A model training method, device, equipment, medium and program product

Abstract

Embodiments of the application disclose a model training method and device, equipment, medium and program product. The method comprises: obtaining historical advertisement data; constructing a sample set according to the historical advertisement data; training a redirection correlation model by using the sample set, and the trained redirection correlation model is used to predict redirection correlation between two advertisements; and the redirection correlation between the two advertisements is used to represent a probability that the other advertisement is triggered and converted when one of the two advertisements is triggered. The embodiments of the application can construct a sample pair to train a network model, ensure a training process of the network model, and further ensure performance of the trained network model.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to the field of artificial intelligence, and particularly to a model training method, a model training device, a computer equipment, a computer-readable storage medium, and a computer program product. Background Technology

[0002] With the development of social informatization, advertising has gradually become an important way to guide user behavior. For example, when advertising distribution platforms present ads in an immersive way in feeds (such as a continuously updated information stream that presents information to users), they can guide users to click on and / or convert to ads, thereby increasing advertising revenue. Advertising distribution platforms mainly rely on network models to realize the function of advertising distribution.

[0003] In existing technologies, the sample data used to train network models consists of multiple independent sample advertisements. This limits the trained network model to predicting only a single advertisement, such as the click-through rate or conversion rate. Therefore, how to construct sample data for training network models while ensuring the performance of the network model in distributing advertisements has become a hot research topic. Summary of the Invention

[0004] This application provides a model training method, apparatus, device, medium, and program product that can construct sample pairs to train a network model, ensuring the training process of the network model and thus ensuring the performance of the trained network model.

[0005] On the one hand, embodiments of this application provide a model training method, the method comprising:

[0006] Retrieve historical advertising data, which includes the M ads that were triggered, the trigger time for each ad, and the triggering object for each ad, where M is an integer;

[0007] A sample set is constructed based on historical advertising data. The sample set contains N sample pairs. Any one of the N sample pairs is represented as a target sample pair. The target sample pair includes ad Di, ad Di+, and ad Di-. Ad Di refers to any ad among the M ads that is triggered by the triggered object but has not been converted. Ad Di+ refers to any ad among the M ads that is triggered by the same triggered object and is converted after ad Di is triggered. Ad Di- refers to any ad among the M ads that is triggered by the same triggered object but has not been converted after ad Di is triggered. N and i are both integers, and i∈[1,N].

[0008] A retargeting relevance model is trained using a sample set. The trained retargeting relevance model is used to predict the retargeting relevance between two ads. The retargeting relevance between two ads is used to represent the probability that the other ad will be triggered and convert when one ad is triggered.

[0009] On the other hand, embodiments of this application provide a model training apparatus, which includes:

[0010] The acquisition unit is used to acquire historical advertising data, which includes M triggered ads, the trigger time for each ad, and the trigger object for each ad, where M is an integer.

[0011] The processing unit is used to construct a sample set based on historical advertising data. The sample set contains N sample pairs, and any one of the N sample pairs is represented as a target sample pair. The target sample pair includes ad Di, ad Di+, and ad Di-. Ad Di refers to any ad among the M ads that is triggered by a triggered object but has not been converted. Ad Di+ refers to any ad among the M ads that is triggered by the same triggered object and converted after ad Di is triggered. Ad Di- refers to any ad among the M ads that is triggered by the same triggered object but has not been converted after ad Di is triggered. N and i are both integers, and i∈[1,N].

[0012] The processing unit is used to train a retargeting relevance model using a sample set. The trained retargeting relevance model is used to predict the retargeting relevance between two ads. The retargeting relevance between two ads is used to represent the probability that the other ad will be triggered and converted when one ad is triggered.

[0013] In one implementation, the processing unit, used to train the retargeting correlation model using a sample set, is specifically used for:

[0014] Obtain the association information of each advertisement in the target sample pair;

[0015] The retargeting relevance model to be trained is invoked to extract features from the association information of each advertisement in the target sample pair, thereby obtaining the feature data of each advertisement.

[0016] Based on the feature data of Ad Di and Ad Di+, the retargeting relevance between Ad Di and Ad Di+ is predicted to obtain the first relevance result. The first relevance result is used to represent the probability that Ad Di+ will be triggered and converted when Ad Di is triggered.

[0017] Based on the feature data of Ad Di and Ad Di-, the retargeting relevance between Ad Di and Ad Di- is predicted to obtain a second relevance result. The second relevance result is used to represent the probability that Ad Di- will be triggered and converted when Ad Di is triggered.

[0018] The retargeting relevance model is optimized based on the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-.

[0019] In one implementation, the retargeting relevance model includes a vector transformation network and a feature classification network; the processing unit, used to call the retargeting relevance model to be trained to perform feature extraction processing on the association information of each advertisement in the target sample pair, specifically for obtaining the feature data of each advertisement, is used to:

[0020] The vector transformation network is invoked to perform vector transformation processing on the association information of each advertisement in the target sample pair, so as to obtain the expression vector of each advertisement in the target sample pair;

[0021] The feature classification network is invoked to classify the expression vectors of each advertisement in the target sample pair, thereby obtaining the feature data of each advertisement in the target sample pair.

[0022] In one implementation, the retargeting relevance model further includes a first relevance calculation network and a second relevance calculation network; the first relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di+ based on the feature data of ad Di and the feature data of ad Di+, to obtain the first relevance result; the second relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di- based on the feature data of ad Di and the feature data of ad Di-, to obtain the second relevance result.

[0023] In one implementation, the processing unit, when optimizing the retargeting relevance model according to the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-, specifically performs the following:

[0024] Obtain the loss function of the retargeting relevance model, which includes the target parameter;

[0025] Based on the first correlation result and the second correlation result, calculate the loss value of the loss function; where, when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is less than the target parameter, the loss value is a non-preset value; when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is greater than the target parameter, the loss value is a preset value.

[0026] If the loss value is not a preset value, the retargeting relevance model is iteratively trained.

[0027] If the loss value is preset, the retargeting correlation model reaches the convergence condition, and the trained retargeting correlation model is obtained.

[0028] In one implementation, the processing unit, when constructing a sample set based on historical advertising data, specifically performs the following functions:

[0029] Select ad Di from ad M, and obtain the trigger time and trigger object corresponding to ad Di;

[0030] Based on the trigger time and triggering object corresponding to ad Di, identify ad Di+ from M ads that are triggered by the same triggering object and converted after ad Di is triggered, and construct ad Di and ad Di+ as positive sample candidate pairs;

[0031] Based on the triggering time and triggering object corresponding to ad Di, identify ad Di- from M ads that were triggered by the same triggering object but did not convert, and construct ad Di and ad Di- as negative sample candidate pairs;

[0032] Construct target sample pairs based on positive and negative candidate sample pairs; and,

[0033] Add the target sample pair to the sample set.

[0034] In one implementation, the processing unit is further used for:

[0035] In response to an ad distribution request for a target audience, retrieve the first ad that has been triggered by the target audience. The first ad is an ad that was triggered by the target audience within a historical time period but has not been converted.

[0036] Retrieve the set of ads to be distributed, which contains one or more second ads to be distributed;

[0037] The trained retargeting relevance model is invoked to predict the retargeting relevance between the first ad and each second ad in the ad set to be distributed;

[0038] Based on the retargeting relevance between the first ad and each second ad in the ad set to be distributed, the target ad is determined from the ad set to be distributed;

[0039] Distribute targeted ads to the target audience.

[0040] In one implementation, the processing unit, when determining the target advertisement from the ad set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed, specifically performs the following:

[0041] Sort the second ads in the ad set to be distributed in descending order of relevance;

[0042] Target ads are determined from the set of ads to be distributed based on the sorting results.

[0043] In one implementation, the processing unit, when determining the target advertisement from the ad set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed, specifically performs the following:

[0044] Obtain the revenue forecast for each second ad in the ad set to be distributed;

[0045] Based on the retargeting relevance between the first ad and each second ad in the ad set to be distributed, and the revenue forecast of each second ad, the distribution value of each second ad is calculated.

[0046] Sort the second ads in the ad set to be distributed in descending order of their distribution values;

[0047] Target ads are determined from the set of ads to be distributed based on the sorting results.

[0048] In one implementation, the processing unit, when determining the target advertisement from the set of advertisements to be distributed according to the sorting result, specifically performs the following:

[0049] Get the number of ad distributions requested in the ad distribution request;

[0050] Based on the sorting results, the second advertisement with the highest number of distributions in the order of the advertisements to be distributed is identified as the target advertisement.

[0051] In one implementation, the processing unit, when acquiring the set of advertisements to be distributed, specifically performs the following functions:

[0052] Obtain the attribute information of the target object, including object attribute information and / or behavioral attribute information;

[0053] Based on the attribute information, search the database for one or more advertisements that match the attribute information;

[0054] Filter out the second ad from the matching ads and add the second ad to the set of ads to be distributed.

[0055] In one implementation, the processing unit, when filtering out a second advertisement to be distributed from the matching advertisements, specifically performs the following:

[0056] For each of the matched ads, perform click-through rate (CTR) prediction processing to obtain the CTR prediction result for each ad; and

[0057] Perform conversion rate prediction processing on each of the one or more matching ads to obtain the conversion rate prediction result for each ad;

[0058] Based on the click-through rate prediction results, conversion rate prediction results, and bidding information, calculate the revenue prediction value for each advertisement. The bidding information includes the electronic resources that the advertiser bids for to reach the advertisement.

[0059] Among one or more matching ads, the ad with a revenue forecast greater than a revenue threshold is identified as the second ad to be distributed.

[0060] On the other hand, embodiments of this application provide a computer device, the computer device comprising:

[0061] A processor, adapted to execute computer programs;

[0062] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the model training method described above.

[0063] On the other hand, embodiments of this application provide a computer-readable storage medium storing computer instructions, the computer program being adapted to be loaded by a processor and executed as described above in the model training method.

[0064] On the other hand, embodiments of this application provide a computer program product or computer program that includes 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 and executes the computer instructions, causing the computer device to perform the aforementioned model training method.

[0065] In this embodiment, a pair of comparative samples can be constructed based on M ads included in historical advertising data. The comparative relationship of the sample pairs can be represented as follows: the relationship between ad Di (triggered but not converted by a triggered object) and ad Di+ (triggered and converted by the same triggered object after ad Di is triggered among the M ads), and ad Di- (triggered and not converted by the same triggered object after ad Di is triggered among the M ads). Since the sample pairs are constructed using the comparative relationship between ads that perform different behaviors (e.g., triggered and converted, triggered but not converted) by the triggered object, the retargeting relevance model trained based on these comparative sample pairs can be used to predict the retargeting relevance between two ads. For example, it can be used to predict the probability that a target object triggers and converts a target ad when a target object triggers an ad, ensuring the performance of the retargeting relevance model in distributing ads, and thus enabling the distribution of more relevant ads to the target object through the retargeting relevance model. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This application illustrates an exemplary embodiment of an advertising distribution system.

[0068] Figure 2 A schematic flowchart of a model training method provided in an exemplary embodiment of this application is shown;

[0069] Figure 3 This illustration shows a schematic diagram of obtaining a sample advertisement set according to an exemplary embodiment of this application;

[0070] Figure 4 This illustration shows a schematic diagram of the model structure used during training of a retargeting relevance model provided in an exemplary embodiment of this application;

[0071] Figure 5 This illustration shows a schematic diagram of the model structure used in the application of a redirection relevance model provided in an exemplary embodiment of this application;

[0072] Figure 6 This illustration shows an exemplary embodiment of the present application, which provides a method for generating an advertisement distribution request through a trigger entry and triggering the display of a target service interface;

[0073] Figure 7 This illustration shows a schematic diagram of generating an advertisement distribution request through a refresh operation, provided by an exemplary embodiment of this application.

[0074] Figure 8 This application illustrates an exemplary embodiment of a method for distributing targeted advertisements to a target object based on a search field input by the target object;

[0075] Figure 9a This illustration shows a schematic diagram of obtaining a set of advertisements to be distributed, provided by an exemplary embodiment of this application;

[0076] Figure 9b This illustration shows a schematic diagram of obtaining a set of advertisements to be distributed, provided by an exemplary embodiment of this application;

[0077] Figure 10 This illustration shows a schematic diagram of the model structure used in prediction by a correlation model provided in an exemplary embodiment of this application;

[0078] Figure 11 A schematic flowchart of a model training method provided in an exemplary embodiment of this application is shown;

[0079] Figure 12 This invention provides a schematic diagram of the structure of a prediction model according to an exemplary embodiment of the present application.

[0080] Figure 13 This invention provides a schematic diagram of the structure of a model training apparatus according to an exemplary embodiment of the present application.

[0081] Figure 14 A schematic diagram of the structure of a computer device provided in an exemplary embodiment of this application is shown. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] This application proposes a model training scheme that involves artificial intelligence technologies such as machine learning, wherein:

[0084] I. Artificial Intelligence (AI).

[0085] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to have perception, reasoning, and decision-making capabilities. AI technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0086] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning. Machine learning can be viewed as a task whose goal is to enable machines (computers in a broad sense) to acquire human-like intelligence through learning. For example, humans play Go, and computer programs (AlphaGo or AlphaGo Zero) are designed to possess Go knowledge and be able to play Go. Various methods can be used to achieve machine learning tasks, such as neural networks, linear regression, decision trees, support vector machines, Bayesian classifiers, reinforcement learning, probabilistic graphical models, and clustering.

[0087] Neural networks are a method for implementing machine learning tasks. In the field of machine learning, discussions of neural networks generally refer to "neural network learning." It is a network structure composed of many simple elements, similar to a biological nervous system, used to simulate the interaction between organisms and their natural environment. The more complex the network structure, the richer its functionality tends to be. Neural networks are a broad concept, and for different learning tasks such as speech, text, and images, neural network models more suitable for specific learning tasks have been developed, such as Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Fully Convolutional Neural Networks (FCNNs), etc.

[0088] II. Advertising.

[0089] Advertising refers to a promotional method that introduces information about commerce, services, or cultural and sports programs to users via the internet. For example, advertisements about products can be displayed on a webpage, which can be used to introduce relevant information about the product (such as its style, price, and specifications). Advertising can include search ads and online ads. Search ads refer to ads found based on search fields entered by the user; for example, if a user enters "treadmill" into a search engine, the search engine can find ads related to treadmills. Online ads, also known as internet ads, refer to ads displayed in ad slots on internet platforms (such as any application). The display of online ads does not depend on the search fields entered by the user but is actively pushed to ad slots on the internet platform. Here, an internet platform can refer to an application that supports the display of ads, and an application can refer to a computer program that performs one or more specific tasks. Applications can be categorized according to different dimensions (such as how the application operates, its functions, etc.) to obtain the types of the same application under different dimensions. ① According to the way the application operates, applications can include, but are not limited to: clients installed on a terminal, small programs that can be used without downloading and installation, web applications opened through a browser, etc. ② According to the functional type of the application, applications may include, but are not limited to: IM (Instant Messaging) applications, content interaction applications, etc. Instant messaging applications refer to applications that facilitate instant messaging and social interaction via the internet, and may include, but are not limited to: social applications with communication functions, map applications with social interaction functions, game applications, etc. Content interaction applications refer to applications capable of content interaction, such as online banking, sharing platforms, personal spaces, news applications, etc. This application embodiment does not limit the specific type of application mentioned above for the internet platform; this is only a description of the specific application type.

[0090] For example, an internet platform is a social application with social functions. This social application includes a social service interface (such as any service interface with content distribution capabilities). The social service interface is used to display a social information stream (or feed stream), which is an information stream that continuously updates and displays social information to the user. Usually, only a refresh operation is needed to obtain a large amount of social information, which constitutes the social information stream. When a user opens and uses the social application, the social service interface of the social application is output. Advertisements can be actively pushed to this social service interface, making the advertisement a type of social information in the feed stream. This makes it convenient for users to see advertisements while browsing social information. For ease of explanation, online advertisements will be used as an example in the following description, but the embodiments of this application are also applicable to search advertisement scenarios, and are explained here.

[0091] In this context, advertisements displayed in ad slots on internet platforms belong to the advertiser; that is, the advertisements are provided or placed by the advertiser. An advertiser is the entity that displays its advertisements through ad slots on an internet platform. This can be an individual user or a group of users (such as Company XX). Advertisers often have specific advertising goals (or objectives) when placing their ads. These goals may include, but are not limited to, triggering goals, conversion goals, and target audience goals.

[0092] ① A trigger goal can be simply understood as a goal targeting the click-through rate (CTR) of an advertisement. CTR refers to the click-through rate of an advertisement, which is the actual number of clicks divided by the number of ad impressions. For example, if an ad is shown 1000 times, and 200 of those impressions are clicked, then the CTR is 200 / 1000 * 100% = 20%. CTR is an important metric for measuring advertising effectiveness.

[0093] ② The conversion goal can refer to the target conversion rate of the advertisement. Conversion Rate (CVR) is also a metric for measuring advertising effectiveness, which refers to the probability that a user will convert after triggering (or clicking) an advertisement. Further, conversion refers to the deeper-level behavior that a user exhibits after triggering an advertisement. For example, if the advertisement is a client-side advertisement, then after a user clicks on that client-side advertisement, if there is a behavioral chain such as download → activation → payment (or download → activation, or download), then it is determined that the user has converted after triggering the advertisement. This conversion chain is download → activation → payment. Therefore, the conversion rate can be understood as the proportion of a user who clicks on an advertisement and becomes a valid activated, registered, or paying user, that is, the actual number of conversions for the advertisement divided by the number of clicks on the advertisement.

[0094] ③Target audience refers to the target audience for the ad's reach (or display). Target audience attributes may include the target audience's age, gender, region, etc. For example, if the ad is a beauty ad, the advertiser may prefer to target the ad to female audiences.

[0095] It should be noted that the embodiments in this application are only illustrated using the above three types of delivery targets as examples. In actual application scenarios, there may be other delivery targets, which will not be described in detail here.

[0096] Based on the foregoing description of advertising conversion rates, it is clear that the higher the conversion rate, the greater the revenue the advertisement can bring to the advertiser. Therefore, when distributing advertisements, it is often hoped that the advertisements distributed to the target audience (such as any user) are more likely to be triggered and converted by the user, in order to increase the revenue of the advertisement. Based on this, this application provides a retargeting relevance model to distribute more relevant target advertisements to the target audience. Here, "more relevant target advertisements" can be defined according to the advertising goals set by the advertiser when placing the advertisement, that is, what kind of conversion behavior the advertiser expects the target audience to produce in response to the advertisement placed by the advertiser. For example, if the advertiser's advertising goal for the advertisement is to generate a download behavior for the advertisement, then the so-called "more relevant target advertisement" can refer to an advertisement that is likely to be triggered and downloaded by the target audience. As another example, if the advertiser's advertising goal for the advertisement is to generate a registration behavior for the advertisement, then the "more relevant target advertisement" can refer to an advertisement that is likely to be triggered and registered by the target audience.

[0097] In specific implementation, the training process of the retargeting relevance model may include: First, acquiring historical advertising data, which includes M triggered ads, the trigger time for each ad, and the triggering object for each ad, where M is an integer; then, constructing a sample set based on the historical advertising data, which contains N sample pairs, where any one of the N sample pairs is represented as a target sample pair, including ad Di, ad Di+, and ad Di-; ad Di refers to any ad among the M ads that is triggered by the triggering object but not converted; ad Di+ refers to any ad among the M ads that is triggered by the same triggering object and converted after ad Di is triggered; ad Di- refers to any ad among the M ads that is triggered by the same triggering object but not converted after ad Di is triggered; N and i are both integers, and i∈[1,N]; finally, training the retargeting relevance model using the sample set, where the trained retargeting relevance model is used to predict the retargeting relevance between two ads; the retargeting relevance between two ads is used to represent the probability that the other ad is triggered and converted when one ad is triggered. In the above scheme, since the sample pairs in the sample set are constructed by comparing the relationship between ads that are triggered and convert, or triggered but not converted, the retargeting relevance model trained based on such sample pairs with comparative relationships can be used to predict the retargeting relevance between two ads, ensuring the performance of the retargeting relevance model in distributing ads, and thus enabling the distribution of more relevant ads to the target object through the retargeting relevance model.

[0098] This application embodiment can use an advertising distribution platform to train the retargeting relevance model. The trained retargeting relevance model can run on the advertising distribution platform, so that it can be called to distribute target advertisements to target objects in the advertising distribution scenario. Furthermore, the advertising distribution platform can be deployed on a computer device. For ease of explanation, the following description uses a computer device to execute the model training scheme provided in this application embodiment as an example. Here, the advertising distribution platform (Ad-Exchange, ADX) refers to the entity that connects media owners (SSP) and advertisers (DSP). It places the advertiser's advertisements on the ad slots provided by the media owner. The media owner can be an object belonging to the aforementioned Internet platform; in some advertising distribution scenarios, both the advertising distribution platform and the Internet platform can belong to the media owner, which is not limited in this application embodiment.

[0099] To facilitate understanding of the model training scheme provided in the embodiments of this application, the following is combined with... Figure 1 The advertising distribution system shown provides a brief introduction to scenarios where a retargeting relevance model is used for advertising distribution; for example... Figure 1As shown, the advertising distribution system includes a target terminal 101 and a computer device 102. The target terminal 101 refers to the terminal used by the target user, which may include, but is not limited to, smartphones (such as Android phones, iOS phones, etc.), tablets, portable personal computers, mobile internet devices (MIDs), smart TVs, in-vehicle devices, head-mounted devices, and other touchscreen smart devices. The terminal runs an internet platform. When the target user opens and uses the internet platform, the advertising slots on the internet platform can display target advertisements recommended to the target user for browsing. The computer device 102 may refer to the server corresponding to the internet platform running on the target terminal 101, or the server corresponding to the advertising distribution platform, used to provide computing and application service support for the target terminal 101 or the advertising distribution platform. This application embodiment does not limit the number of computer devices 102 in the advertising distribution system; for example, the number of computer devices 102 can be two, including the server of the Internet platform and the server of the advertising distribution platform. In specific implementation, after the advertising distribution platform calls its running retargeting relevance model to determine the target advertisement to be distributed to the target object, it sends the target advertisement to the server of the Internet platform, so that the server of the Internet platform can send the target advertisement to the target terminal 101 for display. The server may include, but is not limited to, devices with complex computing capabilities such as data processing servers, web servers, and application servers. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. The target terminal 101 and the computer device 102 can communicate directly or indirectly through wired or wireless means. This application embodiment does not limit the connection method between the target terminal and the computer device.

[0100] based on Figure 1 The advertising distribution system illustrated in this application uses an internet platform as the target application and the advertising space on the internet platform as the social service interface of the target application as an example. The general process of calling a trained retargeting relevance model to distribute target advertisements to the target object is as follows: First, when the target object opens and uses the target application, the social service interface of the target application is displayed. This social service interface is used to display social information streams, and an advertising distribution request is generated at this time. This advertising distribution request is used to request the distribution of advertisements to the target object. Second, the target terminal sends the advertising distribution request to a computer device (such as the server corresponding to the advertising distribution platform), so that the computer device responds to the advertising distribution request and calls the trained retargeting relevance model to predict the target advertisement. Finally, the computer device sends the target advertisement to the target terminal, so that the target application outputs the target advertisement (such as...) in the advertising space of the social service interface. Figure 1Advertisement 103 shown.

[0101] It should be noted that when the embodiments of this application are applied to specific products or technologies, such as when obtaining historical advertising data, it is inevitable to obtain the attribute information of the triggering object (such as object attribute information and / or behavioral attribute information). In this case, permission or consent from the target object is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0102] Based on the model training scheme described above, this application proposes a more detailed model training method. The model training method proposed in this application will be described in detail below with reference to the accompanying drawings.

[0103] Figure 2 The illustration shows a flowchart of a model training method provided in an exemplary embodiment of this application; the model training method can be executed by a computer device, and the model training method may include, but is not limited to, steps S201-S203:

[0104] S201: Obtain historical advertising data.

[0105] Historical advertising data includes M triggered ads, the trigger time for each ad, and the triggering object for each ad, where M is an integer. The M triggered ads can refer to ads triggered within a target time period; the target time period is a preset duration earlier than the time the historical advertising data is retrieved. For example, if the retrieval of historical advertising data is triggered at 12:00 on December 1st, the target time period could be from 12:00 on November 24th to 12:00 on November 31st. The trigger time for each ad can refer to the time when the ad is triggered by the triggering object within the target time period. If any ad is triggered by the same or different triggering objects at the first and second moments within the target time period, then the trigger time for that ad includes both the first and second moments. The triggering object for each ad can refer to the user who triggers the ad after it is reached. For example, if any ad is reached and is triggered by a first triggering object at the first moment within the target time period, and then triggered by a second triggering object at the second moment within the target time period, and the first and second triggering objects are different, then the triggering objects for that ad include both the first and second triggering objects.

[0106] Specifically, this application embodiment supports obtaining historical advertising data for multiple trigger objects that have performed retargeting behavior within a target time period. That is, trigger objects that have performed retargeting behavior on advertisements within the target time period are added to an object set. Then, advertisements triggered by each trigger object within the object set during the target time period are added as historical advertisements to the historical advertising data, thereby achieving the acquisition of historical advertising data. The retargeting behavior performed on advertisements can include: triggered but not converted behavior, and triggered and converted behavior; the conversion of an advertisement can refer to: download, registration, payment, etc.; this application embodiment uses download and payment conversions as examples for illustration. For example: if a trigger object performs a triggered and downloaded behavior on any advertisement within the target time period, then it is determined that the trigger object has performed retargeting behavior within the target time period, i.e., the trigger object belongs to the object set. It should be noted that this application embodiment does not limit the duration of the target time period; for example, the duration of the target time period (e.g., the target time period) can be 7 days (or 8 days, etc.).

[0107] S202: Construct a sample set based on historical advertising data.

[0108] It should be noted that the main purpose of this application's embodiments is to learn, through modeling, the comparison relationship between: triggering ad A and then triggering and converting ad B, and triggering ad A and then triggering but not converting ad C; that is, in the process of training the retargeting relevance model, the retargeting relevance model should be able to predict that the retargeting relevance between ad A and ad B is higher than the retargeting relevance between ad A and ad B; that is, in training the retargeting relevance model, the training score of ads with conversion behavior is higher than the training score of ads without conversion behavior.

[0109] Based on this, the process of constructing a sample set based on historical advertising data in this application embodiment can be understood as: constructing sample pairs with a comparative relationship based on historical advertising data; thus, the sample set constructed based on historical advertising data may include N sample pairs, where N is an integer. This application embodiment supports representing the sample set as:

[0110]

[0111] Where i represents the i-th advertisement, i = 1, ..., N, i.e., i ∈ [1, N], and N represents the total number of samples. Ad Di refers to any advertisement among the M advertisements in the historical ad data that was triggered by a triggering object but did not convert. Ad Di+ refers to any advertisement among the M advertisements that, after Ad Di was triggered, was triggered by the same triggering object and converted. Ad Di- refers to any advertisement among the M advertisements that, after Ad Di was triggered, was triggered by the same triggering object but did not convert.

[0112] Understandably, the construction process for each sample pair in the sample set built from historical advertising data is similar. If any one of the N sample pairs is represented as the target sample pair, then the advertisements contained in that target sample pair can include: Ad Di, Ad Di+, and Ad Di-. The following section combines... Figure 3 The process of constructing target sample pairs based on historical advertising data is illustrated below. The steps for constructing target sample pairs based on M advertisements may include steps s11-s14, where:

[0113] s11: Select ad Di from the M ads contained in the historical ad data, and obtain the trigger time and corresponding trigger object for ad Di. For example, the trigger time and corresponding trigger object for ad Di may include: ad Di is triggered by the first trigger object at the first moment, and triggered by the second trigger object at the second moment; the first moment and the second moment are different, and the first trigger object and the second trigger object are different. Then, one trigger time and corresponding trigger object can be selected from the multiple trigger times corresponding to ad Di, such as selecting the first moment as the trigger time corresponding to ad Di and the first trigger object as the corresponding trigger object, and continue to perform subsequent operations.

[0114] s12: Based on the trigger time and corresponding triggering object of ad Di, determine the ad sequence triggered by the same triggering object after ad Di is triggered from M ads. Specifically, starting from the moment ad Di is triggered (e.g., the first moment), obtain one or more ads that are triggered and converted by the same triggering object within the observation period (e.g., 7 days). For example, if ad Di is ad A, the ad sequence consisting of ads triggered by the first triggering object after ad A is triggered at the first moment within the observation period may include: [ad B, ad C, ad D, ad E]; among which, the ads that are triggered and converted by the first triggering object within the observation period include: ad D, and the ads that are triggered by the first triggering object but not converted within the observation period include: ad B, ad C, and ad E.

[0115] s13: Based on the trigger time and corresponding triggering object of ad Di, determine ad Di+ from the M ads that is triggered and converted by the same triggering object after ad Di is triggered, and construct ad Di and ad Di+ as a positive sample candidate pair. Specifically, determine the ad that is triggered and converted by the first triggering object from the ad sequence in step s12. If the ad is ad D, then ad D is determined as ad Di+; then, construct ad Di and ad Di+ as a positive sample candidate pair (e.g., [ad A => ad D]). This positive sample candidate pair indicates that when ad A is triggered by the first triggering object but does not convert, ad D is triggered and converted by the first triggering object.

[0116] s14: Based on the trigger time and corresponding triggering object of ad Di, determine ad Di- from the M ads that were triggered by the same triggering object but did not convert after ad Di was triggered, and construct ad Di and ad Di- as a negative sample candidate pair. Specifically, determine the ads that were triggered by the first triggering object but did not convert from the ad sequence in step s11. If the ads include ad B, ad C, and ad D, then any one of ad B, ad C, and ad D can be determined as ad Di-; then, construct ad Di and ad Di- as a negative sample candidate pair, such as [ad A => ad B]. This negative sample candidate pair indicates that when ad A was triggered by the first triggering object but did not convert, ad B was triggered by the first triggering object but did not convert.

[0117] s15: Construct target sample pairs based on positive and negative candidate sample pairs; and add the target sample pairs to the sample set. Specifically, the target sample pairs constructed based on the positive candidate pair [Ad A => Ad D] and the negative candidate pair [Ad A => Ad B] can be represented as: Ad A => (Ad D, Ad B).

[0118] Based on the above description of steps s12-s14, a sample pair corresponding to the first triggering object can be constructed based on the historical triggering ad sequence of the first triggering object. It is worth noting that the ads triggered by the first triggering object but not converted in step s14 include not only ad B, but also ad C and ad E. Therefore, ad B, ad C, and ad D can all be regarded as ad Di-. At this time, the negative sample candidate pairs constructed by ad Di and ad Di- are [ad A => ad B, ad A => ad C, ad A => ad E]. Then, based on the positive sample candidate pair [ad A => ad D] and the negative sample candidate pair [ad A => ad B, ad A => ad C, ad A => ad E], the number of target sample pairs constructed is 3, namely: ad A => (ad D, ad B), ad A => (ad D, ad C), and ad A => (ad D, ad E). This allows for the construction of multiple sample pairs corresponding to the same trigger object when selecting an ad Di from M ads, thereby improving the efficiency of sample pair construction and thus enhancing model training efficiency.

[0119] S203: Use a sample set to train the retargeting correlation model.

[0120] The retargeting relevance model, also known as the ranking model, provided in this application is a model trained using a pairwise architecture. This model can be used to predict the retargeting relevance between ad Di and ad Di+, and between ad Di and ad Di-, in each pair of samples in the sample set. Pairwise approaches solve the ranking problem by approximating it as a classification problem. The model structure of the retargeting relevance model during pairwise training can be found in [link to relevant documentation]. Figure 4 .like Figure 4 As shown, the retargeting relevance model can be understood as a binary classifier. This binary classifier learns from the input pair AB (such as ad A and ad B). Based on whether the relevance of ad A is better than that of ad B, the binary classifier can output a classification label of 1 or 0. In this way, multiple ads can be classified, and thus the ranking relationship of multiple ads can be obtained.

[0121] See also Figure 4 The retargeting relevance model may include: a vector transformation network, a feature classification network, a first relevance calculation network, and a second relevance calculation network. Specifically: ① The vector transformation network can be used to transform input identifier-type features (such as ID-type features) into feature vectors (or embedding vectors); here, identifier-type features can refer to the ID-type features of advertisements in the input sample pair, such as the advertisement's identifier (e.g., an ID used to uniquely identify an advertisement), the advertiser's object identifier (e.g., an ID used to uniquely identify the advertiser), etc. ② The feature classification network can refer to the fully connected layer network of the aforementioned neural network model. This feature classification network can be used to perform some non-limited transformations on the input feature vectors to achieve classification of the feature vectors and obtain the prediction results corresponding to the input advertisements. ③ The first relevance calculation network receives the feature data of ad Di and ad Di+ output by the feature classification network, and calculates the retargeting relevance between ad Di and ad Di+ based on the feature data of ad Di and ad Di+. Similarly, ④ the second relevance calculation network receives the feature data of ad Di and ad Di- output by the feature classification network, and calculates the retargeting relevance between ad Di and ad Di- based on the feature data of ad Di and ad Di-. It should be noted that... Figure 4 The model structure of the redirection relevance model shown is only an example. In real-world applications, this redirection relevance model may include other structures, such as a concatenation network for concatenating multiple embedding vectors.

[0122] Based on the above introduction to the model structure of the redirection relevance model, the following describes the implementation process of training the redirection relevance model using a sample set. Specifically, taking the target sample pair as an example, the training process of the redirection relevance model is illustrated, and this process may include steps s21-s24; wherein:

[0123] s21: Obtain the association information of each advertisement in the target sample pair.

[0124] In this target sample pair, each advertisement is designated as Ad Di, Ad Di+, and Ad Di-. The associated information for each advertisement may include, but is not limited to: the advertisement's identifier (e.g., ad id), the identifier of the advertiser to which the advertisement belongs (e.g., the advertiser's nickname, id, etc.), the advertiser's target audience for the advertisement, the corresponding trigger time, and the attribute information of the triggering object (e.g., object attribute information (e.g., the triggering object's nickname, id, etc.) and behavioral attribute information (e.g., the behavior information of the triggering object in response to the advertisement)). Specifically, the associated information of the advertisement can be obtained from the database based on the advertisement's identifier (e.g., the id used to uniquely identify the advertisement). This application embodiment does not limit the type of database or the data stored, which is not described here.

[0125] In practice, after obtaining the association information of each advertisement in the target sample pair, this association information can be used as input information for the retargeting relevance model to be trained, so that the model can perform prediction processing on the input information. For example... Figure 4 As shown, the input information for the retargeting relevance model includes: the association information of the triggered ad Di and the attribute information of the triggered object (such as...). Figure 4 (the far left); related information of the Di+ advertisement (such as...) Figure 4 The middle section shown); the rightmost section represents the associated information of the advertisement Di- (such as...). Figure 4 (As shown on the far right).

[0126] s22: The retargeting relevance model to be trained is invoked to extract features from the association information of each advertisement in the target sample pair, obtaining feature data for each advertisement. Specifically, the vector transformation network and feature classification network included in the retargeting relevance model are used to extract features from the association information of each advertisement in the target sample pair. It should be noted that the network structure and function of the vector transformation network and feature classification network included in the retargeting relevance model can be found in the aforementioned descriptions and will not be repeated here. In the specific implementation, the vector transformation network is invoked to perform vector transformation processing on the association information of each advertisement in the target sample pair, obtaining the expression vector of each advertisement in the target sample pair; then, the feature classification network is invoked to classify the expression vector of each advertisement in the target sample pair, obtaining feature data for each advertisement in the target sample pair, which can be represented by vectors.

[0127] s23: The retargeting relevance model to be trained is invoked to predict the retargeting relevance between ad Di and ad Di+, yielding a first relevance result; and the retargeting relevance between ad Di and ad Di- is also predicted, yielding a second relevance result. Specifically, the first and second relevance calculation networks contained in the retargeting relevance model are invoked to predict the retargeting relevance between ads; the network structure and function of the first and second relevance calculation networks can be found in the aforementioned descriptions and will not be repeated here.

[0128] The first relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di+ based on the feature data of ad Di and ad Di+, obtaining the first relevance result. The first relevance result represents the probability that ad Di+ will be triggered and convert when ad Di is triggered. The first relevance calculation network can be represented as follows:

[0129]

[0130] Indicates the advertisement D in the target sample pair i Ads that lead to conversion The first relevance result (or simply relevance score) between the target sample pairs is used to determine the relevance of the advertisement D in the target sample pair. The higher the value of the first relevance result, the stronger the relevance of the advertisement D in the target sample pair. i Ads that lead to conversion The higher the relevance between ads D, i.e., when ad D... i When triggered, the advertisement The greater the probability of being triggered and converted.

[0131] Similarly, the second relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di- based on the feature data of ad Di and ad Di-, obtaining the second relevance result. The second relevance result represents the probability that ad Di- will be triggered and convert when ad Di is triggered. The second relevance calculation network can be represented as:

[0132]

[0133] Indicates the advertisement D in the target sample pair i Ads that lead to conversion The second prediction result (or simply relevance score) between the target sample pairs is used to determine the relevance of the advertisement D. The higher the value of this second relevance score, the better the relevance of the advertisement D in the target sample pair. i Ads that lead to conversion The higher the relevance between ads D, i.e., when ad D... i When triggered, the advertisement The greater the probability of being triggered and converted.

[0134] It should be noted that, based on the above representations of the first and second relevance calculation networks, the calculation logic of the first and second relevance calculation networks is similar, only the data used is different; for example, the first relevance calculation network calculates the feature data of advertisement Di and the feature data of advertisement Di+; and the second relevance calculation network calculates the feature data of advertisement Di and the feature data of advertisement Di-.

[0135] s24: Following the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-, the retargeting relevance model is optimized. It is understood that, in order for the trained relevance model to predict ads with higher relevance to the first ad already triggered by the target object—that is, to predict ads more likely to be triggered and converted by the target object after being reached—the training objective of the retargeting relevance model in this application embodiment is: to make the retargeting relevance of the retargeting relevance model between ad Di and ad Di+ in the target sample pair higher than the retargeting relevance between ad Di and ad Di-, so that the output result of the retargeting relevance model... As large as possible, Keep it as small as possible.

[0136] In practice, the loss function of the retargeting correlation model is used to calculate the loss value, and the retargeting correlation model is continuously optimized based on the loss value to obtain a trained retargeting correlation model. The loss function (or optimization function) of the retargeting correlation model can be expressed as:

[0137]

[0138] Where ε is the target parameter, or hyperparameter, included in the loss function; in this embodiment, the hyperparameter ε can take the value 1. i, N, D i , Sim(D i D + ) and Sim(D i D - For related concepts, please refer to the above description, which will not be repeated here. Indicates from 0 and Select the largest value from the list. This means adding up the maximum values ​​from the sample ad set.

[0139] The following describes the process of calculating the loss value based on the relevance results between advertisements in a target sample pair and optimizing the retargeting relevance model based on the loss value, using the example of the relevance results between advertisements in the target sample pair. First, the loss function of the retargeting relevance model is obtained, which includes the target parameter ε. Then, the loss value of the loss function is calculated based on the first relevance result between advertisement Di and advertisement Di+ in the target sample pair, and the second relevance result between advertisement Di and advertisement Di- in the target sample pair. When the first relevance result is greater than the second relevance result, and the correlation amplitude between them is less than the target parameter, the loss value is not a preset value. In this case, the retargeting relevance model is iteratively trained. Specifically, the network parameters of the retargeting relevance model are optimized using this loss value. If the number of training iterations is less than the preset total number of iterations, other samples in the sample set are used to continue training the optimized retargeting relevance model. When the first relevance result is greater than the second relevance result, and the correlation amplitude between them is greater than the target parameter, the loss value is the preset value. In this case, the retargeting relevance model is considered to have reached convergence, and the trained retargeting relevance model is obtained.

[0140] For example, suppose the objective parameter ε = 1, if for Greater than If the amplitude is 0.3, then it is determined. Greater than If the magnitude of the error is lower than the target parameter ε, then the loss function's loss value is a non-preset value (e.g., not equal to 0), indicating that the retargeting correlation model will produce a certain error loss. Therefore, iterative training of the retargeting correlation model is performed. Conversely, if for... Greater than The amplitude is 8, determined. Greater than If the magnitude of the error is higher than the target parameter ε, then the loss value of the loss function is equal to the preset value (e.g., equal to 0), indicating that the retargeting correlation model does not produce error loss, that is, the retargeting correlation model has reached stability. Then, the retargeting correlation model after this training is determined as the trained retargeting correlation model.

[0141] In summary, this application's embodiments, during modeling, compare and model ads Di, Di+, and Di- that have been clicked by the same triggering object. This better utilizes the relationship between ads in the ad sequence (e.g., [Ad B, Ad C, Ad D, Ad E]) after the triggering object triggers ad Di and the ads that trigger ad Di. Through this comparative modeling approach, ads with conversion behavior (e.g., ad Di+) achieve higher training scores than non-converting ads (e.g., ad Di-). This allows the trained retargeting relevance model to more accurately predict the retargeting relevance between the ads to be distributed and the ads already triggered by the target object during the application phase of the retargeting relevance model. This enables the ranking of multiple ads to be distributed, allowing for the distribution of more relevant ads to the target object.

[0142] The above Figure 2 The illustrated embodiment mainly introduces the implementation process of training a retargeting relevance model. The following section will combine... Figure 5 The implementation process of the model application is given. Figure 5 The illustration shows a flowchart of a model training method provided in an exemplary embodiment of this application; the model training method can be executed by a computer device, and the model training method may include, but is not limited to, steps S501-S508:

[0143] S501: Obtain historical advertising data.

[0144] S502: Construct a sample set based on historical advertising data.

[0145] S503: Use the sample set to train the retargeting correlation model to obtain the trained retargeting correlation model.

[0146] It should be noted that the specific implementation process shown in steps S501-S503 can be found in the aforementioned... Figure 2 The specific implementation process shown in steps S201-S203 of the illustrated embodiment will not be repeated here.

[0147] S504: In response to an ad delivery request for the target audience, retrieve the first ad that has been triggered for the target audience.

[0148] An ad distribution request for a target object is used to request the distribution of one or more target ads (such as ads highly relevant to the target object) to the target object. These one or more target ads can be displayed in the target application used by the target object, specifically in ad slots on the service interface provided by the target application. In specific implementation, the target terminal used by the target object can detect whether there are triggering conditions for generating an ad distribution request; if so, it determines that an ad distribution request has been generated and sends the ad distribution request to the computer device, so that the computer device can respond to the ad distribution request, obtain the first ad that the target object has triggered, and distribute the target ads to the target object based on the first ad that the target object has triggered.

[0149] The first advertisement triggered by the target audience can refer to an advertisement that was triggered by the target audience within a historical time period but did not convert. The historical time period can refer to the target time period (e.g., 3 days) prior to the moment the advertisement distribution request was received. It should be noted that if the target audience triggers only one advertisement within the historical time period, that advertisement is determined as the first advertisement; if the target audience triggers two or more advertisements within the historical time period, one advertisement can be selected as the first advertisement according to the advertisement selection rules. The advertisement selection rules may include, but are not limited to: random selection, such as randomly selecting one advertisement from two or more advertisements; or, determining the advertisement with the latest trigger time as the first advertisement, such as two or more advertisements including advertisement A and advertisement B, where the trigger time of advertisement A is 12:00 on December 3rd and the trigger time of advertisement B is 13:00 on December 3rd, then advertisement B is determined to have the latest trigger time, and thus advertisement B is determined as the first advertisement; etc. The specific implementation process of the advertisement selection rules is not limited in this application embodiment.

[0150] As described above, advertisements can include search ads and internet ads. Therefore, depending on the ad distribution scenario, the triggering conditions for generating ad distribution requests on the target terminal may differ. Below, we provide exemplary triggering conditions for generating ad distribution requests, using internet ads and search ads as examples:

[0151] 1) Advertising distribution scenario for internet advertising. As described above, internet advertising search does not depend on the search fields entered by the target audience, but is actively distributed to the target audience by the advertising distribution platform. In this implementation, the triggering conditions for generating an advertising distribution request may include: the target audience triggering the display of the target service interface (such as any service interface) containing advertising space in the target application.

[0152] In the specific implementation, when the trigger entry used to trigger the display of the target service interface is triggered, an ad distribution request is determined, and the display of the target service interface is triggered. An exemplary implementation process for generating an ad distribution request and triggering the display of the target service interface through the trigger entry can be found in [link to relevant documentation]. Figure 6 ;like Figure 6 As shown, assuming that the discovery service interface 601 of the target application includes a trigger entry 6011, which is used to trigger the display of the target service interface 602; then when the trigger entry 6011 is triggered, the target terminal determines that an advertisement distribution request for the target object has been generated, and triggers the display of the target service interface 602. The target service interface 602 can be a social service interface for displaying the social information stream of the target application, which contains a social information stream containing one or more social information, such as... Figure 6 It includes social information 6011, social information 6012, ...; any social information can be sent by any object that has a communication relationship with the target object (such as a friend relationship, a stranger relationship, a colleague relationship, etc.); the social service interface contains advertising space, and the advertisement can be displayed as a kind of social information in the social information flow.

[0153] It should be noted that this application embodiment does not limit the location of the ad slot in the target service interface, but specifically the location of the ad slot in the social message stream. For example, an ad slot can be set every target number of social messages in the social message stream (such as 3, 10, etc., this application embodiment does not limit the specific value of the target number), so that an ad can be viewed every target number of social messages. Since the display area of ​​the target terminal's display screen (or terminal screen, display screen) is limited, the ad slot set in the target service interface may be hidden when the target service interface is triggered to be displayed; then, the ad contained in the hidden ad slot can be displayed by sliding along the sliding axis included in the target service interface, or the ad contained in the ad slot can be displayed by pressing and sliding at any position in the target service interface. This application embodiment does not limit the specific implementation process of displaying the ad contained in the ad slot.

[0154] Furthermore, if the target application is already open and displaying the target service interface, an ad distribution request is determined when an information update event is detected in the target service interface. An information update event refers to an event that occurs when an information refresh operation is detected in the target service interface and the number of new social information pieces obtained meets the target conditions. Specifically, an information refresh operation refers to an operation that updates the displayed social information in the target service interface (such as gesture operations, voice input operations, triggering the refresh button, etc.). This involves obtaining new social information and displaying it in the target service interface (e.g., displaying it before or above already displayed social information) so that the target audience can quickly and conveniently browse the latest social information. Here, "already displayed social information" refers to social information that the target terminal has already obtained and can display on the terminal screen; "new social information" refers to social information whose publication time (i.e., the time the editor publishes the social information) is later than the latest displayed social information. The new social information meeting the target condition can mean that the number of new social information is greater than the target number; or, the number of new social information is less than the target number, but the sum of the new social information and the previously displayed social information between the last displayed advertisement is greater than or equal to the target number. Of course, this application does not limit the generation of an advertisement distribution request to only when the number of new social information meets the target condition; for example, an advertisement distribution request can also be generated whenever an information refresh operation is detected (or the trigger entry of the target service interface is triggered), which is hereby explained.

[0155] like Figure 7 As shown, assuming the social information stream displayed in the target service interface 701 contains social information 7011 and social information 7012, where social information 7011 is an advertisement placed by advertiser A; if an information refresh operation is detected in the target service interface 701, such as the refresh control 7016 in the target service interface 701 being triggered, or a downward swipe gesture operation, then new social information is obtained, such as social information 7013 and social information 7014; it can be seen that: the number of new social information is 2, which is less than the target number (e.g., the target number is 3), but the sum of the new social information and the previously displayed social information (i.e., social information 7012) between the most recently displayed advertisement (i.e., social information 7011) is 3, which is equal to the target number, then an advertisement distribution request is determined to be generated.

[0156] It should be noted that according to the relationship between the sum of the displayed social information between the new social information and the most recently displayed advertisement and the target quantity, the advertisement distribution request can be used to request the distribution quantity of the target advertisement. For example, if the target quantity is 3 and the sum of the displayed social information between the new social information and the most recently displayed advertisement is 6, it is determined that the advertisement distribution request is used to request 2 target advertisements for the target object. That is, this advertisement distribution request is used to request to distribute 2 target advertisements for the target object, and these two target advertisements are inserted into different positions in the social information stream respectively.

[0157] To sum up, in the advertisement distribution scenario where the advertisement is an Internet advertisement, the operation of the target object to open or refresh the target service interface can be used as the trigger condition for generating the advertisement distribution request. Then, the target advertisement is distributed to the target object, and the target advertisement is inserted into the social information stream as social information for display. To a certain extent, it can be realized that the target object receives the advertisement without feeling, which is beneficial to the promotion of the advertisement and also enriches the types of information that the target object can browse (such as advertisement types, social information types, etc.).

[0158] 2) Advertisement distribution scenario for search advertisements. As described above, a search advertisement is an advertisement obtained by searching depending on the search field input by the target object; among them, the search field can include a string, and the string contains one or more characters, and the character can include at least one of the following: Chinese characters (i.e., Chinese words), English characters (i.e., letters), numbers, and punctuation marks (such as commas ",", full stops ".", square brackets "【】"), and so on. In this implementation manner, the trigger condition for generating the advertisement distribution request can include: detecting a search operation for the search field.

[0159] In a specific implementation, if the target application contains a search box that supports the target object to input a search field; when the search field input by the target object is obtained in the search box and the search option used to trigger the start of the search operation is triggered, it is determined that an advertisement distribution request is generated. At this time, this advertisement distribution request is used to request to distribute a target advertisement that matches the search field for the target object. For example, if the search field is "bicycle", then the target advertisement distributed to the target object is related to "bicycle", such as an advertisement for selling bicycles, or a game advertisement containing bicycles, and so on. A schematic diagram of an example of distributing a target advertisement according to the search field input by the target object can be seen Figure 8 , such as Figure 8 shown, the target service interface contains a search box 801. If the search field is obtained in the search box 801 and the search option in the target service interface is triggered, an advertisement distribution request carrying the search field is generated, and this advertisement distribution request is used to request to distribute a target advertisement that matches the search field for the target object.

[0160] The above describes the triggering conditions for generating ad distribution requests in different ad distribution scenarios. For ease of explanation, the following description will use ad distribution requests generated in ad distribution scenarios where the ad is an internet ad as an example to illustrate the embodiments of this application.

[0161] S505: Get the set of ads to be distributed.

[0162] The set of advertisements to be distributed contains one or more second advertisements to be distributed. These second advertisements are stored in the database of the computer device and retrieved from the database when the computer device responds to an advertisement distribution request. In other words, the database of the computer device stores advertisements placed by one or more advertisers. When the computer device receives an advertisement distribution request sent by the target terminal, it can retrieve some advertisements from the database as second advertisements in response to the advertisement distribution request and add the second advertisements to the set of advertisements to be distributed.

[0163] In the specific implementation, firstly, in response to an ad distribution request targeting a specific object, the target object's attribute information can be obtained. This attribute information may include object attribute information and / or behavioral attribute information. Object attribute information may be information related to the target object's identity, including but not limited to: age, gender, location, and education information. Behavioral attribute information may be information related to the target object's historical behavior, including but not limited to: the type and number of ads triggered historically. Then, based on the target object's attribute information—specifically, based on the target object's object attribute information and / or behavioral attribute information—one or more ads matching the attribute information are searched (or retrieved) from the database. For example, if the target object's object attribute information includes: the target object's location is city XX, then all or part of the ads retrieved from the database based on this object attribute information (such as the location of the advertiser distributing the ads) are located in city XX. Finally, a second ad to be distributed is selected from the one or more ads matching the attribute information and added to the set of ads to be distributed.

[0164] This application's embodiments support using multiple methods to filter second advertisements to be distributed from one or more advertisements that match attribute information. For example, advertisements matching attribute information can be directly used as second advertisements. Figure 9aAs shown, in response to an ad distribution request, the target object's attribute information is first obtained, which may be stored in a storage space used to store object information; then, based on the target object's attribute information, one or more ads (or distribution ads) that match the target object's attribute information are retrieved from the database storing ads; these one or more ads are identified as second ads and added to the ad set to be distributed, thus obtaining the ad set to be distributed related to the target object.

[0165] For example, a subset of ads can be selected as second ads from one or more ads matching the attribute information. These second ads can be understood as those that are more closely matched to the target audience. In other words, after retrieving one or more ads based on the target audience's attribute information, further filtering can be performed on these ads based on filtering rules, adding the filtered ads as second ads to the ad set to be distributed. It's understandable that the specific implementation of filtering the one or more ads matching the attribute information will differ depending on the filtering rules. For instance, filtering rules could include filtering based on click-through rate (CTR) and conversion rate (PCR). This allows for the estimation of the CTR and PCR of one or more ads retrieved based on the target audience's attribute information, and the selection of a subset of ads as second ads from the one or more ads matching the attribute information based on the estimated CTR and PCR. The estimated CTR corresponds to the aforementioned CTR; it represents the probability that an ad will be clicked under certain circumstances after it has been delivered. Estimated conversion rate is the probability that an ad will convert after it has been clicked under certain circumstances, as estimated by the ad distribution platform. An illustration of filtering second ads based on click-through rate and conversion rate can be found here. Figure 9b .

[0166] S506: Call the trained retargeting relevance model to predict the retargeting relevance between the first ad and each second ad in the ad set to be distributed.

[0167] As described above, the retargeting correlation model is a model based on a pairwise architecture. However, the model structure used for prediction in a pairwise architecture differs from the model structure used for training the correlation model. The model structure used during training can be found in [link to training model]. Figure 3 As shown, the model structure for prediction can be found in [reference needed]. Figure 10 ,like Figure 3 as well as Figure 10As shown, the model structure during prediction is similar to using half of the model structure from training. It's worth noting that when applying the retargeting relevance model online, only the association information of the first ad already triggered by the target object, the association information of the second ad in the ad set to be distributed, and the attribute information of the target object are used as input information for the retargeting relevance model. This is because the embedding (vector) features related to the ads are shared features during offline training of the retargeting relevance model. Therefore, when applying the retargeting relevance model online, the retargeting relevance between the first and second ads can be predicted based on the shared embedding features. Specifically, in the application phase of the retargeting relevance model, the sharing of embedding features is achieved through the vector transformation network, feature classification network, and relevance calculation network (such as the first relevance calculation network or the second relevance calculation network) included in the trained retargeting relevance model.

[0168] The following describes the specific implementation process of predicting the retargeting relevance between the first ad and any second ad within the ad set to be distributed, using a pre-trained retargeting relevance model as an example. Specifically, the association information of the first ad, the association information of any second ad, and the attribute information of the target object are used as input to the pre-trained retargeting relevance model. Then, the vector transformation model within the pre-trained retargeting relevance model is invoked to perform vector transformation processing on the association information of the first ad and the association information of any second ad, respectively, to obtain the expression vector (e.g., embedding vector) of the first ad and the expression vector of any second ad. Finally, a feature classification network is invoked to classify the association information of the first ad and the association information of any second ad, respectively, to obtain the feature data of the first ad and the feature data of any second ad. Finally, the relevance calculation network is invoked to predict the retargeting relevance between the first advertisement and the second advertisement based on the feature data of the first advertisement and the feature data of any second advertisement, to obtain the target relevance result. The target relevance result is used to indicate the probability that the second advertisement will be triggered and converted by the target object after reaching the advertisement position when the first advertisement is triggered by the target object. The larger the target relevance result, the more likely the second advertisement is to be triggered and converted by the target object after being reached.

[0169] S507: Determine the target ad from the ad set to be distributed based on the retargeting relevance between the first ad and each second ad in the ad set to be distributed.

[0170] S508: Distribute targeted ads to the target audience.

[0171] In steps S507-S508, based on the foregoing description, the retargeting relevance between the first ad and any second ad of the target audience can be used to indicate the probability that any second ad might be triggered and converted by the target audience if the first ad is triggered by the target audience. Therefore, the greater the retargeting relevance of the second ad (i.e., the retargeting relevance between the second ad and the first ad), the more likely the second ad is to be triggered and converted after being displayed in the ad slot. Thus, ad distribution platforms often prefer to use second ads with higher retargeting relevance as target ads to increase the likelihood of ad conversion, improve the ad experience for the target audience, and ultimately increase ad revenue.

[0172] Based on this, embodiments of this application support sorting each second advertisement in the ad set to be distributed in descending order of retargeting relevance; and determining the target advertisement from the ad set to be distributed according to the sorting result. For example, assuming the ad set to be distributed contains second advertisement 1, second advertisement 2, second advertisement 3, second advertisement 4, and second advertisement 5, and the retargeting relevance prediction result obtained by calling the trained retargeting relevance model is 0.5 for the retargeting relevance between the first advertisement and second advertisement 1, 0.3 for the retargeting relevance between the first advertisement and second advertisement 2, 0.2 for the retargeting relevance between the first advertisement and second advertisement 3, 0.7 for the retargeting relevance between the first advertisement and second advertisement 4, and 0.6 for the retargeting relevance between the first advertisement and second advertisement 5, then the sequence obtained by sorting according to the retargeting relevance from high to low is: second advertisement 4 → second advertisement 5 → second advertisement 1 → second advertisement 2 → second advertisement 3; then second advertisement 4 can be determined as the target advertisement to be distributed to the target object.

[0173] In this embodiment, in response to an ad distribution request to a target object, a first ad and a set of ads to be distributed to the target object can be obtained, and the retargeting relevance between the first ad and each second ad in the set of ads to be distributed can be calculated. Based on the retargeting relevance, a target ad to be distributed to the target object is determined from the second ad dataset. The retargeting relevance between the first ad and any second ad can be used to indicate the probability that any second ad will be triggered and converted when the first ad is triggered. In the above scheme, the target ad to be distributed to the target object can be determined by calculating the relevance between the first ad and the second ad that the target object has triggered. This ensures that the target ad reaching the target object is not only relevant to the first ad but is also more likely to be triggered and converted by the target object, thus effectively distributing more relevant ads to the target object and improving the ad conversion rate.

[0174] Figure 11 The illustration shows a flowchart of a model training method provided in an exemplary embodiment of this application; the model training method can be executed by a computer device, and the model training method may include, but is not limited to, steps S1101-S1109:

[0175] S1101: Obtain historical advertising data.

[0176] S1102: Construct a sample set based on historical advertising data.

[0177] S1103: Use a sample set to train the retargeting correlation model.

[0178] It should be noted that the specific implementation process shown in steps S1101-S1103 can be found in the aforementioned... Figure 2 The specific implementation process shown in steps S201-S203 of the illustrated embodiment will not be repeated here.

[0179] S1104: In response to an ad distribution request for the target object, retrieve the first ad that has been triggered for the target object.

[0180] It should be noted that the specific implementation process shown in step S1104 can be found in the aforementioned... Figure 5 The specific implementation process shown in step S504 of the illustrated embodiment will not be repeated here.

[0181] S1105: Obtain the set of advertisements to be distributed.

[0182] As described above, in response to an ad distribution request targeting a specific audience, the method for obtaining the ad set to be distributed can include various approaches. For example, one or more ads retrieved from a database based on the target audience's attribute information can be directly used as the second ad in the ad set to be distributed. Alternatively, one or more ads retrieved from a database based on the target audience's attribute information can be filtered according to filtering rules, and the filtered portion of ads can be determined as the second ad in the ad set to be distributed. These filtering rules include filtering based on the ad's estimated click-through rate, estimated conversion rate, and the advertiser's bid information.

[0183] The following example illustrates the process of selecting a second advertisement from one or more advertisements that match attribute information, using filtering rules based on click-through rate and conversion rate.

[0184] (1) Perform click-through rate (CTR) prediction processing on each of the one or more ads that match the attribute information to obtain the CTR prediction result for each ad. The CTR prediction result for any ad is used to indicate the probability that a user will click or trigger the ad after it is reached. Then, perform conversion rate prediction processing on each of the one or more ads that match the attribute information to obtain the conversion rate prediction result for each ad. The conversion rate prediction result for any ad is used to indicate the probability that a user will trigger and convert the ad after it is reached.

[0185] This application supports the use of predictive models to estimate the click-through rate (CTR) and conversion rate (CTR) of one or more advertisements that match the attribute information of a target object. Specifically, a CTR model can be used to estimate the CTR of distributed advertisements, and a conversion rate model can be used to estimate the conversion rate of distributed advertisements. The predictive model can be either a CTR model or a conversion rate model. Both CTR and conversion rate models can be treated as binary classification optimization modeling problems, i.e., predicting whether the target object will click or convert to the distributed advertisement. Here, pCXR can represent the probability of a click or conversion; when X is T, pCXR is pCTR, and when X is V, pCXR is pCVR.

[0186] The model structures for click-through rate (CTR) and conversion rate (CTR) models can be found in [reference needed]. Figure 12 ,like Figure 12 As shown, the prediction model uses machine learning algorithms such as deep neural networks. Based on multi-dimensional raw features (represented by x), including real-time contextual features (such as time or ad location), target object features, and historical ad performance features (such as ad clicks and impressions), it treats different objectives (such as predicting click-through rate or conversion rate) as binary classification problems. The binary classification problem has two categories: click and no click (for conversion rate models, conversion or no conversion). The optimization function or cross-entropy function of the prediction model can be found in the following definition:

[0187]

[0188] Where q represents the q-th distributed advertisement, q = 1, ..., Q, Q represents the total number of advertisements that match the attribute information of the target object, and y represents the actual click behavior that occurred; the pCTR obtained by using a prediction model to estimate the advertisement is denoted by z, and the entire prediction model can be represented as f(x), which can be expressed mathematically as z = f(x) ∈ R. CIn the pCTR model, C represents the number of classes, such as C=2, and R represents a real number. Regarding the optimization function Loss, for click-through rate (CTR) models, historical exposure-click data is typically used, where y is set to 1 for clicked samples and 0 for samples with exposure but no clicks. Batch gradient descent is then used to optimize the model. Similarly, for conversion rate (CTR) models, historical click-conversion data is typically used, where y is set to 1 for converted samples and 0 for samples with clicks but no conversions. Batch gradient descent is then used to optimize the model.

[0189] (2) Based on the click-through rate prediction results, conversion rate prediction results and bidding information, calculate the revenue prediction value for each advertisement. The revenue prediction value represents the revenue that the advertisement can bring to the advertiser when it is reached 1,000 times. The bidding information (Bid) of any distribution advertisement may include the electronic resources bid by the advertiser to reach the advertisement. In other words, when placing distribution advertisements, advertisers often expect to obtain the conversion of distribution advertisements with a certain amount of electronic resources. The larger the amount of electronic resources that advertisers spend on distribution advertisements, the more likely the advertisements of that advertiser are to be reached.

[0190] Specifically, based on the ad's click-through rate prediction, conversion rate prediction, and bidding information, the formula for calculating the ad's revenue forecast (i.e., eCPM) is as follows:

[0191] eCPM = bid × pCTR × pCVR

[0192] Here, bid refers to the advertiser's bid information for the advertisement; the higher the bid value, the more likely the advertiser's advertisement is to be reached. pCTR refers to the click probability obtained by predicting the click-through rate of the advertisement. pCVR refers to the conversion probability obtained by predicting the conversion rate of the advertisement. eCPM (effective cost per mille) represents the revenue per thousand impressions of the advertisement (such as the advertising revenue that can be obtained per thousand impressions).

[0193] (3) Among one or more advertisements that match the attribute information, the advertisement with a revenue prediction value greater than the revenue threshold is identified as the second advertisement to be distributed and added to the advertisement set to be distributed, thus obtaining the advertisement set to be distributed related to the target object. Specifically, one or more advertisements that match the attribute information of the target object can be sorted in descending order of revenue prediction value, and the advertisement with a revenue prediction value greater than the revenue threshold in the sorted sequence is identified as the second advertisement; the specific value of the revenue threshold is not limited in this embodiment of the application, but is explained here.

[0194] For example, suppose one or more ads that match the attribute information of the target object include: Ad A, Ad B, Ad C, Ad D, and Ad E. After performing the above steps on one or more ads, the eCPM of Ad A is 2, Ad B is 3, Ad C is 5, Ad D is 7, and Ad E is 1. Then, the one or more ads are sorted in descending order of revenue prediction, resulting in the sequence: Ad D → Ad C → Ad B → Ad A → Ad E. If the revenue threshold is 1.7, then Ad D, Ad C, Ad B, and Ad A, which are ranked first, are determined as the second ad and added to the ad set to be distributed.

[0195] It should be noted that the process of retrieving one or more advertisements from the database based on the attribute information of the target object is essentially a rough selection of advertisements that match the target object from the database; furthermore, the process of forming a set of advertisements to be distributed from the top one or more advertisements selected by eCPM value is essentially a fine selection process of the rough selection of advertisements, and the resulting set of advertisements to be distributed can be called the fine selection advertisement queue.

[0196] S1106: Call the trained retargeting relevance model to predict the retargeting relevance between the first ad and each second ad in the ad set to be distributed.

[0197] It should be noted that the specific implementation process shown in step S1106 can be found in [reference needed]. Figure 5 The specific implementation process of step S506 in the illustrated embodiment will not be repeated here.

[0198] S1107: Obtain the revenue forecast for each second ad in the ad set to be distributed.

[0199] It should be noted that the specific implementation process for calculating the revenue forecast value of each second advertisement can be found in the relevant description of the specific implementation process shown in step S1105 above, that is, calling the click-through rate model and / or conversion rate model to calculate the revenue forecast value of the second advertisement, which will not be elaborated here.

[0200] S1108: Determine the target advertisement from the ad set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed, and the revenue forecast of the second advertisement.

[0201] S1109: Distribute targeted ads to the target audience.

[0202] In steps S1108-S1109, this embodiment of the application performs retargeting relevance prediction processing on each second advertisement in the selected advertisement queue (i.e., the set of advertisements to be distributed sorted by eCPM value), to obtain the relevance result (or relevance score) between the first advertisement and each second advertisement in the set of advertisements to be distributed. Then, the selected advertisement queue (such as the set of advertisements to be distributed) is rearranged based on the relevance result, which can sustainably maximize the revenue of online advertising in the long run. In specific implementation, this embodiment of the application supports calculating the distribution value of each second advertisement based on the retargeting relevance between the first advertisement and each second advertisement in the set of advertisements to be distributed and the revenue prediction value (eCPM value) of each second advertisement; then, the second advertisements in the set of advertisements to be distributed are sorted in descending order of distribution value; and the target advertisement is determined from the set of advertisements to be distributed according to the sorting result.

[0203] It should be noted that in calculating the distribution value of the second ad based on the retargeting relevance between the first and second ads, and the predicted revenue of the second ad, it is necessary to calculate the retargeting relevance between the first and second ads; specifically, this retargeting relevance is calculated based on the hyperparameter γ. The formula for calculating the retargeting quality between the first and second ads is as follows:

[0204] retarget_quality=sim(D,D j )*γ

[0205] Where j represents any second advertisement in the set of advertisements to be distributed, j = 1, ..., and is an integer. γ is a hyperparameter, with a default value of 1.0, meaning that the relevance between the first and second advertisements can be directly expressed as the retargeting relevance sim(D, D) between the first and second advertisements. j In the process of reordering the second ad in the ad set to be distributed based on retargeting relevance, the higher the retargeting relevance of the second ad, the greater the retarget_quality of the relevance result. This ensures that the second ad in the selected ad queue with high relevance to the first ad most recently clicked by the target audience is more competitive during reordering; the second ad with higher competitiveness can be ranked at the front of the ad queue.

[0206] Furthermore, the distribution value of the second advertisement can be calculated based on its eCPM value (i.e., revenue forecast) and relevance results, such as using the following formula to calculate the distribution value (e.g., score):

[0207] score = eCPM + retarget_quality

[0208] The calculation methods for eCPM and retarget_quality values ​​can be found in the preceding description.

[0209] Furthermore, as described above, an ad distribution request targeting a specific object can be used to request the distribution of multiple target ads to that object. In this implementation, the number of ads requested in the ad distribution request can be obtained, and then, according to the sorting result, the second ad with the highest number of ads in the ad set to be distributed can be identified as the target ad. For example, if the ad distribution request is used to request the distribution of two target ads to a specific object, and the sorting sequence of the second ads in the ad set to be distributed is: ad D → ad C → ad B → ad A → ad E, then ad D and ad C, which are ranked highest in the sorting sequence, can be identified as target ads, and ad D and ad C can be distributed to the specific object.

[0210] In this embodiment, the retargeting correlation between the first and second advertisements triggered by the target object within a historical time period can be calculated, and the retargeting correlation result can be introduced into the ranking as a conversion intention factor to achieve the reordering of the second advertisements in the set of advertisements to be distributed. This makes the second advertisements that are more relevant to the first advertisement and more likely to be converted by the target object appear higher in the reordered advertisement queue, thereby optimizing the recommendation strategy of the advertisement distribution platform and improving advertising revenue.

[0211] The methods of the embodiments of this application have been described in detail above. In order to facilitate better implementation of the above solutions of the embodiments of this application, the apparatus of the embodiments of this application is provided below.

[0212] Figure 13 This illustration shows a schematic diagram of a model training apparatus provided in an exemplary embodiment of this application; the model training apparatus can be used as a computer program (including program code) running in a computer device, for example, the model training apparatus can be an advertising distribution platform in the computer device; the model training apparatus can be used to execute... Figure 2 , Figure 5 and Figure 11 Some or all of the steps in the method embodiments shown. Please refer to [link / reference]. Figure 13 The model training device includes the following units:

[0213] The acquisition unit 1301 is used to acquire historical advertising data, which includes M triggered advertisements, the trigger time for each advertisement, and the trigger object for each advertisement, where M is an integer.

[0214] Processing unit 1302 is used to construct a sample set based on historical advertising data. The sample set contains N sample pairs, and any one of the N sample pairs is represented as a target sample pair. The target sample pair includes ad Di, ad Di+, and ad Di-. Ad Di refers to any ad among the M ads that is triggered by a triggered object but has not been converted. Ad Di+ refers to any ad among the M ads that is triggered by the same triggered object and converted after ad Di is triggered. Ad Di- refers to any ad among the M ads that is triggered by the same triggered object but has not been converted after ad Di is triggered. N and i are both integers, and i∈[1,N].

[0215] Processing unit 1302 is used to train a retargeting relevance model using a sample set. The trained retargeting relevance model is used to predict the retargeting relevance between two ads. The retargeting relevance between two ads is used to represent the probability that the other ad will be triggered and converted when one ad is triggered.

[0216] In one implementation, the processing unit 1302, when training the retargeting correlation model using a sample set, is specifically used for:

[0217] Obtain the association information of each advertisement in the target sample pair;

[0218] The retargeting relevance model to be trained is invoked to extract features from the association information of each advertisement in the target sample pair, thereby obtaining the feature data of each advertisement.

[0219] Based on the feature data of Ad Di and Ad Di+, the retargeting relevance between Ad Di and Ad Di+ is predicted to obtain the first relevance result. The first relevance result is used to represent the probability that Ad Di+ will be triggered and converted when Ad Di is triggered.

[0220] Based on the feature data of Ad Di and Ad Di-, the retargeting relevance between Ad Di and Ad Di- is predicted to obtain a second relevance result. The second relevance result is used to represent the probability that Ad Di- will be triggered and converted when Ad Di is triggered.

[0221] The retargeting relevance model is optimized based on the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-.

[0222] In one implementation, the retargeting relevance model includes a vector transformation network and a feature classification network; the processing unit 1302, used to call the retargeting relevance model to be trained to perform feature extraction processing on the association information of each advertisement in the target sample pair, specifically for obtaining the feature data of each advertisement, is used to:

[0223] The vector transformation network is invoked to perform vector transformation processing on the association information of each advertisement in the target sample pair, so as to obtain the expression vector of each advertisement in the target sample pair;

[0224] The feature classification network is invoked to classify the expression vectors of each advertisement in the target sample pair, thereby obtaining the feature data of each advertisement in the target sample pair.

[0225] In one implementation, the retargeting relevance model further includes a first relevance calculation network and a second relevance calculation network; the first relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di+ based on the feature data of ad Di and the feature data of ad Di+, to obtain the first relevance result; the second relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di- based on the feature data of ad Di and the feature data of ad Di-, to obtain the second relevance result.

[0226] In one implementation, the processing unit 1302, when optimizing the retargeting relevance model according to the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-, specifically performs the following:

[0227] Obtain the loss function of the retargeting relevance model, which includes the target parameter;

[0228] Based on the first correlation result and the second correlation result, calculate the loss value of the loss function; where, when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is less than the target parameter, the loss value is a non-preset value; when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is greater than the target parameter, the loss value is a preset value.

[0229] If the loss value is not a preset value, the retargeting relevance model is iteratively trained.

[0230] If the loss value is preset, the retargeting correlation model reaches the convergence condition, and the trained retargeting correlation model is obtained.

[0231] In one implementation, the processing unit 1302, when constructing a sample set based on historical advertising data, specifically performs the following functions:

[0232] Select ad Di from ad M, and obtain the trigger time and trigger object corresponding to ad Di;

[0233] Based on the trigger time and triggering object corresponding to ad Di, identify ad Di+ from M ads that are triggered by the same triggering object and converted after ad Di is triggered, and construct ad Di and ad Di+ as positive sample candidate pairs;

[0234] Based on the triggering time and triggering object corresponding to ad Di, identify ad Di- from M ads that were triggered by the same triggering object but did not convert, and construct ad Di and ad Di- as negative sample candidate pairs;

[0235] Construct target sample pairs based on positive and negative candidate sample pairs; and,

[0236] Add the target sample pair to the sample set.

[0237] In one implementation, the processing unit 1302 is further configured to:

[0238] In response to an ad distribution request for a target audience, retrieve the first ad that has been triggered by the target audience. The first ad is an ad that was triggered by the target audience within a historical time period but has not been converted.

[0239] Retrieve the set of ads to be distributed, which contains one or more second ads to be distributed;

[0240] The trained retargeting relevance model is invoked to predict the retargeting relevance between the first ad and each second ad in the ad set to be distributed;

[0241] Based on the retargeting relevance between the first ad and each second ad in the ad set to be distributed, the target ad is determined from the ad set to be distributed;

[0242] Distribute targeted ads to the target audience.

[0243] In one implementation, when processing unit 1302 determines the target advertisement from the ad set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed, it specifically performs the following:

[0244] Sort the second ads in the ad set to be distributed in descending order of relevance;

[0245] Target ads are determined from the set of ads to be distributed based on the sorting results.

[0246] In one implementation, when processing unit 1302 determines the target advertisement from the ad set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed, it specifically performs the following:

[0247] Obtain the revenue forecast for each second ad in the ad set to be distributed;

[0248] Based on the retargeting relevance between the first ad and each second ad in the ad set to be distributed, and the revenue forecast of each second ad, the distribution value of each second ad is calculated.

[0249] Sort the second ads in the ad set to be distributed in descending order of their distribution values;

[0250] Target ads are determined from the set of ads to be distributed based on the sorting results.

[0251] In one implementation, the processing unit 1302, when determining the target advertisement from the set of advertisements to be distributed according to the sorting result, specifically performs the following:

[0252] Get the number of ad distributions requested in the ad distribution request;

[0253] Based on the sorting results, the second advertisement with the highest number of distributions in the order of the advertisements to be distributed is identified as the target advertisement.

[0254] In one implementation, the processing unit 1302, when acquiring the set of advertisements to be distributed, specifically performs the following functions:

[0255] Obtain the attribute information of the target object, including object attribute information and / or behavioral attribute information;

[0256] Based on the attribute information, search the database for one or more advertisements that match the attribute information;

[0257] Filter out the second ad from the matching ads and add the second ad to the set of ads to be distributed.

[0258] In one implementation, when processing unit 1302 filters out a second advertisement to be distributed from matching advertisements, it specifically performs the following functions:

[0259] For each of the matched ads, perform click-through rate (CTR) prediction processing to obtain the CTR prediction result for each ad; and

[0260] Perform conversion rate prediction processing on each of the one or more matching ads to obtain the conversion rate prediction result for each ad;

[0261] Based on the click-through rate prediction results, conversion rate prediction results, and bidding information, calculate the revenue prediction value for each advertisement. The bidding information includes the electronic resources that the advertiser bids for to reach the advertisement.

[0262] Among one or more matching ads, the ad with a revenue forecast greater than a revenue threshold is identified as the second ad to be distributed.

[0263] According to one embodiment of this application, Figure 13 The various units in the model training device shown can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effect of the embodiments of this application. The above units are based on logical function division. In practical applications, the function of one unit can also be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the model training device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented by multiple units working together. According to another embodiment of this application, the device can be run on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), and is capable of performing functions such as... Figure 2 , Figure 5 and Figure 11 The computer program (including program code) for each step involved in the corresponding method shown, to construct such... Figure 13 The model training apparatus shown herein, and the model training method for implementing the embodiments of this application, are described. The computer program may be recorded on, for example, a computer-readable recording medium, loaded onto the aforementioned computing device via the computer-readable recording medium, and run therein.

[0264] In this embodiment, the processing unit 1302 can construct a pair of samples with a comparative relationship based on M ads included in the historical advertising data. The comparative relationship of the sample pairs can be represented as follows: the relationship between ad Di (triggered but not converted by the triggered object) and ad Di+ (triggered and converted by the same triggered object after ad Di is triggered in the M ads), and ad Di- (triggered and not converted by the same triggered object after ad Di is triggered in the M ads). Since the sample pairs are constructed using the comparative relationship between ads that perform different behaviors (e.g., triggered and converted, triggered but not converted) by the triggered object, the retargeting relevance model trained based on these comparative sample pairs can be used to predict the retargeting relevance between two ads. For example, it can be used to predict the probability that a target object triggers and converts a target ad when a target object triggers an ad, ensuring the performance of the retargeting relevance model in distributing ads, and thus enabling the distribution of more relevant ads to the target object through the retargeting relevance model.

[0265] Figure 14 A schematic diagram of the structure of a computer device provided in an exemplary embodiment of this application is shown. Please refer to... Figure 14 The computer device includes a processor 1401, a communication interface 1402, and a computer-readable storage medium 1403. The processor 1401, communication interface 1402, and computer-readable storage medium 1403 can be connected via a bus or other means. The communication interface 1402 is used to receive and send data. The computer-readable storage medium 1403 can be stored in the computer device's memory and is used to store computer programs, including program instructions. The processor 1401 is used to execute the program instructions stored in the computer-readable storage medium 1403. The processor 1401 (or CPU (Central Processing Unit)) is the computing and control core of the computer device, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve corresponding method flows or corresponding functions.

[0266] This application embodiment also provides a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the processing system of the computer device. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by the processor 1401, which may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here may be high-speed RAM memory or non-volatile memory, such as at least one disk storage device; optionally, it may also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0267] In one embodiment, the computer device may be the server mentioned in the foregoing embodiments; the computer-readable storage medium stores one or more instructions; the processor 1401 loads and executes one or more instructions stored in the computer-readable storage medium to implement the corresponding steps in the above-described model training method embodiments; specifically, the one or more instructions in the computer-readable storage medium are loaded by the processor 1401 and executed as follows:

[0268] Retrieve historical advertising data, which includes the M ads that were triggered, the trigger time for each ad, and the triggering object for each ad, where M is an integer;

[0269] A sample set is constructed based on historical advertising data. The sample set contains N sample pairs. Any one of the N sample pairs is represented as a target sample pair. The target sample pair includes ad Di, ad Di+, and ad Di-. Ad Di refers to any ad among the M ads that is triggered by the triggered object but has not been converted. Ad Di+ refers to any ad among the M ads that is triggered by the same triggered object and is converted after ad Di is triggered. Ad Di- refers to any ad among the M ads that is triggered by the same triggered object but has not been converted after ad Di is triggered. N and i are both integers, and i∈[1,N].

[0270] A retargeting relevance model is trained using a sample set. The trained retargeting relevance model is used to predict the retargeting relevance between two ads. The retargeting relevance between two ads is used to represent the probability that the other ad will be triggered and convert when one ad is triggered.

[0271] In one implementation, when one or more instructions in the computer-readable storage medium are loaded by the processor 1401 and executed to train a retargeting correlation model using a sample set, the following steps are specifically performed:

[0272] Obtain the association information of each advertisement in the target sample pair;

[0273] The retargeting relevance model to be trained is invoked to extract features from the association information of each advertisement in the target sample pair, thereby obtaining the feature data of each advertisement.

[0274] Based on the feature data of Ad Di and Ad Di+, the retargeting relevance between Ad Di and Ad Di+ is predicted to obtain the first relevance result. The first relevance result is used to represent the probability that Ad Di+ will be triggered and converted when Ad Di is triggered.

[0275] Based on the feature data of Ad Di and Ad Di-, the retargeting relevance between Ad Di and Ad Di- is predicted to obtain a second relevance result. The second relevance result is used to represent the probability that Ad Di- will be triggered and converted when Ad Di is triggered.

[0276] The retargeting relevance model is optimized based on the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-.

[0277] In one implementation, the retargeting relevance model includes a vector transformation network and a feature classification network; one or more instructions in a computer-readable storage medium are loaded by the processor 1401 and, when executing the retargeting relevance model to be trained to extract features from the association information of each advertisement in the target sample pair to obtain the feature data of each advertisement, the following steps are specifically performed:

[0278] The vector transformation network is invoked to perform vector transformation processing on the association information of each advertisement in the target sample pair, so as to obtain the expression vector of each advertisement in the target sample pair;

[0279] The feature classification network is invoked to classify the expression vectors of each advertisement in the target sample pair, thereby obtaining the feature data of each advertisement in the target sample pair.

[0280] In one implementation, the retargeting relevance model further includes a first relevance calculation network and a second relevance calculation network; the first relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di+ based on the feature data of ad Di and the feature data of ad Di+, to obtain the first relevance result; the second relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di- based on the feature data of ad Di and the feature data of ad Di-, to obtain the second relevance result.

[0281] In one implementation, when one or more instructions in a computer-readable storage medium are loaded by processor 1401 and executed to optimize the retargeting relevance model according to the principle of increasing the retargeting relevance between advertisement Di and advertisement Di+, and decreasing the retargeting relevance between advertisement Di and advertisement Di-, the following steps are specifically performed:

[0282] Obtain the loss function of the retargeting relevance model, which includes the target parameter;

[0283] Based on the first correlation result and the second correlation result, calculate the loss value of the loss function; where, when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is less than the target parameter, the loss value is a non-preset value; when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is greater than the target parameter, the loss value is a preset value.

[0284] If the loss value is not a preset value, the retargeting relevance model is iteratively trained.

[0285] If the loss value is preset, the retargeting correlation model reaches the convergence condition, and the trained retargeting correlation model is obtained.

[0286] In one implementation, one or more instructions in the computer-readable storage medium are loaded by the processor 1401 and, when executing to construct a sample set based on historical advertising data, specifically perform the following steps:

[0287] Select ad Di from ad M, and obtain the trigger time and trigger object corresponding to ad Di;

[0288] Based on the trigger time and triggering object corresponding to ad Di, identify ad Di+ from M ads that are triggered by the same triggering object and converted after ad Di is triggered, and construct ad Di and ad Di+ as positive sample candidate pairs;

[0289] Based on the triggering time and triggering object corresponding to ad Di, identify ad Di- from M ads that were triggered by the same triggering object but did not convert, and construct ad Di and ad Di- as negative sample candidate pairs;

[0290] Construct target sample pairs based on positive and negative candidate sample pairs; and,

[0291] Add the target sample pair to the sample set.

[0292] In one implementation, one or more instructions in a computer-readable storage medium are loaded by processor 1401 and the following steps are also performed:

[0293] In response to an ad distribution request for a target audience, retrieve the first ad that has been triggered by the target audience. The first ad is an ad that was triggered by the target audience within a historical time period but has not been converted.

[0294] Retrieve the set of ads to be distributed, which contains one or more second ads to be distributed;

[0295] The trained retargeting relevance model is invoked to predict the retargeting relevance between the first ad and each second ad in the ad set to be distributed;

[0296] Based on the retargeting relevance between the first ad and each second ad in the ad set to be distributed, the target ad is determined from the ad set to be distributed;

[0297] Distribute targeted ads to the target audience.

[0298] In one implementation, when one or more instructions in a computer-readable storage medium are loaded by processor 1401 and executed to determine a target advertisement from the advertisement set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the advertisement set to be distributed, the following steps are specifically performed:

[0299] Sort the second ads in the ad set to be distributed in descending order of relevance;

[0300] Target ads are determined from the set of ads to be distributed based on the sorting results.

[0301] In one implementation, when one or more instructions in a computer-readable storage medium are loaded by processor 1401 and executed to determine a target advertisement from the advertisement set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the advertisement set to be distributed, the following steps are specifically performed:

[0302] Obtain the revenue forecast for each second ad in the ad set to be distributed;

[0303] Based on the retargeting relevance between the first ad and each second ad in the ad set to be distributed, and the revenue forecast of each second ad, the distribution value of each second ad is calculated.

[0304] Sort the second ads in the ad set to be distributed in descending order of their distribution values;

[0305] Target ads are determined from the set of ads to be distributed based on the sorting results.

[0306] In one implementation, when one or more instructions in the computer-readable storage medium are loaded by the processor 1401 and, upon execution, determine the target advertisement from the set of advertisements to be distributed according to the sorting result, the following steps are specifically performed:

[0307] Get the number of ad distributions requested in the ad distribution request;

[0308] Based on the sorting results, the second advertisement with the highest number of distributions in the order of the advertisements to be distributed is identified as the target advertisement.

[0309] In one implementation, one or more instructions in the computer-readable storage medium are loaded by the processor 1401 and, when executing the process of obtaining the set of advertisements to be distributed, specifically perform the following steps:

[0310] Obtain the attribute information of the target object, including object attribute information and / or behavioral attribute information;

[0311] Based on the attribute information, search the database for one or more advertisements that match the attribute information;

[0312] Filter out the second ad from the matching ads and add the second ad to the set of ads to be distributed.

[0313] In one implementation, when one or more instructions in a computer-readable storage medium are loaded by processor 1401 and executed to select a second advertisement to be distributed from matching advertisements, the following steps are performed:

[0314] For each of the matched ads, perform click-through rate (CTR) prediction processing to obtain the CTR prediction result for each ad; and

[0315] Perform conversion rate prediction processing on each of the one or more matching ads to obtain the conversion rate prediction result for each ad;

[0316] Based on the click-through rate prediction results, conversion rate prediction results, and bidding information, calculate the revenue prediction value for each advertisement. The bidding information includes the electronic resources that the advertiser bids for to reach the advertisement.

[0317] Among one or more matching ads, the ad with a revenue forecast greater than a revenue threshold is identified as the second ad to be distributed.

[0318] In this embodiment of the application, the processor 1401 can determine the target advertisement to be distributed to the target object by calculating the correlation between the first advertisement and the second advertisement triggered by the target object. This makes the target advertisement that reaches the target object not only related to the first advertisement, but also more likely to be triggered and converted by the target object. This can effectively distribute more relevant advertisements to the target object, thereby improving the conversion rate of the target advertisement.

[0319] This application also provides a computer program product or computer program, which includes 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 and executes the computer instructions, causing the computer device to perform the aforementioned model training method.

[0320] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0321] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0322] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this invention should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A model training method, characterized in that, include: Obtain historical advertising data, which includes M triggered ads, the trigger time for each ad, and the triggering object for each ad, where M is an integer; A sample set is constructed based on the historical advertising data. The sample set contains N sample pairs. Any one of the N sample pairs is represented as a target sample pair. The target sample pairs include: Ad Di and Ad Di+, Ad Di and Ad Di-. Ad Di and Ad Di+ are positive sample candidate pairs, and Ad Di and Ad Di- are negative sample candidate pairs. Ad Di refers to any ad among the M ads that is triggered by a triggered object but not converted. Ad Di+ refers to any ad among the M ads that is triggered by the same triggered object and converted after Ad Di is triggered. Ad Di- refers to any ad among the M ads that is triggered by the same triggered object but not converted after Ad Di is triggered. N and i are both integers, and i∈[1,N]. The sample set is used to train a retargeting relevance model, which is then used to predict the retargeting relevance between two ads. The retargeting relevance between the two ads represents the probability that the other ad will be triggered and convert when one ad is triggered.

2. The method as described in claim 1, characterized in that, The step of training the retargeting relevance model using the sample set includes: Obtain the association information of each advertisement in the target sample pair; The retargeting relevance model to be trained is invoked to perform feature extraction processing on the association information of each advertisement in the target sample pair, so as to obtain the feature data of each advertisement. Based on the feature data of Ad Di and Ad Di+, the retargeting correlation between Ad Di and Ad Di+ is predicted to obtain a first correlation result. The first correlation result is used to represent the probability that Ad Di+ will be triggered and converted when Ad Di is triggered. Based on the feature data of ad Di and ad Di-, the retargeting correlation between ad Di and ad Di- is predicted to obtain a second correlation result. The second correlation result is used to represent the probability that ad Di- will be triggered and converted when ad Di is triggered. The retargeting relevance model is optimized according to the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-.

3. The method as described in claim 2, characterized in that, The retargeting relevance model includes a vector transformation network and a feature classification network; the retargeting relevance model to be trained is invoked to perform feature extraction processing on the association information of each advertisement in the target sample pair to obtain the feature data of each advertisement, including: The vector transformation network is invoked to perform vector transformation processing on the association information of each advertisement in the target sample pair, so as to obtain the expression vector of each advertisement in the target sample pair. The feature classification network is invoked to classify the expression vectors of each advertisement in the target sample pair, thereby obtaining the feature data of each advertisement in the target sample pair.

4. The method as described in claim 3, characterized in that, The retargeting relevance model further includes a first relevance calculation network and a second relevance calculation network; the first relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di+ based on the feature data of ad Di and ad Di+, to obtain a first relevance result; the second relevance calculation network is used to predict the retargeting relevance between ad Di and ad Di- based on the feature data of ad Di and ad Di-, to obtain a second relevance result.

5. The method as described in claim 2, characterized in that, The optimization process for the retargeting relevance model, based on the principle of increasing the retargeting relevance between ad Di and ad Di+, and decreasing the retargeting relevance between ad Di and ad Di-, includes: Obtain the loss function of the retargeting relevance model, wherein the loss function includes the target parameter; Based on the first correlation result and the second correlation result, the loss value of the loss function is calculated; wherein, when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is less than the target parameter, the loss value is a non-preset value; when the first correlation result is greater than the second correlation result and the correlation amplitude between the two is greater than the target parameter, the loss value is a preset value. If the loss value is not a preset value, then the retargeting relevance model is iteratively trained; If the loss value is a preset value, then the redirection correlation model reaches the convergence condition, and the trained redirection correlation model is obtained.

6. The method as described in claim 1, characterized in that, The step of constructing a sample set based on the historical advertising data includes: Select advertisement Di from the M advertisements, and obtain the trigger time and the corresponding trigger object of advertisement Di; Based on the triggering time and triggering object corresponding to the advertisement Di, determine the advertisement Di+ that is triggered and converted by the same triggering object after the advertisement Di is triggered from the M advertisements, and construct the advertisement Di and the advertisement Di+ as positive sample candidate pairs; Based on the triggering time and triggering object corresponding to the advertisement Di, determine from the M advertisements the advertisement Di- that was triggered by the same triggering object but did not convert after the advertisement Di was triggered, and construct the advertisement Di and the advertisement Di- as a negative sample candidate pair; Based on the positive sample candidate pairs and the negative sample candidate pairs, the target sample pairs are constructed; and, Add the target sample pair to the sample set.

7. The method as described in claim 1, characterized in that, The method further includes: In response to an ad distribution request for a target object, obtain a first ad that has been triggered by the target object, wherein the first ad is an ad that was triggered by the target object within a historical time period but was not converted; Obtain a set of advertisements to be distributed, wherein the set of advertisements to be distributed contains one or more second advertisements to be distributed; The trained retargeting relevance model is invoked to predict the retargeting relevance between the first advertisement and each second advertisement in the advertisement set to be distributed; Based on the retargeting relevance between the first advertisement and each second advertisement in the advertisement set to be distributed, the target advertisement is determined from the advertisement set to be distributed; Distribute the target advertisement to the target audience.

8. The method as described in claim 7, characterized in that, The step of determining the target advertisement from the ad set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed includes: The second ads in the set of ads to be distributed are sorted in descending order of relevance to the retargeting. Target ads are determined from the set of ads to be distributed based on the sorting results.

9. The method as described in claim 7, characterized in that, The step of determining the target advertisement from the ad set to be distributed based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed includes: Obtain the revenue prediction value for each second advertisement in the set of advertisements to be distributed; Based on the retargeting relevance between the first advertisement and each second advertisement in the ad set to be distributed, and the revenue prediction value of each second advertisement, the distribution value of each second advertisement is calculated; The second advertisements in the set of advertisements to be distributed are sorted in descending order of their distribution values. Target ads are determined from the set of ads to be distributed based on the sorting results.

10. The method as described in claim 7 or 8, characterized in that, The step of determining the target advertisement from the set of advertisements to be distributed according to the sorting result includes: Obtain the number of ad distributions requested in the ad distribution request; Based on the sorting results, the second advertisement with the highest number of distributions in the advertisement distribution group is determined as the target advertisement.

11. The method as described in claim 7, characterized in that, The process of obtaining the set of advertisements to be distributed includes: Obtain the attribute information of the target object, the attribute information including object attribute information and / or behavioral attribute information; Based on the attribute information, search the database for one or more advertisements that match the attribute information; Select a second advertisement from the matching advertisements and add it to the set of advertisements to be distributed.

12. The method as described in claim 11, characterized in that, The step of filtering out the second advertisement to be distributed from the matching advertisements includes: Click-through rate (CTR) prediction is performed on each of the one or more matching ads to obtain a CTR prediction result for each ad; and Perform conversion rate prediction processing on each of the one or more matching ads to obtain the conversion rate prediction result for each ad; Based on the click-through rate prediction results, conversion rate prediction results, and bidding information, the revenue prediction value of each advertisement is calculated. The bidding information includes the electronic resources that the advertiser to which the advertisement belongs bid for in order to reach the advertisement. Among the matched one or more advertisements, the advertisement with a revenue prediction value greater than the revenue threshold is identified as the second advertisement to be distributed.

13. A model training device, characterized in that, include: The acquisition unit is used to acquire historical advertising data, which includes M triggered advertisements, the trigger time corresponding to each advertisement, and the triggering object corresponding to each advertisement, where M is an integer; The processing unit is configured to construct a sample set based on the historical advertising data. The sample set contains N sample pairs, and any one of the N sample pairs is represented as a target sample pair. The target sample pair includes: ad Di and ad Di+, ad Di and ad Di-, ad Di and ad Di+ are positive sample candidate pairs, and ad Di and ad Di- are negative sample candidate pairs. Ad Di refers to any ad among the M ads that is triggered by a triggered object but not converted. Ad Di+ refers to any ad among the M ads that is triggered by the same triggered object and converted after ad Di is triggered. Ad Di- refers to any ad among the M ads that is triggered by the same triggered object but not converted after ad Di is triggered. N and i are both integers, and i∈[1,N]. The processing unit is also used to train a retargeting relevance model using the sample set. The trained retargeting relevance model is used to predict the retargeting relevance between two ads. The retargeting relevance between the two ads is used to represent the probability that the other ad will be triggered and converted when one ad is triggered.

14. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program that, when executed by the processor, implements the model training method as described in any one of claims 1-12.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in claims 1-12.

16. A computer program product, characterized in that, The computer program product includes computer instructions that, when executed by a processor, implement the model training method as described in claims 1-12.

Images