two training samples, which are used by the model during training. These samples can consist of feature data, extracted from data generated after the push recipient interacts with historical push messages. The push recipient is an object that can interact with the received historical push messages. Historical messages refer to messages pushed at historical moments, which can be text messages, image messages, voice messages, video messages, etc. The training sample can be a target sample or a reference sample. A target sample is the training sample obtained when pushing historical messages in a target order, while a reference sample is the training sample obtained when pushing historical messages in a reference order. The target order refers to the message push order that requires server resources to determine. For example, the target order could be an interest order, a relevance order, etc. An interest order is determined by the subscribers' level of interest in the messages, requiring server resources to calculate this interest. A relevance order is determined by the degree of relevance between the subscribers and the messages, requiring server resources to calculate this relevance. The reference order refers to the message push order that can be determined directly without consuming server resources. For example, the reference order can be chronological order or subscription order. Chronological order refers to the order determined by the message push time, which can be pre-set and does not require server resource calculation. Subscription order refers to the order determined by the message subscription time, which is determined when the message is subscribed and also does not require server resource calculation. The sample pair interaction difference label is used to characterize the difference between the target sample and the reference sample in their interactions with the historical push messages. This sample pair interaction difference label can be pre-set.

[0071] Specifically, the server can retrieve training sample pairs and corresponding interaction difference labels from the database. Training sample pairs include target samples and reference samples. Target samples are obtained when historical messages are pushed in the target order, and reference samples are obtained when historical messages are pushed in the reference order. The server can also retrieve training sample pairs and corresponding interaction difference labels uploaded by the terminal. The server can also obtain training sample pairs and corresponding interaction difference labels from the service provider that provides the data service.

[0072] In one embodiment, the server can obtain a target sample set and a reference sample set from a database, perform sample matching based on the target sample set and the reference sample set to obtain a training sample pair set, and determine training sample pairs from the training sample pair set. The server can also obtain the target interaction label corresponding to the target sample in the target sample set and the reference interaction label corresponding to the reference sample in the reference sample set, and then use the target interaction label and the reference interaction label to calculate the interaction difference label of the sample pair. Here, the target interaction label is a label used to characterize whether there is interaction with historical messages pushed in the target order, and the reference interaction label is a label used to characterize whether there is interaction with historical messages pushed in the reference order.

[0073] Step 204: Input the target sample into the initial push difference prediction model to predict the interaction probability difference and obtain the interaction difference degree corresponding to the target sample. Input the reference sample into the initial push difference prediction model to predict the interaction probability difference and obtain the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order. The reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0074] The initial push difference prediction model refers to the push difference prediction model with initialized model parameters. These parameters can be initialized to zero, randomly, or using a Gaussian distribution, etc. The push difference prediction model can be an artificial intelligence model built using a neural network to predict the degree of interaction difference. The degree of interaction difference characterizes the difference between the target interaction probability and the reference interaction probability. The target interaction probability characterizes the interaction probability obtained when messages are pushed in the target order, and the reference interaction probability characterizes the interaction probability obtained when messages are pushed in the reference order. The interaction probability characterizes the probability that the push recipient will interact with the push message. The higher the interaction probability, the more likely the push recipient is to interact with the push message. This interaction includes, but is not limited to, viewing, commenting on, forwarding, and sharing messages.

[0075] Specifically, the server inputs the target sample into the initial push difference prediction model. The model predicts the target interaction probability and the reference interaction probability corresponding to the target sample. Then, it calculates the difference between these two probabilities to obtain the degree of interaction difference for the target sample. Next, the server inputs a reference sample into the initial push difference prediction model. This model then predicts the target interaction probability and the reference interaction probability for the reference sample, and again calculates the difference between these two probabilities to obtain the degree of interaction difference for the reference sample.

[0076] Step 206: Calculate the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample to obtain the training interaction difference, and calculate the loss on the interaction difference label based on the training interaction difference and the sample to obtain the training loss information.

[0077] Among them, training interaction difference refers to the difference in the degree of interaction between the target sample and the reference sample. It is used to characterize whether the interaction result of pushing according to the target order is positive or negative compared to pushing according to the reference order. A positive interaction result indicates that pushing according to the target order will not waste server resources, while a negative interaction result indicates that pushing according to the target order will waste server resources. Training loss information is used to characterize the error between the training interaction difference obtained during training and the actual sample pair interaction difference labels.

[0078] Specifically, the server calculates the difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample to obtain the training interaction difference. Then, it uses a preset loss function to calculate the error between the training interaction difference and the interaction difference label of the sample pair to obtain the training loss information. The preset loss function can be the cross-entropy loss function.

[0079] Step 208: Update the initial push difference prediction model based on the training loss information, and return to the step of obtaining training sample pairs and corresponding sample pair interaction difference labels. Iterate until the training completion condition is met, and obtain the target push difference prediction model. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference.

[0080] Among them, the training completion condition refers to the pre-set conditions for the model training to be completed, including but not limited to reaching the maximum number of iterations, the model parameters no longer changing, or the model loss being less than a preset loss threshold. The target interaction difference degree refers to the interaction difference obtained from the push object feature data input to the target push difference prediction model. This target interaction difference degree is used to characterize the interaction difference between pushing messages to the push object in the target order and pushing messages in the reference order.

[0081] Specifically, the server determines whether training has met the completion conditions, such as whether the training loss information is less than a preset loss threshold or whether the maximum number of iterations has been reached. If the training completion conditions are not met, the server uses the training loss information to back-update the model parameters of the initial push difference prediction model. Gradient descent can be used for this back-up update to obtain the updated push difference prediction model. This updated model is then used as the initial push difference prediction model, and the steps of obtaining training sample pairs and corresponding sample pair interaction difference labels are iteratively executed until the training completion conditions are met. The initial push difference prediction model at the point of completion is then used as the target push difference prediction model. The target push difference prediction model can then be deployed and used. During use, the feature data of the push object can be input into the target push difference prediction model to obtain the output target interaction difference degree. The server then determines whether to push messages in the target order based on the magnitude of the target interaction difference degree. For example, if the target interaction difference degree exceeds a preset difference degree threshold, messages can be pushed in the target order; if the target interaction difference degree does not exceed the preset difference degree threshold, messages can be pushed in a reference order.

[0082] The aforementioned push model training method trains an initial push difference prediction model using target samples and reference samples to obtain training interaction differences. Then, it calculates the loss using the training interaction differences and sample interaction difference labels to obtain training loss information. This training loss information is used to update the initial push difference prediction model, and this process is iterated until the training completion condition is met, resulting in a target push difference prediction model. The target push difference prediction model is then used to calculate the loss and update the initial push difference prediction model using training interaction differences and sample interaction difference labels, thereby improving the accuracy of the obtained target push difference prediction model. This target push difference prediction model is then used to predict the degree of target interaction difference when pushing messages according to the target order versus the reference order, thus improving the accuracy of the target interaction difference degree. Pushing messages according to the target interaction difference degree further improves message push accuracy and avoids the problem of wasting server resources.

[0083] In one embodiment, such as Figure 3 As shown, in step 202, the target sample is input into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the target sample. Similarly, the reference sample is input into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the reference sample. This includes:

[0084] Step 302: Input the target sample into the initial push difference prediction model, and use the initial push difference prediction model to predict the target interaction probability of the target sample to obtain the target interaction probability corresponding to the target sample.

[0085] Step 304: Predict the reference interaction probability of the target sample using the initial push difference prediction model to obtain the reference interaction probability corresponding to the target sample.

[0086] The target interaction probability corresponding to the target sample is used to characterize the interaction probability of the target sample when historical messages are pushed in the target order. The reference interaction probability corresponding to the target sample is used to characterize the interaction probability of the target sample when historical messages are pushed in the reference order.

[0087] Specifically, the initial push difference prediction model can be a two-branch neural network, comprising a reference interaction probability prediction branch network and a target interaction probability prediction branch network. The server inputs the target sample into the initial push difference prediction model. This model can first extract the target depth features corresponding to the target sample using a feature extraction deep neural network, and then input these target depth features into both the reference interaction probability prediction branch network and the target interaction probability prediction branch network. These branches can be classification prediction networks, such as fully connected networks. The target interaction probability prediction branch network then uses the target depth features to predict the target interaction probability, obtaining the target interaction probability corresponding to the target sample. Finally, the reference interaction probability prediction branch network uses the target depth features to predict the reference interaction probability, obtaining the reference interaction probability corresponding to the target sample.

[0088] Step 306: Calculate the difference between the target interaction probability corresponding to the target sample and the reference interaction probability corresponding to the target sample using the initial push difference prediction model, and obtain the interaction difference degree corresponding to the target sample.

[0089] Among them, the degree of interaction difference corresponding to the target sample is used to characterize the difference in the interaction probability of the target sample being pushed messages according to the target order compared to being pushed according to the reference order.

[0090] Specifically, the initial push difference prediction model inputs the target interaction probability corresponding to the target sample and the reference interaction probability corresponding to the target sample into the interaction probability difference calculation network for calculation, and obtains the interaction difference degree corresponding to the target sample as the output. This interaction probability difference calculation network can be a fully connected network.

[0091] Step 308: Input the reference sample into the initial push difference prediction model, and use the initial push difference prediction model to predict the target interaction probability of the reference sample to obtain the target interaction probability corresponding to the reference sample.

[0092] Step 310: Predict the reference interaction probability of the reference sample using the initial push difference prediction model to obtain the reference interaction probability corresponding to the reference sample.

[0093] The target interaction probability corresponding to the reference sample is used to characterize the interaction probability of the reference sample when historical messages are pushed in the target order. The reference interaction probability corresponding to the reference sample is used to characterize the interaction probability of the reference sample when historical messages are pushed in the reference order.

[0094] Specifically, the server inputs reference samples into the initial push difference prediction model. This model first extracts reference depth features corresponding to the reference samples using a feature extraction deep neural network. Then, these reference depth features are input into the reference interaction probability prediction branch network and the target interaction probability prediction branch network, respectively. The target interaction probability prediction branch network uses the reference depth features to predict the target interaction probability, obtaining the target interaction probability corresponding to the reference sample. The reference interaction probability prediction branch network uses the reference depth features to predict the reference interaction probability, obtaining the reference interaction probability corresponding to the reference sample.

[0095] Step 312: Calculate the difference between the target interaction probability corresponding to the reference sample and the reference interaction probability corresponding to the reference sample using the initial push difference prediction model, and obtain the interaction difference degree corresponding to the reference sample.

[0096] The degree of interaction difference corresponding to the reference sample is used to characterize the difference in the likelihood of interaction between the reference sample pushing messages in the target order and pushing them in the reference order.

[0097] Specifically, the initial push difference prediction model inputs the target interaction probability corresponding to the target sample and the reference interaction probability corresponding to the target sample into the interaction probability difference calculation network for calculation, and obtains the interaction difference degree corresponding to the reference sample.

[0098] In a specific embodiment, such as Figure 4The diagram shows the architecture of the push difference prediction model. Specifically, this model includes a deep network for feature extraction, a target probability prediction network, a reference probability prediction network, and a probability difference calculation network. During training, after initializing the model parameters, the training samples are used as input. The deep network, target probability prediction network, reference probability prediction network, and probability difference calculation network in the initial push difference prediction model perform predictions to obtain the interaction difference level corresponding to the training samples output by the model.

[0099] In the above embodiments, the reference interaction probability prediction and target interaction probability prediction are performed on the target sample and the reference sample respectively by the initial push difference prediction model, so as to obtain the target interaction probability and reference interaction probability corresponding to the target sample and the reference sample respectively. Then, the difference is calculated to obtain the interaction difference degree, thereby improving the accuracy of the obtained interaction difference degree.

[0100] In one embodiment, the target sample includes target time features, target object features, and target order features, and the reference sample includes reference time features, reference object features, and reference order features;

[0101] Step 202 involves inputting the target sample into the initial push difference prediction model to predict the interaction probability difference, obtaining the interaction difference degree corresponding to the target sample, and inputting the reference sample into the initial push difference prediction model to predict the interaction probability difference, obtaining the interaction difference degree corresponding to the reference sample. This includes the following steps:

[0102] The target time features, target object features, and target order features are input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the degree of interaction difference of the target samples at the target time. Similarly, the reference time features, reference object features, and reference order features are input into the initial push difference prediction model to predict the interaction probability difference, thus obtaining the degree of interaction difference of the reference samples at the reference time.

[0103] Among them, the target time feature refers to the time feature corresponding to the target sample, and the reference time feature refers to the time feature corresponding to the reference sample. This time feature is a characteristic representing a specific moment, which can be a specific point in time or a time period. This time feature is used to characterize the degree of interaction difference that the model wants to predict; it represents the degree of interaction difference at the time corresponding to this time feature. The participating time and the target time can be the same. The target object feature refers to the object feature corresponding to the target sample, and the reference object feature refers to the object feature corresponding to the reference sample. This object feature includes object attribute features and object historical message interaction features. Object attribute features are used to characterize the basic attributes of the object. Object historical message interaction features are used to characterize the message data of historical interactions, which can be historical interaction data, historical interaction category interests, historical interaction times, etc. The target order feature refers to the push order feature corresponding to the target sample, and the reference order feature refers to the push order feature corresponding to the reference sample. This push order feature includes the target order feature and the reference order feature. Each training sample will contain a target sample with the target order feature and a reference sample with the reference order feature.

[0104] Specifically, the server inputs the target time features, target object features, and target order features into the initial push difference prediction model. The model then uses a deep network for feature extraction, a target probability prediction network, a reference probability prediction network, and a probability difference calculation network to predict interaction probability differences, obtaining the degree of interaction difference of the target sample at the target time corresponding to the target time features. Next, the reference time features, reference object features, and reference order features are input into the initial push difference prediction model for interaction probability difference prediction. Again, the model uses a deep network for feature extraction, a target probability prediction network, a reference probability prediction network, and a probability difference calculation network to predict interaction probability differences, obtaining the degree of interaction difference of the reference sample at the reference time corresponding to the reference time features.

[0105] In a specific embodiment, when applied to subscription account message push, training samples are obtained by pre-publishing messages and collecting subscriber interaction information. Specifically, the server obtains the training samples, whose data can be, for example, [time t, confounding variable X, pseudo-variable FT], where "time t" is the time marker. "Confounding variable X" consists of relevant features of subscribers, including basic attributes, historical interaction data, historical interaction category interests, and historical interaction times. The pseudo-variable FT is used to characterize whether the training sample is a reference sample or a target sample; for example, when FT = 1, the training sample can be a target sample, and when FT = 0, the training sample can be a reference sample. Then, [time t, confounding variable X, pseudo-variable FT] is used as input to the initial push difference prediction model, predicting the target interaction probability and the reference interaction probability of this input. The difference between the target interaction probability and the reference interaction probability is then calculated to obtain the training interaction difference at time t. Using time t as input data, the training interaction difference at time t can be predicted, thereby improving the accuracy of the obtained training interaction difference. One approach is to divide a day into 12 time periods, using these 12 time periods as time features to obtain the training interaction differences at these 12 time periods. This allows the trained target push difference prediction model to predict the degree of target interaction differences at these 12 time periods throughout the day.

[0106] In the above embodiments, by inputting the target time features, target object features, and target order features into the initial push difference prediction model to predict the interaction probability difference, the degree of interaction difference of the target sample at the target time is obtained. By inputting the reference time features, reference object features, and reference order features into the initial push difference prediction model to predict the interaction probability difference, the degree of interaction difference of the reference sample at the reference time can be obtained, which can improve the accuracy of obtaining the degree of interaction difference.

[0107] In one embodiment, such as Figure 5 As shown, step 202 involves obtaining the training sample pairs and their corresponding sample pair interaction difference labels, including:

[0108] Step 502: Obtain the target sample set and the corresponding target sample push interaction result set, as well as the reference sample set and the corresponding reference sample push interaction result set.

[0109] The target sample push interaction result set includes the target sample push interaction results for each target sample in the target sample set. The target sample push interaction results refer to the interaction results obtained after pushing historical messages to the target sample in the target order. These results include those with interactive behavior and those without. The reference sample push interaction result set includes the reference sample push interaction results for each reference sample in the reference sample set. The reference sample push interaction results refer to the interaction results obtained after pushing historical messages to the reference sample in the reference order. These results include those with interactive behavior and those without. Results with interactive behavior indicate that the push recipient has interacted with the pushed message; such interactions can include viewing, commenting, forwarding, sharing, etc. Results without interactive behavior indicate that the push recipient has not interacted with the pushed message.

[0110] Specifically, the server can retrieve the target sample set and its corresponding target sample push interaction result set, as well as the reference sample set and its corresponding reference sample push interaction result set, from the database. The server can also retrieve these from the data service provider. Furthermore, the server can retrieve the target sample set and its corresponding target sample push interaction result set, as well as the reference sample set and its corresponding reference sample push interaction result set, from the message push service provider. Finally, the server can retrieve the target sample set and its corresponding target sample push interaction result set, as well as the reference sample set and its corresponding reference sample push interaction result set, uploaded via the terminal.

[0111] Step 502: Match each target sample in the target sample set with each target sample in the reference sample set to obtain each training sample pair.

[0112] Step 502: Determine the push interaction result pairs corresponding to each training sample pair based on the target sample push interaction result set and the reference sample push interaction result set.

[0113] The push interaction results include the push interaction results corresponding to the target sample and the push interaction results corresponding to the reference sample in the training sample pair.

[0114] Specifically, the server randomly pairs each target sample in the target sample set with each target sample in the reference sample set, using each matched target sample and reference sample as a training sample pair. Then, it obtains the push interaction results of the reference sample and the target sample from each training sample pair to obtain the push interaction result pair.

[0115] Step 502: Based on the push interaction results, compare the push interaction results of the target sample and the push interaction results of the reference sample to obtain the historical interaction differences, and determine the corresponding sample pair interaction difference label for each training sample pair based on the historical interaction differences.

[0116] Among them, historical interaction differences are used to characterize the differences in interaction results obtained based on historical message pushes.

[0117] Specifically, the server compares each push interaction result with the target sample push interaction result and the reference sample push interaction result. Based on the comparison results, it obtains the historical interaction difference and then determines the corresponding sample pair interaction difference label for each training sample pair. For example, if the target sample pushes historical messages according to interest and obtains an interaction result with interaction behavior, while the reference sample pushes historical messages in chronological order and obtains an interaction result without interaction behavior, the historical interaction difference is positive, and the sample pair interaction difference label can be set to 1. If the target sample pushes historical messages according to interest and obtains an interaction result without interaction behavior, while the reference sample pushes historical messages in chronological order and obtains an interaction result with interaction behavior, the historical interaction difference is negative, and the sample pair interaction difference label can be set to 0. In other cases, it is considered that the interaction difference resulting from pushing in the target order relative to pushing in the reference order is uncertain, and the sample pair interaction difference label can be set to 0.5.

[0118] Step 502: Obtain the training sample pairs and their corresponding interaction difference labels from each training sample pair and the corresponding sample pair interaction difference labels.

[0119] Specifically, the server can randomly select the training sample pairs and their corresponding interaction difference labels to be used during the current training from each training sample pair and its corresponding interaction difference labels. Alternatively, the server can select the training sample pairs and their corresponding interaction difference labels according to the type of the interaction difference labels. For example, it can select interaction difference labels representing positive interactions and their corresponding training sample pairs, or it can select interaction difference labels representing negative interactions and their corresponding training sample pairs.

[0120] In one specific embodiment, examples of training sample pairs and corresponding sample pair interaction difference labels can be shown in Table 1 below.

[0121] Table 1 Examples of Training Samples

[0122] [u1, T=1, y=1, s1] [u5, T=0, y=0, s5] [1, s1-s5] [u2, T=1, y=1, s2] [u6, T=0, y=1, s6] [0.5, s2-s6] [u3, T=1, y=0, s3] [u7, T=0, y=0, s7] [0.5, s3-s7] [u4, T=1, y=0, s4] [u8, T=0, y=1, s8] [0, s4-s8]

[0123] Where ui represents the recommended object in the i-th training sample, where i is a positive integer. For example, i can be 1, meaning u1 represents the recommended object in the first training sample. T represents the message push order. T=1 indicates message push according to the target order. T=0 indicates message push according to the reference order. Si represents the interaction difference obtained by inputting the i-th training sample into the initial push difference prediction model. When the training sample is the target sample, the interaction difference corresponding to the target sample is obtained. When the training sample is the reference sample, the interaction difference corresponding to the reference sample is obtained. Subtracting the interaction difference corresponding to the target sample from the interaction difference corresponding to the reference sample yields the training interaction difference. For example, s1-s5 represents the training interaction difference obtained during training when the sample pair consisting of target sample 1 and target sample 5 is used. A 1 in the sample pair indicates that the interaction difference is positive when the target sample is pushed according to the target order compared to when the reference sample is pushed according to the time order. A 0 in the sample pair indicates that the interaction difference is negative when the target sample is pushed according to the target order compared to when the reference sample is pushed according to the time order. The 0.5 in the sample pair indicates that the degree of interaction difference when the target sample pushes messages in the target order versus the reference sample pushes messages in chronological order is uncertain. The 0, 1, and 0.5 in the sample pair are the interaction difference labels.

[0124] In the above embodiments, a target sample set and its corresponding target sample push interaction result set, as well as a reference sample set and its corresponding reference sample push interaction result set, are obtained. Then, based on the target sample push interaction results and the reference sample push interaction results, the interaction difference labels of the corresponding training sample pairs are determined, thereby improving the accuracy of the obtained interaction difference labels of the sample pairs, and consequently improving the accuracy of the obtained training sample pairs and their corresponding interaction difference labels.

[0125] In one embodiment, such as Figure 6 As shown, step 502 involves obtaining the target sample set and the corresponding target sample push interaction result set, as well as the reference sample set and the corresponding reference sample push interaction result set, including:

[0126] Step 602: Obtain each initial push object, and determine each target push object and each reference push object from each initial push object.

[0127] The initial push target refers to the object that can receive messages at the beginning. This object can be a real object, such as a person, or a virtual object, such as a virtual character or a virtual broadcaster. The target push target refers to the object that receives messages according to the target order, and the reference push target refers to the object that receives messages according to the reference order.

[0128] Specifically, the server can retrieve initial push objects from the database. These initial push objects can be pre-configured by the corresponding terminal to receive messages. Then, it can randomly select objects from these initial push objects to push messages in a target order, resulting in target push objects. Alternatively, it can randomly select objects from these initial push objects to push messages in a reference order, resulting in reference push objects. These reference push objects can be the same initial push objects as the target push objects, or they can be different initial push objects.

[0129] Step 604: Obtain the message to be pushed corresponding to each target push object, and push the message to be pushed corresponding to each target push object to the terminal corresponding to each target push object in the target order.

[0130] Here, "messages to be pushed" refers to messages that need to be pushed, and these messages can include multiple messages. Each target push object has a corresponding message to be pushed, which can be the same message or different messages.

[0131] Specifically, the server can retrieve the messages to be pushed to each target object from the database. Then, the server can calculate the interest level between each message and the target object, sort the messages according to the interest level, obtain a message sequence, and then push each message to the terminal of the corresponding target object according to the message sequence. After receiving each message, the terminal of the target object displays each message in sequence.

[0132] Step 606: Obtain the target interaction data returned by the terminal corresponding to each target push object, and obtain the target sample set and the corresponding target sample push interaction result set based on each target push object and the target interaction data.

[0133] Among them, target interaction data refers to the data generated after the target push object interacts with various push messages through the terminal. It can include message data with interaction behavior, message data without interaction behavior, interaction time data, interaction category interest data, etc.

[0134] Specifically, after the target push object interacts with the message displayed on the terminal, target interaction data is generated. The terminal can then upload this target interaction data to the server. The server obtains the target interaction data returned by the terminal for each target push object. Then, based on the target push object and the corresponding target interaction data, it obtains the target sample pairs and the corresponding target sample push interaction results. By traversing each target push object, it obtains the target sample set and the corresponding target sample push interaction result set.

[0135] Step 608: Obtain the message to be pushed corresponding to each reference push object, and push the message to be pushed corresponding to each reference push object to the terminal corresponding to each reference push object in the reference order.

[0136] Specifically, the server can retrieve the messages to be pushed to each reference push object from the database, then retrieve the push reference order from the database, and then push each message to the terminal of the corresponding reference push object according to the reference order. After receiving each message, the terminal of the reference push object displays each message in chronological order.

[0137] Step 610: Obtain the reference interaction data returned by the terminal corresponding to each reference push object, and obtain the reference sample set and the corresponding reference sample push interaction result set based on each reference push object and the reference interaction data.

[0138] Among them, reference interaction data refers to the data generated after the reference push object interacts with the various push messages through the terminal.

[0139] Specifically, after a reference push object interacts with a message displayed on a terminal, reference interaction data is generated. The terminal can then upload this data to the server. The server retrieves the reference interaction data returned by the terminal for each reference push object. Then, based on the reference push object and its corresponding reference interaction data, it obtains reference sample pairs and corresponding reference sample push interaction results. By iterating through each reference push object, it obtains a reference sample set and its corresponding reference sample push interaction result set.

[0140] In a specific embodiment, to predict the difference in interaction degree when pushes according to the target order relative to the reference order, i.e., the increment of the preset intervention, training samples can be obtained through A / B experiments. That is, each initial push object is obtained, and then randomly sampled to obtain each push object of the experimental group and each push object of the control group. Then, each push object of the experimental group is pushed with messages according to the target order, i.e., random order delivery. Random order delivery refers to the entire process of push service, including message recall, coarse ranking, fine ranking and mixed ranking, and then push. Then, the push interaction results corresponding to each target push object are obtained, and thus the target sample set and the target sample push interaction result set are obtained. At the same time, each push object of the control group is pushed with messages according to the reference order, and then the push interaction results corresponding to each reference push object are obtained, and thus the reference sample set and the reference sample push interaction result set are obtained. Then, a push difference prediction model is established by conditional average causal effect and trained with training samples. The conditional average causal effect can be defined as the difference in average interaction probability caused by the different push order, as shown in the following formula (1).

[0141] CATE:τ=E[Y i (1)-Y i (0)|X i ] Formula (1)

[0142] Where i represents the index of the training sample. X i Y represents the confounding variable for the i-th training sample. Confounding variables refer to relevant features of the target object, including basic attributes, historical interaction data, historical interaction categories and interests, and historical interaction times. i (1) indicates that the push interaction result is obtained after the i-th training sample pushes messages in the target order. i (0) indicates the push interaction result obtained after the i-th training sample is pushed according to the reference order. CATE:τ refers to the average interaction probability difference obtained using conditional average causal effect. E refers to expectation.

[0143] In the above embodiments, after determining each target push object and each reference push object from each initial push object, the corresponding message to be pushed is obtained and the corresponding interaction data is obtained. Then, the push object and the interaction data are used to obtain the sample set and the corresponding sample push interaction result set, which improves the accuracy of obtaining the sample set and the corresponding sample push interaction result set.

[0144] In one embodiment, step 602, determining each target push object and each reference push object from each initial push object, includes the following steps:

[0145] Each initial target push object and each initial reference push object are determined from each initial push object; a homogeneity check is performed on each initial target push object and each initial reference push object. When the homogeneity check passes, each initial target push object is used as the target push object, and each initial reference push object is used as the reference push object.

[0146] The initial target push object refers to the target push object that needs to undergo homogeneity testing. The initial reference push object refers to the reference push object that needs to undergo homogeneity testing. Homogeneity testing is used to check whether the distribution of indicators across important dimensions is consistent.

[0147] Specifically, the server can randomly sample from each initial push object to obtain initial target push objects and initial reference push objects. Then, a homogeneity check algorithm is used to check the homogeneity of each initial target push object and each initial reference push object. This homogeneity check algorithm can be a balance check algorithm. If the homogeneity check fails, random sampling needs to be performed again from each initial push object to obtain updated initial target push objects and initial reference push objects. The homogeneity check is then performed again until the homogeneity check passes. Finally, the initial target push objects are designated as the target push objects, and the initial reference push objects are designated as the reference push objects.

[0148] In the above embodiments, by performing homogeneity checks on each initial target push object and each initial reference push object, when the homogeneity check passes, each initial target push object is used as the target push object and each initial reference push object is used as the reference push object. This ensures that each target push object and each reference push object are homogeneous, thereby ensuring that the final push interaction probability is entirely determined by whether the push is performed in the target order and the reference order. Then, each target push object and each reference push object are used for model training, thereby improving the accuracy of model training.

[0149] In one embodiment, a homogeneity check is performed based on each initial target push object and each initial reference push object, including the following steps:

[0150] Obtain the initial target object features corresponding to each initial target push object and the initial reference object features corresponding to each initial reference push object; calculate the standard mean difference based on the initial target object features and the initial reference object features to obtain the target value; when the target value does not exceed the preset threshold, determine that the homogeneity test of each initial target push object and each initial reference push object has passed; when the target value exceeds the preset threshold, determine that the homogeneity test of each initial target push object and each initial reference push object has failed.

[0151] The initial target object features refer to the characteristics corresponding to the initial target push object. These features can include the target push object's attribute features and historical message interaction features. Attribute features refer to the basic attributes of the target push object, such as age, gender, and region. Historical message interaction features are used to characterize the message data of historical interactions, which can include specific messages, historical interaction categories / interests, and historical interaction times. The initial reference object features refer to the characteristics corresponding to the initial target push object, and these features can include the reference push object's attribute features and historical message interaction features. The target value is a value used to characterize the homogeneity between each initial target push object and each initial reference retirement object. The smaller the target value, the higher the homogeneity between the initial target push objects and each initial reference retirement object. The larger the target value, the lower the homogeneity between the initial target push objects and each initial reference retirement object. The preset threshold is a pre-set value used to distinguish the passing of the homogeneity test; for example, the preset threshold can be set to 0.2.

[0152] In one specific embodiment, the difference between the standard mean of each initial target push object and each initial reference push object can be calculated to determine whether the homogeneity test passes. That is, the homogeneity test can be performed using the formula (2) shown below.

[0153]

[0154] Where T=1 represents the push objects pushed according to the target order, i.e., each initial target push object. X refers to the confounding variable, representing the relevant characteristics of the push object, including basic attributes, historical interaction data, historical interaction category interests, historical interaction time, etc. T=0 represents the push objects pushed according to the reference order, i.e., each initial reference push object. mean(X,T=1) calculates the mean of the confounding variable for each initial target push object. mean(X,T=0) calculates the mean of the confounding variable for each initial reference push object. var(X,T=1) calculates the variance of the confounding variable for each initial target push object. var(X,T=0) calculates the variance of the confounding variable for each initial reference push object. For example, if the confounding variable X can be age, then the mean and variance of the age of each initial target push object are calculated, then the mean and variance of the age of each initial reference push object are calculated, and then the standard mean difference is calculated. When the standard mean difference does not exceed the preset threshold, the homogeneity test passes; when the standard mean difference does not exceed the preset threshold, the homogeneity test fails. The preset threshold can be set to 0.2. When the difference between the standard means does not exceed 0.2, it indicates that the flow of the control group and the experimental group in the age dimension of the confounding variable is uniform, and the homogeneity test is passed.

[0155] In the above embodiments, by calculating the initial target object features and the initial reference object features, and then calculating the standard mean difference, the target value is obtained. The homogeneity test is then judged based on the target value, thereby improving the accuracy of the homogeneity test.

[0156] In one specific embodiment, after determining each target push object and each reference push object from each initial push object, the method further includes the following steps:

[0157] Each candidate interaction feature is acquired and used as a result feature. The target order corresponding to each target push object is used as a cause feature. Causal discovery is performed based on each result feature and cause feature to obtain a causal relationship graph. Based on the causal relationship graph, the push interaction result feature is determined from each candidate interaction feature.

[0158] Candidate interaction features refer to candidate features used to characterize the results of message interactions. Outcome features refer to features that serve as results in causal discovery. Causal features refer to features that serve as causes in causal discovery. A causal relationship graph is a directed acyclic graph used to represent causal relationships, where one event leads to another.

[0159] Specifically, the server can retrieve candidate interaction features from the database and directly use them as result features. Alternatively, the server can retrieve candidate interaction features uploaded by the terminal or obtained from the push service provider. Simultaneously, pushing according to the target order is set as the cause, meaning the target order corresponding to each target push object is used as the cause feature. This determines whether pushing events according to the target order led to the occurrence of the interaction event corresponding to the candidate interaction feature. The server then uses the target order corresponding to each target push object as the cause feature and employs a causal discovery algorithm to perform causal discovery using the result features and cause features, obtaining a causal relationship graph. This causal discovery algorithm can be a condition-constrained algorithm, an algorithm based on a causal function model, or a hybrid method, etc. Finally, based on the causal relationship graph, the push interaction result features are determined from the candidate interaction features.

[0160] Based on each target push object and target interaction data, the target sample set and the corresponding target sample push interaction result set are obtained, including the following steps:

[0161] Determine the feature values ​​corresponding to the push interaction results from the target interaction data, and use the feature values ​​as target samples to push interaction results.

[0162] Specifically, when determining the labels corresponding to training samples, the server can either determine the feature values ​​corresponding to the push interaction result features from the target interaction data and use these feature values ​​as the target sample to push the interaction result, or it can determine the feature values ​​corresponding to the push interaction result features from the reference interaction data and use these feature values ​​as the reference sample to push the interaction result. Then, the target sample push interaction result and the reference sample push interaction result are used to determine the interaction difference labels for the sample pairs.

[0163] In a specific implementation, after passing the homogeneity test, it is necessary to identify the interaction variables that have a causal relationship with the target order. In the application scenario of message subscription, candidate interaction variables can be obtained, including message box exposure, message box entry, message box interaction, message interaction, and video account exposure. Then, causal discovery is used to determine the interaction variables that truly affect the target order, that is, to identify sensitive target groups through the maximum causal effect of the interaction variables. The Peter Clark algorithm (PC) based on conditional constraints can be used to build a causal relationship graph. Figure 7 The diagram illustrates the structure of a causal relationship graph. Specifically, the server uses message box exposure, message box entry, message box interaction, message interaction, video account exposure, and the target order of each target push object in the experimental group as nodes in the causal relationship graph. Then, the dependencies between nodes are confirmed using the Fisher conditional independence test. For any two nodes in the graph, given other sets, the test determines whether the two nodes are independent, ultimately outputting an undirected graph. Next, the dependency directions of each node are confirmed using the D-separation principle to determine the dependency directions of edges in the graph, expanding the undirected graph into a directed acyclic graph, thus obtaining the causal relationship graph. Based on causal discovery, the result variables besides affecting the target order include message interaction and message publisher exposure. Furthermore, the only parent node affecting the message interaction stage is the experimental group; that is, the only reason that can influence the push object to perform message interaction is whether the message is pushed according to the target order. Therefore, message interaction is used as a feature of the push interaction result.

[0164] In the above embodiments, the push interaction result features are determined from each candidate interaction feature through causal discovery, thereby improving the accuracy of the obtained push interaction result features. Then, the feature values ​​of the push interaction result features are obtained to get the push interaction result, thereby improving the accuracy of obtaining the push interaction results of the training samples.

[0165] In one embodiment, such as Figure 8 As shown, a message push method is provided, which can be applied to... Figure 1Taking a server as an example, it can be understood that this method can also be applied to servers, and also to systems including terminals and servers, and implemented through the interaction between the terminal and the server. This embodiment includes the following steps:

[0166] Step 802: Obtain the message push request, which carries the object to be pushed.

[0167] In this context, a push notification request refers to a request triggered by the target device on its corresponding terminal and sent to the server. This push notification request is used to request a push message. The target device refers to the object that needs to receive a push notification; this object can be a real object or a virtual object.

[0168] Specifically, the terminal corresponding to the target object receives a push notification instruction. This instruction is received when the terminal detects a push notification event triggered by the target object. This event can be triggered by an action on the target object, such as a click, swipe, or voice command. The terminal then responds to the push notification instruction by sending a push notification request to the server. The server receives the request, parses it, and obtains the target object.

[0169] Step 804: Obtain the target interaction difference degree based on the object to be pushed. The target interaction difference degree is obtained by inputting the feature data corresponding to the object to be pushed into the target push difference prediction model to predict the interaction probability difference. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels, inputs the target sample and reference sample distribution in the training sample pairs into the initial push difference prediction model to predict the interaction probability difference, obtains the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample, and calculates the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample to obtain the training interaction difference. Based on the training interaction difference and the sample pair interaction difference labels, the loss is calculated to obtain the training loss information. Based on the training loss information, the initial push difference prediction model is trained. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order, and the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0170] The target interaction difference degree refers to the interaction difference degree corresponding to the object to be pushed, which is obtained by inputting the feature data corresponding to the object to be pushed into the target push difference prediction model to predict the interaction probability difference. The feature data corresponding to the object to be pushed may include the attribute feature data of the object to be pushed and the historical message interaction feature data of the object to be pushed. The target push difference prediction model can be trained using any of the above-mentioned push model training methods. For example, the server can obtain training sample pairs and corresponding sample pair interaction difference labels, input the distribution of target samples and reference samples in the training sample pairs into the initial push difference prediction model to predict the interaction probability difference, obtain the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample, calculate the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample, obtain the training interaction difference, calculate the loss based on the training interaction difference and sample pair interaction difference labels, obtain training loss information, update the initial push difference prediction model based on the training loss information, and return to the step of obtaining training sample pairs and corresponding sample pair interaction difference labels iteratively until the training completion condition is met, and obtain the target push difference prediction model.

[0171] Specifically, the server can directly retrieve the target interaction difference level corresponding to the object to be pushed from the database. That is, this target interaction difference level is obtained by pre-inputting the attribute feature data of the object to be pushed and its historical message push interaction feature data into the target push difference prediction model to predict the interaction probability difference. This historical message interaction feature data may include specific message features of historical interactions, message category and interest features of historical interactions, and time features of historical interactions.

[0172] In one specific embodiment, the server can also obtain the feature data corresponding to the object to be pushed, and then input the feature data corresponding to the object to be pushed into the target push difference prediction model. The target push difference prediction model uses a dual-branch network to predict the target interaction probability and reference interaction probability corresponding to the object to be pushed, and then calculates the difference between the target interaction probability and the reference interaction probability corresponding to the object to be pushed to obtain the degree of target interaction difference corresponding to the object to be pushed.

[0173] Step 806: When the degree of difference in target interaction meets the preset target order push conditions, respond to the message push request, obtain each message to be pushed corresponding to the object to be pushed and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the object to be pushed according to the target push order.

[0174] The preset target order push condition refers to the pre-set threshold for the degree of interaction difference when pushing messages in a target order, which can be set according to needs. The target push order refers to the order obtained by sorting the messages to be pushed according to information such as interest or relevance calculated using server resources.

[0175] Specifically, the server determines whether the degree of difference in the target interaction meets the preset target order push conditions. For example, the server can compare the degree of difference in the target interaction with a preset threshold for the degree of difference in the interaction. When the degree of difference in the target interaction exceeds the preset threshold, it means that pushing according to the target order increases the likelihood of interaction compared to pushing according to the reference order. At this time, the degree of difference in the target interaction meets the preset target order push conditions, and the server responds to the message push request. The server can directly retrieve each message to be pushed corresponding to the target object and the target push order corresponding to each message from the database, and push each message to be pushed to the terminal corresponding to the target object according to the target push order. In one embodiment, when the server can retrieve each message to be pushed corresponding to the target object from the database, it can calculate the target object's interest or relevance to each message. This calculation can be performed using a neural network model. After obtaining the calculation results, the messages to be pushed are sorted according to the magnitude of interest or relevance to obtain the target push order. Then, each message to be pushed is pushed to the terminal corresponding to the target object according to the target push order. The terminal receives each message to be pushed according to the target push order and displays them sequentially.

[0176] The aforementioned message push method obtains a message push request, which carries the target object to be pushed to. Based on the target object, it obtains the corresponding target interaction difference degree. When the target interaction difference degree meets the preset target order push conditions, it responds to the message push request, obtains each message to be pushed corresponding to the target object and the target push order for each message, and pushes each message to the terminal corresponding to the target object in the target push order. Since the target interaction difference degree is obtained using a target push difference prediction model, its accuracy is improved. Furthermore, only when the target interaction difference degree meets the preset target order push conditions are the messages to be pushed to the terminal corresponding to the target object pushed in the target push order, thus avoiding server resource waste and saving server resources.

[0177] In one embodiment, a push notification request includes at least two components; such as... Figure 9 As shown, message push methods also include:

[0178] Step 902: Obtain each message push request, and obtain the corresponding target interaction difference degree based on the push object in each message push request.

[0179] The target interaction difference level to be classified refers to the degree of difference in target interactions among the objects to be pushed to, obtained when multiple push notification requests are received. This target interaction difference level can be used to classify each push notification request.

[0180] Specifically, when at least two terminals with target objects send push notification requests to the server, the server simultaneously receives all push notification requests, each carrying a corresponding target object. At this point, the server can directly retrieve the target interaction difference degree corresponding to the target object in each push notification request from the database, thus obtaining the target interaction difference degree to be segmented for each target object in each push notification request. Alternatively, the server can obtain the feature data corresponding to the target object in each push notification request, input the feature data for each target object into the target push difference prediction model for prediction, obtain the output target interaction difference degree, and then use the target interaction difference degree corresponding to each target object as the target interaction difference degree to be segmented.

[0181] Step 904: Determine the target message push request and non-target message push request from each message push request according to the degree of difference in the target interaction to be divided.

[0182] Target message push requests refer to message push requests that require messages to be pushed in a target order. Non-target message push requests refer to message push requests that do not require messages to be pushed in a target order, but can be pushed in a reference order.

[0183] Specifically, the server sorts the push requests according to the degree of difference in the target interactions to be categorized. The top-ranked push requests are then designated as target push requests, while the bottom-ranked requests are designated as non-target push requests. For example, the top 50% of push requests can be selected as target push requests, and the bottom 50% as both target and non-target push requests. Sorting the push requests based on the degree of difference in target interactions improves the accuracy of the sorting results, thereby improving the accuracy of identifying target and non-target push requests.

[0184] Step 906: Respond to the target message push request, obtain each target message to be pushed corresponding to the target message push request and the target push order corresponding to each target message to be pushed, and push each target message to be pushed to the terminal corresponding to the target message push request in the target push order.

[0185] Here, the target message to be pushed refers to the message to be pushed corresponding to the target message push request. This target message to be pushed needs to be pushed in a target order. A target message push request can correspond to multiple target messages to be pushed. The target push order refers to the target order corresponding to each target message to be pushed, which needs to be calculated using server resources. Different target message push requests correspond to different target messages to be pushed, and different target messages to be pushed have different target push orders.

[0186] Specifically, the server responds to the target message push request and locates the corresponding target messages based on the target objects specified in the request. Then, it obtains the target push order for each target message. This can be achieved by calculating the interest level between the target objects and each target message, and then sorting the messages according to this interest level. For example, the server can use a neural network model to calculate the interest level between the target objects and each target message. Alternatively, the server can calculate the relevance between the target objects and each target message, and then sort the messages according to this relevance. Again, the server can use a neural network model to calculate the relevance between the target objects and each target message. Finally, the server pushes each target message to the terminal corresponding to the target message push request according to the target push order. The terminal receives and displays the target messages sent by the server according to the target push order.

[0187] Step 908: Respond to the non-target message push request, obtain each non-target message to be pushed corresponding to the non-target message push request and the non-target push order corresponding to each non-target message to be pushed, and push each non-target message to be pushed to the terminal corresponding to the non-target message push request in the non-target push order.

[0188] Among them, non-target push messages refer to the messages to be pushed corresponding to a non-target message push request. These non-target push messages need to be pushed according to a reference order. A non-target message push request can correspond to multiple non-target push messages. The non-target push order refers to the reference order corresponding to each non-target push message, which can be obtained without calculating using server resources. Different non-target message push requests correspond to different non-target push messages, and different non-target push messages have different non-target push orders.

[0189] Specifically, the server responds to non-target message push requests, locates the corresponding non-target messages to be pushed based on the objects to be pushed in the requests, and then obtains the non-target push order for each message. This non-target push order can be based on chronological order. The terminal then receives and displays the non-target messages sent by the server according to the non-target push order.

[0190] In the above embodiments, when multiple message push requests are obtained simultaneously, the message push requests can be divided according to the degree of difference in the target interaction corresponding to the object to be pushed in the message push request, resulting in target message push requests and non-target message push requests. Then, the target message push requests are responded to and the messages are pushed in the target order, while the non-target message push requests are responded to and the messages are pushed in the reference order. This avoids pushing message push requests with small differences in the degree of interaction in the target order, thereby saving server resources while ensuring the quality of push service.

[0191] In one embodiment, prior to step 902 and before obtaining each message push request, the following step is also included:

[0192] During historical idle periods, feature data corresponding to each historical object is acquired. The feature data includes the target time. The feature data corresponding to each historical object is input into the target push difference prediction model to predict the difference in interaction probability. The interaction difference degree of each historical object at the target time is obtained. The interaction difference degree at the target time is used to characterize the target interaction difference degree at the target time. The interaction difference degree of each historical object and the corresponding target time is saved.

[0193] In this context, "historical idle time" refers to a period when the server is relatively idle, meaning its resource utilization is low and it has spare resources. For example, a historical idle time could be a low point in the number of push message requests, during which the server has idle resources to process these requests. "Historical object" refers to the object to be pushed during a historical idle time. "Target time" refers to a period when the server is relatively busy, meaning its resource utilization is high and it has few or no idle resources. "Target time interaction difference" refers to the degree of difference in target interactions corresponding to historical objects at the target time.

[0194] Specifically, during historical idle periods, the server can directly retrieve the feature data corresponding to each historical object from the database, including the target time. Then, the server inputs the feature data corresponding to each historical object into the target push difference prediction model to predict the interaction probability difference, obtaining the interaction difference degree corresponding to the target time for each historical object. The server then associates and saves each historical object with the corresponding interaction difference degree at the target time.

[0195] Step 902: Obtain each message push request, and based on the target object to be pushed in each message push request, obtain the corresponding target interaction difference degree, including the following steps:

[0196] At the target time, obtain each message push request, and based on the correlation between each historical object and the degree of interaction difference at the corresponding target time, obtain the degree of target interaction difference corresponding to the object to be pushed in each message push request.

[0197] The correlation between the degree of interaction difference between each historical object and its corresponding target time is calculated and saved during historical idle time.

[0198] Specifically, when the server receives each message push request at the target time, it can match the object to be pushed in each message push request with each historical object. Based on the saved correlation between each historical object and the corresponding target time interaction difference, it obtains the target time interaction difference corresponding to the matching historical object, thus obtaining the target interaction difference corresponding to the object to be pushed. Then, the found target difference can be used to sort each message push request, improving efficiency.

[0199] In one specific embodiment, by dividing each message push request into target message push requests and non-target message push requests according to the degree of difference in target interaction at a target time, and then responding to the target message push requests and pushing messages in the target order, and responding to the non-target message push requests and pushing messages in the reference order, server resources can be saved while ensuring the quality of push services. Figure 10 The diagram illustrates a comparison of server resources. Dark black represents server resource usage when push requests are categorized into target push requests and non-target push requests. Light black represents server resource usage when all push requests are sent according to the target order. It is clear that by dividing push requests based on the degree of difference in target interactions before responding to them, server resources are maintained in a balanced manner, ensuring the quality of the push service remains unchanged, and saving server resources at the target time.

[0200] In the above embodiments, by acquiring feature data corresponding to each historical object during historical idle times, and using the feature data corresponding to each historical object to predict the interaction probability difference through the target push difference prediction model, the interaction difference degree corresponding to each historical object at the target time is obtained and associated and saved. Then, when each message push request is obtained at the target time, the target interaction difference degree corresponding to the object to be pushed in each message push request can be obtained directly according to the saved association relationship. This can keep the server resources in a balanced state of use, avoid server resource problems at the target time, and thus ensure the quality of push services.

[0201] In a specific embodiment, such as Figure 11 As shown, a push model training method is provided, which specifically includes the following steps:

[0202] Step 1102: Obtain training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples include target time features, target object features and target order features. The reference samples include reference time features, reference object features and reference order features.

[0203] Step 1104: Input the target time features, target object features, and target sequence features into the initial push difference prediction model. Use the initial push difference prediction model to predict the target interaction probability of the target sample and obtain the target interaction probability corresponding to the target sample.

[0204] Step 1106: Predict the reference interaction probability of the target sample using the initial push difference prediction model to obtain the reference interaction probability corresponding to the target sample.

[0205] Step 1108: Calculate the difference between the target interaction probability corresponding to the target sample and the reference interaction probability corresponding to the target sample using the initial push difference prediction model, and obtain the degree of interaction difference of the target sample at the target time.

[0206] Step 1110: Input the reference time features, reference object features, and reference order features into the initial push difference prediction model. Use the initial push difference prediction model to predict the target interaction probability of the reference sample and obtain the target interaction probability corresponding to the reference sample.

[0207] Step 1112: Predict the reference interaction probability of the reference sample using the initial push difference prediction model to obtain the reference interaction probability corresponding to the reference sample.

[0208] Step 1114: Calculate the difference between the target interaction probability corresponding to the reference sample and the reference interaction probability corresponding to the reference sample using the initial push difference prediction model, and obtain the degree of interaction difference corresponding to the reference sample at the participation time.

[0209] Step 1116: Calculate the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample to obtain the training interaction difference, and calculate the loss of the interaction difference label based on the training interaction difference and the sample to obtain the training loss information.

[0210] Step 1118: Update the initial push difference prediction model based on the training loss information, and return to the step of obtaining training sample pairs and corresponding sample pair interaction difference labels. Iterate until the training completion condition is met to obtain the target push difference prediction model.

[0211] In a specific embodiment, the message push method is applied to the subscription account scenario of an instant messaging application. Specifically: when a user uses the instant messaging application, the server detects that there are various messages to be pushed to the user from the subscription account. At this time, the server obtains the user's attribute information and historical subscription account message interaction data, then extracts features from the user's attribute information and historical subscription account message interaction data to obtain the user's feature data. The user's feature data is then input into a pre-trained and deployed target push difference prediction model. The target push difference prediction model predicts the user's reference interaction probability and target interaction probability, and then calculates the difference between the target interaction probability and the reference interaction probability to obtain the target interaction difference degree. The target interaction difference degree is then compared with a preset threshold. When the target interaction difference degree exceeds the preset threshold, the server calculates the user's interest level in each message to be pushed, and then pushes each message to the user's terminal in order of interest level. When the terminal receives the messages to be pushed in order of interest level, an unread message reminder is displayed in the subscription account message section of the instant messaging application. When a user clicks on a subscription account message on their device, the device enters the subscription account message page, displaying each pending message in order of interest. The user can then view each pending message.

[0212] Furthermore, if the instant messaging application's server receives a large number of push notification requests from users simultaneously, it calculates the degree of target interaction difference for each push notification request using a target push difference prediction model. Then, it categorizes the push notification requests according to this degree of target interaction difference, resulting in push notification requests ordered by interest and those ordered by time. When responding to a push notification request ordered by interest, the user's device will display the messages in that order on the instant messaging application's subscription account message page. When responding to a push notification request ordered by time, the user's device will display the messages in that order on the instant messaging application's subscription account message page.

[0213] 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.

[0214] Based on the same inventive concept, this application also provides a push model training device for implementing the push model training method and a message push device for implementing the message push method described above. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more push model training devices and message push devices provided below can be found in the limitations of the push model training method and message push method described above, and will not be repeated here.

[0215] In one embodiment, such as Figure 12 As shown, a push model training device 1200 is provided, including: a sample acquisition module 1202, a prediction module 1204, a loss calculation module 1206, and an iteration module 1208, wherein:

[0216] The sample acquisition module 1202 is used to acquire training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when historical messages are pushed in the target order, and the reference samples are obtained when historical messages are pushed in the reference order.

[0217] The prediction module 1204 is used to input the target sample into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the target sample. It also inputs the reference sample into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order. The reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0218] The loss calculation module 1206 is used to calculate the difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample, to obtain the training interaction difference, and to calculate the loss based on the training interaction difference and the interaction difference label of the sample to obtain the training loss information.

[0219] The iteration module 1208 is used to update the initial push difference prediction model based on the training loss information and return the step of obtaining training sample pairs and corresponding sample pair interaction difference labels to be executed iteratively until the training completion condition is met, and the target push difference prediction model is obtained. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference.

[0220] In one embodiment, the prediction module 1204 is further configured to input a target sample into an initial push difference prediction model, predict the target interaction probability of the target sample using the initial push difference prediction model to obtain the target interaction probability corresponding to the target sample, predict the reference interaction probability of the target sample using the initial push difference prediction model to obtain the reference interaction probability corresponding to the target sample, calculate the difference between the target interaction probability corresponding to the target sample and the reference interaction probability corresponding to the target sample using the initial push difference prediction model to obtain the interaction difference degree corresponding to the target sample; input a reference sample into the initial push difference prediction model, predict the target interaction probability of the reference sample using the initial push difference prediction model to obtain the target interaction probability corresponding to the reference sample, predict the reference interaction probability of the reference sample using the initial push difference prediction model to obtain the reference interaction probability corresponding to the reference sample, calculate the difference between the target interaction probability corresponding to the reference sample and the reference interaction probability corresponding to the reference sample using the initial push difference prediction model to obtain the interaction difference degree corresponding to the reference sample.

[0221] In one embodiment, the target sample includes target time features, target object features, and target order features, and the reference sample includes reference time features, reference object features, and reference order features. The prediction module 1204 is further configured to input the target time features, target object features, and target order features into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree of the target sample at the target time corresponding to the target time features; and to input the reference time features, reference object features, and reference order features into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree of the reference sample at the reference time corresponding to the reference time features.

[0222] In one embodiment, the sample acquisition module 1202 includes:

[0223] The sample pair determination unit is used to acquire the target sample set and the corresponding target sample push interaction result set, as well as the reference sample set and the corresponding reference sample push interaction result set; match each target sample in the target sample set with each target sample in the reference sample set to obtain each training sample pair; determine the push interaction result pair corresponding to each training sample pair based on the target sample push interaction result set and the reference sample push interaction result set; compare the interaction results of the target sample push interaction result and the reference sample push interaction result in the push interaction result pair to obtain historical interaction differences, and determine the sample pair interaction difference label corresponding to each training sample pair based on the historical interaction differences; and obtain the training sample pair and the corresponding sample pair interaction difference label from each training sample pair and the corresponding sample pair interaction difference label.

[0224] In one embodiment, the sample pair determination unit is further configured to: acquire each initial push object, and determine each target push object and each reference push object from each initial push object; acquire the message to be pushed corresponding to each target push object, and push the message to be pushed corresponding to each target push object to the terminal corresponding to each target push object in the target order; acquire the target interaction data returned by the terminal corresponding to each target push object, and obtain the target sample set and the corresponding target sample push interaction result set based on each target push object and the target interaction data; acquire the message to be pushed corresponding to each reference push object, and push the message to be pushed corresponding to each reference push object to the terminal corresponding to each reference push object in the reference order; acquire the reference interaction data returned by the terminal corresponding to each reference push object, and obtain the reference sample set and the corresponding reference sample push interaction result set based on each reference push object and the reference interaction data.

[0225] In one embodiment, the sample pair determination unit is further configured to determine each initial target push object and each initial reference push object from each initial push object; perform a homogeneity test based on each initial target push object and each initial reference push object; when the homogeneity test passes, each initial target push object is used as each target push object, and each initial reference push object is used as each reference push object.

[0226] In one embodiment, the sample pair determination unit is further configured to acquire the initial target object features corresponding to each initial target push object and the initial reference object features corresponding to each initial reference push object; calculate the standard mean difference based on the initial target object features and the initial reference object features to obtain the target value; when the target value does not exceed a preset threshold, determine that the homogeneity test of each initial target push object and each initial reference push object has passed; when the target value exceeds the preset threshold, determine that the homogeneity test of each initial target push object and each initial reference push object has failed.

[0227] In one embodiment, the sample pair determination unit is further configured to acquire each candidate interaction feature, use each candidate interaction feature as each result feature, use the target order corresponding to each target push object as the cause feature, and perform causal discovery based on each result feature and cause feature to obtain a causal relationship graph, and determine the push interaction result feature from each candidate interaction feature based on the causal relationship graph.

[0228] The sample pair determination unit is also used to determine the feature value corresponding to the push interaction result feature from the target interaction data, and use the feature value as the target sample to push the interaction result.

[0229] In one embodiment, such as Figure 13 As shown, a message push device 1300 is provided, including: a request acquisition module 1302, a difference degree acquisition module 1304, and a request response module 1306, wherein:

[0230] The request retrieval module 1302 is used to retrieve message push requests, which carry the object to be pushed.

[0231] The difference degree acquisition module 1304 is used to acquire the target interaction difference degree based on the object to be pushed. The target interaction difference degree is obtained by inputting the feature data corresponding to the object to be pushed into the target push difference prediction model to predict the interaction probability difference. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels, inputs the target sample and reference sample distribution in the training sample pairs into the initial push difference prediction model to predict the interaction probability difference, obtains the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample, and calculates the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample to obtain the training interaction difference. Based on the training interaction difference and the sample pair interaction difference labels, loss is calculated to obtain training loss information. Based on the training loss information, the initial push difference prediction model is trained. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages in the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages in the reference order.

[0232] The request response module 1306 is used to respond to the message push request when the target interaction difference meets the preset target order push conditions, obtain each message to be pushed corresponding to the push object and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the push object according to the target push order.

[0233] In one embodiment, the message push request includes at least two; the message push device 1300 further includes:

[0234] The segmentation module is used to acquire each message push request, obtain the degree of difference in target interaction based on the target objects to be pushed in each message push request, determine the target message push request and the non-target message push request from each message push request according to the magnitude of the degree of difference in target interaction, respond to the target message push request, acquire each target message to be pushed corresponding to the target message push request and the target push order corresponding to each target message to be pushed, and push each target message to be pushed to the terminal corresponding to the target message push request in the target push order; respond to the non-target message push request, acquire each non-target message to be pushed corresponding to the non-target message push request and the non-target push order corresponding to each non-target message to be pushed, and push each non-target message to be pushed to the terminal corresponding to the non-target message push request in the non-target push order.

[0235] In one embodiment, the message push device 1300 further includes:

[0236] The historical prediction module is used to acquire feature data corresponding to each historical object during historical idle time. The feature data includes the target time. The feature data corresponding to each historical object is input into the target push difference prediction model to predict the interaction probability difference. The interaction difference degree of each historical object at the target time is obtained. The interaction difference degree at the target time is used to characterize the target interaction difference degree at the target time. The interaction difference degree of each historical object and the corresponding target time is saved.

[0237] The difference degree acquisition module 1304 is also used to acquire each message push request at the target time; and to acquire the target interaction difference degree corresponding to the object to be pushed in each message push request based on the correlation between each historical object and the corresponding target time interaction difference degree.

[0238] The modules in the aforementioned push model training device and message push device 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, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0239] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores training sample data, sample pair interaction difference labels, messages to be pushed, and historical push messages, etc. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a push model training method or a message push method.

[0240] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 15 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a push model training method or a message push method. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0241] Those skilled in the art will understand that Figure 14 or Figure 15The 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.

[0242] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0243] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0244] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0245] 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, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions. Users can refuse or conveniently refuse push notifications.

[0246] 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.

[0247] 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.

[0248] 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 patent 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 method for training a push model, characterized in that, The method includes: Obtain training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when pushing historical messages in the target order, and the reference samples are obtained when pushing historical messages in the reference order. The target order refers to the message push order that requires server resources to determine, and the reference order refers to the message push order that can be determined directly without consuming server resources. The target sample is input into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the target sample. The reference sample is also input into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages according to the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages according to the reference order. The difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample is calculated to obtain the training interaction difference. Based on the training interaction difference and the sample, the loss of the interaction difference label is calculated to obtain the training loss information. The initial push difference prediction model is updated based on the training loss information, and the step of obtaining training sample pairs and corresponding sample pair interaction difference labels is iteratively executed until the training completion condition is met, thus obtaining the target push difference prediction model. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference. When the degree of target interaction difference exceeds a preset difference threshold, the message is pushed according to the target order; when the degree of target interaction difference does not exceed the preset difference threshold, the message is pushed according to the reference order.

2. The method according to claim 1, characterized in that, The target sample is input into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the target sample. Similarly, the reference sample is input into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the reference sample. This includes: The target sample is input into the initial push difference prediction model, and the target interaction probability of the target sample is predicted by the initial push difference prediction model to obtain the target interaction probability corresponding to the target sample. The reference interaction probability of the target sample is predicted by the initial push difference prediction model to obtain the reference interaction probability corresponding to the target sample. The difference between the target interaction probability corresponding to the target sample and the reference interaction probability corresponding to the target sample is calculated by the initial push difference prediction model to obtain the degree of interaction difference corresponding to the target sample. The reference sample is input into the initial push difference prediction model, and the target interaction probability of the reference sample is predicted by the initial push difference prediction model to obtain the target interaction probability corresponding to the reference sample. The reference interaction probability is predicted for the reference sample by the initial push difference prediction model, and the reference interaction probability corresponding to the reference sample is obtained. The difference between the target interaction probability corresponding to the reference sample and the reference interaction probability corresponding to the reference sample is calculated by the initial push difference prediction model to obtain the interaction difference degree corresponding to the reference sample.

3. The method according to claim 1, characterized in that, The target sample includes target time features, target object features, and target order features; the reference sample includes reference time features, reference object features, and reference order features. The step of inputting the target sample into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the target sample, and inputting the reference sample into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the reference sample, includes: The target time feature, the target object feature, and the target sequence feature are input into the initial push difference prediction model to predict the interaction probability difference, so as to obtain the degree of interaction difference of the target sample at the target time corresponding to the target time feature. The reference time feature, the reference object feature, and the reference order feature are input into the initial push difference prediction model to predict the interaction probability difference, so as to obtain the interaction difference degree of the reference sample at the reference time corresponding to the reference time feature.

4. The method according to claim 1, characterized in that, The step of obtaining training sample pairs and corresponding sample pair interaction difference labels includes: Obtain the target sample set and the corresponding target sample push interaction result set, as well as the reference sample set and the corresponding reference sample push interaction result set; Each target sample in the target sample set is matched with each reference sample in the reference sample set to obtain each training sample pair; Based on the target sample push interaction result set and the reference sample push interaction result set, determine the push interaction result pair corresponding to each training sample pair; Based on the push interaction results, the push interaction results of the target sample and the push interaction results of the reference sample are compared to obtain historical interaction differences. Based on the historical interaction differences, the corresponding sample pair interaction difference label of each training sample pair is determined. The training sample pairs and their corresponding interaction difference labels are obtained from each training sample pair and their corresponding sample pair interaction difference labels.

5. The method according to claim 4, characterized in that, The process of obtaining the target sample set and the corresponding target sample push interaction result set, as well as the reference sample set and the corresponding reference sample push interaction result set, includes: Obtain each initial push object, and determine each target push object and each reference push object from the initial push objects; Obtain the message to be pushed corresponding to each target push object, and push the message to be pushed corresponding to each target push object to the terminal corresponding to each target push object according to the target order; Obtain the target interaction data returned by the terminal corresponding to each target push object, and obtain the target sample set and the corresponding target sample push interaction result set based on each target push object and the target interaction data; Obtain the message to be pushed corresponding to each of the reference push objects, and push the message to be pushed corresponding to each of the reference push objects to the terminal corresponding to each of the reference push objects in the reference order; Obtain the reference interaction data returned by the terminal corresponding to each reference push object, and obtain the reference sample set and the corresponding reference sample push interaction result set based on each reference push object and the reference interaction data.

6. The method according to claim 5, characterized in that, The step of determining each target push object and each reference push object from each initial push object includes: Each initial target push object and each initial reference push object are determined from the aforementioned initial push objects; A homogeneity check is performed on each initial target push object and each initial reference push object. When the homogeneity check passes, each initial target push object is used as the target push object, and each initial reference push object is used as the reference push object.

7. The method according to claim 6, characterized in that, The homogeneity check based on each initial target push object and each initial reference push object includes: Obtain the initial target object features corresponding to each initial target push object and the initial reference object features corresponding to each initial reference push object; The target value is obtained by calculating the standard mean difference based on the initial target object features and the initial reference object features; When the target value does not exceed the preset threshold, it is determined that the homogeneity test of each initial target push object and each initial reference push object has passed; when the target value exceeds the preset threshold, it is determined that the homogeneity test of each initial target push object and each initial reference push object has failed.

8. The method according to claim 5, characterized in that, After determining each target push object and each reference push object from the initial push objects, the method further includes: Each candidate interaction feature is obtained, and these candidate interaction features are used as the result features. The order of the targets corresponding to each target push object is used as the causal feature, and causal discovery is performed based on the result features and the causal features to obtain a causal relationship graph. Based on the causal relationship graph, push interaction result features are determined from each candidate interaction feature; The step of obtaining the target sample set and the corresponding target sample push interaction result set based on each target push object and the target interaction data includes: The feature value corresponding to the push interaction result feature is determined from the target interaction data, and the feature value is used as the target sample push interaction result.

9. A message push method, characterized in that, The method includes: Obtain a message push request, wherein the message push request carries the object to be pushed; Based on the target object to be pushed, the corresponding target interaction difference degree is obtained. This target interaction difference degree is obtained by inputting the feature data corresponding to the target object into a target push difference prediction model to predict the interaction probability difference. The target push difference prediction model obtains training sample pairs and corresponding sample pair interaction difference labels. The distributions of target samples and reference samples in the training sample pairs are input into an initial push difference prediction model to predict the interaction probability difference, obtaining the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample. The difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample is calculated to obtain the training interaction difference. Based on the training interaction difference and the sample pair interaction difference labels, a loss calculation is performed to obtain the training... The training loss information is used to train the initial push difference prediction model. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages according to the target order. The reference interaction probability is used to characterize the interaction probability obtained when pushing messages according to the reference order. The target order refers to the message push order that requires server resources to determine. The reference order refers to the message push order that can be determined directly without consuming server resources. When the target interaction difference degree exceeds a preset difference degree threshold, the message push is performed according to the target order. When the target interaction difference degree does not exceed the preset difference degree threshold, the message push is performed according to the reference order. When the degree of difference in the target interaction meets the preset target order push condition, respond to the message push request, obtain each message to be pushed corresponding to the object to be pushed and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the object to be pushed according to the target push order.

10. The method according to claim 9, characterized in that, The message push request includes at least two; the method further includes: Obtain each message push request, and based on the push object in each message push request, obtain the corresponding target interaction difference degree to be divided; Based on the magnitude of the difference in the target interaction to be divided, target message push requests and non-target message push requests are determined from each message push request; In response to the target message push request, obtain each target message to be pushed corresponding to the target message push request and the target push order corresponding to each target message to be pushed, and push each target message to be pushed to the terminal corresponding to the target message push request according to the target push order; In response to the non-target message push request, obtain each non-target message to be pushed corresponding to the non-target message push request and the non-target push order corresponding to each non-target message to be pushed, and push each non-target message to be pushed to the terminal corresponding to the non-target message push request according to the non-target push order.

11. The method according to claim 10, characterized in that, Before obtaining each message push request, the following is also included: During historical idle periods, feature data corresponding to each historical object is acquired, and the feature data includes the target time. The feature data corresponding to each historical object are respectively input into the target push difference prediction model to predict the interaction probability difference, so as to obtain the target time interaction difference degree corresponding to each historical object. The target time interaction difference degree is used to characterize the target interaction difference degree at the target time. Save the degree of interaction difference between each historical object and its corresponding target time. The step of obtaining each message push request and obtaining the corresponding target interaction difference degree based on the push object in each message push request includes: The message push requests are acquired at the target time. The degree of difference in target interaction between each historical object and its corresponding target time is obtained based on the correlation between the interaction differences between each historical object and its corresponding target time.

12. A push model training device, characterized in that, The device includes: The sample acquisition module is used to acquire training sample pairs and corresponding sample pair interaction difference labels. The training sample pairs include target samples and reference samples. The target samples are obtained when historical messages are pushed in the target order, and the reference samples are obtained when historical messages are pushed in the reference order. The target order refers to the message push order that requires server resources to determine, and the reference order refers to the message push order that can be determined directly without consuming server resources. The prediction module is used to input the target sample into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the target sample; and to input the reference sample into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree corresponding to the reference sample. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages according to the target order; the reference interaction probability is used to characterize the interaction probability obtained when pushing messages according to the reference order. The loss calculation module is used to calculate the difference between the degree of interaction difference corresponding to the target sample and the degree of interaction difference corresponding to the reference sample, to obtain the training interaction difference, and to perform loss calculation on the interaction difference label based on the training interaction difference and the sample to obtain training loss information. An iterative module is used to update the initial push difference prediction model based on the training loss information, and iteratively execute the step of obtaining training sample pairs and corresponding sample pair interaction difference labels until the training completion condition is met, thus obtaining the target push difference prediction model. The target push difference prediction model is used to predict the degree of target interaction difference when pushing messages according to the target order versus pushing messages according to the reference order, and push messages based on the degree of target interaction difference. When the degree of target interaction difference exceeds a preset difference threshold, the message is pushed according to the target order; when the degree of target interaction difference does not exceed the preset difference threshold, the message is pushed according to the reference order.

13. The apparatus according to claim 12, characterized in that, The prediction module is also used to input the target sample into the initial push difference prediction model, and use the initial push difference prediction model to predict the target interaction probability of the target sample to obtain the target interaction probability corresponding to the target sample. The reference interaction probability of the target sample is predicted by the initial push difference prediction model to obtain the reference interaction probability corresponding to the target sample. The difference between the target interaction probability corresponding to the target sample and the reference interaction probability corresponding to the target sample is calculated by the initial push difference prediction model to obtain the interaction difference degree corresponding to the target sample; the reference sample is input into the initial push difference prediction model, and the target interaction probability of the reference sample is predicted by the initial push difference prediction model to obtain the target interaction probability corresponding to the reference sample. The reference interaction probability is predicted for the reference sample by the initial push difference prediction model, and the reference interaction probability corresponding to the reference sample is obtained. The difference between the target interaction probability corresponding to the reference sample and the reference interaction probability corresponding to the reference sample is calculated by the initial push difference prediction model to obtain the interaction difference degree corresponding to the reference sample.

14. The apparatus according to claim 12, characterized in that, The target sample includes target time features, target object features, and target order features; the reference sample includes reference time features, reference object features, and reference order features. The prediction module is further configured to input the target time feature, the target object feature, and the target order feature into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree of the target sample at the target time corresponding to the target time feature; and to input the reference time feature, the reference object feature, and the reference order feature into the initial push difference prediction model to predict the interaction probability difference, thereby obtaining the interaction difference degree of the reference sample at the reference time corresponding to the reference time feature.

15. The apparatus according to claim 12, characterized in that, The sample acquisition module includes: A sample pair determination unit is used to acquire a target sample set and a corresponding target sample push interaction result set, as well as a reference sample set and a corresponding reference sample push interaction result set; match each target sample in the target sample set with each reference target sample in the reference sample set to obtain each training sample pair; determine the push interaction result pair corresponding to each training sample pair based on the target sample push interaction result set and the reference sample push interaction result set; compare the interaction results of the target sample push interaction result and the reference sample push interaction result in the push interaction result pair to obtain historical interaction differences; determine the sample pair interaction difference label corresponding to each training sample pair based on the historical interaction differences; and obtain the training sample pair and the corresponding sample pair interaction difference label from the training sample pair and the sample pair interaction difference label corresponding to the training sample pair.

16. The apparatus according to claim 15, characterized in that, The sample pair determination unit is further configured to: acquire each initial push object; determine each target push object and each reference push object from the initial push objects; acquire the messages to be pushed corresponding to each target push object; push the messages to be pushed corresponding to each target push object to the terminals corresponding to each target push object according to the target order; acquire the target interaction data returned by the terminals corresponding to each target push object; obtain the target sample set and the corresponding target sample push interaction result set based on the target push objects and the target interaction data; acquire the messages to be pushed corresponding to each reference push object; push the messages to be pushed corresponding to each reference push object to the terminals corresponding to each reference push object according to the reference order; acquire the reference interaction data returned by the terminals corresponding to each reference push object; and obtain the reference sample set and the corresponding reference sample push interaction result set based on the reference push objects and the reference interaction data.

17. The apparatus according to claim 16, characterized in that, The sample pair determination unit is further configured to determine each initial target push object and each initial reference push object from each initial push object; perform a homogeneity test based on each initial target push object and each initial reference push object; when the homogeneity test passes, use each initial target push object as each target push object and use each initial reference push object as each reference push object.

18. The apparatus according to claim 17, characterized in that, The sample pair determination unit is further configured to acquire the initial target object features corresponding to each initial target push object and the initial reference object features corresponding to each initial reference push object; calculate the standard mean difference based on the initial target object features and the initial reference object features to obtain a target value; when the target value does not exceed a preset threshold, determine that the homogeneity test of each initial target push object and each initial reference push object has passed; when the target value exceeds the preset threshold, determine that the homogeneity test of each initial target push object and each initial reference push object has failed.

19. The apparatus according to claim 16, characterized in that, The sample pair determination unit is also used to acquire each candidate interaction feature, take each candidate interaction feature as each result feature, take the target order corresponding to each target push object as the cause feature, and perform causal discovery based on each result feature and the cause feature to obtain a causal relationship graph, and determine the push interaction result feature from each candidate interaction feature based on the causal relationship graph. The sample pair determination unit is further configured to determine the feature value corresponding to the push interaction result feature from the target interaction data, and use the feature value as the target sample push interaction result.

20. A message push device, characterized in that, The device includes: The request acquisition module is used to acquire message push requests, wherein the message push requests carry objects to be pushed; The difference degree acquisition module is used to acquire the target interaction difference degree based on the object to be pushed. The target interaction difference degree is obtained by inputting the feature data corresponding to the object to be pushed into a target push difference prediction model to predict the interaction probability difference. The target push difference prediction model is obtained by acquiring training sample pairs and corresponding sample pair interaction difference labels, inputting the distribution of target samples and reference samples in the training sample pairs into an initial push difference prediction model to predict the interaction probability difference, obtaining the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample, and calculating the difference between the interaction difference degree corresponding to the target sample and the interaction difference degree corresponding to the reference sample to obtain the training interaction difference. A loss is calculated based on the training interaction difference and the sample pair interaction difference labels. The training loss information is calculated, and the initial push difference prediction model is trained based on the training loss information. The interaction difference degree is used to characterize the difference between the target interaction probability and the reference interaction probability. The target interaction probability is used to characterize the interaction probability obtained when pushing messages according to the target order. The reference interaction probability is used to characterize the interaction probability obtained when pushing messages according to the reference order. The target order refers to the message push order that requires server resources to determine, and the reference order refers to the message push order that can be determined directly without consuming server resources. When the target interaction difference degree exceeds a preset difference degree threshold, the message push is performed according to the target order. When the target interaction difference degree does not exceed the preset difference degree threshold, the message push is performed according to the reference order. The request-response module is used to respond to the message push request when the target interaction difference meets the preset target order push conditions, obtain each message to be pushed corresponding to the target object and the target push order corresponding to each message to be pushed, and push each message to be pushed to the terminal corresponding to the target object according to the target push order.

21. The apparatus according to claim 20, characterized in that, The message push request includes at least two; the device also includes: The segmentation module is used to acquire each message push request, obtain the corresponding target interaction difference degree based on the push objects in each message push request, determine target message push requests and non-target message push requests from the message push requests according to the magnitude of the target interaction difference degree, respond to the target message push request, acquire each target push message corresponding to the target message push request and the target push order corresponding to each target push message, and push each target push message to the terminal corresponding to the target message push request according to the target push order; respond to the non-target message push request, acquire each non-target push message corresponding to the non-target message push request and the non-target push order corresponding to each non-target push message, and push each non-target push message to the terminal corresponding to the non-target message push request according to the non-target push order.

22. The apparatus according to claim 21, characterized in that, The device further includes: The historical prediction module is used to acquire feature data corresponding to each historical object during historical idle times, the feature data including the target time; input the feature data corresponding to each historical object into the target push difference prediction model to predict the interaction probability difference, and obtain the interaction difference degree of each historical object at the target time, the interaction difference degree at the target time is used to characterize the target interaction difference degree at the target time; and save the interaction difference degree of each historical object and the corresponding target time. The difference degree acquisition module is also used to acquire each message push request at the target time; and to acquire the target interaction difference degree corresponding to the object to be pushed in each message push request based on the correlation between each historical object and the corresponding target time interaction difference degree.

23. 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 11.

24. 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 11.

25. 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 11.

Images