CN112949864A

CN112949864A - Training method and device for pre-estimation model

Info

Publication number: CN112949864A
Application number: CN202110137591.7A
Authority: CN
Inventors: 李爽; 谢乾龙; 王兴星; 王栋
Original assignee: 北京三快在线科技有限公司
Current assignee: Hainan Liangxin Technology Co ltd
Priority date: 2021-02-01
Filing date: 2021-02-01
Publication date: 2021-06-11
Anticipated expiration: 2041-02-01
Also published as: CN112949864B

Abstract

The disclosure provides a training method and device for a pre-estimation model. The method comprises the following steps: acquiring a first training set and a second training set according to historical data of a first service party, and acquiring a third training set according to historical data of a second service party; the first business party is a business party with the exposure amount larger than a first threshold value, the second business party is a business party with the exposure amount smaller than a second threshold value, and the first threshold value is larger than the second threshold value; training the initial estimation model with the target sub-model shielded on the basis of the first training set to obtain a first estimation model; training a target sub-model in the first pre-estimation model based on the second training set to obtain a second pre-estimation model; and training the second estimation model based on the third training set to obtain a target estimation model. The method and the device can improve the long tail flow estimation effect, and further improve the estimation accuracy of the click rate and the conversion rate of a merchant.

Description

Training method and device for pre-estimation model

Technical Field

The embodiment of the disclosure relates to the technical field of model training, in particular to a training method and device for a pre-estimation model.

Background

The exposure distribution in the advertising and recommendation scenarios follows the two eight law in economics, with a greater number of small merchants in the marketing advertising scenario, with about 30% of the head merchants contributing 80% of the exposure and 70% of the long-tailed merchants contributing 20% of the exposure.

The existing estimation of the click rate and the conversion rate is mainly based on a deep neural network model with an Embedding and MLP structure, the deep model has a better optimization effect on head flow, and the exposure of the long-tail flow subdivided to the granularity of a merchant is too little, so that the deep model is not beneficial to the Embedding learning of the merchant, the optimization effect is poor, and the estimation accuracy of the click rate and the conversion rate of the merchant is low.

Disclosure of Invention

The embodiment of the disclosure provides a training method and a training device for an estimation model, which are used for improving estimation accuracy of click rate and conversion rate of merchants.

According to a first aspect of embodiments of the present disclosure, there is provided a training method of a predictive model, including:

acquiring a first training set and a second training set according to historical data of a first service party, and acquiring a third training set according to historical data of a second service party; the first business party is a business party with the exposure amount larger than a first threshold value, the second business party is a business party with the exposure amount smaller than a second threshold value, and the first threshold value is larger than the second threshold value;

training the initial estimation model with the target sub-model shielded on the basis of the first training set to obtain a first estimation model;

training a target sub-model in the first pre-estimation model based on the second training set to obtain a second pre-estimation model;

and training the second estimation model based on the third training set to obtain a target estimation model.

Optionally, the obtaining a first training set and a second training set according to the historical data of the first service party includes:

obtaining a model training set according to the historical data of the first service party;

obtaining a first training sample with a first proportion in the model training set, and obtaining the first training set according to the first training sample;

obtaining a second training sample with a second proportion in the model training set, and obtaining the second training set according to the second training sample;

wherein the first proportion is greater than the second proportion, and the sum of the first proportion and the second proportion is 1.

Optionally, each first training sample in the first training set corresponds to a first initial click rate and a first initial conversion rate, and the training of the initial estimation model with the target sub-model shielded based on the first training set to obtain the first estimation model includes:

adjusting the model parameters of the initial pre-estimated model into first model parameters to shield the target sub-model;

inputting the first training sample into the initial prediction model, and obtaining a first prediction click rate and a first prediction conversion rate corresponding to the first training sample output by the initial prediction model;

calculating to obtain a first loss value corresponding to the initial prediction model according to the first initial click rate, the first predicted click rate, the first initial conversion rate and the first predicted conversion rate;

and under the condition that the first loss value is within a first preset range, taking the trained initial estimation model as the first estimation model.

Optionally, the training the target sub-model in the first pre-estimation model based on the second training set to obtain a second pre-estimation model includes:

adjusting the model parameters of the first pre-estimated model into second model parameters to shield other submodels except the target submodel in the first pre-estimated model;

inputting second training samples in the second training set into the first pre-estimation model, and acquiring a first click vector corresponding to a predicted click rate and a first conversion vector corresponding to a predicted conversion rate which are output by the target sub-model;

inputting a third training sample in the third training set into the first pre-estimation model, and acquiring a second click vector corresponding to a predicted click rate and a second conversion vector corresponding to a predicted conversion rate which are output by the target sub-model;

calculating to obtain a second loss value corresponding to the first pre-estimation model according to the first click vector, the first conversion vector, the second click vector and the second conversion vector;

and under the condition that the second loss value is within a second preset range, taking the trained first estimation model as the second estimation model.

Optionally, each third training sample in the third training set corresponds to a second initial click rate and a second initial conversion rate, and the training the second estimation model based on the third training set to obtain the target estimation model includes:

adjusting the model parameters of the second pre-estimated model to third model parameters;

inputting the third training sample into the second pre-estimation model, and obtaining a second predicted click rate and a second predicted conversion rate corresponding to the third training sample output by the second pre-estimation model;

calculating to obtain a third loss value corresponding to the second pre-estimation model according to the second initial click rate, the second predicted click rate, the second initial conversion rate and the second predicted conversion rate;

and under the condition that the third loss value is within a third preset range, taking the trained second estimation model as a target estimation model.

Optionally, after the training of the second prediction model based on the third training set is performed to obtain a target prediction model, the method further includes:

acquiring the average daily target exposure of a target service party to be estimated in a period which is a set time length away from the current time;

acquiring service party characteristic information of the target service party;

under the condition that the target exposure is larger than a set threshold value, adjusting the model parameters of the target pre-estimation model into target model parameters to shield the target sub-model, and inputting the characteristic information of the business party into the target pre-estimation model to obtain the click rate and the conversion rate corresponding to the target business party output by the target pre-estimation model;

and under the condition that the target exposure is less than or equal to the set threshold, inputting the characteristic information of the business party into the target pre-estimation model to obtain the click rate and the conversion rate corresponding to the target business party output by the target pre-estimation model.

According to a second aspect of the embodiments of the present disclosure, there is provided a training apparatus for a predictive model, including:

the training set acquisition module is used for acquiring a first training set and a second training set according to the historical data of the first service party and acquiring a third training set according to the historical data of the second service party; the first business party is a business party with the exposure amount larger than a first threshold value, the second business party is a business party with the exposure amount smaller than a second threshold value, and the first threshold value is larger than the second threshold value;

the first pre-estimation model acquisition module is used for training the initial pre-estimation model which shields the target sub-model based on the first training set to obtain a first pre-estimation model;

the second estimation model obtaining module is used for training the target sub-model in the first estimation model based on the second training set to obtain a second estimation model;

and the target estimation model acquisition module is used for training the second estimation model based on the third training set to obtain a target estimation model.

Optionally, the training set obtaining module includes:

a training set obtaining unit, configured to obtain a model training set according to historical data of the first service party;

the first training set acquisition unit is used for acquiring a first training sample with a first proportion in the model training set and acquiring the first training set according to the first training sample;

the second training set acquisition unit is used for acquiring a second training sample with a second proportion in the model training set and acquiring the second training set according to the second training sample;

Optionally, each first training sample in the first training set corresponds to a first initial click rate and a first initial conversion rate, and the first pre-estimation model obtaining module includes:

the first model parameter adjusting unit is used for adjusting the model parameters of the initial pre-estimated model into first model parameters so as to shield the target sub-model;

the prediction conversion rate obtaining unit is used for inputting the first training sample into the initial prediction model and obtaining a first prediction click rate and a first prediction conversion rate corresponding to the first training sample output by the initial prediction model;

a first loss value calculating unit, configured to calculate a first loss value corresponding to the initial prediction model according to the first initial click rate, the first predicted click rate, the first initial conversion rate, and the first predicted conversion rate;

and the first estimation model obtaining unit is used for taking the trained initial estimation model as the first estimation model under the condition that the first loss value is within a first preset range.

Optionally, the second prediction model obtaining module includes:

the second model parameter adjusting unit is used for adjusting the model parameters of the first pre-estimated model into second model parameters so as to shield other sub-models except the target sub-model in the first pre-estimated model;

a first conversion vector obtaining unit, configured to input a second training sample in the second training set to the first pre-estimation model, and obtain a first click vector corresponding to a predicted click rate and a first conversion vector corresponding to a predicted conversion rate, which are output by the target sub-model;

a second conversion vector obtaining unit, configured to input a third training sample in the third training set to the first pre-estimation model, and obtain a second click vector corresponding to the predicted click rate and a second conversion vector corresponding to the predicted conversion rate, which are output by the target sub-model;

a second loss value calculation unit, configured to calculate a second loss value corresponding to the first pre-estimation model according to the first click vector, the first conversion vector, the second click vector, and the second conversion vector;

and the second estimation model obtaining unit is used for taking the trained first estimation model as the second estimation model under the condition that the second loss value is within a second preset range.

Optionally, each third training sample in the third training set corresponds to a second initial click rate and a second initial conversion rate, and the target prediction model obtaining module includes:

a third model parameter adjusting unit, configured to adjust the model parameter of the second prediction model to a third model parameter;

the predicted click rate obtaining unit is used for inputting the third training sample into the second prediction model and obtaining a second predicted click rate and a second predicted conversion rate corresponding to the third training sample output by the second prediction model;

a third loss value calculation unit, configured to calculate a third loss value corresponding to the second prediction model according to the second initial click rate, the second predicted click rate, the second initial conversion rate, and the second predicted conversion rate;

and the target estimation model obtaining unit is used for taking the trained second estimation model as the target estimation model under the condition that the third loss value is within a third preset range.

Optionally, the method further includes:

the target exposure acquiring module is used for acquiring the average daily target exposure of a target service party to be estimated in a period which is a set time length away from the current time;

a service party characteristic obtaining module, configured to obtain service party characteristic information of the target service party;

the first click rate acquisition module is used for adjusting the model parameters of the target pre-estimation model into target model parameters to shield the target sub-model under the condition that the target exposure is larger than a set threshold value, and inputting the characteristic information of the business party into the target pre-estimation model to acquire the click rate and the conversion rate corresponding to the target business party output by the target pre-estimation model;

and the second click rate acquisition module is used for inputting the characteristic information of the business party to the target pre-estimation model under the condition that the target exposure is less than or equal to the set threshold value so as to acquire the click rate and the conversion rate corresponding to the target business party output by the target pre-estimation model.

According to a third aspect of embodiments of the present disclosure, there is provided an electronic apparatus including:

the system comprises a processor, a memory and a computer program stored on the memory and capable of running on the processor, wherein the processor realizes the training method of the estimation model in any one of the above items when executing the program.

According to a fourth aspect of embodiments of the present disclosure, there is provided a readable storage medium, wherein instructions, when executed by a processor of an electronic device, enable the electronic device to perform any one of the above training methods for a predictive model.

The embodiment of the disclosure provides a training method and a training device for a pre-estimation model, wherein a first training set and a second training set are obtained according to historical data of a first service party, a third training set is obtained according to historical data of a second service party, the first service party is a service party with exposure larger than a first threshold, the second service party is a service party with exposure smaller than a second threshold, the first threshold is larger than the second threshold, an initial pre-estimation model shielding a target sub-model is trained on the basis of the first training set to obtain the first pre-estimation model, the target sub-model in the first pre-estimation model is trained on the basis of the second training set to obtain the second pre-estimation model, and the second pre-estimation model is trained on the basis of the third training set to obtain the target pre-estimation model. The embodiment of the disclosure improves the generalization capability of the long tail flow by adding the target sub-model in the pre-estimation model, optimizes the training process, uses the head merchant to migrate and learn the target sub-model, can ensure the quality of generating embedding, improves the pre-estimation effect of the long tail flow on the premise of not influencing the optimization effect of the head flow, and further improves the pre-estimation accuracy of the click rate and the conversion rate of the merchant.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the description of the embodiments of the present disclosure will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present disclosure, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without inventive exercise.

Fig. 1 is a flowchart illustrating steps of a training method for a predictive model according to an embodiment of the present disclosure;

FIG. 2 is a flowchart illustrating steps of another predictive model training method according to an embodiment of the present disclosure;

fig. 3 is a schematic structural diagram of a training apparatus for pre-estimation models according to an embodiment of the present disclosure;

fig. 4 is a schematic structural diagram of another training apparatus for predictive models according to an embodiment of the present disclosure.

Detailed Description

Technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the drawings in the embodiments of the present disclosure, and it is apparent that the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments, which can be obtained by a person skilled in the art without making creative efforts based on the embodiments of the present disclosure, belong to the protection scope of the embodiments of the present disclosure.

Example one

Referring to fig. 1, a flowchart illustrating steps of a training method of a predictive model provided by an embodiment of the present disclosure is shown, and as shown in fig. 1, the training method of the predictive model may specifically include the following steps:

step 101: and acquiring a first training set and a second training set according to the historical data of the first service party, and acquiring a third training set according to the historical data of the second service party.

The embodiment of the disclosure can be applied to a scene of training models of click rate and conversion rate of pre-estimated merchants.

In this embodiment, the first threshold and the second threshold may be thresholds preset by business staff and used for determining a head merchant and a long-tailed merchant, where the first threshold is greater than the second threshold, that is, when the exposure of the business party is greater than the first threshold, the business party is regarded as the head merchant, and when the exposure of the business party is less than the second threshold, the business party is regarded as the long-tailed merchant.

In this embodiment, the structure of the pre-estimation model to be trained may include: the main framework sub-model and the Embedding Generator sub-model, wherein the main framework sub-model can be used for predicting the click rate and the conversion rate of head merchants, and the Embedding Generator sub-model can be used for predicting the click rate and the conversion rate of long-tail merchants.

When the estimation model needs to be trained, a first training set and a second training set can be obtained according to historical data of a first business party, and a third training set can be obtained according to historical data of a second business party, wherein training samples in the first training set can be used for training a main framework sub-model in the estimation model, training samples in the second training set can be independently used for training an Embedding generating sub-model in the estimation model, and training samples in the third training set can be used for jointly training the main framework sub-model and the Embedding generating sub-model.

The training samples in the three training sets may include characteristics of the business party, such as business party ID, business party business district information, commodity category information, historical click rate, historical conversion rate, exposure, and other characteristic information.

The process of acquiring the first training set and the second training set may be described in detail in conjunction with the following specific implementation.

In a specific implementation manner of the embodiment of the present disclosure, the step 101 may include:

substep A1: and obtaining a model training set according to the historical data of the first service party.

In this embodiment, a model training set may be obtained according to historical data of the first service party.

Substep A2: obtaining a first training sample with a first proportion in the model training set, and obtaining the first training set according to the first training sample;

substep A3: and obtaining a second training sample of a second proportion in the model training set, and obtaining the second training set according to the second training sample.

After the model training set is obtained, a first training sample of a first proportion may be obtained from the model training set, and a first training set may be formed according to the first training sample, and a second training sample of a second proportion may be obtained from the model training set, and a second training set may be formed according to the second training sample, where the first proportion is greater than the second proportion, and a sum of the first proportion and the second proportion is 1, for example, after the model training set is obtained, 80% of the training samples may be screened from the model training set to form the first training set, and a second training set may be formed according to the remaining 20% of the training samples, and so on.

It should be understood that the above examples are only examples for better understanding of the technical solutions of the embodiments of the present disclosure, and are not to be taken as the only limitation to the embodiments.

After the first, second, and third training sets are obtained, step 102 is performed.

Step 102: and training the initial estimation model with the target sub-model shielded on the basis of the first training set to obtain a first estimation model.

The initial estimation model refers to a model of estimated merchant click rate and conversion rate which is not trained yet.

The first pre-estimation model is obtained after the initial pre-estimation model is trained for the first time by adopting a first training set.

After the first training set, the second training set, and the third training set are obtained, the initial estimation model may be trained by using the first training set.

When the first training set is adopted to train the initial prediction model, a target sub-model in the initial prediction model needs to be shielded, and the initial prediction model with the target sub-model shielded is trained through the first training set to obtain the first prediction model.

After the initial pre-estimation model with the target sub-model shielded is trained based on the first training set to obtain a first pre-estimation model, step 103 is executed.

Step 103: and training the target sub-model in the first pre-estimation model based on the second training set to obtain a second pre-estimation model.

The second estimation model is an estimation model obtained after a target sub-model in the first estimation model is trained by adopting a second training set.

After the initial estimation model with the target sub-model shielded is trained based on the first training set to obtain the first estimation model, at this time, shielding of the target sub-model can be cancelled, other sub-models except the target sub-model in the first estimation model are shielded, and then the target sub-model in the first estimation model is trained based on the second training set, so that the second estimation model can be obtained.

After the second prediction model is obtained, step 104 is performed.

Step 104: and training the second estimation model based on the third training set to obtain a target estimation model.

The target estimation model is an estimation model obtained after the second estimation model is trained by adopting a third training set.

After the target submodel in the first pre-estimated model is trained based on the second training set to obtain the second pre-estimated model, shielding of all submodels in the second pre-estimated model can be cancelled, and the second pre-estimated model is comprehensively trained through the third training set to obtain the target pre-estimated model.

The embodiment of the disclosure improves the generalization capability of the long tail flow by adding the target sub-model in the pre-estimation model, optimizes the training process, uses the head merchant to migrate and learn the target sub-model, can ensure the quality of generating embedding, improves the pre-estimation effect of the long tail flow on the premise of not influencing the optimization effect of the head flow, and further improves the pre-estimation accuracy of the click rate and the conversion rate of the merchant.

The training method of the pre-estimation model provided by the embodiment of the disclosure includes the steps of obtaining a first training set and a second training set according to historical data of a first service party, obtaining a third training set according to historical data of a second service party, wherein the first service party is a service party with exposure larger than a first threshold, the second service party is a service party with exposure smaller than a second threshold, the first threshold is larger than the second threshold, training an initial pre-estimation model shielding a target sub-model based on the first training set to obtain the first pre-estimation model, training the target sub-model in the first pre-estimation model based on the second training set to obtain the second pre-estimation model, and training the second pre-estimation model based on the third training set to obtain the target pre-estimation model. The embodiment of the disclosure improves the generalization capability of the long tail flow by adding the target sub-model in the pre-estimation model, optimizes the training process, uses the head merchant to migrate and learn the target sub-model, can ensure the quality of generating embedding, improves the pre-estimation effect of the long tail flow on the premise of not influencing the optimization effect of the head flow, and further improves the pre-estimation accuracy of the click rate and the conversion rate of the merchant.

Example two

Referring to fig. 2, a flowchart illustrating steps of another estimation model training method provided in an embodiment of the present disclosure is shown, and as shown in fig. 2, the estimation model training method may specifically include the following steps:

step 201: and acquiring a first training set and a second training set according to the historical data of the first service party, and acquiring a third training set according to the historical data of the second service party.

In a specific implementation manner of the embodiment of the present disclosure, the step 201 may include:

substep B1: and obtaining a model training set according to the historical data of the first service party.

Substep B2: obtaining a first training sample with a first proportion in the model training set, and obtaining the first training set according to the first training sample;

substep B3: and obtaining a second training sample of a second proportion in the model training set, and obtaining the second training set according to the second training sample.

After the first, second, and third training sets are acquired, step 202 is performed.

Step 202: and adjusting the model parameters of the initial pre-estimated model into first model parameters to shield the target sub-model.

The first model parameter refers to a model parameter for shielding a target sub-model in the initial prediction model.

After the first training set, the second training set and the third training set are obtained, firstly, the initial prediction model can be trained by the first training set, and at this time, the model parameters of the initial prediction model can be adjusted to the first model parameters so as to shield the target sub-model in the initial prediction model.

After adjusting the model parameters of the initial pre-estimated model to the first model parameters and shielding the target sub-model in the initial pre-estimated model, step 203 is executed.

Step 203: and inputting the first training sample into the initial prediction model, and acquiring a first predicted click rate and a first predicted conversion rate corresponding to the first training sample output by the initial prediction model.

The first training sample refers to a training sample in a first training set, and after the first training set is obtained, a first initial click rate and a first initial conversion rate corresponding to each first training sample in the first training set may be obtained, that is, each first training sample corresponds to one initial click rate and one initial conversion rate.

After the target submodels in the initial prediction model are shielded, the first training samples in the first training set can be input into the initial prediction model, and the first training samples are processed by the initial prediction model, so that a first predicted click rate and a first predicted conversion rate corresponding to the first training samples can be obtained.

After the first predicted click rate and the first predicted conversion rate are obtained, step 204 is performed.

Step 204: and calculating to obtain a first loss value corresponding to the initial prediction model according to the first initial click rate, the first predicted click rate, the first initial conversion rate and the first predicted conversion rate.

After the first predicted click rate and the first predicted conversion rate are obtained, a first loss value corresponding to the initial prediction model can be obtained through calculation according to the first initial click rate, the first initial conversion rate, the first predicted click rate and the first predicted conversion rate.

After calculating the first loss value corresponding to the initial pre-estimated model, step 205 is executed.

Step 205: and under the condition that the first loss value is within a first preset range, taking the trained initial estimation model as the first estimation model.

The first preset range refers to a range preset by a service person and used for determining whether the trained initial estimation model can meet the service requirement, and the specific value of the first preset range may be determined according to the service requirement, which is not limited in this embodiment.

After the first loss value is calculated, it may be determined whether the first loss value is within a first predetermined range.

If the first loss value is not within the first preset range, it indicates that the result after the initial prediction model is trained by using the first training set cannot meet the business requirement, and at this time, more first training samples can be obtained to continue training the initial prediction model which shields the target sub-model.

If the first loss value is within the first preset range, the trained model can meet the business requirement, and at this time, the trained initial estimation model can be used as the first estimation model.

After the first prediction model is obtained, step 206 is performed.

Step 206: and adjusting the model parameters of the first pre-estimation model into second model parameters so as to shield other sub-models except the target sub-model in the first pre-estimation model.

The second model parameter refers to a model parameter for shielding other submodels except the target submodel in the first estimation model.

After the first pre-estimated model is obtained through training, the model parameters of the first pre-estimated model can be adjusted to be the second model parameters so as to shield other sub-models except the target self-model in the first pre-estimated model.

After the model parameters of the first pre-estimation model are adjusted to the second model parameters, and other submodels except the target submodel in the first pre-estimation model are shielded, step 207 is executed.

Step 207: and inputting a second training sample in the second training set into the first pre-estimation model, and acquiring a first click vector corresponding to the predicted click rate and a first conversion vector corresponding to the predicted conversion rate which are output by the target sub-model.

After shielding other submodels except the target submodel in the first prediction model, a second training set and a third training set can be assembled to carry out combined training on the first prediction model, namely, a second training sample in the second training set is input into the first prediction model, so that the predicted click rate and the predicted conversion rate of the second training sample are determined by the target submodel, vector representation is carried out on the predicted click rate and the predicted conversion rate by the target submodel, a first click vector corresponding to the predicted click rate and a first conversion vector corresponding to the predicted conversion rate are obtained, and the first click vector and the first conversion vector are output.

Step 208: and inputting a third training sample in the third training set into the first pre-estimation model, and acquiring a second click vector corresponding to the predicted click rate and a second conversion vector corresponding to the predicted conversion rate which are output by the target sub-model.

Furthermore, a third training sample in a third training set may be input to the first predictive model, so as to determine a predicted click rate and a predicted conversion rate of the third training sample by the target sub-model, and the predicted click rate and the predicted conversion rate are vector-represented by the target sub-model, so as to obtain a second click vector corresponding to the predicted click rate and a second conversion vector corresponding to the predicted conversion rate, and output the second click vector and the second conversion vector.

Step 209: and calculating to obtain a second loss value corresponding to the first pre-estimation model according to the first click vector, the first conversion vector, the second click vector and the second conversion vector.

After the first click vector, the second click vector, the first conversion vector and the second conversion vector are obtained, a second loss value corresponding to the first pre-estimation model can be calculated according to the first click vector, the second click vector, the first conversion vector and the second conversion vector.

Step 210: and under the condition that the second loss value is within a second preset range, taking the trained first estimation model as the second estimation model.

The second preset range is a range preset by a service person and used for determining whether the trained first estimation model can meet the service requirement, and the specific numerical value of the second preset range may be determined according to the service requirement, which is not limited in this embodiment.

After the second loss value is calculated, it may be determined whether the second loss value is within a second preset range.

If the second loss value is outside the second preset range, it indicates that the trained target sub-model cannot meet the business requirement, and at this time, more second training samples and third training samples can be obtained to train the target sub-model continuously.

If the second loss value is within the second preset range, the trained target sub-model can meet the business requirement, and at this time, the trained first estimation model can be used as the second estimation model.

It can be understood that the preset range described in this step and the preset range described in step 205 may be the same range or different ranges, and specifically, may be determined according to business requirements, which is not limited in this embodiment.

For the joint training process, a second training sample and a third training sample can be obtained each time, the second training sample and the third training sample are simultaneously input into the first pre-estimation model, and then a second loss value is obtained through joint calculation by combining a conversion vector and a click vector which respectively correspond to the two training samples.

After the second prediction model is obtained, step 211 is executed.

Step 211: and adjusting the model parameters of the second pre-estimated model to third model parameters.

The third model parameter refers to a model parameter for canceling shielding of a sub-model in the pre-estimated model.

After the second prediction model is obtained, the model parameters of the second prediction model may be adjusted to the third model parameters to cancel the masking operation on the sub-models in the second prediction model.

After adjusting the model parameters of the second prediction model to the third model parameters, step 212 is performed.

Step 212: and inputting the third training sample into the second pre-estimation model, and acquiring a second predicted click rate and a second predicted conversion rate corresponding to the third training sample output by the second pre-estimation model.

After the second pre-estimation model is obtained, the pre-estimation model can be trained again by combining with a third training set so as to realize the adjustment process of the pre-estimation model.

At this time, each third training sample in the third training set corresponds to an initial click rate and an initial conversion rate, which are respectively recorded as a second initial click rate and a second initial conversion rate.

After the model parameter of the second prediction model is adjusted to the third model parameter, the third training sample can be input into the second prediction model, and a second prediction click rate and a second prediction conversion rate corresponding to the third training sample output by the second prediction model are obtained.

After the second predicted click rate and the second predicted conversion rate are obtained, step 213 is performed.

Step 213: and calculating to obtain a third loss value corresponding to the second estimation model according to the second initial click rate, the second predicted click rate, the second initial conversion rate and the second predicted conversion rate.

After the second predicted click rate and the second predicted conversion rate are obtained, a third loss value corresponding to the second prediction model can be calculated by combining the second initial click rate, the second predicted click rate, the second initial conversion rate and the second predicted conversion rate, and then step 214 is executed.

Step 214: and under the condition that the third loss value is within a third preset range, taking the trained second estimation model as a target estimation model.

The third preset range refers to a range preset by a service person and used for determining whether the trained second estimation model meets the service requirement, and a specific numerical value of the third preset range may be determined according to the service requirement, which is not limited in this embodiment.

After the third loss value is calculated, it may be determined whether the third loss value is within a third preset range.

If the third loss value is not within the third preset range, it indicates that the trained second prediction model cannot meet the business requirements, and at this time, more third training samples can be obtained to train the second prediction model continuously.

If the third loss value is within a third preset range, the trained second prediction model can meet the service requirement, at this time, the trained second prediction model can be used as a final target prediction model, and the target prediction model can be used in a subsequent prediction scene of the merchant click rate and the conversion rate.

Step 215: and acquiring the average daily target exposure of the target service party to be estimated in a period which is a set time length away from the current time.

The daily average target exposure amount refers to the exposure amount of the target business party in a period of a set time length from the current time.

When the click rate and the conversion rate of the target service party need to be estimated, the daily average target exposure of the target service party to be estimated in a period of a set time length from the current time can be obtained, and then step 216 is executed.

Step 216: and acquiring the service party characteristic information of the target service party.

The service party characteristic information refers to characteristic information of a target service party, and may include characteristic information such as a service party ID, service party business district information, commodity category information, historical click rate, historical conversion rate, exposure and the like.

After the daily average target exposure is acquired, the business side characteristic information of the target business side may be acquired, and further, step 217 is performed, or step 218 is performed.

Step 217: and under the condition that the target exposure is larger than a set threshold value, adjusting the model parameters of the target pre-estimation model into target model parameters to shield the target sub-model, and inputting the characteristic information of the business party into the target pre-estimation model to obtain the click rate and the conversion rate corresponding to the target business party output by the target pre-estimation model.

Under the condition that the target exposure is larger than the set threshold, the target business party is a head merchant, at the moment, the model parameters of the target estimation model can be adjusted to be target model parameters to shield the target sub-model, the business party characteristic information is input into the target estimation model, and the business party characteristic information is processed through the target estimation model which shields the target sub-model to predict and obtain the click rate and the conversion rate of the target business party.

Step 218: and under the condition that the target exposure is less than or equal to the set threshold, inputting the characteristic information of the business party into the target pre-estimation model to obtain the click rate and the conversion rate corresponding to the target business party output by the target pre-estimation model.

And under the condition that the target exposure is less than or equal to the set threshold, the target business party is a long-tailed merchant, and at the moment, the business party characteristic information of the target business party can be directly input into the target estimation model, so that the business party characteristic information is processed by the target estimation model to predict and obtain the click rate and the conversion rate of the target business party.

EXAMPLE III

Referring to fig. 3, a schematic structural diagram of a training apparatus for a predictive model according to an embodiment of the present disclosure is shown, and as shown in fig. 3, the training apparatus 300 for a predictive model may specifically include the following modules:

a training set obtaining module 310, configured to obtain a first training set and a second training set according to historical data of a first service party, and obtain a third training set according to historical data of a second service party; the first business party is a business party with the exposure amount larger than a first threshold value, the second business party is a business party with the exposure amount smaller than a second threshold value, and the first threshold value is larger than the second threshold value;

a first pre-estimation model obtaining module 320, configured to train an initial pre-estimation model with a target sub-model shielded based on the first training set, to obtain a first pre-estimation model;

a second pre-estimation model obtaining module 330, configured to train a target sub-model in the first pre-estimation model based on the second training set to obtain a second pre-estimation model;

and a target estimation model obtaining module 340, configured to train the second estimation model based on the third training set to obtain a target estimation model.

Optionally, the training set obtaining module 310 includes:

The training device for the pre-estimation model, provided by the embodiment of the disclosure, acquires a first training set and a second training set according to historical data of a first service party, and acquires a third training set according to historical data of a second service party, wherein the first service party is a service party with an exposure amount larger than a first threshold, the second service party is a service party with an exposure amount smaller than a second threshold, the first threshold is larger than the second threshold, an initial pre-estimation model shielding a target sub-model is trained based on the first training set to obtain the first pre-estimation model, the target sub-model in the first pre-estimation model is trained based on the second training set to obtain the second pre-estimation model, and the second pre-estimation model is trained based on the third training set to obtain the target pre-estimation model. The embodiment of the disclosure improves the generalization capability of the long tail flow by adding the target sub-model in the pre-estimation model, optimizes the training process, uses the head merchant to migrate and learn the target sub-model, can ensure the quality of generating embedding, improves the pre-estimation effect of the long tail flow on the premise of not influencing the optimization effect of the head flow, and further improves the pre-estimation accuracy of the click rate and the conversion rate of the merchant.

Example four

Referring to fig. 4, a schematic structural diagram of another training apparatus for a predictive model provided in an embodiment of the present disclosure is shown, and as shown in fig. 4, the training apparatus 400 for a predictive model may specifically include the following modules:

a training set obtaining module 410, configured to obtain a first training set and a second training set according to historical data of a first service party, and obtain a third training set according to historical data of a second service party; the first business party is a business party with the exposure amount larger than a first threshold value, the second business party is a business party with the exposure amount smaller than a second threshold value, and the first threshold value is larger than the second threshold value;

a first pre-estimation model obtaining module 420, configured to train an initial pre-estimation model with a target sub-model shielded based on the first training set, to obtain a first pre-estimation model;

a second pre-estimation model obtaining module 430, configured to train a target sub-model in the first pre-estimation model based on the second training set to obtain a second pre-estimation model;

a target estimation model obtaining module 440, configured to train the second estimation model based on the third training set to obtain a target estimation model;

a target exposure acquiring module 450, configured to acquire a daily average target exposure of a target service party to be estimated in a period that is a set duration from a current time;

a service party characteristic obtaining module 460, configured to obtain service party characteristic information of the target service party;

a first click rate obtaining module 470, configured to, when the target exposure amount is greater than a set threshold, adjust a model parameter of the target pre-estimation model to a target model parameter to shield the target sub-model, and input the service party feature information to the target pre-estimation model to obtain a click rate and a conversion rate corresponding to the target service party output by the target pre-estimation model;

the second click rate obtaining module 480 is configured to, when the target exposure amount is less than or equal to the set threshold, input the service party feature information into the target pre-estimation model to obtain a click rate and a conversion rate corresponding to the target service party output by the target pre-estimation model.

Optionally, each first training sample in the first training set corresponds to a first initial click rate and a first initial conversion rate, and the first prediction model obtaining module 420 includes:

a first model parameter adjusting unit 421, configured to adjust the model parameters of the initial pre-estimated model to first model parameters to shield the target sub-model;

a prediction conversion rate obtaining unit 422, configured to input the first training sample into the initial prediction model, and obtain a first prediction click rate and a first prediction conversion rate corresponding to the first training sample output by the initial prediction model;

a first loss value calculating unit 423, configured to calculate a first loss value corresponding to the initial prediction model according to the first initial click rate, the first predicted click rate, the first initial conversion rate, and the first predicted conversion rate;

a first estimation model obtaining unit 424, configured to use the trained initial estimation model as the first estimation model when the first loss value is within a preset range.

Optionally, the second prediction model obtaining module 430 includes:

a second model parameter adjusting unit 431, configured to adjust the model parameters of the first pre-estimated model to second model parameters, so as to shield other submodels in the first pre-estimated model except the target submodel;

a first conversion vector obtaining unit 432, configured to input a second training sample in the second training set to the first pre-estimation model, and obtain a first click vector corresponding to a predicted click rate and a first conversion vector corresponding to a predicted conversion rate, which are output by the target sub-model;

a second conversion vector obtaining unit 433, configured to input a third training sample in the third training set to the first pre-estimation model, and obtain a second click vector corresponding to a predicted click rate and a second conversion vector corresponding to a predicted conversion rate that are output by the target sub-model;

a second loss value calculating unit 434, configured to calculate a second loss value corresponding to the first pre-estimation model according to the first click vector, the first conversion vector, the second click vector, and the second conversion vector;

and the second estimation model obtaining unit is used for taking the trained first estimation model as the second estimation model under the condition that the second loss value is within a preset range.

Optionally, each third training sample in the third training set corresponds to a second initial click rate and a second initial conversion rate, and the target prediction model obtaining module 440 includes:

a third model parameter adjusting unit 441, configured to adjust the model parameters of the second pre-estimated model to third model parameters;

a predicted click rate obtaining unit 442, configured to input the third training sample to the second prediction model, and obtain a second predicted click rate and a second predicted conversion rate corresponding to the third training sample output by the second prediction model;

a third loss value calculating unit 443, configured to calculate a third loss value corresponding to the second prediction model according to the second initial click rate, the second predicted click rate, the second initial conversion rate, and the second predicted conversion rate;

the target prediction model obtaining unit 444 is configured to, when the third loss value is within a third preset range, use the trained second prediction model as a target prediction model.

An embodiment of the present disclosure also provides an electronic device, including: a processor, a memory and a computer program stored on the memory and executable on the processor, the processor implementing the training method of the predictive model of the foregoing embodiment when executing the program.

Embodiments of the present disclosure also provide a readable storage medium, and when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the training method of the predictive model of the foregoing embodiments.

For the device embodiment, since it is basically similar to the method embodiment, the description is simple, and for the relevant points, refer to the partial description of the method embodiment.

The algorithms and displays presented herein are not inherently related to any particular computer, virtual machine, or other apparatus. Various general purpose systems may also be used with the teachings herein. The required structure for constructing such a system will be apparent from the description above. In addition, embodiments of the present disclosure are not directed to any particular programming language. It is appreciated that a variety of programming languages may be used to implement the teachings of the embodiments of the present disclosure as described herein, and any descriptions of specific languages are provided above to disclose the best modes of the embodiments of the present disclosure.

In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the disclosure may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the disclosure, various features of the embodiments of the disclosure are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the disclosure and aiding in the understanding of one or more of the various inventive aspects. However, the disclosed method should not be interpreted as reflecting an intention that: that is, claimed embodiments of the disclosure require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of an embodiment of this disclosure.

Those skilled in the art will appreciate that the modules in the device in an embodiment may be adaptively changed and disposed in one or more devices different from the embodiment. The modules or units or components of the embodiments may be combined into one module or unit or component, and furthermore they may be divided into a plurality of sub-modules or sub-units or sub-components. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and all of the processes or elements of any method or apparatus so disclosed, may be combined in any combination, except combinations where at least some of such features and/or processes or elements are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.

The various component embodiments of the disclosure may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It will be understood by those skilled in the art that a microprocessor or Digital Signal Processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components in a motion picture generating device according to an embodiment of the present disclosure. Embodiments of the present disclosure may also be implemented as an apparatus or device program for performing a portion or all of the methods described herein. Such programs implementing embodiments of the present disclosure may be stored on a computer readable medium or may be in the form of one or more signals. Such a signal may be downloaded from an internet website or provided on a carrier signal or in any other form.

It should be noted that the above-mentioned embodiments illustrate rather than limit embodiments of the disclosure, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Embodiments of the disclosure may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The usage of the words first, second and third, etcetera do not indicate any ordering. These words may be interpreted as names.

It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described systems, apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.

The above description is only for the purpose of illustrating the preferred embodiments of the present disclosure and is not to be construed as limiting the embodiments of the present disclosure, and any modifications, equivalents, improvements and the like that are made within the spirit and principle of the embodiments of the present disclosure are intended to be included within the scope of the embodiments of the present disclosure.

The above description is only a specific implementation of the embodiments of the present disclosure, but the scope of the embodiments of the present disclosure is not limited thereto, and any person skilled in the art can easily conceive of changes or substitutions within the technical scope of the embodiments of the present disclosure, and all the changes or substitutions should be covered by the scope of the embodiments of the present disclosure. Therefore, the protection scope of the embodiments of the present disclosure shall be subject to the protection scope of the claims.

Claims

1. A training method of a pre-estimation model is characterized by comprising the following steps:

2. The method of claim 1, wherein obtaining the first training set and the second training set based on historical data of the first business party comprises:

3. The method of claim 1, wherein each first training sample in the first training set corresponds to a first initial click rate and a first initial conversion rate, and the training of the initial predictive model with the target sub-model masked based on the first training set to obtain the first predictive model comprises:

4. The method of claim 3, wherein the training a target sub-model in the first pre-estimation model based on the second training set to obtain a second pre-estimation model comprises:

5. The method of claim 1, wherein each third training sample in the third training set corresponds to a second initial click rate and a second initial conversion rate, and the training the second predictive model based on the third training set to obtain the target predictive model comprises:

6. The method of claim 1, wherein after the training the second predictive model based on the third training set to obtain a target predictive model, further comprising:

acquiring service party characteristic information of the target service party;

7. A training device for a pre-estimation model is characterized by comprising:

8. The apparatus of claim 7, wherein the training set acquisition module comprises:

9. The apparatus of claim 7, wherein each first training sample in the first training set corresponds to a first initial click rate and a first initial conversion rate, and the first pre-prediction model obtaining module comprises:

10. The apparatus of claim 9, wherein the second predictive model obtaining module comprises:

11. The apparatus of claim 7, wherein each of the third training samples in the third training set corresponds to a second initial click rate and a second initial conversion rate, and the target predictive model obtaining module comprises:

12. The apparatus of claim 7, further comprising:

13. An electronic device, comprising:

a processor, a memory, and a computer program stored on the memory and executable on the processor, the processor implementing the training method of the predictive model of any of claims 1 to 6 when executing the program.

14. A readable storage medium, wherein instructions in the storage medium, when executed by a processor of an electronic device, enable the electronic device to perform the training method of the predictive model of any of claims 1 to 6.