Data processing method, device and equipment and readable storage medium

Abstract

The invention discloses a data processing method, device and equipment and a readable storage medium. The method comprises the following steps: acquiring sample data for model training; the sample data is determined based on the sample object and the historical resource data of the historical business operation executed by the sample object; calling a recommendation model to carry out probability prediction processing on the sample data to obtain a sample prediction probability; monotonic transformation processing is carried out on the sample prediction probability through the monotonic transformation parameter of each business index, and a sample transformation probability corresponding to each business index is obtained; performing training optimization on the recommendation model through the N sample transformation probabilities; the trained and optimized recommendation model is used for carrying out probability prediction processing on the service data in a resource data pushing scene. The resource data recommendation method and device can be applied to various scenes such as the map field, the traffic field, the automatic driving field, the vehicle-mounted scene, the cloud technology, artificial intelligence, intelligent traffic and auxiliary driving, and the resource data recommendation accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular, to a data processing method, apparatus, device, and readable storage medium. Background Art

[0002] Multi-objective ranking models are widely used in recommendation systems. The multi-objective ranking model obtains additional benefits that cannot be obtained by single-objective models by simultaneously optimizing the effects of multiple business objectives (or business metrics), such as click-through rate, conversion rate, and click-through conversion rate, etc. The multi-objective ranking model includes a multi-objective model and a multi-objective fusion model. Through the multi-objective model, the estimated probability corresponding to each business objective can be output. Through the multi-objective fusion model, the estimated probabilities of each business objective can be fused to obtain the final ranking score of the candidate resource data. Based on this ranking score, each candidate resource data can be ranked, and then the resource data to be finally pushed to the user is selected from the ranked sequence.

[0003] It can be seen that in the architecture of the recommendation system, the multi-objective model and the multi-objective fusion model are relatively independent. The multi-objective model and the multi-objective fusion model break down the ranking problem into two stages. Through the training and optimization of the models, although the multi-objective model can achieve better training effects for individual business objectives, the improvement of the effect of a single business objective often does not correspondingly improve the effect after multi-objective fusion. Even because some objectives are mutually exclusive, after the respective effects are improved through respective training, the effect after multi-objective fusion will instead decrease. The ranking effect of the multi-objective model is not unified with the ranking effect of multi-objective fusion, reducing the accuracy of the final ranking result. Summary of the Invention

[0004] Embodiments of the present application provide a data processing method, apparatus, device, and readable storage medium, which can optimize the data ranking effect and improve the accuracy of data recommendation in the recommendation service.

[0005] On the one hand, an embodiment of the present application provides a data processing method, including:

[0006] Obtain sample data for model training; the sample data is determined based on sample objects and historical resource data on which the sample objects have performed historical business operations;

[0007] Call a recommendation model to perform probability prediction processing on the sample data to obtain sample prediction probabilities; the sample prediction probabilities refer to the probabilities of predicting that the sample objects perform business operations jointly indicated by N business metrics; N is a positive integer;

[0008] Perform monotonic transformation processing on the sample prediction probabilities respectively through the monotonic transformation parameters of each business metric, to obtain the sample transformation probabilities corresponding to each business metric respectively; a sample transformation probability is used to represent the probability that the sample object performs the business operation indicated by the corresponding business metric alone.

[0009] Train and optimize the recommendation model through N sample transformation probabilities; the trained and optimized recommendation model is used to perform probability prediction processing on business data in the resource data push scenario; the business data is determined based on the business object and the candidate push resource data of the business object.

[0010] An embodiment of the present application provides a data processing device on the one hand, including:

[0011] A sample acquisition module, configured to acquire sample data for model training; the sample data is determined based on the sample object and the historical resource data of the historical business operations performed by the sample object.

[0012] A prediction module, configured to call the recommendation model to perform probability prediction processing on the sample data, to obtain sample prediction probabilities; the sample prediction probability refers to the probability that the predicted sample object performs the business operation indicated jointly by N business metrics; N is a positive integer.

[0013] A monotonic transformation module, configured to perform monotonic transformation processing on the sample prediction probabilities respectively through the monotonic transformation parameters of each business metric, to obtain the sample transformation probabilities corresponding to each business metric respectively; a sample transformation probability is used to represent the probability that the sample object performs the business operation indicated by the corresponding business metric alone.

[0014] A training and optimization module, configured to train and optimize the recommendation model through N sample transformation probabilities; the trained and optimized recommendation model is used to perform probability prediction processing on business data in the resource data push scenario; the business data is determined based on the business object and the candidate push resource data of the business object.

[0015] In one embodiment, the specific implementation manner for the sample acquisition module to acquire sample data for model training includes:

[0016] Acquire the feature data of the sample object and the feature data of the historical resource data; the historical resource data refers to the resource data of the historical business operations performed by the sample object; the feature data of the sample object includes the feature data of N1 first feature domains, and the feature data of the historical resource data includes the feature data of N2 second feature domains; both N1 and N2 are positive integers.

[0017] Acquire the first feature transformation vector corresponding to each first feature domain.

[0018] Acquire the second feature transformation vector corresponding to each second feature domain.

[0019] Concatenate N1 first feature transformation vectors and N2 second feature transformation vectors to obtain sample data.

[0020] Among them, the method for obtaining the feature transformation vector includes: performing fixed-length vector transformation on the feature data in the feature domain to obtain the feature transformation vector.

[0021] In one embodiment, the specific implementation manner for the sample acquisition module to perform fixed-length vector transformation on the feature data in the feature domain to obtain the feature transformation vector includes:

[0022] Obtain the domain type to which the feature domain belongs;

[0023] When the domain type is a fixed-length type, perform vector embedding processing on the feature data in the feature domain through a feature embedding network to obtain the feature transformation vector of the feature domain;

[0024] When the domain type is a variable-length type, perform vector mapping processing and pooling processing on the feature data in the feature domain through a sequence network to obtain the feature transformation vector of the feature domain.

[0025] In one embodiment, the specific implementation manner for the monotonic transformation module to perform monotonic transformation processing on the sample prediction probability respectively through the monotonic transformation parameters of each business indicator to obtain the sample transformation probability corresponding to each business indicator includes:

[0026] Determine any one of the N business indicators as the transformation business indicator, and determine the monotonic transformation parameter of the transformation business indicator as the target monotonic transformation parameter;

[0027] Through the scaling transformation function in the monotonic transformation rule, perform scaling transformation on the target monotonic transformation parameter and the sample prediction probability to obtain the initial sample transformation probability of the transformation business indicator;

[0028] Through the activation transformation function in the monotonic transformation rule, perform activation transformation on the initial sample transformation probability to obtain the sample transformation probability corresponding to the transformation business indicator.

[0029] In one embodiment, the target monotonic transformation parameter includes a first trainable parameter and a second trainable parameter;

[0030] The specific implementation manner for the monotonic transformation module to perform scaling transformation on the target monotonic transformation parameter and the sample prediction probability through the scaling transformation function in the monotonic transformation rule to obtain the initial sample transformation probability of the transformation business indicator includes:

[0031] Perform a multiplication operation on the first trainable parameter and the sample prediction probability to obtain a first multiplication result;

[0032] Multiply the first product result by the first trainable parameter to obtain a second product result;

[0033] Sum the second product result and the second trainable parameter to obtain the initial sample transformation probability for the transformed service metric.

[0034] In one embodiment, the specific implementation of the training and optimization module for training and optimizing the recommendation model through N sample transformation probabilities includes:

[0035] Obtain the cross-entropy loss function;

[0036] Obtain the prediction error value corresponding to each sample transformation probability through the cross-entropy loss function;

[0037] Sum the prediction error values corresponding to the N sample transformation probabilities to obtain the total loss value corresponding to the N sample transformation probabilities;

[0038] Train and optimize the model parameters of the recommendation model based on the total loss value.

[0039] In one embodiment, the specific implementation of the training and optimization module for obtaining the prediction error value corresponding to each sample transformation probability through the cross-entropy loss function includes:

[0040] Determine any one of the N sample transformation probabilities as the target sample transformation probability, and determine the service metric corresponding to the target sample transformation probability as the target service metric;

[0041] Based on the historical service operations performed by the sample object on the historical resource data, obtain the true sample label of the sample data under the target service metric;

[0042] Through the cross-entropy loss function, perform error calculation processing on the target sample transformation probability and the true sample label of the sample data under the target service metric to obtain the prediction error value corresponding to the target sample transformation probability.

[0043] In one embodiment, the specific implementation of the training and optimization module for summing the prediction error values corresponding to the N sample transformation probabilities to obtain the total loss value corresponding to the N sample transformation probabilities includes:

[0044] Perform a bias operation on the prediction error value corresponding to each sample transformation probability to obtain N bias error values;

[0045] Sum the N bias error values to obtain the total loss value corresponding to the N sample transformation probabilities.

[0046] In one embodiment, the specific implementation manner in which the training optimization module performs a bias operation on the prediction error values corresponding to each sample transformation probability to obtain N bias error values includes:

[0047] Determine any one of the N sample transformation probabilities as the target sample transformation probability, and determine the service metric corresponding to the target sample transformation probability as the target service metric;

[0048] Obtain the weight coefficient configured for the target service metric; during the process of training and optimizing the recommendation model, the weight coefficient of the target service metric is used to control the degree to which the output result of the recommendation model biases towards the target service metric;

[0049] Perform a multiplication operation on the weight coefficient of the target service metric and the target sample transformation probability to obtain the bias error value corresponding to the target sample transformation probability.

[0050] In one embodiment, the specific implementation manner in which the training optimization module trains and optimizes the model parameters of the recommendation model based on the total loss value includes:

[0051] Obtain the model iteration count of the recommendation model;

[0052] When the model iteration count does not meet the model convergence condition, train and optimize the model parameters of the recommendation model based on the total loss value.

[0053] In one embodiment, after the training optimization module trains and optimizes the recommendation model through N sample transformation probabilities, the data processing device further includes:

[0054] Obtain M candidate push resource data regarding the business object; M is a positive integer;

[0055] Obtain the service data corresponding to each of the M candidate push resource data; the service data corresponding to a candidate push resource data is determined based on the business object and the candidate push resource data;

[0056] Call the trained and optimized recommendation model to perform probability prediction processing on the service data corresponding to each candidate push resource data respectively, to obtain the prediction probability corresponding to each candidate push resource data respectively;

[0057] Based on the M prediction probabilities, determine the push resource data of the business object from the M candidate push resource data, and push the push resource data to the business object.

[0058] In one embodiment, the specific implementation manner in which the data push module 18 determines the push resource data of the business object based on the M prediction probabilities and pushes the push resource data to the business object includes:

[0059] Sort the M candidate push resource data in descending order according to the magnitude order among the M prediction probabilities to obtain a resource data sequence;

[0060] Determine the first K candidate push resource data in the resource data sequence as the push resource data of the service object; K is a positive integer less than or equal to M.

[0061] One aspect of the embodiments of the present application provides a computer device, including: a processor and a memory;

[0062] The memory stores a computer program. When the computer program is executed by the processor, the processor is caused to execute the method in the embodiments of the present application.

[0063] One aspect of the embodiments of the present application provides a computer-readable storage medium. The computer-readable storage medium stores a computer program. The computer program includes program instructions. When the program instructions are executed by the processor, the method in the embodiments of the present application is executed.

[0064] One aspect of the present application provides a computer program product. The computer program product includes a computer program. The computer program is stored in a computer-readable storage medium. The processor of the computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in one aspect of the embodiments of the present application.

[0065] In an embodiment of the present application, an end-to-end multi-objective efficient modeling method with monotonic shared unique sorting is provided, which can improve the consistency of multi-objective training and fusion, optimize the data sorting effect, and improve the accuracy of recommendation results. Specifically, after obtaining the sample data for model training, the present application can use a recommendation model to predict the sample data and output a sample prediction probability, which can reflect the probability that the sample object performs the business operation jointly indicated by N business indicators. That is to say, this sample prediction probability is equivalent to the sorting score after multi-objective (the objective is the business indicator) fusion. Compared with the sorting score after multi-objective fusion obtained through multi-stage processing, the present application adopts an end-to-end manner of directly outputting the sorting score jointly indicated by multiple business indicators, which can greatly reduce the effect loss caused by multiple different stages; then, this sample prediction probability can be used as the unique sorting score shared by each business indicator, and each business indicator can perform a monotonic transformation process on this unique sorting score through corresponding single-label transformation parameters to obtain the corresponding sample transformation probability under each business indicator. Since the monotonic transformation does not change the order of the original values, after each business indicator performs a monotonic transformation on each sample prediction probability, it can still maintain its original sorting situation, and at the same time transform the sample prediction probability into a value that more conforms to the actual situation of the business indicator (that is, the probability used to reflect that the sample object performs the business operation separately indicated by the corresponding business indicator); finally, the recommendation model can be trained and optimized based on the sample transformation probabilities of each business indicator, so that the prediction probability output by the recommendation model can improve the accuracy of the prediction probability under the joint influence of each business indicator, and based on the accurate prediction probability, optimize the sorting result of the resource data and improve the recommendation accuracy of the resource data. In summary, the present application provides an end-to-end multi-objective efficient modeling method with monotonic shared unique sorting, which can optimize the data sorting effect and improve the recommendation accuracy of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0067] Figure 1 It is a schematic diagram of the architecture of a solution system provided by an exemplary embodiment of the present application;

[0068] Figure 2 It is a schematic diagram of a scenario provided by an embodiment of the present application;

[0069] Figure 3It is a schematic flowchart of a data processing method provided by an exemplary embodiment of the present application;

[0070] Figure 4 It is a schematic logical architecture diagram of a multi-objective modeling solution provided by an embodiment of the present application;

[0071] Figure 5 It is a schematic flowchart of training and optimizing a recommendation model provided by an embodiment of the present application;

[0072] Figure 6 It is a schematic structural diagram of a data processing device provided by an embodiment of the present application;

[0073] Figure 7 It is a schematic structural diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0074] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0075] The embodiments of the present application relate to artificial intelligence and related technologies. For ease of understanding, the following will first briefly elaborate on artificial intelligence and related technical terms and concepts.

[0076] 1. Artificial Intelligence (AI)

[0077] Artificial intelligence is to use a digital computer or a machine controlled by a digital computer to simulate, extend, and expand human intelligence, a theory, method, technology, and application system that perceives the environment, acquires knowledge, and uses knowledge to obtain the best results. Artificial intelligence technology is an interdisciplinary subject, involving a wide range of fields, including both hardware-level technologies and software-level technologies. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0078] Further, the embodiments of the present application are mainly related to Machine Learning (ML) in artificial intelligence technology. Among them: Machine learning is an interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. Machine learning specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve its own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning. In the specific implementation of the embodiments of the present application, the machine learning method can be used to train and optimize relevant models (such as the recommendation model mentioned later) to improve the accuracy of its output results.

[0079] 2. Multiple objectives

[0080] Multiple objectives refer to multiple prediction tasks or multiple prediction business indicators. Objectives are equivalent to tasks or business indicators, so they can also be called multiple tasks or multiple indicators, etc. Objectives or business indicators include but are not limited to: click-through rate, conversion rate (such as purchase rate, favorite rate, like rate, favorite rate, share rate, comment rate, etc.), browsing duration, click-through conversion rate (such as click-favorite rate, click-purchase rate, etc.). Taking the business indicator of click-through rate as an example, predicting the click-through rate is to predict the click probability, which means that after a certain resource data (such as advertisement data) is exposed to a user, it is estimated the probability that the user may click on the resource data. Taking the business indicator of conversion rate as an example, predicting the conversion rate is to predict the conversion probability, which means that after a certain resource data is exposed to a user, it is estimated the probability that the user may perform a conversion operation (such as purchase, favorite, like, etc.) on the resource data.

[0081] 3. Multiple-objective prediction

[0082] Multi-object prediction can also be called multi-task prediction. Multi-object prediction is used to predict the occurrence probability under each business metric (i.e., the probability that users may perform corresponding operations on resource data). Based on the occurrence probabilities of these business metrics, multi-object fusion can be performed on them. Specifically, the occurrence probabilities of each business metric can be fused to determine the final ranking score. Multi-object prediction usually can adopt the Multi-gate Mixture-of-Experts (MMoE) model. Each expert network in the MMoE model has the same network structure, such as network structures like Deep Neural Networks (DNN), Factorization Machine (FM), and Deep&Cross Network (DCN). The MMoE model obtains different features based on training samples through multiple expert networks, and then predicts the multi-business metrics of an object based on different features. Among them, the Multi-gate Mixture-of-Experts (MMoE) is a commonly used network structure for multi-object learning. The experts are mostly DNN network structures. Multiple expert networks are used to extract different features, and the gating is used to allocate the weights of each expert network. The expert network can also be called an expert model. Specifically, it can be a neural network model that can be obtained through model training and convergence. The shared expert network is a network jointly used by multiple tasks and is used to obtain corresponding high-order outputs based on the input feature data. The private expert network is used by a single task and is used to obtain corresponding high-order outputs based on the input features.

[0083] In practical applications, recommendation systems usually adopt multi-objective models to output the predicted probabilities of each business metric (i.e., the predicted probabilities that users may perform corresponding operations on certain resource data) based on the input features. Then, the recommendation system will use a multi-objective fusion model to fuse the predicted probabilities of each business metric to obtain the sorted scores after fusion under multiple business metrics, so as to sort the candidate recommendations to each resource data of the user based on the sorted scores, and determine the final resource data to be recommended to the user based on the sorted results. For example, assume that multiple business metrics include click-through rate and like rate, and the resource data candidate pushed to a certain user is video data. For these video data, it is hoped that more users can like them. First, the video features of each video data (video features include but are not limited to: video ID, video theme, video category, video creator, etc.) and the features of the user (user features include but are not limited to: user age, user location information, etc.) can be obtained. Based on the video features and user features of a video data, the combined features corresponding to the video data can be formed. Then, the combined features can be input into a multi-objective prediction model (also called a multi-objective model, which can be an MMoE model here). The Mixture of Experts (MoE) in the MMoE model can calculate and process the combined features to obtain an output result. Then, the independent gate corresponding to the click-through rate in the MMoE model can perform weighted combination on the output of the MoE network and output the high-order output corresponding to the click-through rate. The high-order output can be processed through the tower network (small DNN network) corresponding to the click-through rate, and finally the predicted probability corresponding to the click-through rate is output. Similarly, the independent gate corresponding to the like rate in the MMoE model can perform weighted combination on the output of the MoE network and output the high-order output corresponding to the like rate. The high-order output can be processed through the tower network (small DNN network) corresponding to the like rate, and finally the predicted probability corresponding to the like rate is output. Further, the predicted probability output by the tower network corresponding to the click-through rate and the predicted probability output by the tower network corresponding to the like rate will both be input into the multi-objective fusion model. The multi-objective fusion model will fuse the predicted probability corresponding to the click-through rate and the predicted probability corresponding to the like rate according to specific rules or formulas, and thus the final sorted score corresponding to the video data (i.e., the predicted probability jointly corresponding to the click-through rate and the like rate) can be obtained.For any video data that is a candidate for being pushed to a user, the above method can be used to determine a sorting score. Subsequently, based on the sorting score, the video data to be pushed to the user can be selected from these video data (for example, the first few video data with larger sorting scores can be used as the video data to be pushed to the user). Since these videos are the video data with relatively large estimated sorting scores, there is a high probability that the user will click and like them in actual situations (such as the operations indicated by the click-through rate and the like rate).

[0084] Based on the above, in the process of pushing resource data (such as video data) to a user, it is necessary to first obtain multiple business metrics in different push scenarios. For each candidate resource data to be pushed to the user, multi-objective prediction (that is, the prediction of multiple business metrics) needs to be performed through a multi-objective prediction model to predict the probability that the user may perform corresponding operations under each business metric. That is, it is necessary to perform multi-objective prediction through a multi-objective prediction model to obtain the estimated probability under each business metric. Further, after determining the estimated probability under each business metric, it is necessary to fuse the estimated probabilities of all business metrics through a multi-objective fusion model according to specific fusion rules (the specific fusion rules can be configured and determined based on actual business requirements, and the specific fusion rules include but are not limited to: the rule of adding the estimated probabilities of all business metrics, the rule of multiplying the estimated probabilities of all business metrics, etc.) to obtain the probability that the user may perform corresponding operations under all business metrics, that is, to determine the estimated probability corresponding to all business metrics together. The estimated probability corresponding to all business metrics together will be used as the sorting score of each candidate resource data to be pushed to the user. According to the final sorting scores of each resource data, the resource data to be finally pushed to the user can be selected from them.

[0085] It should be noted that, in order to optimize the sorting effect of each resource data, traditional recommendation techniques will train and optimize a multi-objective prediction model so that the results output by the multi-objective prediction model (i.e., the estimated probabilities of each business indicator) can be more accurate. However, since the multi-objective prediction model and the multi-objective fusion model in traditional techniques are in two stages, even if the prediction result of a single business indicator can be optimized through model training, it does not mean that the result after the fusion of multiple business indicators can also be improved accordingly. Even in the case where some business indicators are mutually exclusive (for example, the click-through rate is the ratio between the number of clicks and the number of exposures, and the conversion rate is the ratio between the number of conversions and the number of clicks. Then, if you want a larger click-through rate, the number of clicks needs to be larger. However, if the number of clicks increases and the click-through rate increases, the conversion rate will decrease due to the increase in the number of clicks. It can be seen that the click-through rate and the conversion rate are mutually exclusive between these two business indicators), through the improvement of the respective results of the business indicators, on the contrary, it will lead to a decline in the effect after the fusion of business indicators, resulting in a decrease in the sorting effect and inaccurate recommendation results.

[0086] Based on this, in order to optimize the sorting effect of resource data under multiple business indicators and improve the recommendation accuracy of resource data, the present application provides an end-to-end multi-objective modeling scheme with monotonic shared unique sorting. A DNN network is used as a recommendation model to process the input features (composed of the features of a user and a resource data) to output a scalar score, and this scalar score can be used as the sorting score commonly corresponding to multiple business indicators; the sorting score output by the DNN can be shared by each business indicator, and each business indicator uses the corresponding monotonic transformation parameter to perform a monotonic transformation on this sorting score to convert it into a value closer to the true label of this business indicator (i.e., the true operation label of the user under this business indicator. For example, assuming the business indicator is the click-through rate, whether the user performs a click operation on the resource data is the true operation label of this click-through rate). Through the values after the transformation of each business indicator and the true operation label, the recommendation model can be trained and optimized to improve the output accuracy of the recommendation model.

[0087] Specifically, the multi-objective modeling solution provided by the embodiments of this application generally includes the following steps: First, sample data for model training can be obtained. Here, the sample data includes one or more samples, and each sample should be constructed and determined based on the feature data of the sample object and the feature data of the historical resource data of the historical business operations performed by the sample object. The sample object can refer to a user, and the historical resource data can refer to the resource data exposed to the user (the resource data can be determined based on the push scenario. When the push scenario is an advertisement push scenario, the resource data can be advertisement data; when the resource data is a video push scenario, the resource data can be video data). The historical business operations can include, but are not limited to, browsing operations, click operations, conversion operations (such as purchase operations, like operations, favorite operations, forward operations, share operations, etc.). That is to say, each resource data exposed to the user can be used as the historical resource data. By combining the feature data of a historical resource data with the feature data of the user, a sample can be formed (each sample can also be called sample data. That is to say, the number of sample data for model training is one or more, and each sample data should be constructed and determined by the feature data of a historical resource data and the feature data of the user).

[0088] After obtaining the sample data, the following processing can be performed on each sample data: Call the recommendation model to perform probability prediction on the sample data to obtain the sample prediction probability corresponding to the sample data. Here, the recommendation model can include, but is not limited to, a DNN network. Through this DNN network, calculations based on a deep network can be performed on the input sample data (such as feature extraction, hidden feature calculation, feature classification output, etc.) to obtain an estimated probability (i.e., the sample prediction probability) for indicating that the sample object may perform a business operation jointly indicated by N (N is a positive integer, and generally, N is a value greater than or equal to 2) business metrics on the historical resource data (the historical resource data indicated by the sample data). Among them, the business operation jointly indicated by N business metrics is equivalent to the business operation indicated after the fusion of multi-business metrics (or multi-objectives) (for example, the above-mentioned click-and-convert operation, or browse-and-click-and-convert operation); then, for this sample prediction probability, it can be shared among each business metric, and each business metric can perform a monotonic transformation process on this sample prediction probability respectively using corresponding monotonic transformation parameters to obtain the sample transformation probability corresponding to each business metric. Through the N sample transformation probabilities and the true sample label of the sample data under this business metric (the true sample label of a certain business metric can be determined based on whether the user has performed the corresponding business operation on the historical resource data), the model parameters of the recommendation model and the monotonic transformation parameters under each business metric can be trained and optimized.

[0089] It should be noted that the monotonic transformation parameters corresponding to each business metric in this application are trainable. They can be trained and optimized prior to the recommendation model, so that the values transformed based on the monotonic transformation parameters can get closer and closer to the true sample labels under the business metrics. Since monotonic transformation is a well-known transformation that does not change the order of the original values, through monotonic transformation, the predicted probabilities of the transformed samples will still maintain their original sorting positions. However, by using different monotonic transformation parameters, the values of the transformed probabilities of the samples can be made to better match the true sample labels of the business metrics themselves. The error between the transformed sample transformation probabilities output by the recommendation model and the true sample labels can be used to train and optimize the recommendation model to output more accurate predicted probability values. In other words, the recommendation model can be optimized under the combined influence of multiple business metrics, and each business metric will not affect the sorting of the recommendation model itself, avoiding the negative impact caused by the mutual exclusion between business metrics.

[0090] The solution provided by the embodiments of this application can be applied to any application scenario that requires resource data recommendation, including but not limited to: short video push scenarios, video playback scenarios (which can be used for scenarios such as watching TV dramas, movies, and variety shows), advertisement push scenarios, and so on.

[0091] The short video push scenario can refer to a scenario where video data is continuously pushed to users. Users can request to update and display the next video data by performing an operation to pull video data (such as swiping the video display interface of a terminal device). In the short video push scenario, users can continuously perform the operation of pulling video data to continuously refresh and view different video data.

[0092] The video playback scenario can refer to a scenario where a user watches video data such as a TV drama, a movie, or a variety show on a certain video playback platform. In the video playback scenario, different video data can be recommended to the user on the home page of the video playback platform.

[0093] In summary, the end-to-end monotonic shared unique sorting multi-objective efficient modeling solution provided by the embodiments of this application can improve the consistency of multi-objective training and fusion, optimize the data sorting effect, and improve the accuracy of recommendation results, effectively improving the business coverage to a certain extent (such as expanding the applicable scenarios).

[0094] It should be noted that the above-mentioned several application scenarios are only examples and will not limit the application scenarios applicable to the solution provided by the embodiments of this application.

[0095] Further, the solution provided by the embodiments of the present application can be executed by a computer device, which may include a terminal or a server, or may further include a terminal and a server. To facilitate understanding of the solution provided by the embodiments of the present application, the application scenarios involved in the embodiments of the present application will be introduced below in conjunction with Figure 1 the system schematic diagram shown; among them, Figure 1 is a schematic diagram of the architecture of a solution system provided by an exemplary embodiment of the present application. As Figure 1 shown, the system includes a terminal 101 and a server 102; where:

[0096] 1) The terminal 101 may include the terminal device used by the user. Of course, according to the different application scenarios and fields to which the present solution is applied, the terminals providing the solution of the embodiments of the present application are different. The terminal device may include, but is not limited to: smart phones (such as smart phones deployed with the Android system, or smart phones deployed with the Internetworking Operating System (IOS)), tablet computers, portable personal computers, Mobile Internet Devices (MIDs), vehicle-mounted devices, head-mounted devices, smart home and smart voice interaction devices, etc. The embodiments of the present application do not limit the type of the terminal device, which is hereby explained.

[0097] For example, in the short video push scenario, the terminal device can be a smart phone; that is to say, in this implementation mode, the solution provided by the embodiments of the present application can be deployed on the smart phone; when the user uses the short video push application on the smart phone, a batch of video data to be pushed to the user is obtained by the smart phone; then, the smart phone can form the service data corresponding to each video data based on the feature data of the user and the feature data of each video data (that is, the feature data of a video data and the feature data of the user can be used to construct a piece of service data); for each piece of service data, the smart phone can call the trained and optimized recommendation model (the recommendation model is deployed in the smart phone) to perform probability prediction on it to obtain the prediction probability of the corresponding video data (the prediction probability can be used to reflect the estimated probability that the user may perform the service operations indicated by multiple service metrics on the video data), through the prediction probabilities of each video data, the smart phone can sort these video data and select the first few video data with larger prediction probabilities as the video data to be finally pushed to the user. Subsequently, the smart phone can push the video data to the user one by one (for example, the user can play these video data one by one by continuously performing the pull operation of the video data). Another example is that in the intelligent vehicle scenario, the application program deployed with the solution provided by the embodiments of the present application is a vehicle-mounted application program; the types of the vehicle-mounted application program can include but are not limited to: music, video, or games, etc.

[0098] Among them, an application program may refer to a computer program for completing one or more specific tasks; classifying application programs according to different dimensions (such as the running mode, function, etc. of the application program), the types of the same application program under different dimensions can be obtained. For example: classified according to the running mode of the application program, the application program may include but is not limited to: the client installed in the terminal, the applet that can be used without downloading and installation (as a subroutine of the client), the World Wide Web (Web) application program opened through the browser, and so on. Another example: classified according to the function type of the application program, the application program may include but is not limited to: the Instant Messaging (IM) application program, the content interaction application program, the audio application program or the video application program, and so on. Among them, the instant messaging application program refers to an application program for instant communication messages and social interaction based on the Internet. The instant messaging application program may include but is not limited to: the application program with communication functions, the map application program with interaction functions, the game application program, and so on. The content interaction application program refers to an application program that can realize content interaction, such as the sharing platform, personal space, news and other application programs. The audio application program refers to an application program that realizes audio functions based on the Internet. The audio application program may include but is not limited to: the music application program with music playing and editing capabilities, the radio application program with radio playing capabilities, or the live broadcast application program with live broadcast capabilities, and so on. The video application program refers to an application program that can play pictures. The video application program may include but is not limited to: the application program with short videos (the video length is often short, such as a few seconds or a few minutes, etc.) (such as the short video push application program), the application program with long videos (such as videos with a long playing time similar to movies or TV dramas), and so on.

[0099] Of course, the solution provided by the embodiments of the present application can be deployed directly on a device (such as a smart phone) or outside the application program as described above, and can also be deployed in the device or the application program in the form of a plug-in. The embodiments of the present application do not limit the carrier of the deployment solution.

[0100] 2) The server 102 can be the server corresponding to the terminal, which is used to interact with the terminal to provide computing and application service support for the terminal. Specifically, this server is the background server corresponding to the application program deployed in the terminal, and is used to interact with the terminal to provide computing and application servers for the application program. Among them, the server 102 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.

[0101] Among them, the terminal 101 and the server 102 can be directly or indirectly connected through wired or wireless communication methods, and this application does not limit this. In addition, the embodiments of this application do not limit the number of terminals and servers; in Figure 1 it is only an example that the numbers of both the terminal 101 and the server 102 are single, and in actual applications, it can include multiple servers with distributed distribution, and this is specifically stated here.

[0102] The following combines Figure 1 the system shown to introduce the general process of the end-to-end monotonic shared unique sorting multi-objective efficient modeling scheme in the application scenario. In specific implementation, first, multiple historical resource data pushed to the user in the past can be obtained. Based on the feature data of one historical resource data and the feature data of the user, a sample data can be constructed and generated. Multiple historical resource data can correspondingly generate multiple sample data; then, the following processing can be performed on each sample data: a recommendation model can be called to perform probability prediction on the sample data to obtain the estimated probability of the operations jointly indicated by N business indicators. This estimated probability can be called the sample prediction probability; for this sample prediction probability, for any one business indicator, a corresponding monotonic transformation parameter can be used to perform monotonic transformation processing on it to obtain the sample transformation probability under this business indicator; through N sample transformation probabilities, the recommendation model can be trained and optimized, and the trained and optimized recommendation model can be used to perform probability prediction on each candidate push resource data to be pushed to the user in the resource data push scenario to obtain the prediction probability corresponding to each candidate push resource data. Then, the computer device can push more user-adapted resource data based on this prediction probability.

[0103] Based on the above-described scheme and system architecture, the following points need to be further explained:

[0104] ① As mentioned above in the embodiments of this application Figure 1The system shown is for more clearly illustrating the technical solution of the embodiment of the present application and does not constitute a limitation on the technical solution provided by the embodiment of the present application. Those of ordinary skill in the art will know that with the evolution of the system architecture and the emergence of new business scenarios, the technical solution provided by the embodiment of the present application is equally applicable to similar technical problems. For example, the above takes the execution subject "computer device" of the embodiment of the present application including a terminal and a server as an example, that is, the solution provided by the embodiment of the present application is jointly executed by the terminal and the server, and an application scenario of the solution is introduced; it should be understood that in actual applications, the computer device can also be a terminal or a server, that is to say, it supports the solution provided by the embodiment of the present application to be executed independently by the terminal or the server.

[0105] ② The embodiment of the present application supports using a recommendation model with prediction ability (such as a model including a Deep Neural Networks (DNN) structure) to implement the solution described above. Among them, if the computer device used to execute the solution provided by the embodiment of the present application is a terminal, then the model can be deployed in the terminal. If the computer device used to execute the solution provided by the embodiment of the present application is a server, then the model is deployed in the server; in this case, the candidate push resource data used by the user from the terminal to the user is transmitted to the server for probability prediction processing based on multiple service metrics.

[0106] ③ In the embodiment of the present application, the collection and processing of relevant data should be strictly in accordance with the requirements of relevant laws and regulations. Obtaining personal information requires the informed consent of the individual subject (or having a legal basis for information acquisition), and subsequent data use and processing behaviors should be carried out within the scope authorized by laws and regulations and the individual information subject. For example, when the embodiment of the present application is applied to a specific product or technology, such as obtaining the feature data of a user, the permission or consent of the user needs to be obtained, and the collection, use, and processing of relevant data (such as the processing of the user's feature data, the recommendation processing of resource data, etc.) need to comply with the relevant laws, regulations, and standards of the relevant region.

[0107] Based on the solution described above, for the convenience of understanding its application scenario, please also refer to Figure 2 , Figure 2 which is a schematic diagram of a scenario provided by the embodiment of the present application. Among them, as Figure 2 shown, the scenario is described by taking the resource data as video data and N service metrics including click-through rate and conversion rate as examples. As Figure 2As shown, when user a uses the instant messaging application deployed on the terminal, the instant messaging application can display a short video push control at the bottom of the corresponding interface of the application. User a can enter the short video push sub-application through the triggering operation on this short video push control and play the video data in an immersive manner in the short video push sub-application. Among them, the short video push sub-application can continuously push new video data to user a, and user a can request to display new video data by performing an operation to pull video data. It should be noted that the short video push sub-application can highlight the short video push control in different forms (such as red heart display, red dot display, etc.) to prompt user a that there is an update of relevant content currently.

[0108] As Figure 2 shown, at the bottom of the home page interface 2000 of the instant messaging application, there is a small world control, which is the short video push control in this application. After user a generates a triggering operation on this control, the terminal can respond to this triggering operation of user a, jump to the video push sub-application, and display the video display interface 2001. As Figure 2 shown, in the video display interface 2001 of the terminal, the video data currently exposed to the user is video data 200a (that is, the video data that user a is currently playing and watching is video data 200a. The creator of this video data 200a is an object named "La La Loves Dancing", and the video copywriting of this video data 200a is "Teach you how to quickly learn a dance in three minutes"). In this video display interface 2001, there are a like control, a comment control, and a share control, which can be used for user a to like, comment on, and share this video data 200a. In addition, user a can perform a sliding operation in the video display interface 2001 to request to refresh and display new video data (video data that user a has not played or is not currently playing).

[0109] It should be noted that in practical applications, the background server of the short video push sub-application can, in advance, based on the video data that the user is currently playing or has played, search the database for multiple video data with similar or identical video themes to these video data (for example, the video themes are all <Pets>, <Cooking>, <Food store exploration>, <Emotional counseling>, etc.), and use them as candidate push video data to be subsequently pushed to this user. For each candidate push video data, the server can determine the probability that the user may perform a click and conversion operation on it. Then, the server can sort each candidate push video data in descending order according to the probabilities between them. After obtaining the video sequence, the server can select the first 5 candidate push video data in the sequence as the video data to be finally exposed to user a. For these 5 video data, they can all be called push video data. The server can transcode each push video data, and the terminal will pre-load the transcoded video data. Then, after the user performs the operation of pulling the video data, the terminal can push these downloaded transcoded video data to the user in sequence for the user to browse and watch.

[0110] In specific implementation, for each candidate push video data, the server can perform the following processing: First, the server can obtain the feature data of the candidate push video data and the feature data of the user; the server can combine the feature data of the candidate push video data and the feature data of the user, that is, perform a fusion process, to generate a business data containing the feature data of the candidate push video data and the feature data of the user; then, the server can call the trained and optimized recommendation model (the recommendation model can be trained and optimized in the manner described above) to perform a probability prediction process on this business data to output the prediction probability corresponding to this business data. This prediction probability can be used to reflect the probability that after the candidate push video data is exposed to user a, user a may perform a click and conversion operation on this candidate push video data. This prediction probability can be used as the prediction probability corresponding to the candidate push video data. It should be understood that the server can call the recommendation model to output the prediction probability corresponding to each candidate push video data respectively. Through the prediction probabilities corresponding to each candidate push video data respectively, the server can select the video data that user a is more likely to perform a click and conversion operation on as the push video data. For each push video data, the server can sort them according to the prediction probability to obtain a video data sequence.

[0111] After user a performs the operation of pulling the video data, the terminal can update and display the video data from video data 200a to the video data ranked first in the video data sequence (that is, the video data with the highest prediction probability) in the video display interface 2001. For example, Figure 2As shown, assume that the video data sequence is {video data 200b, video data 200c, video data 200d, video data 200e, video data f}, and user a generates a sliding operation in the video display interface 2001 (the sliding direction is the direction indicated by the arrow as shown in Figure 2 ). The terminal can respond to this sliding operation and update the display of the video data in the video display interface 2001 to the transcoded video data 200b (the creator of this video data 200b is an object named "Meimei Cooks Delicious Food", and the video copywriting of this video data 200b is "Buy crucian carp and cook it like this, with a fragrant smell!"). If user a generates another sliding operation in the video display interface 2001 (the sliding direction is the direction indicated by the arrow as shown in Figure 2 ), then the terminal can respond to this sliding operation and update the display of the video data displayed in the video display interface 2001 to video data 200c... That is to say, the user can request to update and display new video data by continuously performing sliding operations. When user a plays and browses video data 200f, the server can obtain a new batch of candidate video data to be pushed to user a, determine several video data with relatively high prediction probabilities from them as the final push video data to be pushed to user a, and then the server can sort this batch of push video data. The sorted video data sequence can be sent to the terminal, and the terminal will preload it to ensure that after user a generates a sliding operation, various video data can be quickly viewed. It should be noted that after determining the video data to be finally pushed to the user, the present application may not limit the playing order of each video data. For example, the push video data can be sorted in a random sorting manner to obtain a video data sequence.

[0112] Based on the above-described solution and application scenario, the embodiments of the present application propose a more detailed data processing method. The following will introduce the data processing method proposed in the embodiments of the present application in detail with reference to the accompanying drawings.

[0113] Please refer to Figure 3 , Figure 3 which is a flowchart of a data processing method provided by an exemplary embodiment of the present application. This process may refer to the process of the end-to-end monotonic shared unique sorting multi-objective modeling scheme provided by the embodiments of the present application. This data processing method can be executed by the computer device in the aforementioned system, such as the computer device being a terminal and / or a server; this data processing method can at least include the following steps S101 - S104:

[0114] Step S101, obtain sample data for model training; the sample data is determined based on the sample object and the historical resource data of the historical business operations performed by the sample object.

[0115] In this application, the sample object may refer to a user, the historical resource data may refer to the resource data exposed to the user, and the resource data includes but is not limited to: media data, advertisement data, items (such as clothes, shoes, etc.). The historical business operation may refer to the business operation performed by the user on a certain historical resource data, and the business operation includes but is not limited to: browsing operation, click operation, various conversion operations (such as purchase operation, like operation, favorite operation, share operation, forward operation, etc.). Based on the feature data of the user and the feature data of a historical resource data, a sample data can be constructed and generated, that is, each historical resource data can correspond to a sample data, and the sample data corresponding to each historical resource data can be used to train and optimize the recommendation model mentioned later.

[0116] Taking the determination of the sample data corresponding to a historical resource data as an example, in specific implementation, the specific implementation process for obtaining the sample data for model training may include but is not limited to: First, the feature data of the sample object and the feature data of the historical resource data (as known above, the historical resource data refers to the resource data on which the sample object has performed one or more historical business operations) can be obtained; among them, whether it is the feature data of the sample object or the feature data of the historical resource data, both can contain the feature data of multiple feature domains. For the convenience of distinction, this application may refer to the feature domain under the sample object as the first feature domain, and the feature domain under the historical resource data as the second feature domain. That is to say, the feature data of the sample object includes N1 (N1 is a positive integer) feature data of the first feature domain, and the feature data of the historical resource data includes N2 (N2 is a positive integer) feature data of the second feature domain. For example, the feature data of the sample object may include but is not limited to: user identifier (such as user ID or number), user age, historical business operation sequence, etc. The feature data of the historical resource data may include but is not limited to: resource identifier (such as the ID or number of the resource data, the producer of the resource data, the category or theme to which the resource data belongs, context information (such as the time of exposure to the user, the current time), etc. feature domains. Whether it is the feature data of the sample object or the feature data of the historical resource data, the feature data of each feature domain can be converted into a fixed-length vector to obtain the feature conversion vector corresponding to each feature domain. Then, the feature conversion vectors under each feature domain can be fused (such as concatenated), and the fused feature conversion vector can be used as the sample data corresponding to this historical resource data.

[0117] That is to say, the present application can specifically perform fixed-length vector conversion on the feature data of each first feature domain, thereby obtaining a feature conversion vector corresponding to each first feature domain (which can be called the first feature conversion vector); it can also perform fixed-length vector conversion on the feature data of each second feature domain, thereby obtaining a feature conversion vector corresponding to each second feature domain (which can be called the second feature conversion vector). After obtaining the first feature conversion vector corresponding to each first feature domain and the second feature conversion vector corresponding to each second feature domain, N1 first feature conversion vectors and N2 second feature conversion vectors can be obtained. Then, the N1 first feature conversion vectors and the N2 second feature conversion vectors can be vector-concatenated to obtain the sample data.

[0118] Based on the above, in specific implementation, the method for obtaining the feature transformation vector of a certain feature domain may include, but is not limited to: performing fixed-length vector transformation on the feature data of the feature domain to obtain the feature transformation vector (for example, performing fixed-length vector transformation on the feature data of any first feature domain to obtain the first feature transformation vector corresponding to the first feature domain; performing fixed-length vector transformation on the feature data of any second feature domain to obtain the second feature transformation vector corresponding to the second feature domain). For any feature domain (such as any first feature domain or second feature domain), the specific implementation process of performing fixed-length vector transformation on the feature data of the feature domain to obtain the feature transformation vector corresponding to the feature domain (such as the first feature transformation vector corresponding to the first feature domain or the second feature transformation vector corresponding to the second feature domain) may include, but is not limited to: First, the domain type to which the feature domain belongs can be obtained; among them, the domain type may include a fixed-length type and a variable-length type. The fixed-length type means that the feature data under this feature domain is data with a fixed length. Correspondingly, the variable-length type means that the feature data under this feature domain is data without a fixed length. For example, feature data under feature domains such as user identification (such as user ID), resource identification (such as resource ID), user age, and user location information are usually data with a unified fixed format (for example, the IDs of different users have the same format; the ages of different users also have the same format), and their lengths are fixed values, so the domain type to which the user identification or resource identification belongs can be determined as the fixed-length type; another example is that feature data under feature domains such as the historical business operation sequence executed by the user and the historical resource data sequence played by the user. Different users have different historical business operations, so the historical business operation sequences of different users are also different (the number of historical business operations included is different. Some users' historical business operation sequences contain 10 historical business operations, and some users' historical business operation sequences contain 7 historical business operations). Similarly, different users have different historical resource data played, so the lengths of the historical resource data sequences of different users (that is, the number of historical resource data included) are also different. After obtaining the domain type to which the feature domain belongs, the vector transformation method of the feature data of the feature domain can be determined according to the domain type.

[0119] Specifically, for the feature data of the fixed-length type feature domain, since the length of its feature data is already a unified length, it only needs to be directly vector-embedded to convert it into a vector; while for the feature data of the variable-length type feature domain, since the length of its feature data is not unified, it needs to be converted into a fixed-length vector. That is, when the domain type is determined to be of the fixed-length type, the feature data of the feature domain can be processed by a feature embedding network (for example, an embedding network layer) for vector embedding to obtain the feature conversion vector of the feature domain; while when the domain type is of the variable-length type, the feature data of the feature domain can be processed by a sequence network for vector mapping and pooling to obtain the feature conversion vector of the feature domain. Among them, the sequence network here can actually also refer to the embedding network layer. Through the embedding network layer, each element in the feature data of the feature domain can be subjected to vector mapping processing (vector embedding processing) to convert each element into a vector representation; then, the vector sequence composed of the vector representations of each element can be subjected to pooling processing (such as summation or averaging processing), and the result obtained after pooling processing can be used as the fixed-length feature conversion vector corresponding to the feature data of this feature domain.

[0120] It should be understood that based on the above-described method, assuming that we have obtained a total of N fixed-length feature conversion vectors (including the first feature conversion vectors corresponding to N1 first feature domains and the second feature conversion vectors corresponding to N2 second feature domains), for the nth feature conversion vector, we can record it as x n , and the dimension of this feature conversion vector can be recorded as d n , then after concatenating these N fixed-length feature conversion vectors, the complete sample input vector (that is, a sample data) can be as shown in formula (1):

[0121] X = […, x n , …] Formula (1)

[0122] Among them, the dimension of the sample input vector X is

[0123] Step S102, call the recommendation model to perform probability prediction processing on the sample data to obtain the sample prediction probability; the sample prediction probability refers to the probability that the predicted sample object performs the business operation jointly indicated by N business indicators; N is a positive integer.

[0124] In this application, the recommendation model can refer to a model containing a DNN network structure. This application can end-to-end consider the influence of sorting factors of multiple service indicators on the final estimated probability, and can use the results output by the recommendation model to sort each resource data, thereby avoiding the loss of effect and the engineering complexity caused by the two-stage decoupling of first estimating the probabilities of multiple service indicators and then performing multi-objective fusion. At the same time, since only the results output by the recommendation model are used for sorting, the model structure will be greatly streamlined compared with the multi-objective prediction model, and thus the resource consumption during model offline training and online service can also be reduced. Based on this, after constructing the sample data, the recommendation model can be called. In the recommendation model, the sample data can be correspondingly calculated and processed through each network layer of the DNN, that is, the recommendation model can perform probability prediction processing on the sample data, and the results output by the recommendation model can be used as the prediction probability corresponding to the sample data (referred to as the sample prediction probability). Since the results output by the recommendation model can be used for the final sorting, the results output by the recommendation model (i.e., the sample prediction probability) can be determined as the probability that the sample object may perform the service operation jointly indicated by N service indicators on the historical resource data (i.e., the historical resource data indicated by the sample data).

[0125] For example, assuming that the N service indicators include click-through rate and collection rate, the service operation indicated by the click-through rate alone is the click operation, and the service operation indicated by the collection rate alone is the collection operation. Then the service operation jointly indicated by the N service indicators can refer to the operation of clicking and collecting. After inputting the sample data into the recommendation model, the sample prediction probability output by the recommendation model can represent the estimated probability that the sample object may perform the click-and-collect operation on a certain historical resource data.

[0126] Step S103: Perform monotonic transformation processing on the sample prediction probability through the monotonic transformation parameters of each service indicator to obtain the sample transformation probability corresponding to each service indicator; a sample transformation probability is used to represent the probability that the sample object performs the service operation indicated by the corresponding service indicator alone.

[0127] In this application, the sample prediction probabilities output by the above-mentioned recommendation model can be shared by N business metrics at the same time, and the business metrics are trained based on this unique value. In a specific implementation, for each business metric, this application can configure different trainable parameters as its monotonic transformation parameters for it, and the business metric can perform a monotonic transformation process on the sample prediction probability through the monotonic transformation parameters to obtain the sample transformation probability corresponding to the business metric. It should be understood that the sample transformation probability obtained through the corresponding monotonic transformation parameter under a business metric should be close to the true sample label under the business metric. For example, taking the business metric including click-through rate as an example, the business operation uniquely indicated by the click-through rate refers to the click operation. Assuming that for a certain historical resource data, the sample object has performed a click operation on it, then for the sample data constructed based on the feature data of the historical resource data, it can be determined as a positive sample under the business metric of click-through rate, and its true sample label in the click-through rate is determined as the label value corresponding to the positive sample (such as the value 1); conversely, if the sample object has not performed a click operation on the historical resource, then for the sample data constructed based on the feature data of the historical resource data, it can be determined as a negative sample under the business metric of click-through rate, and its true sample label in the click-through rate is determined as the label value corresponding to the negative sample (such as the value 0). After performing a monotonic transformation process on the sample prediction probability through the monotonic transformation parameter corresponding to the click-through rate, the obtained sample transformation probability should be close to the true sample label corresponding to the sample data. For example, when the true sample label is the value 1, then the sample transformation probability should be as close to 1 as possible, and when the true sample label is the value 0, the sample transformation probability should be as close to 0 as possible. Thus, the sample transformation probability after the monotonic transformation can conform to the true situation of the business metric itself.

[0128] In other words, the sample transformation probability obtained after performing a monotonic transformation process on a certain business metric is determined based on the sample prediction probability and can be used to represent the estimated probability that the sample object performs the business operation uniquely indicated by the business metric. This estimated probability should be as close as possible to the true sample label of the sample data under the business metric.

[0129] In a specific implementation, the following processing can be performed for any business metric: First, for the sake of easy understanding and explanation, any one of the N business metrics can be determined as the transformation business metric (referred to as the transformation business metric), and the monotonic transformation parameter of this transformation business metric can be determined as the target monotonic transformation parameter; then, the scaling transformation function in the monotonic transformation rule can be obtained. The scaling transformation function can be a function that can perform scaling transformation on values. Through the scaling transformation function in the monotonic transformation rule, the target monotonic transformation parameter and the sample prediction probability can be scaled and transformed to obtain the initial sample transformation probability of the transformation business metric; further, the activation transformation function in the monotonic transformation rule can be obtained. Through the activation transformation function in the monotonic transformation rule, the process of adding a bias to the value can be performed. Then, through this activation transformation function, the initial sample transformation probability is activated and transformed (such as the process of adding a bias), and the result obtained after the activation transformation can be used as the sample transformation probability corresponding to this transformation business metric.

[0130] For the specific implementation process of scaling and transforming the sample prediction probability, it can be as shown in formula (2):

[0131] s t =w t ·w t ·s + b t Formula (2)

[0132] Among them, the function shown in formula (2) can be used to represent the scaling transformation function. As shown in formula (2), s can be used to represent the sample prediction probability corresponding to the sample data; w t can be used to represent a trainable parameter corresponding to the t-th business metric (such as the transformation business metric) (for the sake of distinction, it can be called the first trainable parameter), and b t can be used to represent another trainable parameter corresponding to the t-th business metric (which can be called the second trainable parameter); s t can be used to represent the initial sample transformation probability of the t-th business metric obtained through the scaling transformation (that is, the initial sample transformation probability of the above transformation business metric).

[0133] It can be seen from formula (2) that in a specific implementation, the target monotonic transformation parameter of a certain business metric (such as the transformation business metric) includes the first trainable parameter and the second trainable parameter. Through the scaling transformation function in the monotonic transformation rule, the process of scaling and transforming the target monotonic transformation parameter and the sample prediction probability to obtain the initial sample transformation probability of the transformation business metric can include but is not limited to: First, the first trainable parameter (such as the above w tAfter performing a multiplication operation with the sample prediction probability to obtain the first product result, it is necessary to perform another multiplication operation on this first product result with the first trainable parameter to obtain the second product result; then, the second product result can be summed with the second trainable parameter (such as b as described above) t ) The result obtained from the summation operation can be used as the initial sample transformation probability of the transformation service metric.

[0134] Furthermore, after calculating the initial sample transformation probability, the initial sample transformation probability can be subjected to an activation transformation (such as a sigmoid transformation), and the result obtained from the transformation can be used as the sample transformation probability corresponding to the transformation service metric (i.e., the final estimated probability corresponding to the transformation service metric).

[0135] For the specific implementation process of performing an activation transformation (sigmoid transformation) on the initial sample transformation probability, it can be as shown in formula (3):

[0136]

[0137] where the function shown in formula (3) can be used to represent the activation transformation function; s as shown in formula (3) t can be used to represent the initial sample transformation probability of the t-th service metric obtained after scaling transformation (i.e., the initial sample transformation probability of the transformation service metric described above); p t can be used to represent the sample transformation probability of the t-th service metric obtained after activation transformation.

[0138] It should be noted that after the recommendation model outputs the sample prediction probability of the sample data, each service metric applies a monotonic transformation to it respectively. Since the monotonic transformation does not change the order of its transformation input value, then under each service metric, although the sample prediction probability output by the recommendation model is scaled and a bias is added to meet the needs of a certain service metric, the sorting of the transformed values (i.e., the sample transformation probability) is still the same as the sorting of the values before the transformation (i.e., the sample prediction probability). The sorting position of each value before or after the transformation remains unchanged, and multiple service metrics perform a monotonic sharing of the unique sorting of the sample prediction probability.

[0139] For ease of understanding, a specific example is used for illustration. Suppose there are 4 historical resource data, and their corresponding sample data also include 4 (sample data 1, sample data 2, sample data 3, and sample data 4). For the 4 sample data, the sample prediction probabilities respectively output by the recommendation model are 94, 86, 75, and 60. Among them, 94 is the sample prediction probability of sample data 1, 86 is the sample prediction probability of sample data 2, 75 is the sample prediction probability of sample data 3, and 60 is the sample prediction probability of sample data 4. For these 4 sample prediction probabilities, after sorting them in descending order, the obtained probability sequence is {94, 86, 75, 60}. For a certain business metric, it is necessary to perform a monotonic transformation on each sample prediction probability. For the sample prediction probability 94, the sample transformation probability 1 can be obtained through the monotonic transformation. For the sample prediction probability 86, the sample transformation probability 2 can be obtained through the monotonic transformation. For the sample prediction probability 75, the sample transformation probability 3 can be obtained through the monotonic transformation. For the sample prediction probability 60, the sample transformation probability 4 can be obtained through the monotonic transformation. For these 4 sample transformation probabilities, after sorting them in descending order of probability, the obtained sequence should be {sample transformation probability 1, sample transformation probability 2, sample transformation probability 3, sample transformation probability 4}. It can be seen that the sorting position of the sample prediction probability 94 is the 1st. After the monotonic transformation, the sorting position of its corresponding sample transformation probability 1 is also the 1st. That is to say, after each business metric performs a monotonic transformation on the sample prediction probability, the sorting position of the obtained sample transformation probability is the same as the sorting position of the value before the transformation. The monotonic transformation will not change the original order of the input values. This application only transforms them into values that meet the requirements of the business metric (i.e., values close to the true sample labels).

[0140] Step S104, training and optimizing the recommendation model through N sample transformation probabilities; the trained and optimized recommendation model is used to perform probability prediction processing on business data in the resource data push scenario; the business data is determined based on the business object and the candidate push resource data of the business object.

[0141] In this application, after obtaining the sample transformation probabilities corresponding to each business metric, the recommendation model can be trained and optimized based on the N sample transformation probabilities. In a specific implementation, the following processing can be performed on the sample transformation probability of each business metric: First, the true sample label of the sample data under this business metric can be obtained (as can be seen from the above, this true sample label can be used to reflect whether the sample object has performed the business operation indicated by this business metric alone on the historical resource data); then, through a loss function (such as the cross-entropy loss function), the prediction error value between the sample transformation probability of this business metric and this true sample label can be calculated; through this prediction error value, the monotonic transformation parameters corresponding to this business metric (at least including the above first trainable parameter and second trainable parameter) can be trained and optimized so that the monotonic transformation parameters corresponding to this business metric can have high accuracy. For any business metric, based on the corresponding sample transformation probability and true sample label, the prediction error value under the business metric can be calculated. This application can fuse the prediction error values of each business metric (such as performing a summation operation) to obtain a total prediction error value, and this total prediction error value can be used as the total loss value. Based on this total loss value, the model parameters of the recommendation model can be trained and optimized so that the recommendation model can optimize the output result of the recommendation model under the influence of the sorting factors of each business metric.

[0142] It should be noted that the monotonic transformation parameters corresponding to each business metric in this application can be trained and optimized together with the recommendation model, that is, after determining the prediction error values of each business metric, through the prediction error value of a single business metric, the monotonic transformation parameters of this business metric can be optimized; through the prediction error values of all business metrics, a total loss value obtained by fusion can be used to optimize the recommendation model, and the training and optimization are carried out until the model convergence condition of the recommendation model is met (the model convergence condition here can refer to that the total loss value reaches the minimum or the number of model iterations has reached the preset iteration threshold). For the specific training and optimization process of the recommendation model, reference can be made to the description in the corresponding embodiments Figure 4 described later.

[0143] It should be understood that after training and optimizing the recommendation model through N sample transformation probabilities, the obtained recommendation model already has high accuracy and can be applied to the resource data push scenario to determine the prediction probabilities of different candidate push resource data. Taking a certain resource data push scenario as an example, assume that currently, it is necessary to push some resource data to a certain user (referred to as the business object). In this scenario, the above-mentioned trained and optimized recommendation model can be applied. In specific applications, first, M (M is a positive integer) candidate push resource data regarding the business object can be obtained (a candidate push resource data refers to the resource data that is to be pushed to the business object as a candidate); then, for any candidate push resource data, the feature data of this candidate push resource data needs to be obtained. Through the feature data of the candidate push resource data and the feature data of the business object, a business data can be constructed and generated (the content included in the feature data of the candidate push resource data is the same as the content included in the feature data of the historical resource data described above. The method of generating business data can also refer to the method of generating a sample data described above, which will not be elaborated here). Based on this, the business data corresponding to each of the M candidate push resource data can be obtained (the business data corresponding to a candidate push resource data is determined based on the business object and the candidate push resource data); then, the trained and optimized recommendation model can be called to perform probability prediction processing on the business data corresponding to each candidate push resource data respectively to obtain the prediction probability corresponding to each candidate push resource data; from the definition description of the sample prediction probability corresponding to the above-mentioned one sample data, it can be known that the prediction probability corresponding to a candidate push resource data here can be used to represent the estimated probability that the business object performs the business operation jointly indicated by N business indicators on the corresponding candidate push resource data; the larger the prediction probability, the more likely it is that the business object will perform the business operation jointly indicated by N business indicators on the candidate push resource data. Then, based on the M prediction probabilities, the push resource data of the business object can be determined from the M candidate push resource data, and the push resource data can be pushed to the business object. For example, the candidate push resource data can be sorted in descending order according to the magnitudes of the M prediction probabilities (that is, sorted according to the order from large to small among the M prediction probabilities), obtaining a resource data sequence. Then, the first few (such as the first K, K should be a positive integer less than or equal to M) candidate push resource data in the resource data sequence can be determined as the push resource data of the business object. Since the prediction probabilities of these push resource data are relatively large, it is very likely that the business object will perform the business operation jointly indicated by the business indicators on these push resource data, and thus it can also be ensured that the resource data pushed to the business object can meet N business indicators.

[0144] In an embodiment of the present application, an end-to-end multi-objective efficient modeling method with monotonic shared unique sorting is provided, which can improve the consistency of multi-objective training and fusion, optimize the data sorting effect, and improve the accuracy of recommendation results. Specifically, after obtaining the sample data for model training, the present application can use a recommendation model to predict the sample data and output a sample prediction probability, which can reflect the probability that the sample object performs the business operation jointly indicated by N business indicators. That is to say, this sample prediction probability is equivalent to the sorting score after multi-objective (the objective is the business indicator) fusion. Compared with the sorting score after multi-objective fusion obtained through multi-stage processing, the present application adopts the method of directly outputting the sorting score jointly indicated by multiple business indicators in an end-to-end manner, which can greatly reduce the effect loss caused by multiple different stages. Then, this sample prediction probability can be used as the unique sorting score shared by each business indicator, and each business indicator can perform a monotonic transformation process on this unique sorting score through corresponding single-label transformation parameters to obtain the corresponding sample transformation probability under each business indicator. Since the monotonic transformation does not change the order of the original values, after each business indicator performs a monotonic transformation on each sample prediction probability, it can still maintain its original sorting situation, and at the same time transform the sample prediction probability into a value that more conforms to the actual situation of the business indicator (that is, the probability used to reflect that the sample object performs the business operation separately indicated by the corresponding business indicator). Finally, the recommendation model can be jointly trained and optimized based on the sample transformation probabilities of each business indicator, so that the prediction probability output by the recommendation model can improve the accuracy of the prediction probability under the joint influence of each business indicator, and based on the accurate prediction probability, optimize the sorting result of the resource data and improve the recommendation accuracy of the resource data.

[0145] For a better understanding of the logical process of the multi-objective modeling solution provided by the present application, please also refer to Figure 4 , Figure 4 which is a schematic diagram of the logical architecture of a multi-objective modeling solution provided by an embodiment of the present application. As Figure 4 shown, at least the following components need to be included in the logical architecture of this solution: a feature input component, a feature embedding layer, a splicing component, a recommendation model, and a monotonic transformation component corresponding to each business indicator. The following will briefly describe each component in this solution:

[0146] Feature input component: The feature input component can be used to obtain the feature data of the sample object and the feature data of the historical resource data (the historical resource data refers to the resource data for which the sample object has performed historical business operations), and input them into the feature embedding layer.

[0147] Feature Embedding Layer: After obtaining the feature data of the sample object input by the feature input component and the feature data of the historical resource data, for the feature data of each historical resource data, the feature embedding layer can perform the following processing: perform fixed-length vector conversion processing on the feature data of the sample object in each feature domain, and perform fixed-length vector conversion processing on the feature data of the historical resource data in each feature domain to obtain feature conversion vectors corresponding to each feature domain respectively. After obtaining the feature conversion vectors corresponding to each feature domain respectively, the feature embedding layer can input the feature conversion vectors corresponding to each feature domain respectively into the splicing component.

[0148] Splicing Component: The splicing component can be used to splice the feature conversion vectors corresponding to each feature domain respectively, and thus a splicing vector corresponding to a historical resource data can be spliced. This splicing vector can be used as the sample data corresponding to this historical resource data. Multiple historical resource data can generate multiple corresponding sample data. The splicing component can input each sample data into the recommendation model.

[0149] Recommendation Model: For any one sample data, the recommendation model can perform probability prediction processing on it to obtain the sample prediction probability corresponding to this sample data (such as Figure 4 the S shown).

[0150] Monotonic Transformation Component Corresponding to a Certain Business Indicator: The sample prediction probability S output by the recommendation model can be shared by the monotonic transformation components of each business indicator. Any monotonic transformation component of a business indicator can obtain the monotonic transformation parameter corresponding to this business indicator, and perform monotonic transformation processing on the sample prediction probability S based on this monotonic transformation parameter, so as to transform the sample prediction probability S into a value that meets the requirements of the business indicator without changing its original order. This value can be called the sample transformation probability.

[0151] Further, after obtaining the sample transformation probabilities calculated by the monotonic transformation components through transformation, the total loss value for training and optimizing the recommendation model can be calculated based on these sample transformation probabilities and the true sample labels of the sample data under the business indicator. According to this total loss value, the recommendation model can be trained and optimized.

[0152] Further, to better illustrate the process of training and optimizing the recommendation model, the following will elaborate on the specific process of model training and optimization in combination with the accompanying drawings. Please refer to Figure 5 , Figure 5 which is a schematic flowchart of training and optimizing a recommendation model provided by an embodiment of the present application. Among them, this process can correspond to the process of training and optimizing the recommendation model through N sample transformation probabilities in the corresponding embodiment of the above Figure 3 as shown.Figure 5 As shown in Figure 5 , the process may at least include the following steps S501 - S504:

[0153] Step S501, obtain the cross - entropy loss function.

[0154] Specifically, for any business metric, after obtaining the corresponding sample transformation probability, the cross - entropy loss function can be used to calculate the prediction error value of the sample transformation probability.

[0155] Step S502, obtain the prediction error value corresponding to each sample transformation probability through the cross - entropy loss function.

[0156] Specifically, in the specific implementation, the specific implementation process of obtaining the prediction error value corresponding to each sample transformation probability through the cross - entropy loss function may include but is not limited to: First, any one of the N sample transformation probabilities can be determined as the target sample transformation probability, and the business metric corresponding to the target sample transformation probability can be determined as the target business metric; based on the historical business operations performed by the sample object on the historical resource data, the true sample label of the sample data under the target business metric can be obtained; for example, for a certain historical resource data, the sample object has performed click operations, like operations, and favorite operations on it. The click operations, like operations, and favorite operations can all be used as historical business operations. Suppose the target business metric is the click - through rate. Then, based on these historical business operations, it can be known that the sample object has performed a click operation on this historical resource data. Through this historical business operation, the true sample label of this sample data under the click - through rate business metric can be determined as 1 (1 can be used to represent that the click operation has been performed); and suppose the target business metric is the comment rate. Then, based on these historical business operations, it can be known that the sample object has not performed a comment operation on this historical resource data. Through the performed historical business operations, the true sample label of this sample data under the comment rate business metric can be determined as 0. That is to say, this application needs to first obtain the business operation uniquely indicated by the target business metric, and then determine whether the sample object has performed the business operation uniquely indicated by the target business metric on the corresponding historical resource data. If it has been performed, the true sample label can be determined as the value representing that the operation has been performed (such as the value 1); if it has not been performed, the true sample label can be determined as the value representing that the operation has not been performed (such as the value 0). Further, when the sample transformation probability and the true sample label of the sample data under the target business metric are both known, the cross - entropy loss function can be used to perform error calculation processing on the target sample transformation probability and the true sample label of the sample data under the target business metric to obtain the prediction error value corresponding to the target sample transformation probability.

[0157] For the specific implementation process of calculating the error between the transformed probability of the target sample and the true sample label of the sample data under the target business metric through the cross-entropy loss function to obtain the prediction error value corresponding to the transformed probability of the target sample, it can be shown as in formula (4):

[0158] L t = -y t logp t - (1 - y t ) log(1 - p t ) Formula (4)

[0159] Among them, the function shown in formula (4) can be used to represent the cross-entropy loss function; p shown in formula (4) t can be used to represent the transformed probability of the sample corresponding to the t-th business metric (such as the target business metric) (such as the transformed probability of the target sample); y t can be used to represent the true sample label of the sample data under the t-th business metric; L t can be used to represent the calculated prediction error value corresponding to the t-th business metric.

[0160] It should be understood that for the transformed probability of any business metric, the corresponding prediction error value can be calculated in the manner shown in formula (4).

[0161] Step S503, perform a summation operation on the prediction error values corresponding to the N transformed probabilities respectively to obtain the total loss value corresponding to the N transformed probabilities together.

[0162] Specifically, after calculating the prediction error value corresponding to each business metric (i.e., each transformed probability) respectively, the N prediction error values can be summed to obtain the total loss value corresponding to the N transformed probabilities together. It should be noted that this application can configure different weight coefficients for each business metric to control the degree to which the output result of the recommendation model biases towards a certain business metric. Then, after obtaining the prediction error values corresponding to each business metric respectively, the bias operation can be performed on the prediction error value corresponding to each transformed probability based on the weight coefficient of the corresponding business metric first to obtain N bias error values (a bias error value is obtained by performing a bias operation on the corresponding prediction error value using the weight coefficient of a business metric), and then, the N bias error values are summed to obtain the total loss value corresponding to the N transformed probabilities together.

[0163] In a specific implementation, for the prediction error values corresponding to each sample transformation probability, the specific implementation process of obtaining N bias error values may include, but is not limited to: First, for any sample transformation probability, the sample transformation probability can be determined as the target sample transformation probability, and the service indicator corresponding to the target sample transformation probability can be determined as the target service indicator; then, the weight coefficient configured for the target service indicator can be obtained; based on the above, in the process of training and optimizing the recommendation model, the weight coefficient of the target service indicator is used to control the degree to which the output result of the recommendation model biases towards the target service indicator; further, the weight coefficient of the target service indicator can be multiplied by the target sample transformation probability, and thus the bias error value corresponding to the target sample transformation probability can be obtained.

[0164] In summary, the specific process of generating the total loss value can be as shown in formula (5):

[0165]

[0166] Among them, L shown in formula (5) t can be used to represent the prediction error value corresponding to the t-th service indicator calculated by formula (4); α t can be used to represent the weight coefficient corresponding to the t-th service indicator, which can be set according to online effects or business needs, and is mainly used to control the degree to which the output result of the recommendation model training biases towards the sorting factor of a certain service indicator. The value of α t is greater than 0, and can all be initially set to 1 and then adjusted relatively; T shown in formula (5) can be used to represent the total number of service indicators; L can be used to represent the calculated total loss value.

[0167] Step S504, train and optimize the model parameters of the recommendation model based on the total loss value.

[0168] Specifically, based on the total loss value, the model parameters of the recommendation model can be trained and optimized. For example, after obtaining the total loss value, the model iteration number of the current recommendation model can be obtained first. If the model iteration number does not meet the model convergence condition, the total loss value can be backpropagated to optimize the model parameters of the recommendation model based on the total loss value. Of course, if the model iteration number of the recommendation model has met the model convergence condition, then the current recommendation model that meets the model convergence condition can be directly determined as the model that has been trained and optimized, and there is no need to calculate the loss value for optimization.

[0169] It should be noted that after calculating the prediction error value corresponding to a certain business indicator in this application, the prediction error value under this business indicator can be used to train and optimize the monotonic transformation parameter of this business indicator, so that the sample transformation probability after monotonic transformation can get closer and closer to the true sample label of the sample data under this business indicator. For the monotonic transformation parameters under each business indicator, they can be trained and optimized synchronously with the recommendation model, or they can be trained and optimized prior to the recommendation model. In the case of being trained and optimized prior to the recommendation model, the monotonic transformation parameters of each business indicator already have a sufficiently high accuracy. After calculating the prediction error values of each business indicator, it is only necessary to calculate the total loss value in the above-described manner to train and optimize the recommendation model; while in the case of being trained and optimized synchronously with the recommendation model, after calculating the prediction error values of each business indicator, it is also necessary to train and optimize the monotonic transformation parameter of this business indicator based on the prediction error value of a single business indicator to synchronously optimize this monotonic transformation parameter until the recommendation model meets the model convergence condition.

[0170] In summary, it can be seen that the monotonic transformation in the embodiments of this application does not change the order of the original values. Then, after the business indicator performs a monotonic transformation on each sample prediction probability, the original sorting situation can still be maintained, and at the same time, the sample prediction probability is transformed into a value that is more in line with the true situation of the business indicator (that is, the probability used to reflect the business operation indicated by the sample object executing the corresponding business indicator alone); finally, the recommendation model can be trained and optimized based on the sample transformation probabilities of each business indicator, so that the prediction probability output by the recommendation model can improve the accuracy of the prediction probability under the joint influence of each business indicator, and based on the accurate prediction probability, optimize the sorting result of the resource data and improve the recommendation accuracy of the resource data.

[0171] Further, please refer to Figure 6 , Figure 6 which is a schematic structural diagram of a data processing device provided by an embodiment of this application. This data processing device can be a computer program (including program code) running in a computer device. For example, this data processing device is an application software; this data processing device can be used to execute Figure 3 the method shown. As Figure 6 shown, this data processing device 1 may include: a sample acquisition module 11, a prediction module 12, a monotonic transformation module 13, and a training optimization module 14.

[0172] The sample acquisition module 11 is used to acquire sample data for model training; the sample data is determined based on the sample object and the historical resource data of the historical business operations performed by the sample object;

[0173] The prediction module 12 is configured to call a recommendation model to perform probability prediction processing on sample data, and obtain a sample prediction probability; the sample prediction probability refers to the probability that the sample object performs the business operation indicated by N business indicators; N is a positive integer;

[0174] The monotonic transformation module 13 is configured to perform monotonic transformation processing on the sample prediction probability respectively through the monotonic transformation parameters of each business indicator, and obtain a sample transformation probability corresponding to each business indicator; a sample transformation probability is used to characterize the probability that the sample object performs the business operation indicated by the corresponding business indicator alone;

[0175] The training and optimization module 14 is configured to train and optimize the recommendation model through N sample transformation probabilities; the trained and optimized recommendation model is used to perform probability prediction processing on business data in a resource data push scenario; the business data is determined based on the business object and the candidate push resource data of the business object.

[0176] Among them, for the specific implementation manners of the sample acquisition module 11, the prediction module 12, the monotonic transformation module 13, and the training and optimization module 14, reference can be made to the descriptions of steps S101 - S104 in the corresponding embodiments above, and details will not be elaborated here. Figure 3 The description of steps S101 - S104 in the corresponding embodiments above will not be repeated here.

[0177] In one embodiment, the specific implementation manner for the sample acquisition module 11 to obtain the sample data for model training includes:

[0178] Obtain the feature data of the sample object and the feature data of the historical resource data; the historical resource data refers to the resource data for which the sample object has performed historical business operations; the feature data of the sample object includes the feature data of N1 first feature domains, and the feature data of the historical resource data includes the feature data of N2 second feature domains; both N1 and N2 are positive integers;

[0179] Obtain the first feature transformation vector corresponding to each first feature domain;

[0180] Obtain the second feature transformation vector corresponding to each second feature domain;

[0181] Perform vector splicing on the N1 first feature transformation vectors and the N2 second feature transformation vectors to obtain the sample data.

[0182] Among them, the method for obtaining the feature transformation vector includes: performing fixed-length vector transformation on the feature data of the feature domain to obtain the feature transformation vector.

[0183] In one embodiment, the specific implementation manner for the sample acquisition module 11 to perform fixed-length vector transformation on the feature data of the feature domain to obtain the feature transformation vector includes:

[0184] Obtain the domain type to which the feature domain belongs;

[0185] When the domain type is a fixed-length type, perform vector embedding processing on the feature data of the feature domain through a feature embedding network to obtain a feature transformation vector of the feature domain;

[0186] When the domain type is a variable-length type, perform vector mapping processing and pooling processing on the feature data of the feature domain through a sequence network to obtain a feature transformation vector of the feature domain.

[0187] In one embodiment, the specific implementation manner in which the monotonic transformation module 13 performs monotonic transformation processing on the sample prediction probability through the monotonic transformation parameter of each business indicator to obtain the sample transformation probability corresponding to each business indicator respectively includes:

[0188] Determine any one of the N business indicators as the transformation business indicator, and determine the monotonic transformation parameter of the transformation business indicator as the target monotonic transformation parameter;

[0189] Through the scaling transformation function in the monotonic transformation rule, perform scaling transformation on the target monotonic transformation parameter and the sample prediction probability to obtain the initial sample transformation probability of the transformation business indicator;

[0190] Through the activation transformation function in the monotonic transformation rule, perform activation transformation on the initial sample transformation probability to obtain the sample transformation probability corresponding to the transformation business indicator.

[0191] In one embodiment, the target monotonic transformation parameter includes a first trainable parameter and a second trainable parameter;

[0192] The specific implementation manner in which the monotonic transformation module 13 performs scaling transformation on the target monotonic transformation parameter and the sample prediction probability through the scaling transformation function in the monotonic transformation rule to obtain the initial sample transformation probability of the transformation business indicator includes:

[0193] Perform a multiplication operation on the first trainable parameter and the sample prediction probability to obtain a first multiplication result;

[0194] Perform a multiplication operation on the first multiplication result and the first trainable parameter to obtain a second multiplication result;

[0195] Perform a summation operation on the second multiplication result and the second trainable parameter to obtain the initial sample transformation probability of the transformation business indicator.

[0196] In one embodiment, the specific implementation manner in which the training optimization module 14 trains and optimizes the recommendation model through N sample transformation probabilities includes:

[0197] Obtain the cross-entropy loss function;

[0198] Obtain the prediction error values corresponding to the transformation probabilities of each sample through the cross-entropy loss function;

[0199] Perform a summation operation on the prediction error values corresponding to the transformation probabilities of N samples to obtain the total loss value corresponding to the transformation probabilities of N samples;

[0200] Train and optimize the model parameters of the recommendation model based on the total loss value.

[0201] In one embodiment, the specific implementation manner for the training and optimization module 14 to obtain the prediction error values corresponding to the transformation probabilities of each sample through the cross-entropy loss function includes:

[0202] Determine any one of the transformation probabilities of N samples as the target transformation probability, and determine the service metric corresponding to the target transformation probability as the target service metric;

[0203] Based on the historical service operations performed by the sample object on the historical resource data, obtain the true sample label of the sample data under the target service metric;

[0204] Through the cross-entropy loss function, perform an error calculation process on the target transformation probability and the true sample label of the sample data under the target service metric to obtain the prediction error value corresponding to the target transformation probability.

[0205] In one embodiment, the specific implementation manner for the training and optimization module 14 to perform a summation operation on the prediction error values corresponding to the transformation probabilities of N samples to obtain the total loss value corresponding to the transformation probabilities of N samples includes:

[0206] Perform a bias operation on the prediction error value corresponding to each sample transformation probability respectively to obtain N bias error values;

[0207] Perform a summation operation on the N bias error values to obtain the total loss value corresponding to the transformation probabilities of N samples.

[0208] In one embodiment, the specific implementation manner for the training and optimization module 14 to perform a bias operation on the prediction error value corresponding to each sample transformation probability respectively to obtain N bias error values includes:

[0209] Determine any one of the transformation probabilities of N samples as the target transformation probability, and determine the service metric corresponding to the target transformation probability as the target service metric;

[0210] Obtain the weight coefficient configured for the target service metric; during the process of training and optimizing the recommendation model, the weight coefficient of the target service metric is used to control the degree to which the output result of the recommendation model biases towards the target service metric;

[0211] Multiply the weight coefficient of the target business metric by the target sample transformation probability to obtain the bias error value corresponding to the target sample transformation probability.

[0212] In one embodiment, the specific implementation manner in which the training and optimization module 14 trains and optimizes the model parameters of the recommendation model based on the total loss value includes:

[0213] Obtain the model iteration count of the recommendation model;

[0214] When the model iteration count does not meet the model convergence condition, train and optimize the model parameters of the recommendation model based on the total loss value.

[0215] In one embodiment, after the training and optimization module 14 trains and optimizes the recommendation model through N sample transformation probabilities, the data processing device 1 further includes: a candidate data acquisition module 15, a service data acquisition module 16, a model prediction module 17, and a data push module 18.

[0216] The candidate data acquisition module 15 is configured to acquire M candidate push resource data regarding a service object; M is a positive integer;

[0217] The service data acquisition module 16 is configured to acquire the service data respectively corresponding to the M candidate push resource data; the service data corresponding to a candidate push resource data is determined based on the service object and the candidate push resource data;

[0218] The model prediction module 17 is configured to call the trained and optimized recommendation model to perform probability prediction processing on the service data corresponding to each candidate push resource data respectively, to obtain the prediction probability respectively corresponding to each candidate push resource data;

[0219] The data push module 18 is configured to determine the push resource data of the service object from the M candidate push resource data based on the M prediction probabilities, and push the push resource data to the service object.

[0220] Among them, for the specific implementation manners of the candidate data acquisition module 15, the service data acquisition module 16, the model prediction module 17, and the data push module 18, reference can be made to the relevant descriptions in step S104 in the corresponding embodiment above, and details will not be elaborated here. Figure 3 The relevant description in step S104 of the corresponding embodiment will not be repeated here.

[0221] In one embodiment, the specific implementation manner in which the data push module 18 determines the push resource data of the service object based on the M prediction probabilities and pushes the push resource data to the service object includes:

[0222] Sort the M candidate push resource data in descending order according to the magnitude order among the M prediction probabilities to obtain a resource data sequence;

[0223] Determine the first K candidate push resource data in the resource data sequence as the push resource data of the service object; K is a positive integer less than or equal to M.

[0224] The monotonic transformation in the embodiments of the present application does not change the order of the original values. Therefore, after the service metrics perform monotonic transformation on the sample prediction probabilities, the original sorting situation can still be maintained. At the same time, the sample prediction probabilities are transformed into values that more conform to the true situation of the service metrics (that is, the probabilities used to reflect the probabilities of sample objects performing service operations indicated by the corresponding service metrics alone). Finally, the recommendation model can be jointly trained and optimized based on the sample transformation probabilities of each service metric, so that the prediction probabilities output by the recommendation model can improve the accuracy of the prediction probabilities under the joint influence of each service metric, and optimize the sorting result of the resource data based on the accurate prediction probabilities to improve the recommendation accuracy of the resource data.

[0225] Further, please refer to Figure 7 , Figure 7 which is a schematic structural diagram of a computer device provided by an embodiment of the present application. As Figure 7 shown, the above computer device 8000 may include: a processor 8001, a network interface 8004, and a memory 8005. In addition, the above computer device 8000 further includes: a user interface 8003 and at least one communication bus 8002. Among them, the communication bus 8002 is used to realize the connection and communication between these components. Among them, the user interface 8003 may include a display screen (Display) and a keyboard (Keyboard). Optionally, the user interface 8003 may further include a standard wired interface and a wireless interface. The network interface 8004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 8005 may be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. Optionally, the memory 8005 may further be at least one storage device located far from the aforementioned processor 8001. As Figure 7 shown, the memory 8005, as a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application program.

[0226] In Figure 7 the computer device 8000 shown, the network interface 8004 can provide network communication functions; while the user interface 8003 is mainly used to provide an input interface for users; and the processor 8001 can be used to call the device control application program stored in the memory 8005 to implement:

[0227] Obtain sample data for model training; the sample data is determined based on the sample object and the historical resource data on which the sample object has performed historical business operations;

[0228] Call the recommendation model to perform probability prediction processing on the sample data to obtain a sample prediction probability; the sample prediction probability refers to the probability of predicting that the sample object will perform the business operation indicated jointly by N business indicators; N is a positive integer;

[0229] Perform monotonic transformation processing on the sample prediction probability through the monotonic transformation parameters of each business indicator respectively to obtain the sample transformation probability corresponding to each business indicator; a sample transformation probability is used to characterize the probability that the sample object will perform the business operation indicated separately by the corresponding business indicator;

[0230] Train and optimize the recommendation model through N sample transformation probabilities; the trained and optimized recommendation model is used to perform probability prediction processing on business data in the resource data push scenario; the business data is determined based on the business object and the candidate push resource data of the business object.

[0231] It should be understood that the computer device 8000 described in the embodiments of the present application can execute the description of the data processing method in the corresponding embodiments mentioned above, and can also execute the description of the data processing device 1 in the corresponding embodiments mentioned above, which will not be elaborated here. In addition, the description of the beneficial effects of using the same method will not be elaborated either. Figures 3 to 5 The description of the data processing method in the corresponding embodiments mentioned above, and can also execute the description of the data processing device 1 in the corresponding embodiments mentioned above, which will not be elaborated here. In addition, the description of the beneficial effects of using the same method will not be elaborated either. Figure 6 In addition, it should be noted here that: the embodiments of the present application also provide a computer-readable storage medium, and the computer-readable storage medium stores a computer program executed by the computer device 8000 for data processing mentioned above, and the computer program includes program instructions. When the above-mentioned processor executes the above-mentioned program instructions, it can execute the description of the above-mentioned data processing method in the corresponding embodiments mentioned above. Therefore, it will not be elaborated here. In addition, the description of the beneficial effects of using the same method will not be elaborated either. For the technical details not disclosed in the embodiments of the computer-readable storage medium involved in the present application, please refer to the description of the method embodiments of the present application.

[0232] Figures 3 to 5 The description of the data processing method in the corresponding embodiments mentioned above, and can also execute the description of the data processing device 1 in the corresponding embodiments mentioned above, which will not be elaborated here. In addition, the description of the beneficial effects of using the same method will not be elaborated either. For the technical details not disclosed in the embodiments of the computer-readable storage medium involved in the present application, please refer to the description of the method embodiments of the present application.

[0233] ​The above computer-readable storage medium may be the data processing device provided in any of the foregoing embodiments or the internal storage unit of the above computer device, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device. Further, the computer-readable storage medium may also include both the internal storage unit and the external storage device of the computer device. The computer-readable storage medium is used to store the computer program and other programs and data required by the computer device. The computer-readable storage medium may also be used to temporarily store the data that has been output or is to be output.

[0234] In one aspect of the present application, a computer program product is provided. The computer program product includes a computer program, and the computer program is stored in a computer-readable storage medium. The processor of the computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in one aspect of the embodiments of the present application.

[0235] In the description of the embodiments of the present application, the terms "first", "second", etc. in the specification, claims and drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the term "comprising" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or equipment that includes a series of steps or units is not limited to the listed steps or modules, but may optionally further include steps or modules not listed, or may optionally further include other step units inherent to these processes, methods, devices, products or equipment.

[0236] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0237] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other related parts to achieve a predetermined goal, and can be fully or partially implemented by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the function of the module or unit.

[0238] The methods and related devices provided in the embodiments of the present application are described with reference to the method flowcharts and / or structural schematic diagrams provided in the embodiments of the present application. Specifically, each process and / or block of the method flowchart and / or structural schematic diagram, and the combination of the processes and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the function specified in Figure 1 one process or multiple processes and / or structural schematic Figure 1 one block or multiple blocks. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the function specified in Figure 1 one process or multiple processes and / or structural schematic Figure 1 one block or multiple blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the function specified in Figure 1 one process or multiple processes and / or structural schematic one block or multiple blocks.

[0239] The above-disclosed are only the preferred embodiments of the present application. Of course, the scope of the rights of the present application cannot be limited thereby. Therefore, equivalent changes made according to the claims of the present application still fall within the scope covered by the present application.

Claims

1. A data processing method, characterized in that, The method includes: Obtaining sample data for model training; the sample data is determined based on the sample object and the historical resource data of the historical business operations performed by the sample object; Invoking a recommendation model to perform probability prediction processing on the sample data to obtain a sample prediction probability; the sample prediction probability refers to the probability of predicting that the sample object performs the business operation jointly indicated by N business indicators; N is a positive integer; Performing monotonic transformation processing on the sample prediction probability through the monotonic transformation parameter of each business indicator respectively to obtain a sample transformation probability corresponding to each business indicator; one sample transformation probability is used to characterize the probability of the sample object performing the business operation separately indicated by the corresponding business indicator; Training and optimizing the recommendation model through N sample transformation probabilities; the trained and optimized recommendation model is used to perform probability prediction processing on business data in a resource data push scenario; the business data is determined based on a business object and the candidate push resource data of the business object.

2. The method according to claim 1, wherein The obtaining of the sample data for model training includes: Obtaining the feature data of the sample object and the feature data of the historical resource data; the historical resource data refers to the resource data of the historical business operations performed by the sample object; the feature data of the sample object includes the feature data of N1 first feature domains, and the feature data of the historical resource data includes the feature data of N2 second feature domains; N1 and N2 are both positive integers; Obtaining a first feature transformation vector corresponding to each first feature domain; Obtaining a second feature transformation vector corresponding to each second feature domain; Performing vector splicing on the N1 first feature transformation vectors and the N2 second feature transformation vectors to obtain the sample data; Among them, the method for obtaining the feature transformation vector includes: performing fixed-length vector transformation on the feature data of the feature domain to obtain the feature transformation vector.

3. The method according to claim 2, characterized in that, The performing of fixed-length vector transformation on the feature data of the feature domain to obtain the feature transformation vector includes: Obtaining the domain type to which the feature domain belongs; When the domain type is a fixed-length type, performing vector embedding processing on the feature data of the feature domain through a feature embedding network to obtain the feature transformation vector of the feature domain; When the domain type is a variable-length type, performing vector mapping processing and pooling processing on the feature data of the feature domain through a sequence network to obtain the feature transformation vector of the feature domain.

4. The method according to claim 1, wherein The performing of monotonic transformation processing on the sample prediction probability through the monotonic transformation parameter of each business indicator respectively to obtain a sample transformation probability corresponding to each business indicator includes: Determining any one of the N business indicators as the transformation business indicator and determining the monotonic transformation parameter of the transformation business indicator as the target monotonic transformation parameter; Performing scaling transformation on the target monotonic transformation parameter and the sample prediction probability through the scaling transformation function in the monotonic transformation rule to obtain the initial sample transformation probability of the transformation business indicator; Through the activation transformation function in the monotonic transformation rule, the initial sample transformation probability is activated and transformed to obtain the sample transformation probability corresponding to the transformed service metric.

5. The method according to claim 4, characterized in that The target monotonic transformation parameter includes a first trainable parameter and a second trainable parameter; The process of obtaining the initial sample transformation probability of the transformed service metric by scaling and transforming the target monotonic transformation parameter and the sample prediction probability through the scaling transformation function in the monotonic transformation rule includes: Performing a multiplication operation on the first trainable parameter and the sample prediction probability to obtain a first multiplication result; Performing a multiplication operation on the first multiplication result and the first trainable parameter to obtain a second multiplication result; Performing a summation operation on the second multiplication result and the second trainable parameter to obtain the initial sample transformation probability of the transformed service metric.

6. The method according to claim 1, characterized in that, The process of training and optimizing the recommendation model with N sample transformation probabilities includes: Obtaining a cross-entropy loss function; Obtaining the prediction error value corresponding to each sample transformation probability through the cross-entropy loss function; Performing a summation operation on the prediction error values corresponding to the N sample transformation probabilities to obtain the total loss value jointly corresponding to the N sample transformation probabilities; Training and optimizing the model parameters of the recommendation model based on the total loss value.

7. The method according to claim 6, characterized in that, The process of obtaining the prediction error value corresponding to each sample transformation probability through the cross-entropy loss function includes: Determining any one of the N sample transformation probabilities as the target sample transformation probability, and determining the service metric corresponding to the target sample transformation probability as the target service metric; Based on the historical service operations performed by the sample object on the historical resource data, obtaining the true sample label of the sample data under the target service metric; Through the cross-entropy loss function, performing an error calculation process on the target sample transformation probability and the true sample label of the sample data under the target service metric to obtain the prediction error value corresponding to the target sample transformation probability.

8. The method according to claim 6, characterized in that, The process of performing a summation operation on the prediction error values corresponding to the N sample transformation probabilities to obtain the total loss value jointly corresponding to the N sample transformation probabilities includes: Performing a bias operation on the prediction error value corresponding to each sample transformation probability respectively to obtain N bias error values; Performing a summation operation on the N bias error values to obtain the total loss value jointly corresponding to the N sample transformation probabilities.

9. The method according to claim 8, wherein The process of performing a bias operation on the prediction error value corresponding to each sample transformation probability respectively to obtain N bias error values includes: Determining any one of the N sample transformation probabilities as the target sample transformation probability, and determining the service metric corresponding to the target sample transformation probability as the target service metric; Obtaining the weight coefficient configured for the target service metric; during the process of training and optimizing the recommendation model, the weight coefficient of the target service metric is used to control the degree to which the output result of the recommendation model biases towards the target service metric. Multiply the weight coefficient of the target business metric by the target sample transformation probability to obtain the bias error value corresponding to the target sample transformation probability.

10. The method according to claim 6, characterized in that, The training and optimization of the model parameters of the recommendation model based on the total loss value includes: Obtain the model iteration count of the recommendation model; When the model iteration count does not meet the model convergence condition, train and optimize the model parameters of the recommendation model based on the total loss value.

11. The method according to claim 1, wherein After training and optimizing the recommendation model with N sample transformation probabilities, the method further includes: Obtain M candidate push resource data for a business object; M is a positive integer; Obtain the business data corresponding to each of the M candidate push resource data; the business data corresponding to one candidate push resource data is determined based on the business object and the candidate push resource data. Call the trained and optimized recommendation model to perform probability prediction processing on the business data corresponding to each candidate push resource data, respectively, to obtain the prediction probability corresponding to each candidate push resource data. Based on the M prediction probabilities, determine the push resource data for the business object from the M candidate push resource data, and push the push resource data to the business object.

12. The method according to claim 11, wherein The determining the push resource data for the business object based on the M prediction probabilities and pushing the push resource data to the business object includes: Sort the M candidate push resource data in descending order according to the magnitude order among the M prediction probabilities to obtain a resource data sequence; Determine the first K candidate push resource data in the resource data sequence as the push resource data for the business object; K is a positive integer less than or equal to M.

13. A data processing device, characterized in that, including: A sample acquisition module for acquiring sample data for model training; The sample data is determined based on a sample object and historical resource data on which the sample object has performed historical business operations; A prediction module for calling a recommendation model to perform probability prediction processing on the sample data to obtain a sample prediction probability; the sample prediction probability refers to the probability of predicting that the sample object performs the business operation jointly indicated by N business metrics; N is a positive integer; A monotonic transformation module for performing monotonic transformation processing on the sample prediction probability through the monotonic transformation parameter of each business metric, respectively, to obtain the sample transformation probability corresponding to each business metric; one sample transformation probability is used to characterize the probability that the sample object performs the business operation separately indicated by the corresponding business metric. A training and optimization module for training and optimizing the recommendation model with N sample transformation probabilities; the trained and optimized recommendation model is used to perform probability prediction processing on business data in a resource data push scenario; the business data is determined based on a business object and the candidate push resource data of the business object.

14. A computer device, characterized in that, including: A processor, a memory, and a network interface; The processor is connected to the memory and the network interface. Among them, the network interface is used to provide network communication functions, the memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device executes the method described in any one of claims 1-12.

15. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, and the computer program is adapted to be loaded and executed by a processor to execute the method described in any one of claims 1-12.

16. A computer program product, characterized in that, The computer program product includes a computer program. The computer program is stored in a computer-readable storage medium and is adapted to be read and executed by a processor so that a computer device having the processor executes the method described in any one of claims 1-12.

Images