CN116957685A

CN116957685A - Advertisement recommendation method, device, equipment and medium

Info

Publication number: CN116957685A
Application number: CN202310125188.1A
Authority: CN
Inventors: 王唯康
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2023-02-02
Filing date: 2023-02-02
Publication date: 2023-10-27

Abstract

The application provides an advertisement recommendation method, device, equipment and medium, relating to the field of machine learning, wherein the method comprises the following steps: extracting a text label set corresponding to the advertisement text, wherein the text label set is used for representing keywords of the content of the advertisement text; calculating a first probability distribution and a second probability distribution corresponding to each text label in the text label set, wherein the first probability distribution is used for representing the clicking probability distribution of the accounts with the text label set on different types of advertisement texts, and the second probability distribution is used for representing the clicking probability distribution of the whole accounts on different types of advertisement texts; updating the text label set according to the first probability distribution and the second probability distribution to obtain an updated text label set; and obtaining recommended advertisements according to the updated text label set. The method can highlight the preference of the user, so that the updated text label set is more accurate in describing the portrait of the user, and the recommended advertisement is more suitable for the user.

Description

Advertisement recommendation method, device, equipment and medium

Technical Field

The present application relates to the field of machine learning, and in particular, to an advertisement recommendation method, apparatus, device, and medium.

Background

With the development of media, advertisements become an integral part of modern life, and how to recommend advertisements that they need to users is an important activity.

In the prior art, after a user reads an advertisement text, a system generates a user tag according to keywords of the advertisement text, and the generated user tag is input into an advertisement recommendation model to search recommended advertisements for the user. For example, when a user reads a text of a reservation-related advertisement, the user tag generated by the system is a reservation, a master, english and education system, and the system obtains the advertisement related to overseas reservation according to the user tag.

In the prior art, when the user tag is extracted, the obtained user tag is not accurate enough, and the extracted user tag cannot be ensured to be consistent with the advertisement service scene. For example, in the above example, the "educational regime" does not have sufficient differentiation within the educational field that the user tab has no instructive effect on the advertisement recommendation of educational products. Therefore, the user tag obtained in the prior art can affect the accuracy of advertisement recommendation.

Disclosure of Invention

The embodiment of the application provides an advertisement recommendation method, device, equipment and medium. The technical scheme comprises the following steps:

According to an aspect of the present application, there is provided an advertisement recommendation method, the method including:

extracting a text label set corresponding to an advertisement text, wherein the text label set is used for representing keywords of the content of the advertisement text;

calculating a first probability distribution and a second probability distribution corresponding to each text label in the text label set, wherein the first probability distribution is used for representing the click probability distribution of accounts with the text label set on different types of advertisement texts, and the second probability distribution is used for representing the click probability distribution of all accounts on different types of advertisement texts;

updating the text label set according to the first probability distribution and the second probability distribution to obtain an updated text label set;

and obtaining recommended advertisements according to the updated text label set.

According to another aspect of the present application, there is provided an advertisement recommendation apparatus including:

the extraction module is used for extracting a text label set corresponding to the advertisement text, wherein the text label set is used for representing keywords of the content of the advertisement text;

the computing module is used for computing a first probability distribution and a second probability distribution corresponding to each text label in the text label set, wherein the first probability distribution is used for representing the click probability distribution of accounts with the text label set on different types of advertisement texts, and the second probability distribution is used for representing the click probability distribution of all accounts on different types of advertisement texts;

The updating module is used for updating the text label set according to the first probability distribution and the second probability distribution to obtain an updated text label set;

and the recommending module is used for obtaining recommended advertisements according to the updated text label set.

In an alternative design, the updating module is further configured to calculate a behavior distribution difference according to a difference between the first probability distribution and the second probability distribution, where the behavior distribution difference is used to represent a click behavior difference between the account having the text label set and the all accounts; and removing the text labels with the behavior distribution differences smaller than a behavior difference distribution threshold value in the text label set to obtain the updated text label set.

In an optional design, the updating module is further configured to calculate a relative entropy of the first probability distribution and the second probability distribution, so as to obtain the behavior distribution difference; alternatively, the degree of deviation between the first probability distribution and the second probability distribution is calculated by using chi-square test, and the behavior distribution difference is obtained.

In an optional design, the computing module is further configured to, for a kth type advertisement text corresponding to the text label set, count a number of first accounts for clicking the kth type advertisement text and browsing text including a jth text label, where k, j is a positive integer; counting the number of second accounts for browsing the text containing the j-th text label; and obtaining the first probability distribution of the kth type advertisement text according to the ratio of the first account number to the second account number.

In an optional design, the computing module is further configured to count, for a kth type advertisement text corresponding to the text label set, a third account number of clicking the kth type advertisement text; and obtaining the second probability distribution of the kth type advertisement text according to the ratio of the number of the third accounts to the total number of the accounts.

In an optional design, the extracting module is further configured to encode text characters in the advertisement text to obtain feature vectors corresponding to the text characters; and predicting the position of the text label in the advertisement text according to the feature vector to obtain the text label set.

In an optional design, the extracting module is further configured to classify the feature vector to obtain a position probability of the text character; extracting the text labels from the advertisement text according to the position probability to obtain the text label set; the position probabilities include at least one of a first position probability, a second position probability and a third position probability, wherein the first position probability is used for representing the probability that the text character is located at the starting position of the text label, the second position probability is used for representing the probability that the text character is located at the middle position of the text label, and the third position probability is used for representing the probability that the text character does not belong to the text label.

In an optional design, the recommendation module is further configured to invoke a recall model, and perform screening processing on the candidate advertisement set according to the updated text label set to obtain a screened advertisement set; calling a sorting model, and sorting the screening advertisement set according to the updated text label set to obtain a recommended advertisement sorting list; and acquiring the recommended advertisements from the recommended advertisement ordered list.

In an alternative design, the recall model includes p, the p representing the number of elements of the updated set of text labels; the recommending module is further used for taking the q-th text label in the updated text label set, wherein q is an integer with an initial value of 1; invoking the q-th recall model, and performing coding operation on the q-th text label and the candidate advertisement set to obtain a q-th label characteristic and a candidate advertisement characteristic; calculating a first similarity between the q-th tag feature and the candidate advertisement feature; placing the candidate advertisements, of which the first similarity is smaller than a first similarity threshold value, into the screening advertisement set; and updating q to q+1, and repeating the four steps p times to obtain the screening advertisement set.

In an optional design, the recommendation module is further configured to invoke the ranking model, and perform coding operation on the updated text tag set and the screening advertisement set to obtain tag features and candidate advertisement features; calculating a second similarity between the tag feature and each of the candidate advertisement features; and sequencing the recommended advertisements according to the second similarity to obtain the recommended advertisement sequencing.

According to another aspect of the present application, there is provided a computer apparatus comprising: a processor and a memory having stored therein at least one instruction, at least one program, code set or instruction set that is loaded and executed by the processor to implement the advertisement recommendation method as described in the above aspect.

According to another aspect of the present application, there is provided a computer storage medium having stored therein at least one program code loaded and executed by a processor to implement the advertisement recommendation method as described in the above aspect.

According to another aspect of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, the processor executing the computer instructions, causing the computer device to perform the advertisement recommendation method as described in the above aspect.

The technical scheme provided by the embodiment of the application has the beneficial effects that at least:

according to the scheme, the first probability distribution and the second probability distribution of each text label in the text label set are obtained, the text label set is updated according to the first probability distribution and the second probability distribution, and the updated text label set is used for obtaining recommended advertisements. As the first probability distribution reflects the preference of the user, the second probability distribution reflects the preference of the whole crowd, the difference between the first probability distribution and the second probability distribution can reflect the difference between the user and the whole crowd, the preference of the user is highlighted, the updated text label set is more accurate in describing the portrait of the user, and the recommended advertisement is more suitable for the user.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic diagram of an advertisement recommendation method according to an embodiment of the present application;

FIG. 2 is a schematic diagram of a computer system according to an embodiment of the present application;

FIG. 3 is a schematic flow chart of an advertisement recommendation method according to an embodiment of the present application;

FIG. 4 is a schematic diagram of keyword extraction according to an embodiment of the present application;

FIG. 5 is a schematic flow chart of an advertisement recommendation method according to an embodiment of the present application;

FIG. 6 is a schematic flow chart of a recall model calling method according to an embodiment of the present application;

FIG. 7 is a schematic flow chart of a method for calling an ordering model according to an embodiment of the present application;

FIG. 8 is a schematic diagram illustrating a structure of an advertisement recommendation apparatus according to an embodiment of the present application;

fig. 9 shows a schematic diagram of a computer device according to an embodiment of the present application.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present application more apparent, the embodiments of the present application will be described in further detail with reference to the accompanying drawings.

First, the nouns involved in the embodiments of the present application will be described:

artificial intelligence (Artificial Intelligence, AI): theory, methods, techniques and application systems that utilize digital computers or digital computer-controlled machines to simulate, extend and extend human intelligence, sense the environment, acquire knowledge and use knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision.

The artificial intelligence technology is a comprehensive subject, and relates to the technology with wide fields, namely the technology with a hardware level and the technology with a software level. Artificial intelligence infrastructure technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and other directions.

Machine Learning (ML): is a multi-domain interdisciplinary, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.

It should be noted that, before and during the process of collecting the relevant data of the user, the present application may display a prompt interface, a popup window or output voice prompt information, where the prompt interface, popup window or voice prompt information is used to prompt the user to collect the relevant data currently, so that the present application only starts to execute the relevant step of obtaining the relevant data of the user after obtaining the confirmation operation of the user to the prompt interface or popup window, otherwise (i.e. when the confirmation operation of the user to the prompt interface or popup window is not obtained), the relevant step of obtaining the relevant data of the user is finished, i.e. the relevant data of the user is not obtained. In other words, all user data collected by the present application is collected with the consent and authorization of the user, and the collection, use and processing of relevant user data requires compliance with relevant laws and regulations and standards of the relevant country and region.

The scheme can recommend advertisements to the user in real time through clicking or reading actions of the user on the advertisement text. After clicking on the advertisement text 102, the user 101 acquires a text label set 103 of the advertisement text 102, and the text label set 103 represents keywords of the content of the advertisement text 102. After that, the first type account 104 and the total account 105 are determined, and the first type account 104 refers to an account having the text label set 103, where the text label set 103 also plays a role of portrait for the user. Then, a first probability distribution 106 and a second probability distribution 107 are calculated, wherein the first probability distribution 106 refers to the probability distribution of clicking of the first type account 104 on different types of advertisement texts, and the second probability distribution 107 refers to the probability distribution of clicking of the whole account 105 on different types of advertisement texts. Behavior distribution differences 108 are calculated based on the first probability distribution 106 and the second probability distribution 107. When the behavior distribution difference 108 is sufficiently large, the behavior distribution difference 108 may represent a difference between the advertisement click behavior of the first type account 104 and the advertisement click behavior of the overall account 105, that is, the advertisement click behavior of the account having the text label set 103 has a tendency, and the text label set 103 has a strong association with the behavior of clicking the advertisement text 102, so that the text label set 103 can explicitly characterize the characteristics possessed by the user 101 who clicks the advertisement text 102. And removing the text labels with the behavior distribution differences 108 smaller than the behavior difference distribution threshold value in the text label set 103 to obtain an updated text label set 109. The recommended advertisements 111 are derived by the advertisement recommendation model 110 and the updated text label collection 109. The advertisement recommendation model 110 comprises a recall model and a sort model, wherein the recall model performs rough screening, and single text labels in the updated text label set 109 are used for screening to obtain a screened advertisement set; the ranking model performs a finer screening using all text labels in the updated set of text labels 109 to obtain recommended advertisements 111.

Fig. 2 shows a schematic diagram of a computer system according to an exemplary embodiment of the present application. The computer system 200 includes: a terminal 220 and a server 240.

The terminal 220 has installed thereon an application program that is associated with advertisement presentation. In some embodiments, the foregoing model employs at least one of a TSFM (Temporal Spatial Fusion Model, space-time fusion model), LSTM (Long Short-Term Memory structure), DNN (Deep Neural Networks, deep neural network), CNN (Convolutional Neural Networks, convolutional neural network). The application may be an applet in an app, a specialized application, or a web client. The terminal 220 is at least one of a smart phone, a tablet computer, an electronic book reader, an MP3 player, an MP4 player, a laptop portable computer, and a desktop computer.

The terminal 220 is connected to the server 240 through a wireless network or a wired network.

The server 240 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or may be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs (Content Delivery Network, content delivery networks), basic cloud computing services such as big data and artificial intelligence platforms, and the like. Server 240 is used to provide background services for model training. Optionally, the server 240 takes on primary computing work and the terminal 220 takes on secondary computing work; alternatively, the server 240 takes on secondary computing work and the terminal 220 takes on primary computing work; alternatively, both the server 240 and the terminal 220 perform collaborative computing using a distributed computing architecture.

Fig. 3 shows a flowchart of an advertisement recommendation method according to an embodiment of the present application. The method may be performed by the computer system shown in fig. 2. The method comprises the following steps:

step 302: and extracting a text label set corresponding to the advertisement text, wherein the text label set is used for representing keywords of the content of the advertisement text.

Alternatively, the advertisement text is an advertisement that the user clicks or reads, and the candidate advertisement text is taken as the advertisement text used in this step in response to a click operation on the candidate advertisement text. Alternatively, the candidate advertisement text is taken as the advertisement text used in this step in response to a click operation on the candidate advertisement text and the dwell time of the candidate advertisement text is greater than the upper time limit.

The text label set includes at least one text label.

Optionally, coding text characters in the advertisement text to obtain feature vectors corresponding to the text characters; and predicting the position of the text label in the advertisement text according to the feature vector to obtain a text label set. In some embodiments, extraction of the text label set is accomplished using a BERT (Bidirectional Encoder Representation from Transformers, bi-directionally encoded representation from the encoder) model.

Optionally, the feature vector corresponding to the text character is determined by a vector lookup table. The vector lookup table is used for representing the corresponding relation between the text characters and the feature vectors. For example, the vector look-up table records the character "he" corresponding feature vector (1, 5), and the character "low" corresponding feature vector (4, 54, 69).

Optionally, classifying the feature vectors to obtain the position probability of the text characters; extracting text labels from advertisement texts according to the position probability to obtain a text label set; the position probabilities include at least one of a first position probability, a second position probability and a third position probability, the first position probability is used for representing the probability that the text character is located at the starting position of the text label, the second position probability is used for representing the probability that the text character is located at the middle position of the text label, and the third position probability is used for representing the probability that the text character does not belong to the text label. Optionally, the type of text character is determined according to the maximum value of the first position probability, the second position probability and the third position probability. For example, if the probability of the second position corresponding to the text character "learn" is the maximum value, then "learn" is considered to be located in the middle of the text label.

Illustratively, the position probability is obtained by the following calculation:

x _i ＝LP(emb _i )；

p _i ＝oftmax(x _i )；

wherein emb _i Representing the corresponding feature vector of the ith text character, MLP (x) represents a multi-layer perceptron, p _i Representing the position probability of the ith text character, softmax represents the normalization function.

In some embodiments, please refer to fig. 4, when the advertisement text is "create personalized leave-on solution for you, there are no doubt benefits in reading foreign books. The first is school … … ", where" leave-on "is described by way of example, where" left-on "is predicted as" B "indicating that the character is the starting position of the text label," learn "is predicted as" I "indicating that the character is the intermediate position of the text label," square "and" case "are predicted as" O "indicating that the character does not belong to the text label.

In some embodiments, text labels may also be used to represent click preferences or preferences of the account. By way of example, the text label includes "education", and the user corresponding to the account is stated to prefer to browse the articles of the education class, so that the user has greater click will on the articles of the education class.

Step 304: and calculating a first probability distribution and a second probability distribution corresponding to each text label in the text label set, wherein the first probability distribution is used for representing the clicking probability distribution of the account with the text label set on different types of advertisement texts, and the second probability distribution is used for representing the clicking probability distribution of the whole account on different types of advertisement texts.

The accounts corresponding to the first probability distribution all have text label sets. The total accounts in the second probability distribution refer to all accounts stored by the computer system.

Optionally, the first probability distribution is in a discrete distribution form, an abscissa of the first probability distribution represents text labels in the text label set, and an ordinate of the first probability distribution represents click probabilities of accounts having the text label set on advertisement text having the text labels.

Optionally, the second probability distribution is in a discrete distribution form, an abscissa of the second probability distribution represents the text labels in the text label set, and an ordinate of the second probability distribution represents the click probability of the whole account on the advertisement text with the text labels.

In some embodiments, there is a clear distinction between the first probability distribution and the second probability distribution. For the same text label a, the difference between the click probability corresponding to the first probability distribution and the click probability corresponding to the second probability distribution is larger than the preset probability difference.

Optionally, for the kth type advertisement text corresponding to the text label set, counting the number of first accounts for clicking the kth type advertisement text and browsing the text containing the jth text label, wherein k and j are positive integers; counting the number of second accounts for browsing the text containing the j text label; and obtaining the first probability distribution of the advertisement text of the kth type according to the ratio of the number of the first accounts to the number of the second accounts.

The first accounts corresponding to the first account number need to meet two conditions simultaneously: 1. the user corresponding to the account clicks on the kth type of advertising text within a first preset time period, which may be set by the technician himself (e.g., the first preset time period is set to the past 7 days); 2. the account number contains the j-th text label. The second account corresponding to the number of the second accounts refers to the account requiring the inclusion of the j-th text label.

Optionally, counting the number of third accounts clicking the kth type advertisement text for the kth type advertisement text corresponding to the text label set; and obtaining a second probability distribution of the advertisement text of the kth type according to the ratio of the number of the third accounts to the total number of the accounts.

The third accounts corresponding to the third account number click the k type advertisement text in a second preset time period, and the second preset time period can be set by a technician.

Step 306: and updating the text label set according to the first probability distribution and the second probability distribution to obtain an updated text label set.

Optionally, updating the text label set according to the difference between the first probability distribution and the second probability distribution to obtain an updated text label set. The difference value is used to represent the distance between the first probability distribution and the second probability distribution.

Illustratively, the behavior distribution difference is calculated from the difference between the first probability distribution and the second probability distribution. And removing text labels with behavior distribution differences smaller than a behavior difference distribution threshold value in the text label set to obtain an updated text label set. The behavior difference distribution is used to measure the distance between the first probability distribution and the second probability distribution. When the behavior difference distribution score of the tag is higher than the behavior difference distribution threshold, the fact that the tag and the advertisement clicking behaviors have a stronger association relation is indicated, and the corresponding advertisement is recommended to enable the user to have stronger clicking will.

Step 308: and obtaining recommended advertisements according to the updated text label set.

Optionally, a recommended advertisement is determined from the candidate advertisement set based on the updated set of text labels. Since in practical situations there are a large number of advertisements in the candidate advertisement set, how to select recommended advertisements among the large number of advertisements is one aspect that needs to be considered by the present application. In the embodiment of the application, recommended advertisements are obtained by adopting secondary screening, and in the primary screening, a screening advertisement set is determined from a candidate advertisement set; in the second filtering, recommended advertisements are determined from the set of filtered advertisements. The first screening carries out rough screening on the candidate advertisement sets to obtain screened advertisement sets, and the first screening can reduce ten thousand counts of candidate advertisement sets to hundred counts of screened advertisement sets; and carrying out one-time detailed screening on the screening advertisement set for the second screening to obtain recommended advertisements.

In the embodiment of the application, a recall ordering model is adopted to process the candidate advertisement sets, and the recall ordering model comprises two parts, namely a recall model and an ordering model. And calling a recall model during the first screening, and screening the candidate advertisement sets according to the updated text label sets to obtain screened advertisement sets. In some embodiments, the recall models include q, each recall model for processing a different type of text label.

In the second screening, calling a sequencing model, and sequencing the screened advertisement set according to the updated text label set to obtain a recommended advertisement sequencing list; and acquiring the recommended advertisements from the recommended advertisement ordered list. The ranked list of recommended advertisements is used to represent the degree of recommendation for each of the screening advertisements in the set of screening advertisements. Optionally, the recommended advertisements are obtained according to the recommendation degree of each screening advertisement in the recommended advertisement sequencing list.

In summary, the embodiment obtains the first probability distribution and the second probability distribution of each text label in the text label set, updates the text label set according to the first probability distribution and the second probability distribution, and obtains the recommended advertisement by using the updated text label set. As the first probability distribution reflects the preference of the user, the second probability distribution reflects the preference of the whole crowd, the difference between the first probability distribution and the second probability distribution can reflect the difference between the user and the whole crowd, the preference of the user is highlighted, the updated text label set is more accurate in describing the portrait of the user, and the recommended advertisement is more suitable for the user.

In the following embodiments, it is necessary to calculate the difference between the first probability distribution and the second probability distribution, resulting in a behavior distribution difference. If the difference of the behavior distribution is large enough, the difference of the two distributions of the first probability distribution and the second probability distribution is large, and the advertisement clicking behavior and the overall crowd difference of the users corresponding to the accounts defined by a certain label are large. The label and the advertisement clicking behavior have stronger association relation, and the user can have stronger clicking willingness by recommending the corresponding advertisement.

Fig. 5 shows a flowchart of an advertisement recommendation method according to an embodiment of the present application. The method may be performed by the computer system shown in fig. 2. The method comprises the following steps:

step 501: and extracting a text label set corresponding to the advertisement text.

Wherein the text label set is used for representing keywords of the content of the advertisement text.

The text label set includes at least one text label.

Optionally, coding text characters in the advertisement text to obtain feature vectors corresponding to the text characters; and predicting the position of the text label in the advertisement text according to the feature vector to obtain a text label set.

Optionally, coding text characters in the advertisement text to obtain feature vectors corresponding to the text characters; and predicting the position of the text label in the advertisement text according to the feature vector to obtain a text label set. In some embodiments, extraction of the text label set is accomplished using a BERT model.

Optionally, the feature vector corresponding to the text character is determined by a vector lookup table. The vector lookup table is used for representing the corresponding relation between the text characters and the feature vectors.

Optionally, classifying the feature vectors to obtain the position probability of the text characters; and extracting text labels from the advertisement text according to the position probability to obtain a text label set.

The position probabilities include at least one of a first position probability, a second position probability and a third position probability, the first position probability is used for representing the probability that the text character is located at the starting position of the text label, the second position probability is used for representing the probability that the text character is located at the middle position of the text label, and the third position probability is used for representing the probability that the text character does not belong to the text label.

Step 502: and calculating a first probability distribution corresponding to each text label in the text label set.

Illustratively, a triplet (u _i ,t _j ,a _k ) Recording clicking action of account number, u _i Represents the ith account number, t _j Represents the j-th text label, a _k Representing the kth type of advertising text. The tripletCan be regarded as an account u _i The corresponding user reads a text-containing tag t _j While the user also clicks on the advertisement text belonging to a _k Category text advertisements. The first probability distribution is calculated as follows:

Wherein, sigma _i Count(u _i ,t _j ,a _k ) Representing simultaneous click of a _k Category text advertisement and simultaneous reading of text advertisement containing tag t _j Account number, sigma of advertisement text _i,k Count(u _i ,t _j ,a _k ) Then it means that the containing tag t is read _j Account number of advertisement text.

Step 503: and calculating a second probability distribution corresponding to each text label in the text label set.

Illustratively, a triplet (u _i ,t _j ,a _k ) Recording the clicking behavior of the account, and calculating the second probability distribution as follows:

wherein, sigma _i,j,k Count(u _i ,t _j ,a _k ) Representing triples (u) _i ,t _j ,a _k ) All numbers, sigma _i,j,k Count(u _i ,t _j ,a _k ) Also represents the total number of accounts, Σ _i,j Count(u _i ,t _j ,a _k ) Indicating click of a _k Account number of text advertisements for a category. u (u) _i Represents the ith account number, t _j Represents the j-th text label, a _k Representing the kth type of advertising text.

Step 504: the behavior distribution difference is calculated from the difference between the first probability distribution and the second probability distribution.

Optionally, the relative entropy of the first probability distribution and the second probability distribution is calculated to obtain the behavior distribution difference.

Illustratively, if the first probability distribution is p and the second probability distribution is q, the behavior distribution is differentiated score (t _i ) The following formula can be used for calculation:

wherein KL (x) represents the calculated relative entropy, u _i Represents the ith account number, t _j Represents the j-th text label, a _k Representing the kth type of advertising text.

Alternatively, the degree of deviation between the first probability distribution and the second probability distribution is calculated using chi-square test, resulting in a behavior distribution difference. The chi-square test is to count the deviation degree between the actual observed value and the theoretical inferred value of the sample, the deviation degree between the actual observed value and the theoretical inferred value determines the chi-square value, and if the chi-square value is larger, the deviation degree of the actual observed value and the theoretical inferred value is larger; conversely, the smaller the deviation of the two; if the two values are completely equal, the chi-square value is 0, indicating that the theoretical value is completely in line.

Step 505: and removing text labels with behavior distribution differences smaller than a behavior difference distribution threshold value in the text label set to obtain an updated text label set.

The behavior difference distribution measures the distance between the first probability distribution and the second probability distribution, and if the two distributions are large in difference, the advertisement clicking behaviors and the overall crowd difference of users corresponding to the accounts defined by a certain label are large. That is, the tag has a sufficient degree of differentiation for the user's behavior on the advertising system. Therefore, when the behavior difference distribution score of the tag is higher than the behavior difference distribution threshold, the tag and the advertisement clicking behavior have stronger association relation, and the corresponding advertisement is recommended to have stronger clicking will for the user.

Step 506: and obtaining recommended advertisements according to the updated text label set.

Optionally, a recommended advertisement is determined from the candidate advertisement set based on the updated set of text labels. In the embodiment of the application, recommended advertisements are obtained by adopting secondary screening, and in the primary screening, a screening advertisement set is determined from a candidate advertisement set; in the second filtering, recommended advertisements are determined from the set of filtered advertisements. The first screening carries out rough screening on the candidate advertisement sets to obtain screened advertisement sets, and the first screening can reduce ten thousand counts of candidate advertisement sets to hundred counts of screened advertisement sets; and carrying out one-time detailed screening on the screening advertisement set for the second screening to obtain recommended advertisements.

And calling a recall model during the first screening, and screening the candidate advertisement sets according to the updated text label sets to obtain screened advertisement sets. In some embodiments, the recall models include q, each recall model for processing a different type of text label.

When the difference of the behavior distribution is large enough, the two distributions of the first probability distribution and the second probability distribution are large, and the advertisement clicking behaviors and the overall crowd difference of the users corresponding to the accounts defined by a certain label are large. The label and the advertisement clicking behavior have stronger association relation, and the user can have stronger clicking willingness by recommending the corresponding advertisement.

In the following example, determining recommended advertisements needs to be done in two passes, the first being a coarse screen and the second being a fine screen. The first coarse screening can reduce the range of selecting recommended advertisements, and the second fine screening can realize finer screening to select more proper recommended advertisements. The step of determining recommended advertisements is as follows:

1. and calling a recall model, and screening the candidate advertisement sets according to the updated text label set to obtain screened advertisement sets.

Step 601: and taking the q-th text label in the updated text label set, wherein q is an integer with an initial value of 1.

Wherein the recall model includes p, p representing the number of elements of the updated text label set.

Note that the number of the q-th text labels is smaller than p. Illustratively, when p is 5, q may be 1 or 2. In the following embodiment, the q-th text label is considered to include only one text label.

Step 602: and calling a q recall model, and performing coding operation on the q text labels and the candidate advertisement sets to obtain q label features and candidate advertisement features.

Optionally, calling the q recall model, and performing coding operation on the q text label to obtain the q label feature. Illustratively, the q-th recall model is invoked, and the vector of each character in the q-th text label is determined through a vector lookup table, so that the q-th label feature is obtained. The vector lookup table is used for recording the corresponding relation between the characters and the vectors. For example, the vector corresponding to the character "pair" is [4,8], and the vector corresponding to the character "square" is [8, 89].

Optionally, calling a q-th recall model, and performing coding operation on the candidate advertisement set to obtain candidate advertisement characteristics. The candidate advertisement features include at least one feature. Similarly, the q-th recall model is called, and the vector of each character in the candidate advertisement set is determined through the vector lookup table, so that the candidate advertisement characteristics are obtained.

In some embodiments, the recall model includes a first encoder for representing a mapping relationship between text symbols and feature vectors. For example, the mapping relationship includes the text symbol "up" and feature vector [45, 48].

Step 603: a first similarity between the q-th tag feature and the candidate advertisement feature is calculated.

The first similarity is used to represent a vector distance between the qth tag feature and the candidate advertisement feature.

Illustratively, the first similarity is calculated by the Euclidean distance. For example, the q-th tag feature is (1,2,0), the candidate advertisement features include (4, 2, 4) and (1, 2, 3), and two euclidean distances of 5 and 3 are obtained, with a first similarity of 5+3=8.

Step 604: and placing the candidate advertisements with the first similarity smaller than the first similarity threshold value in the screening advertisement set.

The first similarity threshold may be set by the skilled person at his own discretion. The first similarity threshold is a constant.

The screening advertisement set is used for storing candidate advertisements meeting the requirement. In some embodiments, the set of screening advertisements stores at least one of a name, a number, a URL (Uniform Resource Locator ), an author, a type, a category, and a kind of screening advertisement.

Step 605: and updating q to q+1, and repeating the four steps p times to obtain the screening advertisement set.

Repeating the four steps p times, all text labels in the updated text label set can be traversed. And further determines the required screening advertisement from all text labels.

2. Calling a sorting model, and sorting the screened advertisement set according to the updated text label set to obtain a recommended advertisement sorting list; and acquiring the recommended advertisements from the recommended advertisement ordered list.

Step 701: and calling a sequencing model, and performing coding operation on the updated text label set and the screening advertisement set to obtain label characteristics and candidate advertisement characteristics.

Optionally, calling the ordering model, and performing coding operation on the updated text label set to obtain label characteristics. Note that, this step uses all the tags in the updated text tag set. The tag features include at least one feature.

Optionally, calling a sequencing model, and performing coding operation on the screening advertisement set to obtain candidate advertisement characteristics.

In some embodiments, the ordering model includes a second encoder for representing a mapping relationship between text symbols and feature vectors.

Step 702: a second similarity between the tag characteristic and each of the candidate advertisement characteristics is calculated.

The second similarity is used to represent a vector distance between the tag feature and each candidate advertisement feature.

Optionally, the second similarity is calculated by euclidean distance. For example, if the tag signature is (1,2,0) and the candidate advertisement signature includes (6, 2, 12) and (1, 2, 3), then two euclidean distances 13 and 3 are obtained, and the first similarity is 13+3=16.

Step 703: and sorting the screening candidate advertisements according to the second similarity to obtain recommended advertisement sorting.

In some embodiments, the filtered advertisements are ranked from large to small according to the second similarity, resulting in a recommended advertisement ranking; the first a screening advertisements are taken as recommended advertisements, a is a positive integer, and the value of a can be set by a technician.

In summary, the embodiment of the application adopts the secondary screening to select the recommended advertisements, so that the advertisement screening efficiency can be improved, and the recommended advertisements are more suitable for the preference of the user.

In the following embodiment, a method of updating a text label, which can accurately depict a portrait of a user, will be described. And providing proper recommended advertisements for the users, providing favorite or preferential contents for the users, and improving the click rate of the users on the recommended advertisements.

Step 1: acquiring a text label;

the method comprises the steps of encoding text characters of advertisement texts, finally obtaining characterization of each text character, and finally predicting the text segment position of each advertisement text. Then splice two special symbols "[ CLS ] [ SEP ]". After the processing is completed, the text character sequence can be sent to a BERT model, and the BERT model obtains the feature vector of each text character. The probability that each text character is predicted as a BIO is then obtained by the following formula, where B represents the starting position of the text label to which the corresponding text character belongs, I represents the intermediate position of the text label to which the corresponding text character belongs, and O represents that the corresponding text character does not belong to a portion of the text label:

x _i ＝MLP(emb _i )；

p _i ＝softmax(x _i )；

Wherein emb _i Representing feature vectors of the ith text character in the bert model, MLP (x) representing the multi-layer perceptron, p _i Representing the probability that the ith text character is predicted as a BIO.

Step 2: calculating clicking behavior distribution of advertisement texts under each label account;

after the labels of each article in the information flow scene are obtained, a preliminary label system T can be constructed. In addition, a triplet (u) can be generated by the browsing behavior of the user in the information flow scene and the clicking behavior in the advertisement system _i ,t _j ,a _k ). Wherein u is _i Represents the ith user, t _j Representing the jth tag, a, in the original information stream tag system _k Representing the kth advertising category defined by the advertising system. The meaning of the triplet is that user u, within a certain time window (e.g. last 90 days) _i Reads a certain inclusion tag t _j While he also clicks on the information stream article belonging to a _k An advertisement for a category. Based on the user's reading behavior in the information flow scene and clicking behavior in the advertisement system, we can generate such triples for all users. Based on a large number of triple co-occurrence relations, we can count the information flow label t _j Advertisement category a under a limited group of people _k Probability of clicked:

wherein, sigma _i Count(u _i ,t _j ,a _k ) Representing simultaneous click advertisement category a _k And at the same time read the tag t _j Account number, sigma of advertisement text _i,k Count(u _i ,t _j ,a _k ) Then it means that the containing tag t is read _j Account number of advertisement text. u (u) _i Represents the ith user, t _j Representing the jth tag, a, in the original information stream tag system _k Representing the kth advertising category defined by the advertising system. Based on the formula, the click probability distribution of the user on different types of advertisements under each information flow article label can be obtained.

Step 3, calculating clicking behavior distribution of advertisement text under the whole account;

similar to step 2, the click distribution of the whole crowd under the advertisement can be obtained by adopting the following formula:

wherein, sigma _i,j,k Count(u _i ,t _j ,a _k ) Representing triples (u) _i ,t _j ,a _k ) All numbers, sigma _i,j Count(u _i ,t _j ,a _k ) Indicating that advertisement category a is clicked _k Is a number of users of the system. q (a) _k ) The prior probability of different category advertisements being clicked in the advertisement system is represented. u (u) _i Represents the ith user, t _j Representing the jth tag, a, in the original information stream tag system _k Representing the kth advertising category defined by the advertising system.

Step 4, counting the behavior distribution difference of the label account and the whole account;

as shown in fig. 3, step 2 is performed by the tag t _i The limiting population gets the probability p (a _k |t _j ) Step 3, obtaining the prior probability distribution q (a) of each advertisement category clicked on without any limitation _k ). Finally, we derive the importance of each tag by the scoring function:

wherein score (t _i ) The distance between the two distributions of p and q is measured, and if the two distributions differ greatly, this indicates that for the tag t _i The defined users have large differences in their advertisement click behavior and overall crowd. That is, the tag t _i And has enough differentiation degree for the behavior of the user on the advertising system. Therefore, when the label t _i The importance score of (2) is higher than a certain threshold value, which indicates that a stronger association exists between the tag and the advertisement clicking behavior. For the initial tag system T in step 1, if T is _i The score of (2) is larger and remains, otherwise, is filtered out. The filtered set of tags will be used for characterization of the user's features in the advertisement recommendation.

Fig. 8 shows a schematic diagram of an advertisement recommendation device according to an embodiment of the present application. The apparatus 800 includes:

an extracting module 801, configured to extract a text label set corresponding to an advertisement text, where the text label set is used to represent keywords of content of the advertisement text;

a calculation module 802, configured to calculate a first probability distribution and a second probability distribution corresponding to each text label in the text label set, where the first probability distribution is used to represent a probability distribution of clicking of an account with the text label set on different types of advertisement texts, and the second probability distribution is used to represent a probability distribution of clicking of a whole account on different types of advertisement texts;

an updating module 803, configured to update the text label set according to the first probability distribution and the second probability distribution, to obtain an updated text label set;

And a recommending module 804, configured to obtain a recommended advertisement according to the updated text label set.

In an alternative design, the updating module 803 is further configured to calculate a behavior distribution difference according to a difference between the first probability distribution and the second probability distribution, where the behavior distribution difference is used to represent a click behavior difference between the account having the text label set and the overall account; and removing the text labels with the behavior distribution differences smaller than a behavior difference distribution threshold value in the text label set to obtain the updated text label set.

In an alternative design, the updating module 803 is further configured to calculate a relative entropy of the first probability distribution and the second probability distribution, to obtain the behavior distribution difference; alternatively, the degree of deviation between the first probability distribution and the second probability distribution is calculated by using chi-square test, and the behavior distribution difference is obtained.

In an alternative design, the calculating module 802 is further configured to, for a kth type advertisement text corresponding to the text label set, count a number of first accounts that click on the kth type advertisement text and browse text including a jth text label, where k, j is a positive integer; counting the number of second accounts for browsing the text containing the j-th text label; and obtaining the first probability distribution of the kth type advertisement text according to the ratio of the first account number to the second account number.

In an alternative design, the calculating module 802 is further configured to count, for a kth text tag in the text tag set, a third account number of clicking on a kth type advertisement text; and obtaining the second probability distribution of the kth text label according to the ratio of the number of the third accounts to the total number of the accounts.

In an optional design, the extracting module 801 is further configured to count, for a kth type advertisement text corresponding to the text label set, a third account number of clicking the kth type advertisement text; and obtaining the second probability distribution of the kth type advertisement text according to the ratio of the number of the third accounts to the total number of the accounts.

In an optional design, the extracting module 801 is further configured to classify the feature vector to obtain a position probability of the text character; extracting the text labels from the advertisement text according to the position probability to obtain the text label set; the position probabilities include at least one of a first position probability, a second position probability and a third position probability, wherein the first position probability is used for representing the probability that the text character is located at the starting position of the text label, the second position probability is used for representing the probability that the text character is located at the middle position of the text label, and the third position probability is used for representing the probability that the text character does not belong to the text label.

In an optional design, the recommendation module 804 is further configured to invoke a recall model, and perform screening processing on the candidate advertisement set according to the updated text label set to obtain a screened advertisement set; calling a sorting model, and sorting the screening advertisement set according to the updated text label set to obtain a recommended advertisement sorting list; and acquiring the recommended advertisements from the recommended advertisement ordered list.

In an alternative design, the recall model includes p, the p representing the number of elements of the updated set of text labels; the recommending module 804 is further configured to take a q-th text tag in the updated text tag set, where q is an integer with an initial value of 1; invoking the q-th recall model, and performing coding operation on the q-th text label and the candidate advertisement set to obtain a q-th label characteristic and a candidate advertisement characteristic; calculating a first similarity between the q-th tag feature and the candidate advertisement feature; placing the candidate advertisements, of which the first similarity is smaller than a first similarity threshold value, into the screening advertisement set; and updating q to q+1, and repeating the four steps p times to obtain the screening advertisement set.

In an optional design, the recommendation module 804 is further configured to invoke the ranking model, and perform an encoding operation on the updated text tag set and the screening advertisement set to obtain tag features and candidate advertisement features; calculating a second similarity between the tag feature and each of the candidate advertisement features; and sequencing the recommended advertisements according to the second similarity to obtain the recommended advertisement sequencing.

Fig. 9 is a schematic diagram of a computer device, according to an example embodiment. The computer apparatus 900 includes a central processing unit (Central Processing Unit, CPU) 901, a system Memory 904 including a random access Memory (Random Access Memory, RAM) 902 and a Read-Only Memory (ROM) 903, and a system bus 905 connecting the system Memory 904 and the central processing unit 901. The computer device 900 also includes a basic Input/Output system (I/O) 906, which helps to transfer information between various devices within the computer device, and a mass storage device 907, for storing an operating system 913, application programs 914, and other program modules 915.

The basic input/output system 906 includes a display 908 for displaying information and an input device 909, such as a mouse, keyboard, etc., for user input of information. Wherein the display 908 and the input device 909 are connected to the central processing unit 901 via an input output controller 910 connected to the system bus 905. The basic input/output system 906 can also include an input/output controller 910 for receiving and processing input from a number of other devices, such as a keyboard, mouse, or electronic stylus. Similarly, the input-output controller 910 also provides output to a display screen, a printer, or other type of output device.

The mass storage device 907 is connected to the central processing unit 901 through a mass storage controller (not shown) connected to the system bus 905. The mass storage device 907 and its associated computer device-readable media provide non-volatile storage for the computer device 900. That is, the mass storage device 907 may include a computer device readable medium (not shown) such as a hard disk or a compact disk-Only (CD-ROM) drive.

The computer device readable medium may include computer device storage media and communication media without loss of generality. Computer device storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer device readable instructions, data structures, program modules or other data. Computer device storage media includes RAM, ROM, erasable programmable read-Only Memory (Erasable Programmable Read Only Memory, EPROM), electrically erasable programmable read-Only Memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), CD-ROM, digital video disk (Digital Video Disc, DVD), or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer device storage medium is not limited to the ones described above. The system memory 904 and mass storage device 907 described above may be collectively referred to as memory.

According to various embodiments of the present disclosure, the computer device 900 may also operate through a network, such as the Internet, to remote computer devices on the network. I.e., the computer device 900 may be connected to the network 912 through a network interface unit 911 coupled to the system bus 905, or other types of networks or remote computer device systems (not shown) may be coupled using the network interface unit 911.

The memory further includes one or more programs stored in the memory, and the central processor 901 implements all or part of the steps of the advertisement recommendation method by executing the one or more programs.

In an exemplary embodiment, there is also provided a computer device having stored therein at least one instruction, at least one program, a set of codes, or a set of instructions, which are loaded and executed by a processor to implement the advertisement recommendation method provided by the above-described respective method embodiments.

The application also provides a computer readable storage medium, wherein at least one instruction, at least one section of program, code set or instruction set is stored in the storage medium, and the at least one instruction, the at least one section of program, the code set or instruction set is loaded and executed by the processor to realize the advertisement recommendation method provided by the embodiment of the method.

Optionally, the present application also provides a computer program product containing instructions which, when run on a computer device, cause the computer device to perform the advertisement recommendation method of the above aspects.

The foregoing embodiment numbers of the present application are merely for the purpose of description, and do not represent the advantages or disadvantages of the embodiments.

It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program for instructing relevant hardware, where the program may be stored in a computer readable storage medium, and the storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

The foregoing description of the preferred embodiments of the present application is not intended to limit the application, but rather, the application is to be construed as limited to the appended claims.

Claims

1. An advertisement recommendation method, the method comprising:

2. The method of claim 1, wherein updating the text label set according to the first probability distribution and the second probability distribution, resulting in an updated text label set, comprises:

calculating a behavior distribution difference according to the difference between the first probability distribution and the second probability distribution, wherein the behavior distribution difference is used for representing the click behavior difference of the account with the text label set and the whole account;

and removing the text labels with the behavior distribution differences smaller than a behavior difference distribution threshold value in the text label set to obtain the updated text label set.

3. The method of claim 2, wherein said calculating a behavior distribution difference from a difference between said first probability distribution and said second probability distribution comprises:

calculating the relative entropy of the first probability distribution and the second probability distribution to obtain the behavior distribution difference;

Or alternatively, the process may be performed,

and calculating the deviation degree between the first probability distribution and the second probability distribution by using chi-square test to obtain the behavior distribution difference.

4. A method according to any one of claims 1 to 3, wherein said calculating a first probability distribution for each text label in said set of text labels comprises:

counting the number of first accounts for clicking the kth type advertisement text and browsing the text containing the jth text label for the kth type advertisement text corresponding to the text label set, wherein k and j are positive integers;

counting the number of second accounts for browsing the text containing the j-th text label;

and obtaining the first probability distribution of the kth type advertisement text according to the ratio of the first account number to the second account number.

5. A method according to any one of claims 1 to 3, wherein said calculating a second probability distribution for each text label in said set of text labels comprises:

counting the number of third accounts clicking the kth type advertisement text for the kth type advertisement text corresponding to the text label set;

and obtaining the second probability distribution of the kth type advertisement text according to the ratio of the number of the third accounts to the total number of the accounts.

6. A method according to any one of claims 1 to 3, wherein the extracting a set of text labels corresponding to advertisement text comprises:

encoding text characters in the advertisement text to obtain feature vectors corresponding to the text characters;

and predicting the position of the text label in the advertisement text according to the feature vector to obtain the text label set.

7. The method of claim 6, wherein predicting the location of text labels based on the feature vector results in the set of text labels, comprising:

classifying the feature vectors to obtain the position probability of the text characters;

extracting the text labels from the advertisement text according to the position probability to obtain the text label set;

the position probabilities include at least one of a first position probability, a second position probability and a third position probability, wherein the first position probability is used for representing the probability that the text character is located at the starting position of the text label, the second position probability is used for representing the probability that the text character is located at the middle position of the text label, and the third position probability is used for representing the probability that the text character does not belong to the text label.

8. A method according to any one of claims 1 to 3, wherein said deriving recommended advertisements from said updated set of text labels comprises:

calling a recall model, and screening candidate advertisement sets according to the updated text label set to obtain screened advertisement sets;

calling a sorting model, and sorting the screening advertisement set according to the updated text label set to obtain a recommended advertisement sorting list;

and acquiring the recommended advertisements from the recommended advertisement ordered list.

9. The method of claim 8, wherein the recall model comprises p, the p representing a number of elements of the updated set of text labels;

and calling a recall model, screening the candidate advertisement set according to the updated text label set to obtain a screened advertisement set, wherein the method comprises the following steps of:

taking the q-th text label in the updated text label set, wherein q is an integer with an initial value of 1;

invoking the q-th recall model, and performing coding operation on the q-th text label and the candidate advertisement set to obtain a q-th label characteristic and a candidate advertisement characteristic;

Calculating a first similarity between the q-th tag feature and the candidate advertisement feature;

placing the candidate advertisements, of which the first similarity is smaller than a first similarity threshold value, into the screening advertisement set;

and updating q to q+1, and repeating the four steps p times to obtain the screening advertisement set.

10. The method of claim 8, wherein the invoking the ranking model to rank the set of filtered advertisements based on the updated set of text labels to obtain a ranked list of recommended advertisements comprises:

invoking the ordering model, and performing coding operation on the updated text label set and the screening advertisement set to obtain label characteristics and candidate advertisement characteristics;

calculating a second similarity between the tag feature and each of the candidate advertisement features;

and sorting the screening advertisements according to the second similarity to obtain the recommended advertisement sorting.

11. An advertisement recommendation device, the device comprising:

12. A computer device, the computer device comprising: a processor and a memory having stored therein at least one instruction, at least one program, code set or instruction set that is loaded and executed by the processor to implement the advertisement recommendation method of any of claims 1 to 10.

13. A computer readable storage medium having stored therein at least one program code that is loaded and executed by a processor to implement the advertisement recommendation method of any one of claims 1 to 10.

14. A computer program product comprising a computer program or instructions which, when executed by a processor, implements the advertisement recommendation method of any one of claims 1 to 10.