Advertising recall methods, devices, electronic equipment, program products and media

Abstract

This application discloses an advertising recall method, apparatus, electronic device, program product, and medium, which can be applied to the field of data processing technology. The method includes: obtaining advertising embedding features for each advertising data point generated based on advertising association information; determining a recommendation weight for each advertising data point based on recommendation evaluation information; weighting the advertising embedding features of each advertising data point based on its recommendation weight to obtain weighted embedding features; and selecting recall advertising data for the target object based on the feature differences between the target object's object embedding features and the weighted embedding features of each advertising data point. Using this application helps improve the accuracy of advertising recall. This application can also be applied to various scenarios such as cloud technology, blockchain, artificial intelligence, smart transportation, assisted driving, and smart home appliances.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to advertising recall methods, devices, electronic devices, program products, and media. Background Technology

[0002] Currently, ad recall is widely used in various software, websites, and systems. The ads requiring recall often differ across business scenarios. For example, during specific time periods (such as major e-commerce promotions), it's often necessary to recall ads that differ significantly from daily business needs. However, the inventors have found in practice that using the same ad recall method for different business scenarios leads to lower accuracy in some scenarios. Summary of the Invention

[0003] This application provides an advertising recall method, apparatus, electronic device, program product, and medium, which helps improve the accuracy of advertising recall.

[0004] On one hand, embodiments of this application disclose an advertising recall method, which includes:

[0005] Obtain at least one piece of advertising data, and obtain the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data;

[0006] Obtain recommendation evaluation information for each ad data point, and determine the recommendation weight for each ad data point based on the recommendation evaluation information for each ad data point; the recommendation evaluation information for any ad data point includes information used to assess the probability of recommending any ad data point to an audience;

[0007] Based on the recommendation weight of each ad data, the ad embedding features of each ad data are weighted to obtain the weighted embedding features of each ad data.

[0008] Obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from at least one advertising data based on the feature differences between the object embedding features and the weighted embedding features of each advertising data.

[0009] On one hand, embodiments of this application disclose a data processing apparatus, which includes:

[0010] An acquisition unit is used to acquire at least one piece of advertising data and to acquire the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data.

[0011] The acquisition unit is also used to acquire recommendation evaluation information for each piece of advertising data, and to determine the recommendation weight for each piece of advertising data based on the recommendation evaluation information for each piece of advertising data; the recommendation evaluation information for any piece of advertising data includes information for assessing the probability of recommending any piece of advertising data to an object;

[0012] The processing unit is used to perform weighted processing on the advertising embedding features of each advertising data based on the recommendation weight of each advertising data, so as to obtain the weighted embedding features of each advertising data.

[0013] The processing unit is also used to obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from at least one advertising data according to the feature differences between the object embedding features and the weighted embedding features of each advertising data.

[0014] On one hand, embodiments of this application provide an electronic device, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to perform the following steps:

[0015] Obtain at least one piece of advertising data, and obtain the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data;

[0016] Obtain recommendation evaluation information for each ad data point, and determine the recommendation weight for each ad data point based on the recommendation evaluation information for each ad data point; the recommendation evaluation information for any ad data point includes information used to assess the probability of recommending any ad data point to an audience;

[0017] Based on the recommendation weight of each ad data, the ad embedding features of each ad data are weighted to obtain the weighted embedding features of each ad data.

[0018] Obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from at least one advertising data based on the feature differences between the object embedding features and the weighted embedding features of each advertising data.

[0019] On one hand, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, perform the following steps:

[0020] Obtain at least one piece of advertising data, and obtain the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data;

[0021] Obtain recommendation evaluation information for each ad data point, and determine the recommendation weight for each ad data point based on the recommendation evaluation information for each ad data point; the recommendation evaluation information for any ad data point includes information used to assess the probability of recommending any ad data point to an audience;

[0022] Based on the recommendation weight of each ad data, the ad embedding features of each ad data are weighted to obtain the weighted embedding features of each ad data.

[0023] Obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from at least one advertising data based on the feature differences between the object embedding features and the weighted embedding features of each advertising data.

[0024] On one hand, embodiments of this application provide a computer program product or computer program that includes computer instructions that, when executed by a processor, can implement the method provided in the above-mentioned aspect.

[0025] By employing the embodiments of this application, the recommendation weight of each advertisement data can be determined based on the recommendation evaluation information of each advertisement data. Then, the advertisement embedding features of the advertisement data are weighted based on the recommendation weight of each advertisement data to obtain weighted embedding features. Finally, recall advertisement data for the target object is determined based on the weighted embedding features of each advertisement data and the object embedding features of the target object. This allows for the determination of different recommendation weights for advertisement data with different recommendation evaluation information. The recall advertisement data is then determined based on the weighted advertisement embedding features obtained through recommendation weights. Essentially, the adjustment of the probability of advertisement recall based on recommendation evaluation information, which is required for different business scenarios, is transformed into processing based on recommendation weights. This allows for flexible adjustment of the probability of advertisement data with different recommendation evaluation information being recalled according to different business scenarios, helping to improve the accuracy and flexibility of advertisement recall. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the structure of an advertising recall system provided in an embodiment of this application;

[0028] Figure 2 This is a schematic diagram of an advertising recall scenario provided in an embodiment of this application;

[0029] Figure 3 This is a flowchart illustrating an advertising recall method provided in an embodiment of this application;

[0030] Figure 4 This is a schematic diagram illustrating the effect of a weighted embedding feature provided in an embodiment of this application;

[0031] Figure 5 This is a flowchart illustrating an advertising recall method provided in an embodiment of this application;

[0032] Figure 6 This is a schematic diagram of the structure of a feature generation network provided in an embodiment of this application;

[0033] Figure 7 This is a flowchart illustrating an advertising recall method provided in an embodiment of this application;

[0034] Figure 8 This is a schematic diagram of the structure of an advertising recall device provided in an embodiment of this application;

[0035] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0036] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0037] This application proposes an ad recall scheme that determines the recommendation weight of each ad data point based on its recommendation evaluation information. Then, it weights the ad embedding features of each ad data point based on its recommendation weight to obtain weighted embedding features. Finally, it determines the recall ad data for the target object based on the weighted embedding features of each ad data point and the object embedding features of the target object. This allows for different recommendation weights to be determined for ad data with different recommendation evaluation information. By using the weighted embedding features obtained from the recommendation weights to determine the recall ad data, the scheme essentially transforms the adjustment of the ad recall probability based on recommendation evaluation information for different business scenarios into a process based on recommendation weights. This allows for flexible adjustment of the recall probability of ad data with different recommendation evaluation information according to different business scenarios, improving the accuracy and flexibility of ad recall.

[0038] In one possible implementation, embodiments of this application can be applied to an advertising recall system. See also... Figure 1 , Figure 1This is a schematic diagram of an advertising recall system provided in an embodiment of this application. The advertising recall system may include a client and a server. The server executes the aforementioned advertising recall scheme, which involves determining the recommendation weight of each advertising data point based on its recommendation evaluation information. Then, based on the recommendation weight of each advertising data point, the server weights the advertising embedding features of the advertising data to obtain weighted embedding features. Based on the weighted embedding features of each advertising data point and the object embedding features of the target object, the server determines the recall advertising data for the target object. The server can then push the determined recall advertising data to the client corresponding to the target object, such as sending the advertising image, advertising title, etc., to the client for display. The client may be associated with a corresponding object, which can be an application on the object's device. After receiving the advertising data pushed by the server, the client can output the pushed advertising data, such as displaying the pushed advertisement.

[0039] In one embodiment, the aforementioned advertising data can be various types of advertisements. For example, the advertising data could be advertisements for products that can be recommended, accessed, or ordered within the client, such as clothing products, insurance products, etc.; the advertising data could also be advertisements for products recommended or accessed within the client, such as software application advertisements, brand advertisements, etc., without limitation. It is understood that if an object is interested in the advertisements displayed on the client, it can perform certain related actions through the client. These related actions could include clicking on the corresponding advertisement to view its details; clicking on the corresponding advertisement and making a purchase; or liking, saving, or searching for similar advertisements, etc., without limitation.

[0040] In one embodiment, the above-mentioned advertising recall scheme can also be applied to scenarios where other resources are recalled, such as video data (e.g., short video data in short video applications), audio data, graphic and text information data (e.g., blog posts, news information, etc. on social media platforms), etc. That is, the advertising data in the above-mentioned advertising recall scheme can be replaced with other resource data, thereby using the above method to achieve resource recall of video data, audio data, graphic and text information data, etc.

[0041] In some scenarios, this application can be applied to ad recall scenarios; please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram illustrating an ad recall scenario provided in an embodiment of this application. In the ad recall scenario, some or all of the ads in the application corresponding to the client can be retrieved, such as... Figure 2As shown in 201, it is possible to obtain advertisement a, advertisement b, advertisement c, etc.; then, the advertisement embedding features of the advertisements can be determined, and the recommendation weight can be determined based on the advertisement recommendation evaluation information. For example, the advertisement embedding feature a and recommendation weight a corresponding to advertisement a can be obtained, the advertisement embedding feature b and recommendation weight b corresponding to advertisement b can be obtained, and so on, to obtain the advertisement embedding features and recommendation weights corresponding to each advertisement. Then, a weighted processing is performed based on the advertisement embedding features and the corresponding recommendation weights (e.g., ...). Figure 2 As shown in 202), the corresponding weighted embedding features can be obtained (such as...). Figure 2 (As shown in 203).

[0042] For example, weighted embedding feature a is obtained by weighting the ad embedding feature a corresponding to ad a and the recommendation weight a. Weighted embedding feature b is obtained by weighting the ad embedding feature b corresponding to ad b and the recommendation weight b. And so on, the weighted embedding features corresponding to each ad can be obtained. Thus, the ad embedding features of each ad can be weighted by the recommendation weight, thereby influencing the feature difference between the ad embedding features and the object embedding features through the recommendation weight, so as to adjust the likelihood of the ad being recalled.

[0043] In a recall scenario, it is possible to obtain the objects for ad recall and ad recommendation (e.g., Figure 2 As shown in 204), such as object 1, object 2, object 3, etc.; then the object embedding features of each object can be obtained (e.g., Figure 2 As shown in 205), for example, obtaining object embedding feature 1 corresponding to object 1, object embedding feature 2 corresponding to object 2, and so on. Furthermore, for each object embedding feature, the feature difference between the object embedding feature and each weighted embedding feature can be calculated, thereby determining the corresponding recall advertising data (such as...) from multiple advertising data sets. Figure 2 As shown in Figure 206, for example, for object 1, the feature difference between object embedding feature 1 and each weighted embedding feature can be calculated, and the ad data with smaller feature differences can be used as the recall ad data; for object 2, the feature difference between object embedding feature 2 and each weighted embedding feature can be calculated, and the ad data with smaller feature differences can be used as the recall ad data, and so on, so as to calculate the recall ad data corresponding to each object that needs ad recall. Since the feature difference between the weighted embedding feature and the object embedding feature is used to determine the recall ad, it actually affects the feature difference between the ad embedding feature and the object embedding feature through recommendation evaluation information, thereby affecting the probability of the ad being recalled, making the recalled ad more in line with the needs of the business scenario, and helping to improve the accuracy of ad recall.

[0044] In one possible implementation, the embodiments of this application can be applied to the field of artificial intelligence (AI) technology. AI is the theory, methods, technology, and application system that uses 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 obtain 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 type of intelligent machine that can react in a way similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0045] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including 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 primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0046] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user-related data (such as object association information). These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their relevant data is being collected. This ensures that the application only begins the steps for collecting user-related data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without receiving confirmation from the user), the steps for collecting user-related data end, meaning no user-related data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of relevant user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0047] The technical solution of this application can be applied to electronic devices, such as the server mentioned above. The electronic device can be a terminal, a server, or other devices used for advertising recall; this application does not limit the scope. Optionally, the server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Terminals include, but are not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, smart speakers, and other smart home appliances.

[0048] It is understood that the above scenarios are merely examples and do not constitute a limitation on the application scenarios of the technical solutions provided in the embodiments of this application. The technical solutions of this application can also be applied to other scenarios. For example, as those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0049] Based on the above description, this application proposes an advertisement recall method. Please refer to [link to relevant documentation]. Figure 3 , Figure 3 This is a flowchart illustrating an advertising recall method provided in an embodiment of this application. The method can be executed by the aforementioned electronic device. The advertising recall method may include the following steps.

[0050] S301. Obtain at least one piece of advertising data, and obtain the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data.

[0051] The at least one piece of advertising data can be all or part of the advertisements in the advertisement library corresponding to the client. That is, the at least one piece of advertising data can be a part of the advertisements selected from the advertisement library according to certain filtering conditions. The filtering conditions can be: the industry category to which the advertisement belongs is a preset industry category, the publication time of the advertisement is within the target time range, the popularity of the advertisement (such as being reflected by the number of likes, reposts, comments, and visits) is greater than or equal to a threshold, etc., and there are no restrictions here.

[0052] The ad embedding features of advertising data can refer to the features determined based on the ad association information of the advertising data. This ad association information can be information associated with the advertising data itself, such as the ad's description, title, image, etc., without limitation. Optionally, the ad embedding features of the advertising data can be represented as a feature vector or a feature matrix, without limitation.

[0053] In one embodiment, an ad embedding feature can be generated based on ad association information of ad data by calling a trained feature generation network. Specifically, obtaining the ad embedding feature for each ad data point can include the following steps: obtaining ad association information for each ad data point, and calling the trained feature generation network to generate ad embedding features for each ad data point based on this ad association information. The trained feature generation network can be trained on a pair of sample data, which may include object association information of sample objects and ad association information of sample ads. This allows for the acquisition of accurate features representing ad data based on the trained feature generation network, improving the accuracy of subsequent ad recall.

[0054] S302. Obtain the recommendation evaluation information for each advertising data, and determine the recommendation weight for each advertising data based on the recommendation evaluation information for each advertising data.

[0055] The recommendation evaluation information for any advertising data includes information used to assess the probability of recommending any advertising data to a target audience. This recommendation evaluation information can be used to determine the recommendation weight of the advertising data, and then, based on the determined recommendation weight, to assess the probability (also known as likelihood) of recommending any advertising data to a target audience. It is understandable that in some scenarios, a higher value in the recommendation evaluation information of advertising data indicates a higher probability of the advertising data being recommended; conversely, in other scenarios, a higher value in the recommendation evaluation information indicates a lower probability of the advertising data being recommended. This depends specifically on the actual business needs. Therefore, the recommendation evaluation information for advertising data can be determined based on actual business needs, thereby assessing the probability of recommending any advertising data to a target audience.

[0056] In one possible implementation, the recommended evaluation information of the required indicator categories can be obtained according to the actual business scenario. The recommended evaluation information of each advertising data includes at least one of the following: advertising value information of the advertising data, advertising impact of the advertising data, or advertising category information of the advertising data.

[0057] The advertising value information refers to the agreed-upon value required to push the advertisement. For example, before placing an advertisement, the advertiser and the advertiser need to agree on the price for placing the advertisement, and the agreed-upon price is a kind of advertising value information.

[0058] The ad impact score describes the effect of an ad after it is published. This score can be represented by various metrics, such as likes, shares, favorites, comments, conversion rate, number of orders, or click-through rate. It can also be a combination of multiple metrics (e.g., adding multiple metrics according to certain weights). There are no restrictions here.

[0059] The advertising category information describes the category to which the advertisement belongs. Generally, each advertisement can be labeled with a corresponding category tag by humans or machines. The corresponding advertising category information can be determined based on the advertising category tag. For example, the advertising category information can be used to indicate the industry to which the advertisement belongs (such as manufacturing industry, Internet industry, energy industry, etc.), the style of the advertisement (such as lively style, serious style, neutral style, etc.), the theme of the advertisement (such as historical theme, current event theme, animal theme, lifestyle theme, etc.), the product category of the advertisement (such as clothing advertisement, furniture advertisement, electronic product advertisement, etc.), etc. There are no restrictions here.

[0060] The recommendation weight of ad data can be used to adjust the likelihood of ad data being recalled. A higher recommendation weight increases the likelihood of ad data being recalled, while a lower recommendation weight decreases the likelihood. Understandably, since a recommendation weight is introduced for each ad data point during subsequent ad recall, ad data with a higher recommendation weight is essentially "upgraded" compared to ad data with a lower recommendation weight, and vice versa. Therefore, this recommendation weight can be used to upgrade or downgrade ad data, thereby adjusting the likelihood of ad data being recalled.

[0061] In one possible implementation, determining the recommendation weight for each piece of advertising data based on the recommendation evaluation information for each piece of advertising data may specifically include the following steps:

[0062] ① Obtain the weight transformation function for the recommendation evaluation information of each ad data point, and obtain the associated evaluation parameters for the recommendation evaluation information of each ad data point. The weight transformation function is the function used to determine the recommendation weight based on the recommendation evaluation information. The associated evaluation parameters for the recommendation evaluation information of ad data refer to the specific values ​​indicated in the recommendation evaluation information. When the recommendation evaluation information is ad value information, the associated evaluation parameters can be specific value values. For example, if the ad value information of a certain ad data point is obtained, and the specific value value corresponding to that ad data is 500, then the associated evaluation parameter for this type of recommendation evaluation information is 500. When the recommendation evaluation information is ad influence, the associated evaluation parameters can be specific values ​​of influence, such as the number of likes, the number of shares, the conversion rate, etc., without restriction. For example, if the ad influence of a certain ad data point is obtained, and the number of likes corresponding to that ad data is 300, then the associated evaluation parameter for this type of recommendation evaluation information is 300. When the recommendation evaluation information is advertising category information, the associated evaluation parameter of the recommendation evaluation information can be a preset weight value corresponding to the advertising category. For example, the preset weight value corresponding to an advertising category can be set to 0.

[0063] ② Based on the weight transformation function, the correlation evaluation parameters corresponding to each ad data point are processed using parameter calculations to obtain the recommendation weight for each ad data point. It can be understood that processing the correlation evaluation parameters corresponding to each ad data point using the weight transformation function is equivalent to using the correlation evaluation parameters as values ​​in the weight transformation function for calculation.

[0064] In one embodiment, the weight transformation function can be a monotonically increasing function, that is, the larger the correlation evaluation parameter corresponding to the advertising data, the larger the corresponding recommendation weight, and the smaller the correlation evaluation parameter corresponding to the advertising data, the smaller the corresponding recommendation weight. Thus, the recommendation weight determined by the monotonically increasing function can increase the likelihood of advertising data with large correlation evaluation parameters being recalled, and reduce the likelihood of advertising data with small correlation evaluation parameters being recalled.

[0065] In one embodiment, the weight transformation function can be a monotonically decreasing function, meaning that the larger the correlation evaluation parameter corresponding to the advertising data, the smaller the corresponding recommendation weight, and the smaller the correlation evaluation parameter corresponding to the advertising data, the larger the corresponding recommendation weight. Thus, the recommendation weight determined by the monotonically decreasing function can increase the likelihood of advertising data with small correlation evaluation parameters being recalled, and reduce the likelihood of advertising data with large correlation evaluation parameters being recalled.

[0066] In one possible implementation, determining the recommendation weight for each advertisement based on its recommendation evaluation information can specifically include the following steps: ① Determining the initial recommendation weight for each advertisement based on its recommendation evaluation information. This initial recommendation weight refers to the weight without standardization. The method for obtaining the initial recommendation weight specifically includes: obtaining the weight transformation function for the recommendation evaluation information of each advertisement, and obtaining the associated evaluation parameters for the recommendation evaluation information of each advertisement; performing parameter calculations on the associated evaluation parameters corresponding to each advertisement based on the weight transformation function to obtain the initial recommendation weight for each advertisement. In other words, the value obtained directly from the weight transformation function and the associated evaluation parameters is the initial recommendation weight. Furthermore, the initial recommendation weight needs to be standardized to obtain the final recommendation weight, thereby unifying the data scale and facilitating related calculations based on the recommendation weight. The relevant descriptions of the weight transformation function and the associated evaluation parameters can be found above and will not be repeated here.

[0067] ② The initial recommendation weights for each ad data point are standardized to obtain the recommendation weight for each ad data point. It is understood that the purpose of standardizing the initial recommendation weights for each ad data point (referred to as standardization) is to ensure that the recommendation weights can be uniformly measured for application in various scenarios and to facilitate calculation. In one embodiment, standardizing the initial recommendation weights for each ad data point can convert the recommendation weights to a range of 0-1, thereby unifying the calculation standard and facilitating subsequent conversion of ad embedding features. For example, standardizing the initial recommendation weights can obtain the maximum and minimum values ​​(or the possible maximum or minimum values ​​of the initial recommendation weights for ad data) among multiple ad data points, and calculate the difference between the maximum and minimum values ​​to obtain the target weight difference. Therefore, when standardizing the initial recommendation weights to obtain the recommendation weights, the minimum value can be subtracted from the initial recommendation weights and then divided by the target weight difference to obtain the recommendation weights.

[0068] In some scenarios, when recalling ads, it's necessary to prioritize ads with high value. This means giving higher weight to ads with larger correlation evaluation parameters indicated by their ad value information. A monotonically increasing function can be used to achieve this. For example, the following function (as shown in Equation 1) can be selected as the weight transformation function to determine the recommendation weight.

[0069] f(x) = log(1 + x) t ) Formula 1

[0070] Here, x can represent the correlation evaluation parameter corresponding to the advertising data. That is, the correlation evaluation parameter corresponding to each advertising data is processed by the weight transformation function, and the correlation evaluation parameter is used as x in the weight transformation function for calculation. t can be a preset hyperparameter. Therefore, by adjusting the value of t, the smoothness of the weight transformation function can be changed. Thus, by controlling the value of t, the recommendation weight is prevented from being excessively boosted when boosting the weight of high-value advertising data.

[0071] For example, if t is 1, the weight transformation function is f(x) = log(1+x). If the correlation index of the advertising value information corresponding to advertising data a is 500, then the initial recommendation weight is log(1+500) ≈ 2.7. If the correlation index of the advertising value information corresponding to advertising data b is 999, then the initial recommendation weight is log(1+999) = 3, and so on, to obtain the initial recommendation weight for each advertising data. Furthermore, the initial recommendation weight for each advertising data can be standardized, such as converting the initial recommendation weight to the 0-1 range to obtain the corresponding recommendation weight. For example, among the initial recommendation weights corresponding to multiple ad data, the largest initial recommendation weight is 8 and the smallest initial recommendation weight is 1. After standardizing the initial recommendation weight of 2.7, the recommendation weight obtained is (2.7-1) / (8-1) = 0.24. After standardizing the initial recommendation weight of 3, the recommendation weight obtained is (3-1) / (8-1) = 0.28. And so on, the recommendation weight corresponding to each ad data can be obtained, which will not be elaborated here.

[0072] In some scenarios, when recalling ads, it's necessary to reduce the likelihood of videos with high impact (such as orders, likes, favorites, etc.) being recalled. This prevents high-impact ads from being overexposed while low-impact ads receive insufficient exposure, leading to even lower impact. In other words, videos with high correlation evaluation parameters indicated by ad impact need to be downweighted. This can be achieved by using a monotonically decreasing function. For example, the following function (as shown in Equation 2) can be used as a weight transformation function to determine the recommendation weight.

[0073] f(x) = (ax + 1) t Formula 2

[0074] Where a is a constant, a < 0; x can represent the correlation evaluation parameter corresponding to the advertising data, that is, the correlation evaluation parameter corresponding to each advertising data is processed by the weight transformation function, and the correlation evaluation parameter is used as x in the weight transformation function for calculation. t can be a preset hyperparameter, which needs to make f(x) a monotonically decreasing function when x is greater than zero. Therefore, the smoothness of the weight transformation function can be changed by adjusting the value of t, thereby controlling the value of t to avoid excessive reduction in the weight of recommendations with high influence when reducing their weight.

[0075] For example, if t is 1 and a is -1, then the weight transformation function is f(x) = -x + 1. If the correlation index information (i.e., the number of likes) of the ad influence corresponding to ad data c is 80, then the recommendation weight is -80 + 1 = -79. If the correlation index information of the ad value information corresponding to ad data d is 120, then the recommendation weight is -120 + 1 = -119. If the correlation index information of the ad value information corresponding to ad data d is 0, then the recommendation weight is 1, and so on, to obtain the recommendation weight corresponding to each ad data. Furthermore, the initial recommendation weight of each ad data can be standardized, such as converting the initial recommendation weight to the 0-1 range to obtain the corresponding recommendation weight. For example, among the initial recommendation weights corresponding to multiple ad data, the largest initial recommendation weight is 1 and the smallest initial recommendation weight is -300. After standardizing the initial recommendation weight of -79, the resulting recommendation weight is (-79+300) / (1-(-300)) = 0.73. After standardizing the initial recommendation weight of -119, the resulting recommendation weight is (-119+300) / (1-(-300)) = 0.60. And so on, the recommendation weight corresponding to each ad data can be obtained, which will not be elaborated here.

[0076] In some scenarios, when conducting ad recall, it is only necessary to reduce the likelihood of ads in the target ad category being recalled, with videos having higher impact receiving a greater reduction. In other words, the recommendation weights for ad data in all ad categories other than the target ad category should be the maximum. For ad data in the target ad category, videos with high correlation evaluation parameters can be downgraded based on ad impact. This allows for the integration of one or more of the aforementioned recommendation evaluation information to determine recommendation weights, enabling flexible allocation of recommendation weight determination methods according to actual needs, which helps improve user experience and the accuracy of ad recall.

[0077] For example, to reduce the weight of videos with high correlation evaluation parameters based on advertising impact, a monotonically decreasing function can be selected to calculate the correlation evaluation parameters of advertising data under the target advertising category. For instance, formula 2 above can be used for calculation, and the correlation evaluation parameters of advertising data under other advertising categories besides the target advertising category can be set to 0. This ensures that the recommendation weights of advertising data under other advertising categories besides the target advertising category are all the maximum recommendation weights, which is equivalent to not reducing the weights of advertising data under other advertising categories. This achieves the customization of recommendation weights for advertising data.

[0078] In one possible implementation, embodiments of this application can fuse multiple recommendation evaluation information to determine recommendation weights, thereby improving the flexibility of the process for determining recommendation weights. In one embodiment, a corresponding weight transformation function can be applied to different recommendation evaluation information to obtain corresponding initial recommendation weights. Then, the initial recommendation weights corresponding to each type of recommendation evaluation information are fused to obtain the recommendation weight corresponding to the advertising data. For example, the initial recommendation weights corresponding to each type of recommendation evaluation information can be added, averaged, or weighted averaged to obtain the recommendation weight corresponding to the advertising data. Furthermore, fusing the initial recommendation weights corresponding to each type of recommendation evaluation information to obtain the recommendation weight corresponding to the advertising data also includes standardizing the fused initial recommendation weights corresponding to each type of recommendation evaluation information to obtain the recommendation weight corresponding to the advertising data.

[0079] For example, in some scenarios, when recalling ads, it's necessary to prioritize recalling high-value ads and reduce the likelihood of ads with high order volumes being recalled. This avoids overexposure of high-order-volume ads, which essentially means increasing the weight of ads with high correlation evaluation parameters indicated by ad value information and decreasing the weight of videos with high correlation evaluation parameters indicated by ad influence. Therefore, a monotonically increasing function can be used as the weight transformation function corresponding to ad value information to determine the initial recommendation weight, and a monotonically decreasing function can be used as the weight transformation function corresponding to ad influence to determine the initial recommendation weight. The initial recommendation weight of an advertising data can be obtained by fusing the initial recommendation weight corresponding to the advertising value information and the initial recommendation weight corresponding to the advertising influence. For example, for advertising data a, the initial recommendation weight corresponding to the advertising value information can be calculated as x1 based on Formula 1 above, and the initial recommendation weight corresponding to the advertising influence can be calculated as x2 based on Formula 2 above. Then, x1 + x2 can be used to fuse the initial recommendation weights corresponding to various recommendation evaluation information. The fused initial recommendation weights are then standardized to obtain the recommendation weight of advertising data a. By analogy, the recommendation weight corresponding to each advertising data can be obtained.

[0080] S303. Based on the recommendation weight of each advertising data, the advertising embedding features of each advertising data are weighted to obtain the weighted embedding features of each advertising data.

[0081] The weighted embedding feature refers to the advertising embedding feature after weighting the advertising embedding feature with recommendation weights.

[0082] In one possible implementation, weighting the ad embedding features can be achieved by multiplying the ad embedding features by their corresponding recommendation weights. Optionally, to facilitate calculation, the weighted ad embedding features need to be standardized to obtain the weighted embedding features for each ad data point. For example, the ad embedding features can be divided by their corresponding modulus, and the modulus of the weighted embedding features can be made equal to 1 by concatenating the corresponding values.

[0083] For example, see Figure 4 , Figure 4 This is a schematic diagram illustrating the effect of a weighted embedding feature provided in an embodiment of this application. For example... Figure 4 As shown, the advertising embedding feature of advertising data can be b (e.g., Figure 4 As shown in 401), the weighted embedding features obtained from the advertising embedding features can be as follows: Figure 4 As shown in 402. When converting ad embedding features into weighted embedding features, they can be multiplied by... Where q is the initial recommendation weight mentioned above, and k is the difference between the maximum and minimum values ​​of the initial recommendation weight (i.e., the target weight difference), that is... The representation is the recommendation weight after standardization; it can be divided by ||b||, that is, divided by the modulus corresponding to the ad embedding feature; it can also be concatenated so that the modulus of the weighted embedding feature is 1, that is, concatenated.

[0084] S304. Obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from at least one advertising data according to the feature differences between the object embedding features and the weighted embedding features of each advertising data.

[0085] The object embedding feature of the target object can refer to the features determined based on the object association information of the target object. This object association information can be information associated with the recommended object, such as information inherent to the object itself, like its age, industry, style preferences, etc. It can also be contextual information, such as the frequency of the advertisements the object has recently liked, shared, or searched for, and the frequency of its interactions with advertisements (e.g., liking, sharing, placing orders, etc.). There are no restrictions here. Optionally, the initial object embedding feature of the recommended object can be represented as a feature vector or a feature matrix. It is understood that the target object can be any object in the client's application, a registered object, or an object operated as a guest. There are no restrictions here.

[0086] In one embodiment, a pre-trained feature generation network can be invoked to generate object embedding features of the target object based on its object association information. Specifically, obtaining the object embedding features of the target object can include the following steps: obtaining the object association information of the target object, and invoking the pre-trained feature generation network to generate the object embedding features of the target object based on this object association information. The pre-trained feature generation network can be trained on sample data pairs, which may include object association information of sample objects and ad association information of sample ads. Therefore, by obtaining the object embedding features of the target object based on the pre-trained feature generation network, accurate object embedding features representing the target object can be obtained, improving the accuracy of subsequent ad recall.

[0087] In one possible implementation, obtaining the object embedding features of the target object can also involve: calling a trained feature generation network to generate initial object embedding features of the target object based on the object association information of the target object; and standardizing the initial object embedding features to obtain the final object embedding features. It is understood that standardizing the initial object embedding features is necessary to perform calculations with the standardized weighted embedding features to determine the ads to be recalled. Standardizing the initial object embedding features involves dividing the ad embedding feature by its modulus and appending one zero dimension to maintain consistency with the dimension of the weighted embedding features. It is understood that the initial object embedding features obtained from the trained feature generation network have the same dimension as the ad embedding features obtained from the trained feature generation network, and the standardized weighted embedding features also have the same dimension as the object embedding features; this is essential for the subsequent ad recall process.

[0088] For example, such as Figure 4As shown, the initial object embedding feature of the target object can be a (e.g., Figure 4 As shown in 403 (in the diagram), the object embedding features obtained after standardizing the initial object embedding features can be as follows: Figure 4 As shown in 404. When standardizing the initial object embedding features, you can divide by ||a||, that is, divide by the modulus corresponding to the initial object embedding features; you can also concatenate one dimension of 0 and add 1 to the dimension, thereby obtaining object embedding features with the same dimension as the weighted embedding features.

[0089] In one possible implementation, the recalled advertising data refers to advertisements recommended to the target audience. Based on the object embedding features and the weighted embedding features of each advertisement data point, recalled advertising data for the target audience is selected from at least one advertisement data point. This is equivalent to using the advertisement data corresponding to the weighted embedding features whose feature differences with the object embedding features satisfy a feature condition as the recalled advertising data for the target audience. In one embodiment, the feature difference between the weighted embedding features and the object embedding features can be characterized by the distance (such as Euclidean distance or cosine distance) or inner product between the weighted embedding features and the object embedding features; this is not limited here. It is understood that the larger the distance between the weighted embedding features and the object embedding features, the larger the feature difference; the smaller the distance, the smaller the feature difference. Similarly, the larger the inner product, the smaller the feature difference; and the smaller the inner product, the larger the feature difference. In one embodiment, satisfying the feature condition can be based on the ranking position of the feature difference or the numerical value of the feature difference meeting certain conditions.

[0090] For example, when the feature difference refers to distance, the feature condition can be that the sorting position is greater than a threshold when sorted by distance from smallest to largest, or the feature condition can be that the distance value is less than or equal to the threshold; as another example, when the feature difference refers to inner product, the feature condition can be that the sorting position is greater than a threshold when sorted by inner product from largest to smallest, or the feature condition can be that the inner product value is greater than or equal to the threshold; or the feature condition can also be that both the sorting position of the feature difference and the magnitude of the feature difference satisfy the condition, without any restrictions here.

[0091] Understandably, the recalled advertising data can usually be determined based on the feature differences between the features corresponding to the advertisement and the features corresponding to the object. However, the features used in this application to ultimately determine the recalled advertising data are weighted embedding features obtained by weighting the features with recommendation weights. This allows the recalled advertisements to be influenced by the recommendation weights, that is, by the recommendation evaluation information of the advertising data. With the same object and advertising data, if the recommendation weight is larger, the feature difference between the weighted embedding features of the advertising data and the object embedding features is smaller (smaller distance, larger inner product); if the recommendation weight is smaller, the feature difference between the weighted embedding features of the advertising data and the object embedding features is larger (larger distance, smaller inner product).

[0092] By employing the embodiments of this application, the recommendation weight of each advertisement data can be determined based on the recommendation evaluation information of each advertisement data. Then, the advertisement embedding features of the advertisement data are weighted based on the recommendation weight of each advertisement data to obtain weighted embedding features. Finally, recall advertisement data for the target object is determined based on the weighted embedding features of each advertisement data and the object embedding features of the target object. This allows for the determination of different recommendation weights for advertisement data with different recommendation evaluation information. The recall advertisement data is then determined based on the weighted advertisement embedding features obtained through recommendation weights. Essentially, the adjustment of the probability of advertisement recall based on recommendation evaluation information, which is required for different business scenarios, is transformed into processing based on recommendation weights. This allows for flexible adjustment of the probability of advertisement data with different recommendation evaluation information being recalled according to different business scenarios, helping to improve the accuracy and flexibility of advertisement recall.

[0093] Please see Figure 5 , Figure 5 This is a flowchart illustrating an advertising recall method provided in an embodiment of this application. The method can be executed by the aforementioned electronic device. The advertising recall method may include the following steps.

[0094] S501. Obtain at least one piece of advertising data, and obtain the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data.

[0095] Step S501 can be referred to the relevant description of step S301 above, and will not be repeated here.

[0096] In one possible implementation, as described above, the ad embedding features of each ad data point can be obtained based on a trained feature generation network. Before obtaining the ad embedding features of each ad data point, the feature generation network can be trained first. This allows the trained feature generation network to generate corresponding object embedding features based on object association information, and also to generate corresponding ad embedding features based on ad association information. Specifically, the training process for the feature generation network can include the following steps:

[0097] ① Obtain sample data pairs. Each sample data pair contains object association information for the sample object and ad association information for the sample ad. The sample data pair has a sample tag, which indicates whether the sample object has a target behavior or not in relation to the sample ad. This target behavior can be any one or more of the aforementioned associated behaviors, such as clicking on the corresponding ad to view ad details, clicking on the corresponding ad and making a purchase, liking or favorited the ad, searching for similar ads, etc., which will not be elaborated here. In one embodiment, if the sample object has a target behavior in relation to the sample ad, the sample data pair can be called a positive sample data pair; if the sample object does not have a target behavior in relation to the sample ad, the sample data pair can be called a negative sample data pair.

[0098] In one embodiment, targeted sample data pairs can be selected based on the application scenario of ad recall. For example, if the application scenario is ad recall based on click-through rate (CTR), then ads with click records on the target object can be used as positive sample data pairs, and ads that were exposed to the target object but had no click records can be used as negative sample data pairs. Similarly, if the application scenario is ad recall based on gross merchandise volume (GMV), then ads with order records on the target object can be used as positive sample data pairs, and ads that were exposed to the target object but had no order records or ads with shallow conversion records can be used as negative sample data pairs. Selecting targeted sample data pairs based on the application scenario can improve certain metrics, such as AUC (an evaluation metric) or recall, thereby improving ad accuracy when using features generated by the feature generation network for ad recall in real-world scenarios.

[0099] ② The feature generation network is invoked to generate sample object embedding features based on the object association information of the sample objects, and the feature generation network is invoked to generate sample ad embedding features based on the ad association information of the sample ads. This feature generation network can be a dual-tower model structure, meaning that the networks generating object embedding features and ad embedding features are different networks. For example, please refer to [link to example]. Figure 6 , Figure 6This is a schematic diagram of the structure of a feature generation network provided in an embodiment of this application. For example... Figure 6 The 601 in the diagram represents the network used to generate object embedding features, such as... Figure 6 Figure 602 shows the network used to generate advertising embedding features.

[0100] Specifically, the network used to generate object embedding features may include an information input layer to receive object association information (such as...) Figure 6 As shown in 603), it can then be encoded and represented using a custom network, such as using an SE Block (a type of neural network). Further representation learning can then be performed using a series of custom networks, such as multiple fully connected layers (FC), to obtain the corresponding object embedding features (e.g., ...). Figure 6 (As shown in 604).

[0101] The network used to generate ad embedding features may include an information input layer to receive ad-related information from the ad data (such as...). Figure 6 As shown in 605), it can then be encoded and represented using a customized network, such as SEBlock (a type of neural network). Further representation learning can then be performed using a series of customized networks, such as multi-layer fully connected (FC) layers, to obtain the corresponding advertising embedding features (e.g., ...). Figure 6 (As shown in 606).

[0102] ③ Based on the sample object embedding features, sample ad embedding features, and sample labels, the network parameters of the feature generation network are corrected to obtain a trained feature generation network. The trained feature generation network is used to generate features for both object and ad data. It can be understood that when the feature generation network is called to generate ad embedding features, the network used to generate ad embedding features in the trained feature generation network is mainly used; similarly, when the feature generation network is called to generate object embedding features, the network used to generate object embedding features in the trained feature generation network is mainly used.

[0103] In one embodiment, the network parameters of the feature generation network are corrected based on the sample object embedding features, sample advertisement embedding features, and sample labels to obtain a trained feature generation network. Specifically, this may include the following steps: obtaining the feature differences between the sample object embedding features and the sample advertisement embedding features; obtaining the feature generation deviation of the feature generation network for the sample object embedding features and the sample advertisement embedding features based on the feature differences between the sample object embedding features and the sample advertisement embedding features and the sample labels; and correcting the network parameters of the feature generation network based on the feature generation deviation to obtain a trained feature generation network.

[0104] Here, the feature difference can refer to the difference between the embedded features of the sample object and the embedded features of the sample advertisement. This feature difference can be characterized by distance (such as Euclidean distance or cosine distance) or inner product, etc., without restriction. This feature generation bias can be calculated by calling the loss function, and this feature generation bias can also be called the loss value. It is understandable that in the process of correcting the network parameters of the feature generation network based on the feature generation bias, the feature generation bias needs to be gradually reduced until it converges. During the training of the feature generation network, if the sample label indicates that the sample object has the target behavior for the sample advertisement (i.e., the sample data pair is a positive sample data pair), then the feature difference between the embedded features of the sample object and the embedded features of the sample advertisement needs to be gradually reduced, that is, the distance gradually decreases or the inner product gradually increases; if the sample label indicates that the sample object does not have the target behavior for the sample advertisement (i.e., the sample data pair is a negative sample data pair), then the feature difference between the embedded features of the sample object and the embedded features of the sample advertisement needs to be gradually increased, that is, the distance gradually increases or the inner product gradually decreases. In this way, the feature generation network can accurately generate features of objects and advertisements, which is beneficial to improving the accuracy of advertisement recall.

[0105] For example, such as Figure 6 As shown, after obtaining the sample object embedding features of the sample object and the sample advertisement embedding features of the sample advertisement based on the feature generation network, the feature differences between the sample advertisement embedding features and the sample object embedding features can be determined (e.g., Figure 6 As shown in 607 in the figure, the feature generation bias can be obtained based on the feature difference and sample label (as shown in 608 in the figure). The network parameters of the feature generation network can be corrected based on the feature generation bias to obtain the trained feature generation network.

[0106] Understandably, the training process of the feature generation network does not involve introducing recommendation weights for advertisements. Instead, recommendation weights are introduced when determining which advertisements need to be recalled. These weights are then applied to the advertisement embedding features obtained from the trained feature generation network. This allows the embedded features of advertisements and objects obtained from the trained feature generation network to be applied to different recall scenarios without needing to retrain the feature generation network to adapt to each scenario. In other words, regardless of the scenario, the corresponding recommendation weights can be obtained by adjusting the recommendation evaluation information and weight transformation function required for the recall scenario. This allows for flexible weighting or deweighting of advertisement data to adjust the likelihood of advertisement data being recalled, thereby determining advertisement data that is more suitable for the corresponding recall scenario and improving the accuracy of the advertisement recall process.

[0107] S502. Obtain the recommendation evaluation information for each advertising data, and determine the recommendation weight for each advertising data based on the recommendation evaluation information for each advertising data.

[0108] S503. Based on the recommendation weight of each advertising data, the advertising embedding features of each advertising data are weighted to obtain the weighted embedding features of each advertising data.

[0109] S504. Obtain object embedding features generated based on object association information of the target object.

[0110] The relevant descriptions of steps S502-S504 can be found in steps S302-S304, and will not be repeated here.

[0111] S505. Obtain the feature differences between the weighted embedding features of each advertisement data and the object embedding features.

[0112] As mentioned above, the feature differences here can be represented by the inner product or distance between the weighted embedding features and the object embedding features, which will not be elaborated here.

[0113] S506. Sort at least one set of advertising data in ascending order of corresponding feature differences to obtain sorted advertising data.

[0114] Specifically, after sorting at least one set of advertising data in ascending order of corresponding feature differences, the advertising data ranked higher have smaller feature differences, meaning the distance between the weighted embedding feature and the object embedding feature is smaller and the inner product is larger; the advertising data ranked lower have larger feature differences, meaning the distance between the weighted embedding feature and the object embedding feature is larger and the inner product is smaller.

[0115] S507 identifies the top R ads in the sorted ad data as recall ad data for the target audience.

[0116] Where R is a positive integer. The recalled advertising data refers to the top R ranked advertising data, which are advertising data with small distances and large inner products between the weighted embedding features and the object embedding features. This means that the ranking position of the aforementioned feature differences meets the conditions, thereby determining the recalled advertising data corresponding to the target object. Since recommendation weights are introduced when determining the recalled advertising data based on the object features and advertising features, the determined recalled advertising data can better match the business scenario, thus improving the accuracy of the determined recalled advertising data.

[0117] In one embodiment, the method of determining recall ad data based on the inner product between weighted embedding features and object embedding features can be called Maximum Inner Product Retrieval (MIPS retrieval). This involves first calculating the inner product between the object embedding feature and each weighted embedding feature, then sorting by inner product to directly find the ad data at the top of the list. In real-world scenarios, online MIPS computation is challenging because ad data, even after database sharding, typically reaches millions of records. Converting inner product operations to matrix multiplication is often too complex, necessitating computational acceleration. For example, the LSH algorithm (a computational acceleration algorithm) can be used. This involves grouping features that are close in space (i.e., small distance, large inner product) into the same bucket, while features that are far apart are likely placed in different buckets. When determining recall ad data, it is only necessary to accurately compare and query ad data whose features are in the same bucket, significantly improving the efficiency of identifying recall ad data.

[0118] In one possible implementation, the recalled advertising data can also be pushed to the target device, causing the target device to output the recalled advertising data. This enables the recommendation of the determined recalled advertising data. The target device refers to the device corresponding to the target object. When the target device outputs the recalled advertising data, it can output information such as images, descriptions, and titles of the recalled advertising data; there are no limitations here.

[0119] In one possible implementation, in actual ad recall scenarios, to reduce online computation and improve ad recall efficiency, the ad embedding features of each ad data can be calculated and stored offline. However, since the object embedding features of an object need to consider some contextual information of the object, the object embedding features of the object need to be determined online. Optionally, to improve the efficiency of determining the object embedding features of objects online, the association information features for each type of object association information can be obtained in advance. Therefore, when determining the object embedding features of an object online, the corresponding object embedding features can be directly determined based on the association information features corresponding to the object association information. This is because when training the feature generation network, the association information features of the sample object association information can be generated based on the feature generation network, and then the corresponding object embedding features can be generated based on the association information features of the sample object association information. This allows for training with more samples, enabling the feature generation network to obtain the association information features corresponding to each type of object association information. Thus, by establishing the association information features for each type of object association information in advance, the process of obtaining association information features online is reduced, improving the efficiency of object embedding feature acquisition, thereby improving the efficiency of the entire ad recall process.

[0120] In one possible implementation, this application embodiment can further perform database segmentation processing on the various advertisements on the client side. That is, corresponding recall ad data is determined from advertisements in different ad databases, so that the recall ad data corresponding to each ad database is used as the final recall ad, ensuring that some ads in each ad database are recalled. In other words, the ad data in each ad database can be used as at least one ad data as described above, to determine the corresponding recall ad data from at least one ad data in each ad database, and the recall ad data corresponding to each ad database is pushed to the target device. Optionally, the number of recall ad data determined in each ad database can be flexibly set according to the actual scenario, and is not limited here.

[0121] In one embodiment, ads can be categorized into different ad libraries based on their impact, which can be metrics such as likes, shares, or orders, etc. For example, ads with an impact greater than or equal to a threshold can be grouped into one library, while those with an impact less than the threshold can be grouped into another. This prevents ads with low impact, due to recent release or insufficient exposure, from failing to be recalled in the next round of ad recall, thus avoiding a vicious cycle.

[0122] In one embodiment, ads can also be categorized based on their publication time. For example, ads published at or above a time threshold can be grouped into one ad group, and ads published below the time threshold into another. This increases the likelihood of newly published ads being recalled. Newly published ads lack interaction records with the target audience, so recommendations are often based on generalized features, resulting in poor scoring and a low probability of recommendation. Furthermore, because newly published ads lack exposure and consumption records, it's difficult to determine the appropriate weighting factors through recommendation evaluation information. This puts newly published ads at a disadvantage when competing with existing ads, severely compressing their chances of winning in the recall phase and advancing to the next stage, impacting final exposure and consumption, creating a vicious cycle. Therefore, by categorizing ads into different groups and using different methods to determine recall data for each group, the likelihood of newly published ads being recalled can be increased.

[0123] In one possible implementation, after determining the recalled advertising data, coarse sorting and fine sorting operations can be performed based on the determined recalled advertising data, thereby selecting fewer and more accurate advertising data to recommend to the target audience and improving the accuracy of advertising recommendations.

[0124] In some scenarios, the entire ad recall process is illustrated here with diagrams. For example, please see... Figure 7 , Figure 7 This is a flowchart illustrating an advertising recall method provided in an embodiment of this application. Figure 7As shown, firstly, a well-trained feature generation network (such as...) can be used. Figure 7 As shown in 701, the object embedding features (such as...) of the objects that need to be recalled for advertising are obtained. Figure 7 (As shown in 702). Specifically, a pre-trained feature generation network can be invoked to generate object embedding features based on object association information; and advertising embedding features of at least one advertising data generated by the pre-trained feature network can be obtained (e.g., ...). Figure 7 As shown in 703, specifically, it can call the trained feature generation network to generate corresponding ad embedding features based on the ad association information of each ad data; it can also obtain the recommendation weight of each ad data (such as...). Figure 7 As shown in 704), this recommendation weight can be determined based on the corresponding recommendation evaluation information. Therefore, based on the recommendation weight and the ad embedding features, a weighted embedding feature of the ad data is obtained (such as...). Figure 7 As shown in 705), recall advertising data (such as...) can be determined based on the object embedding features and weighted embedding features of the object. Figure 7 As shown in 706, specifically, the feature difference between the object embedding feature and each weighted embedding feature can be determined based on the object embedding feature and weighted embedding feature of the object. Thus, the recalled advertising data can be determined based on the feature difference. Due to the introduction of recommendation weight, the recalled advertising data can be more in line with the application scenario, which helps to improve the accuracy and flexibility of advertising recall.

[0125] By employing the embodiments of this application, the recommendation weight of each advertisement data can be determined based on the recommendation evaluation information of each advertisement data. Then, the advertisement embedding features of the advertisement data are weighted based on the recommendation weight of each advertisement data to obtain weighted embedding features. Finally, recall advertisement data for the target object is determined based on the weighted embedding features of each advertisement data and the object embedding features of the target object. This allows for the determination of different recommendation weights for advertisement data with different recommendation evaluation information. The recall advertisement data is then determined based on the weighted advertisement embedding features obtained through recommendation weights. Essentially, the adjustment of the probability of advertisement recall based on recommendation evaluation information, which is required for different business scenarios, is transformed into processing based on recommendation weights. This allows for flexible adjustment of the probability of advertisement data with different recommendation evaluation information being recalled according to different business scenarios, helping to improve the accuracy and flexibility of advertisement recall.

[0126] Please see Figure 8 , Figure 8 This is a schematic diagram of an advertising recall device provided in an embodiment of this application. Optionally, the advertising recall device can be installed in the aforementioned electronic device. Figure 8 As shown, the advertising recall device described in this embodiment may include:

[0127] The acquisition unit 801 is used to acquire at least one piece of advertising data and acquire the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data;

[0128] The acquisition unit 801 is also used to acquire recommendation evaluation information for each piece of advertising data, and to determine the recommendation weight for each piece of advertising data based on the recommendation evaluation information for each piece of advertising data; the recommendation evaluation information for any piece of advertising data includes information for evaluating the probability of recommending any piece of advertising data to an object;

[0129] Processing unit 802 is used to perform weighted processing on the advertising embedding features of each advertising data based on the recommendation weight of each advertising data, so as to obtain the weighted embedding features of each advertising data.

[0130] The processing unit 802 is further configured to obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from at least one advertising data according to the feature differences between the object embedding features and the weighted embedding features of each advertising data.

[0131] In one implementation, the processing unit 802 is specifically used for:

[0132] Obtain the weight transformation function for the recommendation evaluation information of each ad data point, and obtain the associated evaluation parameters for the recommendation evaluation information of each ad data point;

[0133] Based on the weight transformation function, the associated evaluation parameters corresponding to each ad data are processed to obtain the recommendation weight for each ad data.

[0134] In one implementation, the processing unit 802 is further configured to:

[0135] Obtain sample data pairs; sample data pairs contain object association information of sample objects and advertising association information of sample advertisements. Sample data pairs have sample tags, which are used to indicate whether the sample object has a target behavior or not in relation to the sample advertisement.

[0136] The feature generation network is invoked to generate sample object embedding features of sample objects based on the object association information of sample objects, and the feature generation network is invoked to generate sample ad embedding features based on the ad association information of sample ads.

[0137] The network parameters of the feature generation network are corrected based on the sample object embedding features, sample advertisement embedding features, and sample labels to obtain a trained feature generation network; the trained feature generation network is used to generate features for object and advertisement data.

[0138] In one implementation, the processing unit 802 is specifically used for:

[0139] Obtain the feature differences between sample object embedding features and sample advertisement embedding features;

[0140] Based on the feature differences and sample ad embedding features between sample object embedding features and sample ad embedding features, the feature generation bias of the feature generation network for sample object embedding features and sample ad embedding features is obtained;

[0141] By correcting the network parameters of the feature generation network based on the feature generation bias, a well-trained feature generation network can be obtained.

[0142] In one implementation, the processing unit 802 is specifically used for:

[0143] Obtain the ad association information for each ad data, and call the trained feature generation network to generate ad embedding features for each ad data based on the ad association information for each ad data;

[0144] Obtain object embedding features generated based on object association information of the target object, including:

[0145] Obtain the object association information of the target object, and call the trained feature generation network to generate object embedding features of the target object based on the object association information.

[0146] In one implementation, the processing unit 802 is specifically used for:

[0147] Determine the initial recommendation weight for each ad data point based on the recommendation evaluation information for each ad data point;

[0148] The initial recommendation weights for each ad data point are standardized to obtain the recommendation weights for each ad data point.

[0149] In one implementation, the processing unit 802 is specifically used for:

[0150] Obtain the feature differences between the weighted embedding features of each advertisement data and the object embedding features;

[0151] Sort at least one set of advertising data in ascending order of the differences in corresponding features to obtain sorted advertising data;

[0152] The top R ads in the sorted ad data are identified as the recall ads for the target audience; R is a positive integer.

[0153] Processing unit 802 is also used for:

[0154] The recall advertising data is pushed to the target device, causing the target device to output the recall advertising data.

[0155] In one implementation, the recommendation evaluation information for each piece of advertising data includes at least one of the following: advertising value information of the advertising data, advertising impact of the advertising data, or advertising category information of the advertising data.

[0156] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device described in this embodiment includes: a processor 901 and a memory 902. Optionally, the electronic device may also include a network interface or a power supply module, etc. The processor 901 and the memory 902 can exchange data.

[0157] The processor 901 described above can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0158] The aforementioned network interface may include input devices and / or output devices. For example, the input device may be a control panel, microphone, receiver, etc., and the output device may be a display screen, transmitter, etc., which will not be listed here.

[0159] The aforementioned memory 902 may include read-only memory and random access memory, and provides program instructions and data to the processor 901. A portion of memory 902 may also include non-volatile random access memory. The processor 901 executes program instructions when it calls them.

[0160] Obtain at least one piece of advertising data, and obtain the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data;

[0161] Obtain recommendation evaluation information for each ad data point, and determine the recommendation weight for each ad data point based on the recommendation evaluation information for each ad data point; the recommendation evaluation information for any ad data point includes information used to assess the probability of recommending any ad data point to an audience;

[0162] Based on the recommendation weight of each ad data, the ad embedding features of each ad data are weighted to obtain the weighted embedding features of each ad data.

[0163] Obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from at least one advertising data based on the feature differences between the object embedding features and the weighted embedding features of each advertising data.

[0164] In one implementation, processor 901 is specifically used for:

[0165] Obtain the weight transformation function for the recommendation evaluation information of each ad data point, and obtain the associated evaluation parameters for the recommendation evaluation information of each ad data point;

[0166] Based on the weight transformation function, the associated evaluation parameters corresponding to each ad data are processed to obtain the recommendation weight for each ad data.

[0167] In one implementation, processor 901 is further used for:

[0168] Obtain sample data pairs; sample data pairs contain object association information of sample objects and advertising association information of sample advertisements. Sample data pairs have sample tags, which are used to indicate whether the sample object has a target behavior or not in relation to the sample advertisement.

[0169] The feature generation network is invoked to generate sample object embedding features of sample objects based on the object association information of sample objects, and the feature generation network is invoked to generate sample ad embedding features based on the ad association information of sample ads.

[0170] The network parameters of the feature generation network are corrected based on the sample object embedding features, sample advertisement embedding features, and sample labels to obtain a trained feature generation network; the trained feature generation network is used to generate features for object and advertisement data.

[0171] In one implementation, processor 901 is specifically used for:

[0172] Obtain the feature differences between sample object embedding features and sample advertisement embedding features;

[0173] Based on the feature differences and sample ad embedding features between sample object embedding features and sample ad embedding features, the feature generation bias of the feature generation network for sample object embedding features and sample ad embedding features is obtained;

[0174] By correcting the network parameters of the feature generation network based on the feature generation bias, a well-trained feature generation network can be obtained.

[0175] In one implementation, processor 901 is specifically used for:

[0176] Obtain the ad association information for each ad data, and call the trained feature generation network to generate ad embedding features for each ad data based on the ad association information for each ad data;

[0177] Obtain object embedding features generated based on object association information of the target object, including:

[0178] Obtain the object association information of the target object, and call the trained feature generation network to generate object embedding features of the target object based on the object association information.

[0179] In one implementation, processor 901 is specifically used for:

[0180] Determine the initial recommendation weight for each ad data point based on the recommendation evaluation information for each ad data point;

[0181] The initial recommendation weights for each ad data point are standardized to obtain the recommendation weights for each ad data point.

[0182] In one implementation, processor 901 is specifically used for:

[0183] Obtain the feature differences between the weighted embedding features of each advertisement data and the object embedding features;

[0184] Sort at least one set of advertising data in ascending order of the differences in corresponding features to obtain sorted advertising data;

[0185] The top R ads in the sorted ad data are identified as the recall ads for the target audience; R is a positive integer.

[0186] Processor 901 is also used for:

[0187] The recall advertising data is pushed to the target device, causing the target device to output the recall advertising data.

[0188] In one implementation, the recommendation evaluation information for each piece of advertising data includes at least one of the following: advertising value information of the advertising data, advertising impact of the advertising data, or advertising category information of the advertising data.

[0189] Optionally, when the program instructions are executed by the processor, other steps of the method in the above embodiments can also be implemented, which will not be described in detail here.

[0190] This application also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the methods described above, such as the methods performed by the electronic device described above, which will not be elaborated here.

[0191] Optionally, the storage medium involved in this application, such as a computer-readable storage medium, may be non-volatile or volatile.

[0192] Optionally, the computer-readable storage medium may primarily include a stored program area and a stored data area. The stored program area may store the operating system, at least one application program required for a given function, etc.; the stored data area may store data created based on the use of blockchain nodes, etc. Here, the blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and cryptographic algorithms. A blockchain is essentially a decentralized database, a chain of data blocks linked using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain may include a blockchain underlying platform, a platform product service layer, and an application service layer, etc.

[0193] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0194] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0195] This application also provides a computer program product or computer program that includes computer instructions that, when executed by a processor, can implement some or all of the steps in the methods described above. For example, the computer instructions are stored in a computer-readable storage medium. The processor of a computer device (i.e., the aforementioned electronic device) reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps executed in the embodiments of the methods described above. For example, the computer device can be a terminal or a server.

[0196] The above provides a detailed description of an advertising recall method, apparatus, electronic device, and storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An advertising recall method, characterized in that, The method includes: Obtain at least one piece of advertising data, and obtain the advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data; Obtain recommendation evaluation information for each piece of advertising data, associated evaluation parameters of the recommendation evaluation information, and weight transformation function corresponding to the recommendation evaluation information; the recommendation evaluation information for any piece of advertising data includes information for evaluating the probability of recommending any piece of advertising data to an object. Based on the weight transformation function, the associated evaluation parameters corresponding to each piece of advertising data are processed by parameter calculation to obtain the initial recommendation weight for each piece of advertising data. If each piece of advertising data contains multiple recommendation evaluation information, the initial recommendation weights obtained based on the various recommendation evaluation information are fused to obtain the fused initial recommendation weights. The initial recommendation weights after fusion or the individual initial recommendation weights are standardized to obtain the recommendation weights for each piece of advertising data. Based on the recommendation weight of each ad data, the ad embedding features of each ad data are weighted to obtain the weighted embedding features of each ad data. Obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from the at least one advertising data according to the feature differences between the object embedding features and the weighted embedding features of each advertising data.

2. The method according to claim 1, characterized in that, The method further includes: Obtain sample data pairs; the sample data pairs contain object association information of sample objects and advertisement association information of sample advertisements, the sample data pairs have sample tags, and the sample tags are used to indicate whether the sample object has the target behavior or does not have the target behavior in response to the sample advertisement; The feature generation network is invoked to generate sample object embedding features of the sample object based on the object association information of the sample object, and the feature generation network is invoked to generate sample advertisement embedding features based on the advertisement association information of the sample advertisement; Based on the sample object embedding features, the sample advertisement embedding features, and the sample labels, the network parameters of the feature generation network are corrected to obtain a trained feature generation network; the trained feature generation network is used to generate features for object and advertisement data.

3. The method according to claim 2, characterized in that, The process of correcting the network parameters of the feature generation network based on the sample object embedding features, the sample advertisement embedding features, and the sample labels to obtain a trained feature generation network includes: Obtain the feature differences between the sample object embedding features and the sample advertisement embedding features; Based on the feature differences between the sample object embedding features and the sample advertisement embedding features and the sample label, the feature generation bias of the feature generation network for the sample object embedding features and the sample advertisement embedding features is obtained; The network parameters of the feature generation network are corrected based on the feature generation bias to obtain the trained feature generation network.

4. The method according to claim 2, characterized in that, The step of obtaining the advertising embedding features of each advertising data generated based on the advertising association information of each advertising data includes: Obtain the ad association information for each ad data, and call the trained feature generation network to generate ad embedding features for each ad data based on the ad association information for each ad data; The process of obtaining object embedding features generated based on object association information of the target object includes: Obtain the object association information of the target object, and call the trained feature generation network to generate the object embedding feature of the target object based on the object association information of the target object.

5. The method according to claim 1, characterized in that, The step of selecting recall advertising data for the target object from the at least one set of advertising data based on the feature differences between the object embedding features and the weighted embedding features of each set of advertising data includes: Obtain the feature differences between the weighted embedding features of each advertisement data and the object embedding features; The at least one set of advertising data is sorted in ascending order of corresponding feature differences to obtain sorted advertising data; The top R ads in the sorted ad data are identified as the recall ad data for the target audience; R is a positive integer. The method further includes: The recall advertising data is pushed to the target device, causing the target device to output the recall advertising data.

6. The method according to any one of claims 1-5, characterized in that, The recommendation evaluation information for each piece of advertising data includes at least one of the following: advertising value information of the advertising data, advertising impact of the advertising data, or advertising category information of the advertising data.

7. An advertising recall device, characterized in that, The device includes: An acquisition unit is configured to acquire at least one piece of advertising data and acquire advertising embedding features of each piece of advertising data generated based on the advertising association information of each piece of advertising data; The acquisition unit is further configured to acquire recommendation evaluation information for each piece of advertising data, associated evaluation parameters of the recommendation evaluation information, and a weight transformation function corresponding to the recommendation evaluation information; the recommendation evaluation information for any piece of advertising data includes information for evaluating the probability of recommending any piece of advertising data to an object; The processing unit is configured to perform parameter calculations on the correlation evaluation parameters corresponding to each piece of advertising data based on the weight transformation function to obtain an initial recommendation weight for each piece of advertising data; if each piece of advertising data has multiple recommendation evaluation information, the initial recommendation weights obtained based on the various recommendation evaluation information are fused to obtain a fused initial recommendation weight; the fused initial recommendation weight or the individual initial recommendation weight is standardized to obtain a recommendation weight for each piece of advertising data. The processing unit is configured to perform weighted processing on the advertising embedding features of each advertising data based on the recommendation weight of each advertising data, so as to obtain the weighted embedding features of each advertising data. The processing unit is further configured to obtain object embedding features generated based on object association information of the target object, and select recall advertising data for the target object from the at least one advertising data according to the feature differences between the object embedding features and the weighted embedding features of each advertising data.

8. An electronic device, characterized in that, The system includes a processor and a memory, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the method as described in any one of claims 1-6.

9. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the method as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1-6.

Images