CN114331499A

CN114331499A - Method and device for determining media information, storage medium and electronic equipment

Info

Publication number: CN114331499A
Application number: CN202111496589.5A
Authority: CN
Inventors: 严超
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-12-08
Filing date: 2021-12-08
Publication date: 2022-04-12

Abstract

The invention discloses a method and a device for determining media information, a storage medium and electronic equipment. Wherein, the method comprises the following steps: acquiring target object characteristics of a target account; determining a media information subset recalled by a target account in the media information set according to the target object characteristics, the estimated click rate of each media information in the media information set, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information; and determining target media information issued to the target account in the media information subset. The invention solves the technical problem of low accuracy of the media information issued to the user.

Description

Method and device for determining media information, storage medium and electronic equipment

Technical Field

The present invention relates to the field of computers, and in particular, to a method and an apparatus for determining media information, a storage medium, and an electronic device.

Background

The collection of advertisements from a large library of advertisements is called an advertisement recall, also called an advertisement search. Several advertisements are retrieved by coarse and fine selection in the recalled advertisement set and distributed to the user.

In traditional targeted advertising, an advertiser usually selects an original target autonomously, and the original target refers to a basic target (such as the gender, age, region and the like of a user) of an advertising audience customized by the advertiser. Advertisers may select a base target for the audience of the ad via the targeting tag. Traditional targeted tag selection relies on a great deal of manual prior knowledge of the advertiser and requires repeated manual tuning by the advertiser. On one hand, manual tuning is difficult, trial and error cost is high, and on the other hand, distribution mutation of advertisement audience population is easily caused in the process of adding and deleting labels for tuning, so that the advertisement issued to the user is not necessarily the advertisement really interested by the user. Making the advertisements published to the user less accurate. In addition, the current traditional targeted advertisement does not consider the revenue of the advertising platform.

In view of the above problems, no effective solution has been proposed.

Disclosure of Invention

The embodiment of the invention provides a method and a device for determining media information, a storage medium and electronic equipment, which are used for at least solving the technical problem of low accuracy of media information issued to a user.

According to an aspect of an embodiment of the present invention, there is provided a method for determining media information, including: acquiring target object characteristics of a target account; determining a media information subset recalled by the target account in the media information set according to the target object characteristics, the estimated click rate of each media information in the media information set, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information, wherein, the media information set comprises the media information to be issued, the estimated click rate of each media information represents the probability that each media information is clicked by the target account when each media information is issued to the target account, the pre-estimated conversion rate of each media message represents the probability that the target event corresponding to each media message is executed by the target account when each media message is issued to the target account, the pre-estimated resource transfer quantity of each media information represents the quantity of resources triggered to be transferred when each media information is issued to the target account; and determining target media information issued to the target account in the media information subset.

Optionally, the determining, according to the target object feature, the estimated click rate of each piece of media information in the media information set, the estimated conversion rate of each piece of media information, and the estimated resource transfer quantity of each piece of media information, a subset of media information to be issued to the target account in the media information set includes: inputting the target object features and the media information features of each media information into a target multitask model; respectively determining the estimated click rate of each media information, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information according to the target object characteristics and the media information characteristics of each media information through the target multitask model, and determining the value of the recall parameter of each media information according to the estimated click rate of each media information, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information; and determining the media information subset recalled for the target account in the media information set according to the recall parameter value of each piece of media information.

Optionally, the method further comprises: acquiring a sample object feature set corresponding to a sample account set, a first media information feature set corresponding to a first media information set issued to the sample account set, an actual click rate set corresponding to the first media information set, an actual conversion rate set corresponding to the first media information set, and an actual resource transfer quantity set corresponding to the first media information set; training a multi-task model to be trained by using the sample object feature set and the first media information feature set until a target loss value corresponding to the multi-task model to be trained meets a preset loss condition, and ending the training to obtain the target multi-task model, wherein the target loss value is a loss value determined according to a click rate loss value, a conversion rate loss value and a resource transfer quantity loss value; the click rate loss value represents a loss value between an estimated click rate and an actual click rate corresponding to the actual click rate set, and the estimated click rate is determined by the multitask model to be trained according to sample object features corresponding to the sample object feature set and sample media information features corresponding to the first media information feature set; the conversion rate loss value represents a loss value between a pre-estimated conversion rate and a corresponding actual conversion rate in the actual conversion rate set, and the pre-estimated conversion rate is determined by the multitask model to be trained according to corresponding sample object features in the sample object feature set and corresponding sample media information features in the first media information feature set; the resource transfer quantity loss value represents a loss value between a pre-estimated resource transfer quantity and a corresponding actual resource transfer quantity in the actual resource transfer quantity rate set, and the pre-estimated resource transfer quantity is determined by the multitask model to be trained according to a corresponding sample object feature in the sample object feature set and a corresponding sample media information feature in the first media information feature set.

Optionally, the training a multi-task model to be trained by using the sample object feature set and the first media information feature set until a target loss value corresponding to the multi-task model to be trained satisfies a preset loss condition, and ending the training to obtain the target multi-task model includes: acquiring a current sample object characteristic of a current sample account in the sample object characteristic set, acquiring a current media information characteristic of current media information issued to the current sample account in the first media information set characteristic, acquiring a current actual click rate corresponding to the current sample account and the current media information in the actual click rate set, acquiring a current actual conversion rate corresponding to the current sample account and the current media information in the actual conversion rate set, and acquiring a current actual resource transfer quantity corresponding to the current sample account and the current media information in the actual resource transfer quantity set; inputting the current sample object characteristics and the current media information characteristics into the multi-task model to be trained, and respectively determining the corresponding current pre-estimated click rate, current pre-estimated conversion rate and current pre-estimated resource transfer quantity according to the current sample object characteristics and the current media information characteristics through the multi-task model to be trained; determining a current click rate loss value according to the current estimated click rate and the current actual click rate, determining a current conversion rate loss value according to the current estimated conversion rate and the current actual conversion rate, and determining a current resource transfer quantity loss value according to the current estimated resource transfer quantity and the current actual resource transfer quantity; determining a current loss value according to the weighted sum of the current click rate loss value, the current conversion rate loss value and the current resource transfer quantity loss value; when the current loss value meets the preset loss condition, ending training, and determining the multi-task model to be trained as the target multi-task model when training is ended; and when the current loss value does not meet the preset loss condition, adjusting parameters in the multi-task model to be trained.

Optionally, the determining, by the multitask model to be trained, a corresponding current pre-estimated click rate, a current pre-estimated conversion rate, and a current pre-estimated resource transfer quantity according to the current sample object feature and the current media information feature respectively includes: determining the current estimated click rate according to the current sample object characteristics and the current media information characteristics through a click rate double-tower structure in the multi-task model to be trained; determining the current pre-estimated conversion rate according to the current sample object characteristics and the current media information characteristics through a conversion rate double-tower structure in the multi-task model to be trained; and determining the current estimated resource transfer quantity according to the current sample object characteristic and the current media information characteristic through a resource transfer quantity double-tower structure in the multi-task model to be trained.

Optionally, the determining a current loss value according to the weighted sum of the current click rate loss value, the current conversion rate loss value, and the current resource transfer quantity loss value includes: determining the weighted sum of the current click rate loss value, the current conversion rate loss value and the current resource transfer quantity loss value as the current loss value; or when the current conversion rate loss value comprises a shallow conversion rate loss value and a deep conversion rate loss value, determining the current click rate loss value, the shallow conversion rate loss value, the deep conversion rate loss value and the weighted sum of the current resource transfer quantity loss value as the current loss value.

Optionally, the determining, by the multitask model to be trained, a corresponding current pre-estimated click rate, a current pre-estimated conversion rate, and a current pre-estimated resource transfer quantity according to the current sample object feature and the current media information feature respectively includes: determining the current estimated click rate according to the current sample object characteristics and the current media information characteristics through a click rate double-tower structure in the multi-task model to be trained; when the current conversion rate loss value comprises a shallow conversion rate loss value and a deep conversion rate loss value, determining the shallow pre-estimated conversion rate according to the current sample object characteristic and the current media information characteristic through a shallow conversion rate double-tower structure in the multi-task model to be trained, and determining the deep pre-estimated conversion rate according to the current sample object characteristic and the current media information characteristic through a deep conversion rate double-tower structure in the multi-task model to be trained; and determining the current estimated resource transfer quantity according to the current sample object characteristic and the current media information characteristic through a resource transfer quantity double-tower structure in the multi-task model to be trained.

Optionally, the determining, by the resource transfer amount double tower structure in the multitask model to be trained, the current pre-estimated resource transfer amount according to the current sample object feature and the current media information feature includes: generating an object vector according to the current sample object characteristics through an object tower structure; generating a media information vector according to the current media information characteristics through a media information tower structure, wherein the resource transfer quantity double-tower structure comprises the object tower structure and the media information tower structure; and determining the current pre-estimated resource transfer quantity as the dot product of the object vector and the media information vector.

Optionally, the determining, according to the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information, and the estimated resource transfer quantity of each piece of media information, a value of a recall parameter of each piece of media information includes: determining the value of the recall parameter of each piece of media information as the weighted sum of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information and the estimated resource transfer quantity of each piece of media information; or determining the value of the recall parameter of each piece of media information to be equal to the product of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information and the estimated resource transfer quantity of each piece of media information.

Optionally, the determining the value of the recall parameter of each piece of media information to be equal to a weighted sum of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information, and the estimated resource transfer quantity of each piece of media information includes: acquiring a click rate weight value, a conversion rate weight value and a resource transfer quantity weight value corresponding to each piece of media information, wherein the resource transfer quantity weight value is greater than the click rate weight value and the conversion rate weight value, or the resource transfer quantity weight value is greater than or equal to a preset weight value; and weighting and summing the estimated click rate of each media message, the estimated conversion rate of each media message and the estimated resource transfer quantity of each media message with a click rate weight value, a conversion rate weight value and a resource transfer quantity weight value corresponding to each media message respectively to obtain a value of a recall parameter of each media message.

Optionally, the determining, according to the value of the recall parameter of each piece of media information, the subset of media information recalled from the target account in the set of media information includes: and according to the arrangement of the values of the recall parameters of each piece of media information from large to small, determining the media information with the value of the recall parameter ranked N before as the subset of the media information in the media information set, wherein N is a positive integer greater than or equal to 1.

Optionally, determining, in the media information subset, target media information published to the target account includes: acquiring a first object characteristic of the target account, and acquiring a first media information characteristic of each media information in the media information subset; determining a first estimated click rate set corresponding to the media information subset according to the first object characteristics and the first media information characteristics of each piece of media information in the media information subset through a first click rate estimation model; determining a first pre-estimated conversion rate set corresponding to the media information subset according to the first object characteristic and the first media information characteristic of each media information in the media information subset through a first conversion rate pre-estimated model; determining a rough-arranged media information subset in the media information subset according to the first pre-estimated click rate set, the first pre-estimated conversion rate set and a current resource transfer quantity set corresponding to the media information subset; acquiring a second object characteristic of the target account, and acquiring a second media information characteristic of each media information in the rough media information subset, wherein the number of characteristics in the first object characteristic is less than that in the second object characteristic, and the number of characteristics in the first media information characteristic is less than that in the second media information characteristic; determining a second estimated click rate set corresponding to the roughly arranged media information subset according to the second object characteristics and second media information characteristics of each piece of media information in the roughly arranged media information subset through a second click rate estimation model; determining a second pre-estimated conversion rate set corresponding to the roughly arranged media information subset according to the second object characteristics and the second media information characteristics of each piece of media information in the roughly arranged media information subset through a second conversion rate pre-estimated model; and determining the target media information in the rough arranged media information subset according to the second pre-estimated click rate set, the second pre-estimated conversion rate set and the current resource transfer quantity set corresponding to the rough arranged media information subset.

Optionally, the determining, according to the first pre-estimated click rate set, the first pre-estimated conversion rate set, and the current resource transfer quantity set corresponding to the media information subset, a coarsely arranged media information subset in the media information subset includes: correspondingly multiplying the estimated click rate in the first estimated click rate set, the estimated conversion rate in the first estimated conversion rate set and the current resource transfer quantity in the current resource transfer quantity set to obtain a value of a first screening parameter corresponding to the media information in the media information subset; and according to the rank of the values of the first screening parameters from high to low, determining the top M bits of media information in the media information subsets as the coarse-ranking media information subsets, wherein M is a positive integer greater than or equal to 2.

Optionally, the determining the target media information in the rough-ranking media information subset according to the second pre-estimated click rate set, the second pre-estimated conversion rate set, and the current resource transfer quantity set corresponding to the rough-ranking media information subset includes: correspondingly multiplying the estimated click rate in the second estimated click rate set, the estimated conversion rate in the second estimated conversion rate set and the current resource transfer quantity in the current resource transfer quantity set corresponding to the rough media information subset to obtain the value of a second screening parameter corresponding to the media information in the rough media information subset; and according to the rank of the values of the second screening parameters from high to low, determining the P-bit media information before ranking in the rough media information subset as the target media information, wherein P is a positive integer greater than or equal to 1.

According to another aspect of the embodiments of the present invention, there is also provided a device for determining media information, including: the acquisition module is used for acquiring the target object characteristics of the target account; a first determining module, configured to determine, according to the target object feature, an estimated click rate of each media information in a media information set, an estimated conversion rate of each media information, and an estimated resource transfer quantity of each media information, a media information subset recalled by the target account in the media information set, where the media information set includes media information to be published, the estimated click rate of each media information indicates a probability that each media information is clicked by the target account when the media information is published to the target account, the estimated conversion rate of each media information indicates a probability that a target event corresponding to each media information is executed by the target account when the media information is published to the target account, and the estimated resource transfer quantity of each media information indicates a quantity of resources triggered to be transferred when each media information is published to the target account An amount; and the second determining module is used for determining the target media information issued to the target account in the media information subset.

According to still another aspect of the embodiments of the present invention, there is also provided a computer-readable storage medium having a computer program stored therein, wherein the computer program is configured to execute the above-mentioned method for determining media information when running.

According to yet another aspect of embodiments herein, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the method of determining media information as above.

According to still another aspect of the embodiments of the present invention, there is also provided an electronic device, including a memory and a processor, where the memory stores therein a computer program, and the processor is configured to execute the above-mentioned method for determining media information by the computer program.

In the embodiment of the invention, according to the target object characteristics of the target account, the pre-estimated click rate, the pre-estimated conversion rate and the pre-estimated resource transfer quantity of each piece of media information in the media information set are combined, a media information subset recalled by the target account is determined in the media information set, and the target media information issued to the target account is determined in the media information subset. In the embodiment of the invention, the estimated click rate, the estimated conversion rate and the estimated resource transfer quantity of the media information are comprehensively considered to push the media information to the target account. The media information issued to the target account can be determined in the media information set through the computer, prior knowledge of an advertiser is not needed, the purpose of improving the accuracy of the media information issued to the user is achieved, and the technical problem that the accuracy of the media information issued to the user is low is solved. In addition, in the embodiment of the invention, the predicted resource transfer quantity of the media information is combined, and the benefit of the media information to the publishing platform is considered. The three-party win-win situation of the user, the advertiser and the publishing platform is achieved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the invention without limiting the invention. In the drawings:

fig. 1 is a schematic diagram of an application environment of an alternative media information determination method according to an embodiment of the present invention;

FIG. 2 is a flow chart of an alternative method of determining media information according to an embodiment of the invention;

FIG. 3 is a schematic diagram of an alternative advertising system architecture according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of an alternative automated diffusion process according to an embodiment of the present invention;

FIG. 5 is an alternative system preferred schematic according to an embodiment of the invention;

FIG. 6 is a schematic diagram of an alternative smart ad page in accordance with an embodiment of the present invention;

FIG. 7 is a schematic diagram of an alternative smart ad page according to an embodiment of the present invention;

FIG. 8 is a schematic view of yet another alternative smart ad page in accordance with an embodiment of the present invention;

FIG. 9 is a schematic view of an alternative smart ad page according to an embodiment of the present invention;

FIG. 10 is a block diagram of an alternative target multitasking model according to an embodiment of the present invention;

FIG. 11 is a schematic diagram of an alternative MMoE architecture in accordance with embodiments of the present invention;

FIG. 12 is a schematic illustration of an alternative double tower configuration in accordance with an embodiment of the present invention;

FIG. 13 is a schematic illustration of an alternative double column configuration according to an embodiment of the present invention;

FIG. 14 is a block diagram of an alternative multitasking model according to an embodiment of the present invention;

FIG. 15 is an alternative overall frame diagram in accordance with an embodiment of the present invention;

FIG. 16 is a diagram of an alternative retrieval architecture according to an embodiment of the present invention;

fig. 17 is a schematic structural diagram of an alternative media information determining apparatus according to an embodiment of the present invention;

FIG. 18 is a schematic diagram of an alternative electronic device according to an embodiment of the invention;

fig. 19 schematically shows a structural block diagram of a computer system of an electronic device for implementing the embodiment of the present application.

Detailed Description

In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

Alternatively, the following explains key terms involved in the embodiments of the present invention:

subscribing: the advertisement playing system subscribes the information related to the advertisement from the service database of the launching end to become a source for playing the advertisement set on line.

Intelligent targeted advertising: the advertisement system automatically finds crowd targeting suitable for advertisements through model strategies.

The system is preferably as follows: in one type of intelligent targeted advertisement, an advertisement system searches for accurate crowds under wide targeting set by an advertiser through a model strategy.

Automatic amplification: in one type of intelligent targeted advertising, the advertising system seeks more and better people through model strategies under the narrow targeting configured by the advertiser.

Recall/retrieve: in the advertisement playing system, an advertisement set meeting the current interest of a user is selected from a massive advertisement library, which is called as a recall and is also called as a retrieval.

Coarse draining: in the advertisement playing system, a process of preliminarily sorting a large number of recalled advertisements and selecting N top-ranked advertisements (N can be arbitrarily set, such as 100, 120, 150, etc.).

Fine discharging: in the advertisement playing system, the process of precisely ordering and selecting the top M (M < N, such as 2, 3, 5, etc.) advertisements from the N advertisements selected in the rough arrangement is performed.

Full advertisement: the total set of existing advertisements.

Incremental advertisement: and adding a new advertisement set in real time.

LiteCTR: the click rate estimation model is used in the rough ranking of the advertising system, and the used characteristics are less than those in the fine ranking due to the performance consideration.

LiteCVR: the conversion rate estimation model is used in the rough ranking of the advertising system, and the used characteristics are less than those of the fine ranking due to the performance consideration.

pCTR: the click rate accurate estimation model is used in fine scheduling of an advertisement system, the use characteristics are more than those of rough scheduling, and the estimation accuracy is higher.

pCVR: the conversion rate accurate estimation model is used in the fine scheduling of the advertisement system, the use characteristics are more than those of the rough scheduling, and the estimation precision is higher.

oCPA (optimized click through): the oCPA advertisement is an advertisement form adopting a new bidding mode, and specifically means that an advertiser sets an expected cost price target _ cpa for a certain specific conversion behavior, and the platform is responsible for controlling the bidding of each exposure so as to achieve that the average cost of each conversion is within 1.2 of the expected cost price target _ cpa of the advertiser.

The oCPA bid is as follows: the advertiser expects a cost price, target _ cpa, for each particular target conversion.

Cost: the average consumption per conversion of the oCPA ad is the cost for the conversion target set by the advertiser.

Cost run-away: the cost of the oCPA ad is not between 0.8-1.2 times the advertiser's expected cost price, target _ cpa.

The following steps are achieved: if the conversion cost of the oCPA advertisement meets 0.8-1.2 times of the expected cost price of the advertiser, target _ cpa, the conversion cost is called the achievement of the advertisement.

The achievement rate is as follows: in all the cpa advertisement sets, the rate of advertisement achievement becomes the achievement rate.

Lambda: the pricing factor is used for enabling the oCPA advertisement to be achieved in the fine-ranking of the advertisement system, the advertisement system multiplies lambda by the advertisement bid, and therefore the bid is guaranteed to be reduced through lambda adjustment when the advertisement cost is higher; bids are raised through lambda adjustment when advertising costs are low.

Double-target advertisement: the oCPA advertisement is provided with two optimization targets, including a shallow optimization target and a deep optimization target. For example, the oCPA dual-target advertisement sets a shallow target to 'download' and a deep target to 'order' at the same time.

And (4) new advertisement: after the ad is created, it is on the first impression and 3 natural days thereafter.

The weighing is successful: and after the new advertisement is put, the total number of the complete 3 advertisements with the conversion number of more than 30 in the natural day is the success of the starting amount.

The volume percentage: the percentage of successful ad campaigns in a new ad set is called the campaign rate.

eCPM: effective Cost per Mille (total deduction per thousand exposures), an index of bid ranking in an advertisement system, and advertisement with high eCPM mean that more revenue can be brought by an advertisement platform and exposure can be obtained preferentially.

GMV: gross Merchandisc Volume (commodity transaction total), an index for measuring the advertiser transaction total in an advertisement system, a calculation formula is the product of the number of conversions of advertisement putting of the advertiser and conversion bids, and high GMV means that the value of the advertisement platform for the advertiser is higher.

Commercial value: the eCPM index concerned by the advertising platform is used as a measure of the commercial value through eCPM. The higher the eCPM, the higher the commercial value for the ad platform.

Sample selection bias: and (5) the model training space is inconsistent with the prediction space. For example, if the recalled prediction space is the whole advertisement library, if the training samples are only from the exposed advertisements, the problem that the training space and the prediction space are inconsistent exists, namely the problem of 'sample selection bias'.

And (3) ANN: approximate Nearest Neighbors (similar Nearest neighbor search algorithms), a class of algorithms for quickly finding the Nearest neighbor vector of a certain query vector in a certain vector set. The method is commonly used in the recalling links of advertisement, search and recommendation and is used for improving the retrieval efficiency.

According to an aspect of the embodiments of the present invention, a method for determining media information is provided, and optionally, as an optional implementation manner, the method for determining media information may be applied, but not limited to, in an application environment as shown in fig. 1. The application environment includes a user device 102, a network 110, and a server 112, wherein the user device 102 includes a memory 104, a processor 106, and a display 108, wherein the memory is used for storing data, including but not limited to the target object characteristics. The processor is used for processing data, including but not limited to processing the request data of the target account. The display includes, but is not limited to, a display for displaying the target media information issued to the target account.

Optionally, in this embodiment, the user equipment may be a terminal device configured with a target client, and may include, but is not limited to, at least one of the following: mobile phones (such as Android phones, iOS phones, etc.), notebook computers, tablet computers, palm computers, MID (Mobile Internet Devices), PAD, desktop computers, smart televisions, etc. The target client may be a video client, an instant messaging client, a browser client, a game client, an audio client, etc., on which advertisements may be published to users.

Optionally, the network 110 may include, but is not limited to: a wired network, a wireless network, wherein the wired network comprises: a local area network, a metropolitan area network, and a wide area network, the wireless network comprising: bluetooth, WIFI, and other networks that enable wireless communication.

The server 112 includes a database 114 and a processing engine 116, the database 114 being used to store data including, but not limited to, advertisements placed by advertisers, representations of users, and the like. The processing engine is used for processing data, and includes but is not limited to executing the following steps:

step S102, determining a media information subset recalled by the target account in the media information set according to the target object characteristics, the estimated click rate of each media information in the media information set, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information;

step S104, determining the target media information issued to the target account in the media information subset.

The server may be a single server, a server cluster composed of a plurality of servers, or a cloud server. The above is merely an example, and this is not limited in this embodiment.

Optionally, as an optional implementation manner, as shown in fig. 2, the method for determining media information includes:

step S202, acquiring target object characteristics of a target account;

the target account is an account of a target client for a user to log in, and the target object features include but are not limited to a basic portrait feature of a user side, a historical behavior statistical feature of the user side, a behavior sequence feature of the user side, a behavior interest mining feature of the user side and the like.

User-side base portrait features include, but are not limited to: the user's age, gender, province, occupation, consumption status, love and marriage status, academic calendar, etc.

The user-side historical behavior statistical characteristics include, but are not limited to: the advertisement click number of the user in the last month, the number of click advertisement detail pages in the last month, the video click attention number in the last month, the click uninteresting number in the last month and the like, and the average exposure advertisement times in the last month; and statistics for up to three months and up to 6 months.

User-side behavioral sequence features include, but are not limited to: user recent exposure, clicks, converted advertisements, client, etc.

User-side behavioral interest mining features include, but are not limited to: and (4) mining the label characteristics of long and short term categories, keywords, interests and the like from the original behavior sequence of the user.

Step S204, according to the target object characteristics, the estimated click rate of each media information in the media information set, the estimated conversion rate of each media information, and the estimated resource transfer quantity of each media information, determining the media information subset recalled from the target account in the media information set, wherein, the media information set comprises the media information to be issued, the estimated click rate of each media information represents the probability that each media information is clicked by the target account when each media information is issued to the target account, the pre-estimated conversion rate of each media message represents the probability that the target event corresponding to each media message is executed by the target account when each media message is issued to the target account, the pre-estimated resource transfer quantity of each media information represents the quantity of resources triggered to be transferred when each media information is issued to the target account;

the media information may be an advertisement. The set of media information may be a full volume advertisement, including advertisements that advertisers publish on the platform. The estimated click rate represents the estimated probability of being clicked by the target account under the condition that the advertisement in the media information set is issued to the target account. The target event corresponding to the media information may refer to an event such as purchasing a commodity in the advertisement or downloading a game in the advertisement. The estimated conversion rate represents the estimated probability of the target account executing the target event when the advertisement is issued to the target account. For example, the probability of the target account purchasing an item in the advertisement, the probability of the target account downloading a game in the advertisement. The predicted resource transfer amount may represent the revenue of the advertising platform after the platform issues the advertisement to the target account, including but not limited to the fee paid by the advertiser to the platform. The advertisements in the subset of media information are those recalled from the full size of advertisements (tens of thousands or millions of advertisements) to the target account (thousands of advertisements).

Step S206, determining the target media information issued to the target account in the media information subset.

Several advertisements (for example, 2 or 3 advertisements) can be selected from the media information subsets (thousands of advertisement numbers) recalled to the target account through rough ranking and fine ranking as the advertisements issued to the target account, and the advertisements can be issued on the target client logged in by the target account.

Alternatively, in the advertisement system shown in fig. 3, the advertisement with top rank (for example, top N, N may be arbitrarily set, for example, 1, 2, etc.) is selected to be presented to the user through offline ranking to determine the high-quality advertisement (referred to as online index advertisement) capable of entering the index through the full amount of included advertisements, then through directional matching or model recall, retrieving the advertisement that can be recalled by the user according to the current traffic, then through rough ranking, multi-way merging ranking (merging and ranking the advertisements in the multi-way rough ranking), and final fine ranking.

Optionally, a brief description of intelligent targeted advertising follows:

in traditional ad targeting, advertisers select user audience targeting of ads based on prior knowledge, providing corresponding crowd packs. The intelligent targeted advertisement allows an advertiser to only provide basic targeting (such as gender, age, region and the like) of an advertisement audience, and in addition, the advertiser is given to an advertisement system to automatically select a user group meeting the basic targeting through a model strategy, and only when the model considers that a high-quality user request comes, the advertisement corresponding to the advertiser is recalled for bidding and striving for an advertisement exposure opportunity.

There are two forms of intelligent targeted advertising: automatic amplification and system optimization.

1) Automatic amplification: it is usually used with a precise narrow targeting, which refers to the original narrow targeting selected by the advertiser itself, such as A and B and C and D in FIG. 4 (A/B/C/D refers to the conventional label targeting condition described above). When the advertiser starts the automatic expansion, the non-breakthrough part in the original orientation can be set, namely the non-breakthrough orientation, if the target is A and B, the expansion strategy crowd is replaced by E, and finally the advertisement target crowd is overlapped with A and B and E on the basis of the original orientation A and B and C and D, so that the expansion effect is achieved.

2) The system is preferably used together with the wide-direction advertising, wherein the wide-direction advertising refers to the original wide-direction advertising selected by the advertiser, such as A and B in FIG. 5, the preferred strategy crowd is replaced by F, and the final advertising targeted crowd is A and B and F, so that the effect of preferred volume adjustment is achieved.

For the product form of the intelligent targeted advertisement, the following processes can be included:

1) creating intelligent targeted advertisements: in the ad format of the ad publishing platform that the advertiser logs in, there is a "smart ad" button as shown in FIG. 6, and if "on," the current ad is set as a smart targeted ad.

2) Setting of intelligent directional advertisement: in terms of intelligent targeting, unlike general advertising, system preferences under the intent of behavioral interest may be additionally selected. In the newly created targeting page shown in FIG. 7, the advertiser can custom select a base targeting, including geographic location, age, gender, etc. as shown in the figure. The behavioral interest intention can be customized and system preferences can be selected. Or turn on "auto-augmentation" while an orientation setting that is not breachable may be selected.

3) The setting of the intelligent targeted advertisement on the position and the schedule can be various clients, such as a news client, an instant messaging client, a video client, an audio client and the like shown in fig. 8. The schedule setting may select a long-term impression within the impression days shown in fig. 9 or select a designated impression start date and end date.

4) Bidding of intelligent targeted advertisements: the intelligent targeted advertisement can be an oCPA advertisement, so that the advertisement button of oCPC (advertiser bids according to conversion and platform charges according to click) or oCPM (advertiser bids according to conversion and platform charges according to exposure) in a bid mode is selected, the optimization target and the bid can be seen to be filled by the advertiser, and the advertiser fills the optimization target and the bid, so that the bid setting is completed.

5) Intelligent targeted advertising effectiveness data feedback page: after the advertiser successfully puts the intelligent targeted advertisement, the feedback data of the advertisement can be seen through the advertisement effect monitoring page, and the dimension comprises various aspects which the advertiser wants to know, such as exposure, click, conversion, click rate, consumption and the like. One of the more important indicators is the average conversion cost, i.e., the cost for the conversion target set by the advertiser, so that the advertiser can monitor whether the cost is as expected. For example, if the conversion target set by the advertiser is form reservation, the corresponding amount of form reservation and cost of form reservation are of great concern to the advertiser.

The intelligent advertisement is a crowd who is suitable for the advertisement through the automatic searching of advertisement system, and the advantage that exists has: 1) the efficiency of the advertiser on the directional setting is improved, trial and error tuning is not needed to be performed repeatedly in the direction, and most of the complicated process of directional selection is handed to the system. 2) The advertisement platform can find out a plurality of potential high-quality crowds through a model strategy, and the advertisement effect of an advertiser is improved. 3) Because the advertisement platform searches for the crowd through the model, the mutation of adding and deleting labels does not exist, the change of the crowd orientation is more gentle compared with the change of the crowd orientation, and the estimated bias fluctuation of a downstream sequencing model can be reduced. The accuracy of advertisement recalling is improved.

As an alternative embodiment, as shown in the target multitask model diagram shown in fig. 10, the user & context input module in the model structure is the user and context feature part in the "feature module". For inputting user characteristics (target object characteristics). The ad input module in the model structure is the advertisement characteristic part in the characteristic module. For entering advertising characteristics (media information characteristics).

The Expert & Gate module in the object multitasking model structure, including (Expert 0-3, CTR Gate, CVR Gate, Deep CVR Gate, ECPM Gate module), may be a MMoE (Multi-Gate texture-of-Experts) structure. FIG. 11 is a schematic diagram of the MMoE structure, in which the Expert module can adopt a DNN (full link network) structure. The Gate structure adopts a softmax structure.

The structures of a CTR double tower, a CVR double tower and a Deep CVR double tower in the target multitask model are shown in figure 12, the inner parts of the double towers adopt multilayer DNN, an activating function adopts ReLU, and finally the final estimated click probability or estimated conversion probability is obtained through a sigmoid output layer through dot product.

The ECPM double tower structure in the objective multi-tasking model is shown in fig. 13, and different from the structure adopting multiple layers of FC + ReLu in the CTR, CVR and DeepCVR double towers, a bn (batch normalization) structure is added after the ReLu structure of each layer in the ECPM double tower to improve the convergence rate of the model. Meanwhile, in a model output layer, the estimated resource transfer quantity ECPM is obtained only by dot product of user embedding and ad embedding, and a sigmoid output layer is not needed.

The media information features of the media information include, but are not limited to: contextual side features, ad side features. The contextual side features include, but are not limited to, ad spot information (e.g., ad spot identification, ad spot material specifications, etc.), device information (device operating system, device networking type), ad spot context information, and the like. Ad-side features include, but are not limited to: advertisement identification, creative identification, merchandise identification, advertiser identification, advertisement category, creative content keywords, advertisement keywords, and the like.

Determining the estimated click rate of the advertisement according to the target object characteristics (user characteristics) of the target account and the characteristics of each media information (advertisement) in the media information set through the CTR double towers in the target multitask model; determining the pre-estimated conversion rate of the advertisement according to the target object characteristics (user characteristics) of the target account and the characteristics of each media information (advertisement) in the media information set through the CVR double tower and the Deep CVR double tower in the target multitask model; and determining the pre-estimated resource transfer quantity of the media information according to the target object characteristics (user characteristics) of the target account and the characteristics of each media information (advertisement) in the media information set through the ECPM double towers in the target multitask model.

According to the estimated resource transfer quantity, the estimated conversion rate and the estimated click rate of each media information (advertisement) in the media information set, the value of the recall parameter of each advertisement in the media information set can be determined, the value of the recall parameter is used for representing the probability of the advertisement being recalled, and the higher the value of the recall parameter is, the higher the probability of being recalled is. The recall parameter value can be the weighted sum of the estimated click rate output by the CTR double tower, the estimated conversion rate output by the CVR double tower and the Deep CVR double tower and the estimated resource transfer quantity output by the ECPM double tower.

And determining the media information recalled in the target account in the media information set according to the recall parameter value of each media information (advertisement) in the media information set. Specifically, media information with a value of the recall parameter larger than a preset value (which can be set according to actual conditions) may be selected as the recall media information, and top N (which can be set according to actual conditions, for example, 100, 1000, 2000, and the like) media information with a value of the recall parameter ranked as top N may also be selected from the media information set as the recall media information in the media information subset.

As an optional implementation manner, the sample account set is an account sample when the multitask model is trained, and sample accounts in the account sample set are accounts used when different sample users log in the client. The sample object feature set includes features of the sample user, including but not limited to the user portrait features described above. The first set of media information is media information (e.g., advertisements) published to sample accounts in the set of sample accounts.

As an optional implementation manner, the actual click rate in the actual click rate set represents an actual result of whether the sample media information is clicked by the sample account when the sample media information corresponding to the first media information set is issued to the sample account corresponding to the sample account set. For example, the media information a in the first media information set is issued to account 1, and whether account 1 clicked on the media information a or not, if clicked, the actual click rate is 1, and if not clicked, the actual click rate is 0.

The actual conversion rate in the actual conversion rate set represents an actual result of whether a sample event corresponding to the sample media information is executed by the sample account when the sample media information is issued to the sample account. The sample events include, but are not limited to, purchasing commodities in the sample media information, downloading games, videos and other events in the sample media information. For example, the media information a in the first media information set is issued to account 1, and whether account 1 is for a good in the media information a, the actual conversion rate is 1 if the good is purchased, and the actual conversion rate is 0 if the good is not purchased.

The actual resource transfer quantity in the actual resource transfer quantity set represents the quantity of the actually transferred resource triggered by the sample media information issued to the sample account. For example, the media information a in the first media information is published to the account a, the media information a is published to a publishing platform (for example, even a communication client, a short video client, and other media information publishing clients), and the amount paid by the media information a to the publishing platform is the actual resource transfer amount.

As an optional implementation manner, in training the multitask model, the multitask model is trained through sample object features (user features, such as the user portraits) corresponding to the sample account, media information features (advertisement features) published to the sample account, an actual click rate and an actual conversion rate of the media information published to the sample account by the sample account, and an actual profit of the publishing platform when the media information is published to the sample account by the publishing platform.

The sample account set includes a plurality of sample accounts (e.g., the account A, B, C used by the sample object), which are the accounts used by the sample object. The first media information set includes a plurality of pieces of media information (ads), for example, a plurality of ads (e.g., ad 1, ad 2, and ad …. ad N) issued to the account A, B, C, where N may be 5 as an example (in practical applications, a large number of samples are required for training the multitask model). For example,

advertisements

1, 2, 3, 4, 5 are distributed to account C. And according to whether the advertisements issued to the account A, the account B and the account C are clicked or not, the actual click rate of each advertisement can be obtained. For example, account A is clicked on advertisement 1, account B is clicked on advertisement 3. The actual click rate set includes whether each account clicks on the advertisement distributed to it. Similarly, the actual conversion rate set includes whether each account executes a target event corresponding to the advertisement issued to the account, such as whether to purchase a commodity in the advertisement, and whether to download a game in the advertisement. The actual resource transfer amount set includes the revenue of the platform (the fee the advertiser pays to the platform) after the corresponding advertisement is issued to each account.

As an optional implementation manner, when the multitask model is trained, the training is ended and the target multitask model is obtained until the loss value output by the multitask model meets the preset loss condition. The loss value can be obtained by the following loss function:

Loss＝aloss_ctr+bloss_{shallow_cvr}+closs_{deep_cvr}+dloss_aux

loss in the above formula_ctrIs the loss of click Rate value, loss_{shallow_cvr}And loss_{deep_cvr}Is the loss of conversion, loss_auxIs the resource transfer quantity loss value. a. b, c and d are preset weights.

Click rate loss values and conversion rate loss values can be calculated by the following cross entropy loss function:

loss_x＝-∑[y_i*log P_i+(1-y_i)*log(1-P_i)]

wherein, y_iRepresenting an estimated value, P_iRepresenting the actual value. Loss in the above formula for the loss in click Rate value_xRepresents the click Rate loss value, y_iIndicates the estimated click rate, P_iRepresenting the actual click rate. And the estimated click rate is an estimated probability value of whether the advertisement is clicked or not by the CTR double towers according to the user characteristics of the sample users when the multitask model is trained. Loss in the above equation for the loss in conversion value_xDenotes the loss of conversion value, y_iRepresents the estimated conversion, P_iRepresenting the actual conversion. Wherein, the pre-estimation conversion rate is that when the multitask model is trained, the CVR double tower and the Deep CVR double tower pre-estimate whether the advertisement executes the target event or not according to the user characteristics of the sample userProbabilities, such as whether to download a game in the advertisement, and whether to purchase an item in the advertisement.

The Huber Loss function in regression Loss can be used for the resource transfer quantity Loss value as follows:

loss_aux＝HuberLoss(predict,ecpm)

where predict is the predicted resource transfer amount and ecpm is the actual resource transfer amount. The pre-estimated resource transfer quantity is a pre-estimated value of the profit of the publishing platform when the multitask model trains and publishes the advertisement with the advertisement characteristic to the user with the user characteristic according to the user characteristic and the advertisement characteristic.

Huber Loss is defined as follows:

the starting point for the Huber Loss is taken here to take into account that the sample Loss to prevent high eCPM leads to unstable model convergence.

As an alternative embodiment, the current sample account number is any account number in the set of sample account numbers. The current sample object features are object features using the current sample account, including features such as the user profile. The current media information is the media information published to the current sample account. For example, in the above embodiment, advertisement 1 (current media information) is issued to account a (current sample account). The current media information features are features of the media information issued to the current account, and include advertisement side features and the like in the above embodiments.

The current actual click rate is used to indicate whether the current sample account clicks the current media information after the current media information (advertisement) is issued to the current sample account. For example, after advertisement 1 is released to account a, whether account a clicked on advertisement 1.

The current actual conversion rate is used to indicate whether the current sample account executes the target event corresponding to the current media information after the current media information (advertisement) is issued to the current sample account. For example, if account a purchased the product in advertisement 1 or downloaded the game in advertisement 1 after advertisement 1 was released to account a.

The current actual conversion rate and the actual click rate may be represented by 0 and 1, where 0 represents no click and 1 represents click. 1 indicates that the current sample account performed the target event corresponding to the current media information, such as purchasing a good in the advertisement or downloading a game in the advertisement. 0 indicates that the target event corresponding to the current media information is not performed.

Inputting the characteristics of the current sample object into a User & context input in the graph 14, inputting the characteristics of the current media information into an Ad input, and outputting a current estimated click rate, a current estimated conversion rate and a current estimated resource transfer quantity by the multitask model when the multitask model is trained;

inputting the current estimated click rate and the current actual click rate into the loss_x＝-∑[y_i*log P_i+(1-y_i)*log(1-P_i)]Obtaining a current click rate loss value;

inputting the current estimated conversion rate and the current actual conversion rate into loss_x＝-∑[y_i*log P_i+(1-y_i)*log(1-P_i)]Obtaining a current conversion rate loss value;

inputting the current estimated resource transfer quantity and the current actual resource transfer quantity into the loss_auxObtaining a current resource transfer quantity loss value, namely HuberLoss (predict, ecpm);

inputting the obtained current click rate Loss value, the current conversion rate Loss value and the current resource transfer quantity Loss value into Loss which is aloss_ctr+bloss_{shallow_cvr}+closs_{deep_cvr}+dloss_auxObtaining a current loss value;

when the current Loss value meets a preset Loss condition, for example, the current Loss value Loss is smaller than a preset value (which may be set according to an actual situation, for example, 0.9, 0.8, etc.), ending the training, and determining the multitask model to be trained when the training is ended as a target multitask model; and when the current loss value does not meet the preset loss condition, adjusting parameters in the multi-task model to be trained.

As an alternative embodiment, the click rate double tower is a CTR double tower as shown in fig. 10, and the CTR double tower structure is a structure as shown in fig. 12. Inputting the characteristics of the current sample object (such as the User portrait characteristics) into User Input in the object tower, inputting the characteristics of current media information (advertisement characteristics) into Ad Input in the media information tower, and obtaining a current estimated click rate through a CTR double-tower structure shown in fig. 12, where the current estimated conversion rate is used to indicate that current media information corresponding to the current media information characteristics is issued to a current sample object corresponding to the characteristics of the current sample object, and the estimated probability of being clicked by the current sample object may be any value from 0 to 1.

The above-mentioned conversion double column includes CVR double column and Deep CVR double column of FIG. 10, and the structures of CVR double column and Deep CVR double column are shown in FIG. 12. The current sample object feature (e.g., the User portrait feature) is Input to the User Input in the object tower, the current media information feature (advertisement feature) is Input to the Ad Input in the media information tower, a current estimated conversion rate is obtained through a CVR double tower structure or a Deep CVR double tower structure shown in fig. 12, the current estimated conversion rate is used to indicate that the current media information corresponding to the current media information feature is distributed to the current sample object corresponding to the current sample object feature, and the probability that the current event in the current media information is executed by the current sample object (e.g., purchasing a commodity in the current media information, downloading a game in the current media information, etc.) may be any value from 0 to 1 to indicate the estimated probability.

Resource transfer number twin towers the ECPM twin towers in fig. 10, ECPM twin towers having the structure shown in fig. 13. Inputting the characteristics of the current sample object (such as the User portrait characteristics) into User Input in the object tower, inputting the characteristics of current media information (advertisement characteristics) into Ad Input in the media information tower, and obtaining the current estimated resource transfer quantity through an ECPM double-tower structure shown in FIG. 13, wherein the current estimated resource transfer quantity is used for representing that the current media information corresponding to the characteristics of the current media information is issued to the current sample object corresponding to the characteristics of the current sample object, and the advertisement owner corresponding to the current media information pays estimated money to the issuing platform.

As an optional implementation manner, since each publishing platform publishes media information to a user, a new sample account and sample characteristics corresponding to the sample account, and sample media information published to the sample account and characteristics of the sample media information are generated every day, and thus, new training samples are generated every day, and the new training samples update the multitask model in a streaming manner.

The current loss value is calculated by the loss function:

Loss＝aloss_ctr+bloss_cvr+dloss_aux

loss in the above formula_ctrIs the current click-through rate loss value, los_cvrIs the current conversion loss value, loss_auxIs the current resource transfer quantity loss value. a. b and d are preset weights.

Or the current loss value is calculated by the following loss function:

Loss＝aloss_ctr+bloss_{shallow_cvr}+closs_{deep_cvr}+dloss_aux

loss in the above formula_ctrIs the current click-through rate loss value, loss_{shallow_cvr}Is the superficial conversion loss value, loss_{deep_cvr}Is the deep conversion loss value, loss_auxIs the current resource transfer quantity loss value. a. b, c and d are preset weights.

As an alternative embodiment, the click rate double tower structure may include a shallow conversion double tower structure and a Deep conversion double tower structure, such as the Deep CVR double tower in fig. 10 is a Deep conversion double tower, and the CVR double tower is a shallow conversion double tower structure. The twin tower configuration of Deep CVR twin tower and CVR twin tower is shown as 12.

The following cross entropy loss functions are adopted for both the shallow layer conversion loss value and the deep layer conversion loss value:

loss_x＝-∑[y_i*log P_i+(1-y_i)*log(1-P_i)]

inputting the current sample object characteristics into a User & context input in fig. 10, inputting the current media information characteristics into an Ad input shown in fig. 10, and determining the shallow prediction conversion rate according to the current sample object characteristics and the current media information characteristics through a CVR double tower with a shallow conversion rate double tower structure shown in fig. 10. Determining a Deep pre-estimated conversion rate according to the characteristics of the current sample object and the characteristics of the current media information through a Deep conversion rate double tower Deep CVR double tower in the graph 10; the ECPM double tower of the resource transfer quantity double tower structure in fig. 10 determines the current pre-estimated resource transfer quantity according to the current sample object feature and the current media information feature.

As an optional implementation, the dual-tower structure includes an object tower and a media information tower, such as the current sample object feature in fig. 13 is Input to the object tower through User Input, and the current media information feature is Input to the media information tower through Ad Input. The object vector user embedding can be obtained through the object tower, and the media information vector ad embedding can be obtained through the media information tower. The current estimated resource transfer quantity is the dot product of the object vector and the media information vector. After the latest advertisement features are obtained from the advertisement data stream, the ad embedding output of the advertisement ECPM double tower is calculated by using the latest online trained model, and the online user requests the corresponding user embedding to search. The ad embedding is well calculated in advance off-line, in order to capture the latest advertisement embedding in time, the advertisement data flow module can acquire the latest features of the advertisement from the delivery database in real time, is convenient for the latest vector embedding to be calculated subsequently, uses the database for storing the total set of the real-time latest delivered advertisement and relevant attributes thereof on line, and the database is also updated frequently due to frequent operation of an advertiser.

As an optional implementation manner, the value of the recall parameter of each piece of media information is determined to be equal to the weighted sum of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information, and the estimated resource transfer quantity of each piece of media information, and the weight can be set according to the actual situation. The value of each media information recall parameter can also be determined to be equal to the product of the estimated click rate of each media information, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information.

As an optional implementation manner, the click rate weight value, the conversion rate weight value, and the resource transfer amount weight value corresponding to each piece of media information may be determined according to an actual situation. For example, the resource transfer quantity weight value is greater than the click rate weight value and the conversion rate weight value. Assume that the resource transfer quantity weight is 0.6, the click rate weight is 0.2, and the conversion rate weight is 0.2.

Alternatively, the weight value of the resource transfer quantity is greater than or equal to a preset weight value, and the preset weight value may be determined according to the actual situation, for example, 0.5. And weighting and summing the estimated click rate of each media message, the estimated conversion rate of each media message and the estimated resource transfer quantity of each media message with a click rate weight value, a conversion rate weight value and a resource transfer quantity weight value corresponding to each media message respectively to obtain a value of the recall parameter of each media message.

As an optional implementation manner, the media information in the media information set is sorted according to the recall parameter value of each media information, and the media information of N before ranking is selected as the recall media information of the target account, where the value of N may be selected according to an actual situation, for example, 100, 1000, 1500, and the like.

As an optional implementation, in the recall module, for the intelligent targeted advertisement, a two-tower model is used to retrieve the recall based on ANN, an online user embedding is used to retrieve the neighbor ad embedding, the model modeling target determines retrieval indexes (such as conventional retrieval indexes CTR, CVR and the commercial value eCPM retrieval index of the present invention), and the retrieved advertisement further enters the rough ranking module for ranking.

In the rough arrangement module, the click rate and the conversion rate of each advertisement are estimated, and specifically, the click rate and the conversion rate of each advertisement can be estimated through a first click rate estimation model and a first conversion rate estimation model. And obtaining the estimated click rate and the estimated conversion rate of each advertisement. The eCPM for each advertisement is calculated by the following formula:

eCPM (current number of resource transfers) litecr (estimated click rate) litecrv (estimated conversion rate). And roughly ranking the advertisements in the recalled media information subsets according to the ranking of each advertisement eCPM, and selecting the roughly ranked media information subsets, wherein the roughly ranked media information subsets can be advertisements at the top M of the eCPM, such as 100, 200 and the like.

In the fine ranking module, the advertisement system firstly obtains pCTR (accurate estimated click rate) and pCVR (accurate estimated conversion rate) of all advertisement which are coarsely ranked, when the pCTR and pCVR of all advertisement are estimated, the number of used user features is larger than the number of used user features used in coarse ranking (the number of features in the first object feature is smaller than the number of features in the second object feature), and the number of used media information features is larger than the number of used media information features (the number of features in the first media information feature is smaller than the number of features in the second media information feature). For example, the first object feature includes 10 features of the user, and the second object feature includes 100 features of the user. The first media information characteristic includes 20 characteristics of the media information, and the second media information characteristic includes 200 user characteristics of the media information.

By the following formula: calculating the CPM value of each media information in the roughly arranged media information subset obtained by rough arrangement:

eCPM2 (current number of resource transfers) pCTR (accurate estimated click rate) pCVR (accurate estimated conversion rate)

And selecting the top P (for example, 1, 3 and the like) advertisements in the rough media information subset according to the height of the eCPM2 as the target media information to be presented to the target account.

The online recall link of the embodiment of the invention provides the recall of the advertisement set for the subsequent rough ranking and the fine ranking of the whole advertisement system, influences the set of the subsequent competition of the whole advertisement system and is a very important link.

The value eCPM (current resource transfer quantity) litecr (estimated click rate) litecrv (estimated conversion rate) of the first screening parameter. And roughly ranking the advertisements in the recalled media information subsets according to the ranking of each advertisement eCPM, and selecting the roughly ranked media information subsets, wherein the roughly ranked media information subsets can be advertisements at the top M of the eCPM, such as 100, 200 and the like.

The value eCPM2 of the second screening parameter is bid (current resource transfer quantity) pCTR (accurate estimated click rate) pCVR (accurate estimated conversion rate). And selecting the top P (for example, 1, 3 and the like) advertisements in the rough media information subset according to the height of the eCPM2 as the target media information to be presented to the target account.

FIG. 15 is an overall framework diagram of an intelligent targeted ad recall scheme based on full link business value multitasking modeling. The method mainly comprises the following steps: the method comprises three modules of commercial value model training, commercial value model prediction and on-line advertisement system recalling link effective. The commercial value model training module is divided into three parts of log analysis, feature construction and model training, and is mainly responsible for acquiring various features (including advertisement features, user features and advertisement competition environment features) required by the model from an original log and a feature platform of an advertisement system, and further constructing a full-link commercial value multi-task modeling model by using the features for training. The commercial value model prediction module mainly comprises three modules, namely a release DB maintenance module, an advertisement data stream acquisition module, an advertisement embedding update module and a model training module, wherein the three modules are mainly used for acquiring the real-time latest advertisement state and calculating the latest advertisement embedding by using the latest model. The advertisement system recalling link validation module is mainly used for acquiring the latest model and the latest advertisement embedding, calculating the latest user embedding by using the model for the current user request, and then retrieving the corresponding advertisement embedding by using the user embedding through an ANN (ANN search algorithm (specific meaning is defined by the initial abbreviation of the text), so that the corresponding advertisement is recalled, the recalled advertisement is subjected to subsequent rough arrangement and fine arrangement of the advertisement system until bidding succeeds, and the exposure opportunity on the current user request is acquired.

And the log analysis module is responsible for analyzing the original exposure log of the advertisement system to obtain the training data of the model. As the most important data source of the advertising system, the exposure log and the refinement log of the advertising system are the most important samples for training the recall model. The module mainly periodically draws the online real-time log and analyzes the online real-time log into a standardized sample, so that the subsequent characteristic module can conveniently process the standard sample.

Commercial value eCPM defines: for the oCPA advertisement, the calculation formula of the fine ranking eCPM is as follows: eCPM ═ bid × pCTR × pCVR × λ

Wherein bid is bid of the advertisement, and pCTR and pCVR are estimated click rate and conversion rate of the current user on the current advertisement at the advertisement refinement stage; lambda (lambda) is the tuning factor (the specific meaning is defined with reference to the acronym in the text). Since the fine ECPM is a precise pre-estimate that measures the value of an advertising business, the fine ECPM can be selected here as a modeling target for recalls in order to improve the consistency of the recall-coarse-fine link. A regression model was constructed in the recall to directly fit the fine-ranked eCPM as a commercial value model.

After the model scheme of the pre-estimated fine-emission ECPM is adopted, a retrieval framework for recalling needs to be further determined. Since the recall prediction set is a global advertisement library, the time for the online request to be left for the advertisement recall is limited, it is impossible to traverse all advertisements by using a full link model, and in order to balance the effect and efficiency of the model, the scheme uses a classical ANN retrieval architecture, as shown in fig. 16. Namely, the model structure adopts a double-tower structure, the user embedding and the advertisement embedding are separated, the advertisement embedding can be offline and calculated in advance, the user embedding is online and calculated in real time, when a user request comes, the user embedding is calculated in real time, and then the adjacent advertisement embedding is searched out by utilizing the user embedding based on an ANN search algorithm (such as HNSW).

In the recalling link of the advertising system, the method directly models the recalling model by taking the commercial value as a target, so that the recalling link can be aligned with a core sequencing index (eCPM) of an advertising platform, and the commercial value income of the whole advertising platform is improved from the recalling link. In consideration of the problems of 'sample selection deviation' in recalling and high fitting difficulty of eCPM direct modeling, the full-link commercial value multi-task modeling scheme is provided, the model fitting difficulty is reduced, and the model fitting accuracy is improved.

While, for purposes of simplicity of explanation, the foregoing method embodiments have been described as a series of acts or combination of acts, it will be appreciated by those skilled in the art that the present invention is not limited by the illustrated ordering of acts, as some steps may occur in other orders or concurrently in accordance with the invention. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required by the invention.

According to another aspect of the embodiment of the present invention, there is also provided a media information determination apparatus for implementing the above media information determination method. As shown in fig. 17, the apparatus includes: an obtaining module 1702, configured to obtain a target object characteristic of a target account; a first determining module 1704, configured to determine, according to the target object feature, the estimated click rate of each piece of media information in a media information set, the estimated conversion rate of each piece of media information, and the estimated resource transfer quantity of each piece of media information, a media information subset recalled by the target account in the media information set, where the media information set includes the piece of media information to be published, the estimated click rate of each piece of media information indicates a probability that each piece of media information is clicked by the target account when the piece of media information is published to the target account, the estimated conversion rate of each piece of media information indicates a probability that a target event corresponding to each piece of media information is executed by the target account when the piece of media information is published to the target account, and the estimated resource transfer quantity of each piece of media information indicates a probability of a resource triggered to be transferred when the piece of media information is published to the target account The number of the particles; a second determining module 1706, configured to determine, in the media information subset, the target media information issued to the target account.

Optionally, the apparatus is further configured to input the target object feature and the media information feature of each media information into a target multitasking model; respectively determining the estimated click rate of each media information, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information according to the target object characteristics and the media information characteristics of each media information through the target multitask model, and determining the value of the recall parameter of each media information according to the estimated click rate of each media information, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information; and determining the media information subset recalled for the target account in the media information set according to the recall parameter value of each piece of media information.

Optionally, the apparatus is further configured to obtain a sample object feature set corresponding to a sample account set, a first media information feature set corresponding to a first media information set published to the sample account set, an actual click rate set corresponding to the first media information set, an actual conversion rate set corresponding to the first media information set, and an actual resource transfer quantity set corresponding to the first media information set; training a multi-task model to be trained by using the sample object feature set and the first media information feature set until a target loss value corresponding to the multi-task model to be trained meets a preset loss condition, and ending the training to obtain the target multi-task model, wherein the target loss value is a loss value determined according to a click rate loss value, a conversion rate loss value and a resource transfer quantity loss value; the click rate loss value represents a loss value between an estimated click rate and an actual click rate corresponding to the actual click rate set, and the estimated click rate is determined by the multitask model to be trained according to sample object features corresponding to the sample object feature set and sample media information features corresponding to the first media information feature set; the conversion rate loss value represents a loss value between a pre-estimated conversion rate and a corresponding actual conversion rate in the actual conversion rate set, and the pre-estimated conversion rate is determined by the multitask model to be trained according to corresponding sample object features in the sample object feature set and corresponding sample media information features in the first media information feature set; the resource transfer quantity loss value represents a loss value between a pre-estimated resource transfer quantity and a corresponding actual resource transfer quantity in the actual resource transfer quantity rate set, and the pre-estimated resource transfer quantity is determined by the multitask model to be trained according to a corresponding sample object feature in the sample object feature set and a corresponding sample media information feature in the first media information feature set.

Optionally, the apparatus is further configured to obtain a current sample object feature of a current sample account in the sample object feature set, obtain a current media information feature of current media information issued to the current sample account in the first media information set feature, obtain a current actual click rate corresponding to the current sample account and the current media information in the actual click rate set, obtain a current actual conversion rate corresponding to the current sample account and the current media information in the actual conversion rate set, and obtain a current actual resource transfer quantity corresponding to the current sample account and the current media information in the actual resource transfer quantity set; inputting the current sample object characteristics and the current media information characteristics into the multi-task model to be trained, and respectively determining the corresponding current pre-estimated click rate, current pre-estimated conversion rate and current pre-estimated resource transfer quantity according to the current sample object characteristics and the current media information characteristics through the multi-task model to be trained; determining a current click rate loss value according to the current estimated click rate and the current actual click rate, determining a current conversion rate loss value according to the current estimated conversion rate and the current actual conversion rate, and determining a current resource transfer quantity loss value according to the current estimated resource transfer quantity and the current actual resource transfer quantity; determining a current loss value according to the weighted sum of the current click rate loss value, the current conversion rate loss value and the current resource transfer quantity loss value; when the current loss value meets the preset loss condition, ending training, and determining the multi-task model to be trained as the target multi-task model when training is ended; and when the current loss value does not meet the preset loss condition, adjusting parameters in the multi-task model to be trained.

Optionally, the device is further configured to determine the current estimated click rate according to the current sample object feature and the current media information feature through a click rate double-tower structure in the multi-task model to be trained; determining the current pre-estimated conversion rate according to the current sample object characteristics and the current media information characteristics through a conversion rate double-tower structure in the multi-task model to be trained; and determining the current estimated resource transfer quantity according to the current sample object characteristic and the current media information characteristic through a resource transfer quantity double-tower structure in the multi-task model to be trained.

Optionally, the device is further configured to determine a weighted sum of the current click rate loss value, the current conversion rate loss value, and the current resource transfer quantity loss value as the current loss value; or when the current conversion rate loss value comprises a shallow conversion rate loss value and a deep conversion rate loss value, determining the current click rate loss value, the shallow conversion rate loss value, the deep conversion rate loss value and the weighted sum of the current resource transfer quantity loss value as the current loss value.

Optionally, the device is further configured to determine the current estimated click rate according to the current sample object feature and the current media information feature through a click rate double-tower structure in the multi-task model to be trained; when the current conversion rate loss value comprises a shallow conversion rate loss value and a deep conversion rate loss value, determining the shallow pre-estimated conversion rate according to the current sample object characteristic and the current media information characteristic through a shallow conversion rate double-tower structure in the multi-task model to be trained, and determining the deep pre-estimated conversion rate according to the current sample object characteristic and the current media information characteristic through a deep conversion rate double-tower structure in the multi-task model to be trained; and determining the current estimated resource transfer quantity according to the current sample object characteristic and the current media information characteristic through a resource transfer quantity double-tower structure in the multi-task model to be trained.

Optionally, the apparatus is further configured to generate an object vector according to the current sample object feature through an object tower structure; generating a media information vector according to the current media information characteristics through a media information tower structure, wherein the resource transfer quantity double-tower structure comprises the object tower structure and the media information tower structure; and determining the current pre-estimated resource transfer quantity as the dot product of the object vector and the media information vector.

Optionally, the apparatus is further configured to determine a value of the recall parameter of each piece of media information as a weighted sum of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information, and the estimated resource transfer quantity of each piece of media information; or determining the value of the recall parameter of each piece of media information to be equal to the product of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information and the estimated resource transfer quantity of each piece of media information.

Optionally, the apparatus is further configured to obtain a click rate weight value, a conversion rate weight value, and a resource transfer quantity weight value corresponding to each piece of media information, where the resource transfer quantity weight value is greater than the click rate weight value and the conversion rate weight value, or the resource transfer quantity weight value is greater than or equal to a preset weight value; and weighting and summing the estimated click rate of each media message, the estimated conversion rate of each media message and the estimated resource transfer quantity of each media message with a click rate weight value, a conversion rate weight value and a resource transfer quantity weight value corresponding to each media message respectively to obtain a value of a recall parameter of each media message.

Optionally, the apparatus is further configured to rank, according to a size of the recall parameter of each piece of media information, the pieces of media information that are N bits before the recall parameter is ranked in the media information set are determined as the media information subset, where N is a positive integer greater than or equal to 1.

Optionally, the apparatus is further configured to obtain a first object feature of the target account, and obtain a first media information feature of each piece of media information in the media information subset; determining a first estimated click rate set corresponding to the media information subset according to the first object characteristics and the first media information characteristics of each piece of media information in the media information subset through a first click rate estimation model; determining a first pre-estimated conversion rate set corresponding to the media information subset according to the first object characteristic and the first media information characteristic of each media information in the media information subset through a first conversion rate pre-estimated model; determining a rough-arranged media information subset in the media information subset according to the first pre-estimated click rate set, the first pre-estimated conversion rate set and a current resource transfer quantity set corresponding to the media information subset; acquiring a second object characteristic of the target account, and acquiring a second media information characteristic of each media information in the rough media information subset, wherein the number of characteristics in the first object characteristic is less than that in the second object characteristic, and the number of characteristics in the first media information characteristic is less than that in the second media information characteristic; determining a second estimated click rate set corresponding to the roughly arranged media information subset according to the second object characteristics and second media information characteristics of each piece of media information in the roughly arranged media information subset through a second click rate estimation model; determining a second pre-estimated conversion rate set corresponding to the roughly arranged media information subset according to the second object characteristics and the second media information characteristics of each piece of media information in the roughly arranged media information subset through a second conversion rate pre-estimated model; and determining the target media information in the rough arranged media information subset according to the second pre-estimated click rate set, the second pre-estimated conversion rate set and the current resource transfer quantity set corresponding to the rough arranged media information subset.

Optionally, the device is further configured to multiply the estimated click rate in the first estimated click rate set, the estimated conversion rate in the first estimated conversion rate set, and the current resource transfer quantity in the current resource transfer quantity set correspondingly to obtain a value of a first screening parameter corresponding to the media information in the media information subset; and according to the rank of the values of the first screening parameters from high to low, determining the top M bits of media information in the media information subsets as the coarse-ranking media information subsets, wherein M is a positive integer greater than or equal to 2.

Optionally, the device is further configured to multiply the estimated click rate in the second estimated click rate set, the estimated conversion rate in the second estimated conversion rate set, and the current resource transfer quantity in the current resource transfer quantity set corresponding to the rough-ranking media information subset correspondingly to obtain a value of a second screening parameter corresponding to the media information in the rough-ranking media information subset; and according to the rank of the values of the second screening parameters from high to low, determining the P-bit media information before ranking in the rough media information subset as the target media information, wherein P is a positive integer greater than or equal to 1.

According to another aspect of the embodiment of the present invention, there is also provided an electronic device for implementing the method for determining media information, where the electronic device may be a terminal device or a server shown in fig. 1. The present embodiment takes the electronic device 18 as an example for explanation. As shown in fig. 18, the electronic device comprises a memory 1802 having stored therein a computer program, and a processor 1804 arranged to execute the steps of any of the above-described method embodiments by means of the computer program.

Optionally, in this embodiment, the electronic device may be located in at least one network device of a plurality of network devices of a computer network.

Optionally, in this embodiment, the processor may be configured to execute the following steps by a computer program:

s1, acquiring the target object characteristics of the target account;

s2, according to the target object characteristics, the estimated click rate of each media information in the media information set, the estimated conversion rate of each media information, and the estimated resource transfer quantity of each media information, determining the media information subset recalled from the target account in the media information set, wherein, the media information set comprises the media information to be issued, the estimated click rate of each media information represents the probability that each media information is clicked by the target account when each media information is issued to the target account, the pre-estimated conversion rate of each media message represents the probability that the target event corresponding to each media message is executed by the target account when each media message is issued to the target account, the pre-estimated resource transfer quantity of each media information represents the quantity of resources triggered to be transferred when each media information is issued to the target account;

and S3, determining the target media information issued to the target account in the media information subset.

Alternatively, it can be understood by those skilled in the art that the structure shown in fig. 18 is only an illustration, and the electronic device may also be a terminal device such as a smart phone (e.g., an Android phone, an iOS phone, etc.), a tablet computer, a palmtop computer, a Mobile Internet Device (MID), a PAD, and the like. Fig. 18 is a diagram illustrating a structure of the electronic device. For example, the electronics may also include more or fewer components (e.g., network interfaces, etc.) than shown in FIG. 18, or have a different configuration than shown in FIG. 18.

The memory 1802 may be used to store software programs and modules, such as program instructions/modules corresponding to the method and apparatus for determining media information in the embodiments of the present invention, and the processor 1804 executes the software programs and modules stored in the memory 1802, so as to execute various functional applications and data processing, that is, to implement the above-mentioned method for determining media information. The memory 1802 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 1802 can further include memory located remotely from the processor 1804, which can be connected to the terminals over a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof. The memory 1802 may be specifically, but not limited to, used for storing information such as sample characteristics of an item and a target virtual resource account number. As an example, as shown in fig. 18, the memory 1802 may include, but is not limited to, an obtaining module 1702, a first determining module 1704, and a second determining module 1706 of the apparatus for determining media information. In addition, other module units in the above-mentioned device for determining media information may also be included, but are not limited to this, and are not described in detail in this example.

Optionally, the transmitting device 1806 is configured to receive or transmit data via a network. Examples of the network may include a wired network and a wireless network. In one example, the transmission device 1806 includes a Network adapter (NIC) that can be connected to a router via a Network cable and other Network devices to communicate with the internet or a local area Network. In one example, the transmission device 1806 is a Radio Frequency (RF) module, which is used for communicating with the internet in a wireless manner.

In addition, the electronic device further includes: a display 1808, configured to display the to-be-processed order information; and a connection bus 1810 for connecting the respective module components in the above-described electronic apparatus.

In other embodiments, the terminal device or the server may be a node in a distributed system, where the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting a plurality of nodes through a network communication. Nodes can form a Peer-To-Peer (P2P, Peer To Peer) network, and any type of computing device, such as a server, a terminal, and other electronic devices, can become a node in the blockchain system by joining the Peer-To-Peer network.

According to an aspect of the application, there is provided a computer program product comprising a computer program/instructions containing program code for performing the method illustrated by the flow chart. In such embodiments, the computer program may be downloaded and installed from a network via communications portion 1909 and/or installed from removable media 1911. When executed by the central processing unit 1901, the computer program performs various functions provided by the embodiments of the present application.

The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments.

It should be noted that the computer system 1900 of the electronic device shown in fig. 19 is only an example, and should not bring any limitation to the functions and the scope of use of the embodiments of the present application.

As shown in fig. 19, the computer system 1900 includes a Central Processing Unit (CPU) 1901 that can perform various appropriate actions and processes in accordance with a program stored in a Read-Only Memory (ROM) 1902 or a program loaded from a storage section 1908 into a Random Access Memory (RAM) 1903. In the random access memory 1903, various programs and data necessary for system operation are also stored. The cpu 1901, the rom 1902, and the ram 1903 are connected to each other via a bus 1904. An Input/Output interface 1905(Input/Output interface, i.e., I/O interface) is also connected to the bus 1904.

The following components are connected to the input/output interface 1905: an input section 1906 including a keyboard, a mouse, and the like; an output section 1907 including a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and the like, a speaker, and the like; a storage section 1908 including a hard disk and the like; and a communications portion 1909 that includes a network interface card, such as a local area network card, modem, and the like. The communication section 1909 performs communication processing via a network such as the internet. A driver 1910 is also connected to the input/output interface 1905 as needed. A removable medium 1911 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 1910 as necessary, so that a computer program read out therefrom is mounted in the storage section 1908 as necessary.

In particular, according to embodiments of the present application, the processes described in the various method flowcharts may be implemented as computer software programs. For example, embodiments of the present application include a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code for performing the method illustrated by the flow chart. In such embodiments, the computer program may be downloaded and installed from a network via communications portion 1909 and/or installed from removable media 1911. When executed by the central processor 1901, performs various functions defined in the system of the present application.

According to an aspect of the present application, there is provided a computer-readable storage medium from which a processor of a computer device reads computer instructions, the processor executing the computer instructions to cause the computer device to perform the method provided in the above-mentioned various alternative implementations.

Alternatively, in the present embodiment, the above-mentioned computer-readable storage medium may be configured to store a computer program for executing the steps of:

s1, acquiring the target object characteristics of the target account;

Alternatively, in this embodiment, a person skilled in the art may understand that all or part of the steps in the methods of the foregoing embodiments may be implemented by a program instructing hardware associated with the terminal device, where the program may be stored in a computer-readable storage medium, and the storage medium may include: flash disks, Read-Only memories (ROMs), Random Access Memories (RAMs), magnetic or optical disks, and the like.

The integrated unit in the above embodiments, if implemented in the form of a software functional unit and sold or used as a separate product, may be stored in the above computer-readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing one or more computer devices (which may be personal computers, servers, network devices, etc.) to execute all or part of the steps of the method according to the embodiments of the present invention.

In the above embodiments of the present invention, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

In the several embodiments provided in the present application, it should be understood that the disclosed client may be implemented in other manners. The above-described embodiments of the apparatus are merely illustrative, and for example, the division of the units is only one type of division of logical functions, and there may be other divisions when actually implemented, for example, a plurality of units or components may be combined or may be integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, units or modules, and may be in an electrical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.

The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations should also be regarded as the protection scope of the present invention.

Claims

1. A method for determining media information, comprising:

acquiring target object characteristics of a target account;

determining a media information subset recalled by the target account in the media information set according to the target object characteristics, the estimated click rate of each media information in the media information set, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information, wherein, the media information set comprises the media information to be issued, the estimated click rate of each media information represents the probability that each media information is clicked by the target account when each media information is issued to the target account, the pre-estimated conversion rate of each media message represents the probability that the target event corresponding to each media message is executed by the target account when each media message is issued to the target account, the pre-estimated resource transfer quantity of each media information represents the quantity of resources triggered to be transferred when each media information is issued to the target account;

and determining target media information issued to the target account in the media information subset.

2. The method of claim 1, wherein the determining a subset of media information to be published to the target account in the set of media information according to the target object characteristics, the estimated click-through rate of each media information in the set of media information, the estimated conversion rate of each media information, and the estimated resource transfer amount of each media information comprises:

inputting the target object features and the media information features of each media information into a target multitask model;

respectively determining the estimated click rate of each media information, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information according to the target object characteristics and the media information characteristics of each media information through the target multitask model, and determining the value of the recall parameter of each media information according to the estimated click rate of each media information, the estimated conversion rate of each media information and the estimated resource transfer quantity of each media information;

and determining the media information subset recalled for the target account in the media information set according to the recall parameter value of each piece of media information.

3. The method of claim 2, further comprising:

acquiring a sample object feature set corresponding to a sample account set, a first media information feature set corresponding to a first media information set issued to the sample account set, an actual click rate set corresponding to the first media information set, an actual conversion rate set corresponding to the first media information set, and an actual resource transfer quantity set corresponding to the first media information set;

training a multi-task model to be trained by using the sample object feature set and the first media information feature set until a target loss value corresponding to the multi-task model to be trained meets a preset loss condition, and ending the training to obtain the target multi-task model, wherein the target loss value is a loss value determined according to a click rate loss value, a conversion rate loss value and a resource transfer quantity loss value;

the click rate loss value represents a loss value between an estimated click rate and an actual click rate corresponding to the actual click rate set, and the estimated click rate is determined by the multitask model to be trained according to sample object features corresponding to the sample object feature set and sample media information features corresponding to the first media information feature set;

the conversion rate loss value represents a loss value between a pre-estimated conversion rate and a corresponding actual conversion rate in the actual conversion rate set, and the pre-estimated conversion rate is determined by the multitask model to be trained according to corresponding sample object features in the sample object feature set and corresponding sample media information features in the first media information feature set;

the resource transfer quantity loss value represents a loss value between a pre-estimated resource transfer quantity and a corresponding actual resource transfer quantity in the actual resource transfer quantity rate set, and the pre-estimated resource transfer quantity is determined by the multitask model to be trained according to a corresponding sample object feature in the sample object feature set and a corresponding sample media information feature in the first media information feature set.

4. The method according to claim 3, wherein the training of the multi-tasking model to be trained by using the sample object feature set and the first media information feature set is performed until a target loss value corresponding to the multi-tasking model to be trained satisfies a preset loss condition, and the ending of the training to obtain the target multi-tasking model comprises:

acquiring a current sample object characteristic of a current sample account in the sample object characteristic set, acquiring a current media information characteristic of current media information issued to the current sample account in the first media information set characteristic, acquiring a current actual click rate corresponding to the current sample account and the current media information in the actual click rate set, acquiring a current actual conversion rate corresponding to the current sample account and the current media information in the actual conversion rate set, and acquiring a current actual resource transfer quantity corresponding to the current sample account and the current media information in the actual resource transfer quantity set;

inputting the current sample object characteristics and the current media information characteristics into the multi-task model to be trained, and respectively determining the corresponding current pre-estimated click rate, current pre-estimated conversion rate and current pre-estimated resource transfer quantity according to the current sample object characteristics and the current media information characteristics through the multi-task model to be trained;

determining a current click rate loss value according to the current estimated click rate and the current actual click rate, determining a current conversion rate loss value according to the current estimated conversion rate and the current actual conversion rate, and determining a current resource transfer quantity loss value according to the current estimated resource transfer quantity and the current actual resource transfer quantity;

determining a current loss value according to the weighted sum of the current click rate loss value, the current conversion rate loss value and the current resource transfer quantity loss value;

when the current loss value meets the preset loss condition, ending training, and determining the multi-task model to be trained as the target multi-task model when training is ended;

and when the current loss value does not meet the preset loss condition, adjusting parameters in the multi-task model to be trained.

5. The method of claim 4, wherein the determining, by the multitask model to be trained, a corresponding current estimated click rate, a current estimated conversion rate and a current estimated resource transfer number according to the current sample object characteristic and the current media information characteristic, respectively, comprises:

determining the current estimated click rate according to the current sample object characteristics and the current media information characteristics through a click rate double-tower structure in the multi-task model to be trained;

determining the current pre-estimated conversion rate according to the current sample object characteristics and the current media information characteristics through a conversion rate double-tower structure in the multi-task model to be trained;

and determining the current estimated resource transfer quantity according to the current sample object characteristic and the current media information characteristic through a resource transfer quantity double-tower structure in the multi-task model to be trained.

6. The method of claim 4, wherein determining a current loss value based on a weighted sum of the current click-through rate loss value, the current conversion rate loss value, and the current resource transfer quantity loss value comprises:

determining the weighted sum of the current click rate loss value, the current conversion rate loss value and the current resource transfer quantity loss value as the current loss value; or

And when the current conversion rate loss value comprises a superficial conversion rate loss value and a deep conversion rate loss value, determining the current click rate loss value, the superficial conversion rate loss value, the deep conversion rate loss value and the weighted sum of the current resource transfer quantity loss value as the current loss value.

7. The method as claimed in claim 6, wherein the determining, by the multitask model to be trained, a corresponding current estimated click rate, a current estimated conversion rate and a current estimated resource transfer quantity according to the current sample object feature and the current media information feature respectively comprises:

when the current conversion rate loss value comprises a shallow conversion rate loss value and a deep conversion rate loss value, determining the shallow pre-estimated conversion rate according to the current sample object characteristic and the current media information characteristic through a shallow conversion rate double-tower structure in the multi-task model to be trained, and determining the deep pre-estimated conversion rate according to the current sample object characteristic and the current media information characteristic through a deep conversion rate double-tower structure in the multi-task model to be trained;

8. The method according to claim 5 or 7, wherein the determining the current pre-estimated resource transfer amount according to the current sample object feature and the current media information feature by the resource transfer amount double tower structure in the multitask model to be trained comprises:

generating an object vector according to the current sample object characteristics through an object tower structure;

generating a media information vector according to the current media information characteristics through a media information tower structure, wherein the resource transfer quantity double-tower structure comprises the object tower structure and the media information tower structure;

and determining the current pre-estimated resource transfer quantity as the dot product of the object vector and the media information vector.

9. The method of claim 2, wherein the determining the recall parameter value of each media message according to the estimated click rate of each media message, the estimated conversion rate of each media message, and the estimated resource transfer amount of each media message comprises:

determining the value of the recall parameter of each piece of media information as the weighted sum of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information and the estimated resource transfer quantity of each piece of media information; or

And determining the value of the recall parameter of each piece of media information as equal to the product of the estimated click rate of each piece of media information, the estimated conversion rate of each piece of media information and the estimated resource transfer quantity of each piece of media information.

10. The method of claim 9, wherein the determining the recall parameter of each media message as a weighted sum of the predicted click through rate of each media message, the predicted conversion rate of each media message, and the predicted number of resource transfers of each media message comprises:

acquiring a click rate weight value, a conversion rate weight value and a resource transfer quantity weight value corresponding to each piece of media information, wherein the resource transfer quantity weight value is greater than the click rate weight value and the conversion rate weight value, or the resource transfer quantity weight value is greater than or equal to a preset weight value;

and weighting and summing the estimated click rate of each media message, the estimated conversion rate of each media message and the estimated resource transfer quantity of each media message with a click rate weight value, a conversion rate weight value and a resource transfer quantity weight value corresponding to each media message respectively to obtain a value of a recall parameter of each media message.

11. The method of claim 2, wherein the determining the subset of media information in the set of media information that is recalled to the target account according to the recall parameter of each media information includes:

and according to the arrangement of the values of the recall parameters of each piece of media information from large to small, determining the media information with the value of the recall parameter ranked N before as the subset of the media information in the media information set, wherein N is a positive integer greater than or equal to 1.

12. The method according to any one of claims 1 to 7 and 9 to 11, wherein determining the target media information published to the target account in the media information subset comprises:

acquiring a first object characteristic of the target account, and acquiring a first media information characteristic of each media information in the media information subset;

determining a first estimated click rate set corresponding to the media information subset according to the first object characteristics and the first media information characteristics of each piece of media information in the media information subset through a first click rate estimation model; determining a first pre-estimated conversion rate set corresponding to the media information subset according to the first object characteristic and the first media information characteristic of each media information in the media information subset through a first conversion rate pre-estimated model;

determining a rough-arranged media information subset in the media information subset according to the first pre-estimated click rate set, the first pre-estimated conversion rate set and a current resource transfer quantity set corresponding to the media information subset;

acquiring a second object characteristic of the target account, and acquiring a second media information characteristic of each media information in the rough media information subset, wherein the number of characteristics in the first object characteristic is less than that in the second object characteristic, and the number of characteristics in the first media information characteristic is less than that in the second media information characteristic;

determining a second estimated click rate set corresponding to the roughly arranged media information subset according to the second object characteristics and second media information characteristics of each piece of media information in the roughly arranged media information subset through a second click rate estimation model; determining a second pre-estimated conversion rate set corresponding to the roughly arranged media information subset according to the second object characteristics and the second media information characteristics of each piece of media information in the roughly arranged media information subset through a second conversion rate pre-estimated model;

and determining the target media information in the rough arranged media information subset according to the second pre-estimated click rate set, the second pre-estimated conversion rate set and the current resource transfer quantity set corresponding to the rough arranged media information subset.

13. The method of claim 12, wherein said determining a coarse media information subset from the media information subsets according to the first set of pre-estimated click rates, the first set of pre-estimated conversion rates, and the set of current resource transfer amounts corresponding to the media information subsets comprises:

correspondingly multiplying the estimated click rate in the first estimated click rate set, the estimated conversion rate in the first estimated conversion rate set and the current resource transfer quantity in the current resource transfer quantity set to obtain a value of a first screening parameter corresponding to the media information in the media information subset;

and according to the rank of the values of the first screening parameters from high to low, determining the top M bits of media information in the media information subsets as the coarse-ranking media information subsets, wherein M is a positive integer greater than or equal to 2.

14. The method of claim 12, wherein determining the target media information in the coarse media information subset according to the second set of pre-estimated click rates, the second set of pre-estimated conversion rates, and the set of current resource transfer amounts corresponding to the coarse media information subset comprises:

correspondingly multiplying the estimated click rate in the second estimated click rate set, the estimated conversion rate in the second estimated conversion rate set and the current resource transfer quantity in the current resource transfer quantity set corresponding to the rough media information subset to obtain the value of a second screening parameter corresponding to the media information in the rough media information subset;

and according to the rank of the values of the second screening parameters from high to low, determining the P-bit media information before ranking in the rough media information subset as the target media information, wherein P is a positive integer greater than or equal to 1.

15. An apparatus for determining media information, comprising:

the acquisition module is used for acquiring the target object characteristics of the target account;

a first determining module, configured to determine, according to the target object feature, an estimated click rate of each media information in a media information set, an estimated conversion rate of each media information, and an estimated resource transfer quantity of each media information, a media information subset recalled by the target account in the media information set, where the media information set includes media information to be published, the estimated click rate of each media information indicates a probability that each media information is clicked by the target account when the media information is published to the target account, the estimated conversion rate of each media information indicates a probability that a target event corresponding to each media information is executed by the target account when the media information is published to the target account, and the estimated resource transfer quantity of each media information indicates a quantity of resources triggered to be transferred when each media information is published to the target account An amount;

and the second determining module is used for determining the target media information issued to the target account in the media information subset.

16. A computer-readable storage medium, comprising a stored program, wherein the program when executed performs the method of any of claims 1 to 14.

17. A computer program product comprising computer program/instructions, characterized in that the computer program/instructions, when executed by a processor, implement the steps of the method of any of claims 1 to 14.

18. An electronic device comprising a memory and a processor, characterized in that the memory has stored therein a computer program, the processor being arranged to execute the method of any of claims 1 to 14 by means of the computer program.