Data promotion method, related equipment, readable medium and program product

Abstract

The invention discloses a data promotion method, related equipment, a readable medium and a program product, and the method is used for obtaining promotion content features of to-be-promoted data and behavior content features of a promotion object, and predicting promotion response features of the to-be-promoted data under a promotion response dimension based on the promotion content features. The promotion response feature comprises feature information of an object having response preference to the to-be-promoted data, and the response preference feature of the promotion object under the promotion response dimension is predicted based on the behavior content feature of the promotion object, and the response preference feature comprises feature information of promotion data of preference response of the promotion object. And finally, according to the promotion content feature and the promotion response feature of the to-be-promoted data and the behavior content feature and the response preference feature of the promotion object, determining whether to push the to-be-promoted data to the promotion object, so that the promotion data pushed to the promotion object can accurately meet the actual demand of the corresponding object.

Description

Technical Field

[0001] This application relates to the fields of deep learning technology and computer technology, and in particular to a data propagation method, related equipment, readable medium and program product. Background Technology

[0002] In the information age, internet users frequently receive all sorts of promotional data, but not every piece of data is viewed. In reality, promotional data that is sent but not viewed often doesn't meet the current needs of the target audience. Pushing such data to the target audience not only fails to effectively convert promotional efforts into benefits but can also easily evoke negative emotions, ultimately leading to user churn for the data promotion platform or provider. Therefore, in the current field of data promotion, accurately selecting promotional data that matches the current needs or interests of the target audience is a necessary research topic. Summary of the Invention

[0003] This application provides a data promotion method, related equipment, readable medium, and program product, which can accurately select promotion data that meets the current needs or interests of the target audience for promotion.

[0004] On the one hand, embodiments of this application provide a data promotion method, including:

[0005] The promotion content features of the data to be promoted are obtained, and the behavioral content features of the target object are obtained. The behavioral content features are obtained by content understanding of the behavioral record data of the target object.

[0006] Based on the promotion content features, a first feature prediction is performed under the promotion response dimension to obtain the promotion response features of the data to be promoted. The promotion response features include feature information of objects that have a response preference to the data to be promoted.

[0007] Based on the behavioral content characteristics of the target audience, a second feature prediction is performed under the promotion response dimension to obtain the response preference characteristics of the target audience. The response preference characteristics include the feature information of the promotion data of the target audience's preference response.

[0008] Based on the promotion content features and promotion response features of the data to be promoted, as well as the behavioral content features and response preference features of the target audience, a matching result is determined between the data to be promoted and the target audience. The matching result is used to indicate whether to push the data to be promoted to the target audience.

[0009] Furthermore, embodiments of this application provide a data promotion device, comprising:

[0010] The acquisition unit is used to acquire the promotion content features of the data to be promoted and to acquire the behavioral content features of the target object. The behavioral content features are obtained by content understanding of the behavioral record data of the target object.

[0011] The first feature prediction unit is used to perform first feature prediction under the promotion response dimension based on the promotion content features to obtain the promotion response features of the data to be promoted. The promotion response features include feature information of objects that have a response preference to the data to be promoted.

[0012] The second feature prediction unit is used to predict the second feature under the promotion response dimension based on the behavioral content features of the promotion object, so as to obtain the response preference features of the promotion object. The response preference features include the feature information of the promotion data of the promotion object's preference response.

[0013] The promotion prediction unit is used to determine the matching result between the data to be promoted and the target audience based on the promotion content characteristics and promotion response characteristics of the data to be promoted, as well as the behavioral content characteristics and response preference characteristics of the target audience. The matching result is used to indicate whether to push the data to be promoted to the target audience.

[0014] In one implementation, when the promotion prediction unit determines the matching result between the data to be promoted and the target audience based on the promotion content features and promotion response features of the data to be promoted, and the behavioral content features and response preference features of the target audience, it can specifically perform the following:

[0015] The promotion content features and the promotion response features are fused to obtain promotion fusion features, and the behavioral content features and the response preference features are fused to obtain object fusion features;

[0016] Based on the feature similarity between the promotion fusion feature and the object fusion feature, the matching result between the data to be promoted and the promotion object is determined.

[0017] In another embodiment, the behavior recording data includes recording data of the promoted object under at least two behaviors; when the acquisition unit acquires the behavioral content characteristics of the promoted object, it can specifically be used to perform:

[0018] Content understanding is performed on the behavioral records of the target audience under each behavior to obtain reference behavioral characteristics of the target audience under each behavior.

[0019] Obtain the feature similarity between reference behavioral features under different behaviors, and based on the obtained feature similarity, select target behavioral features from each reference behavioral feature corresponding to the promotion object. The feature similarity between each selected target behavioral feature is less than or equal to the similarity threshold.

[0020] Based on the selected target behavioral characteristics, the behavioral content characteristics of the promotion object are generated.

[0021] In another implementation, when acquiring the promotional content features of the data to be promoted, the acquisition unit may specifically perform the following:

[0022] Obtain the description text of the data to be promoted;

[0023] Based on the description text, target business data is selected from candidate business data, and the semantic similarity between the target business data and the description text is greater than a similarity threshold.

[0024] The category to which the target business data belongs is used as the reference category of the data to be promoted, and promotional content features of the data to be promoted are generated based on the reference category and the description text.

[0025] In another embodiment, the data promotion device further includes a model optimization unit, wherein the first feature prediction and the second feature prediction are executed by calling an optimized recall model, and the model optimization unit can be used to execute:

[0026] Obtain training samples, which include behavioral content features of sample objects, promotional content features of promotional samples, and reference matching results between the sample objects and the promotional samples;

[0027] The first feature prediction network in the recall model is used to predict the first feature of the promotion content feature of the promotion sample to obtain the promotion response feature of the promotion sample.

[0028] The second feature prediction network in the recall model is used to predict the behavioral content features of the sample object to obtain the response preference features of the sample object.

[0029] Based on the promotion content features and promotion response features of the promotion sample, and the behavioral content features and response preference features of the sample object, the predicted matching result between the promotion sample and the sample object is determined;

[0030] With the goal of reducing the difference between the predicted matching result and the reference matching result, the parameters of the first feature extraction network and the second feature extraction network are optimized to obtain the optimized recall model.

[0031] In another embodiment, the promotion sample is promotional data that has been pushed to the sample object, and the reference matching result is used to indicate whether the sample object responds to the promotion sample after receiving it; or, the promotion sample is promotional data contained in a promotional data sequence determined for the sample object, and the reference matching result is used to indicate the order of the promotion sample in the promotional data sequence of the sample object.

[0032] In another implementation, the promotional content features and the behavioral content features are extracted by calling an optimized content understanding model, and the model optimization unit can also be used to perform:

[0033] Retrieve the search history of the sample object;

[0034] First business data is filtered out from the business data associated with the search record, and positive samples are generated based on the search record and the first business data; the first business data is contained in the browsing record corresponding to the search record, and the category to which the first business data belongs matches the category searched by the search record;

[0035] The second business data is selected from the candidate business data, and a first negative sample is generated based on the search record and the second business data; the category to which the second business data belongs is different from the category searched by the search record, and the description information of the second business data contains the keywords in the search record;

[0036] The content understanding model is optimized using the set of positive and negative samples to obtain the optimized content understanding model, wherein the set of negative samples includes the first negative sample.

[0037] In another implementation, the model optimization unit can also be used to perform:

[0038] A third business data is selected from the business data associated with the search record, and a second negative sample is generated based on the third business data and the search record. The third business data is not included in the browsing records corresponding to the search record.

[0039] Randomly sample the candidate business data to obtain fourth business data, and generate a third negative sample based on the fourth business data and the search record;

[0040] Add the second negative sample and the third negative sample to the negative sample set.

[0041] In another aspect, embodiments of this application provide a computer device, including:

[0042] A memory, wherein a computer program is stored;

[0043] A processor for loading the computer program to implement the data promotion method as described in the first aspect.

[0044] In another aspect, embodiments of this application provide a readable medium storing a computer program adapted to be loaded by a processor and executed as the data promotion method of the first aspect.

[0045] In another aspect, embodiments of this application provide a program product comprising a computer program adapted to be loaded by a processor and executed as the data promotion method of the first aspect.

[0046] In determining whether to push data to be promoted to the target audience, this application embodiment utilizes not only the promotion content features of the data to be promoted and the promotion response features of the data to be promoted in the promotion response dimension, but also the behavioral content features of the target audience and the response preference features of the target audience in the promotion response dimension. Since the promotion response features contain the feature information of the target audience that has a response preference to the data to be promoted, and the response preference features contain the feature information of the promotion data that the target audience prefers to respond to, the matching process between the data to be promoted and the target audience can refer to the feature information of different feature levels and different feature dimensions, thereby ensuring that the promotion data pushed to the target audience can be as close as possible to the needs or interests of the target audience. Attached Figure Description

[0047] 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.

[0048] Figure 1 This is a schematic diagram of a data promotion system provided in an embodiment of this application;

[0049] Figure 2 This is a schematic flowchart illustrating a data promotion method provided in an embodiment of this application;

[0050] Figure 3 This is a schematic diagram illustrating the working principle of a dual-tower model provided in an embodiment of this application;

[0051] Figure 4 This is a schematic diagram of a data promotion process provided in an embodiment of this application;

[0052] Figure 5This is a schematic diagram illustrating the feature extraction principle of a content domain provided in an embodiment of this application;

[0053] Figure 6 This is a schematic diagram of the structure of a data promotion device provided in an embodiment of this application;

[0054] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0055] It should be noted in advance that, in order to enable those skilled in the art to better understand the technical solutions proposed in the embodiments of this application, the embodiments of this application will be described clearly and completely in conjunction with one or more accompanying drawings. Furthermore, the accompanying drawings shown in the embodiments of this application are merely illustrative examples; for instance, the execution order of each step in the drawings can be adaptively adjusted according to the actual application scenario.

[0056] Furthermore, in the embodiments of this application, the block diagrams, modules, and units shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. Each module or unit can be part of a larger module or unit that includes the functionality of that module or unit. That is, the terms "module" or "unit" mentioned in the embodiments of this application refer to a computer program or part of a computer program with a predetermined function, which can work together with other related parts to achieve a predetermined goal. It can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof, or implemented in different network and / or processor devices and / or microcontroller devices. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units.

[0057] Specifically, this application provides a data promotion scheme for data promotion scenarios. The principle of this scheme is as follows: when promoting data to be promoted, it is necessary to obtain the characteristics of the data to be promoted in both the advertising and content domains, as well as the characteristics of the target audience in both the advertising and content domains. By combining the characteristics of the data to be promoted with the characteristics of the target audience, it is determined whether the data to be promoted should be pushed to the target audience. The characteristics in the content domain primarily indicate the features exhibited by the content itself. For example, the characteristics of the target audience in the content domain (or behavioral content characteristics) can indicate the behavioral characteristics exhibited by the target audience's behavioral record data, and the characteristics of the data to be promoted in the content domain (or promotional content characteristics) can indicate the content characteristics exhibited by the content contained in the data to be promoted. The characteristics in the advertising domain primarily indicate features related to the promotion response dimension. For example, the characteristics of the target audience in the advertising domain can indicate the feature information contained in the promotional data for which the target audience has a response preference, and the characteristics of the data to be promoted in the advertising domain can indicate the characteristics possessed by the target audience who has a response preference for the data to be promoted.

[0058] By leveraging the characteristics of the target audience and the data to be promoted within the content domain, it's possible to assess whether the data aligns with the target audience's interests from a content perspective. Conversely, by considering the characteristics of the target audience and the data to be promoted within the advertising domain, it's possible to assess whether the target audience exhibits a preference for a particular type of promotional data from a promotional response perspective. In data recommendation scenarios, the characteristics of the content dimension and the characteristics of the promotional response dimension may not necessarily be correlated (e.g., an audience may prefer searching for type X information but may not be interested in promoting type X products). This application, by comprehensively considering the characteristics of both dimensions for data promotion, enriches the feature information during data recommendation, thereby ensuring that the promotional data received by the target audience is as closely aligned as possible with their characteristics, thus meeting their needs or interests.

[0059] In one implementation, the technical solution proposed in this application can be combined with a neural network model to enable efficient portability to various data recommendation scenarios. For example, an optimized content understanding model can be used to extract features in the content domain, and an optimized recall model can be used to extract features in the advertising domain. The optimization of the content understanding model is achieved based on data recall in the content domain, while the optimization of the recall model is achieved based on data recall combining the content and advertising domains. Specifically, when optimizing the content understanding model, the optimization goal can be to enable the model to accurately determine whether the promotional data matches the interests of the target audience based on the promotional data and the behavioral record data of the target audience. When optimizing the recall model, the optimization goal can be to enable the model to accurately determine whether the promotional data has been responded to by the target audience (e.g., clicks, purchase of related products, etc.) based on the features of the promotional data in the content domain and the features of the behavioral record data of the target audience in the content domain.

[0060] In one implementation, this application can be deployed in, for example, Figure 1 The data promotion system shown is used for application. For example... Figure 1 The system can, for example, include n data providers (marked with 101), a data promotion platform (marked with 102), and m terminal devices (marked with 103), where m and n are both positive integers. The objects using the terminal devices can be referred to as promotional targets. Data providers supply promotional data to the data promotion platform, which selects promotional data that matches the needs of the promotional targets from the various promotional data and pushes the promotional data to the terminal devices used by the promotional targets, enabling the promotional targets to browse or respond to the promotional data through their terminal devices.

[0061] It is understandable that a data promotion platform can run on computer devices such as terminal devices and / or servers. When the computer device includes a terminal device, this terminal device can be used by the target audience, and other applications and / or clients can run within it. Specific examples of terminal devices include smartphones, laptops, desktop computers, in-vehicle terminals, smart home appliances, and wearable devices. When the computer device includes a server, the server can be a physical server (or server cluster) or a cloud server (or server cluster) providing cloud services. The server can be used to provide the data promotion platform with the necessary data computing and data storage services, thereby improving data processing efficiency during the data promotion process through superior data processing performance.

[0062] Based on the above-mentioned data promotion principles, this application specifically proposes a data promotion method, the schematic flowchart of which can be found in [reference needed]. Figure 2 Furthermore, this method can continue to be executed by the aforementioned computer devices. For example... Figure 2 The method may include steps S201-S204:

[0063] S201. Obtain the promotion content features of the data to be promoted, and obtain the behavioral content features of the target object. The behavioral content features are obtained by understanding the content of the behavioral record data of the target object.

[0064] In a specific embodiment, the data to be promoted refers to data that has promotional demand and can be disseminated through the network. The promotional content characteristics of the data to be promoted refer to the features of the data in terms of content, which can be obtained through content understanding of the data to be promoted and / or its descriptive information. Optionally, the data to be promoted may include, but is not limited to, one or more types of data such as program products, text, images, audio and video. When the data to be promoted contains text, the promotional content characteristics may, for example, include the semantic features, grammatical features, and / or language type (such as the language of region A). In specific application scenarios, the data to be promoted can be placed on a data promotion platform so that the platform can select suitable targets for promotion.

[0065] Generally, data-driven promotion platforms can be used to provide one or more business services (such as social services, entertainment services, etc.), and the target audience includes users of at least the various business services offered on the platform. With the permission of the target audience, the platform can record their various behaviors (such as querying and purchasing) within the platform, thus obtaining behavioral data. Computer devices can then perform content understanding on this behavioral data to determine the behavioral content characteristics of the target audience. These characteristics typically indicate the target audience's interests and preferences, such as a preference for searching type X information or never browsing type Y information.

[0066] In one feasible implementation, when a computer device needs to acquire the promotional content features of data to be promoted, if the amount of data to be promoted is large, understanding the content of the data itself can easily lead to significant resource consumption, resulting in reduced data processing efficiency. To overcome this problem, the computer device can acquire the promotional content features based on the descriptive information of the data to be promoted. Optionally, the descriptive information can include image information, text information, audio information, and tag information, etc. Among them, image descriptive information can include promotional images and video keyframes of the data to be promoted; text descriptive information can include promotional text and titles of the data to be promoted; and tag descriptive information can include the category to which the data to be promoted belongs and the data provider.

[0067] For example, the category to which the data to be promoted belongs can be selected from the various categories to which the business data of the data promotion platform belongs, so that the promotional content characteristics obtained by the computer device can contain some characteristic information of the corresponding business data. Then, during the data promotion process, the computer device can, to a certain extent, refer to the target audience's preferences for the business data in the data promotion platform to determine whether the data to be promoted is suitable to be pushed to the target audience, thereby ensuring that the promotional data received by the target audience can highly match the corresponding target audience's needs or interests.

[0068] As an exemplary implementation, when it is necessary to determine the category of the data to be promoted, the computer device can first obtain the descriptive text of the data to be promoted, and then, based on the descriptive text, filter out target business data from candidate business data whose semantic similarity to the descriptive text is greater than a similarity threshold. The category to which the target business data belongs is then used as the category of the data to be promoted (for ease of distinction, this application refers to the category determined for the data to be promoted as the reference category). It is understood that, in this case, the promotional content features of the data to be promoted can be generated by the computer device based on the reference category and the descriptive text.

[0069] In one feasible implementation, when a computer device needs to obtain behavioral content features of a target audience, to ensure that these behavioral content features accurately describe the target audience, the behavioral record data obtained by the computer device may include record data generated by the target audience under at least two behaviors, so as to extract behavioral content features based on the record data under each behavior. Specifically, the record data under one behavior can correspond to the extraction of one behavioral feature of the target audience (hereinafter referred to as a reference behavioral feature for ease of distinction). After extracting the reference behavioral features under each behavior, the computer device can generate behavioral content features based on one or more of these reference behavioral features.

[0070] Specifically, when the behavioral record data includes records of the promoted object under at least two behaviors, if the computer device needs to obtain behavioral content features, it can first perform content understanding on the recorded data of the promoted object under each behavior to obtain reference behavioral features of the promoted object under each behavior. Then, it can obtain the feature similarity between the reference behavioral features under different behaviors. Based on the obtained feature similarity, it can select dissimilar target behavioral features from the various reference behavioral features, and then generate the behavioral features of the promoted object based on the selected target behavioral features. It can be understood that dissimilar target behavioral features refer to any two target behavioral features whose feature similarity is less than or equal to a similarity threshold.

[0071] In another feasible implementation, the computer device can also invoke the optimized content understanding model to extract promotional content features and behavioral content features. The content understanding model is used to perform content understanding on the corresponding data to obtain the feature information reflected in the content. Depending on the application scenario, the type of content understanding model varies. For example, in natural language processing scenarios, the content understanding model can be a Bidirectional Encoder Representation from Transformers (BERT), Recurrent Neural Networks (RNN), etc.; in image processing scenarios, the content understanding model can be a Convolutional Neural Networks (CNN), Generative Adversarial Networks (GANs), etc. This application does not impose any limitations on this.

[0072] It is worth mentioning that, in optimizing the content understanding model in this application embodiment, the training samples used may include positive samples and at least one negative sample, so that the optimized content understanding model can accurately determine whether the promotional data matches the interests of the target audience. Specifically, when constructing training samples, the computer device can first obtain the search records of the sample object on the data promotion platform (or other platforms), then filter out the first business data viewed by the sample object from the business data associated with the search records (i.e., the business data searched accordingly), and then generate positive samples based on the search records and the first business data.

[0073] In this process, the category of the first business data must match the category searched in the corresponding search record. For example, suppose the sample object has N search records, where the i-th search record was initiated using the keyword Z, and the category involved in keyword Z is "toys". If the search results based on keyword Z contain M business data entries, and the j-th business data entry was viewed by the sample object, then keyword Z can be used as the i-th search record, and the category "toys" involved in keyword Z can be used as the category searched in the i-th search record. When the category of the j-th business data entry matches (e.g., the same) the category searched in the i-th search record, the j-th business data entry can be used as the first business data entry. The j-th business data entry and the i-th search record (i.e., keyword Z) can form a positive sample.

[0074] By constructing a first negative sample to optimize the content understanding model, the optimized model can extract different feature representations for similar business data, thus accurately distinguishing similar business data. In other words, by constructing a first negative sample to optimize the content understanding model, the optimized model can have a stronger feature representation capability.

[0075] In one implementation, the negative samples in the training samples may include a first negative sample. The first negative sample can be generated by a computer device after filtering second business data from candidate business data, based on the search record and the second business data. The candidate business data includes business data contained in data promotion platforms and / or other platforms. The category to which the second business data belongs is different from the category corresponding to the search record, and the description information of the second business data contains keywords corresponding to the search record.

[0076] For example, suppose there are N search records for the sample object. The i-th search record is initiated by the keyword Z, and the category involved by keyword Z is "toy". The platform corresponding to the i-th search record contains P business data, and the description information (such as the title) of the k-th business data contains the keyword Z. If the category to which the k-th business data belongs is different from the category involved by keyword Z (such as the title of the k-th business data is "Characteristics of cats who love toys"), then the k-th business data can be used as the second business data. The k-th business data and the i-th search record (i.e., keyword Z) can form a first negative sample.

[0077] In another implementation, the negative samples in the training data can also include a second negative sample. The second negative sample can be generated by a computer device selecting third business data that has not been viewed by the sample object from the business data associated with the search records, and then using this third business data and the search records. In other words, the third business data is data outside the business data contained in the browsing records. For example, suppose the sample object has N search records, and the search results corresponding to the i-th search record contain M business data. If the h-th business data among the M business data has not been viewed by the sample object, then the h-th business data can be used as the third business data, and a second negative sample can be generated based on the h-th business data and the i-th search record.

[0078] In another implementation, the negative samples in the training data may also include a third negative sample. The third negative sample may consist of fourth business data and a search record. The fourth business data is obtained by randomly sampling candidate business data using a computer device. For example, a third negative sample may contain one search record and one piece of fourth business data. The explanation of candidate business data can be found in the aforementioned description of the first negative sample, and will not be repeated here.

[0079] S202. Based on the promotion content features, predict the first feature under the promotion response dimension to obtain the promotion response features of the data to be promoted. The promotion response features contain the feature information of the objects that have a response preference to the data to be promoted.

[0080] In a specific embodiment, the first feature prediction under the promotion response dimension mainly predicts which (or which) promotion object's response preferences the data to be promoted with the promotion content feature might match, and based on the promotion content feature, generates promotion response features that can express the feature information of these promotion objects for the data to be promoted.

[0081] For example, suppose that target audience A prefers to click on and browse promotional data of type X. It can be assumed that target audience A has a responsive preference for promotional data of type X. If the current data to be promoted also belongs to type X, then after performing the first feature prediction based on the promotional content features of the data to be promoted, the resulting promotional response features can contain the relevant feature information of target audience A. Thus, the target audience matched by the computer device based on the promotional response features can have a certain responsive preference for the data to be promoted, so that the data to be promoted can highly match the interests or needs of the target audience.

[0082] S203. Based on the behavioral content characteristics of the target audience, predict the second feature under the promotion response dimension to obtain the response preference features of the target audience. The response preference features contain the feature information of the promotion data of the target audience's preference response.

[0083] In a specific embodiment, the second feature prediction under the promotion response dimension mainly predicts which (or which) promotion data a promotion object with the behavioral content feature may have a response preference to, and generates a response preference feature that can express the feature information of these promotion data for the promotion object based on the content feature of the object.

[0084] For example, suppose that advertisers who have watched game highlight videos all prefer to click on and browse electronic product advertising data. If the behavioral content characteristics of advertiser B indicate that the advertiser has frequently searched for multiple game highlight videos, the response preference characteristics obtained by predicting the second feature based on the behavioral content characteristics can include relevant feature information of electronic product advertising data. This allows the advertising data matched by the computer device based on the response preference characteristics to arouse the interest of the advertiser or meet the needs of the advertiser.

[0085] It is worth mentioning that the computer device can invoke the optimized recall model to perform the first feature prediction in step S202 and the second feature prediction in step S203. Optionally, the recall model can be a dual-tower model, which refers to a model with two independent but interacting sub-networks. One sub-network processes user-side data (such as behavioral record data of the target audience), and the other sub-network processes product-side data (such as data to be promoted). For example, Figure 3 This illustrates the working principle of a dual-tower text processing model (Moka Massive Mixed Embedding, M3E). Figure 3 The subnet labeled 301 (hereinafter referred to as subnet 301) and the subnet labeled 302 (hereinafter referred to as subnet 302) are two independent data processing networks. Figure 3 If the first input data is user-side data and the second input data is product-side data, then sub-network 301 can store the feature expression u after extracting the feature expression u of the first input data. Similarly, sub-network 302 can also store the feature expression v after extracting the feature expression v of the second input data. Subsequently, when it is necessary to calculate the feature similarity between feature expression u and feature expression v, it can be done by querying feature expression u and feature expression v without re-extracting features.

[0086] It is easy to see that when the user base of a data promotion platform is large and the types of products are numerous (such as video websites or e-commerce platforms), it usually needs to respond quickly to a large number of recommendation requests. The dual-tower model enables the computer equipment to directly look up the corresponding features for calculation when it needs to predict the feature similarity between user-side features and product-side features, thereby improving data processing efficiency.

[0087] Assuming that the two sub-networks included in the recall model in this application embodiment are referred to as the first feature prediction network and the second feature prediction network, then an exemplary way for a computer device to optimize the recall model is as follows:

[0088] First, training samples are obtained. These training samples include behavioral content features of the sample objects, promotional content features of the promotional samples, and reference matching results between the sample objects and the promotional samples. Optionally, the promotional samples can be promotional data pushed to the sample objects, and the reference matching results can be used to indicate whether the sample objects respond to the promotional samples after receiving them. Alternatively, the promotional samples can be promotional data contained in a promotional data sequence determined for the sample objects, and the reference matching results are used to indicate the order of the promotional samples in the promotional data sequence of the sample objects; this is not limited here.

[0089] It should be noted that, in one implementation of this application, the sample object can be a user of the data promotion platform, and the behavioral content characteristics of the sample object can be extracted based on the behavioral record data of the sample object in the data promotion platform or other platforms, enabling the computer device to push diversified promotional data to the promotion object. The promotional sample is the promotional data existing in the data promotion platform, and the reference matching result is determined based on the sample object's response behavior to the promotional sample in the data promotion platform (such as clicking, browsing, purchasing, sharing, etc.). In this way, the optimized recall model can predict the promotional response characteristics and response preference characteristics closely related to the current data promotion platform, thereby ensuring that the promotional data pushed to the promotion object is diverse and meets the needs of the promotion object in the current data promotion platform.

[0090] After obtaining the training samples, the first feature prediction network in the recall model is used to predict the first feature of the promotional content features of the promotional samples, thus obtaining the promotional response features of the promotional samples. Then, the second feature prediction network in the recall model is used to predict the second feature of the behavioral content features of the sample objects, thus obtaining the response preference features of the sample objects. Both the first and second feature prediction networks can essentially be feedforward artificial neural networks, and as an example, both can be multilayer perceptrons (MLPs). MLPs aim to further process the input features (such as feature abstraction, feature crossing, and feature combination) to obtain deeper or more dimensional feature representations.

[0091] Finally, based on the promotion content features and promotion response features of the promotion samples, as well as the behavioral content features and response preference features of the sample objects, the predicted matching results between the promotion samples and the sample objects are determined. With the goal of reducing the difference between the predicted matching results and the reference matching results, the parameters of the first feature extraction network and the second feature extraction network in the recall model are optimized until a recall model that meets the expected optimization effect is obtained. This model is the optimized recall model.

[0092] S204. Based on the promotion content characteristics and promotion response characteristics of the data to be promoted, as well as the behavioral content characteristics and response preference characteristics of the target audience, determine the matching result between the data to be promoted and the target audience. The matching result is used to indicate whether to push the data to be promoted to the target audience.

[0093] In a specific embodiment, the computer device can perform feature fusion on promotional content features and promotional response features to obtain promotional fusion features, and perform feature fusion on behavioral content features and response preference features to obtain object fusion features. Then, based on the feature similarity between the promotional fusion features and the object fusion features, the matching result between the data to be promoted and the promotional object is determined. As an example, when the feature similarity is greater than or equal to a preset similarity threshold, it can be considered that the data to be promoted matches the promotional object, which means that the data to be promoted can be pushed to the promotional object.

[0094] Since response preference features are obtained by further processing behavioral content features, and promotion response features are obtained by further processing promotion content features, matching the data to be promoted with the promotion target based on these four features is equivalent to referring to feature information with different expressive capabilities and different expressive dimensions during the matching process. This can help computer devices better integrate features at different levels, avoid shallow features being completely covered or lost in deep networks, improve the expressive capability of the features obtained by computer devices, and enhance the computer devices' resistance to noise and outliers, thereby improving the robustness and generalization ability of computer devices.

[0095] In this embodiment, when determining whether to push data to be promoted to the target audience, the computer device utilizes not only the promotion content features of the data to be promoted and the promotion response features of the data to be promoted in the promotion response dimension, but also the behavioral content features of the target audience and the response preference features of the target audience in the promotion response dimension. Since the promotion response features contain the feature information of the target audience that has a response preference to the data to be promoted, and the response preference features contain the feature information of the promotion data that the target audience prefers to respond to, the matching process between the data to be promoted and the target audience can refer to the feature information of different feature levels and different feature dimensions, thereby ensuring that the promotion data pushed by the computer device to the target audience can be as close as possible to the needs or interests of the target audience.

[0096] To facilitate a clearer understanding of the embodiments of this application by those skilled in the art, the data promotion process of the embodiments of this application is briefly described below with reference to specific examples and accompanying drawings. In this example, the features of the content domain (i.e., the behavioral content features of the promotion target and the promotion content features of the data to be promoted) are extracted using an optimized content understanding model, the features of the advertising domain (i.e., the response preference features of the promotion target and the promotion response features of the data to be promoted) are extracted using an optimized recall model, and the matching of the data to be promoted and the promotion target is also achieved using an optimized recall model.

[0097] When optimizing the content understanding model, positive samples can be <search keywords, text titles of clicked search results>, and negative samples can include <search keywords, text titles of unclicked search results>, <search keywords, text titles of similar business data to the search results>, and <search keywords, text titles of randomly sampled business data>. As an example, similar business data can be understood as business data whose text titles contain the search keywords corresponding to the search results, but whose categories are not the same as those search keywords.

[0098] For example, suppose the data promotion platform is an e-commerce platform, and the business data is the product data on the platform. After a sample user enters the search keyword "mobile phone" on the platform, they get search result 1 and search result 2. Search result 1 has a product title of "Latest xxx Mobile Phone," and search result 2 has a product title of "Foldable Screen Mobile Phone xxx Next-Day Delivery." If the sample user only clicks on search result 1 to view details, a positive sample <mobile phone, latest xxx mobile phone> can be formed based on the search keyword "mobile phone" and the product title of search result 1, and a negative sample <mobile phone, foldable screen mobile phone xxx next-day delivery> can be formed based on the search keyword "mobile phone" and the product title of search result 2, "Foldable Screen Mobile Phone xxx Next-Day Delivery." Furthermore, suppose the e-commerce platform contains multiple products. Product i has a product title of "Fluffy Phone Case xx Popular Style." Since the product title contains the search keyword "mobile phone," but the product category should be phone cases, not the category involved in the search keyword (i.e., mobile phone), product i can be considered similar data to search result 1, thus forming a negative sample <mobile phone, fluffy phone case xx popular style>. In addition, products can be randomly sampled from a variety of goods, and then negative samples <mobile phone, mango-flavored xxx candy> can be constructed based on the search keywords and the product titles corresponding to the randomly sampled products (assuming it is "mango-flavored xxx candy").

[0099] By optimizing the content understanding model using the above four training samples (one positive sample and three negative samples), a model with accurate understanding and expression capabilities for text data can be obtained. Using this model to extract features from the data to be promoted and the behavioral record data, promotion content features and behavioral content features with strong feature expression capabilities can be obtained.

[0100] When optimizing the recall model, it can be based on promotional content features and behavioral content features. Positive samples can be <the behavioral content features of the sample object, plus the promotional content features of the promotional data that the sample object clicked>, while negative samples can be <the behavioral content features of the sample object, plus the promotional content features of the promotional data that the sample object closed>. For example, suppose a sample object, upon receiving a mobile promotion, viewed the promotion details, and upon receiving a game promotion, ignored the promotional information. Then, the computer device can construct positive samples based on the sample object's behavioral content features and the mobile promotion's promotional content features, and construct negative samples based on the sample object's behavioral content features and the game promotion's promotional content features.

[0101] Using the positive and negative samples constructed in the above manner to optimize the recall model allows the model optimization process to take into account the behavioral signals of users in response to the promotional data itself, thereby ensuring that the optimized recall model can recall promotional data (or recall promotional targets) from the perspective of improving promotional revenue.

[0102] Specifically, the data promotion process corresponding to the recall model when used for data promotion can be found in [reference needed]. Figure 4 The process for optimizing the recall model can be as follows:

[0103] The MLP included in the first sub-model of the recall model is used to further abstract the behavioral content features, and the resulting features are called response preference features. Then, the response preference features and behavioral content features are combined (or fused) to obtain the object fusion features. Similarly, the MLP included in the second sub-model of the recall model can be used to further abstract the promotional content features, and the resulting features are called promotional response features. Then, the promotional response features and promotional content features are combined (or fused) to obtain the promotional fusion features. Finally, by calculating the similarity between the object fusion features and the promotional fusion features, the sample type of the training sample is predicted to be either positive or negative. The model parameters in the recall model are optimized to reduce the difference between the predicted results and the true results until a recall model that meets the optimization expectations is obtained. As an example, the loss function used for optimizing the recall model can be the Binary Cross Entropy Loss (BCELoss) function shown in Equation 1.

[0104]

[0105] Among them, Loss bce Let y represent the loss value, sum represent the total number of samples in the current training batch, and y represent the total number of samples in the current training batch. iz is the true label of the i-th training sample (i.e., the true result, with positive samples labeled 1 and negative samples labeled 0). i σ(z) refers to the feature similarity between the object fusion feature and the generalized fusion feature of the i-th training sample. i Let σ(z) represent the probability (or prediction result) of the response to the i-th training sample. i It can be calculated as shown in Equation 2.

[0106]

[0107] It is worth mentioning that in other implementations, the training samples used to optimize the recall model can also be samples of the recommendation order. That is, the training samples can contain a sequence of behavioral content features and promotional content features of the sample objects, which includes the promotional content features of each promotional sample sorted by recommendation value. In this case, the Bayesian Personalized Ranking Loss (BPR Loss) can be used to calculate the loss value. The optimization objective is to reduce the loss value to the expected value to ensure the consistency between the predicted promotional sample sequence for the sample objects and the actual promotional sample sequence as much as possible. The actual promotional sample sequence refers to the promotional sample sequence corresponding to the promotional content feature sequence in the training samples. For example, the BPR Loss can be calculated as shown in Equation 3. In Equation 3, r i It is the recommendation value (or logits) of the i-th promoted sample, r j It is the recommended value of the j-th promotion sample.

[0108] Loss bpr =-∑ i,j log(σ(r i -r j Formula 3

[0109] After optimizing the models according to the above optimization methods to obtain the optimized content understanding model and the optimized recall model, the formal data promotion can be carried out according to the following process:

[0110] First, please see Figure 5 . Figure 5 This illustrates the feature extraction principle of an optimized content understanding model, such as... Figure 5After acquiring recent data on n behaviors of the target audience, the computer device can invoke the first sub-network of the optimized content understanding model to perform content understanding on the data for each behavior, obtaining the corresponding behavioral features. Then, based on the similarity between these behavioral features, feature filtering is performed on the n behaviors, retaining only those with low similarity (the retained behavioral features are...). Figure 5 The system uses various reference behavioral features (as defined in the model) and then generates behavioral content features based on these retained behavioral features, ensuring the diversity of feature information reflected in the behavioral content features. Furthermore, after acquiring the data to be promoted, the computer device can further obtain the descriptive information (such as category, title, and summary) of the data to be promoted. This descriptive information is then used to call the second sub-network in the optimized content understanding model to perform content understanding on the data to be promoted, thus obtaining the promotional content features. The category of the data to be promoted can be the category corresponding to training samples similar to the data to be promoted. After obtaining the promotional content features and behavioral content features, the computer device can call the optimized content understanding model to send the behavioral content features and promotional content features to the optimized recall model.

[0111] Next, please refer to the above. Figure 4 The process is shown. (For example...) Figure 4 As can be seen, the promotional content features of the data to be promoted and the behavioral content features of the target audience can be input into the first and second sub-models of the optimized recall model, respectively. The first sub-model can normalize the promotional content features, and the second sub-model can normalize the behavioral content features, to eliminate the influence of units on feature representation, making the interaction results between features (such as the calculated similarity) more reliable.

[0112] As an example, feature normalization can be L2-Norm, and the specific method can be found in Equation 4.

[0113]

[0114] In Equation 4, v represents the features before normalization, usually in vector form, and therefore can also be called eigenvectors, such as an n-dimensional eigenvector v = [v1, v2, ..., v2]. n ]. This is the normalized eigenvector, where |v|² represents the magnitude of the eigenvector v. The magnitude can be calculated according to Equation 5.

[0115]

[0116] In Equation 5, v i This represents the i-th eigenvalue in the eigenvector v, where i is less than or equal to n.

[0117] Then, for the first sub-model, after obtaining the normalized promotional content features, these features can be input into the MLP (equivalent to performing the first feature prediction under the promotional response dimension), and the features output by the MLP are used as promotional response features. These features are then normalized, and the normalized promotional response features are combined with the normalized promotional content features to obtain the promotional fusion features of the data to be promoted. Correspondingly, for the second sub-model, after obtaining the normalized behavioral content features, these features can be input into the MLP (equivalent to performing the second feature prediction under the promotional response dimension), and the features output by the MLP are used as response preference features. These features are then normalized, and the normalized behavioral content features are combined with the normalized response preference features to obtain the object fusion features of the promoted object. Finally, the optimized recall model is called to calculate the feature similarity between the object fusion features and the promotional fusion features, and the feature similarity is used to determine whether to push the data to be promoted to the promoted object.

[0118] By employing the above process for data promotion, on the one hand, the feature information reflected in advertising behavior (i.e., response behavior to promotional data) can be utilized during the feature matching process. This ensures a high degree of matching between the promotional targets matched by the recall model and the promotional data, thereby improving the effectiveness of promotional data delivery and reducing user negative feedback rates. In the long run, this can help improve the balanced transaction volume indicator (or ad load indicator) of the promotional scenario. On the other hand, when providing promotional data, the provider's interests may not align with those of the users. For example, a user's interest may be shopping, but shopping-related promotional data may not necessarily be delivered to e-commerce platforms or similar data promotion platforms. The embodiments of this application can accurately predict the interests or needs of various users on the current data promotion platform, thereby providing delivery suggestions and guidance to the promotional data provider. This optimizes both the conversion effect and the delivery ecosystem of the promotional data delivered by the provider.

[0119] Based on the aforementioned data promotion method, this application also provides a data promotion device. Specifically, please refer to... Figure 6 , Figure 6 The structure of the device is shown; this device can be mounted on a computer device and used to achieve the above. Figure 2 Some or all of the functions described in the method embodiments. For example... Figure 6 The device may include an acquisition unit 601, a first feature prediction unit 602, a second feature prediction unit 603, and a generalization prediction unit 604, wherein:

[0120] The acquisition unit 601 is used to acquire the promotion content features of the data to be promoted and to acquire the behavioral content features of the object to be promoted. The behavioral content features are obtained by content understanding of the behavioral record data of the object to be promoted.

[0121] The first feature prediction unit 602 is used to perform first feature prediction under the promotion response dimension based on the promotion content features to obtain the promotion response features of the data to be promoted. The promotion response features include feature information of objects that have a response preference to the data to be promoted.

[0122] The second feature prediction unit 603 is used to predict the second feature under the promotion response dimension based on the behavioral content features of the promotion object, so as to obtain the response preference features of the promotion object. The response preference features include the feature information of the promotion data of the promotion object's preference response.

[0123] The promotion prediction unit 604 is used to determine the matching result between the data to be promoted and the target audience based on the promotion content features and promotion response features of the data to be promoted, as well as the behavioral content features and response preference features of the target audience. The matching result is used to indicate whether to push the data to be promoted to the target audience.

[0124] In one implementation, when the promotion prediction unit 604 determines the matching result between the data to be promoted and the target audience based on the promotion content features and promotion response features of the data to be promoted, and the behavioral content features and response preference features of the target audience, it can specifically perform the following:

[0125] The promotion content features and the promotion response features are fused to obtain promotion fusion features, and the behavioral content features and the response preference features are fused to obtain object fusion features;

[0126] Based on the feature similarity between the promotion fusion feature and the object fusion feature, the matching result between the data to be promoted and the promotion object is determined.

[0127] In another embodiment, the behavior recording data includes recording data of the promoted object under at least two behaviors; when acquiring the behavioral content characteristics of the promoted object, the acquisition unit 601 can specifically be used to perform:

[0128] Content understanding is performed on the behavioral records of the target audience under each behavior to obtain reference behavioral characteristics of the target audience under each behavior.

[0129] Obtain the feature similarity between reference behavioral features under different behaviors, and based on the obtained feature similarity, select target behavioral features from each reference behavioral feature corresponding to the promotion object. The feature similarity between each selected target behavioral feature is less than or equal to the similarity threshold.

[0130] Based on the selected target behavioral characteristics, the behavioral content characteristics of the promotion object are generated.

[0131] In another embodiment, when acquiring the promotional content features of the data to be promoted, the acquisition unit 601 may specifically be used to perform:

[0132] Obtain the description text of the data to be promoted;

[0133] Based on the description text, target business data is selected from candidate business data, and the semantic similarity between the target business data and the description text is greater than a similarity threshold.

[0134] The category to which the target business data belongs is used as the reference category of the data to be promoted, and promotional content features of the data to be promoted are generated based on the reference category and the description text.

[0135] In another embodiment, the data promotion device further includes a model optimization unit 605, wherein the first feature prediction and the second feature prediction are executed by calling an optimized recall model, and the model optimization unit 605 can be used to execute:

[0136] Obtain training samples, which include behavioral content features of sample objects, promotional content features of promotional samples, and reference matching results between the sample objects and the promotional samples;

[0137] The first feature prediction network in the recall model is used to predict the first feature of the promotion content feature of the promotion sample to obtain the promotion response feature of the promotion sample.

[0138] The second feature prediction network in the recall model is used to predict the behavioral content features of the sample object to obtain the response preference features of the sample object.

[0139] Based on the promotion content features and promotion response features of the promotion sample, and the behavioral content features and response preference features of the sample object, the predicted matching result between the promotion sample and the sample object is determined;

[0140] With the goal of reducing the difference between the predicted matching result and the reference matching result, the parameters of the first feature extraction network and the second feature extraction network are optimized to obtain the optimized recall model.

[0141] In another embodiment, the promotion sample is promotional data that has been pushed to the sample object, and the reference matching result is used to indicate whether the sample object responds to the promotion sample after receiving it; or, the promotion sample is promotional data contained in a promotional data sequence determined for the sample object, and the reference matching result is used to indicate the order of the promotion sample in the promotional data sequence of the sample object.

[0142] In another embodiment, the promotional content features and the behavioral content features are extracted by calling an optimized content understanding model, and the model optimization unit 605 can also be used to perform:

[0143] Retrieve the search history of the sample object;

[0144] First business data is filtered out from the business data associated with the search record, and positive samples are generated based on the search record and the first business data; the first business data is contained in the browsing record corresponding to the search record, and the category to which the first business data belongs matches the category searched by the search record;

[0145] The second business data is selected from the candidate business data, and a first negative sample is generated based on the search record and the second business data; the category to which the second business data belongs is different from the category searched by the search record, and the description information of the second business data contains the keywords in the search record;

[0146] The content understanding model is optimized using the set of positive and negative samples to obtain the optimized content understanding model, wherein the set of negative samples includes the first negative sample.

[0147] In yet another implementation, the model optimization unit 605 can also be used to perform:

[0148] A third business data is selected from the business data associated with the search record, and a second negative sample is generated based on the third business data and the search record. The third business data is not included in the browsing records corresponding to the search record.

[0149] Randomly sample the candidate business data to obtain fourth business data, and generate a third negative sample based on the fourth business data and the search record;

[0150] Add the second negative sample and the third negative sample to the negative sample set.

[0151] It should be noted that, Figure 6The various units in the illustrated device can be individually or entirely combined into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. In other words, the above units are based on logical functional division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, Figure 6 The device shown may also include other units, and in practical applications, these functions may also be implemented with the assistance of other units, and may be implemented by multiple units working together.

[0152] According to another embodiment of this application, the following can be executed by running on a computing device including processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). Figure 2 The computer program (including program code) for each step involved in the corresponding method shown, to construct such... Figure 6 The apparatus shown, and the method for implementing the embodiments of this application. A computer program may be recorded on, for example, a computer-readable storage medium (hereinafter referred to as a readable medium), loaded into the apparatus described above through the computer-readable storage medium, and executed therein.

[0153] Based on the descriptions of the above method and apparatus embodiments, this application also provides a computer device. Specifically, please refer to... Figure 7 , Figure 7 This application provides a schematic diagram of the structure of a computer device, as shown in the embodiment of the present application. Figure 7 As shown, the computer device may include a processor 701, a memory 702, and a communication interface 703, and the processor 701, the memory 702, and the communication interface 703 may be connected by a bus or other means.

[0154] The processor 701 (or Central Processing Unit, CPU) is the computing and control core of a computer device. It can parse various instructions within the computer device and process various types of data. For example, the CPU can parse the data sharing task creation instructions sent by the data requester to the computer device and control the computer device to perform response operations; the CPU can also transfer various types of data between internal structures of the computer device, and so on.

[0155] Memory 702 is a storage device in a computer device used to store programs and data. It is understood that memory 702 here can include both the computer device's built-in memory and any extended memory supported by the computer device.

[0156] The communication interface 703 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 701; the communication interface 703 can also be used for the transmission and interaction of data within the computer device.

[0157] In one embodiment, processor 701 may load and execute one or more computer programs stored in memory 702 to achieve the aforementioned... Figure 2 The corresponding method steps in the method embodiments shown.

[0158] In a specific implementation, one or more computer programs in memory 702 are loaded and executed by processor 701:

[0159] The promotion content features of the data to be promoted are obtained, and the behavioral content features of the target object are obtained. The behavioral content features are obtained by content understanding of the behavioral record data of the target object.

[0160] Based on the promotion content features, a first feature prediction is performed under the promotion response dimension to obtain the promotion response features of the data to be promoted. The promotion response features include feature information of objects that have a response preference to the data to be promoted.

[0161] Based on the behavioral content characteristics of the target audience, a second feature prediction is performed under the promotion response dimension to obtain the response preference characteristics of the target audience. The response preference characteristics include the feature information of the promotion data of the target audience's preference response.

[0162] Based on the promotion content features and promotion response features of the data to be promoted, as well as the behavioral content features and response preference features of the target audience, a matching result is determined between the data to be promoted and the target audience. The matching result is used to indicate whether to push the data to be promoted to the target audience.

[0163] In one implementation, when the processor 701 determines the matching result between the data to be promoted and the target object based on the promotion content features and promotion response features of the data to be promoted, and the behavioral content features and response preference features of the target object, it can specifically load and execute:

[0164] The promotion content features and the promotion response features are fused to obtain promotion fusion features, and the behavioral content features and the response preference features are fused to obtain object fusion features;

[0165] Based on the feature similarity between the promotion fusion feature and the object fusion feature, the matching result between the data to be promoted and the promotion object is determined.

[0166] In another embodiment, the behavior recording data includes recording data of the promoted object under at least two behaviors; when the processor 701 obtains the behavioral content characteristics of the promoted object, it can specifically be used to load and execute:

[0167] Content understanding is performed on the behavioral records of the target audience under each behavior to obtain reference behavioral characteristics of the target audience under each behavior.

[0168] Obtain the feature similarity between reference behavioral features under different behaviors, and based on the obtained feature similarity, select target behavioral features from each reference behavioral feature corresponding to the promotion object. The feature similarity between each selected target behavioral feature is less than or equal to the similarity threshold.

[0169] Based on the selected target behavioral characteristics, the behavioral content characteristics of the promotion object are generated.

[0170] In another embodiment, when the processor 701 acquires the promotional content features of the data to be promoted, it can specifically be used to load and execute:

[0171] Obtain the description text of the data to be promoted;

[0172] Based on the description text, target business data is selected from candidate business data, and the semantic similarity between the target business data and the description text is greater than a similarity threshold.

[0173] The category to which the target business data belongs is used as the reference category of the data to be promoted, and promotional content features of the data to be promoted are generated based on the reference category and the description text.

[0174] In another implementation, the first feature prediction and the second feature prediction are executed by invoking an optimized recall model, which can be loaded and executed by processor 701:

[0175] Obtain training samples, which include behavioral content features of sample objects, promotional content features of promotional samples, and reference matching results between the sample objects and the promotional samples;

[0176] The first feature prediction network in the recall model is used to predict the first feature of the promotion content feature of the promotion sample to obtain the promotion response feature of the promotion sample.

[0177] The second feature prediction network in the recall model is used to predict the behavioral content features of the sample object to obtain the response preference features of the sample object.

[0178] Based on the promotion content features and promotion response features of the promotion sample, and the behavioral content features and response preference features of the sample object, the predicted matching result between the promotion sample and the sample object is determined;

[0179] With the goal of reducing the difference between the predicted matching result and the reference matching result, the parameters of the first feature extraction network and the second feature extraction network are optimized to obtain the optimized recall model.

[0180] In another embodiment, the promotion sample is promotional data that has been pushed to the sample object, and the reference matching result is used to indicate whether the sample object responds to the promotion sample after receiving it; or, the promotion sample is promotional data contained in a promotional data sequence determined for the sample object, and the reference matching result is used to indicate the order of the promotion sample in the promotional data sequence of the sample object.

[0181] In another embodiment, the promotional content features and the behavioral content features are extracted by calling an optimized content understanding model, and the processor 701 can also be used to load and execute them.

[0182] Retrieve the search history of the sample object;

[0183] First business data is filtered out from the business data associated with the search record, and positive samples are generated based on the search record and the first business data; the first business data is contained in the browsing record corresponding to the search record, and the category to which the first business data belongs matches the category searched by the search record;

[0184] The second business data is selected from the candidate business data, and a first negative sample is generated based on the search record and the second business data; the category to which the second business data belongs is different from the category searched by the search record, and the description information of the second business data contains the keywords in the search record;

[0185] The content understanding model is optimized using the set of positive and negative samples to obtain the optimized content understanding model, wherein the set of negative samples includes the first negative sample.

[0186] In yet another implementation, the processor 701 can also be used to load and execute:

[0187] A third business data is selected from the business data associated with the search record, and a second negative sample is generated based on the third business data and the search record. The third business data is not included in the browsing records corresponding to the search record.

[0188] Randomly sample the candidate business data to obtain fourth business data, and generate a third negative sample based on the fourth business data and the search record;

[0189] Add the second negative sample and the third negative sample to the negative sample set.

[0190] Furthermore, embodiments of this application also provide a readable medium (memory), which is a memory device in a computer device used to store programs and data. It is understood that the readable medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The readable medium provides storage space that stores the processing system of the computer device. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by the processor 701, which can be one or more computer programs (including program code). It should be noted that the readable medium here can be high-speed RAM memory or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0191] This application also provides a program product that includes computer instructions, which can be loaded by the processor of a computer device to implement the above-described method embodiments.

[0192] Based on the same inventive concept, the principles and beneficial effects of the data promotion device, computer equipment, readable medium and program product provided in the embodiments of this application are similar to those described in the foregoing corresponding method embodiments. Therefore, the corresponding method implementation principles and beneficial effects can be referred to, which will not be repeated here for the sake of brevity.

[0193] It should be further noted that the steps in the methods of this application embodiment can be adjusted, merged, and deleted according to actual needs, and the modules in the device of this application embodiment can be merged, divided, and deleted according to actual needs. Furthermore, 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. It should also be particularly emphasized that when the above embodiments of this application are applied to specific products or technologies, the data acquisition involved in each specific implementation of this application requires the permission or consent of the relevant parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0194] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art will understand that implementing all or part of the processes of the above embodiments and making equivalent changes in accordance with the claims of this application are still within the scope of this application.

Claims

1. A data promotion method, characterized in that, include: The promotion content features of the data to be promoted are obtained, and the behavioral content features of the target object are obtained. The behavioral content features are obtained by content understanding of the behavioral record data of the target object. Based on the promotion content features, a first feature prediction is performed under the promotion response dimension to obtain the promotion response features of the data to be promoted. The promotion response features include feature information of objects that have a response preference to the data to be promoted. Based on the behavioral content characteristics of the target audience, a second feature prediction is performed under the promotion response dimension to obtain the response preference characteristics of the target audience. The response preference characteristics include the feature information of the promotion data of the target audience's preference response. Based on the promotion content features and promotion response features of the data to be promoted, as well as the behavioral content features and response preference features of the target audience, a matching result is determined between the data to be promoted and the target audience. The matching result is used to indicate whether to push the data to be promoted to the target audience.

2. The method according to claim 1, characterized in that, The process of determining the matching result between the data to be promoted and the target audience based on the promotion content features and promotion response features of the data to be promoted, and the behavioral content features and response preference features of the target audience, includes: The promotion content features and the promotion response features are fused to obtain promotion fusion features, and the behavioral content features and the response preference features are fused to obtain object fusion features; Based on the feature similarity between the promotion fusion feature and the object fusion feature, the matching result between the data to be promoted and the promotion object is determined.

3. The method according to claim 1 or 2, characterized in that, The behavioral record data includes record data of the promoted object under at least two behaviors; obtaining the behavioral content characteristics of the promoted object includes: Content understanding is performed on the behavioral records of the target audience under each behavior to obtain reference behavioral characteristics of the target audience under each behavior. Obtain the feature similarity between reference behavioral features under different behaviors, and based on the obtained feature similarity, select target behavioral features from each reference behavioral feature corresponding to the promotion object. The feature similarity between each selected target behavioral feature is less than or equal to the similarity threshold. Based on the selected target behavioral characteristics, the behavioral content characteristics of the promotion object are generated.

4. The method according to claim 1 or 2, characterized in that, The promotional content features for obtaining the data to be promoted include: Obtain the description text of the data to be promoted; Based on the description text, target business data is selected from candidate business data, and the semantic similarity between the target business data and the description text is greater than a similarity threshold. The category to which the target business data belongs is used as the reference category of the data to be promoted, and promotional content features of the data to be promoted are generated based on the reference category and the description text.

5. The method according to claim 1, characterized in that, The first feature prediction and the second feature prediction are performed by calling the optimized recall model; the method further includes: Obtain training samples, which include behavioral content features of sample objects, promotional content features of promotional samples, and reference matching results between the sample objects and the promotional samples; The first feature prediction network in the recall model is used to predict the first feature of the promotion content feature of the promotion sample to obtain the promotion response feature of the promotion sample. The second feature prediction network in the recall model is used to predict the behavioral content features of the sample object to obtain the response preference features of the sample object. Based on the promotion content features and promotion response features of the promotion sample, and the behavioral content features and response preference features of the sample object, the predicted matching result between the promotion sample and the sample object is determined; With the goal of reducing the difference between the predicted matching result and the reference matching result, the parameters of the first feature extraction network and the second feature extraction network are optimized to obtain the optimized recall model.

6. The method according to claim 5, characterized in that, The promotion sample is promotional data that has been pushed to the sample object, and the reference matching result is used to indicate whether the sample object responds to the promotional sample after receiving it; or, the promotion sample is promotional data contained in the promotional data sequence determined for the sample object, and the reference matching result is used to indicate the order of the promotion sample in the promotional data sequence of the sample object.

7. The method according to claim 1, characterized in that, The promotional content features and the behavioral content features are extracted by calling an optimized content understanding model. The method further includes: Retrieve the search history of the sample object; First business data is filtered out from the business data associated with the search record, and positive samples are generated based on the search record and the first business data; the first business data is contained in the browsing record corresponding to the search record, and the category to which the first business data belongs matches the category searched by the search record; The second business data is selected from the candidate business data, and a first negative sample is generated based on the search record and the second business data; the category to which the second business data belongs is different from the category searched by the search record, and the description information of the second business data contains the keywords in the search record; The content understanding model is optimized using the set of positive and negative samples to obtain the optimized content understanding model, wherein the set of negative samples includes the first negative sample.

8. The method according to claim 7, characterized in that, The method further includes: A third business data is selected from the business data associated with the search record, and a second negative sample is generated based on the third business data and the search record. The third business data is not included in the browsing records corresponding to the search record. Randomly sample the candidate business data to obtain fourth business data, and generate a third negative sample based on the fourth business data and the search record; Add the second negative sample and the third negative sample to the negative sample set.

9. A data promotion device, characterized in that, include: The acquisition unit is used to acquire the promotion content features of the data to be promoted and to acquire the behavioral content features of the target object. The behavioral content features are obtained by content understanding of the behavioral record data of the target object. The first feature prediction unit is used to perform first feature prediction under the promotion response dimension based on the promotion content features to obtain the promotion response features of the data to be promoted. The promotion response features include feature information of objects that have a response preference to the data to be promoted. The second feature prediction unit is used to predict the second feature under the promotion response dimension based on the behavioral content features of the promotion object, so as to obtain the response preference features of the promotion object. The response preference features include the feature information of the promotion data of the promotion object's preference response. The promotion prediction unit is used to determine the matching result between the data to be promoted and the target audience based on the promotion content characteristics and promotion response characteristics of the data to be promoted, as well as the behavioral content characteristics and response preference characteristics of the target audience. The matching result is used to indicate whether to push the data to be promoted to the target audience.

10. A computer device, characterized in that, include: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the data promotion method according to any one of claims 1-8.

11. A readable medium, characterized in that, The readable medium stores a computer program adapted to be loaded by a processor and execute the data promotion method according to any one of claims 1-8.

12. A program product, characterized in that, The program product includes a computer program adapted to be loaded by a processor and execute the data promotion method according to any one of claims 1-8.

Images