Push processing method, related device and medium

Abstract

The invention provides a push processing method, a related device and a medium. The method comprises the following steps: acquiring a plurality of pushing basic features; inputting the plurality of push basic features into a multi-task prediction model to obtain first probabilities corresponding to the plurality of tasks; based on the first probabilities corresponding to the tasks, determining fusion parameters through a fusion parameter determination model; based on the first probabilities corresponding to the multiple tasks and the fusion parameters, determining a post-fusion probability; and pushing to-be-pushed content to the target object based on the fused probability, the fusion parameter determination model being trained based on a distribution position of a first sample probability, corresponding to the task, of each target sample in the target sample set in the first sample probability of each sample corresponding to the task, and pushing the to-be-pushed content to the target object based on the fusion parameter determination model. The first sample probability is obtained by inputting the sample pushing basic features of the target sample into the multi-task prediction model. According to the embodiment of the invention, the accuracy of push processing is improved. The embodiment of the invention can be applied to scenes such as content recommendation and data screening.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and in particular to a push processing method, related devices, and media. Background Art

[0002] Currently, in Internet content push, the characteristics of the target object (such as the person to be pushed) and the characteristics of the content to be pushed (such as the short video to be pushed) are usually input into a multi-task prediction model to obtain the probability of the target object completing multiple tasks (such as clicking in, liking, commenting, etc.) on the content to be pushed. These probabilities are used to determine whether to push the content to be pushed to the target object. Since the importance of each task to the platform or object is different, these probabilities of different tasks in the existing technology are often fused through a multi-task fusion module to reflect the importance of each task. Finally, the decision on whether to push is made based on the fusion result.

[0003] Current multi-task fusion modules generally employ formula fusion or model fusion. Regardless of the fusion method, multi-task fusion does not retrain the multi-task prediction model. Instead, corresponding fusion parameters (weights) are assigned to different tasks. Once assigned, these fusion parameters remain unchanged. However, due to the continuous shift in data distribution, the output probabilities of the multi-task prediction model may have undergone significant overall changes, and the previous fusion parameters may no longer be applicable. Furthermore, the weighting of the benefits of each task in the push of each sample may vary. For example, a multi-task prediction model may output a probability of 0.6 for Task A and a probability of 0.6 for Task B. However, for Task A, the predicted probabilities may be generally high, and 0.6 is not considered high. For Task B, a probability of 0.6 is quite high. Although the predicted probabilities are the same, the fusion results may also be the same. However, a probability of 0.6 should be given greater weight for each task. This is impossible with existing technologies that use relatively fixed fusion parameters. Consequently, existing technologies lack fine-grained fusion, resulting in poor push processing accuracy. Summary of the Invention

[0004] The embodiments of the present disclosure provide a push processing method, related devices, and media, which can improve the accuracy of push processing.

[0005] According to one aspect of the present disclosure, a push processing method is provided, comprising:

[0006] Acquire a plurality of basic push features, wherein the plurality of basic push features include target object features and to-be-pushed content features;

[0007] Inputting the plurality of push basic features into a multi-task prediction model to obtain first probabilities corresponding to the plurality of tasks for pushing the content to be pushed to the target object;

[0008] determining a fusion parameter by a fusion parameter determination model based on the first probabilities corresponding to the plurality of tasks;

[0009] Determining a fused probability of pushing the to-be-pushed content to the target object based on the first probabilities corresponding to the plurality of tasks and the fusion parameter;

[0010] Based on the fused probability, the content to be pushed is pushed to the target object, wherein the fusion parameter determination model is trained based on the distribution position of the first sample probability corresponding to the task for each target sample in the target sample set in the first sample probability of each sample corresponding to the task, and the first sample probability is obtained by inputting the sample push basic features of the target sample into the multi-task prediction model.

[0011] According to one aspect of the present disclosure, a push processing device is provided, comprising:

[0012] An acquisition unit, configured to acquire a plurality of basic push features, wherein the plurality of basic push features include target object features and features of content to be pushed;

[0013] A prediction unit, configured to input the plurality of push basic features into a multi-task prediction model to obtain a first probability of pushing the content to be pushed to the target object corresponding to the plurality of tasks;

[0014] a first determining unit, configured to determine a fusion parameter by a fusion parameter determination model based on the first probabilities corresponding to the plurality of tasks;

[0015] a second determining unit, configured to determine a fused probability of pushing the content to be pushed to the target object based on the first probabilities corresponding to the plurality of tasks and the fusion parameter;

[0016] A pushing unit, configured to push the content to be pushed to the target object based on the fused probability;

[0017] Among them, the fusion parameter determination model is trained based on the distribution position of the first sample probability corresponding to the task of each target sample in the target sample set in the first sample probability of each sample corresponding to the task, and the first sample probability is obtained by inputting the sample push basic features of the target sample into the multi-task prediction model.

[0018] Optionally, the fusion parameter determination model is trained in the following manner:

[0019] Acquire the target sample set, where the target sample set includes a plurality of target samples, each of the target samples has a plurality of sample push basic features, and the plurality of sample push basic features include a sample object feature and a sample content feature;

[0020] Inputting the plurality of sample push basic features into the multi-task prediction model to obtain second probabilities corresponding to the plurality of tasks for pushing the sample content to the sample object;

[0021] Inputting the second probabilities corresponding to the plurality of tasks into the fusion parameter determination model to obtain sample fusion parameters, and determining a sample fusion probability of pushing the sample content to the sample object based on the second probabilities corresponding to the plurality of tasks and the fusion parameters;

[0022] Obtaining a distribution position of a first sample probability of the target sample corresponding to the task in the first sample probabilities of each sample corresponding to the task;

[0023] Based on the sample fusion probability and the distribution position, a first sample loss function is calculated, and the fusion parameter determination model is trained based on the first sample loss function.

[0024] Optionally, the calculating a first sample loss function based on the sample fusion probability and the distribution position includes:

[0025] Performing a monotonic transformation on the sample fusion probability to obtain the sample monotonic transformation probability corresponding to the task;

[0026] Determining a first sample sub-loss function corresponding to the task based on the monotonically transformed probability of the sample corresponding to the task;

[0027] Determine the first sample loss function based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task.

[0028] Optionally, performing a monotonic transformation on the sample fusion probability to obtain the sample monotonic transformation probability corresponding to the task includes:

[0029] Determining a first fusion score based on the sample fusion probability;

[0030] Performing a monotonic transformation on the first fusion score to obtain a second fusion score corresponding to the task;

[0031] Based on the second fusion score corresponding to the task, a probability of the sample corresponding to the task after monotonous transformation is determined.

[0032] Optionally, determining a first fusion score based on the sample fusion probability includes:

[0033] Determine a first difference between 1 and the probability of the sample after fusion;

[0034] The logarithm of a first quotient of the sample fusion probability and the first difference is determined as the first fusion score.

[0035] Optionally, performing a monotonic transformation on the first fused score to obtain a second fused score corresponding to the task includes:

[0036] Obtaining a monotonic transformation coefficient and a monotonic transformation offset corresponding to the task;

[0037] Based on the monotonic transformation coefficient and monotonic transformation offset corresponding to the task, the first fusion score is monotonically transformed to obtain the second fusion score corresponding to the task, wherein the monotonic transformation coefficient and monotonic transformation offset corresponding to each task are jointly trained with the fusion parameter determination model.

[0038] Optionally, performing a monotonic transformation on the first fused score based on a monotonic transformation coefficient and a monotonic transformation offset corresponding to the task to obtain the second fused score corresponding to the task includes:

[0039] multiplying the square of the monotonic transformation coefficient by the first fusion score to obtain a first product;

[0040] The second fused score is determined based on the first product and the monotonic transformation offset.

[0041] Optionally, determining the probability of the sample corresponding to the task after monotonic transformation based on the second fusion score corresponding to the task includes:

[0042] determining a first power value with a first constant as a base and a negation of the second fused component as an exponent;

[0043] determining a first sum of the first power value and a second constant;

[0044] The reciprocal of the first sum is used as the probability of the sample corresponding to the task after monotonically transformed.

[0045] Optionally, determining a first sample sub-loss function corresponding to the task based on the probability of the sample corresponding to the task after monotonically transformed includes:

[0046] Determine a second product of a first logarithm of a probability of the sample corresponding to the task after monotonically transformed and the second probability corresponding to the task;

[0047] Determine a first complement of the second probability and a second complement of the probability of the sample after monotonically transformed, wherein the first complement is the difference between 1 and the second probability, and the second complement is the difference between 1 and the probability of the sample after monotonically transformed;

[0048] determining a third product of a second logarithm of the second complement and the first complement;

[0049] Based on the second product and the third product, a first sample sub-loss function corresponding to the task is determined.

[0050] Optionally, determining the first sample loss function based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task includes:

[0051] Determine a fourth product of the first sample sub-loss function corresponding to the task and the distribution position corresponding to the task;

[0052] The fourth products of the plurality of tasks are added together to obtain the first sample loss function.

[0053] Optionally, determining a fourth product of a first sample sub-loss function corresponding to the task and a distribution position corresponding to the task includes:

[0054] Get the prior weight corresponding to the task;

[0055] Determine the fourth product of the first sample sub-loss function corresponding to the task, the distribution position corresponding to the task, and the prior weight corresponding to the task.

[0056] Optionally, the training the fusion parameter determination model based on the first sample loss function includes:

[0057] Determining a second sample loss function based on the second probabilities corresponding to the plurality of tasks;

[0058] Determining a total loss function based on the first sample loss function and the second sample loss function;

[0059] Based on the total loss function, the fusion parameter determination model is trained.

[0060] Optionally, determining a total loss function based on the first sample loss function and the second sample loss function includes:

[0061] Determining a total sample loss function of the target sample based on the first sample loss function and the second sample loss function;

[0062] The total loss function is determined based on the total sample loss function of each target sample in the target sample set.

[0063] Optionally, the distribution position is a multiple of a standard deviation of a first sample probability of the target sample corresponding to the task, compared to an average of the first sample probabilities of the samples corresponding to the task;

[0064] The obtaining of the distribution position of the first sample probability of the target sample corresponding to the task in the first sample probabilities of each sample corresponding to the task includes:

[0065] Acquire multiple samples corresponding to the task;

[0066] Determining an average value and a standard deviation of the first sample probabilities of the plurality of samples corresponding to the task;

[0067] Determine a second difference between the first sample probability of the target sample corresponding to the task and the average value;

[0068] A second quotient of the second difference and the standard deviation is determined as the distribution position.

[0069] Optionally, the fusion parameter includes a first fusion parameter and a second fusion parameter; and the second determining unit is specifically configured to:

[0070] determining a second sum of the first probability and the first fusion parameter corresponding to the task;

[0071] determining a probability contribution value corresponding to the task based on the second sum corresponding to the task and the second fusion parameter;

[0072] The probability contribution values ​​corresponding to the multiple tasks are multiplied to obtain the fused probability.

[0073] Optionally, determining the probability contribution value corresponding to the task based on the second sum corresponding to the task and the second fusion parameter includes:

[0074] Determine the third sum of 1 and the maximum value of the first fusion parameter;

[0075] determining a third quotient of the second sum and the third sum;

[0076] A power value having the third quotient as a base and the second fusion parameter as an exponent is determined as the probability contribution value corresponding to the task.

[0077] Optionally, the first fusion parameter is between 1 and a maximum value of the first fusion parameter, and the second fusion parameter is between 1 and a maximum value of the second fusion parameter; the maximum value of the first fusion parameter and the maximum value of the second fusion parameter are determined in the following manner:

[0078] Obtaining a first basic scalar corresponding to the first fusion parameter, a second basic scalar corresponding to the second fusion parameter, and a third constant greater than 1;

[0079] performing exponential normalization on the first basic scalar to obtain a first exponential normalized value;

[0080] performing exponential normalization on the second basic scalar to obtain a second exponential normalized value;

[0081] determining a maximum value of the first fusion parameter based on the third constant and the first exponential normalized value;

[0082] Based on the third constant and the second exponential normalization value, a second fusion parameter maximum value is determined, wherein the first basic scalar and the second basic scalar are jointly trained with the fusion parameter determination model.

[0083] Optionally, the fusion parameter includes a first fusion parameter and a second fusion parameter, the first fusion parameter is between 1 and a maximum value of the first fusion parameter, the second fusion parameter is between 1 and a maximum value of the second fusion parameter, and the fusion parameter determination model includes a first importance weight vector generation model and a second importance weight vector generation model;

[0084] The first determining unit is specifically configured to:

[0085] Inputting the first probabilities corresponding to the plurality of tasks into the first importance weight vector generation model to obtain a first importance weight vector;

[0086] Obtaining a first importance weight corresponding to the task from the first importance weight vector, and determining the first fusion parameter based on the maximum value of the first fusion parameter and the first importance weight corresponding to the task;

[0087] Inputting the first probabilities corresponding to the plurality of tasks into a second importance weight vector generation model to obtain a second importance weight vector;

[0088] The second importance weight corresponding to the task is obtained from the second importance weight vector, and the second fusion parameter is determined based on the second fusion parameter maximum value and the second importance weight corresponding to the task.

[0089] Optionally, the first importance weight vector generation model includes a first compression model and a first stretching model;

[0090] The first determining unit is further configured to:

[0091] generating a first probability vector based on the first probabilities corresponding to the plurality of tasks;

[0092] Inputting the first probability vector into the first compression model to obtain a first compression vector, where the dimension of the first compression vector is smaller than the number of the tasks;

[0093] The first compression vector is input into the first stretching model to obtain the first importance weight vector, where the dimension of the first importance weight vector is equal to the number of the tasks.

[0094] Optionally, the first compression model includes a first matrix and a nonlinear activation function; and the first determining unit is further configured to:

[0095] determining a first product vector based on the first matrix and the first probability vector;

[0096] Applying the nonlinear activation function to the first product vector to obtain the first compressed vector.

[0097] Optionally, the first stretching model includes a second matrix and an exponential normalization function; and the first determining unit is further configured to:

[0098] determining a second product vector based on the second matrix and the first compressed vector;

[0099] Apply the exponential normalization function to the second product vector to obtain the first importance weight vector.

[0100] Optionally, the second importance weight vector generation model includes a second compression model and a second stretching model; and the first determining unit is specifically configured to:

[0101] generating a first probability vector based on the first probabilities corresponding to the plurality of tasks;

[0102] Inputting the first probability vector into the second compression model to obtain a second compression vector, where the dimension of the second compression vector is smaller than the number of the tasks;

[0103] The second compression vector is input into the second stretching model to obtain the second importance weight vector, where the dimension of the second importance vector is equal to the number of the tasks.

[0104] Optionally, the second compression model includes a third matrix and a nonlinear activation function; and the first determining unit is further configured to:

[0105] determining a third product vector based on the third matrix and the first probability vector;

[0106] Applying the nonlinear activation function to the third product vector to obtain the second compressed vector.

[0107] Optionally, the second stretching model includes a fourth matrix and an exponential normalization function; and the first determining unit is further configured to:

[0108] determining a fourth product vector based on the fourth matrix and the second compressed vector;

[0109] Apply the exponential normalization function to the fourth product vector to obtain the second importance weight vector.

[0110] According to one aspect of the present disclosure, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the push processing method described above when executing the computer program.

[0111] According to one aspect of the present disclosure, a computer-readable storage medium is provided, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the push processing method described above is implemented.

[0112] According to one aspect of the present disclosure, a computer program product is provided. The computer program product includes a computer program. The computer program is read and executed by a processor of a computer device, so that the computer device executes the push processing method described above.

[0113] In the disclosed embodiment, multiple push basic features are input into a multi-task prediction model to obtain first probabilities of pushing content to a target object, corresponding to multiple tasks. Since the importance of each task may vary, fusion is required. When determining the fusion parameters for each task, rather than setting fixed fusion parameters for each task, the fusion parameters are determined using a fusion parameter determination model. The fusion parameter determination model is trained based on the distribution of the first sample probabilities corresponding to the task, as predicted by the multi-task prediction model, for target samples. Thus, for a medium first sample probability, if the first sample probability is generally low and varies relatively slowly across all samples corresponding to the task, it has a significant advantage across all samples for that task. The fusion parameter determination model will learn to assign a larger fusion parameter to that task for that target sample. For the same medium first sample probability, if the first sample probability is generally high and fluctuates significantly across all samples corresponding to the task, it will not have a significant advantage across all samples for that task. The fusion parameter determination model will learn to assign a smaller fusion parameter to that task for that target sample. The fusion parameter determination model trained in this way can accurately determine the fusion parameters of each task for each specific sample, thereby reducing the impact of large overall changes in the first probability distribution output by the multi-task prediction model due to continuous changes in data distribution, and improving the accuracy of the final content push.

[0114] Other features and advantages of the present disclosure will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present disclosure. The purposes and other advantages of the present disclosure can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] The accompanying drawings are used to provide a further understanding of the technical solution of the present disclosure and constitute a part of the specification. Together with the embodiments of the present disclosure, they are used to explain the technical solution of the present disclosure and do not constitute a limitation to the technical solution of the present disclosure.

[0116] Figure 1 is an architecture diagram of a system to which the push processing method according to an embodiment of the present disclosure is applied;

[0117] Figure 2A-2C This is a schematic diagram of an interface in a scenario where an embodiment of the present disclosure is applied to an instant messaging application to view a video content push message;

[0118] Figure 3 is a flowchart of a push processing method according to an embodiment of the present disclosure;

[0119] Figure 4It is an overall structural diagram of a push processing method according to an embodiment of the present disclosure.

[0120] Figure 5 is a schematic diagram of specific components of push basic features according to an embodiment of the present disclosure;

[0121] Figure 6 is an overall structural diagram of a push processing method according to another embodiment of the present disclosure;

[0122] Figure 7 yes Figure 3 Flowchart of step 330 in FIG.

[0123] Figure 8 is a schematic diagram of determining a first fusion parameter and a second fusion parameter using a fusion parameter determination model according to an embodiment of the present disclosure;

[0124] Figure 9 yes Figure 7 Flowchart of step 710 in FIG.

[0125] Figure 10 is a schematic diagram of determining a first importance weight vector using a first compression model and a first stretching model according to an embodiment of the present disclosure;

[0126] Figure 11 yes Figure 7 Flowchart of step 730 in FIG.

[0127] Figure 12 is a schematic diagram of determining a second importance weight vector using a second compression model and a second stretching model according to an embodiment of the present disclosure;

[0128] Figure 13 yes Figure 3 Flowchart of step 340;

[0129] Figure 14 yes Figure 13 Flowchart of step 1320;

[0130] Figure 15 is a flowchart of determining a maximum value of a first fusion parameter and a maximum value of a second fusion parameter according to one embodiment of the present disclosure;

[0131] Figure 16 yes Figure 3 Schematic diagram of a specific implementation of content push based on the fused probability in step 350;

[0132] Figure 17 is a flowchart of a training fusion parameter determination model according to one embodiment of the present disclosure;

[0133] Figure 18is a schematic diagram of a specific implementation of determining a distribution location according to an embodiment of the present disclosure;

[0134] Figure 19 yes Figure 17 Flowchart of step 1750 in FIG.

[0135] Figure 20 yes Figure 19 Flowchart of step 1910 in FIG.

[0136] Figure 21 yes Figure 20 Flowchart of step 2020 in FIG.

[0137] Figure 22 yes Figure 19 Flowchart of step 1930 in FIG.

[0138] Figure 23 is a flowchart of a model for determining fusion parameters based on a first sample loss function training according to an embodiment of the present disclosure;

[0139] Figure 24 is a schematic diagram of an overall implementation of a push processing method according to an embodiment of the present disclosure;

[0140] Figure 25 is a module diagram of a push processing device according to an embodiment of the present disclosure;

[0141] Figure 26 According to the embodiment of the present disclosure Figure 3 The terminal structure diagram of the push processing method shown;

[0142] Figure 27 According to the embodiment of the present disclosure Figure 3 The server structure diagram of the push processing method shown. DETAILED DESCRIPTION

[0143] In order to make the purpose, technical solutions and advantages of the present disclosure more clearly understood, the present disclosure is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present disclosure and are not intended to limit the present disclosure.

[0144] Before further explaining the embodiments of the present disclosure in detail, the nouns and terms involved in the embodiments of the present disclosure are explained. The nouns and terms involved in the embodiments of the present disclosure are subject to the following interpretations:

[0145] A deep neural network is a multi-layer, unsupervised neural network. It uses the output features of the previous layer as input to the next layer for feature learning. Through layer-by-layer feature mapping, it maps the features of existing spatial samples to another feature space, thereby learning to better represent the existing input. Deep neural networks have multiple nonlinear feature transformations, enabling them to fit highly complex functions.

[0146] Gradient descent algorithm: A method for minimizing the loss function that calculates the gradient of the loss function at the current position and updates the parameters in the direction of gradient descent to find the optimal solution that minimizes the value of the loss function.

[0147] System architecture and scenario description of the application of the embodiments of the present disclosure

[0148] Figure 1 1 is a system architecture diagram of a push processing method according to an embodiment of the present disclosure, which includes: a target terminal 110 , the Internet 120 , a gateway 130 , and a push processing server 140 .

[0149] The target terminal 110 is the device used by the target to view the corresponding reminder message for the pushed content. It can be a desktop computer, laptop computer, PDA (Personal Digital Assistant), mobile phone, in-vehicle terminal, home theater terminal, dedicated terminal, and other various forms. Furthermore, it can be a single device or a collection of multiple devices. For example, multiple devices connected via a local area network and sharing a common display device can collectively constitute a terminal. The target terminal 110 can also communicate with the Internet 120 via wired or wireless means to exchange data.

[0150] Gateway 130, also known as a gateway or protocol converter, implements network interconnection at the transport layer and is a computer system or device that performs a conversion function. It acts as a translator between two systems using different communication protocols, data formats, or languages, or even completely different architectures. Gateway 130 also provides filtering and security functions. Messages sent by target terminals 110 to push processing servers 140 are sent to the corresponding push processing servers 140 via gateway 130. Messages sent by push processing servers 140 to target terminals 110 are also sent to the corresponding target terminals 110 via gateway 130.

[0151] The push processing server 140 is a computer system that provides content message push services to the target terminal 110. Compared to the target terminal 110, the push processing server 140 has higher requirements in terms of stability, security, and performance. The push processing server 140 can be a single high-performance computer within a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines). The push processing server 140 can also communicate with the internet 120 via wired or wireless means to exchange data.

[0152] The embodiments of the present disclosure can be applied in various scenarios, such as Figures 2A-2C The scenario shown is of viewing video content push messages in an instant short messaging application, etc.

[0153] like Figure 2A As shown, in the interface of the instant short message application in the target terminal 110, the instant short message application can be used to communicate with multiple contacts and can also be used to watch videos. After triggering the "Message" option in the function bar below, the short message list is displayed. The short message list contains message bars with multiple contacts. At 11:30 am, there is a content prompt ( Figure 2A The small black dot in the upper right corner of "Short Video" in the navigation bar reminds the recipient that new content is available. The recipient can watch the recommended content by triggering the "Short Video" option in the function bar below.

[0154] like Figure 2B As shown, at 12:00 am, the subject views the "short video" in the interface of the instant short message application and selects to trigger the "short video" in the function bar below.

[0155] like Figure 2CAs shown in the figure, after triggering the "Short Video" option, the instant messaging app opens the content push interface and plays the pushed content, which is a video V. The content push interface also displays the number of likes (67,000), comments (2,450), and favorites (1,422) the video received before 12:00 AM. The subject can also like, comment, and favorite the video in the interface. The video content pushed to the subject is determined by the push processing server. Generally, the push processing server calculates the estimated probability of the target subject for multiple tasks (such as likes, comments, and favorites) based on the characteristics of the target subject and the content to be pushed. Then, based on the fusion weights, the estimated probabilities of multiple tasks are fused to obtain the fused probability. Finally, based on the fused probability, the content to be pushed is pushed to the target subject. However, the fusion weights of each task vary for different subjects. For example, for subjects who use the instant messaging app less frequently, the fusion weights of click tasks and like tasks are higher, while for subjects who use the app more frequently, the fusion weights of favorite tasks and comment tasks are higher. For the same object, the fusion parameters of each task vary at different push stages. For example, in a stage where the frequency of use of instant messaging applications is relatively low, the fusion weights of click tasks and like tasks are higher, while in a stage where the frequency of use of instant messaging applications is relatively high, the fusion weights of collection tasks and comment tasks are lower. The disclosed embodiments can dynamically set the fusion weights for different objects, or for the same object at different push stages, thereby improving the personalization of push processing and the accuracy of push.

[0156] General description of the embodiments of the present disclosure

[0157] According to an embodiment of the present disclosure, a push processing method is provided.

[0158] The push processing method is the process of pushing the content that the target object is interested in to the object terminal 110 of the target object. The target object is the object that wants to view the content to be pushed. The embodiment of the present disclosure can be applied to the push of different types of content, and the content to be pushed can be pictures, videos, articles, applications and other types. If the target object opens the video recommendation software, the push processing method of the embodiment of the present disclosure is used to predict the video content that the target object may be interested in, and push the video content to the object terminal 110 of the target object through the video recommendation software. When determining the content to be pushed, the completion status of the target object in terms of different tasks for pushing the content must also be considered. For example, whether the target object will like the content to be pushed, whether the target object will collect the content to be pushed, etc. The order in which the content to be pushed is determined based on the target object's completion status in terms of the content to be pushed.

[0159] The push processing method of the embodiment of the present disclosure is executed in the push processing server 140. After the execution is completed, the obtained push content is transmitted to the target terminal 110 through the gateway 130 and the Internet 120, and the target terminal 110 displays the push content to the target object.

[0160] like Figure 3 As shown, according to one embodiment of the present disclosure, the push processing method includes:

[0161] Step 310: Acquire multiple basic push features, including target object features and features of the content to be pushed;

[0162] Step 320: Input the multiple push basic features into the multi-task prediction model to obtain a first probability of pushing the content to be pushed to the target object corresponding to the multiple tasks;

[0163] Step 330: Determine a fusion parameter using a fusion parameter determination model based on first probabilities corresponding to the multiple tasks;

[0164] Step 340: Determine a fused probability of pushing the content to be pushed to the target object based on the first probabilities corresponding to the multiple tasks and the fusion parameter;

[0165] Step 350: Push the content to be pushed to the target object based on the fused probability.

[0166] The above steps 310-350 are generally described below.

[0167] Figure 4 It is an overall structural diagram of the push processing method according to an embodiment of the present disclosure.

[0168] Figure 4 The lowest level in the push basic features is the basic push features in step 310. There are multiple basic push features, which are used to predict the target object's completion of different tasks on the content to be pushed. In one example, the basic push features include target object features of the target object and content features of the content to be pushed. In another example, the basic push features also include target scene features of the target scene in which the target object is located.

[0169] Target object features refer to features related to the target object. Target object features can be features of the target object itself. For example, target object features include features such as the target object's years of work experience and education level. Target scene features refer to features of the target scene in which the target object views the pushed content. For example, target scenes include scenes such as the workplace, home, or supermarket. Target scene features include features such as the target object at work, the target object at home, and the target object in the supermarket. Features of the content to be pushed are features of the content to be pushed. For example, features of the content to be pushed include features such as the content type and length of the content to be pushed.

[0170] Step 320 is performed on the basic features of push, and multi-task prediction is performed through the multi-task prediction model to obtain a first probability (not shown in the figure) corresponding to multiple tasks for pushing the content to be pushed to the target object. The first probability is the probability that the target object completes the task of pushing the content. The greater the probability that the target object completes the task of pushing the content, the greater the first probability, and thus the greater the probability of pushing the content to be pushed to the target object. The multi-task prediction model refers to a model that predicts the probability of the target object completing different tasks for the pushed content based on the basic features of push. For example, the multi-task prediction model is the deep neural network described above. A task refers to a matter related to the content to be pushed, which is used to weigh the target object's interest in the pushed content from multiple aspects. For example, the task is that the target object likes the pushed content, or the target object collects the content to be pushed, or the target object comments on the content to be pushed, etc. Each task corresponds to a first probability. Figure 4 The first probability P corresponding to task A is shown in A The first probability P corresponding to task B B .

[0171] Step 330 is performed for the first probabilities corresponding to the multiple tasks, and a fusion parameter is determined by a fusion parameter determination model (not shown in the figure). The fusion parameter indicates the weight of each task in predicting the interest of the target object in the content to be pushed. The fusion parameter determination model refers to a model that predicts the fusion parameter based on the probabilities corresponding to the multiple tasks. For example, the fusion parameter determination model is the deep neural network described above. In one example, Figure 4 As shown, the first probability P corresponding to task A A The first probability P corresponding to task B B Input the first splicing layer to obtain a spliced ​​vector, and then input the spliced ​​vector into the fusion parameter determination model to obtain the fusion parameters. The first splicing layer is a layer used to splice multiple input features. In the disclosed embodiment, the input feature is a first probability, and the first splicing layer is used to splice multiple first probabilities to obtain a spliced ​​vector.

[0172] In order to meet the requirements of the push processing method of the disclosed embodiment and to dynamically set fusion parameters during model application, a specific training method is required to train the fusion parameter determination model. In one embodiment, the fusion parameter determination model is trained based on the distribution position of the first sample probability corresponding to each target sample in the target sample set within the first sample probability of each sample corresponding to the task. The first sample probability is obtained by inputting the sample push basic features of the target sample into the multi-task prediction model.

[0173] A target sample set is a set of samples consisting of multiple samples. The samples are used to train the fusion parameter determination model. The multiple samples include a target sample. The samples include multiple basic sample push features. These basic sample push features are used to predict the sample subject's completion of different tasks for the sample's content to be pushed. The basic sample push features of the multiple target samples are input into the multi-task prediction model to obtain multiple first sample probabilities corresponding to each task. For each task, there are multiple first sample probabilities corresponding to each sample. Each sample can be used as a target sample in turn. The first sample probability of each target sample has a corresponding distribution position within the first sample probabilities of each sample. This distribution position reflects the weight or importance of each target sample's first sample probability within the multiple first sample probabilities corresponding to the task. Therefore, the fusion parameter determination model trained based on this distribution position can reflect the weight or importance of each task's first probability within the multiple first probabilities of the tasks. During the model usage phase, after the first probabilities corresponding to the multiple tasks are input into the fusion parameter determination model, the fusion parameters output by the fusion parameter determination model can reflect the weight of each task in predicting the target subject's interest in the content to be pushed.

[0174] Step 340 is performed on the first probabilities and fusion parameters corresponding to multiple tasks to obtain a fused probability of pushing the content to be pushed to the target object. In one example, for the target object, the first probability of completing Task A for content R1 to be pushed is 0.7, and the first probability of completing Task B for content R1 to be pushed is 0.3. Assuming that the fusion parameter for Task A is 0.6 and the fusion parameter for Task B is 0.4, the fused probability is 0.7*0.6+0.3*0.4=0.54. For the target object, the first probability of completing Task A for content R2 to be pushed is 0.8, and the first probability of completing Task B for content R2 to be pushed is 0.2. Assuming that the fusion parameter for Task A is 0.7 and the fusion parameter for Task B is 0.3, the fused probability is 0.8*0.7+0.2*0.3=0.62.

[0175] As can be seen from the above example, for the content to be pushed R1, the fusion parameter of task A is 0.6, while for the content to be pushed R2, the fusion parameter of task A is 0.7, and the two are not the same. This is because, when the push basic features include the content features to be pushed of the content features R1, the multi-task prediction model predicts that the first probability corresponding to the target object completing task A is 0.7, and when the push basic features include the content features to be pushed of the content to be pushed R2, the multi-task prediction model predicts that the first probability of the target object completing task A is 0.7. The fusion parameter determination model of the embodiment of the present disclosure can capture the subtle differences between different tasks completed by different push basic features (such as different target object features, and / or different content features to be pushed, and / or different target scene features), can fully reflect the differentiation between different push basic features, and improve the accuracy of push processing.

[0176] After obtaining the fused probability, step 350 is executed. In one embodiment, the contents to be pushed are sorted from high to low according to the fused probability to obtain a sorting result; and the contents to be pushed are pushed to the target object in the order of the sorting results. In one example, the fused probability of content R1 to be pushed is 0.54, the fused probability of content R2 to be pushed is 0.62, the fused probability of content R3 to be pushed is 0.32, the fused probability of content R4 to be pushed is 0.82, and the fused probability of content R5 to be pushed is 0.22. Then, these five contents to be pushed are sorted from high to low according to the fused probability, and the sorting result is: content R4 to be pushed, content R2 to be pushed, content R1 to be pushed, content R3 to be pushed, and content R5 to be pushed. Content R4 to be pushed is pushed to the target object first because the target object has the highest fused probability of content R4 to be pushed, which means that the target object is most likely to be interested in content R4 among all the contents to be pushed. Next, the content to be pushed R2, the content to be pushed R3, and the content to be pushed R5 are pushed to the target object.

[0177] The advantage of the embodiment of steps 310-350 is that multiple push basic features are input into a multi-task prediction model to obtain first probabilities of pushing content to a target object corresponding to multiple tasks. Since the importance of each task may vary, fusion is required. When determining the fusion parameters for each task, rather than setting fixed fusion parameters for each task, the fusion parameters are determined using a fusion parameter determination model. The fusion parameter determination model is trained based on the distribution of the first sample probabilities corresponding to the task, as predicted by the multi-task prediction model, for target samples. Thus, for a medium first sample probability, if the first sample probability is generally low and fluctuates relatively slowly across all samples corresponding to the task, it has a significant advantage across all samples of the task. The fusion parameter determination model will learn to assign a larger fusion parameter to the task for that target sample. For the same medium first sample probability, if the first sample probability is generally high and fluctuates significantly across all samples corresponding to the task, it will not have a significant advantage across all samples of the task. The fusion parameter determination model will learn to assign a smaller fusion parameter to the task for that target sample. The fusion parameter determination model trained in this way can accurately determine the fusion parameters of each task for each specific sample, thereby reducing the impact of large overall changes in the first probability distribution output by the multi-task prediction model due to continuous changes in data distribution, and improving the accuracy of the final content push.

[0178] The above steps 310-350, the training process of the multi-task prediction model, and the training process of the fusion parameter determination model are described in detail below.

[0179] Detailed description of step 310

[0180] In step 310, the push-based features are used to predict the fused probabilities of the target subject's preference for push content across multiple tasks. Based on these fused probabilities, the content to be pushed is determined. Therefore, the push-based features include multiple aspects that can influence the target subject's preference for push content.

[0181] In one embodiment, multiple push basic features are obtained, including:

[0182] Obtain target object features of the target object;

[0183] Obtaining target scene features of a target scene where a target object is located;

[0184] Taking the features of each content to be pushed in the content set to be pushed as the features of the content to be pushed in turn;

[0185] The target object features, target scene features, and features of the content to be pushed are used as multiple push basic features.

[0186] Regarding target object features, target objects that usually share a common set of target object features may exhibit certain commonalities in viewing target content. Therefore, using target object features for target content prediction can help improve the effect of push processing. Figure 5 As shown, the target object characteristics include education level characteristics, working years characteristics, job characteristics, hobby characteristics, historical viewing content characteristics, etc.

[0187] Regarding education level, subjects with the same education level tend to be more likely to choose content related to their education level. For example, if a subject with a college education tends to watch videos introducing graduate schools, then another subject with a college education may also tend to watch videos introducing graduate schools.

[0188] Regarding the characteristics of years of work experience, subjects with the same years of work experience are more likely to choose the same type of content related to their years of work experience. For example, if a subject with 15 years of work experience tends to watch workplace-related videos, then another subject with 15 years of work experience may also tend to watch the same workplace-related videos.

[0189] For work, users who have the same job are more likely to choose the same content that they often need to watch for their job. For example, students tend to watch educational or growth-oriented videos.

[0190] Subjects with the same hobbies may tend to view the same content. For example, subjects who like mathematics tend to view mathematics-related content; for subjects who like painting, painting-related content is more likely to be recommended and displayed.

[0191] For the historical viewing content feature, the subject may like to watch content with the same type of historical viewing content feature. For example, a subject tends to watch video content with a content length of less than 100 seconds, or a subject tends to watch video content with more than 1200 likes.

[0192] It should be understood that when obtaining the target subject's target characteristics, the target subject's consent must be obtained in advance. Furthermore, the collection, use, and processing of these target subject characteristics will comply with relevant laws, regulations, and standards. When obtaining the target subject's consent, individual permission or consent can be obtained through a pop-up window or by redirecting to a confirmation page.

[0193] In one embodiment, the target object's object characteristics can be obtained through registration information. This registration information can be the target object's registration information on the target terminal 110, or the registration information in the software used to push content. When the target object uses the target terminal 110 for the first time, the target terminal 110 may require the target object to register, which requires filling in information such as hobbies as registration information. In this way, the corresponding object characteristics can be obtained from the target object's terminal 110. When the target object uses the software used to push content for the first time, the target object also needs to register, which may require the target object to fill in information such as hobbies as registration information. Therefore, the object characteristics can also be obtained through the registration information in the software. The advantages of obtaining object characteristics through registration information are convenience, speed, and high efficiency. However, as mentioned above, obtaining the target object's registration information requires prior consent from the target object, which is not detailed here to conserve space.

[0194] In another embodiment, the target subject's target characteristics can be obtained using the target subject's historical viewing logs within a predetermined time period prior to the current time point. Historical viewing logs refer to records of pushed content viewed by the target subject within a predetermined time period prior to the current time point. The predetermined time period can be one hour, one day, one week, one month, etc. The predetermined time period can be determined based on the target subject's frequency of viewing pushed content. If the target subject frequently views pushed content, a shorter viewing time period can be sufficient to obtain the target subject's current characteristics. If the target subject rarely views pushed content, a longer viewing time period is required to ensure sufficient data for obtaining the target subject's characteristics. The predetermined time period can also be determined based on the type of pushed content. For example, if the pushed content is video, a longer time period can be used to capture the target subject's long-term preferences. If the pushed content is news, a shorter time period is required due to its timeliness. Obtaining the target subject's historical viewing logs also requires the target subject's prior consent.

[0195] The advantage of the embodiment of obtaining object characteristics through the historical viewing log of the target object within a predetermined time period before the current time point is that the target object's interest level in the content it has watched can be accurately obtained through the historical viewing log, and the target object's preferences can be predicted based on the interest level, thereby improving the accuracy of obtaining the target object's object characteristics.

[0196] Regarding target scene features, the target user is likely to view similar content in the same scene. For example, if the target scene is an office setting, the target user may be more interested in viewing work-related push content; if the target scene is at home, the target user may be more interested in viewing push content related to their preferences. Therefore, using target scene features for target content prediction helps improve the effectiveness of push processing. Target scene features generally include geographical area features, current input day type features (weekday, weekend, or holiday), weather features, etc.

[0197] The characteristics of the current geographical area refer to the characteristics of the geographical area where the target object is currently located. This geographical area can be an administrative geographical area, such as M City, N City, etc., or it can be a geographical area divided by entities on the map, such as the XX shopping mall area, XX hospital area, etc. For administrative geographical areas, the push content that the target object is interested in has an impact. For example, if the target object is in M ​​City, the push content can be tourist attractions or food culture related to M City. For geographical areas divided by entities, the push content that the target object is interested in also has an impact. For example, if the target object is in the XX shopping mall area, the push content can be event information or special items in the mall, etc.; if the target object is in the XX hospital area, the push content can be content related to health and wellness.

[0198] The current input day type feature refers to the characteristics of the day type when the target object views the pushed content. It is divided into weekdays, weekends, and holidays. The type of pushed content the target object wants to view may be different on different types of days. For example, on weekdays (Monday to Friday), the target object is often at work, and the pushed content they want to view may be work-related; on weekends (Saturday and Sunday), when the target object does not need to work, the pushed content they want to view may be related to their hobbies; on holidays, the target object may want to view pushed content related to holiday culture.

[0199] When obtaining the target scene features of the target object, the consent of the target object must be obtained in advance, which is similar to the above and will not be repeated here.

[0200] Regarding the characteristics of the content to be pushed, usually the content to be pushed with the same characteristics is easy to be viewed by the same target object, so using the characteristics of the content to be pushed for prediction is helpful to improve the effect of push processing. Figure 5 As shown, the features of the content to be pushed include content type features, content length features, like count features, comment count features, collection count features, etc.

[0201] It should be noted that the content type feature indicates the content type of the content to be pushed. The content type feature can be movies, music, books, pets, etc. The content length feature indicates the content length of the content to be pushed. The content length feature can be 10 seconds, 20 seconds, 30 seconds, 70 seconds, etc. The likes feature indicates the number of likes received by the content to be pushed. The likes feature can be any number in the range [0, +∞], such as 100, 200, 400, etc. The comments feature indicates the number of comments received by the content to be pushed. The comments feature can be any number in the range [0, +∞], such as 100, 200, 400, etc. The collection feature indicates the number of collections received by the content to be pushed. The collection feature can be any number in the range [0, +∞], such as 100, 200, 400, etc. The content to be pushed feature is a feature of the content to be pushed, and the content to be pushed is generally stored in the content set to be pushed. The content set to be pushed refers to a collection for storing multiple content to be pushed, and the features of each content to be pushed in the content set to be pushed are used as the features of the content to be pushed in turn.

[0202] After obtaining the target object features, target scene features, and features of the content to be pushed, the target object features, target scene features, and features of the content to be pushed are used as multiple push basic features. Specifically, one target object feature is used as a push basic feature, one target scene feature is used as a push basic feature, and one feature of the content to be pushed is used as a push basic feature. In one example, the target object has 5 target object features, the target scene in which the target object is located has 3 target scene features, and one feature of the content to be pushed includes 4 features of the content to be pushed, so a total of 12 push basic features can be obtained. The 12 push basic features are uniformly input into the multi-task prediction model to obtain the first probability corresponding to the multiple tasks for pushing the content to be pushed to the target object, thereby calculating the fused probability based on the multiple first probabilities, and then pushing the content to be pushed to the target object based on the fused probability.

[0203] The above embodiment combines the target object characteristics of the target object, the target scene characteristics of the scene, and the characteristics of the content to be pushed to form the push basic characteristics, so that the constructed push basic characteristics cover all aspects that affect the fusion probability of the target object to treat the pushed content, thereby improving the comprehensiveness of obtaining the push basic characteristics.

[0204] Detailed description of step 320

[0205] In step 320 , the plurality of push basic features are input into the multi-task prediction model to obtain first probabilities of pushing the content to be pushed to the target object corresponding to the plurality of tasks.

[0206] Figure 6 It is an overall structural diagram of a push processing method according to another embodiment of the present disclosure.

[0207] Figure 6 and Figure 4 The difference is, Figure 6 The model structure of the vectorization layer, the second splicing layer, and the multi-task prediction model is also shown. The multi-task prediction model includes multiple expert sub-models and gating nodes and prediction sub-models corresponding to each task. The expert sub-model is a model for multi-task modeling of the push basic features. The gating node is used to generate high-order features corresponding to the task based on the output of the expert sub-model. The prediction sub-model is a model for probabilistically predicting the high-order features corresponding to the task. Both the expert sub-model and the prediction sub-model can be deep neural networks. It has been explained here Figure 6 and Figure 4 The difference between the embodiments shown, Figure 6 The other parts shown are the same Figure 4 Same, no more details.

[0208] In one embodiment, step 320 includes:

[0209] Inputting multiple push basic features into the vectorization layer to obtain multiple push basic feature vectors;

[0210] Inputting the multiple push basic feature vectors into the second concatenation layer to obtain a cascade push basic feature vector;

[0211] Input the cascade push basic feature vector into the expert sub-model to obtain the model output of the expert sub-model;

[0212] The gating weights corresponding to the gating nodes are used to perform weighted sum on the model outputs of multiple expert sub-models to obtain the weighted model output;

[0213] The weighted sum model output is input into the prediction sub-model to obtain a first probability of pushing the content to be pushed to the target object corresponding to the task.

[0214] Each push basic feature corresponds to a push basic feature vector. In one embodiment, the push basic feature vector corresponding to the push basic feature can be determined by table lookup. In another embodiment, if the push basic feature is a character feature, the character feature is input into the embedding layer to obtain the push basic feature vector corresponding to the character feature; if the push basic feature is a numerical feature, the numerical feature is numerically mapped to obtain the push basic feature vector corresponding to the numerical feature.

[0215] A vector is an array of values ​​in different dimensions; it represents a point in multidimensional space. The line segment between this point and the origin in the multidimensional coordinate system has a magnitude and direction, which are the magnitude and direction of the vector. Each of the above values ​​is the point value projected onto the corresponding coordinate axis in the multidimensional coordinate system, i.e., a vector element. Vector elements can be numerical or symbolic. The vector elements of the push basic feature vector in the disclosed embodiment are numerical.

[0216] In one embodiment, the dimension of the push basic feature vector for each push basic feature does not necessarily have to be the first threshold. For example, for a historically viewed content feature characterized as "content sequences viewed for more than 10 seconds," the length of the content sequence is not fixed, so the dimension of the corresponding target push basic feature vector is uncertain. To mitigate the instability caused by varying dimensions, dimensionality adjustment is required.

[0217] In a specific implementation of this embodiment, after the multiple push basic features are input into the vectorization layer to obtain multiple push basic feature vectors, step 320 further includes:

[0218] Determining that a dimension of the pushed basic feature vector is greater than a first threshold;

[0219] Pooling is performed on the pushed basic feature vector so that the dimension of the pooled pushed basic feature vector is equal to the first threshold.

[0220] Pooling, also known as pooling, is essentially sampling. Pooling uses a certain method to reduce the dimensionality of the input push basic features to speed up the calculation. Pooling includes average pooling and maximum pooling. Specifically, assume that the push basic features are target object features, and the target object features are "content sequences watched for more than 10 seconds." After vectorizing the "content sequences watched for more than 10 seconds," the target push basic feature vector is "[70,90,80,10,40,50,30,20,80,90,100,50]" with a dimension of 12. If the first threshold set in advance is 8, the target push basic feature vector with a dimension of 12 needs to be pooled to obtain a target push basic feature vector with a dimension of 8.

[0221] In an example, assume that there are N push basic feature vectors, the dimensions of which are all set to d, and the nth feature is recorded as x n . Cascade / join N push basic feature vectors to obtain the cascade push basic feature vector: X = […, x n ,…]. The dimension of the cascade push basic feature vector is:

[0222] The cascade push basic feature vector is input into the expert sub-model. For example, after obtaining the fixed-length cascade push basic feature vector, multiple expert sub-models process the cascade push basic feature vector to obtain multiple expert model outputs. Each expert model output is a high-order representation vector. Assuming there are M expert sub-models, the expert sub-model processes the cascade push basic feature vector to obtain M high-order representation vectors. For the mth first expert model, its high-order representation vector is recorded as: e m (X), dimension is denoted as d e .

[0223] In an example, assuming there are T tasks, for task t, the gating weight of its gating node can be calculated as follows: G t (X) = softmax(W t ·X T ). Among them, G t (X) refers to the gating weight of the gating node corresponding to task t, W t The dimension is M×d X The high-order representation vectors input to the gated nodes are weighted and combined by the gating weights corresponding to task t to obtain the weighted sum model output, i.e. in, Refers to the weighted model output corresponding to task t. Input into the prediction sub-model corresponding to task t, and further obtain the first probability of task t, that is, Among them, p t refers to the first probability of task t, θ t (·) represents the prediction processing of the prediction sub-model for task t. Gating nodes corresponding to different tasks have different gating node weights, and prediction sub-models corresponding to different tasks also have different model structures and / or model parameters. As a result, the gated nodes and prediction sub-models calculate different results for the model outputs of multiple expert sub-models.

[0224] The advantage of this embodiment is that the comprehensiveness of feature extraction can be improved by using multiple expert sub-models to uniformly extract different aspects of the basic features of the push, and then the probability prediction of a single task is performed using the gated node and the prediction sub-model, which is conducive to obtaining more accurate multi-task prediction results and improving the accuracy of determining the push content.

[0225] Detailed description of step 330

[0226] In step 330, based on the first probabilities corresponding to the plurality of tasks, a fusion parameter is determined by a fusion parameter determination model. In one example, the tasks include task A and task B. Figure 4 or Figure 6As shown, the first probability P corresponding to task A A The first probability P corresponding to task B B The first concatenated layer is input to obtain a concatenated vector; the concatenated vector is input to a fusion parameter determination model to obtain a fusion parameter.

[0227] In the above example, the fusion parameter includes one parameter, for example, the fusion parameter of task A is 0.6. In one embodiment, the fusion parameter includes a first fusion parameter and a second fusion parameter, a total of two parameters. The fusion probability is determined based on the two parameters and the first probabilities corresponding to multiple tasks, which can improve the accuracy and flexibility of determining the fusion probability. The first fusion parameter is between 1 and the maximum value of the first fusion parameter, and the second fusion parameter is between 1 and the maximum value of the second fusion parameter. The fusion parameter determination model includes a first importance weight vector generation model and a second importance weight vector generation model.

[0228] When this embodiment is specifically implemented, refer to Figure 7 , step 330 includes:

[0229] Step 710: Input the first probabilities corresponding to the multiple tasks into a first importance weight vector generation model to obtain a first importance weight vector;

[0230] Step 720: Obtain the first importance weight corresponding to the task from the first importance weight vector, and determine the first fusion parameter based on the maximum value of the first fusion parameter and the first importance weight corresponding to the task;

[0231] Step 730: Input the first probabilities corresponding to the multiple tasks into a second importance weight vector generation model to obtain a second importance weight vector;

[0232] Step 740: Obtain the second importance weight corresponding to the task from the second importance weight vector, and determine the second fusion parameter based on the second fusion parameter maximum value and the second importance weight corresponding to the task.

[0233] Figure 8 It is a schematic diagram of determining a first fusion parameter and a second fusion parameter by using a fusion parameter determination model according to an embodiment of the present disclosure.

[0234] Figure 8The lowest level in step 320 is the first probability corresponding to multiple tasks. One task corresponds to one first probability. For example, if there are T tasks in total, there are T first probabilities in total. Step 710 is performed on the first probabilities corresponding to multiple tasks, and the first importance weight vector can be obtained using the first importance weight vector generation model. The first importance weight vector generation model is a model used to generate the first importance weight vector. The first importance weight vector generation model can be a deep neural network, such as a Senet network. The first importance weight vector is a vector comprising multiple elements. Each element indicates the importance weight of a task. The dimension of the first importance weight vector is the same as the number of tasks. If there are T tasks in total, the first importance weight vector includes T elements. In one example, there are three tasks, namely Task A, Task B, and Task C. The first importance weight vector is [0.4, 0.8, 0.3], with a total of three elements. Among them, the first element is 0.4, indicating that the first importance weight corresponding to Task A is 0.4. The second element is 0.8, indicating that the first importance weight corresponding to task B is 0.8. The third element is 0.3, indicating that the first importance weight corresponding to task C is 0.3.

[0235] After obtaining the first importance weight vector, step 720 is executed to determine the first fusion parameter based on the maximum value of the first fusion parameter and the first importance weight corresponding to the task. The first fusion parameter is a parameter that indicates the degree of contribution of the first probability of the task to the fused probability. The larger the first fusion parameter, the greater the contribution of the first probability of the task to the fused probability. The maximum value of the first fusion parameter refers to the largest first fusion parameter among multiple first fusion parameters. The first fusion parameter can be determined based on the first basic scalar and a third constant greater than 1. The third constant is a constant that can be set in advance as required, ensuring that it is greater than 1. The first basic scalar is a trainable real scalar. During the training phase, the first basic scalar needs to be jointly trained with the first importance weight vector generation model. The specific training process will be elaborated below and is omitted here. During the use phase, the first basic scalar has been determined, and the maximum value of the first fusion parameter can be determined based on the first basic scalar and the third constant.

[0236] Based on the above example, the first importance weight vector is [0.4, 0.8, 0.3]. Assume that the maximum value of the first fusion parameter is 1.5 and the third constant is 1. The first fusion parameter of Task A is 0.4*1.5+1=1.6. The first fusion parameter of Task B is 0.8*1.5+1=2.2. The first fusion parameter of Task C is 0.3*1.5+1=1.45.

[0237] Step 730 is performed on the first probabilities corresponding to multiple tasks, and a second importance weight vector can be obtained using the second importance weight vector generation model. The second importance weight vector generation model is a model used to generate the second importance weight vector. The second importance weight vector generation model can be a deep neural network, such as a Senet network. The second importance weight vector is a vector comprising multiple elements. Each element indicates the importance weight of a task. The dimension of the second importance weight vector is the same as the number of tasks. If there are T tasks in total, the second importance weight vector includes T elements. In one example, there are three tasks: Task A, Task B, and Task C. The second importance weight vector is [0.5, 0.2, 0.7], with a total of three elements. The first element is 0.5, indicating that the second importance weight corresponding to Task A is 0.5. The second element is 0.2, indicating that the second importance weight corresponding to Task B is 0.2. The third element is 0.7, indicating that the second importance weight corresponding to Task C is 0.7.

[0238] After obtaining the second importance weight vector, step 740 is executed to determine the second fusion parameter based on the maximum value of the second fusion parameter and the second importance weight corresponding to the task. The second fusion parameter, similar to the first fusion parameter, also indicates the contribution of the task's first probability to the fused probability. A larger second fusion parameter indicates a greater contribution of the task's first probability to the fused probability. It should be noted that the second fusion parameter and the first fusion parameter are two different parameters. While both indicate the contribution of the task's first probability to the fused probability, the factors influencing the contribution are different. For example, in one embodiment, the first fusion parameter is added to the first probability to obtain a sum. This sum is used as the base, and the second fusion parameter is used as the exponent to determine the contribution probability value of the task to the fused probability. The contribution probability value is a numerical value used to represent the degree of contribution. In this embodiment, a larger first fusion parameter results in a larger sum, i.e., a larger base, and thus a larger contribution probability value. A larger second fusion parameter results in a larger exponent, and thus a larger contribution probability value. It can be seen that the first fusion parameter affects the contribution degree at the base level, while the second fusion parameter affects the contribution degree at the exponential level.

[0239] The maximum value of the second fusion parameter refers to the largest second fusion parameter among multiple second fusion parameters. The second fusion parameter can be determined based on the second basic scalar and a third constant greater than 1. The third constant is a constant that can be set in advance according to needs, and it is sufficient to ensure that it is greater than 1. The second basic scalar is a trainable real scalar. During the training phase, the second basic scalar needs to be jointly trained with the second importance weight vector generation model. The specific training process will be expanded below and is omitted here. During the use phase, the second basic scalar has been determined, and the maximum value of the second fusion parameter can be determined based on the second basic scalar and the third constant.

[0240] Based on the above example, the second importance weight vector is [0.5, 0.2, 0.7]. Assume that the maximum value of the second fusion parameter is 1.6 and the third constant is 1. The second fusion parameter of Task A is 0.5*1.6+1=1.8. The first fusion parameter of Task B is 0.2*1.6+1=1.32. The first fusion parameter of Task C is 0.7*1.6+1=2.12.

[0241] The advantage of the embodiment of steps 710-740 is that the first fusion parameter is generated by the first importance weight vector generation model, and the second fusion parameter is generated by the second importance weight vector generation model, which can achieve separate determination of the fusion parameters, achieve parallel generation, and improve generation efficiency. In addition, the first fusion parameter and the second fusion parameter are generated using the first probabilities corresponding to multiple tasks as inputs to the model, which can indicate the degree of contribution of the first probabilities of different tasks to the fused probability. Since the first fusion parameter and the second fusion parameter are generated based on different generation models, the separation of the first fusion parameter and the second fusion parameter is achieved, especially in the training stage, which can reduce the difficulty of model adjustment.

[0242] In one embodiment, referring to Figure 9 , step 710 includes:

[0243] Step 910: Generate a first probability vector based on first probabilities corresponding to multiple tasks;

[0244] Step 920: Input the first probability vector into a first compression model to obtain a first compression vector, where the dimension of the first compression vector is smaller than the number of tasks;

[0245] Step 930: Input the first compression vector into the first stretching model to obtain a first importance weight vector, where the dimension of the first importance weight vector is equal to the number of tasks.

[0246] Figure 10 It is a schematic diagram of determining a first importance weight vector using a first compression model and a first stretching model according to an embodiment of the present disclosure.

[0247] Figure 10 The lowest layer is the first probabilities corresponding to the multiple tasks in step 320. Step 910 is executed on the first probabilities corresponding to the multiple tasks to obtain a first probability vector. Specifically, the first probabilities corresponding to the multiple tasks are input into the first concatenation layer, and the first concatenation layer concatenates the multiple first probabilities to obtain a first probability vector. In one example, assuming there are T tasks, and the first probability of the t-th task is p t , then the first probability vector is Z=[…,p t ,…].

[0248] In one embodiment, the first fusion parameter corresponding to the multiple tasks is […,α t ,…]=α max SENet α (Z). α t Refers to the first fusion parameter corresponding to the t-th task. max Refers to the maximum value of the first fusion parameter. SENet α (.) refers to the processing process of the first importance weight vector generation model. Z refers to the first probability vector. The first importance weight vector generation model includes a first compression model and a first stretching model.

[0249] Reference Figure 10 After obtaining the first probability vector, step 920 is executed, and the first compressed vector can be obtained through the first compression model. The first compression model refers to a model used to reduce the dimension of the first probability vector. The first compressed vector refers to the vector obtained after the dimension reduction of the first probability vector. The dimension of the first probability vector is equal to the number of tasks. The dimension of the first compressed vector is smaller than the dimension of the first probability vector, that is, the dimension of the first compressed vector is smaller than the number of tasks. For example, assuming that there are 4 tasks in total, the first probability vector is [0.4, 0.9, 0.3, 0.2], and the dimension is 4. After compressing the first probability vector using the first compression model, the obtained first compressed vector is [0.5, 0.8, 0.4], and the dimension is 3.

[0250] In one embodiment, the first compression model includes a first matrix and a nonlinear activation function; step 920 includes:

[0251] Determining a first product vector based on the first matrix and the first probability vector;

[0252] A nonlinear activation function is applied to the first product vector to obtain a first compressed vector.

[0253] In this embodiment, the first matrix refers to a matrix used to reduce the dimension of the first probability vector. The number of rows of the first matrix is ​​the dimension of the first compression vector, and the number of columns of the first matrix is ​​the number of tasks. The first matrix is ​​multiplied by the first probability vector to obtain a first product vector. The various vector elements in the first product vector are all linearly related. Considering that in practice, the various vector elements are generally nonlinearly related. The nonlinear activation function is introduced to increase the nonlinearity of the first compression vector. Therefore, a nonlinear activation function is applied to the first product vector to obtain a first compression vector. Nonlinear activation functions include Sigmoid, Tanh, ReLU, LReLU, ELU, GeLu, Softmax and other functions.

[0254] In one example, the first compressed vector is R sq =f sq (Z)=a1(W sq ·Z T ), where R sq Refers to the first compression vector, the dimension is T / r, T represents the number of tasks, r represents the compression multiple, f sq (Z) represents the process of compressing the first probability vector. sq is the first matrix, with dimensions Z T It refers to the transpose of the first probability vector, and a1 is a nonlinear activation function.

[0255] For example, assume that the first probability vector is [0.4, 0.9, 0.3, 0.2], and the first matrix W sq for W sq The dimension of is 3×4. Then the first product vector is: right Apply the Sigmoid function After that, the first compressed vector R is obtained sq is [0.55,0.8,0.85].

[0256] The advantage of this embodiment is that the first matrix is ​​used to achieve dimensionality reduction compression of the first probability vector, and a nonlinear activation function is used to increase the nonlinear relationship of the first product vector, thereby improving the accuracy of generating the first compressed vector.

[0257] Reference Figure 10After obtaining the first compression vector, step 930 is executed, and the first importance weight vector can be obtained through the first stretching model. The first stretching model refers to a model used to perform dimensionality upgrade on the first compression vector. The first importance weight vector refers to the vector obtained after the dimensionality upgrade of the first compression vector. The dimension of the first importance weight vector is equal to the number of tasks. Based on the above example, the first compression vector is [0.5, 0.8, 0.4], and the dimension is 3. Assuming that there are 4 tasks in total, after stretching the first compression vector using the first stretching model, the first importance weight vector obtained is [0.6, 0.9, 0.2, 0.7], and the dimension is 4.

[0258] In one embodiment, the first stretching model includes a second matrix and an exponential normalization function; step 930 includes:

[0259] determining a second product vector based on the second matrix and the first compressed vector;

[0260] Applying an exponential normalization function to the second product vector yields a first importance weight vector.

[0261] In this embodiment, the second matrix refers to a matrix used to increase the dimension of the first compressed vector. The number of rows of the second matrix is ​​the dimension of the first importance weight vector, and the number of columns of the second matrix is ​​the dimension of the first compressed vector. The second matrix is ​​multiplied by the first compressed vector to obtain a second product vector. The various vector elements in the second product vector are all in a linear relationship. Considering that in practice, the various vector elements are generally in a nonlinear relationship. The exponential normalization function is introduced to increase the nonlinearity of the first importance weight vector. Therefore, the exponential normalization function is applied to the second product vector to obtain the first importance weight vector. Exponential normalization functions include Sigmoid, Softmax and other functions.

[0262] In one example, based on the first compression vector R in the above example sq , the first importance weight vector is R ex =f ex (R sq )=a2(W ex ·[R sq ] T ). Among them, R ex is the first importance weight vector, with dimension T. ex is the second matrix with dimensions a2 is the sigmoid function Directly obtain the first importance weight between 0 and 1.

[0263] For example, assuming that the first compressed vector R sq is [0.55, 0.8, 0.85], and the second matrix W ex for W sq The dimension of is 4×3. Then the second product vector is: right After applying the sigmoid function, the first importance weight vector is [0.6, 0.58, 0.59, 0.56].

[0264] The advantage of this embodiment is that the first stretching matrix is ​​used to achieve dimensionality-enhancing stretching of the first compressed vector, and the exponential normalization function is used to increase the nonlinear relationship of the first importance weight vector, thereby improving the accuracy of generating the first importance weight vector.

[0265] The advantage of the embodiment of steps 910-930 is that the first importance weight vector is generated using the first compression model and the first stretching model, which can fully reflect the connection and difference between the first probabilities corresponding to multiple tasks, thereby improving the accuracy of generating the first importance weight vector.

[0266] In one embodiment, referring to Figure 11 , step 730 includes:

[0267] Step 1110: Generate a first probability vector based on first probabilities corresponding to multiple tasks;

[0268] Step 1120: Input the first probability vector into a second compression model to obtain a second compression vector, where the dimension of the second compression vector is smaller than the number of tasks;

[0269] Step 1130: Input the second compression vector into the second stretching model to obtain a second importance weight vector, where the dimension of the second importance weight vector is equal to the number of tasks.

[0270] Figure 12 It is a schematic diagram of determining a second importance weight vector using a second compression model and a second stretching model according to an embodiment of the present disclosure.

[0271] Figure 12 The lowest layer is the first probabilities corresponding to the multiple tasks in step 320. Step 1110 is executed on the first probabilities corresponding to the multiple tasks to obtain a first probability vector. Specifically, the first probabilities corresponding to the multiple tasks are input into the first concatenation layer, and the first concatenation layer concatenates the multiple first probabilities to obtain a first probability vector. In one example, assuming there are T tasks, and the first probability of the t-th task is p t , then the first probability vector is Z=[…,p t ,…].

[0272] In one embodiment, the second fusion parameters corresponding to the multiple tasks are […, β t ,…]=βmax SENet β (Z). β t Refers to the second fusion parameter corresponding to the t-th task. max Refers to the maximum value of the second fusion parameter. SENet β (.) refers to the processing process of the second importance weight vector generation model. Z refers to the first probability vector. The second importance weight vector generation model includes a second compression model and a second stretching model.

[0273] Reference Figure 12 After obtaining the first probability vector, step 1120 is performed to obtain a second compressed vector using a second compression model. The second compression model is a model used to perform dimensionality reduction on the first probability vector. The second compressed vector is the vector obtained after dimensionality reduction of the first probability vector. The dimension of the first probability vector is equal to the number of tasks. The dimension of the second compressed vector is smaller than the dimension of the first probability vector, that is, the dimension of the second compressed vector is smaller than the number of tasks.

[0274] In one embodiment, the second compression model includes a third matrix and a nonlinear activation function; step 1120 includes:

[0275] determining a third product vector based on the third matrix and the first probability vector;

[0276] A nonlinear activation function is applied to the third product vector to obtain a second compressed vector.

[0277] In this embodiment, the third matrix refers to a matrix used to reduce the dimension of the first probability vector. The number of rows of the third matrix is ​​the dimension of the second compressed vector, and the number of columns of the third matrix is ​​the number of tasks. The third matrix is ​​multiplied by the first probability vector to obtain a third product vector. Each vector element in the third product vector is in a linear relationship. Considering that in practice, the relationship between each vector element is generally nonlinear. The nonlinear activation function is introduced to increase the nonlinearity of the second compressed vector. Therefore, a nonlinear activation function is applied to the third product vector to obtain a second compressed vector. Nonlinear activation functions include Sigmoid, Tanh, ReLU, LReLU, ELU, GeLu, Softmax and other functions.

[0278] The advantage of this embodiment is that the third matrix is ​​used to achieve dimensionality reduction compression of the first probability vector, and a nonlinear activation function is used to increase a nonlinear relationship, thereby improving the accuracy of generating the second compressed vector.

[0279] Reference Figure 12After obtaining the second compressed vector, step 1130 is performed to obtain a second importance weight vector using a second stretching model. The second stretching model is a model used to perform dimensionality upscaling on the second compressed vector. The second importance weight vector is the vector obtained after the dimensionality upscaling of the second compressed vector. The dimension of the second importance weight vector is equal to the number of tasks.

[0280] In one embodiment, the second stretching model includes a fourth matrix and an exponential normalization function; step 1130 includes:

[0281] determining a fourth product vector based on the fourth matrix and the second compressed vector;

[0282] An exponential normalization function is applied to the fourth product vector to obtain a second importance weight vector.

[0283] In this embodiment, the fourth matrix refers to a matrix used to increase the dimension of the second compressed vector. The number of rows of the fourth matrix is ​​the dimension of the second importance weight vector, and the number of columns of the fourth matrix is ​​the dimension of the second compressed vector. The fourth matrix is ​​multiplied by the second compressed vector to obtain a fourth product vector. Each vector element in the fourth product vector is in a linear relationship. Considering that in practice, the relationship between each vector element is generally nonlinear. The exponential normalization function is introduced to increase nonlinearity. Therefore, the exponential normalization function is applied to the fourth product vector to obtain the second importance weight vector. Exponential normalization functions include Sigmoid, Softmax and other functions.

[0284] The advantage of this embodiment is that the fourth stretching matrix is ​​used to achieve dimensionality-enhancing stretching of the second compressed vector, and the exponential normalization function is used to increase the nonlinear relationship of the second importance weight vector, thereby improving the accuracy of generating the second importance weight vector.

[0285] The advantage of the embodiment of steps 1110-1130 is that the second importance weight vector is generated using the second compression model and the second stretching model, which can fully reflect the connection and difference between the first probabilities corresponding to multiple tasks, thereby improving the accuracy of generating the second importance weight vector.

[0286] Detailed description of step 340

[0287] In one embodiment, the fusion parameter includes a first fusion parameter and a second fusion parameter; Figure 13 , step 340 includes:

[0288] Step 1310: Determine a second sum of the first probability corresponding to the task and the first fusion parameter;

[0289] Step 1320: Determine a probability contribution value corresponding to the task based on the second sum corresponding to the task and the second fusion parameter;

[0290] Step 1330: Multiply the probability contribution values ​​corresponding to multiple tasks to obtain the fused probability.

[0291] The first fusion parameter indicates the contribution of the first probability of a task to the fused probability. The second fusion parameter also indicates the contribution of the first probability of a task to the fused probability. The first and second fusion parameters affect the contribution at different levels.

[0292] In step 1310, the first probability corresponding to the task is added to the first fusion parameter to obtain a second sum. Each task corresponds to one second sum. Assuming there are T tasks, there are a total of T first sums.

[0293] In step 1320, the probability contribution value corresponding to a task is a value used to measure the contribution of the task's first probability to the fused probability. The larger the probability contribution value corresponding to a task, the greater the contribution of the task's first probability to the fused probability.

[0294] In one embodiment, referring to Figure 14 , step 1320 includes:

[0295] Step 1410: determine the third sum of 1 and the maximum value of the first fusion parameter;

[0296] Step 1420: Determine a third quotient of the second sum and the third sum;

[0297] Step 1430: Determine a power value with the third quotient as the base and the second fusion parameter as the exponent as the probability contribution value corresponding to the task.

[0298] The maximum first fusion parameter is the largest first fusion parameter among multiple first fusion parameters. Adding 1 to the maximum first fusion parameter yields the third sum. Each task corresponds to one third sum. Assuming there are T tasks, there are T third sums. Comparing the second sum with the third sum reveals that the second probability in the second sum is less than 1 in the third sum, and that the first fusion parameter in the second sum is less than the maximum first fusion parameter in the third sum. Therefore, the third quotient obtained by dividing the second sum by the third sum is a value less than 1. Assuming there are T tasks, there are T third quotients. When calculating the probability contribution of each task, the third quotient is raised to the power of the second fusion parameter, effectively raising the third quotient to the power of the second fusion parameter to obtain the probability contribution value. The probability contribution value represents the contribution of the task's first probability to the fused probability. A larger probability contribution value indicates a greater contribution of the task's first probability to the fused probability.

[0299] The advantage of the embodiment of steps 1410-1430 is that by using the third quotient as the base and the second fusion parameter as the power value of the exponent, the probability contribution values ​​of each task are within the same numerical range, which can more effectively reflect the influence of the first probabilities of different tasks on the determination of the fused probability, thereby improving the accuracy of the determined fused probability.

[0300] In step 1330, the probability contribution values ​​corresponding to the multiple tasks are multiplied to obtain the fused probability. The calculation process of the fused probability is shown in Formula 1:

[0301]

[0302] In Formula 1, T refers to the number of tasks, and t refers to the t-th task. t Refers to the first probability of the t-th task. t Refers to the first fusion parameter of the t-th task. max Refers to the maximum value of the first fusion parameter. t Refers to the second fusion parameter of the t-th task. ∏(.) refers to the multiplication process. α t ∈(1,α max ), α t >1,β t ∈(1,β max ), β t >1.

[0303] The advantage of the embodiment of steps 1310-1330 is that the first sum is first determined using the first fusion parameter, and then the probability contribution value corresponding to each task is determined using the first sum and the second fusion parameter. While determining the probability contribution value, the first fusion parameter and the second fusion parameter are relatively independent, which improves the flexibility of the determination. Furthermore, the probability contribution values ​​of multiple tasks are integrated using a multiplication method, which reduces the difficulty of determination and improves the flexibility of determination while determining the fused probability.

[0304] In one embodiment, the first fusion parameter is between 1 and the maximum value of the first fusion parameter, and the second fusion parameter is between 1 and the maximum value of the second fusion parameter. Figure 15 The maximum value of the first fusion parameter and the maximum value of the second fusion parameter are determined by the following method:

[0305] Step 1510: Obtain a first basic scalar corresponding to the first fusion parameter, a second basic scalar corresponding to the second fusion parameter, and a third constant greater than 1;

[0306] Step 1520: Perform exponential normalization on the first basic scalar to obtain a first exponential normalized value;

[0307] Step 1530: Perform exponential normalization on the second basic scalar to obtain a second exponential normalized value;

[0308] Step 1540: Determine a maximum value of the first fusion parameter based on the third constant and the first exponential normalized value;

[0309] Step 1550: Determine the maximum value of the second fusion parameter based on the third constant and the second exponential normalization value, wherein the first basic scalar and the second basic scalar are jointly trained with the fusion parameter determination model.

[0310] In step 1510, each task corresponds to a first basic scalar and a second basic scalar. The first basic scalar and the third constant jointly determine the first fusion parameter. The second basic scalar and the third constant jointly determine the second fusion parameter. The third constant is a constant that can be set in advance as needed, ensuring that it is greater than 1. The first basic scalar is a trainable real scalar. During the training phase, the first basic scalar needs to be jointly trained with the fusion parameter determination model. The specific training process will be expanded below and is omitted here. During the use phase, the first basic scalar has been determined, and the maximum value of the first fusion parameter can be determined based on the first basic scalar and the third constant.

[0311] In step 1520, an exponential normalization function may be introduced to perform exponential normalization on the first basic scalar so that the first exponential normalized value falls within a specific numerical range, making it easier to observe the differences in the impact of different tasks on the first fusion parameter.

[0312] In step 1530, similar to step 1520, an exponential normalization function may be introduced to perform exponential normalization on the second basic scalar so that the second exponential normalized value falls within a specific numerical range, making it easier to observe the differences in the impact of different tasks on the second fusion parameter.

[0313] In step 1540, a maximum value of the first fusion parameter is determined based on the third constant and the first exponential normalized value. In one embodiment, step 1540 includes: subtracting 1 from the third constant to obtain a first difference; multiplying the first difference by the first exponential normalized value to obtain a first multiplication result; and adding 1 to the first multiplication result to obtain the maximum value of the first fusion parameter. The process of calculating the maximum value of the first fusion parameter is shown in Formula 2:

[0314] α max =(K-1)·sigmoid(v1)+1 (Formula 2).

[0315] In formula 2, α max refers to the maximum value of the first fusion parameter, K refers to the third constant (K>1), v1 refers to the first basic scalar, and sigmoid(·) refers to the exponential normalization process.

[0316] In step 1550, a second fusion parameter maximum value is determined based on the third constant and the second exponential normalized value. In one embodiment, step 1550 includes: subtracting 1 from the third constant to obtain a second difference; multiplying the second difference by the second exponential normalized value to obtain a second multiplication result; and adding 1 to the second multiplication result to obtain the second fusion parameter maximum value. The process of calculating the second fusion parameter maximum value is shown in Formula 3:

[0317] β max =(K-1)·sigmoid(v2)+1 (Formula 3).

[0318] In formula 3, β max refers to the maximum value of the second fusion parameter, K refers to the third constant (K>1), v2 refers to the second basic scalar, and sigmoid(·) refers to the exponential normalization process.

[0319] The advantage of the embodiment of steps 1510-1550 is that, by performing exponential normalization on the first basic scalar and then determining the first fusion parameter based on the first exponential normalization value, while determining the first fusion parameter based on the first basic scalar, the first fusion parameters corresponding to different tasks are made to fall within a specific numerical range, which can reflect the differences in the first fusion parameter between different tasks. Similarly, this embodiment, while determining the second fusion parameter based on the second basic scalar, makes the first fusion parameters corresponding to different tasks fall within a specific numerical range, which can reflect the differences in the second fusion parameter between different tasks. In addition, the first basic scalar and the second basic scalar are jointly trained with the fusion parameter determination model, which can reduce the training overhead while ensuring the training effect.

[0320] It should be understood that joint training refers to the process of adjusting multiple trainable scalars or model parameters. For example, a loss function is determined based on the first basic scalar, the second basic scalar, and the fusion parameter determination model. Based on the loss function, the parameters of the first basic scalar, the second basic scalar, and the fusion parameter determination model are adjusted until the loss function falls below a loss threshold, completing the joint training.

[0321] Detailed description of step 350

[0322] In step 350 , based on the fused probability, the content to be pushed is pushed to the target object.

[0323] like Figure 16As shown, the target object features of the target object include: "university" and "movie". Among them, "university" refers to the educational level of the target object, and "movie" refers to the hobby of the target object. The content features of the content R1 to be pushed in the content set to be pushed include: "picture", "1000 likes", and the probability after fusion is 50%. The content features of the content R2 to be pushed in the content set to be pushed include: "movie", "length 70 seconds", "1300 likes", "500 comments", and "100 collections", and the probability after fusion is 90%. The content features of the content R3 to be pushed in the content set to be pushed include: "movie", "length 3600 seconds", "30 likes", and the probability after fusion is 66%. In this way, based on the probability after fusion, the content R2 to be pushed can be used as the target content A, and the content R3 to be pushed can be used as the target content B. The target content A and the target content B constitute the target content set, and the target content in the target content set is pushed to the target object.

[0324] The training process of the multi-task prediction model in the embodiment of the present disclosure

[0325] In one embodiment, the multi-task prediction model is trained by:

[0326] Get the sample push feature set;

[0327] Extracting a first sample and a second sample from the sample push feature;

[0328] The first sample and the second sample extracted from each sample push feature are input into the multi-task prediction model for training.

[0329] The sample push feature set includes multiple sample push features. The sample push features include sample object features, sample content features, and sample scene features. The sample push feature also has a pushed training label, which indicates whether the sample content is pushed to the sample object when the sample object is in the sample scene. For example, a sample push feature has a training label of {Task A: 1, Task B; 0}, where a training label of 1 represents that the task is completed, and a training label of 0 represents that the task is not completed. That is to say, in this example, when the sample object is in the sample scene, the sample object completes Task A for the sample content, but does not complete Task B.

[0330] The first sample extracted from the sample push feature contains the sample object feature and the sample content feature to be pushed; the second sample contains the sample object feature, the sample content feature to be pushed, and the sample scene feature. The reason for extracting the first and second samples is that in the actual use of the multi-task prediction model, if the scene features of the target object are not obtained, the push effect of the multi-task prediction model will be affected. Therefore, it is necessary to use the first and second samples of the same sample push feature to train the multi-task prediction model.

[0331] In one embodiment, the first sample and the second sample extracted from each sample push feature are input into a multi-task prediction model for training, including:

[0332] For each of the first sample and the second sample, determining a second sample sub-loss function for each task based on the training label and the second probability predicted by the multi-task prediction model;

[0333] Based on the sum of the second sample sub-loss functions of each task, a second sample loss function is obtained;

[0334] Based on the second sample loss function, a multi-task prediction model is trained.

[0335] Specifically, in one embodiment, the second sample sub-loss function can be determined as follows:

[0336] L t =-y t logp t -(1-y t )log(1-p t ) Formula 4.

[0337] In formula 4, L t is the second sample loss function of the t-th task, y t is the training label of the t-th task, p t is the second probability of the tth task. If the training label indicates that the task is completed, then the training label is 1, otherwise, the training label is 0.

[0338] In one embodiment, the second sample loss function may be determined as follows:

[0339]

[0340] In Formula 5, L refers to the second sample loss function, T refers to the number of tasks, and L t is the second sample loss function of the t-th task.

[0341] It should be noted that the above formulas 4 and 5 are not exclusive. For example, a logarithmic operation or a linear operation may be added to formula 4. This embodiment does not specifically limit this.

[0342] After determining the second sample loss function, the multi-task prediction model can be trained based on the second sample loss function, that is, the parameters of the multi-task prediction model can be adjusted. Specifically, a first threshold can be pre-set. When the error function is less than the first threshold, the training process ends. When the second sample loss function is greater than or equal to the first threshold, the parameters of the multi-task prediction model are adjusted until the second sample loss function is less than the first threshold.

[0343] This embodiment has the advantage of improving the prediction accuracy of the trained model by constructing a second sample loss function to train the multi-task prediction model. Furthermore, the second sample sub-loss function corresponding to each task contributes equally to determining the second sample loss function, improving fairness across multiple tasks when training the model.

[0344] The training process of the fusion parameter determination model in the embodiment of the present disclosure

[0345] In one embodiment, the fusion parameter determination model is trained based on the distribution position of the first sample probability corresponding to the task of each target sample in the target sample set among the first sample probabilities of each sample corresponding to the task. The first sample probability is obtained by inputting the basic features of the sample push of the target sample into the multi-task prediction model.

[0346] The target sample set refers to a sample set containing multiple samples. Samples are data used to train the fusion parameter determination model. Multiple samples include target samples. Samples include multiple sample push basic features. Sample push basic features are features used to predict the completion of different tasks performed by the sample object on the sample content to be pushed. The sample push basic features of multiple target samples are input into the multi-task prediction model to obtain multiple first sample probabilities corresponding to each task. For each task, there are first sample probabilities corresponding to multiple samples. Each sample can be used as a target sample in turn. The first sample probability of each target sample has a corresponding distribution position in the first sample probabilities of each sample.

[0347] The distribution position can be the difference between the first sample probability of the target sample corresponding to the task and the average value of the first sample probabilities of each sample corresponding to the task. For example, for a certain task, there are 3 first sample probabilities: 0.5, 0.6, and 0.7. The average value is (0.5+0.6+0.7) / 3=0.6. The difference between 0.5 and the average value of 0.6 is 0.1, so the distribution position is 0.1. The difference between 0.6 and the average value of 0.6 is 0, so the distribution position is 0. The difference between 0.7 and the average value of 0.6 is 0.1, so the distribution position is 0.1.

[0348] The distribution position can also be the first sample probability of the target sample corresponding to the task, which is a multiple of the standard deviation of the first sample probability of each sample corresponding to the task compared to the average of the first sample probabilities of each sample corresponding to the task. The average value refers to the average value of the probability calculated based on the first sample probabilities of multiple samples. The standard deviation refers to the standard deviation / variance of the probability calculated based on the first sample probabilities of multiple samples. For example, assuming the average value is 0.4 and the standard deviation is 0.2. If the first probability is 0.5, the distribution position is (0.5-0.4) / 0.2=0.5. If the first probability is 0.4, the distribution position is (0.4-0.4) / 0.2=0.

[0349] It should be understood that the larger the distribution position of the first sample probability of the target sample corresponding to the task, the greater the proportion or importance of the first probability in the multiple first sample probabilities.

[0350] The distribution position can reflect the proportion or importance of the first sample probability of each target sample among the first sample probabilities of multiple target samples. Therefore, the fusion parameter determination model trained based on the distribution position can reflect the proportion or importance of the first probability of each task among the first probabilities of multiple tasks. During the model usage phase, after the first probabilities corresponding to multiple tasks are input into the fusion parameter determination model, the fusion parameter output by the fusion parameter determination model can reflect the weight of each task in predicting the target object's interest in the content to be pushed.

[0351] In one embodiment, referring to Figure 17 , the fusion parameter determination model is trained in the following way:

[0352] Step 1710: Acquire a target sample set, where the target sample set includes multiple target samples, each target sample having multiple sample push basic features, and the multiple sample push basic features include sample object features and sample content features;

[0353] Step 1720: Input the multiple sample push basic features into the multi-task prediction model to obtain a second probability of pushing sample content for the sample object corresponding to the multiple tasks;

[0354] Step 1730: Input the second probabilities corresponding to the multiple tasks into the fusion parameter determination model to obtain sample fusion parameters, and determine the sample fusion probability of pushing sample content to the sample object based on the second probabilities corresponding to the multiple tasks and the fusion parameters;

[0355] Step 1740: Obtain the distribution position of the first sample probability of the target sample corresponding to the task in the first sample probabilities of each sample corresponding to the task;

[0356] Step 1750: Calculate the first sample loss function based on the probability and distribution position of the sample after fusion, and train the fusion parameter to determine the model based on the first sample loss function.

[0357] In step 1710, multiple sample push basic features can be first obtained, and then a target sample set can be constructed based on the multiple sample push basic features. The sample push basic features are similar to the push basic features in step 310, except that one is in the model training phase and the other is in the model usage phase. Similarly, the sample object features are similar to the target object features in step 310, and the sample content features are similar to the content features to be pushed in step 310. The process of obtaining sample push basic features is similar to the process of obtaining multiple push basic features in step 310 and will not be repeated here.

[0358] In step 1720, a multi-task prediction model is used to perform multi-task predictions, and a second probability corresponding to multiple tasks and pushing sample content to the sample object is obtained. The second probability is the probability that the sample object completes the task for the sample content. The greater the probability that the sample object completes the task for the sample content, the greater the second probability, and thus the greater the probability of pushing the sample content to the sample object. The multi-task prediction model is a model that predicts the probability of the sample object completing different tasks for the sample content based on the basic features of the sample push. For example, the multi-task prediction model is the deep neural network described above. Tasks refer to matters related to the sample content that are used to weigh the sample object's interest in the sample content from multiple aspects. For example, a task is a sample object liking the sample content, or a sample object collecting the sample content, or a sample object commenting on the sample content. Each task corresponds to a second probability.

[0359] In step 1730, the sample fusion parameters are determined using a fusion parameter determination model. The sample fusion parameters indicate the weight of each task in predicting the sample subject's interest in the sample content. The fusion parameter determination model is a model that predicts the sample fusion parameters based on the probabilities corresponding to multiple tasks. For example, the fusion parameter determination model is the deep neural network described above. The process of determining the post-fusion probability of the sample based on the second probability and the fusion parameters is similar to the process of determining the post-fusion probability based on the first probability and the fusion parameters in step 340, and will not be repeated here.

[0360] In step 1740 , the first sample probability corresponding to the task of the target sample whose distribution position is the average of the first sample probabilities of the samples corresponding to the task deviates from the multiple of the standard deviation of the first sample probabilities of the samples corresponding to the task.

[0361] In one embodiment, step 1740 includes:

[0362] Get multiple samples corresponding to the task;

[0363] Determine the mean and standard deviation of the first sample probabilities of the plurality of samples corresponding to the task;

[0364] Determine a first sample probability of the target sample corresponding to the task and a second difference from the mean value;

[0365] The second quotient of the second difference and the standard deviation is determined as the distribution position.

[0366] In this embodiment, the calculation process for determining the distribution position is shown in Formula 6:

[0367]

[0368] In Formula 6, η t Refers to the distribution position, and λ is a parameter set in advance that is greater than 0 and is used to control the impact of the probability deviation on the distribution position. t μ refers to the probability of the first sample of the tth task of the target sample. t Refers to the average of the first sample probabilities of multiple samples corresponding to the t-th task, σ t It refers to the standard deviation of the probability of the first sample of multiple samples corresponding to the t-th task. B refers to the number of samples corresponding to the t-th task. i refers to the i-th sample. [p t ] i It refers to the first sample probability of the i-th sample in the B samples of the t-th task.

[0369] In one example, if Figure 18 As shown, the sample set corresponding to the task includes sample Y1, sample Y2, sample Y3, sample Y4, and sample Y5. The first sample probability of sample Y1 is 0.5, the first sample probability of sample Y2 is 0.4, the first sample probability of sample Y3 is 0.6, the first sample probability of sample Y4 is 0.3, and the first sample probability of sample Y5 is 0.2. The average value is (0.5 + 0.4 + 0.6 + 0.3 + 0.2) / 5 = 0.4.

[0370] The standard deviation is The probability of the first sample of the target sample corresponding to the task is 0.7, and the distribution position can be (0.7-0.4) / 0.45=0.67.

[0371] In another example, combined with the above example, assuming that λ is 1, the distribution position can also be sigmoid(1·0.67)=0.66.

[0372] The advantage of the above embodiment is that the mean and standard deviation are determined based on the first sample probability of multiple samples corresponding to the task, and the distribution position is determined based on the first sample probability, mean and standard deviation of the target sample, thereby improving accuracy.

[0373] In one embodiment, referring to Figure 19 , step 1750 includes:

[0374] Step 1910: Perform a monotonic transformation on the sample fusion probability to obtain the sample monotonic transformation probability corresponding to the task;

[0375] Step 1920: Determine a first sample sub-loss function corresponding to the task based on the probability of the sample corresponding to the task after monotonically transformed;

[0376] Step 1930: Determine a first sample loss function based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task.

[0377] In step 1910, the sample fusion probability is a numerical value calculated from the first sample probabilities corresponding to multiple tasks and the fusion parameters. The sample fusion probabilities are individually transformed to reversely calculate the sample monotonically transformed probabilities corresponding to each task based on the sample fusion probabilities. The sample monotonically transformed probabilities are not equal to the first sample probabilities. However, similar to the first sample probabilities, the sample monotonically transformed probabilities can represent the probability that the sample object will complete the task for the sample content.

[0378] In one embodiment, referring to Figure 20 , step 1910 includes:

[0379] Step 2010: Determine a first fusion score based on the probability of sample fusion;

[0380] Step 2020: Perform a monotonic transformation on the first fusion score to obtain a second fusion score corresponding to the task;

[0381] Step 2030: Based on the second fusion score corresponding to the task, determine the probability of the sample corresponding to the task after monotonic transformation.

[0382] In step 2010, the post-fusion sample probabilities are inversely transformed to obtain a first fusion score. It should be understood that when calculating the post-fusion sample probabilities, some exponential normalization functions (such as sigmoid and softmax functions) are often used to calculate the fusion score. Therefore, when inversely calculating the fusion score based on the post-fusion sample probabilities, some logical subfunctions (such as the inverse sigmoid function) can be used to achieve the inverse transformation.

[0383] In one embodiment, step 2010 includes:

[0384] Determine the first difference between 1 and the probability of sample fusion;

[0385] The logarithm of the first quotient of the probability of the sample after fusion and the first difference is determined as the first fusion score.

[0386] In this embodiment, the process of calculating the first fusion score is shown in Formula 7:

[0387]

[0388] In Formula 7, s refers to the first fusion score, p refers to the probability of sample fusion, and logit(.) refers to the inverse function of sigmoid.

[0389] The advantage of this embodiment is that it can quickly determine the first fusion score and improve processing efficiency.

[0390] In one embodiment, referring to Figure 21 , step 2020 includes:

[0391] Step 2110: Obtain the monotonic transformation coefficient and monotonic transformation offset corresponding to the task;

[0392] Step 2120: Based on the monotonic transformation coefficient and the monotonic transformation offset corresponding to the task, perform a monotonic transformation on the first fused score to obtain a second fused score corresponding to the task.

[0393] In this embodiment, the monotonic transform coefficient is a trainable scalar. The monotonic transform offset is also a trainable scalar. The monotonic transform coefficient and the monotonic transform offset corresponding to each task are jointly trained with the fusion parameter determination model. Joint training refers to the process of adjusting multiple trainable scalars or model parameters. For example, after determining the first sample loss function, the monotonic transform coefficient, the monotonic transform offset, and the fusion parameter determination model are parameter adjusted based on the first sample loss function until the first sample loss function is lower than a preset threshold, and the joint training is completed. The determination of the first sample loss function will be introduced in the detailed description of step 1750 below.

[0394] After obtaining the monotonic transformation coefficient and the monotonic transformation offset, a monotonic transformation is performed on the first fused score to obtain a second fused score.

[0395] In one embodiment, step 2120 includes:

[0396] Multiplying the square of the monotone transformation coefficient by the first fusion factor to obtain a first product;

[0397] A second fused score is determined based on the first product and the monotonic transformation offset.

[0398] In this embodiment, the process of calculating the second fusion score is shown in Formula 8:

[0399] s t =a t ·a t ·s+b t (Formula 8).

[0400] In formula 8, s t refers to the second fusion score corresponding to the t-th task, a t refers to the monotonic transformation coefficient, b t It refers to the monotonic transformation offset, and s refers to the first fusion score.

[0401] It should be noted that Formula 8 is not unique. For example, a logarithmic operation or a linear operation may be added to Formula 8. This embodiment does not specifically limit this.

[0402] This embodiment has the advantage of multiplying the square of the monotonic transform coefficient by the first fused score, ensuring that the sign of the first product is positive, satisfying the purpose of monotonic transformation while reducing computational complexity. Furthermore, by applying an offset to the first product using the monotonic transform offset, the nonlinearity of the determined second fused score is reduced, thereby improving the accuracy of the determination.

[0403] In step 2030, the second fusion score can be calculated using an exponential normalization function (such as sigmoid, softmax function, etc.) to obtain the probability of the sample after monotonic transformation. For example, refers to the probability of the sample of the t-th task after monotonic transformation, s t It refers to the second fusion score corresponding to the t-th task.

[0404] In one embodiment, step 2030 includes:

[0405] determining a first power value with the first constant as base and the opposite of the second fused fraction as exponent;

[0406] determining a first sum of the first power value and a second constant;

[0407] The reciprocal of the first sum is taken as the probability of the sample corresponding to the task after monotonous transformation.

[0408] In this embodiment, the calculation process for determining the probability of a sample after monotonic transformation is shown in Formula 9:

[0409]

[0410] In formula 9, refers to the probability of the sample of the t-th task after monotonic transformation, s trefers to the second fusion score corresponding to the tth task, k1 refers to the first constant, and k2 refers to the second constant. k1 and k2 are set according to actual needs and are not specifically limited in this embodiment. For example, k2 is set to the natural constant e, and k1 is set to 1.

[0411] The advantage of this embodiment is that the first constant and the second constant can be flexibly set, so that the probability of the sample after monotonic transformation can be flexibly calculated based on the first constant and the second constant, thereby improving flexibility and efficiency.

[0412] In step 1920, the first sample sub-loss function refers to the distance between the monotonically transformed probability of the sample and the task label for a single task. The closer the distance, the smaller the error and the better the prediction effect. The task label indicates whether the sample object completes the task for the sample content. For example, if the sample object completes the task for the sample content, the task label is 1, otherwise, the task label is 0. In one embodiment, the first sample probability corresponding to the task can be used as the task label, so that the first sample sub-loss function corresponding to the task is determined based on the monotonically transformed probability of the sample for the single task and the first sample probability.

[0413] In the specific implementation of this embodiment, step 1920 includes:

[0414] Determine the first logarithm of the monotonically transformed probability of the sample corresponding to the task and the second product of the first logarithm and the second probability corresponding to the task;

[0415] Determine a first complement of the second probability and a second complement of the probability after the monotonically transformed sample, wherein the first complement is the difference between 1 and the second probability, and the second complement is the difference between 1 and the probability after the monotonically transformed sample;

[0416] Determine the third product of the second logarithm of the second complement and the first complement;

[0417] Based on the second product and the third product, the first sample sub-loss function corresponding to the task is determined.

[0418] In this embodiment, the calculation process of determining the first sample sub-loss function is shown in Formula 10:

[0419]

[0420] In formula 10, Refers to the first sample loss function corresponding to the t-th task, p t refers to the second probability corresponding to the t-th task, It refers to the probability of the sample of the t-th task after monotonic transformation.

[0421] The advantage of this embodiment is that there is no need to set additional task labels. The second probability corresponding to the task is used as the task label, and the first sample sub-loss function is jointly determined with the probability after monotonic transformation of the sample. While being able to determine the loss function, the accuracy and efficiency of the determination are improved.

[0422] In step 1930, a first sample loss function is determined based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task. For each task, the distribution position can reflect the weight of the first sample sub-loss function of the task, thereby reflecting the weight of the task on the first sample loss function.

[0423] In one embodiment, referring to Figure 22 , step 1930 includes:

[0424] Step 2210: Determine the fourth product of the first sample sub-loss function corresponding to the task and the distribution position corresponding to the task;

[0425] Step 2220: Add the fourth products of multiple tasks to obtain a first sample loss function.

[0426] In step 2210, the first sample sub-loss function can be multiplied by the distribution position to obtain a fourth product. In actual training, it has been found that certain tasks may be of particular interest to the subject, and these tasks should be given a certain basic weight. Therefore, in one embodiment, in addition to considering the influence of the distribution position on determining the fourth product, a priori weighting is also required.

[0427] In the specific implementation of this embodiment, step 2210 includes:

[0428] Get the prior weight corresponding to the task;

[0429] Determine the fourth product of the first sample sub-loss function corresponding to the task, the distribution position corresponding to the task, and the prior weight corresponding to the task.

[0430] In this embodiment, the prior weight can be set in advance according to the needs, and this embodiment does not limit this. The fourth product is obtained by multiplying the first sample sub-loss function, the distribution position corresponding to the task, and the prior weight corresponding to the task.

[0431] The advantage of this embodiment is that, in addition to the first sample sub-loss function, the fourth product can also be determined jointly by the prior weight and the distribution position, so that the contribution of the first sample sub-loss function to the first sample loss function is not completely determined by the size of the distribution position, thereby ensuring that the first sample sub-loss function of some tasks has a certain contribution to determining the first sample loss function.

[0432] In step 2220, the calculation process of determining the first sample loss function is shown in Formula 11:

[0433]

[0434] In formula 11, refers to the first sample loss function, refers to the first sample sub-loss function corresponding to the t-th task, T refers to the number of tasks, t refers to the t-th task, η t refers to the distribution position, γ t refers to the prior weight.

[0435] The advantage of the embodiment of steps 2210-2220 is that the first sample sub-loss function corresponding to each task contributes the same to determining the first sample loss function, thereby improving fairness for multiple tasks when training the model.

[0436] In one embodiment, referring to Figure 23 The step 1750 of training the fusion parameter determination model based on the first sample loss function includes:

[0437] Step 2310: Determine a second sample loss function corresponding to the task based on the second probability corresponding to the task;

[0438] Step 2320: Determine a total loss function based on the first sample loss function corresponding to the task and the second sample loss function corresponding to the task;

[0439] Step 2330: Based on the total loss function, train the fusion parameter to determine the model.

[0440] In step 2310, based on the training labels and the second probability corresponding to the task, a second sample sub-loss function for each task is determined. In one embodiment, the calculation process for determining the second sample sub-loss function is shown in Formula 12:

[0441] L t =-y t logp t -(1-y t )log(1-p t )(Formula 12).

[0442] In formula 12, L t is the second sample loss function of the t-th task, y t is the training label of the t-th task, p t is the second probability of the tth task. If the training label indicates that the task is completed, then the training label is 1, otherwise, the training label is 0.

[0443] The second sample sub-loss functions of multiple tasks are added together to obtain the second sample loss function. In one embodiment, the calculation process for determining the second sample loss function is shown in Formula 13:

[0444]

[0445] In Formula 13, L refers to the second sample loss function, T refers to the number of tasks, and L t is the second sample loss function of the t-th task.

[0446] In step 2320, the first sample loss function and the second sample loss function may be added or weighted together to obtain a total loss function. In one embodiment, the calculation process for determining the total loss function is shown in Formula 14:

[0447]

[0448] In formula 14, refers to the total loss function, L refers to the second sample loss function, refers to the first sample loss function.

[0449] In one embodiment, step 2320 includes:

[0450] Determine a total sample loss function of the target sample based on the first sample loss function and the second sample loss function;

[0451] Based on the total sample loss function of each target sample in the target sample set, a total loss function is determined.

[0452] In this embodiment, not only one target sample is considered, but also the remaining target samples in the target sample set. The first sample loss function is added to the second sample loss function to obtain the total sample loss function of the target sample. Each target sample corresponds to a total sample loss function. The total sample loss function of multiple target samples is added to obtain the total loss function. For example, if the total sample loss function of target sample B1 is 0.3 and the total sample loss function of target sample B2 is 0.1, then the total loss function is 0.3 + 0.1 = 0.4.

[0453] The advantage of this embodiment is that the total sample loss function of each target sample contributes to the total loss function, and the greater the number of target samples, the better the training effect of the training model based on the total loss function.

[0454] In step 2330, after determining the total loss function, the fusion parameter determination model is trained based on the total loss function. The purpose of training is to reduce the total loss function. A lower total loss function indicates more accurate model output results. Therefore, the training process is to update the parameters in the fusion parameter determination model so that the updated model will reduce the final total loss function. Gradient descent can be used to update the parameters of the fusion parameter determination model in the direction of the gradient descent of the total loss function.

[0455] It should be understood that in one embodiment, after determining the total loss function, the parameters in the fusion parameter determination model, the monotonic transformation coefficients corresponding to each task, the monotonic transformation offset corresponding to each task, the first basic scalar and the second basic scalar are updated based on the total loss function, so that the updated model will reduce the final generated total loss function.

[0456] The advantage of the embodiment of steps 2310-2330 is that the model can be trained based on the first sample loss function and the second sample loss function, thereby improving the accuracy of the trained fusion parameter determination model.

[0457] It should be noted that when training the fusion parameters to determine the model, truncating the training gradient of the fusion part (stop gradient) does not affect the training of the multi-task prediction model itself. When the model is actually used, the fusion probability is used for push, and there is no need to calculate the first fusion score, second fusion score, etc. in subsequent steps.

[0458] Schematic diagram of the overall implementation of the push processing method of the embodiment of the present disclosure

[0459] Refer to the following Figure 24 , which illustrates in detail the implementation details of the push processing method of the embodiment of the present disclosure.

[0460] (1) Obtain the sample object features of the sample object and the sample scene features of the scene in which the sample object is located; take the features of each sample content in the sample content set as the sample content features in turn; take the sample object features, sample scene features, and sample content features as multiple basic features for sample push. For example, the sample object features include the education level of college and the hobby of movies. The sample scene features include the current input day type of holiday and the current geographical area of ​​XX Cinema. The sample content features of the sample content include the content type of movie, the content length of 70 seconds, and the number of likes of the content of 1300.

[0461] (2) Multiple sample push basic features are input into the vectorization layer to obtain multiple sample push basic feature vectors. For example, the above-mentioned features such as the education level of college, the hobby of movies, the current input day type of holiday, the current geographical area of ​​XX theater, the content type of movie, the content length of 70 seconds, and the number of likes of the content of 1300 are input into the vectorization layer to obtain 7 push basic feature vectors.

[0462] (3) Inputting multiple sample push basic feature vectors into the second splicing layer to obtain a sample cascade push basic feature vector. For example, inputting the above 7 sample push basic feature vectors into the second splicing layer to obtain a sample cascade push basic feature vector.

[0463] (4) Input the sample cascade push basic feature vector into the expert sub-model to obtain the model output of the expert sub-model; use the gating weight corresponding to the gating node to perform a weighted sum of the cascade push basic feature vector and the model output of multiple expert sub-models to obtain the weighted model output; input the weighted sum model output into the prediction sub-model to obtain the second probability of pushing the sample content for the sample object corresponding to the task. For example, the prediction sub-model on the left in the multi-task prediction model outputs the second probability P corresponding to task A. A The prediction sub-model on the left in the multi-task prediction model outputs the second probability P corresponding to task B B . Figure 24 Also shown is L A and L B . L A Refers to the second sample loss function corresponding to task A. L B Refers to the second sample sub-loss function corresponding to task B. The second sample loss function is obtained by summing the second sample sub-loss functions corresponding to multiple tasks.

[0464] (5) Inputting the second probabilities corresponding to the plurality of tasks into the first importance weight vector generation model to obtain a first importance weight vector; obtaining the first importance weight corresponding to the task from the first importance weight vector, and determining the first fusion parameter based on the maximum value of the first fusion parameter and the first importance weight corresponding to the task; inputting the second probabilities corresponding to the plurality of tasks into the second importance weight vector generation model to obtain a second importance weight vector; obtaining the second importance weight corresponding to the task from the second importance weight vector, and determining the second fusion parameter based on the maximum value of the second fusion parameter and the second importance weight corresponding to the task. The maximum value of the first fusion parameter is determined based on the first basic scalar, and the maximum value of the second fusion parameter is determined based on the second basic scalar.

[0465] (6) Determine the second probability corresponding to the task and the second sum of the first fusion parameter; determine the probability contribution value corresponding to the task based on the second sum corresponding to the task and the second fusion parameter; multiply the probability contribution values ​​corresponding to multiple tasks to obtain the fused probability. As shown in Formula 1, In Formula 1, T refers to the number of tasks, and t refers to the t-th task. t Refers to the second probability of the t-th task. t Refers to the first fusion parameter of the t-th task. max Refers to the maximum value of the first fusion parameter. t is the second fusion parameter of the t-th task. ∏(.) refers to the multiplication process.

[0466] (7) Obtain the monotonic transformation coefficient and monotonic transformation offset corresponding to the task; multiply the square of the monotonic transformation coefficient by the first fusion score to obtain a first product; determine the second fusion score based on the first product and the monotonic transformation offset. As shown in Formula 8, s t =a t ·a t ·s+b t In formula 8, s t refers to the second fusion score corresponding to the t-th task, a t refers to the monotonic transformation coefficient, b t It refers to the monotonic transformation offset, and s refers to the first fusion score.

[0467] (8) Based on the second fusion score corresponding to the task, determine the probability of the sample corresponding to the task after monotonic transformation. Referring to the above, refers to the probability of the sample of the t-th task after monotonic transformation, s t It refers to the second fusion score corresponding to the t-th task. Figure 24 Shows the probability of the sample corresponding to task A after monotonic transformation And the probability of the sample corresponding to task B after monotonic transformation Figure 24 It also shows and Refers to the first sample sub-loss function corresponding to task A. Refers to the first sample sub-loss function corresponding to task B. The first sample loss function is obtained by summing the first sample sub-loss functions corresponding to multiple tasks.

[0468] (9) Based on the first sample loss function and the second sample loss function, a total loss function is determined; based on the total loss function, a fusion parameter determination model, a first basic scalar, a second basic scalar, a monotonic transformation coefficient corresponding to each task, and a monotonic transformation offset corresponding to each task are jointly trained.

[0469] (10) After training to obtain the fusion parameter determination model, the first basic scalar, the second basic scalar, the monotonic transformation coefficients corresponding to each task, and the monotonic transformation offsets corresponding to each task, steps 310-350 are executed. To save space, they are not further described here.

[0470] Description of the apparatus and device of the present disclosure

[0471] It is to be understood that, although the steps in the above-mentioned flowcharts are shown in sequence according to the arrow representations, these steps are not necessarily performed in sequence according to the order represented by the arrows. Unless otherwise specified in the present embodiment, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the above-mentioned flowcharts may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0472] It should be noted that in each specific embodiment of the present application, when it comes to the need to perform relevant processing based on data related to the characteristics of the object, such as object attribute information or attribute information sets, the permission or consent of the object will be obtained first, and the collection, use and processing of such data will comply with relevant laws, regulations and standards. In addition, when the embodiment of the present application needs to obtain object attribute information, the object's separate permission or separate consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the object's separate permission or separate consent, the necessary object-related data for the normal operation of the embodiment of the present application will be obtained.

[0473] Figure 25 This is a schematic diagram of the structure of a push processing device 2500 provided in an embodiment of the present disclosure. The push processing device 2500 includes:

[0474] An acquiring unit 2510 is configured to acquire a plurality of basic push features, wherein the plurality of basic push features include target object features and features of content to be pushed;

[0475] The prediction unit 2520 is configured to input the plurality of push basic features into the multi-task prediction model to obtain a first probability of pushing the content to be pushed to the target object corresponding to the plurality of tasks;

[0476] A first determining unit 2530 is configured to determine a fusion parameter by using a fusion parameter determination model based on first probabilities corresponding to the plurality of tasks;

[0477] The second determining unit 2540 is configured to determine a fused probability of pushing the content to be pushed to the target object based on the first probabilities corresponding to the multiple tasks and the fusion parameter;

[0478] The push unit 2550 is used to push the content to be pushed to the target object based on the fused probability, wherein the fusion parameter determination model is trained based on the distribution position of the first sample probability corresponding to the task of each target sample in the target sample set in the first sample probability of each sample corresponding to the task, and the first sample probability is obtained by inputting the sample push basic feature of the target sample into the multi-task prediction model.

[0479] Optionally, the fusion parameter determination model is trained by:

[0480] Acquire a target sample set, where the target sample set includes multiple target samples, each target sample has multiple sample push basic features, and the multiple sample push basic features include sample object features and sample content features;

[0481] Inputting the multiple sample push basic features into the multi-task prediction model to obtain a second probability of pushing sample content for the sample object corresponding to the multiple tasks;

[0482] Inputting the second probabilities corresponding to the multiple tasks into the fusion parameter determination model to obtain sample fusion parameters, and determining the sample fusion probability of pushing the sample content to the sample object based on the second probabilities corresponding to the multiple tasks and the fusion parameters;

[0483] Obtaining a distribution position of a first sample probability of a target sample corresponding to the task in the first sample probabilities of each sample corresponding to the task;

[0484] Based on the probability and distribution position of the sample after fusion, the first sample loss function is calculated, and the fusion parameter determination model is trained based on the first sample loss function.

[0485] Optionally, calculating a first sample loss function based on the probability and distribution position of the sample after fusion includes:

[0486] Perform a monotonic transformation on the sample fusion probability to obtain the monotonic transformed probability of the sample corresponding to the task;

[0487] Determine the first sample sub-loss function corresponding to the task based on the monotonically transformed probability of the sample corresponding to the task;

[0488] Determine a first sample loss function based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task.

[0489] Optionally, performing a monotonic transformation on the sample fusion probability to obtain the monotonic transformed probability of the sample corresponding to the task includes:

[0490] Determine a first fusion score based on the probability of sample fusion;

[0491] Perform a monotonic transformation on the first fusion score to obtain the second fusion score corresponding to the task;

[0492] Based on the second fusion score corresponding to the task, the probability of the sample corresponding to the task after monotonic transformation is determined.

[0493] Optionally, determining a first fusion score based on the sample fusion probability includes:

[0494] Determine the first difference between 1 and the probability of sample fusion;

[0495] The logarithm of the first quotient of the probability of the sample after fusion and the first difference is determined as the first fusion score.

[0496] Optionally, performing a monotonic transformation on the first fusion score to obtain a second fusion score corresponding to the task includes:

[0497] Get the monotonic transformation coefficient and monotonic transformation offset corresponding to the task;

[0498] Based on the monotonic transformation coefficient and monotonic transformation offset corresponding to the task, the first fusion score is monotonically transformed to obtain the second fusion score corresponding to the task, wherein the monotonic transformation coefficient and monotonic transformation offset corresponding to each task are jointly trained with the fusion parameter determination model.

[0499] Optionally, performing a monotonic transformation on the first fused score based on a monotonic transformation coefficient and a monotonic transformation offset corresponding to the task to obtain a second fused score corresponding to the task includes:

[0500] Multiplying the square of the monotone transformation coefficient by the first fusion factor to obtain a first product;

[0501] A second fused score is determined based on the first product and the monotonic transformation offset.

[0502] Optionally, determining the probability of a sample corresponding to the task after monotonic transformation based on the second fusion score corresponding to the task includes:

[0503] determining a first power value with the first constant as base and the opposite of the second fused fraction as exponent;

[0504] determining a first sum of the first power value and a second constant;

[0505] The reciprocal of the first sum is taken as the probability of the sample corresponding to the task after monotonous transformation.

[0506] Optionally, determining a first sample sub-loss function corresponding to the task based on the probability of the sample corresponding to the task after monotonic transformation includes:

[0507] Determine the first logarithm of the monotonically transformed probability of the sample corresponding to the task and the second product of the first logarithm and the second probability corresponding to the task;

[0508] Determine a first complement of the second probability and a second complement of the probability after the monotonically transformed sample, wherein the first complement is the difference between 1 and the second probability, and the second complement is the difference between 1 and the probability after the monotonically transformed sample;

[0509] Determine the third product of the second logarithm of the second complement and the first complement;

[0510] Based on the second product and the third product, the first sample sub-loss function corresponding to the task is determined.

[0511] Optionally, determining the first sample loss function based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task includes:

[0512] Determine the fourth product of the first sample sub-loss function corresponding to the task and the distribution position corresponding to the task;

[0513] The fourth products of multiple tasks are added together to obtain the first sample loss function.

[0514] Optionally, determining a fourth product of a first sample sub-loss function corresponding to the task and a distribution position corresponding to the task includes:

[0515] Get the prior weight corresponding to the task;

[0516] Determine the fourth product of the first sample sub-loss function corresponding to the task, the distribution position corresponding to the task, and the prior weight corresponding to the task.

[0517] Optionally, training a fusion parameter determination model based on the first sample loss function includes:

[0518] Determining a second sample loss function based on second probabilities corresponding to the multiple tasks;

[0519] Determine a total loss function based on the first sample loss function and the second sample loss function;

[0520] Based on the total loss function, the fusion parameters are trained to determine the model.

[0521] Optionally, determining a total loss function based on the first sample loss function and the second sample loss function includes:

[0522] Determine a total sample loss function of the target sample based on the first sample loss function and the second sample loss function;

[0523] Based on the total sample loss function of each target sample in the target sample set, a total loss function is determined.

[0524] Optionally, the distribution position is a multiple of a standard deviation of the first sample probability of the target sample corresponding to the task, compared to the average of the first sample probabilities of the samples corresponding to the task;

[0525] Obtaining a distribution position of a first sample probability of a target sample corresponding to the task in the first sample probabilities of each sample corresponding to the task includes:

[0526] Get multiple samples corresponding to the task;

[0527] Determine the mean and standard deviation of the first sample probabilities of the plurality of samples corresponding to the task;

[0528] Determine a first sample probability of the target sample corresponding to the task and a second difference from the mean value;

[0529] The second quotient of the second difference and the standard deviation is determined as the distribution position.

[0530] Optionally, the fusion parameter includes a first fusion parameter and a second fusion parameter; the second determining unit is specifically configured to:

[0531] determining a second sum of a first probability and a first fusion parameter corresponding to the task;

[0532] Determining a probability contribution value corresponding to the task based on the second sum corresponding to the task and the second fusion parameter;

[0533] Multiply the probability contribution values ​​corresponding to multiple tasks to obtain the fused probability.

[0534] Optionally, determining a probability contribution value corresponding to the task based on the second sum corresponding to the task and a second fusion parameter includes:

[0535] Determine the third sum of 1 and the maximum value of the first fusion parameter;

[0536] Determine the third quotient of the second sum and the third sum;

[0537] Determine a power value with the third quotient as the base and the second fusion parameter as the exponent as the probability contribution value corresponding to the task.

[0538] Optionally, the first fusion parameter is between 1 and a maximum value of the first fusion parameter, and the second fusion parameter is between 1 and a maximum value of the second fusion parameter; the maximum value of the first fusion parameter and the maximum value of the second fusion parameter are determined in the following manner:

[0539] Obtain a first basic scalar corresponding to the first fusion parameter, a second basic scalar corresponding to the second fusion parameter, and a third constant greater than 1;

[0540] performing exponential normalization on the first basic scalar to obtain a first exponential normalized value;

[0541] performing exponential normalization on the second basic scalar to obtain a second exponential normalized value;

[0542] determining a first fusion parameter maximum value based on a third constant and the first exponential normalized value;

[0543] Based on the third constant and the second exponential normalized value, a second fusion parameter maximum value is determined, wherein the first basic scalar and the second basic scalar are jointly trained with the fusion parameter determination model.

[0544] Optionally, the fusion parameter includes a first fusion parameter and a second fusion parameter, the first fusion parameter is between 1 and a maximum value of the first fusion parameter, the second fusion parameter is between 1 and a maximum value of the second fusion parameter, and the fusion parameter determination model includes a first importance weight vector generation model and a second importance weight vector generation model;

[0545] The first determining unit 2530 is specifically configured to:

[0546] Inputting first probabilities corresponding to the plurality of tasks into a first importance weight vector generation model to obtain a first importance weight vector;

[0547] Obtaining a first importance weight corresponding to the task from the first importance weight vector, and determining a first fusion parameter based on a first fusion parameter maximum value and the first importance weight corresponding to the task;

[0548] Inputting the first probabilities corresponding to the plurality of tasks into a second importance weight vector generation model to obtain a second importance weight vector;

[0549] The second importance weight corresponding to the task is obtained from the second importance weight vector, and the second fusion parameter is determined based on the second fusion parameter maximum value and the second importance weight corresponding to the task.

[0550] Optionally, the first importance weight vector generation model includes a first compression model and a first stretching model;

[0551] The first determining unit 2530 is further configured to:

[0552] generating a first probability vector based on first probabilities corresponding to the plurality of tasks;

[0553] Inputting the first probability vector into a first compression model to obtain a first compression vector, where the dimension of the first compression vector is smaller than the number of tasks;

[0554] The first compression vector is input into the first stretching model to obtain a first importance weight vector, where the dimension of the first importance weight vector is equal to the number of tasks.

[0555] Optionally, the first compression model includes a first matrix and a nonlinear activation function; the first determining unit 2530 is further configured to:

[0556] Determining a first product vector based on the first matrix and the first probability vector;

[0557] A nonlinear activation function is applied to the first product vector to obtain a first compressed vector.

[0558] Optionally, the first stretching model includes a second matrix and an exponential normalization function; the first determining unit 2530 is further configured to:

[0559] determining a second product vector based on the second matrix and the first compressed vector;

[0560] Applying an exponential normalization function to the second product vector yields a first importance weight vector.

[0561] Optionally, the second importance weight vector generation model includes a second compression model and a second stretching model; the first determination unit 2530 is specifically configured to:

[0562] generating a first probability vector based on first probabilities corresponding to the plurality of tasks;

[0563] Inputting the first probability vector into a second compression model to obtain a second compression vector, wherein the dimension of the second compression vector is smaller than the number of tasks;

[0564] The second compression vector is input into the second stretching model to obtain a second importance weight vector, where the dimension of the second importance vector is equal to the number of tasks.

[0565] Optionally, the second compression model includes a third matrix and a nonlinear activation function; the first determining unit 2530 is further configured to:

[0566] determining a third product vector based on the third matrix and the first probability vector;

[0567] A nonlinear activation function is applied to the third product vector to obtain a second compressed vector.

[0568] Optionally, the second stretching model includes a fourth matrix and an exponential normalization function; the first determining unit 2530 is further configured to:

[0569] determining a fourth product vector based on the fourth matrix and the second compressed vector;

[0570] An exponential normalization function is applied to the fourth product vector to obtain a second importance weight vector.

[0571] Reference Figure 26 , Figure 26The following is a block diagram of the structure of a terminal that implements the push processing method according to an embodiment of the present disclosure. The terminal includes: a radio frequency (RF) circuit 2610, a memory 2615, an input unit 2630, a display unit 2640, a sensor 2650, an audio circuit 2660, a wireless fidelity (WiFi) module 2670, a processor 2680, and a power supply 2690. It will be understood by those skilled in the art that Figure 26 The terminal structure shown does not constitute a limitation on the mobile phone or computer, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0572] The RF circuit 2610 may be used for receiving and sending signals during information transmission or calls. In particular, after receiving downlink information from the base station, it is sent to the processor 2680 for processing. In addition, the designed uplink data is sent to the base station.

[0573] The memory 2615 may be used to store software programs and modules. The processor 2680 executes the software programs and modules stored in the memory 2615 to perform various functional applications and data processing of the content terminal.

[0574] The input unit 2630 may be configured to receive input digital or character information and generate key signal input related to the settings and function control of the content terminal. Specifically, the input unit 2630 may include a touch panel 2631 and other input devices 2632.

[0575] The display unit 2640 may be configured to display input information or provided information and various menus of the content terminal. The display unit 2640 may include a display panel 2641.

[0576] The audio circuit 2660 , the speaker 2661 , and the microphone 2662 may provide an audio interface.

[0577] In this embodiment, the processor 2680 included in the terminal can execute the push processing method of the previous embodiment.

[0578] The terminals of the embodiments of the present disclosure include but are not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc. The embodiments of the present invention can be applied to various scenarios, including but not limited to content recommendation, data screening, etc.

[0579] Figure 27A structural block diagram of part of a server for implementing the push processing method of an embodiment of the present disclosure. The server may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPUs) 2722 (for example, one or more processors) and memories 2732, and one or more storage media 2730 (for example, one or more mass storage devices) for storing application programs 2742 or data 2744. Among them, the memories 2732 and the storage media 2730 can be short-term storage or persistent storage. The program stored in the storage medium 2730 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the server. Furthermore, the central processing unit 2722 can be configured to communicate with the storage medium 2730 to execute a series of instruction operations in the storage medium 2730 on the server.

[0580] The server may also include one or more power supplies 2726, one or more wired or wireless network interfaces 2750, one or more input and output interfaces 2758, and / or one or more operating systems 2741, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0581] The central processor 2722 in the server can be used to execute the push processing method of the embodiment of the present disclosure.

[0582] The embodiments of the present disclosure further provide a computer-readable storage medium, which is used to store program codes, and the program codes are used to execute the push processing methods of the aforementioned embodiments.

[0583] The present disclosure also provides a computer program product, which includes a computer program. A processor of a computer device reads and executes the computer program, so that the computer device executes the above-mentioned push processing method.

[0584] The terms "first," "second," "third," "fourth," and the like (if any) in the specification of the present disclosure and the accompanying drawings are used to distinguish between similar contents and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present disclosure described herein, for example, can be implemented in orders other than those illustrated or described herein. In addition, the terms "comprises" and "comprising," and any variations thereof, are intended to cover non-exclusive inclusions, e.g., a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0585] It should be understood that in the present disclosure, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated content, indicating that there can be three relationships. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and following associated content is in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0586] It should be understood that in the description of the embodiments of the present disclosure, the meaning of multiple (or multiple items) is more than two, greater than, less than, exceed, etc. are understood to exclude the number itself, and above, below, within, etc. are understood to include the number itself.

[0587] In the several embodiments provided in the present disclosure, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0588] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0589] In addition, the functional units in the various embodiments of the present disclosure may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0590] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, server 130, or network device, etc.) to execute all or part of the steps of the various embodiments of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0591] It should also be understood that the various implementations provided in the embodiments of the present disclosure can be combined arbitrarily to achieve different technical effects.

[0592] The above is a specific description of the implementation methods of the present disclosure, but the present disclosure is not limited to the above implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present disclosure. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present disclosure.

Claims

1. A push processing method, characterized in that: include: Acquire a plurality of basic push features, wherein the plurality of basic push features include target object features and to-be-pushed content features; Inputting the plurality of push basic features into a multi-task prediction model to obtain first probabilities corresponding to the plurality of tasks for pushing the content to be pushed to the target object; determining a fusion parameter by a fusion parameter determination model based on the first probabilities corresponding to the plurality of tasks; Determining a fused probability of pushing the to-be-pushed content to the target object based on the first probabilities corresponding to the plurality of tasks and the fusion parameter; Based on the fused probability, the content to be pushed is pushed to the target object, wherein the fusion parameter determination model is trained based on the distribution position of the first sample probability corresponding to the task for each target sample in the target sample set in the first sample probability of each sample corresponding to the task, and the first sample probability is obtained by inputting the sample push basic features of the target sample into the multi-task prediction model.

2. The push processing method according to claim 1, characterized in that: The fusion parameter determination model is trained in the following way: Acquire the target sample set, where the target sample set includes a plurality of target samples, each of the target samples has a plurality of sample push basic features, and the plurality of sample push basic features include a sample object feature and a sample content feature; Inputting the plurality of sample push basic features into the multi-task prediction model to obtain second probabilities corresponding to the plurality of tasks for pushing the sample content to the sample object; Inputting the second probabilities corresponding to the plurality of tasks into the fusion parameter determination model to obtain sample fusion parameters, and determining a sample fusion probability of pushing the sample content to the sample object based on the second probabilities corresponding to the plurality of tasks and the fusion parameters; Obtaining a distribution position of a first sample probability of the target sample corresponding to the task in the first sample probabilities of each sample corresponding to the task; Based on the sample fusion probability and the distribution position, a first sample loss function is calculated, and the fusion parameter determination model is trained based on the first sample loss function.

3. The push processing method according to claim 2, characterized in that: The calculating a first sample loss function based on the sample fusion probability and the distribution position includes: Performing a monotonic transformation on the sample fusion probability to obtain the sample monotonic transformation probability corresponding to the task; Determining a first sample sub-loss function corresponding to the task based on the monotonically transformed probability of the sample corresponding to the task; Determine the first sample loss function based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task.

4. The push processing method according to claim 3, characterized in that: The monotonically transforming the sample fusion probability to obtain the monotonically transformed sample probability corresponding to the task includes: Determining a first fusion score based on the sample fusion probability; Performing a monotonic transformation on the first fusion score to obtain a second fusion score corresponding to the task; Based on the second fusion score corresponding to the task, a probability of the sample corresponding to the task after monotonic transformation is determined.

5. The push processing method according to claim 4, characterized in that: The determining of a first fusion score based on the sample fusion probability includes: Determine a first difference between 1 and the probability of the sample after fusion; The logarithm of a first quotient of the sample fusion probability and the first difference is determined as the first fusion score.

6. The push processing method according to claim 4, characterized in that: The monotonically transforming the first fusion score to obtain a second fusion score corresponding to the task includes: Obtaining a monotonic transformation coefficient and a monotonic transformation offset corresponding to the task; Based on the monotonic transformation coefficient and monotonic transformation offset corresponding to the task, the first fusion score is monotonically transformed to obtain the second fusion score corresponding to the task, wherein the monotonic transformation coefficient and monotonic transformation offset corresponding to each task are jointly trained with the fusion parameter determination model.

7. The push processing method according to claim 6, characterized in that: The monotonic transformation of the first fused score based on the monotonic transformation coefficient and the monotonic transformation offset corresponding to the task to obtain the second fused score corresponding to the task includes: multiplying the square of the monotonic transformation coefficient by the first fusion score to obtain a first product; The second fused score is determined based on the first product and the monotonic transformation offset.

8. The push processing method according to claim 4, characterized in that: The determining, based on the second fusion score corresponding to the task, the probability of the sample corresponding to the task after monotonic transformation, includes: determining a first power value with a first constant as a base and a negation of the second fused component as an exponent; determining a first sum of the first power value and a second constant; The reciprocal of the first sum is used as the probability of the sample corresponding to the task after monotonically transformed.

9. The push processing method according to claim 3, characterized in that: The determining, based on the monotonically transformed probability of the sample corresponding to the task, a first sample sub-loss function corresponding to the task, includes: Determine a second product of a first logarithm of a probability of the sample corresponding to the task after monotonically transformed and the second probability corresponding to the task; Determine a first complement of the second probability and a second complement of the probability of the sample after monotonically transformed, wherein the first complement is the difference between 1 and the second probability, and the second complement is the difference between 1 and the probability of the sample after monotonically transformed; determining a third product of a second logarithm of the second complement and the first complement; Based on the second product and the third product, a first sample sub-loss function corresponding to the task is determined.

10. The push processing method according to claim 3, characterized in that: The determining the first sample loss function based on the distribution position corresponding to the task and the first sample sub-loss function corresponding to the task includes: Determine a fourth product of the first sample sub-loss function corresponding to the task and the distribution position corresponding to the task; The fourth products of the plurality of tasks are added together to obtain the first sample loss function.

11. The push processing method according to claim 2, characterized in that: The step of training the fusion parameter determination model based on the first sample loss function includes: Determining a second sample loss function based on the second probabilities corresponding to the plurality of tasks; Determining a total loss function based on the first sample loss function and the second sample loss function; Based on the total loss function, the fusion parameter determination model is trained.

12. The push processing method according to claim 11, characterized in that: The determining of a total loss function based on the first sample loss function and the second sample loss function includes: Determining a total sample loss function of the target sample based on the first sample loss function and the second sample loss function; The total loss function is determined based on the total sample loss function of each target sample in the target sample set.

13. The push processing method according to claim 2, characterized in that: The distribution position is a multiple of a standard deviation of a first sample probability of the target sample corresponding to the task, compared to an average of the first sample probabilities of the samples corresponding to the task; The obtaining of the distribution position of the first sample probability of the target sample corresponding to the task in the first sample probabilities of each sample corresponding to the task includes: Acquire multiple samples corresponding to the task; Determining an average value and a standard deviation of the first sample probabilities of the plurality of samples corresponding to the task; Determine a second difference between the first sample probability of the target sample corresponding to the task and the average value; A second quotient of the second difference and the standard deviation is determined as the distribution position.

14. The push processing method according to claim 1, characterized in that: The fusion parameters include a first fusion parameter and a second fusion parameter; The determining, based on the first probabilities corresponding to the plurality of tasks and the fusion parameter, a fused probability of pushing the content to be pushed to the target object includes: determining a second sum of the first probability and the first fusion parameter corresponding to the task; determining a probability contribution value corresponding to the task based on the second sum corresponding to the task and the second fusion parameter; The probability contribution values ​​corresponding to the multiple tasks are multiplied to obtain the fused probability.

15. The push processing method according to claim 14, characterized in that: The determining, based on the second sum corresponding to the task and the second fusion parameter, a probability contribution value corresponding to the task, includes: Determine the third sum of 1 and the maximum value of the first fusion parameter; determining a third quotient of the second sum and the third sum; A power value having the third quotient as a base and the second fusion parameter as an exponent is determined as the probability contribution value corresponding to the task.

16. The push processing method according to claim 15, characterized in that: The first fusion parameter is between 1 and the maximum value of the first fusion parameter, and the second fusion parameter is between 1 and the maximum value of the second fusion parameter. The maximum value of the first fusion parameter and the maximum value of the second fusion parameter are determined in the following manner: Obtaining a first basic scalar corresponding to the first fusion parameter, a second basic scalar corresponding to the second fusion parameter, and a third constant greater than 1; performing exponential normalization on the first basic scalar to obtain a first exponential normalized value; performing exponential normalization on the second basic scalar to obtain a second exponential normalized value; determining a maximum value of the first fusion parameter based on the third constant and the first exponential normalized value; Based on the third constant and the second exponential normalization value, a second fusion parameter maximum value is determined, wherein the first basic scalar and the second basic scalar are jointly trained with the fusion parameter determination model.

17. A push processing device, characterized in that: include: An acquisition unit, configured to acquire a plurality of basic push features, wherein the plurality of basic push features include target object features and features of content to be pushed; A prediction unit, configured to input the plurality of push basic features into a multi-task prediction model to obtain a first probability of pushing the content to be pushed to the target object corresponding to the plurality of tasks; a first determining unit, configured to determine a fusion parameter by a fusion parameter determination model based on the first probabilities corresponding to the plurality of tasks; a second determining unit, configured to determine a fused probability of pushing the content to be pushed to the target object based on the first probabilities corresponding to the plurality of tasks and the fusion parameter; A pushing unit, configured to push the content to be pushed to the target object based on the fused probability; Among them, the fusion parameter determination model is trained based on the distribution position of the first sample probability corresponding to the task of each target sample in the target sample set in the first sample probability of each sample corresponding to the task, and the first sample probability is obtained by inputting the sample push basic features of the target sample into the multi-task prediction model.

18. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the push processing method according to any one of claims 1 to 16 is implemented.

19. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the push processing method according to any one of claims 1 to 16 is implemented.

20. A computer program product, comprising a computer program, wherein the computer program is read and executed by a processor of a computer device, so that the computer device executes the push processing method according to any one of claims 1 to 16.

Images