CN115185606A

CN115185606A - Method, device, equipment and storage medium for obtaining service configuration parameters

Info

Publication number: CN115185606A
Application number: CN202210852082.7A
Authority: CN
Inventors: 颜林; 李航; 李忠飞; 郑建嘉
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2022-07-19
Filing date: 2022-07-19
Publication date: 2022-10-14

Abstract

The present disclosure provides a method and an apparatus for obtaining a service configuration parameter, an electronic device, a storage medium, and a computer program product, which relate to the technical field of artificial intelligence, and in particular, to the technical field of artificial intelligence, machine learning, intelligent search, and intelligent recommendation. The specific implementation scheme is as follows: acquiring flow request data; inputting the characteristic data of the flow request data into at least one group of target models to obtain the output parameters of each group of target models; the output parameter representation of each group of target models is a service index gain quantity generated by performing service processing on the flow request data from the original service configuration parameters corresponding to each group of target models to the candidate service configuration parameters corresponding to each group of target models; and determining target service configuration parameters from the original service configuration parameters and the candidate service configuration parameters corresponding to each group of target models based on the service index gain of each group of target models. According to the method and the device, the target service configuration parameters can be accurately acquired.

Description

Method, device, equipment and storage medium for obtaining service configuration parameters

Technical Field

The present disclosure relates to the field of artificial intelligence technology, and in particular, to the field of artificial intelligence, machine learning, intelligent search, and intelligent recommendation, and more particularly, to a method and an apparatus for obtaining service configuration parameters, an electronic device, a storage medium, and a computer program product.

Background

In the research on massive data services such as searching, advertisement putting, product recommendation and the like, it is found that the selection of service configuration parameters may influence service benefits to a certain extent and may influence service experience of users. Therefore, how to provide accurate service configuration parameters becomes a technical problem to be solved urgently.

Disclosure of Invention

The disclosure provides a method, a device, a storage medium and a computer program product for obtaining service configuration parameters.

According to an aspect of the present disclosure, a method for obtaining service configuration parameters is provided, including:

acquiring flow request data; inputting the characteristic data of the flow request data into at least one group of target models to obtain the output parameters of each group of target models; the output parameters of each group of target models represent service index gain generated by changing the service processing of the flow request data from the original service configuration parameters corresponding to each group of target models into the candidate service configuration parameters corresponding to each group of target models; and determining a target service configuration parameter from the original service configuration parameter and the candidate service configuration parameter corresponding to each group of target models based on the service index gain of each group of target models, wherein the target service configuration parameter is used for performing service processing on the object generating the flow request data.

According to another aspect of the present disclosure, an apparatus for obtaining service configuration parameters is provided, including:

a first obtaining unit, configured to obtain traffic request data;

the second acquisition unit is used for inputting the characteristic data of the flow request data into at least one group of target models to obtain the output parameters of each group of target models; the output parameters of each group of target models represent service index gain generated by changing the service processing of the flow request data from the original service configuration parameters corresponding to each group of target models into the candidate service configuration parameters corresponding to each group of target models;

and the determining unit is used for determining a target service configuration parameter from the original service configuration parameter and the candidate service configuration parameter corresponding to each group of target models based on the service index gain of each group of target models, wherein the target service configuration parameter is used for performing service processing on the object generating the flow request data.

According to still 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 memory stores instructions executable by the at least one processor to enable the at least one processor to perform a method in any embodiment of the disclosure.

According to yet another aspect of the present disclosure, there is provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing a computer to perform a method in any of the embodiments of the present disclosure.

According to yet another aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method in any of the embodiments of the present disclosure.

According to the method and the device, the target service configuration parameters are obtained based on the service index gain quantity output by the target model, and the accurate obtaining of the target service configuration parameters can be realized by considering the influence of the service indexes on the determination or selection of the service configuration parameters.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they 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 diagram of a system architecture of a mass data services system of an embodiment of the present disclosure;

fig. 2 is a first flowchart illustrating a method for obtaining service configuration parameters according to an embodiment of the present disclosure;

fig. 3 is a flowchart illustrating a second method for obtaining service configuration parameters according to an embodiment of the present disclosure;

fig. 4 is a schematic diagram of implementation components of a method for obtaining service configuration parameters according to an embodiment of the present disclosure;

FIG. 5 is a schematic diagram of an implementation component implementing a method for obtaining service configuration parameters according to an embodiment of the present disclosure;

FIG. 6 is a schematic diagram of an online decision component implementing a method of obtaining business configuration parameters, in accordance with an embodiment of the present disclosure;

fig. 7 is a schematic composition diagram of an apparatus for obtaining service configuration parameters according to an embodiment of the present disclosure;

fig. 8 is a block diagram of an electronic device for implementing the method for obtaining service configuration parameters according to the embodiment of the present 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 present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

In mass data services, such as searching, advertisement delivery, product recommendation and other services, any automatically configurable parameter can be regarded as a service configuration parameter, and parameters such as the recommended number of products provided for a user, the number of products displayed on a display interface of the user each time and the like can be regarded as service configuration parameters.

Fig. 1 is a system architecture of a mass data service system according to an embodiment of the present disclosure. As shown in fig. 1, the system architecture generally includes a front end 11 and a back end 12. The front end 11 is, for example, any reasonable terminal, and is configured to display or display a search result, an advertisement recommendation result, or a product recommendation result provided by the back end 12 for the user. The terminal can be at least one of a mobile phone, an all-in-one machine, a tablet computer, a wearable intelligent device and the like. The background 12 is a server, and is configured to provide the front end 11 with search results, advertisement display results, or product recommendation results by using a funnel-type architecture, so that the front end 11 displays the search results, the advertisement display results, or the product recommendation results to the user.

The funnel type framework screens mass data layer by layer in a layered mode based on search keywords or product keywords input by a user, and returns a final search or recommendation result to the user at least through several stages of data recall, data rough arrangement, data fine arrangement and the like in sequence.

Taking a product recommendation service as an example, the recall stage is used for placing products which meet the product keywords input by the user in the mass data into the candidate data set according to the request of the user. And the data rough-sorting stage is used for roughly sorting the candidate data in the candidate data set. And in the data fine-ranking stage, the ranking results obtained in the data coarse-ranking stage are ranked finely by combining user characteristics such as user preference, purchasing habits and the like so as to match the requirements of the user as much as possible.

Each stage has certain service configuration parameters, such as the length of a candidate data set in a recall stage; the data rough arrangement stage has data recommendation length (the arrangement length of rough arrangement); in the data fine-sorting stage, there is a data recommendation length (sorting length for fine sorting). Therefore, accurate configuration of the service configuration parameters is performed for each stage, so that product recommendation is certainly more accurate, and good use experience can be brought to a user from a user side. From the system side, the system can provide more accurate service configuration parameters for users, the service capability of the system such as recommendation capability is greatly improved, and the system is easy to popularize in engineering and high in practicability.

The method for obtaining the service configuration parameters can determine or select the service configuration parameters for any one of the stages, and the method considers the influence of the service indexes on the determination or selection of the service configuration parameters, adopts a model (target model) capable of obtaining the service index gain to obtain the service index gain, and realizes the accurate obtaining of the target service configuration parameters based on the service index gain obtained by the target model. The scheme for acquiring the service configuration parameters based on the service index gain quantity can effectively improve the service experience of the user and bring benefits for the service. The following detailed description of the present solution can be specifically referred to.

The processing logic of the method for obtaining the service configuration parameters according to the embodiment of the present disclosure may be deployed in any reasonable server. The servers include ordinary servers, cloud servers, servers used in professional fields such as the aforementioned server used as the background 12 in fig. 1. The device for obtaining the service configuration parameters in the embodiments of the present disclosure may be used as an implementation component in a server, and may also be used as a device that can be implemented separately to implement the processing logic of the method for obtaining the service configuration parameters in the present disclosure separately.

The following describes a method for obtaining service configuration parameters according to an embodiment of the present disclosure. As shown in fig. 2, the method includes:

s201: acquiring flow request data;

the traffic request data may be a search request, a recommendation request, etc. of the user. The user can input a search request and a recommendation request through the front end 11, and in this step, the background 12 can obtain the traffic request data by receiving the search request and the recommendation request sent by the front end 11.

The traffic request data may be a request of a single user or a request of a plurality of users. And if the request is the request of a single user, obtaining the target service configuration parameters determined for the single user based on the subsequent scheme. And if the request is the request of a plurality of users, obtaining the target service configuration parameters determined for each user based on the subsequent scheme. The target service configuration parameters determined for each user may be completely different, partially the same, or completely the same, depending on the actual situation.

S202: inputting the characteristic data of the flow request data into at least one group of target models to obtain the output parameters of each group of target models; the output parameter representation of each group of target models is a service index gain quantity generated by performing service processing on the flow request data from the original service configuration parameter corresponding to each group of target models to the candidate service configuration parameter corresponding to each group of target models;

the characteristic data of the traffic request data includes, but is not limited to, the following: user attributes such as gender, age, net age, etc.; user device attributes such as device model, location data, etc.; user behavior attributes such as access behavior over the last period of time, e.g., the number of accesses to back office 12 over the last 7 days, the number of interactions with the product merchant over the last 1 day, etc.

In this step, each group of target models corresponds to two service configuration parameters: original service configuration parameters and candidate service configuration parameters. The original service configuration parameter is a service configuration parameter used by the original background 12 when performing service processing, for example, an original recommended length provided by the background 12 in a recommended or search service. The candidate service configuration parameter is a service configuration parameter that may be adopted when performing service processing, such as a recommended length of a candidate provided by the background 12 in a recommended or search service.

Each set of target models of the disclosed embodiments may perform two functions. The first function is to adopt the original service configuration parameters corresponding to the first function to perform service processing on the flow request data. The second function is to adopt the candidate service configuration parameter corresponding to itself to process the service of the flow request data. Each group of target models generates an output parameter aiming at the two functions, and the output parameter represents the service index gain quantity generated by performing service processing on the flow request data from the original service configuration parameter corresponding to the group of target models to the candidate service configuration parameter corresponding to the group of target models. That is, the output parameter of each group of target models is characterized as the service index gain.

In the application scenarios of mass data services such as search, advertisement delivery, recommendation, etc., the service index may be an index of service attention, such as click rate, purchase rate, satisfaction, reading rate, conversion rate, etc. The original recommended length and the candidate recommended length may specifically be selectable lengths of candidate data sets in the recall stage of the application scenario, may be selectable data recommended lengths in the rough ranking stage, and may also be selectable data recommended lengths in the fine ranking stage. From the user-level, the user may see the products presented using the original recommended length or the candidate recommended length.

In practical applications, the gain amount may be a positive gain or a negative gain. For an internet company or unit, the gain amount is positive, which indicates that there is a business profit, and the larger the gain amount is positive, the higher the profit.

The original service configuration parameter and the candidate service configuration parameter of the embodiment of the present disclosure may be parameters selected from preset parameters. Taking the service configuration parameters as the data recommendation lengths in the fine ranking stage as an example, the preset data recommendation lengths include lengths with values of 100, 200, 300, 400, 500 and the like, one of the lengths can be selected as an original service configuration parameter, and at least one of the lengths is selected as a candidate service configuration parameter. Illustratively, the data recommendation length with the value of 100 is used as an original service configuration parameter, and the data recommendation lengths with the values of 200 and 300 are used as candidate service configuration parameters.

Since each group of target models corresponds to one candidate service configuration parameter, the number of target model groups can be consistent with the number of candidate service configuration parameters. Selecting several parameters as candidate service configuration parameters, taking a corresponding number of target model sets, and inputting the flow request data into the corresponding number of target model sets.

In practical applications, the target model is any reasonable model capable of achieving the calculation of the gain amount of the service indicator, such as a statistical model for calculating the gain amount based on statistics, such as a gain (uplift) model based on causal inference.

S203: and determining a target service configuration parameter from the original service configuration parameter and the candidate service configuration parameter corresponding to each group of target models based on the service index gain of each group of target models, wherein the target service configuration parameter is used for performing service processing on an object generating flow request data.

In this step, based on the output parameters of each group of target models, target service configuration parameters are selected from the original service configuration parameters and the candidate service configuration parameters corresponding to each group of target models. The object of generating the traffic request data is the user who generated the traffic request data.

In the foregoing S201 to S203, the target service configuration parameter is obtained based on the service index gain amount output by the target model, and in consideration of the influence of the service index on the determination or selection of the service configuration parameter, the model (target model) capable of obtaining the service index gain amount is used to obtain the service index gain amount, so as to accurately obtain the target service configuration parameter. The scheme for acquiring the service configuration parameters based on the service index gain quantity can not only determine accurate service configuration parameters for users, but also improve the service experience of the users. The target service configuration parameters are obtained based on the service index gain, benefits can be brought to the service, and usability and practicability of the method of the embodiment of the disclosure are reflected.

As an alternative embodiment, as shown in fig. 3, the scheme for determining the target service configuration parameter from the original service configuration parameter and the candidate service configuration parameter corresponding to each set of target models based on the service index gain of each set of target models may be implemented as follows:

s303: determining target service index gain amount from the service index gain amount of each group of target models; the target service index gain is at least one of the service index gain with the largest value and reaching a target threshold value in the service index gains of the target models;

s304: and determining target service configuration parameters from the original service configuration parameters and the candidate service configuration parameters corresponding to each group of target models based on the target service index gain.

In fig. 3, the schemes shown in S301 and S302 are understood with reference to the schemes shown in S201 and S202.

In S303 to S304, there are some cases as follows:

and under the condition that the target service index gain is the service index gain with the largest value in the service index gains output by each group of target models, taking the candidate service configuration parameter corresponding to the target model which obtains the service index gain with the largest value in the service index gains output by each group of target models as the target service configuration parameter.

Illustratively, assume that there are two candidate traffic configuration parameters, parameter 1 and parameter 2. The first set of object models corresponds to parameters 1 and the second set of object models corresponds to parameters 2. And in the first group of target models and the second group of target models, if the service index gain output by the first group of target models is the largest, taking the parameter 1 as a target service configuration parameter.

And under the condition that the target service index gain is the service index gain reaching the target threshold value in the service index gain output by each group of target models, taking the candidate service configuration parameter corresponding to the target model which obtains the service index gain reaching the target threshold value in the service index gain output by each group of target models as the target service configuration parameter.

Exemplarily, it is assumed that there are three candidate traffic configuration parameters, parameter 1, parameter 2 and parameter 3. The first set of object models corresponds to parameters 1, the second set of object models corresponds to parameters 2, and the third set of object models corresponds to parameters 3. In the first group-third group target models, the service index gain output by the third group target model reaches a target threshold, and the service index gain output by the second group target model and the first group target model does not reach the target threshold, and then the parameter 3 is used as a target service configuration parameter.

And under the condition that the target service index gain is the service index gain with the maximum value and reaching the target threshold value in the service index gain output by each group of target models, obtaining the candidate service configuration parameters corresponding to the target model with the maximum value and reaching the target threshold value in the service index gain output by each group of target models.

Illustratively, assume that there are three candidate traffic configuration parameters, parameter 1, parameter 2, and parameter 3. The first set of object models corresponds to parameters 1, the second set of object models corresponds to parameters 2, and the third set of object models corresponds to parameters 3. In the first group-third group of target models, if the service index gain output by the second group of target models is the maximum and reaches the target threshold, the parameter 2 is used as the target service configuration parameter.

As can be seen, in the scheme shown in fig. 3, the target service index gain is determined based on the service index gain with the largest value and/or reaching the target threshold, and then the candidate service configuration parameter corresponding to the target model with the largest service index gain and/or the candidate service configuration parameter corresponding to the target model with the service index gain reaching the target threshold are/is used as the target service configuration parameter, so that the target service configuration parameter is accurately obtained under the condition of maximizing the service gain. Easy to implement and easy to use in engineering.

In practical application, the original service configuration parameters corresponding to each group of target models may be the same or different. As an optional implementation, the original service configuration parameters corresponding to each group of target models are the same. Based on the above, the output parameters of each group of target models represent the service index gain generated by changing the service processing of the flow request data from the service processing by using the same original service configuration parameters into the service processing by using the candidate service configuration parameters corresponding to each group of target models. And under the condition that the service index gain quantity output by each group of target models does not reach the target threshold value, responding to that the service index gain quantity output by each group of target models does not reach the target threshold value, and taking the same original service configuration parameters as the target service configuration parameters.

In a popular way, when the service index gain output by each group of target models does not reach the target threshold, the original service configuration parameters are used as the target service configuration parameters. Therefore, the target service configuration parameters are accurately acquired, the accurate service configuration parameters can be determined for the user, and the good experience of the user is effectively improved.

In the foregoing solution, the target threshold may be a value set empirically, such as 0.5, 0.7, etc. In addition, the target threshold may be calculated by the target model group. For example, the target threshold is obtained by screening a threshold from at least one preset candidate threshold based on a historical service index gain obtained by at least one set of target model for historical traffic request data, a calculation power consumption amount under a historical service configuration parameter corresponding to the historical service index gain, and a preset service calculation power constraint condition. Further, the calculated power consumption amount can be restricted under a service calculated power restriction condition, and a target threshold value is screened from at least one preset candidate threshold value so as to realize calculation of the target threshold value. Equivalently, the target threshold is a numerical value obtained based on the computing power resource. It can be understood that the target threshold is calculated to achieve more accurate acquisition of the service configuration parameters. Therefore, the aforementioned calculation of the target threshold implemented based on the calculation power consumption amount and the service calculation power constraint condition and the acquisition of the service configuration parameter based on the target threshold can be regarded as a scheme for implementing accurate acquisition of the service configuration parameter under the action or constraint of the service calculation power constraint condition. On one hand, the method can effectively save the calculation power, and on the other hand, the normal operation of the service is not influenced.

The service computing power constraint condition is a condition related to computing power resources provided by the background 12, and embodies the computing power provided by the background 12 to the service. Which may be any reasonable condition. For example, the background 12 originally provides 100 processors for maintaining computational power consumption. Each time the value of the service configuration parameter is increased by a certain amount, for example, 100 or 150, the background 12 needs to increase certain computing resources, for example, 100 or 150 processors, so as to support the computational power consumption caused by the increase of the service configuration parameter.

In the embodiment of the present disclosure, the calculation power consumption is constrained under the service calculation power constraint condition in consideration of the calculation power provided by the background 12 and the influence of the calculation power consumption on the service configuration parameters, and the service configuration parameters are obtained based on the service index gain. The service effect can be maximized under the constraint condition of service computing power, for example, the candidate configuration parameter corresponding to the target model with the maximum service index gain is selected as the target service configuration parameter, wherein the maximization of the service index gain reflects the maximization of the service effect.

In an application scenario of a massive data service such as search, advertisement delivery, recommendation, etc., the service index may be an index of interest of the service, such as click rate, purchase rate, satisfaction, reading rate, conversion rate, etc. Therefore, the candidate service configuration parameter in the embodiment of the present disclosure may be a candidate recommended length in the search service or the recommended service, and the original service configuration parameter may be an original recommended length in the search service or the recommended service. Based on this, the service index gain amount may be a service gain generated by performing service search or service recommendation on the traffic request data by using each group of target models from an original recommended length corresponding to each group of target models to a candidate recommended length corresponding to each group of target models.

In the scheme, the service recommendation length in the application scenes such as the search service, the advertisement delivery service, the recommendation service and the like can be used as the service configuration parameter, the service search or the service recommendation of the target model on the flow request data from the original recommendation length corresponding to the target model is changed into the service gain generated by the service search or the service recommendation of the candidate recommendation length corresponding to the target model, and the service gain is used as the service index gain, so that the method can be greatly adapted to the application scenes, and has good practicability.

The business profit may refer to changes brought to indexes concerned by the business, such as increment of click through rate, increment of purchase rate, and the like. The method can also be the economic benefit brought by the indexes concerned by the business, such as the economic benefit brought by the increment of the click rate and the economic benefit brought by the increment of the purchase rate.

In an application scenario of mass data services such as search, advertisement delivery, recommendation, and the like, the service index gain based on each group of target models determines, from original service configuration parameters and candidate service configuration parameters corresponding to each group of target models, that a target service configuration parameter may be: and determining a target recommendation length from the original recommendation length and the candidate recommendation length corresponding to each group of target models based on the service income of each group of target models, wherein the target recommendation length is used for performing service search or service recommendation on an object generating flow request data, such as a user.

The target recommendation length is acquired based on the service income in application scenes such as a search service, an advertisement putting service and a recommendation service, so that the acquisition accuracy of the recommendation length can be ensured, the income can be maximized as much as possible, and the economic value is improved.

As an alternative, each set of target models in the embodiment of the present disclosure may be a trained causal inference-based gain model, and each set of causal inference-based gain models is obtained by training through historical traffic request data and tag data acquired for each set of causal inference-based gain models.

The tag data is obtained based on feedback data of service configuration parameters recommended by the object to the historical traffic request data by the historical background 12. Illustratively, the product recommendation length recommended by the history background 12 for the object is 100 or 50, the click condition of the object under the condition that the product recommendation length is 100 or the click condition of the object under the condition that the product recommendation length is 50 is collected, the click rate of the object under the condition that the product recommendation length is 100 or the click rate of the object under the condition that the product recommendation length is 50 is calculated based on the click condition, and the click rate is used as the tag data. And training a gain model (simply called a gain model or an uplift model) based on causal inference by taking historical flow request data and click rate as training samples.

And training the gain model by using the historical traffic request data and the label data to obtain an accurate target model. Accurate business index gain can be obtained by using the accurate target model, and accurate acquisition of target business configuration parameters is further ensured.

The technical solution of the embodiment of the present disclosure is described in detail below with reference to fig. 4 to 6.

As shown in fig. 4, the components for implementing the method for obtaining service configuration parameters according to the embodiment of the present disclosure include an online exploration component, an offline training component, and an online decision component. And the online exploration component is used for acquiring historical traffic request data. And the offline training component is used for taking the historical flow request data and the label data as training samples and training at least one group of uplift models to be trained by utilizing the training samples. And the online decision-making component is used for determining target service configuration parameters for the user generating the flow request data by utilizing the online flow request data, the trained uplift model and the target threshold value provided by the offline training component.

The following takes the data recommendation length of the fine ranking stage in the recommended service application scenario as a service configuration parameter in the recommended service scenario, and details the method according to the embodiment of the present disclosure are described with reference to the above components and shown in fig. 5.

It is understood that for historical traffic request data such as historical recommendation requests, the background 12 may recommend different recommendation lengths for different users, such as 100 recommendation lengths for user 1, 200 recommendation lengths for user 2, 300 recommendation lengths for user 3, and so on. And regarding the different lengths as different gears of the service configuration parameters of the fine ranking stage. Each gear corresponds to a recommended length. For example, for gear 1, the corresponding recommended length is 100; in gear 2, the corresponding recommended length is 200.

The back office 12 may collect historical traffic request data for users in different gear positions. The online exploration component collects or extracts historical flow request data of the same proportion from the historical flow request data of the user of at least one gear. For example, 1% of the flow request data is collected or extracted from the historical flow request data of the user in the gear 1 (the recommended length is 100), 1% of the flow request data is collected or extracted from the historical flow request data of the user in the gear 2 (the recommended length is 200), and 1% of the flow request data is collected or extracted from the historical flow request data of the user in the gear 3 (the recommended length is 300).

And selecting the recommended length of one gear from the recommended lengths of different gears as a control group, and using the recommended lengths of other gears as an experimental group.

Illustratively, if the recommended length with the value of 100 is selected as a control group to be used as an original service configuration parameter (original recommended length), and the recommended lengths with other values are selected as an experimental group to be used as candidate service configuration parameters (candidate recommended lengths).

The number of uplift models to be trained and the number of candidate service configuration parameters are kept consistent.

Illustratively, under the condition that the original recommended length is 100 and the candidate recommended length is 200 or 300, two groups of uplift models to be trained are included. And the candidate recommended length corresponding to the first group of uplift models to be trained in the two groups of uplift models to be trained is 200. The candidate recommended length corresponding to the second group of uplift models to be trained is 300.

And the two groups of uplift models to be trained can process the service by changing the original recommended length into the candidate recommended length corresponding to each other. Illustratively, the first set of uplift models to be trained implements: performing service processing on historical flow request data by adopting a recommended length (original recommended length) with a value of 100; and performing service processing on the historical flow request data by adopting the recommended length (candidate recommended length) with the value of 200. The second group of uplift models to be trained adopt the recommended length (original recommended length) with the value of 100 to perform service processing on the historical flow request data; and performing service processing on the historical traffic request data by adopting the recommended length (candidate recommended length) with the value of 300.

It can be appreciated that the online exploration component sends the collected or extracted historical traffic request data to the offline training component, and the offline training component performs extraction of feature data of the historical traffic request data.

The offline training component collects behavioral data fed back by the user. And calculating the service index based on the behavior data. The behavior data refers to actual click conditions or actual purchase conditions of products recommended by the user with a certain recommended length provided by the background 12 for the user in the past (history), and actual service indexes such as the calculated actual click rate or actual purchase rate are used as tag data. And the offline training component takes the received historical flow request data and the calculated label data as training samples to train each group of uplift models to be trained. As for the two groups of uplift models to be trained.

Taking training of the first group of uplift models to be trained as an example, because the original recommended length of the uplift models to be trained is 100, and the candidate recommended length is 200, the feature data of the historical flow request data input to the uplift models to be trained includes the feature data of the historical flow request data of gear 1 and the feature data of the historical flow request data of gear 2. The tag data thereof may be the actual click rate of the user in the case where the recommendation length is 100 and the actual click rate of the user in the case where the recommendation length is 200.

The training process is roughly as follows: inputting the characteristic data of the two-gear historical flow request data into the to-be-trained uplift model, and outputting a probability value by the to-be-trained uplift model. The probability value represents the change degree of the handling service index, such as click rate or purchase rate, of the service of the uplift model to be trained from adopting the original recommendation length 100 to adopting the candidate recommendation length 200. If the probability value is 30%, the service index, such as the click rate or the purchase rate, of the service processed by adopting the original recommended length 100 to the candidate recommended length 200 is changed by 30% and increased by 30%. And obtaining a predicted value of a service index such as the click rate under the condition that the candidate recommendation length is 200 according to the known actual click rate of the user and the probability value output by the uplift model to be trained under the condition that the recommendation length is 100. Illustratively, knowing that the actual click rate of the user is 20% and the probability value output by the uplift model to be trained is 30% when the recommendation length is 100, the predicted value of the traffic indicator is (1 + probability value) = (1 + 30%) + 20% =26% when the candidate recommendation length is 200. That is, if the object originally recommended with the recommended length of 100 is changed to the object recommended with the recommended length of 200, the click rate of the user is 26%, and the click rate is improved compared with the object recommended with the recommended length of 100.

And calculating the value of the loss function based on the predicted value of the click rate and the actual click rate of the user under the condition that the recommended length is 200. And under the condition that the loss function is as small as possible, namely under the condition that the predicted value of the click rate is fitted to the actual click rate of the user under the condition that the recommended length is 200 as much as possible, the training is completed. The loss function may be any reasonable function, such as a mean square error function, a square error function, and the like.

In the application scenario, the uplift model to be trained may be a Multi-Treatment X-Learner algorithm model. It will be appreciated that the characteristic data input to the uplift model to be trained includes user attributes such as gender, age, and also device attributes such as model, location, etc. of the user device, and also user behavior attributes such as number of visits within the last period of time. If the attributes are considered as input features of the uplift model to be trained, the Multi-Treatment X-leaner algorithm model has the characteristic of being capable of performing feature screening autonomously for the input features. I.e. having the property of being able to screen features that are not useful or useful for model training.

For example, for the input feature of gender, if it is found through analysis, for a male user, the probability value obtained by changing the service processing of the uplift model to be trained from the original recommended length of 100 to the candidate recommended length of 200 is 0.3; for the female user, the probability value obtained by changing the uplift model to be trained from the original recommended length of 100 to the candidate recommended length of 200 for business processing is 0.28. From the viewpoint of gender, the difference between the values of the probability values 0.3 and 0.28 is not large, which indicates that the application meaning of the prediction of the service index from the viewpoint of gender is not large, the input feature of gender is taken as a useless feature, and the input feature of gender can not be considered in the subsequent scheme of continuously carrying out iterative training on the uplift model to be trained. For the input feature of the region, if the input feature is found through analysis, for the shanghai user, the probability value obtained by performing service processing on the uplift model to be trained from the original recommended length of 100 to the candidate recommended length of 200 is 0.7. For the Beijing user, the probability value obtained by changing the uplift model to be trained from the original recommended length of 100 to the candidate recommended length of 200 for service processing is 0.3. From the perspective of the region, the numerical value difference between the probability values 0.3 and 0.28 is large, the input feature of the region serves as a useful feature, and the input feature of the region can be considered in the subsequent scheme of continuing iterative training on the uplift model to be trained.

The uplift model to be trained has the characteristics, so that the calculated amount can be reduced as much as possible in the iterative training or adjusting process of the model on one hand, and the training time can be shortened on the other hand.

It can be understood that the training process of the uplift models to be trained in other groups is referred to the foregoing related description of the first group, and repeated details are omitted.

In the application scenario, each group of target models is obtained after each group of uplift models to be trained is trained. The trained sets of target models can be sent by the offline training component to the online decision-making component for use, or sent to the online decision-making component for use after the evaluation passes.

In the application scenario, the trained target model can be evaluated by the following evaluation method. And the off-line training component acquires a test sample and evaluates each group of target models based on the test sample. Taking an example of an object model 1 (from an object model which performs business processing by using the original recommended length 100 to the candidate recommended length 200) in the multiple sets of object models, the object model 1 may be simply referred to as an object model which changes from a gear 1 (corresponding to the recommended length of 100) to a gear 2 (corresponding to the recommended length of 200). Its test sample includes gear 1 flow request data and gear 2 flow request data. The traffic request data in the test sample may be current traffic request data as opposed to historical traffic request data. The characteristic data of the flow request data of the two gears are extracted, the characteristic data of the flow request data of the two gears are input into the target model 1, the probability value output by the target model 1 is positive and larger, the improvement degree of the service index is larger after the gear 1 is changed into the gear 2, and the economic benefit is higher for an internet company. If the improvement degree of the service index is larger, if the preset improvement threshold is reached, the target model is trained better, and the model is considered to pass the evaluation.

Both the training process and the evaluation process above can be accomplished by an offline training component.

By training the uplift model to be trained by using the historical traffic request data and the label data, an accurate target model can be obtained. Based on the model evaluation method, the accuracy of the target model is further ensured.

After the training of each group of target models is completed or after the training is completed and the evaluation is passed, the offline training component can perform offline simulation of computational power consumption based on historical flow data so as to determine a target threshold value from at least two preset candidate threshold values under the constraint of a business computational power constraint condition.

Specifically, the simulation algorithm is as follows:

A. two or more candidate thresholds are previously configured or set empirically.

For example, three candidate thresholds are configured, and the values are 0, 0.5 and 0.8 respectively.

B. Selecting one candidate threshold value from the three candidate threshold values as a score threshold value of the uplift model;

for example, pick 0 as the score threshold for the model.

C. Inputting the characteristic data of the historical flow request data into each group of target models, and outputting respective probability values by each group of target models;

exemplarily, inputting characteristic data of historical flow request data of the gear 1 into each group of target models;

assume that the target model 1 in each set of target models represents that the service processing is changed from adopting the original recommended length (recommended length in gear 1, value is 100) to adopting the candidate recommended length in gear 2 (value is 200). The target model 2 in each group of target models represents that the service processing is changed from the service processing by adopting the original recommended length to the service processing by adopting the candidate recommended length (the value is 300) of the gear 3. The characteristic data of the historical flow request data of the gear 1 is input into the target model 1, and the obtained output parameter of the target model 1 represents a gain quantity (gain quantity 1) of a service index gain quantity, such as a click rate, obtained by changing from service processing by adopting an original recommended length to service processing by adopting a candidate recommended length of the gear 2. The characteristic data of the historical flow request data of the gear 1 is input into the target model 2, and the obtained output parameter of the target model 2 represents a gain quantity (gain quantity 2) of a service index gain quantity, such as a click rate, obtained by changing from service processing by adopting an original recommended length to service processing by adopting a candidate recommended length of the gear 3. Wherein, the gain amount 1 and the gain amount 2 are used as the historical traffic indicator gain amount.

In all the groups of target models, judging which groups of target models have service index gain exceeding the score threshold of the uplift model; and taking the candidate recommended length corresponding to the target model exceeding the score threshold value of the uplift model in all the groups of target models as the new recommended length of the user generating the historical flow request data of the gear 1.

Exemplarily, assuming two groups of target models including a target model 1 and a target model 2, where a gain 1 of the target model 1 does not exceed a score threshold of the uplift model, and a gain 2 of the target model 2 exceeds the score threshold of the uplift model, a candidate recommended length (300) corresponding to the target model 2 is used as a new recommended length of the user requesting the historical traffic data. The original recommended length of the user 1 generating the historical traffic request data is 100, the new recommended length is 300, and subsequently, the new recommended length 300 can be adopted for service recommendation for the user 1.

And if the gain 1 and the gain 2 both exceed the score threshold of the uplift model, selecting the candidate recommended length corresponding to the target model with the maximum gain as the new recommended length of the user of the historical flow request data of the gear 1. For example, if the gain amount 2 is greater than the gain amount 1, the candidate recommended length (300) corresponding to the target model 2 is selected as the new recommended length for the user of the historical traffic request data for gear 1.

If neither gain 1 nor gain 2 exceeds the score threshold of the uplift model, then it remains at the original recommended length of 100.

In this way, a new recommended length may be recalculated for all users that have generated historical traffic request data for a (historical) period of time or the user may be kept at the original recommended length based on the actual magnitude of the gain amount.

In practical application, the new recommended length may be larger than the original recommended length, or may be smaller than the original recommended length. If the original recommended length is 100, the new recommended length may be 50 or 200. The new recommended length can be used as a historical service configuration parameter corresponding to the historical service index gain.

D. Obtaining the average recommended length of all users generating historical flow request data in the historical time, and converting the average recommended length into the calculated power consumption in the historical time;

here, the computational power consumption amount is the computational power resource situation of the background 12 that is consumed for providing computational power support for all users in the period of time during the historical period.

Assuming that there are 100 users in the historical time period and the original recommended length is 100, the new recommended length calculated for 50 of the users is 50 and the new recommended length calculated for the other 50 users is 150 through the calculation of the step C. Then, the average recommended length = the sum of the new recommended lengths of all users/total number of users = (50 × 50+50 × 150)/100 =100.

Here, it is empirically known that the background 12 is required to provide 100 processors in the case where the average recommended length is 100 to satisfy the computational power consumption in the case where the average recommended length is 100. The background 12 is required to provide 200 processors in the case of an average recommended length of 200 to satisfy the computational power consumption in the case of an average recommended length of 200. Based on the foregoing experience, the calculated power consumption amount to scale the average recommended length 100 is approximately equal to or equal to the calculated power consumption amount of 100 processors.

E. And D, according to the constraint of the service calculation force constraint condition, determining whether a reasonable candidate threshold value needs to be selected from other candidate threshold values as a final score threshold value of the uplift model or not based on the calculation force consumption amount in the step D, namely, the reasonable candidate threshold value is used as a target threshold value. Specifically, one of the candidate threshold values is selected from the other candidate threshold values as a score threshold value of the uplift model, and the steps C, D and E are repeatedly executed until the business calculation force constraint condition is met and a suitable candidate threshold value is found.

The business computing constraints may provide 100 processors for the back office 12 to maintain computing power costs. Each time the value of the recommended length is increased by a certain amount, for example, by 50, the background 12 needs to increase certain computing resources, for example, by 50 processors, so as to support the computational power consumption caused by the increase of the service configuration parameters.

Compared with historical flow request data in the same historical time, when the score threshold value of the uplift model is 0.5, the average recommended length calculated for each user in the historical time is converted into the calculated power consumption, and the background can meet the calculated power consumption without adding a processor, so that the score threshold value of the uplift model of 0.5 is appropriate. If the processor must be added to satisfy the calculated power consumption, the score threshold of the uplift model, which is illustrated as 0.5, is not the most appropriate score threshold, and it is determined whether the most appropriate score threshold exists among the other candidate thresholds according to steps C, D and step E, as described above.

Compared with historical flow request data in the same historical time, the calculated power consumption can be met by a background on the basis of not increasing a processor under the condition that the score threshold of the uplift model is 0.8, and the calculated power consumption can be met only by increasing the processor under the condition that the score threshold of the uplift model is 0.5, which indicates that the score threshold of the uplift model of 0.8 is proper.

Therefore, the setting of the business computing power constraint condition of the disclosure can select one threshold value which can effectively reduce the computing power consumption from a plurality of candidate threshold values as the target threshold value. The influence of calculation power consumption and business calculation power constraint conditions on the selection of the target threshold is considered, so that calculation power resources are saved more when the target threshold is selected.

After the offline training component completes the offline computational power simulation, the optimal uplift score threshold value, namely the target threshold value, and each trained target model are synchronized to the online decision-making component. Wherein each object model can be recorded in the form of a file to the online decision component.

The online decision component loads each target model file, and determines the target recommended length for the user by the following method in combination with fig. 6 when the traffic request data of the user exists.

1. Extracting characteristic data of the flow request data of the user and inputting the characteristic data into each group of target models;

2. each group of target models calculates and outputs service gains such as click rate gain generated by performing service processing on the user from the original recommended length (such as 100) to the candidate recommended length corresponding to each group of recommended models;

3. selecting the service income with the maximum value and reaching the target threshold value from the service income output by each group of target models;

4. taking the candidate recommended length corresponding to the target model of the service income which obtains the maximum service income value and reaches the target threshold value as the target recommended length of the user;

illustratively, assume two sets of object models including object model 1 (corresponding to a recommendation length candidate of 200) and object model 2 (corresponding to a recommendation length candidate of 300). And selecting the business profit output by the target model 1 from the business profit 1 and the business profit 2 as the target business index gain amount if the business profit output by the target model 1 is larger than the business profit output by the target model 2 and the business profit output by the target model 1 reaches the target threshold value. And outputting the target service index gain amount by the target model 1, and if the candidate recommended length corresponding to the target model 1 is 200, taking 200 as a new recommended length of the user, wherein the new recommended length can be used as the target recommended length.

In practical application, if the service income output by each group of target models does not reach the target threshold value or the service income values are the same, the original recommended length of the user is kept, and the original recommended length is continuously adopted for recommending the product for the user.

5. And recommending the product for the user by adopting the target recommendation length.

In the foregoing solution, the service profit with the largest value or reaching the target threshold may also be selected, and the candidate recommended length corresponding to the target model of the service profit is obtained as the target recommended length of the user.

In the above description, the target models are taken as two groups, and in practical application, the number of the target models may also be other values, for example, N groups of target models, where N is a positive integer greater than 2. Alternatively, N may also be 1.

In the scheme, the target recommended length is obtained based on the maximum value of the service income and the target threshold value. The acquisition of the target recommendation length meets the maximization of the business effect on one hand and meets the business calculation force constraint condition on the other hand. In a popular way, the target recommendation length is obtained accurately, and under the condition that the constraint condition of business calculation force is met, business effects such as business income maximization are maximized.

It can be understood that the precise ranking stage in the recommended service application scene is a stage consuming computational power very much, the data recommendation length of the precise ranking stage is in direct proportion to the computational power consumption, and by using the technical scheme, the conversion efficiency between the computational power and the service effect can be effectively improved.

Each set of uplift models to be trained using the Multi-Treatment X-Learner algorithm model is described below.

Each group of uplift models to be trained consists of two regression models. Taking a first group of uplift models to be trained in each group of uplift models to be trained as an example, the first group of uplift models to be trained is used for processing the service from adopting the original recommended length (such as 100) to changing into the candidate recommended length (such as 200).

The first group of uplift models to be trained consists of two regression models mu ₀ And mu ₁ And (4) forming. Wherein,

μ ₀ (x)＝E[Y(0)|X＝x] (1)

μ ₁ (x)＝E[Y(1)|X＝x] (2)

where x represents traffic request data for the same user. Y (0) =100 (original recommended length); y (2) =200 (candidate recommended length). E is a data operation, representing a mathematical expectation. Mu.s ₀ The original recommended length can be regarded as a control regression model. Mu.s ₁ Corresponding to the candidate recommended length, it can be regarded as a trial set regression model.

It can be understood that in each group of uplift models to be trained, historical flow request data of the same user when the recommended length of the user is the candidate recommended length of 200 is input into the regression model of the comparison group, and the obtained result mu ₀ (x ⁰ ) And true tag data Y at the time of candidate recommended length 200 ⁰ Performing residual calculation to obtain D ⁰ . The same one is usedInputting historical flow request data of a user when the recommended length of the user is the candidate recommended length 100 into a test group regression model, and inputting real label data Y when the recommended length of the user is the candidate recommended length 100 ¹ And the result obtained is ₀ (x ¹ ) Residual error calculation is carried out to obtain D ¹ 。

D ¹ ＝Y ¹ -μ ₀ (x ¹ ) (3)

D ⁰ ＝μ ₀ (x ⁰ )-Y ⁰ (4)

The true tag data at the time of the candidate recommended length 200 and the true tag data at the time of the candidate recommended length 100 can be obtained by referring to the method for obtaining the tag data in the training sample.

Calculating the residual error D ¹ 、D ⁰ As label data of the uplift model to be trained, the regression model τ is trained again ₁ (x) And τ ₀ (x) .1. The Wherein the regression model τ ₁ (x) And τ ₀ (x) Expressed mathematically as:

τ ₁ (x)＝E[D ¹ |X＝x] (5)

τ ₀ (x)＝E[D ⁰ |X＝x] (6)

regression model τ ₁ (x) And τ ₀ (x) And carrying out weighted average, wherein the weight g (x) can be a function of input x, or g =0.5, namely obtaining the first group of uplift models to be trained. Expressed mathematically as τ (x):

τ(x)＝g(x)*τ ₀ (x)+(1-g(x))*τ ₁ (x) (7)

the expressions of the uplift model to be trained and equation (7) for the other groups are similar. The training of the uplift model to be trained is just to train the tau of the model ₁ (x) And τ ₀ (x) .1. The The training scheme consists in determining τ ₁ (x) And τ ₀ (x) And the weight g (x) is taken as a value, so that the formula (7) can be applied to calculate the service index gain generated by processing the service by the user from the original recommended length (such as 100) to the candidate recommended length (such as 200).

According to the scheme, the procedures of training, determining the target threshold value, obtaining the target recommended length and the like are boxed in a whitening mode, and the method is very explanatory. Reasonable target service configuration parameters such as target recommendation length in a fine ranking stage can be determined for flow request data of different users, personalized service experience is provided for the users, and user experience is improved. Meanwhile, the calculation power consumption condition is considered, the service index gain is maximized under the condition of meeting the service calculation power constraint condition, and objective economic benefits can be brought.

According to the method and the device, the computational power consumption is simulated off line, the computational power consumption can be estimated off line, the optimal uplift score threshold value is selected from the candidate threshold values according to the limitation of business computational power constraint conditions, the computational power constraint controllability is good, the process is transparent, and the implementation cost is low.

Compared with on-line training, the off-line training can avoid factors such as interference on-line users in the on-line training process, reduce training cost and ensure better user experience.

It should be noted that the above algorithm is only one of them. In the field of causal inference, there are other algorithmic models that can be used as causal inference-based gain models for the present disclosure, including but not limited to: S-Learner, T-Learner, R-Learner, 2SLS, DR Learner, DRIV Learner, CEVAE, dragonNet, uplift Tree, and the like.

Since the embodiment of the present disclosure can customize the target service configuration parameters for the user, it can be regarded as a configuration method for personalized service configuration parameters for the user. Further, the method is a personalized recommendation method of the service configuration parameters based on the causal inference theory. Based on the scheme, not only can the personalized target service configuration parameters of different users be customized in a high-efficiency manner, but also the optimization of service indexes, such as the maximization of service index gain quantity, can be realized on the premise that the overall calculation power consumption is effectively restricted by a service calculation power restriction condition.

According to the technical scheme, based on the target threshold value, the service index gain amount and the target and calculation force constraint threshold value, the optimal service configuration parameters can be automatically allocated to each request for searching, recommending and the like in the Internet.

It can be understood that the target model of the present disclosure is a gain model based on causal inference, and benefits from a causal inference method, and the training process, the selection process of the target threshold, the determination process of the target recommendation length, and the like of the present disclosure can be white-boxed, and compared with a black-box method in which a deep learning model is adopted for processing in the related art, the present disclosure has the advantages of strong explanatory performance, transparent process, and benefit for transplantation to be used in various related fields.

An embodiment of the present disclosure provides an apparatus for obtaining a service configuration parameter, as shown in fig. 7, the apparatus includes:

a first obtaining unit 701, configured to obtain traffic request data;

a second obtaining unit 702, configured to input feature data of the traffic request data into at least one group of target models to obtain output parameters of each group of target models; the output parameters of each group of target models represent service index gain generated by changing the service processing of the flow request data from the original service configuration parameters corresponding to each group of target models into the candidate service configuration parameters corresponding to each group of target models;

a determining unit 703, configured to determine, based on the service index gain of each group of target models, a target service configuration parameter from the original service configuration parameter and the candidate service configuration parameter corresponding to each group of target models, where the target service configuration parameter is used to perform service processing on an object generating the traffic request data.

In an alternative, the determining unit 703 is configured to: determining target business index gain amount from the business index gain amount output by each group of target models; the target service index gain is at least one of the service index gain with the largest value and reaching a target threshold value in the service index gains output by each group of target models; and determining target service configuration parameters from the original service configuration parameters and the candidate service configuration parameters corresponding to each group of target models based on the target service index gain.

In an alternative, the target service configuration parameter is one of the following: obtaining candidate service configuration parameters corresponding to the target model with the service index gain with the largest value from the service index gain output by each group of target models; obtaining candidate service configuration parameters corresponding to the target model of the service index gain reaching the target threshold value from the service index gain output by each group of target models; and obtaining candidate service configuration parameters corresponding to the target model of the service index gain quantity which has the largest value and reaches the target threshold value from the service index gain quantities output by each group of target models.

In an alternative, the determining unit 703 is configured to: and in response to that the service index gain output by each group of target models does not reach the target threshold value, using the same original service configuration parameters as the target service configuration parameters.

In an alternative, the target threshold is obtained by screening at least one preset candidate threshold based on a historical traffic indicator gain obtained by the at least one set of target model for historical traffic request data, a calculated power consumption amount under a historical traffic configuration parameter corresponding to the historical traffic indicator gain, and a preset traffic calculated power constraint condition.

In an alternative, the candidate service configuration parameter is a candidate recommended length, and the original service configuration parameter is an original recommended length; the service index gain is the service gain generated by the target models in each group from performing service search or service recommendation on the flow request data by adopting the original recommended length corresponding to the target models in each group to performing service search or service recommendation by adopting the candidate recommended length corresponding to the target models in each group.

In an alternative, the determining unit 703 is configured to: and determining the target recommendation length from the original recommendation length and the candidate recommendation length corresponding to each group of target models based on the service income of each group of target models.

In one alternative, the at least one set of target models is obtained by training each set of causal inference based gain models through historical traffic request data and tag data acquired for the set of causal inference based gain models.

The functions of each component in the apparatus for obtaining service configuration parameters in the embodiment of the present disclosure may refer to the description of the method for obtaining service configuration parameters, and are not described herein again. In the apparatus for obtaining service configuration parameters according to the embodiment of the present disclosure, since the principle of solving the problem is similar to the method for obtaining service configuration parameters, the implementation process and the implementation principle of the apparatus for obtaining service configuration parameters can be described with reference to the implementation process and the implementation principle of the foregoing related method, and repeated parts are not described again.

According to an embodiment of the present disclosure, there is also provided an electronic device comprising at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the aforementioned method for obtaining service configuration parameters.

The description of the processor and the memory of the electronic device can refer to the related descriptions of the computing unit 801 and the storage unit 808 in fig. 8.

According to an embodiment of the present disclosure, there is also provided a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the aforementioned traffic control method, training method of traffic control model. For a description of the computer-readable storage medium, reference is made to the description in fig. 8.

According to an embodiment of the present disclosure, the present disclosure also provides a computer program product comprising a computer program which, when executed by a processor, enables the obtaining of the aforementioned service configuration parameters. For an explanation of the computer program product, reference is made to the description in connection with fig. 8.

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 good customs of the public order.

Fig. 8 is a block diagram of an electronic device for implementing an apparatus for obtaining service configuration parameters according to an embodiment 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. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular telephones, 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. 8, the electronic apparatus 800 includes a computing unit 801 that can perform various appropriate actions and processes according to a computer program stored in a ROM802 or a computer program loaded from a storage unit 808 into a RAM 803. In the RAM803, various programs and data necessary for the operation of the electronic apparatus 800 can also be stored. The calculation unit 801, the ROM802, and the RAM803 are connected to each other by a bus 804. An input/output (I/O) interface 805 is also connected to bus 804.

A number of components in the electronic device 800 are connected to the I/O interface 805, including: an input unit 806, such as a keyboard, a mouse, or the like; an output unit 807 such as various types of displays, speakers, and the like; a storage unit 808, such as any device that can be used as a memory, e.g., a magnetic disk, an optical disk, etc.; and a communication unit 809 such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows the electronic device 800 to exchange information/data with other devices through a computer network such as the internet and/or various telecommunication networks.

The storage unit 808 in the embodiments of the present disclosure may be embodied as at least one of 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 CD-ROM, an optical storage device, and a magnetic storage device.

Computing unit 801 may be a variety of general and/or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a CPU, a Graphics Processing Unit (GPU), an Artificial Intelligence (AI) computing chip, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, or the like. The calculation unit 801 performs the various methods and processes described above, such as the obtaining method of the service configuration parameters. For example, in some embodiments, the method of obtaining service configuration parameters may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program may be loaded and/or installed onto the electronic device 800 via the ROM802 and/or the communication unit 809. When loaded into RAM803 and executed by the computing unit 801, the computer program may perform one or more steps of the above described method of obtaining service configuration parameters. Alternatively, in other embodiments, the computing unit 801 may be configured to perform the obtaining method of the service configuration parameters in any other suitable manner (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, editable arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), system On Chip (SOCs), load 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 codes may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, or the like, that may execute the computer program codes, such that the program codes, when executed by the processor or controller, cause the functions/acts specified in the flowchart and/or block diagram block or blocks 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 (storage 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 RAM, a ROM, an EPROM or flash memory, an optical fiber, a 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); 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 component, middleware component, or front-end component.

It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present disclosure may be executed in parallel, 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 in accordance with 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 method for obtaining service configuration parameters includes:

acquiring flow request data;

inputting the characteristic data of the flow request data into at least one group of target models to obtain the output parameters of each group of target models; the output parameters of each group of target models represent service index gain generated by changing the service processing of the flow request data from the original service configuration parameters corresponding to each group of target models into the candidate service configuration parameters corresponding to each group of target models;

and determining a target service configuration parameter from the original service configuration parameter and the candidate service configuration parameter corresponding to each group of target models based on the service index gain of each group of target models, wherein the target service configuration parameter is used for performing service processing on the object generating the flow request data.

2. The method of claim 1, wherein the determining a target service configuration parameter from the original service configuration parameter and the candidate service configuration parameter corresponding to each set of target models based on the service index gain of each set of target models comprises:

determining target service index gain amount from the service index gain amount output by each group of target models; the target service index gain is at least one of the service index gain with the largest value and reaching a target threshold value in the service index gains output by each group of target models;

and determining target service configuration parameters from the original service configuration parameters and the candidate service configuration parameters corresponding to each group of target models based on the target service index gain.

3. The method according to claim 1 or 2, wherein the target traffic configuration parameter is one of:

obtaining candidate service configuration parameters corresponding to the target model with the service index gain with the largest value from the service index gain output by each group of target models;

obtaining candidate service configuration parameters corresponding to the target model of the service index gain reaching the target threshold value from the service index gain output by each group of target models;

and obtaining candidate service configuration parameters corresponding to the target model of the service index gain quantity which has the largest value and reaches the target threshold value from the service index gain quantities output by each group of target models.

4. The method according to claim 3, wherein the original service configuration parameters corresponding to the target models of the respective groups are the same;

and in response to that the service index gain output by each group of target models does not reach the target threshold value, using the same original service configuration parameters as the target service configuration parameters.

5. The method according to claim 2, 3 or 4, wherein the target threshold is obtained by filtering from at least one preset candidate threshold based on historical traffic indicator gain amount obtained by the at least one set of target models for historical traffic request data, calculated power consumption amount under historical traffic configuration parameters corresponding to the historical traffic indicator gain amount, and a preset traffic calculated power constraint condition.

6. The method of claim 1, wherein the candidate service configuration parameter is a candidate recommended length, and the original service configuration parameter is an original recommended length;

the service index gain is the service gain generated by the target models in each group from performing service search or service recommendation on the flow request data by adopting the original recommended length corresponding to the target models in each group to performing service search or service recommendation by adopting the candidate recommended length corresponding to the target models in each group.

7. The method of claim 6, wherein the determining a target service configuration parameter from an original service configuration parameter and a candidate service configuration parameter corresponding to each group of target models based on the service index gain of each group of target models comprises:

and determining the target recommendation length from the original recommendation length and the candidate recommendation length corresponding to each group of target models based on the service income of each group of target models.

8. The method of any of claims 1 to 7, wherein the at least one set of goal models is derived by training each set of causal inference based gain models with historical traffic request data and tag data acquired for the set of causal inference based gain models.

9. An apparatus for obtaining service configuration parameters, comprising:

a first obtaining unit, configured to obtain traffic request data;

10. The apparatus of claim 9, wherein the determining unit is configured to:

11. The apparatus according to claim 9 or 10, wherein the target traffic configuration parameter is one of:

12. The apparatus of claim 11, wherein the determining unit is configured to:

13. The apparatus according to claim 10, 11 or 12, wherein the target threshold is obtained by filtering from at least one preset candidate threshold based on a historical traffic indicator gain amount obtained by the at least one set of target models for historical traffic request data, a calculated power consumption amount under a historical traffic configuration parameter corresponding to the historical traffic indicator gain amount, and a preset traffic calculated power constraint.

14. The apparatus according to claim 9, wherein the candidate service configuration parameter is a candidate recommended length, and the original service configuration parameter is an original recommended length;

15. The apparatus of claim 14, wherein the determining unit is configured to:

16. The apparatus of any of claims 9 to 15, wherein the at least one set of target models is derived by training each set of causal inference based gain models with historical traffic request data and tag data acquired for the set of causal inference based gain models.

17. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

18. A non-transitory computer readable storage medium having stored thereon computer instructions for causing a computer to perform the method of any one of claims 1-8.

19. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1-8.