CN115456085A

CN115456085A - Model training method, resource estimation method, device, electronic device and medium

Info

Publication number: CN115456085A
Application number: CN202211132002.7A
Authority: CN
Inventors: 吴学超; 周杨; 白云龙; 刘洁; 郝昌臻
Original assignee: Baidu com Times Technology Beijing Co Ltd
Current assignee: Baidu com Times Technology Beijing Co Ltd
Priority date: 2022-09-16
Filing date: 2022-09-16
Publication date: 2022-12-09

Abstract

The disclosure provides a model training method, a resource estimation device, electronic equipment and a storage medium, and relates to the field of artificial intelligence, in particular to a deep learning technology. The specific implementation scheme comprises the following steps: acquiring a training sample under a specified service scene; wherein a plurality of business targets exist in the designated business farm; aiming at any training sample, extracting general features and business features related to business targets from the training sample; and training the multi-target estimation model according to the general characteristics and the business characteristics related to the business target. The scheme disclosed by the invention achieves the purpose of completing the training of the multi-target pre-estimation model by using the characteristic system of the general characteristic and the business characteristic under the condition that the characteristic system is not uniform, and ensures the pre-estimation effect of the multi-target pre-estimation model.

Description

Model training method, resource estimation method, device, electronic equipment and medium

Technical Field

The present disclosure relates to the field of artificial intelligence, in particular to the field of deep learning technology, and in particular to a multi-objective prediction model training method, a resource prediction method, an apparatus, an electronic device, a storage medium, and a computer program product.

Background

In many internet service (e.g., e-commerce, news reading, video playing, etc.) scenes nowadays, it is more and more common for users to passively browse and consume information recommended by service parties, so that corresponding content needs to be recommended to the users by a recommendation system, and the recommendation system can pre-estimate and rank candidate content to be recommended to provide better recommendation service to the users.

Disclosure of Invention

The disclosure provides a multi-target pre-estimation model training method, a resource pre-estimation method, a device, electronic equipment, a storage medium and a computer program product.

According to an aspect of the disclosure, a multi-target estimation model training method is provided, including:

acquiring a training sample under a specified service scene; wherein a plurality of business targets exist in the designated business farm;

extracting general features and business features related to business targets from the training samples aiming at any training sample;

and training the multi-target pre-estimation model according to the general characteristics and the business characteristics related to the business target.

According to an aspect of the present disclosure, a resource estimation method is provided, including:

acquiring target resources to be estimated;

extracting general characteristics and business characteristics related to the business target from the target resource;

predicting and predicting the general characteristics and the service characteristics through a multi-target prediction model to obtain a prediction result of target resources; the multi-target estimation model is obtained after training according to any model training method in the embodiment.

According to an aspect of the present disclosure, there is provided a multi-objective predictive model training apparatus, including:

the data acquisition module is used for acquiring a training sample under a specified service scene; wherein a plurality of business targets exist in the designated business site;

the feature extraction module is used for extracting general features and business features related to business targets from the training samples aiming at any training sample;

and the training module is used for training the multi-target estimation model according to the general characteristics and the business characteristics related to the business target.

According to an aspect of the present disclosure, there is provided a resource estimation apparatus, including:

the resource acquisition module is used for acquiring target resources to be estimated;

the resource feature extraction module is used for extracting general features and business features related to business targets from target resources;

the pre-estimation module is used for performing prediction and pre-estimation on the general characteristics and the service characteristics through a multi-target pre-estimation model to obtain a pre-estimation result of target resources; the multi-target estimation model is obtained after training according to any model training method in the embodiment.

According to another aspect of the present disclosure, there is provided an electronic device including:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein, the first and the second end of the pipe are connected with each other,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a multi-objective predictive model training method or a resource predictive method of any of the embodiments of the disclosure.

According to another aspect of the present disclosure, there is provided a non-transitory computer readable storage medium storing computer instructions for causing a computer to perform the multi-objective predictive model training method or the resource predictive method of any of the embodiments of the present disclosure.

According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements the multi-objective predictive model training method or the resource predictive method of any of the embodiments of the present disclosure.

According to the technology disclosed by the invention, the aim of completing the training of the multi-target pre-estimation model by using the characteristic system of the universal characteristic + the business characteristic is fulfilled under the condition that the characteristic systems of the training samples corresponding to the targets are not uniform; and the trained multi-target estimation model is used for resource estimation, so that the estimation accuracy can be ensured.

It should be understood that the statements in this section are not intended to identify key or critical features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:

FIG. 1 is a schematic flow chart diagram of a multi-target estimation model training method according to an embodiment of the present disclosure;

FIG. 2a is a schematic flow chart of another multi-target estimation model training method provided in the embodiment of the present disclosure;

FIG. 2b is a schematic structural diagram of a multi-target prediction model provided in the embodiment of the present disclosure;

FIG. 3 is a schematic flow chart diagram illustrating another multi-target estimation model training method provided in the embodiment of the present disclosure;

FIG. 4 is a schematic flowchart of a resource estimation method according to an embodiment of the disclosure;

FIG. 5 is a schematic structural diagram of a multi-target prediction model training apparatus according to an embodiment of the present disclosure;

fig. 6 is a schematic structural diagram of a resource estimation apparatus according to an embodiment of the present disclosure;

FIG. 7 is a block diagram of an electronic device for implementing a multi-objective predictive model training method according to an embodiment of the disclosure.

Detailed Description

Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, in which various details of the embodiments of the disclosure are included to assist understanding, and which are to be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

At present, many application software adopt the form of fed stream to push information such as news information to users. When a user carries out an FEED refreshing action, a client firstly sends a request to the resource aggregation module, and the resource aggregation module sends a request for recalling resources to each downstream queue. And each queue recalls the resources by adopting different strategies according to different resource attributes, scores the resources through stages of rough arrangement, fine arrangement, rearrangement and the like in the sequencing stage, and returns the scores to the resource aggregation module. And the convergence module performs the operations of duplication elimination, fusion, filtration, sequencing and the like on the resources of each queue, generates a final issuing list and returns the final issuing list to the client. And the client presents the resource result which finally accords with the user interest in front of the user. In the whole recommendation process, if a user wants to recommend high-quality resources, the recalled resources need to be pre-estimated and sorted.

In this embodiment, in order to ensure the recommendation effect, a single pre-estimation ranking model is trained for each service target respectively for a plurality of service targets (for example, resource click rate, resource playing duration, and interactive behavior) in the recommendation scene, and the trained pre-estimation ranking models are used to predict the corresponding service targets. However, it has been found through experiments that this method has certain disadvantages: training a single recommendation model for each business target, which occupies excessive system resources; for a business target with a small training sample size, the problem of sparse data features seriously causes insufficient model training and influences estimation accuracy. If a unified multi-target pre-estimation model is trained for a plurality of business targets, the problems of data sparseness and resource waste are solved to a certain extent, but the mode requires consistent characteristic systems and is not beneficial to designing characteristics of all the business targets according to own characteristics. For a specific implementation process, see the following examples.

Fig. 1 is a schematic flow chart of a multi-target estimation model training method according to an embodiment of the present disclosure, which is applicable to a situation where training data corresponding to different business targets are used to train a multi-target estimation model in a recommendation scenario. The method can be executed by a multi-target pre-estimation model training device which is realized in a software and/or hardware mode and is integrated on electronic equipment.

Specifically, referring to fig. 1, the multi-target estimation model training method is as follows:

s101, acquiring a training sample under a specified service scene; wherein a plurality of business targets exist in the designated business arena.

In this embodiment, the specified service scenario is, for example, a recommendation scenario for recommending news information, and may also be any other recommendation scenario, which is not specifically limited herein. The service target is exemplarily at least one of a resource click rate, a resource play duration, and an interactive behavior. Due to the difference between different business targets, training samples used for training the estimation model and corresponding to the different business targets have deviation, so that the training samples under the appointed business scene actually consist of the training samples corresponding to the different business targets.

And S102, aiming at any training sample, extracting general features and business features related to business targets from the training sample.

In this embodiment, a neural network for extracting features may be constructed, where the neural network structure includes a first Embedding layer for extracting general features and a second Embedding layer for extracting personalized service features. Therefore, when the features are extracted, the training samples are only required to be input into the neural network for extracting the features, and then the general features are extracted through the first Embedding layer, and the personalized service features are extracted through the second Embedding layer. It should be noted that in any training sample, there may be business features related to different business targets at the same time, and the extracted general features and business features are represented in the form of feature vectors. In addition, when the universal features are extracted from different training samples, the same feature extraction configuration is adopted, so that the training samples corresponding to all the business targets can be ensured to have uniform feature formats on the universal features.

In this embodiment, the common features include at least one of a user feature (e.g., user identification, age, gender, etc.), an article feature (e.g., article title, category, etc.), a user behavior sequence feature (e.g., usage behavior record of the user, which may be a certain category of videos that the user continuously watches for a period of time), a user request feature (e.g., request address); the service characteristic is at least one of a resource click rate characteristic, a resource play duration characteristic and an interaction characteristic (for example, approval or forwarding of the resource and the like).

S103, training the multi-target estimation model according to the general characteristics and the business characteristics related to the business target.

In this embodiment, the multi-target estimation model is to estimate a plurality of business targets, and the multi-target estimation model needs to include a plurality of estimation subnetworks, where each estimation subnetwork corresponds to a business target. Therefore, when the multi-target estimation model is trained, the estimation sub-network corresponding to the business target is trained only by utilizing the general characteristics and the business characteristics related to the business target.

In the embodiment, the multi-target estimation model is trained by using a feature system of general features and business features, so that each business target can share training data and give consideration to the personalized business features of the business target, and the estimation accuracy of the multi-target estimation model can be ensured.

FIG. 2a is a schematic flow chart of another multi-objective prediction model training method according to an embodiment of the disclosure. Referring to fig. 2a, the multi-target estimation model training method specifically includes:

s201, obtaining a training sample under a specified service scene; wherein a plurality of business targets exist in the designated business arena.

In this embodiment, the specified service scenario is, for example, a recommendation scenario for recommending news information, and may also be any other recommendation scenario, which is not specifically limited herein. The service target is exemplarily at least one of a resource click rate, a resource play duration, and an interactive behavior.

S202, aiming at any training sample, extracting general features and business features related to business targets from the training sample.

The general characteristics comprise at least one of user characteristics, article characteristics, user behavior sequence characteristics and user request characteristics; the service characteristics are at least one of resource click rate characteristics, resource playing time characteristics and interaction characteristics.

In this embodiment, in order to describe the training process of the multi-target estimation model in detail, a network structure of the multi-target estimation model is now described. Referring to FIG. 2b, a schematic diagram of a network structure of the multi-objective prediction model is shown.

The multi-target pre-estimation model can be built by a neural network and is divided into an upper layer network and a lower layer network, wherein the lower layer network comprises a general tower network for receiving general characteristics and a service tower network for receiving service characteristics, the service tower network is divided into a plurality of service tower sub-networks, and different service tower sub-networks are used for receiving different service characteristics; the upper network comprises a pre-estimation sub-network which is used for pre-estimating resources and corresponds to each service target, illustratively, the pre-estimation sub-network comprises a pre-estimation sub-network for pre-estimating the click rate of the resources, a pre-estimation sub-network for pre-estimating the playing time of the resources and a pre-estimation sub-network for pre-estimating the interaction behavior. In the figure, ha, hb, and the like denote a layer of neural network. Thus, by setting the structure of the multi-target estimation model shown in fig. 2b, a foundation is provided for training the model by using the feature system of the general features and the business features. In addition, aiming at training samples corresponding to different business targets, general features are extracted through a general tower network, so that the general tower can share network parameters, and stable updating of model parameters can be guaranteed.

After the structure of the multi-objective predictive model is introduced, the process of training the multi-objective predictive model according to the general features and the business features related to the business objective can be referred to as S203-S206.

And S203, inputting the general features into a general tower network.

S204, aiming at the service characteristics related to any service target, inputting the service characteristics into the service tower network, and determining a first pre-estimating sub-network corresponding to the service target.

In this embodiment, after the service characteristics are input into the service tower network, a dynamic selection layer in the service tower network determines a first estimating subnetwork corresponding to the service target, for example, if the service characteristics are resource click rate characteristics, it is determined that the first estimating subnetwork is an estimating subnetwork used for estimating resource click rate in an upper network; if the service characteristic is a resource playing time characteristic, determining that the first estimating sub-network is an estimating sub-network used for estimating the resource playing time in an upper network; and if the service characteristics are the characteristics of the resource interaction behaviors, determining that the first predictor sub-network is a predictor sub-network used for predicting the resource interaction behaviors in the upper network.

S205, the output result of the general tower network and the output result of the service tower network are transmitted to the first predictor network in a forward direction.

After the first estimating sub-network is determined through S204, the output result of the general tower network and the output result of the service tower network are transmitted to the first estimating sub-network in a forward direction, so that the first estimating sub-network outputs the estimating result, and the estimating error is calculated according to the estimating result and the real result on the sample label, therefore, if the training sample simultaneously comprises the service characteristics related to all targets, the estimating error output by each estimating sub-network can be obtained through the steps.

S206, reversely updating the network parameters of the first prediction sub-network and the network parameters of the underlying network according to the prediction error output by the first prediction sub-network.

After the estimation error is obtained through S205, the overall estimation error of the model is calculated, and then the error gradient is propagated in reverse and the network parameters of the first estimation subnetwork and the network parameters of the underlying network are updated, it should be noted that, when the network parameters of the underlying service tower network are updated, the error gradient is forwarded to the corresponding service tower subnetwork through the dynamic selection layer so as to adjust the parameters of the service tower subnetwork, and the parameters of other service tower subnetworks are not adjusted during this training, so that one training of the multi-objective estimation model is completed.

Further, because the training samples corresponding to different business targets have different scales, in order to balance the influence of the sample data scale difference corresponding to each business target on the model, before updating the network parameters in a reverse direction, the method further includes: and setting corresponding training weights for the estimation errors according to the scale difference of training samples corresponding to different business targets. Wherein the training weights

Wherein, N _i Representing the sample size corresponding to any business target; n is a radical of hydrogen ₀ And the sample sum corresponding to all the service targets in the specified service scene is shown. When setting the training weight, only need to set gamma _i Multiplying the estimated error of each estimated subnetwork output.

Therefore, the overall estimation error loss of the multi-target estimation model can be calculated according to the following formula: loss = ∑ Σ _i∈R γ _i loss _i (ii) a Therein, loss _i Representing the prediction error of the outputs of the different prediction subnetworks. And further updating the network parameters according to the overall estimated error loss.

In the example, the predictor to be trained can be quickly determined according to the business characteristics through the dynamic selection layer, and then the predictor to be trained is trained by utilizing the general characteristics and the business characteristics, so that the efficiency and the accuracy of model training can be ensured.

FIG. 3 is a schematic flow chart diagram of another multi-objective prediction model training method according to an embodiment of the present disclosure. Referring to fig. 3, the model training method is as follows:

s301, acquiring a training sample under a specified service scene; wherein a plurality of business targets exist in the designated business arena.

In this embodiment, the specified service scenario is, for example, a recommendation scenario for recommending news information, and may be any other recommendation scenario, which is not specifically limited herein. The service target is exemplarily at least one of a resource click rate, a resource play duration, and an interactive behavior.

S302, aiming at any training sample, extracting general characteristics and business characteristics related to business targets from the training sample.

The general characteristics comprise at least one of user characteristics, article characteristics, user behavior sequence characteristics and user request characteristics; the service characteristic is at least one of a resource click rate characteristic, a resource playing time length characteristic and an interaction characteristic.

In the example, the underlying network of the multi-target pre-estimation model comprises a general tower network for receiving general characteristics and a service tower network for receiving service characteristics; the upper network of the multi-target estimation model comprises estimation sub-networks which are respectively corresponding to each service target and used for estimating resources. Besides, the underlying network of the multi-target estimation model also comprises a cross-layer network. On the basis, the process of training the multi-target estimation model according to the general characteristics and the business characteristics related to the business target is shown in S303-S307.

And S303, inputting the general features into a general tower network.

S304, aiming at the service characteristics related to any service target, inputting the service characteristics into the service tower network, and determining a first pre-estimating sub-network corresponding to the service target.

The processes of S303 to S304 refer to the above embodiments, which are not described herein again.

The process of propagating the output of the general tower network and the output of the service tower network forward to the first predictor network can be seen in S305-S306.

S305, carrying out cross processing on the output result of the general tower network and the output result of the service tower network through a cross layer network to obtain cross characteristics.

Optionally, the output result of the general tower network and the output result of the service tower network are processed in a cross way through a cross layer network in an element addition way to obtain cross characteristics; or the like, or, alternatively,

performing cross processing on the output result of the general tower network and the output result of the service tower network in an element multiplication mode through a cross layer network to obtain cross characteristics; or the like, or, alternatively,

and performing nonlinear transformation on the output result of the service tower network through a cross layer network (for example, performing nonlinear processing on the output result of the service tower by using a preset excitation function), and performing cross processing on the nonlinear transformation result and the output result of the universal tower network in an element multiplication manner to obtain cross characteristics.

It should be noted that the output result of the general tower network and the output result of the service tower network are both expressed in the form of Embedding vectors, and the addition of elements and the multiplication of elements are both performed on the Embedding vectors. Model training is carried out according to the features after crossing, and the estimation effect of the multi-target estimation model can be improved.

S306, forward propagating the cross features to the first predictor network.

S307, reversely updating the network parameters of the first prediction sub-network and the network parameters of the underlying network according to the prediction error output by the first prediction sub-network.

In the embodiment of the disclosure, the service features and the general features are crossed, so that feature fusion is realized, and then the corresponding predictor network is trained based on the fused features, so that the training efficiency and the training effect can be ensured.

Fig. 4 is a schematic flowchart of a resource estimation method according to an embodiment of the present disclosure, which is applicable to a case of performing estimation sorting on recalled resources in a recommendation system. The method can be executed by a resource pre-training device which is realized by adopting a software and/or hardware mode and is integrated on the electronic equipment.

Specifically, referring to fig. 4, the resource estimation method is as follows:

s401, obtaining target resources to be estimated.

The target resource to be estimated can be a resource recalled by a downstream queue in response to a user refresh request, and can be a video resource, a text resource and the like.

S402, extracting general characteristics and business characteristics related to the business target from the target resource.

Optionally, the constructed feature extraction neural network is used to extract the generic features and the business features related to the business target from the target resource. Wherein the general characteristics comprise at least one of user characteristics (such as user identification, age, gender and the like), article characteristics (such as article titles, categories and the like), user behavior sequence characteristics (such as usage behavior records of the user, which can be videos of a certain category continuously watched by the user for a period of time), and user request characteristics (such as request addresses); the service characteristic is at least one of a resource click rate characteristic, a resource play duration characteristic and an interaction characteristic (for example, approval or forwarding of the resource and the like).

S403, performing prediction and estimation on the general characteristics and the service characteristics through a multi-target prediction model to obtain a prediction result of target resources; the multi-target estimation model is obtained after training according to the model training method in the embodiment.

In this embodiment, the underlying network of the multi-target prediction model includes a general tower network for receiving general characteristics and a service tower network for receiving service characteristics; the upper network of the multi-target estimation model comprises estimation sub-networks which are respectively corresponding to each service target and used for estimating resources. Through a multi-target prediction model, the general characteristics and the service characteristics are predicted and predicted, and the method comprises the following processes: inputting the general characteristics into a general tower network, and inputting the service characteristics into a service tower network; and performing cross processing on the output result of the general tower network and the output result of the service tower, and inputting the cross result into a pre-estimation sub-network corresponding to the service target for pre-estimation. So as to sort and recommend the resources according to the pre-estimated result.

In the embodiment of the disclosure, the target resources are accurately estimated by using the multi-target estimation model trained in the embodiment, so that resource sequencing and recommendation are performed according to estimation results, and thus, the recommendation effect can be ensured.

Fig. 5 is a schematic structural diagram of a multi-target prediction model training device according to an embodiment of the present disclosure, which is applicable to a situation where training data corresponding to different business targets are used to train a multi-target prediction model in a recommendation scenario. Referring to fig. 5, the apparatus includes:

a data obtaining module 501, configured to obtain a training sample in a specified service scene; wherein a plurality of business targets exist in the execution business field;

a feature extraction module 502, configured to extract, for any training sample, a generic feature and a business feature related to a business target from the training sample;

and the training module 503 is configured to train the multi-target estimation model according to the general features and the business features related to the business targets.

On the basis of the above embodiment, optionally, the underlying network of the multi-target prediction model includes a general tower network for receiving general characteristics and a service tower network for receiving service characteristics; the upper network of the multi-target estimation model comprises estimation sub-networks which are respectively corresponding to each service target and used for estimating resources.

On the basis of the foregoing embodiment, optionally, the training module includes:

a first input unit for inputting the generic features into a generic tower network;

the second input unit is used for inputting the service characteristics into the service tower network aiming at the service characteristics related to any service target and determining a first pre-estimated sub-network corresponding to the service target;

the forward propagation unit is used for forward propagating the output result of the general tower network and the output result of the business tower network to the first predictor network;

and the parameter updating unit is used for reversely updating the network parameters of the first estimating sub-network and the network parameters of the underlying network according to the estimating error output by the first estimating sub-network.

On the basis of the above embodiment, optionally, the underlying network of the multi-target estimation model further includes a cross-layer network;

the forward propagation unit includes:

the crossing subunit is used for carrying out crossing processing on the output result of the general tower network and the output result of the service tower network through a crossing layer network to obtain crossing characteristics;

a transmission subunit for propagating the cross signature forward into the first predictor network.

On the basis of the above embodiment, optionally, the crossing subunit is further configured to:

performing cross processing on the output result of the general tower network and the output result of the service tower network in an element addition mode through a cross layer network; or the like, or, alternatively,

performing cross processing on the output result of the general tower network and the output result of the service tower network in an element multiplication mode through a cross layer network; or the like, or a combination thereof,

and carrying out nonlinear transformation on the output result of the service tower network through a cross layer network, and carrying out cross processing on the nonlinear transformation result and the output result of the universal tower network in an element multiplication mode.

On the basis of the above embodiment, optionally, the method further includes:

and the weight adding module is used for setting corresponding training weights for the pre-estimation errors according to the scale difference of the training samples corresponding to different business targets.

On the basis of the above embodiment, optionally, the general features include at least one of a user feature, an article feature, a user behavior sequence feature, and a user request feature; the service characteristics are at least one of resource click rate characteristics, resource playing time characteristics and interaction characteristics.

The multi-target estimation model training device provided by the embodiment of the disclosure can execute the multi-target estimation model training method provided by any embodiment of the disclosure, and has the corresponding functional modules and beneficial effects of the execution method. Reference may be made to the description of any method embodiment of the disclosure for a matter not explicitly described in this embodiment.

Fig. 6 is a schematic structural diagram of a resource estimation apparatus according to an embodiment of the present disclosure, which is applicable to a case of estimating and sorting recalled resources in a recommendation system. Referring to fig. 6, the apparatus includes:

a resource obtaining module 601, configured to obtain a target resource to be pre-estimated;

a resource feature extraction module 602, configured to extract, from a target resource, a generic feature and a service feature related to a service target;

the pre-estimation module 602 is configured to perform prediction and pre-estimation on the general characteristics and the service characteristics through a multi-target pre-estimation model to obtain a pre-estimation result of a target resource; the multi-target pre-estimation model is obtained after training according to any one of the multi-target pre-estimation surface model training methods in the embodiment of the application.

On the basis of the above embodiment, optionally, the underlying network of the multi-target prediction model includes a general tower network for receiving general characteristics and a service tower network for receiving service characteristics; the upper layer of the multi-target estimation model comprises estimation sub-networks which are respectively corresponding to each service target and used for estimating resources;

the estimation module is further used for:

inputting the general characteristics into a general tower network, and inputting the service characteristics into a service tower network;

and performing cross processing on the output result of the universal tower network and the output result of the service tower, and inputting the cross result into a pre-estimating sub-network corresponding to the service target for pre-estimation.

The resource estimation device provided by the embodiment of the disclosure can execute the resource estimation method provided by any embodiment of the disclosure, and has corresponding functional modules and beneficial effects of the execution method. Reference may be made to the description of any method embodiment of the disclosure for a matter not explicitly described in this embodiment.

In the technical scheme of the disclosure, the acquisition, storage, application and the like of the personal information of the related user all accord with the regulations of related laws and regulations, and do not violate the customs of public sequences.

The present disclosure also provides an electronic device, a readable storage medium, and a computer program product according to embodiments of the present disclosure.

FIG. 7 illustrates a schematic block diagram of an example electronic device 700 that can be used to implement embodiments of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the disclosure described and/or claimed herein.

As shown in fig. 7, the device 700 comprises a computing unit 701, which may perform various suitable actions and processes according to a computer program stored in a Read Only Memory (ROM) 702 or a computer program loaded from a storage unit 708 into a Random Access Memory (RAM) 703. In the RAM703, various programs and data required for the operation of the device 700 can also be stored. The computing unit 701, the ROM702, and the RAM703 are connected to each other by a bus 704. An input/output (I/O) interface 705 is also connected to bus 704.

Various components in the device 700 are connected to the I/O interface 705, including: an input unit 706 such as a keyboard, a mouse, or the like; an output unit 707 such as various types of displays, speakers, and the like; a storage unit 708 such as a magnetic disk, optical disk, or the like; and a communication unit 709 such as a network card, modem, wireless communication transceiver, etc. The communication unit 709 allows the device 700 to exchange information/data with other devices via a computer network, such as the internet, and/or various telecommunication networks.

Computing unit 701 may be a variety of general purpose and/or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, and so forth. The calculation unit 701 performs the respective methods and processes described above, such as the model training method or the resource estimation method. For example, in some embodiments, the model training method or the resource estimation method may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as storage unit 708. In some embodiments, part or all of a computer program may be loaded onto and/or installed onto device 700 via ROM702 and/or communications unit 709. When the computer program is loaded into RAM703 and executed by the computing unit 701, one or more steps of the model training method or resource estimation method described above may be performed. Alternatively, in other embodiments, the computing unit 701 may be configured to perform the model training method or the resource prediction method by any other suitable means (e.g., by means of firmware).

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuitry, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), system on a chip (SOCs), complex Programmable Logic Devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program code may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program code, when executed by the processor or controller, causes the functions/acts specified in the flowchart and/or block diagram to be performed. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server with a combined blockchain.

It should be understood that various forms of the flows shown above, reordering, adding or deleting steps, may be used. For example, the steps described in the present disclosure may be executed in parallel or sequentially or in different orders, and are not limited herein as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved.

The above detailed description should not be construed as limiting the scope of the disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made, depending on design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A multi-target pre-estimation model training method comprises the following steps:

aiming at any training sample, extracting general features and business features related to business targets from the training sample;

and training the multi-target estimation model according to the general characteristics and the business characteristics related to the business target.

2. The method of claim 1, wherein the underlying network of multi-objective predictive models includes a general purpose tower network for receiving general features and a business tower network for receiving business features; the upper layer network of the multi-target estimation model comprises estimation sub-networks which are respectively corresponding to each business target and used for estimating resources.

3. The method of claim 2, wherein training the multi-objective predictive model based on the generic features and business features associated with business objectives comprises:

inputting the generic features into the generic tower network;

inputting the service characteristics into the service tower network aiming at the service characteristics related to any service target, and determining a first pre-estimating sub-network corresponding to the service target;

the output result of the general tower network and the output result of the business tower network are transmitted to the first predictor network in a forward direction;

and reversely updating the network parameters of the first prediction sub-network and the network parameters of the underlying network according to the prediction error output by the first prediction sub-network.

4. The method of claim 3, wherein the underlying network of the multi-objective prediction model further comprises a cross-layer network;

forward propagating the output result of the general tower network and the output result of the service tower network to the first predictor network, including:

performing cross processing on the output result of the general tower network and the output result of the service tower network through the cross layer network to obtain cross characteristics;

propagating the cross feature forward into the first predictor network.

5. The method of claim 4, wherein the cross-processing the output of the generic tower network and the output of the traffic tower network by the cross-layer network comprises:

performing cross processing on the output result of the general tower network and the output result of the service tower network in an element addition mode through the cross layer network; or the like, or a combination thereof,

performing cross processing on the output result of the general tower network and the output result of the service tower network in an element multiplication mode through the cross layer network; or the like, or, alternatively,

and carrying out nonlinear transformation on the output result of the service tower network through the cross layer network, and carrying out cross processing on the nonlinear transformation result and the output result of the general tower network in an element multiplication mode.

6. The method of claim 3, further comprising:

and setting corresponding training weights for the pre-estimated errors according to the scale difference of training samples corresponding to different business targets.

7. The method of claim 1, wherein the generic features comprise at least one of user features, article features, user behavior sequence features, user request features; the service characteristic is at least one of a resource click rate characteristic, a resource playing time length characteristic and an interaction characteristic.

8. A resource pre-estimating method comprises the following steps:

acquiring target resources to be estimated;

extracting general characteristics and business characteristics related to a business target from the target resource;

predicting and predicting the general characteristics and the service characteristics through a multi-target prediction model to obtain a prediction result of the target resource; wherein the multi-target prediction model is obtained after training according to the method of any one of claims 1 to 7.

9. The method of claim 1, wherein the underlying network of multi-objective predictive models includes a general purpose tower network for receiving general features and a business tower network for receiving business features; the upper layer network of the multi-target estimation model comprises estimation sub-networks which are respectively corresponding to each business target and used for estimating resources;

through a multi-target pre-estimation model, the general characteristics and the service characteristics are predicted and pre-estimated, and the method comprises the following steps:

inputting the generic features into the generic tower network and inputting the business features into a business tower network;

and performing cross processing on the output result of the universal tower network and the output result of the service tower, and inputting the cross result into a pre-estimation sub-network corresponding to a service target for pre-estimation.

10. A multi-target pre-estimation model training device comprises:

the data acquisition module is used for acquiring a training sample under a specified service scene; wherein a plurality of business targets exist in the designated business farm;

the feature extraction module is used for extracting general features and business features related to business targets from any training sample;

11. The apparatus of claim 10, wherein the underlying network of multi-objective predictive models comprises a general purpose tower network for receiving general features and a business tower network for receiving business features; the upper layer network of the multi-target estimation model comprises estimation sub-networks which are respectively corresponding to each business target and used for estimating resources.

12. The apparatus of claim 11, wherein the training module comprises:

a first input unit for inputting the generic features into the generic tower network;

the second input unit is used for inputting the service characteristics into the service tower network aiming at the service characteristics related to any service target and determining a first pre-estimating sub-network corresponding to the service target;

and the parameter updating unit is used for reversely updating the network parameters of the first estimating sub-network and the network parameters of the underlying network according to the estimation error output by the first estimating sub-network.

13. The apparatus of claim 12, wherein the underlying network of the multi-objective prediction model further comprises a cross-layer network;

the forward propagation unit includes:

the cross subunit is used for performing cross processing on the output result of the general tower network and the output result of the service tower network through the cross layer network to obtain cross characteristics;

a transmission subunit, configured to forward propagate the cross signature into the first predictor network.

14. The apparatus of claim 13, wherein the cross subunit is further configured to:

performing cross processing on the output result of the general tower network and the output result of the service tower network in an element multiplication mode through the cross layer network; or the like, or a combination thereof,

and carrying out nonlinear transformation on the output result of the service tower network through the cross layer network, and carrying out cross processing on the nonlinear transformation result and the output result of the universal tower network in an element multiplication mode.

15. The apparatus of claim 12, further comprising:

and the weight adding module is used for setting corresponding training weights for the estimation errors according to the scale difference of training samples corresponding to different business targets.

16. The apparatus of claim 10, wherein the generic features comprise at least one of a user feature, an article feature, a user behavior sequence feature, a user request feature; the service characteristics are at least one of resource click rate characteristics, resource playing time characteristics and interaction characteristics.

17. A resource projection apparatus, comprising:

the resource feature extraction module is used for extracting general features and business features related to business targets from the target resources;

the prediction module is used for predicting and predicting the general characteristics and the service characteristics through a multi-target prediction model to obtain a prediction result of the target resource; wherein the multi-target prediction model is obtained after training according to the method of any one of claims 1 to 7.

18. An electronic device, comprising:

at least one processor; and

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the model training method of any one of claims 1-7 or the resource estimation method of any one of claims 8-9.

19. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the model training method of any one of claims 1-7 or the resource estimation method of any one of claims 8-9.

20. A computer program product comprising a computer program which, when executed by a processor, implements the model training method of any one of claims 1-7 or the resource prediction method of any one of claims 8-9.