Promotion information pushing method and device, computer device, storage medium and computer program product

Abstract

The application relates to a promotion information pushing method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: obtaining a promotion information pushing request, the promotion information pushing request carrying an object identifier to be pushed; obtaining object attribute features corresponding to the object identifier to be pushed; obtaining current promotion features corresponding to current promotion information, the current promotion features being obtained by fusing promotion attribute features and promotion interaction features corresponding to the current promotion information, the promotion attribute features being obtained by aggregating static attribute features and dynamic attribute features corresponding to the current promotion information, the aggregation weight of the aggregation being associated with the object identifier to be pushed; calculating a matching degree between the object attribute features and the current promotion features, and determining a pushing result of the current promotion information for the object identifier to be pushed based on the matching degree. The method can improve the accuracy of pushing and save pushing resources.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a method, apparatus, computer equipment, storage medium, and computer program product for promoting information push. Background Technology

[0002] With the development of internet technology, promotional information push technology has emerged. This technology can push promotional information to the target devices via the internet, such as advertisements, news, and social media content. Currently, when pushing promotional information, it is usually done by directly vectorizing the information and using the vectorized features to match the target audience to determine whether to push the information.

[0003] However, when using vectorized features to match the objects to be pushed, there is a problem that the vectorized features cannot accurately represent the promotional information, which leads to a decrease in the accuracy of the promotional information push and a waste of push resources. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of push notifications and reduce the waste of push notification resources, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for pushing promotional information. The method includes:

[0006] Obtain a promotional information push request, which carries the identifier of the target audience;

[0007] Obtain the object attribute characteristics corresponding to the identifier of the object to be pushed;

[0008] Obtain the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0009] Calculate the degree of matching between the object's attribute features and the current promotion features, and determine the push result of the current promotion information for the target object based on the degree of matching.

[0010] Secondly, this application also provides a promotional information push device. The device includes:

[0011] The request module is used to obtain promotional information push requests, which carry the identifier of the object to be pushed to.

[0012] The object feature acquisition module is used to acquire the object attribute features corresponding to the identifier of the object to be pushed;

[0013] The promotion feature acquisition module is used to acquire the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0014] The push determination module is used to calculate the degree of matching between the object's attribute features and the current promotion features, and to determine the push result of the current promotion information for the target object based on the degree of matching.

[0015] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0016] Obtain a promotional information push request, which carries the identifier of the target audience;

[0017] Obtain the object attribute characteristics corresponding to the identifier of the object to be pushed;

[0018] Obtain the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0019] Calculate the degree of matching between the object's attribute features and the current promotion features, and determine the push result of the current promotion information for the target object based on the degree of matching.

[0020] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0021] Obtain a promotional information push request, which carries the identifier of the target audience;

[0022] Obtain the object attribute characteristics corresponding to the identifier of the object to be pushed;

[0023] Obtain the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0024] Calculate the degree of matching between the object's attribute features and the current promotion features, and determine the push result of the current promotion information for the target object based on the degree of matching.

[0025] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0026] Obtain a promotional information push request, which carries the identifier of the target audience;

[0027] Obtain the object attribute characteristics corresponding to the identifier of the object to be pushed;

[0028] Obtain the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0029] Calculate the degree of matching between the object's attribute features and the current promotion features, and determine the push result of the current promotion information for the target object based on the degree of matching.

[0030] The aforementioned promotional information push method, apparatus, computer equipment, storage medium, and computer program product obtain the object attribute features corresponding to the identifier of the object to be pushed in the promotional information push request. Then, it obtains the current promotional features corresponding to the current promotional information. These current promotional features are obtained by aggregating static and dynamic attribute features according to the aggregation weight associated with the identifier of the object to be pushed, and further fusing them with the promotional interaction features corresponding to the current promotional information. This allows the obtained current promotional features to accurately represent the current promotional information. Furthermore, by aggregating the promotional attribute features according to the aggregation weight associated with the identifier of the object to be pushed, the aggregated promotional attribute features can better reflect the importance of the current promotional information to the object to be pushed, further improving the accuracy of the obtained current promotional features. Then, the matching degree between the current promotional features and the object attribute features is used to determine the push result of the current promotional information for the identifier of the object to be pushed. That is, using more accurate current promotional features to determine the push result can improve the accuracy of the push result, thereby improving the accuracy of the push and reducing the waste of push resources. Attached Figure Description

[0031] Figure 1 This is an application environment diagram of the information push method in one embodiment;

[0032] Figure 2 This is a flowchart illustrating a promotional information push method in one embodiment;

[0033] Figure 3 This is a schematic diagram of the framework for obtaining the current promotion feature in a specific embodiment;

[0034] Figure 4 This is a schematic diagram of the process for obtaining the current promotion feature in one embodiment;

[0035] Figure 5 This is a schematic diagram of the model architecture for a promotional information push model in a specific embodiment;

[0036] Figure 6 A schematic diagram of the network architecture of an object network in a specific embodiment;

[0037] Figure 7 This is a schematic diagram of the network architecture of a promotional information network in a specific embodiment.

[0038] Figure 8 This is a flowchart illustrating the process of training a promotional information push model in one embodiment;

[0039] Figure 9 This is a schematic diagram of the training architecture of a promotional information push model in a specific embodiment;

[0040] Figure 10This is a flowchart illustrating a promotional information push method in a specific embodiment;

[0041] Figure 11 This is a timing diagram illustrating the promotion of information push in a specific embodiment;

[0042] Figure 12 This is a schematic diagram of the model architecture for real-time push of promotional information in a specific embodiment;

[0043] Figure 13 This is a schematic diagram of the architecture for extracting promotional information features in a specific embodiment;

[0044] Figure 14 This is a structural block diagram of a promotional information push device in one embodiment;

[0045] Figure 15 This is an internal structural diagram of a computer device in one embodiment;

[0046] Figure 16 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] The promotional information push method provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network, and promotion vendor terminal 106 also communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed in the cloud or on another server. The promotion vendor's manager can send a request to server 104 via promotion vendor terminal 106 to push the current promotional information to the target audience. Server 104 receives the request and responds by obtaining the static and dynamic attributes of the current promotional information. Then, server 104 performs feature extraction on the static and dynamic attributes to obtain static attribute features and dynamic attribute features. Server 104 then aggregates the static and dynamic attribute features using an aggregation weight associated with the identifier of the target audience to obtain promotional attribute features. At this point, server 104 obtains the promotional interaction information corresponding to the current promotional information and extracts the features of the promotional interaction information to obtain promotional interaction features. Finally, server 104 fuses the promotion attribute features and promotion interaction features corresponding to the current promotion information to obtain the current promotion features, and saves the current promotion features corresponding to the current promotion information to the data storage system. Then, if server 104 receives a promotion information push request sent by requesting terminal 102, the promotion information push request carries the identifier of the object to be pushed. Server 104 responds to the promotion information push request, obtains the object attributes corresponding to the identifier of the object to be pushed, and extracts the features of the object attributes to obtain the object attribute features. At the same time, it obtains the current promotion features corresponding to the current promotion information from the data storage system. Then, server 104 calculates the matching degree between the object attribute features and the current promotion features, and determines the push result of the current promotion information for the identifier of the object to be pushed based on the matching degree. If the push result is to push, server 104 pushes the current promotion information to requesting terminal 102. Requesting terminal 102 receives the current promotion information and displays it. The requesting terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminals and servers can be connected directly or indirectly through wired or wireless communication, which is not limited herein.

[0049] In one embodiment, such as Figure 2 As shown, a method for pushing promotional information is provided, which can be applied to... Figure 1Taking a server as an example, it can be understood that this method can also be applied to a terminal, and also to a system that includes both a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0050] S202, Obtain promotional information push request, the promotional information push request carries the identifier of the target object to be pushed.

[0051] Among them, the promotion information push request refers to the request that needs to push promotion information. It can be a request sent by the request terminal corresponding to the object to be pushed, that is, the request triggered by the object to be pushed through the request terminal.

[0052] Promotional information refers to information that can be disseminated and promoted via the internet, such as through various websites, web pages, and internet applications. Promotional information is usually pre-set. For example, it can be advertisements, news, social media content, etc. Promotional information can be of different types, such as text, images, videos, audio, or animation. It can also be a combination of multiple media data, such as text and images, text and animation, or text and voice. Promotional information can be provided by the specific promotional company that needs to promote the information; different companies typically provide different promotional information.

[0053] The push notification object identifier is used to uniquely identify the corresponding push notification object, which is the object to which promotional information needs to be pushed. The push notification object identifier can be the name, number, or string of the push notification object. The push notification object can be a virtual object, such as a virtual robot or virtual character, or it can be a real object, such as a person, animal, or robot.

[0054] In this embodiment, when the target object is browsing information on the requesting terminal, it can request the push of promotional information, such as news, self-media content, or advertisements. At this time, the server can obtain the promotional information push request sent by the target object through the requesting terminal, which carries the target object identifier.

[0055] In some embodiments, an object can request a push notification for promotional information from an application on its terminal. For example, the object can log into an application that provides self-media content and browse self-media content within that application. Then, when the self-media content application detects that new self-media content needs to be pushed to the object, it sends a push notification request for promotional information to the application server.

[0056] In some embodiments, an object can also send a promotional information push request through a page. For example, when browsing a product advertisement page to select a product, the object can continuously browse various product advertisements on the page by swiping. When the object scrolls to the bottom of the product advertisement page, it can continue to swipe down. At this time, the product advertisement page responds to the downward swipe operation by sending a promotional information push request to the server of the product advertisement page.

[0057] S204, Obtain the object attribute characteristics corresponding to the identifier of the object to be pushed.

[0058] Among them, object attribute features refer to the characteristics of object attributes, which are used to characterize the attributes of the object to be pushed to the object identifier. For example, if the object is a person, the object attributes may include, but are not limited to, the object's gender, the object's age, the identifier of the promotional information already triggered on the object, and the identifier of the promotional information exposed to the object. When the object is a virtual robot, it may be the virtual robot's role setting information, etc.

[0059] In this embodiment, the server can obtain the object attributes corresponding to the identifier of the object to be pushed, and then extract the features of the object attributes to obtain object attribute features. The object attributes can be vectorized using a vectorization algorithm; for example, an embedding representation lookup table can be used to map the object attributes to low-dimensional continuous embedding representation vectors to obtain object attribute features. The vectorization algorithm can be a neural network algorithm, a hot coding algorithm, etc.

[0060] In some embodiments, the server can first embed the object attributes to obtain embedding representation features, and then perform deep feature extraction on the embedding representation features to obtain object attribute features. Here, deep features refer to features with high discriminative power and semantic information, containing globally relevant and semantically rich information. For example, deep feature extraction can be performed using a trained deep neural network to obtain object attribute features.

[0061] In some embodiments, the server can also obtain historical behavior data corresponding to the object identifier to be pushed, and then use the historical behavior data and object attributes to extract object attribute features, further improving the accuracy of the obtained object attribute features. Here, historical behavior data refers to data on the interaction behavior of the object identifier to be pushed with historically pushed promotional information.

[0062] S206, Obtain the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0063] Here, "current promotional information" refers to the promotional information for which it needs to be determined whether to push it to the target object. "Current promotional features" are the promotional features corresponding to the current promotional information. Promotional features characterize the characteristics of the corresponding promotional information. "Promotional interaction features" characterize the characteristics of the promotional interaction data corresponding to the current promotional information. Promotional interaction data refers to the data obtained after pushing the current promotional information, which includes data on interactions between different objects and the current promotional information. In some embodiments, if the current promotional information is new, the promotional interaction data can be empty. "Promotional attribute features" characterize the attributes corresponding to the current promotional information. "Static attribute features" characterize static attributes. "Dynamic attribute features" characterize dynamic attributes. Static attributes refer to attributes of the current promotional information that typically remain unchanged during long-term pushes. For example, if the current promotional information is an advertisement, the static attributes of the advertisement may include the industry to which the advertisement belongs, the manufacturer to which the advertisement belongs, etc. If the promotional information is news, the static attributes may be the industry to which the news belongs, the news publisher, etc. Dynamic attributes refer to attributes of the current promotional information that change over a long period of time. For example, if the promotional information is an advertisement, dynamic attributes could be the advertisement title, product price, discount level, and specific content. If the promotional information is news, dynamic attributes could be news tags and content. If the promotional information is self-media content, dynamic attributes could be the self-media content title, text content, and images. Aggregate weights are weights used to combine static and dynamic attribute features. The aggregate weight corresponding to the target object identifier is obtained by training on the historical promotional information data corresponding to that target object identifier; that is, the aggregate weight can be learned through a model. Different target object identifiers typically correspond to different historical promotional information data, and therefore, different target object identifiers usually have different aggregate weights.

[0064] In this embodiment, the server can directly retrieve the current promotion features corresponding to the current promotion information from the database. This database stores the promotion features corresponding to the promotion information to be pushed. These promotion features are obtained by pre-fusing the promotion attribute features and promotion interaction features corresponding to the promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static and dynamic attribute features are obtained by extracting features from the static and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0065] S208, calculate the degree of matching between the object attribute features and the current promotion features, and determine the push result of the current promotion information for the target object identifier based on the degree of matching.

[0066] The matching degree characterizes the level of interest of the target audience (identified by the target object) in the current promotional information. A higher matching degree indicates greater interest in the promotional information, increasing the likelihood of it being pushed to that audience. Conversely, a lower matching degree indicates less interest, decreasing the likelihood of it being pushed. The matching degree can be represented by a score between 0 and 100, or by a value between 0 and 1. The push result characterizes whether the promotional information is pushed to the target audience, including "no push required" and "push performed." "No push required" indicates that the promotional information does not need to be pushed to the target audience, while "push performed" indicates that the promotional information needs to be pushed to the target audience.

[0067] In this embodiment, the server can use a similarity algorithm to calculate the matching degree between object attribute features and current promotional features. The similarity algorithm can be a distance similarity algorithm, cosine similarity algorithm, etc. Then, based on the matching degree, the push result of the current promotional information for the target object is determined. Specifically, a pre-set matching degree threshold can be obtained, and the matching degree between the object attribute features and the current promotional features is compared with the matching degree threshold. If the matching degree is greater than the matching degree threshold, the current promotional information is pushed to the target object and sent to the requesting terminal. The requesting terminal receives and displays the pushed promotional information. If the matching degree is greater than the matching degree threshold, it indicates that the target object is not interested in the current promotional information. In this case, if the current promotional information is pushed to the requesting terminal, the target object will not interact with the current promotional information, resulting in wasted push resources. Therefore, the result is that the current promotional information does not need to be pushed.

[0068] The aforementioned promotional information push method obtains the object attribute features corresponding to the target object identifier in the promotional information push request. Then, it obtains the current promotional features corresponding to the current promotional information. These current promotional features are obtained by aggregating static and dynamic attribute features according to the aggregation weight associated with the target object identifier, and further fusing them with the promotional interaction features corresponding to the current promotional information. This ensures that the obtained current promotional features accurately represent the current promotional information. Furthermore, aggregating by the aggregation weight associated with the target object identifier allows the aggregated promotional attribute features to better reflect the importance of the current promotional information to the target object identifier, further improving the accuracy of the obtained current promotional features. Then, the matching degree between the current promotional features and the object attribute features is used to determine the push result of the current promotional information for the target object identifier. That is, using more accurate current promotional features to determine the push result improves the accuracy of the push result, thereby increasing the accuracy of the push and reducing the waste of push resources.

[0069] In some embodiments, S206, obtaining the current promotion feature corresponding to the current promotion information includes the following steps:

[0070] Obtain the static attributes corresponding to the current promotion information and extract the features of the static attributes to obtain static attribute features; obtain the dynamic attributes corresponding to the current promotion information and extract the features of the dynamic attributes to obtain dynamic attribute features; obtain the aggregation weight associated with the identifier of the object to be pushed, and perform aggregation operation on the static attribute features and dynamic attribute features based on the aggregation weight to obtain promotion attribute features; obtain the promotion interaction features corresponding to the current promotion information, and perform fusion operation on the promotion attribute features and promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

[0071] In this embodiment, the server can extract the current promotion features corresponding to the current promotion information in real time. Specifically, after receiving a promotion information push request, the server responds by retrieving the current promotion information that needs to be pushed from the database storing promotion information, and simultaneously obtains the static and dynamic attributes corresponding to the current promotion information. At this point, the server can perform embedding representation mapping on the static attributes to obtain static attribute features, and similarly, embedding representation mapping on the dynamic attributes to obtain dynamic attribute features. This embedding representation mapping transforms discrete or high-dimensional data into low-dimensional continuous feature vectors, thereby capturing the semantic relationships and inherent features between data and ensuring the accuracy of the obtained static and dynamic attribute features. Then, the server searches the aggregation weight database for the aggregation weight associated with the object identifier to be pushed. The aggregation weight database stores the object identifier and associated aggregation weight. Finally, the server can use the aggregation weight to weight and fuse the static and dynamic attribute features to obtain the promotion attribute features. At this point, the server can search the database for the promotion interaction data corresponding to the current promotion information, and then perform embedding representation mapping on the promotion interaction data to obtain the promotion interaction features. Finally, the server fuses the promotion attribute features and promotion interaction features to obtain the current promotion features corresponding to the current promotion information. The fusion process can be the sum of promotion attribute features and promotion interaction features, the product of promotion attribute features and promotion interaction features, or the fusion of promotion attribute features and promotion interaction features through an attention mechanism. In some embodiments, the fusion operation can also be the sum of promotion attribute features and promotion interaction features to obtain initial promotion fusion features, and then deep feature extraction can be performed on the initial promotion fusion features to obtain the current promotion features.

[0072] In a specific embodiment, such as Figure 3The diagram illustrates the framework for obtaining the current promotion features. The service acquires the static and dynamic attributes corresponding to the current promotion information, then performs feature extraction to obtain static and dynamic attribute features. Next, the server uses aggregation weights to aggregate the static and dynamic attribute features, resulting in promotion attribute features. Simultaneously, the server acquires promotion interaction data, extracts features from this data, and obtains promotion interaction features. Finally, the server fuses the promotion interaction features and promotion attribute features to obtain the current promotion features.

[0073] In the above embodiments, by acquiring the static and dynamic attributes corresponding to the current promotion information in real time, performing feature extraction and aggregation operations to obtain promotion attribute features, real-time dynamic attributes can be obtained, ensuring the accuracy of dynamic attributes and thus improving the accuracy of the obtained promotion attribute features. Finally, the promotion interaction features and promotion attribute features are fused to obtain the current promotion features, thereby improving the accuracy of the obtained current promotion features.

[0074] In some embodiments, the aggregation weight includes static aggregation weight and dynamic aggregation weight;

[0075] Aggregate static and dynamic attribute features based on aggregation weights to obtain generalized attribute features, including the following steps:

[0076] Static weighted features are obtained by weighting static attribute features based on static aggregation weights; dynamic weighted features are obtained by weighting dynamic attribute features based on dynamic aggregation weights; and the generalized attribute features are obtained by fusing the static and dynamic weighted features.

[0077] Static aggregation weights are used when aggregating static attribute features. Dynamic aggregation weights are used when aggregating dynamic attribute features. Static weighted features are features obtained by weighting static attribute features using static aggregation weights. Dynamic weighted features are features obtained by weighting dynamic attribute features using dynamic aggregation weights.

[0078] Specifically, the server can directly calculate the product of static aggregation weights and static attribute features to obtain static weighted features, and simultaneously calculate the product of dynamic aggregation weights and dynamic attribute features to obtain dynamic weighted features. Finally, the server can calculate the sum of static and dynamic weighted features to obtain generalized attribute features, or it can calculate the product of static and dynamic weighted features to obtain generalized attribute features. Alternatively, it can fuse static and dynamic weighted features through an attention mechanism to obtain generalized attribute features. For example, static weighted features can be used as query features, and dynamic weighted features can be used as key and value features for attention feature extraction to obtain generalized attribute features.

[0079] In some embodiments, the sum of the dynamic aggregation weight and the static aggregation weight is the target value, which is typically 1. That is, the server can obtain the static aggregation weight, and then calculate the difference between the target value and the static aggregation weight to obtain the dynamic aggregation weight.

[0080] In one specific embodiment, the server can use the following formula (1) to calculate the promotion attribute features.

[0081] new_center = λ * static_center + (1-λ) * dynamic_center formula (1)

[0082] Where `new_center` refers to the generalized attribute feature, `static_center` refers to the static attribute feature, `dynamic_center` refers to the dynamic attribute feature, `λ` refers to the static aggregation weight, and `1-λ` refers to the dynamic aggregation weight. That is, by calculating the product of the static attribute feature and the static aggregation weight, we obtain the static weighted feature; by calculating the product of the dynamic attribute feature and the dynamic aggregation weight, we obtain the dynamic weighted feature; and finally, by calculating the sum of the static and dynamic weighted features, we obtain the generalized attribute feature.

[0083] In the above embodiments, static attribute features are weighted using static aggregation weights, dynamic attribute features are weighted using dynamic aggregation weights, and finally feature fusion is performed to obtain promotional attribute features. That is, different attribute features are weighted using different aggregation weights and then fused to obtain promotional attribute features. This allows the obtained promotional attribute features to take into account the different effects of different attribute features on the obtained attribute features, thereby improving the accuracy of the obtained promotional attribute features.

[0084] In some embodiments, determining the promotion attribute features includes:

[0085] Obtain the promotion information push model associated with the identifier of the object to be pushed; use the network parameters of the aggregation operation network in the promotion information push model as the aggregation weight associated with the identifier of the object to be pushed; perform aggregation operation on static attribute features and dynamic attribute features based on the aggregation weight to obtain promotion attribute features.

[0086] The promotional information push model refers to a trained neural network model used to determine the push of promotional information. This model, associated with the identifier of the target object, includes an aggregation operation network associated with that identifier. The network parameters of the aggregation operation network are pre-trained parameters used to perform aggregation operations on static and dynamic attribute features.

[0087] In this embodiment, the server can obtain a promotional information push model associated with the identifier of the object to be pushed. This promotional information push model can be trained using training data corresponding to the identifier of the object to be pushed. The training data can include training promotional information and training tags. The training promotional information can be promotional information historically pushed to the object of the identifier of the object to be pushed, and the training tags can be determined by the interaction behavior of the object of the identifier of the object to be pushed with the historically pushed promotional information. The server obtains training static attributes, training dynamic attributes, and training interaction data corresponding to the training promotion information. These are input into the initial promotion information push model. The initial model extracts training static attribute features, training dynamic attribute features, and training interaction features. These features are then input into the initial aggregation operation network for aggregation operations to obtain training attribute features. These training attribute features are then fused with the training interaction features to obtain training promotion features. Finally, the training matching degree between the training promotion features and the object attribute features corresponding to the target object identifier is calculated, and the loss information between the training matching degree and the training label is calculated. This loss information is used to update the model parameters in the initial promotion information push model, including the initial aggregation parameters in the initial aggregation operation network, resulting in an updated promotion information push model. This updated model is then iteratively trained again. Upon reaching the training completion condition, a promotion information push model associated with the target object identifier is obtained. At this point, the server directly uses the network parameters of the aggregation operation network in the promotion information push model as the aggregation weights associated with the target object identifier. Then, aggregate weights are used to perform aggregation operations on static and dynamic attribute features to obtain generalized attribute features.

[0088] In some embodiments, the server may also directly obtain the network parameters of the aggregation operation network, and use the static and dynamic attribute features of the network parameters of the aggregation operation network to perform aggregation operations to obtain the promotion attribute features.

[0089] In the above embodiments, the static and dynamic attribute features are aggregated using the network parameters of the aggregation operation network in the promotion information push model to obtain the promotion attribute features. Since the network parameters of the aggregation operation network are trained using historical data, the accuracy of the obtained promotion attribute features is guaranteed.

[0090] In some embodiments, determining the current promotion feature includes:

[0091] The promotion attribute features are linearly mapped through a linear mapping network to obtain linear mapping features. The feature dimensions of the linear mapping features are matched with those of the promotion interaction features. The linear mapping features and the promotion interaction features are then fused to obtain the current promotion features corresponding to the current promotion information.

[0092] Here, linear mapping features refer to the features obtained after linear mapping through a linear mapping network. This linear mapping network can be a pre-trained neural network used to match the feature dimensions of the linear mapping features with those of the generalized interaction features. Feature dimension matching is used to indicate that the feature dimensions of the linear mapping features and the generalized interaction features are the same.

[0093] In this embodiment, since promotional attributes typically include different attributes, such as promotional images and promotional titles, the feature dimensions of promotional attribute features may not match the feature dimensions of promotional interaction features. In this case, the server can input the promotional attribute features into a linear mapping network for linear mapping. This linear mapping network can be a single-layer neural network with only linear weights and no activation function, ensuring that the resulting linearly mapped features match the feature dimensions of the promotional interaction features, such as promotional attribute features. and promotional interaction features The feature dimensions are different. In this case, the feature dimension of the promotion attribute feature is 1*n, and the feature dimension of the promotion interaction feature is 1*m. Furthermore, the dimension of the linear weights in the neural network layer is m*n. Therefore, the product of the promotion attribute feature and the promotion interaction feature is calculated to obtain the linear mapping feature, which has a feature dimension of 1*n. Finally, the server can element-wise add the linear mapping features with the same feature dimension to the promotion interaction feature, that is, add the dimension values ​​of the matching feature dimensions to obtain the current promotion feature corresponding to the current promotion information.

[0094] In the above embodiments, by linearly mapping the promotion attribute features to match the feature dimensions of the promotion interaction features, the accuracy of subsequent fusion can be guaranteed, thereby ensuring the accuracy of the obtained current promotion features.

[0095] In some embodiments, dynamic attributes include content attributes and promotion vendor status attributes corresponding to the current promotion information;

[0096] The determination of static and dynamic attribute features includes:

[0097] Static attributes are embedded and represented to obtain static attribute features; content attributes are embedded and represented to obtain content attribute features; and the promoter status attribute is embedded and represented to obtain promoter status features; the content attribute features and promoter status features are concatenated to obtain dynamic attribute features.

[0098] Content attributes refer to the attributes used to characterize the content of the current promotional information. For example, if the current promotional information is an advertisement, the content attributes could be the advertisement image, advertisement title, etc. The promoter status attribute is used to characterize the status of the promoter corresponding to the current promotional information. This promoter status attribute can include, but is not limited to, the promoter's business status, and the promoter's restrictions on the promotional information. For example, if the promoter is a product manufacturer, the promoter's business status could be the product manufacturer's business status, such as being in a hot-selling state or a clearance state, etc. The promoter's restrictions on the promotional information could be the product manufacturer's budget allocation for product advertising, etc.

[0099] In this embodiment, the server uses a vectorization algorithm to embed and represent static attributes, obtaining static attribute features. This vectorization algorithm can be a text vectorization algorithm, such as one-hot encoding or the bag-of-words model, or an image vectorization algorithm, such as a neural network algorithm. Simultaneously, the server can also embed and represent content attributes to obtain content attribute features, and embed the promoter status attribute to obtain promoter status features. Finally, the server concatenates the content attribute features and the promoter status features to obtain dynamic attribute features.

[0100] In the above embodiments, dynamic attribute features are obtained by extracting content attribute features and promotion vendor status features and concatenating the features. This dynamic attribute feature is not only related to the promotion information itself, but also to the promotion vendor status of the promotion information, thereby improving the accuracy of the obtained dynamic attributes.

[0101] In some embodiments, S204, obtaining the object attribute features corresponding to the identifier of the object to be pushed includes the following steps:

[0102] Obtain the numerical and categorical attributes corresponding to the identifier of the object to be pushed; perform embedding representation mapping on the categorical attributes through a preset embedding lookup table to obtain categorical features; and concatenate the categorical features with the numerical attributes to obtain the object attribute features.

[0103] Numerical attributes refer to attributes of numerical value, such as price attributes and discount attributes. Categorical attributes refer to attributes of non-numeric value, which can be character-based attributes, such as text, product identifiers, object identifiers, etc. The preset embedding lookup table is the core component used to map categorical attributes to continuous vector representations. The preset embedding lookup table is a trained weight matrix, and the vectorized features of the corresponding attribute can be directly obtained through indexing.

[0104] In this embodiment, the object attributes obtained by the server may include numeric attributes and categorical attributes. In this case, the categorical attributes can be embedded and mapped using a preset embedding lookup table to obtain categorical features. Then, the categorical features are concatenated with the numeric attributes to obtain the object attribute features.

[0105] In some embodiments, the server can obtain static and dynamic attributes, which may include numeric and categorical attributes. In this case, the categorical attributes can be embedded and mapped using a predefined embedding lookup table to obtain categorical features. Then, the categorical features are concatenated with the numeric attributes to obtain object attribute features.

[0106] In some embodiments, the server may also concatenate categorical features with numerical attributes to obtain initial concatenated features, and then perform deep feature extraction on the initial concatenated features to obtain object attribute features.

[0107] In the above embodiments, categorical attributes are embedded and represented using a preset embedding lookup table to obtain categorical features, and categorical features are concatenated with numerical attributes to obtain object attribute features. In other words, categorical features are embedded and represented without needing to be embedded and represented by numerical attributes, thereby improving the efficiency of obtaining object attribute features while ensuring the accuracy of the obtained object attribute features.

[0108] In some embodiments, such as Figure 4 As shown, S206, which involves obtaining the current promotional features corresponding to the current promotional information, includes the following steps:

[0109] S402, obtain the static attribute features corresponding to the current promotion information, and obtain the current dynamic attribute features and historical dynamic attribute features corresponding to the current promotion information.

[0110] Here, "current dynamic attribute feature" refers to the dynamic attribute feature currently in use. "Historical dynamic attribute feature" refers to the dynamic attribute feature used in the past, which can be the previous dynamic attribute feature corresponding to the current dynamic attribute feature.

[0111] In this embodiment, the server can pre-save static and dynamic attributes in the format of a dynamic memory network. This dynamic memory network is mainly implemented through multi-layer memory modules, each containing a memory matrix and a controller. The memory matrix stores static and dynamic attributes, while the controller is responsible for reading and writing the memory matrix. The server can obtain the static attributes corresponding to the current promotional information, and then extract the features of the static attributes to obtain static attribute features. Then, since the dynamic attributes of the current promotional information may have changed dynamically, the server can obtain the current dynamic attributes. If the current dynamic attributes are updated based on historical dynamic attributes, the historical dynamic attributes are also obtained. At this time, the server can perform an embedding representation mapping between the current dynamic attributes and the historical dynamic attributes to obtain the current dynamic attribute features and the historical dynamic attribute features.

[0112] In some embodiments, the server may also store dynamic and static attributes in key-value pair format, that is, in the format of attribute name and corresponding attribute value.

[0113] In one specific embodiment, the dynamic and static attributes of the advertisements stored in the dynamic memory network are as follows:

[0114] memory = {

[0115] "industry_1":{"static_center":tensor(256d), "dynamic_center":tensor(256d)},

[0116] "industry_2":{"static_center":tensor(256d), "dynamic_center":tensor(256d)},

[0117] ...}

[0118] Here, "memory" refers to a dynamic memory network. `industry_1` represents the industry to which the current advertisement belongs. `industry_2` represents the industry to which the current advertisement belongs. `"static_center": tensor(256d)` represents static attributes and their corresponding values, stored as key-value pairs. `"dynamic_center": tensor(256d)` also represents static attributes and their corresponding values, stored as key-value pairs. The dynamic and static attributes of advertisements in this network can be updated according to industry attributes, and advertisements within the same industry share memory units, addressing the data sparsity problem for new advertisements.

[0119] S404, calculate the offset between the current dynamic attribute features and the historical dynamic attribute features, and determine the target dynamic attribute features corresponding to the current promotion information based on the offset and the current dynamic attribute features.

[0120] The offset magnitude characterizes the degree of change of the current dynamic attribute relative to the historical dynamic attribute. A larger offset magnitude indicates a greater degree of change in the current dynamic attribute compared to its historical counterpart, while a smaller offset magnitude indicates a smaller degree of change. The target dynamic attribute feature refers to the dynamic attribute feature determined based on the magnitude of the offset magnitude.

[0121] In this embodiment, the server can calculate the offset between the current dynamic attribute feature and historical dynamic attribute features. Then, it uses the offset magnitude and a preset magnitude threshold to determine whether to use the current dynamic attribute feature as the target dynamic attribute feature. The preset magnitude threshold is a pre-set offset magnitude threshold to keep the dynamic attribute stable; this stability can be achieved by the offset magnitude being less than the preset magnitude threshold. That is, if the offset magnitude is less than the preset magnitude threshold, it indicates that the change in the current dynamic attribute is within a preset range, and in this case, the current dynamic attribute feature can be directly used as the target dynamic attribute feature.

[0122] S406, obtain the aggregation weight associated with the identifier of the object to be pushed, and perform aggregation operation on the static attribute features and the target dynamic attribute features based on the aggregation weight to obtain the target promotion attribute features;

[0123] S408: Obtain the promotion interaction features corresponding to the current promotion information, and perform a fusion operation on the target promotion attribute features and promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

[0124] Among them, the target promotion attribute features refer to the promotion attribute features obtained by aggregating the target dynamic attribute features after they have passed the offset amplitude detection.

[0125] In this embodiment, when performing aggregation operations, the server can use the aggregation weight associated with the object identifier to be pushed to aggregate the static attribute features and the target dynamic attribute features after passing the offset amplitude detection. This can be achieved by weighting the static attribute features with static aggregation weights, then weighting the target dynamic attribute features with dynamic aggregation weights, and finally calculating the sum to obtain the target promotion attribute features. At this point, the server can perform a fusion operation between the target promotion attribute features and promotion interaction features to obtain the current promotion feature corresponding to the current promotion information. This can be achieved by calculating the sum of the target promotion attribute features and promotion interaction features to obtain the total promotion feature, and then performing deep feature extraction on the total promotion feature to obtain the current promotion feature.

[0126] In the above embodiments, by calculating the offset between the current dynamic attribute features and the historical dynamic attribute features, the target dynamic attribute features corresponding to the current promotion information are determined based on the offset and the current dynamic attribute features. This ensures the stability of the obtained target dynamic attribute features and avoids the problem that the accuracy of the obtained current promotion features is reduced due to large changes in dynamic attribute features, thereby improving the accuracy of promotion information push.

[0127] In some embodiments, the offset magnitude includes the feature distance;

[0128] S304, Calculate the offset between the current dynamic attribute features and the historical dynamic attribute features, and determine the target dynamic attribute features corresponding to the current promotion information based on the offset and the current dynamic attribute features, including the following steps:

[0129] Calculate the feature distance between the current dynamic attribute feature and the historical dynamic attribute feature; if the feature distance is less than a preset distance threshold, then use the current dynamic attribute feature as the target dynamic attribute feature corresponding to the current promotion information; if the feature distance is greater than or equal to the preset distance threshold, then obtain the updated dynamic attribute feature corresponding to the current dynamic attribute feature, use the updated dynamic attribute feature as the target dynamic attribute feature corresponding to the current promotion information, and the feature distance between the updated dynamic attribute feature and the historical dynamic attribute feature is less than the preset distance threshold.

[0130] The feature distance characterizes the difference between the current dynamic attribute features and historical dynamic attribute features. The preset distance threshold is a pre-set threshold for the feature distance, used to ensure the stability of the obtained target dynamic attribute features and prevent excessive changes in the target dynamic attribute features relative to historical dynamic attribute features. The updated dynamic attribute features are obtained by extracting features from the updated dynamic attribute. The updated dynamic attribute features are obtained by updating the current dynamic attribute. The feature distance between the updated dynamic attribute features and historical dynamic attribute features is less than the preset distance threshold.

[0131] Specifically, the server uses a feature distance algorithm to calculate the feature distance between the current dynamic attribute feature and historical dynamic attribute features. This feature distance algorithm can be Euclidean distance, Manhattan distance, etc. The server then compares the feature distance with a preset distance threshold and determines the target dynamic attribute feature based on the comparison result. If the feature distance is less than the preset distance threshold, the current dynamic attribute feature is used as the target dynamic attribute feature corresponding to the current promotional information. If the feature distance is greater than or equal to the preset distance threshold, the updated dynamic attribute feature corresponding to the current dynamic attribute feature is obtained. The server then uses the updated dynamic attribute feature as the current dynamic attribute feature and iteratively executes the step of calculating the feature distance between the current dynamic attribute feature and historical dynamic attribute features until the feature distance between the obtained updated dynamic attribute feature and historical dynamic attribute features is less than the preset distance threshold. If the updated dynamic attribute feature is less than the preset distance threshold, it is used as the target dynamic attribute feature corresponding to the current promotional information. For example, when the dynamic attribute feature is the product price attribute of a product advertisement, the feature distance between the current product price feature and the historical product price feature can be made less than a preset distance threshold. This ensures that the current product price of the same advertisement will not change too much compared to the historical product price, such as changing from 100 to 10000, thereby ensuring the stability of the pushed product advertisement.

[0132] In the above embodiments, by using a preset distance threshold to constrain the current dynamic attribute features, the feature distance between the current dynamic attribute features and the historical dynamic attribute features is prevented from exceeding the preset distance threshold. This improves the accuracy of the obtained promotion features and avoids pushing promotion information with excessively large dynamic attribute changes, thus ensuring the stability of the pushed promotion information.

[0133] In some embodiments, the promotional information push method further includes the step of:

[0134] The process involves: obtaining the object attributes corresponding to the target object identifier; obtaining the static and dynamic attributes corresponding to the current promotional information; inputting the object attributes, static attributes, and dynamic attributes into the promotional information push model; extracting object attribute features from the object network within the promotional information push model; obtaining promotional interaction features corresponding to the current promotional information through the promotional information network within the promotional information push model; extracting static attribute features from static attributes and dynamic attribute features from dynamic attributes; obtaining the aggregation weight associated with the target object identifier; performing aggregation operations on the static and dynamic attribute features based on the aggregation weights to obtain promotional attribute features; and fusing the promotional attribute features and promotional interaction features to obtain the current promotional features corresponding to the current promotional information. Finally, calculating the matching degree between the object attribute features and the current promotional features through the push prediction network within the promotional information push model, and determining the push result of the current promotional information for the target object identifier based on the matching degree.

[0135] The promotional information push model is a neural network model that predicts whether to push promotional information based on the identifier of the target object. It is pre-trained using historical data through a neural network, which includes at least one of the following: convolutional neural network, recurrent neural network, feedforward neural network, attention neural network, temporal neural network, etc. This promotional information push model is based on a dual-tower model. The object network is a neural network used to extract object attribute features. The promotional information network is a neural network used to extract promotional features. The object network and the promotional information network are parallel neural networks in the promotional information push model. The push prediction network is a neural network used to predict push notifications.

[0136] In this embodiment, a pre-trained promotional information push model is pre-deployed in the server. When a promotional message needs to be pushed, its static and dynamic attributes, as well as the object attributes corresponding to the target object identifier, are obtained. The server then invokes the deployed promotional information push model, inputting the object attributes, static attributes, and dynamic attributes into it. The model inputs the object attributes into the object network to obtain the output object attribute features, and inputs the static and dynamic attributes into the promotional information network to obtain the output current promotional features. Finally, the object attribute features and the current promotional features are input into the push prediction network to calculate the matching degree, and the push result for the target object identifier is determined based on the matching degree.

[0137] In a specific embodiment, such as Figure 5The diagram shows the architecture of a promotional information push model, which includes an object network, a promotional information network, and a push prediction network. Specifically, the server inputs object attributes into the object network for computation to obtain object attribute features. Simultaneously, it inputs the dynamic and static attributes of the current promotional information into the promotional information network for computation to obtain the output current promotional features. Then, the object attribute features and the current promotional features are input into the push prediction network for computation to obtain the output matching degree. The matching degree determines the push result of the current promotional information targeting the target object identifier. For example, ... Figure 6 The diagram shows the network architecture of an object network, which includes an embedding representation layer and a deep neural network. Specifically, the server obtains the object attributes corresponding to the identifier of the object to be pushed, inputs each object attribute into the embedding representation layer for embedding identifier mapping, obtains the object attribute features, and then concatenates the object attribute features to obtain object concatenation features. At this time, the object concatenation features are input into the deep neural network, which can be a DNN neural network, which includes multiple neural network layers, to obtain the output object attribute features.

[0138] In the above embodiments, the matching degree of the object to be pushed to the current promotional information is predicted by using the object network, promotional information network and push prediction network in the promotional information push model. That is, the efficiency of promotional information push prediction can be improved by using a pre-trained and deployed promotional information push model for push prediction.

[0139] In some embodiments, the promotion information network includes: a feature extraction network, an aggregation operation network associated with the identifier of the object to be pushed, and a fusion operation network;

[0140] The promotion interaction features corresponding to the current promotion information are obtained through the promotion information network in the promotion information push model. Static attribute features of static attributes and dynamic attribute features of dynamic attributes are extracted, and the aggregation weight associated with the identifier of the object to be pushed is obtained. Based on the aggregation weight, the static attribute features and dynamic attribute features are aggregated to obtain the promotion attribute features. Finally, the promotion attribute features and promotion interaction features are fused to obtain the current promotion features corresponding to the current promotion information, including:

[0141] The feature extraction network obtains the promotion interaction features corresponding to the current promotion information, and extracts the static attribute features of static attributes and the dynamic attribute features of dynamic attributes. The aggregation operation network performs aggregation operation on the static attribute features and dynamic attribute features to obtain the promotion attribute features. The fusion operation network performs fusion operation on the promotion attribute features and promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

[0142] Among them, the feature extraction network is a neural network used to convert static and dynamic attributes into feature vectors; for example, the feature extraction network can be an embedding representation network. The aggregation operation network for the identifier association of the object to be pushed refers to a neural network that performs aggregation operations on static and dynamic attribute features. The fusion operation network refers to a neural network that performs fusion operations on promotion attribute features and promotion interaction features.

[0143] In this embodiment, the server inputs the static attributes, dynamic attributes, and promotional interaction data corresponding to the current promotional information into the promotional network. The feature extraction network within the promotional network extracts features from the static attributes, dynamic attributes, and promotional interaction data to obtain promotional interaction features, static attribute features, and dynamic attribute features. Then, the static attribute features and dynamic attribute features are input into the aggregation operation network for aggregation operations to obtain the promotional attribute features. Finally, the promotional attribute features and promotional interaction features are input into the fusion operation network for fusion operations to obtain the current promotional features corresponding to the current promotional information.

[0144] In a specific embodiment, such as Figure 7 The diagram illustrates the network architecture of a promotion information network, which includes an embedding representation network, an aggregation operation network, and a fusion operation network. Specifically: The server obtains the dynamic attributes, static attributes, and promotion interaction data corresponding to any promotional information. Then, the dynamic attributes, static attributes, and promotion interaction data are input into the embedding representation network for feature extraction, resulting in promotion interaction features, static attribute features, and dynamic attribute features. Next, the static and dynamic attribute features are input into the aggregation operation network for aggregation operations, yielding promotion attribute features. Finally, the promotion attribute features are concatenated with the promotion interaction data to obtain concatenated promotion features. These concatenated features are then input into the fusion operation network, which can be a deep neural network, to obtain the promotion features corresponding to the output promotional information.

[0145] In the above embodiments, the promotion features corresponding to the promotion information are extracted by the joint action of the feature extraction network, the aggregation operation network associated with the identifier of the object to be pushed, and the fusion operation network. That is, the promotion features are extracted by training neural networks with different functions, thereby improving the accuracy of the obtained promotion features.

[0146] In some embodiments, such as Figure 8 As shown, training the promotional information push model includes the following steps:

[0147] S802, Obtain the pre-trained promotional information push model. The pre-trained promotional information push model is obtained by training the initial promotional information push model using the training data and training labels corresponding to the training object identifier.

[0148] Among them, the pre-trained promotion information push model refers to the promotion information push model that has been trained in advance.

[0149] The training object identifier refers to the object identifier that needs to be used to determine the push of promotional information during model training. In some embodiments, the training object identifier can also be the identifier to be pushed.

[0150] Training data refers to the data used when training the promotional information push model. This training data can include training object attributes and the static, dynamic, and interactive data of historically pushed promotional information. Training labels are the tags used when training the promotional information push model. These training labels can be determined based on the interaction behavior of training object identifiers with historically pushed promotional information. For example, training labels can include tags indicating whether a historically pushed promotional information was clicked. The training data corresponding to the training object identifier can include the training object attributes corresponding to the training object identifier, the static, dynamic, and interactive data of the historically pushed promotional information corresponding to the training object identifier. Notably, the dynamic attributes and interactive data of the same historically pushed promotional information can differ in the training data corresponding to different training object identifiers. The initial promotional information push model refers to the promotional information push model with initialized model parameters.

[0151] In this embodiment, the server can pre-train a pre-trained promotional information push model by training the initial promotional information push model using training data and training labels corresponding to the training object identifiers. Then, an aggregation operation network associated with different object identifiers can be trained using the pre-trained promotional information push model and the training data and training labels corresponding to different object identifiers.

[0152] S804, obtain the training data and training labels corresponding to the identifier of the object to be pushed.

[0153] In this embodiment, the server can obtain the object attributes corresponding to the object identifier to be pushed from the database, as well as the static attributes, dynamic attributes, and promotion interaction data of the historical promotion information corresponding to the object identifier to be pushed, to obtain the training data corresponding to the object identifier to be pushed, and obtain the interaction behavior data of the object identifier to be pushed on the historical promotion information, and determine the training label of the historical promotion information through the interaction behavior data.

[0154] S806, input the training data corresponding to the identifier of the object to be pushed into the pre-trained promotion information push model, and obtain the training push result of the output training promotion information for the identifier of the object to be pushed.

[0155] Among them, the training push result refers to the push result of the training promotion information obtained during model training for the target object identifier.

[0156] In this embodiment, the server inputs the training data corresponding to the identifier of the object to be pushed into the pre-trained promotion information push model. The pre-trained promotion information push model can execute any step of the promotion information push model in the above-described promotion information push method embodiment, thereby obtaining the training push result of the output training promotion information for the identifier of the object to be pushed.

[0157] S808 updates the aggregation operation network in the pre-trained promotion information push model based on the training push results and training labels, obtains the updated aggregation operation network corresponding to the identifier of the object to be pushed, and obtains the updated promotion information push model based on the updated aggregation operation network.

[0158] The updated aggregation operation network refers to the aggregation operation network after its parameters have been updated. This aggregation operation network can be configured with aggregation weights. Specifically, the aggregation operation network in the pre-trained promotional information push model can be trained using the training push results and training labels corresponding to the target object identifier, thereby obtaining the aggregation operation network corresponding to the target object identifier, i.e., obtaining the aggregation weights associated with the target object identifier.

[0159] In this embodiment, the server calculates the error between the training push result and the training label to obtain training loss information. This training loss information can be calculated using the cross-entropy loss function. Then, using the training loss information, the gradient descent algorithm is used to back-update the network parameters of the aggregation operation network in the pre-trained promotion information push model, while keeping other model parameters in the pre-trained promotion information push model unchanged, such as feature extraction parameters, fusion operation parameters, and push prediction parameters. Once the network parameters of the aggregation operation network are updated, the updated aggregation operation network corresponding to the target object identifier is obtained, and the updated promotion information push model is obtained based on the updated aggregation operation network.

[0160] In a specific embodiment, the training loss information can be calculated using the formula (2) shown below to update the aggregation operation network.

[0161] Formula (2)

[0162] Where N is the sum of the number of all samples, and y represents the training label of the i-th training data. This represents the training push result for the i-th training data.

[0163] S810 updates the promotion information push model as a pre-trained promotion information push model, and returns to iteratively execute the steps of obtaining the training data corresponding to the identifier of the object to be pushed and the training labels corresponding to the training data until the training completion condition is met, thus obtaining the promotion information push model including the aggregation operation network associated with the identifier of the object to be pushed.

[0164] In this embodiment, the server iteratively trains the promotional information push model until the training completion condition is met. The promotional information push model that has met the training completion condition is then used as the final trained model, which includes an aggregation computation network associated with the identifier of the target object. The training completion condition includes, but is not limited to, model parameters no longer changing, training loss information reaching a preset threshold, or the number of training iterations reaching the maximum number of iterations. At this point, the server can deploy the promotional information push model including the aggregation computation network associated with the identifier of the target object. Then, if a promotional information push request corresponding to the identifier of the target object is received, the deployed promotional information push model including the aggregation computation network associated with the identifier of the target object can be used to predict whether the promotional information will be pushed.

[0165] In a specific embodiment, such as Figure 9 The diagram illustrates the training architecture of the promotional information push model during training. The training object attributes of the training object identifier are input into the embedding representation layer of the object network for embedding representation mapping, yielding training object attribute features. These object attribute features are then input into a deep neural network for deep feature extraction, resulting in output object attribute features. Simultaneously, the training static attributes, training dynamic attributes, and training promotional interaction data of the historical promotional information corresponding to the training object identifier are input into the embedding representation layer of the promotional network for embedding representation mapping, yielding training static attribute features, training dynamic attribute features, and training promotional interaction features. The training static and dynamic attribute features are then input into an aggregation operation network for aggregation operations, yielding training promotional attribute features. The sum of the training promotional interaction features and the training promotional attribute features is calculated and input into the deep neural network for deep feature extraction, resulting in output training promotional features. Finally, the training object attribute features and the training promotional features are multiplied by a dot product to obtain the output training matching degree. The training push result is then determined based on the training matching degree. Then, the training loss information between the training push results and the training labels is calculated, and the training loss information is used to update the aggregation operation network. The training is iterated continuously until the training completion condition is met, and a promotion information push model including the aggregation operation network associated with the identifier of the object to be pushed is obtained.

[0166] In some embodiments, the server can obtain positive and negative samples for training. Positive samples can be training data pushed to the target terminal and for which click behavior data is obtained, while negative samples can be training data pushed to the target terminal without which click behavior data is obtained.

[0167] In the above embodiments, by obtaining a pre-trained promotional information push model, and then using the training data and training labels corresponding to the identifier of the target object to be pushed to train the aggregation operation network in the pre-trained promotional information push model, a promotional information push model including the aggregation operation network associated with the identifier of the target object to be pushed is obtained. This allows the aggregation operation network in the promotional information push model to be associated with the identifier of the target object, thereby improving the accuracy of model training and thus improving the accuracy of the obtained promotional information push model in predicting whether to push promotional information to the identifier of the target object.

[0168] In some embodiments, S206, obtaining the current promotion feature corresponding to the current promotion information includes the following steps:

[0169] Get the current promotion index corresponding to the current promotion information; based on the current promotion index, search for the corresponding promotion feature in the promotion feature database to obtain the current promotion feature corresponding to the current promotion information. The promotion index and the corresponding promotion feature are pre-stored in the promotion feature database.

[0170] The current promotion index refers to the index for finding the promotion feature corresponding to the current promotion information. This index can be generated by indexing the promotion features; for example, it can be obtained by calculating the hash value of the promotion feature, or by generating a unique identifier corresponding to the promotion feature. The promotion feature database is a database used to store the promotion features corresponding to each promotion information. This promotion feature database also stores the index corresponding to each promotion feature.

[0171] In this embodiment, the server can extract the corresponding promotional features upon receiving promotional information and save these features to a promotional feature database. It can also save the generation time of the promotional features. If the dynamic attributes corresponding to the promotional information change, the server can obtain the changed dynamic attributes and use them to extract updated promotional features. These updated features are then also saved to the promotional feature database. The update time can also be obtained, and the update time and updated promotional features are associated and saved to the promotional feature database. At this point, the promotional feature database stores promotional features with the same index corresponding to different times. Then, if a promotional feature needs to be used, the latest promotional feature can be quickly found and used based on the index and time.

[0172] In the above embodiments, the current promotion feature corresponding to the current promotion information is searched in the database that stores the promotion index and the corresponding promotion feature in advance by using the current promotion index. That is, the efficiency of obtaining the promotion feature can be improved by searching for the corresponding promotion feature through the index.

[0173] In some embodiments, the current promotional information includes at least two:

[0174] S206, Obtain the current promotion features corresponding to the current promotion information, including the following steps:

[0175] Obtain the current promotion features corresponding to at least two current promotion information.

[0176] S208, calculate the matching degree between object attribute features and current promotion features, and determine the push result of the current promotion information for the target object identifier based on the matching degree, including the following steps:

[0177] Calculate the degree of matching between the object attribute features and each current promotion feature, and determine the push results of at least two current promotion information for the object identifier to be pushed based on each degree of matching.

[0178] In this embodiment, after receiving a promotional information push request, the server can obtain multiple promotional information items that need to be pushed. At this time, the server obtains the promotional features corresponding to each promotional information item. Each promotional feature is obtained by fusing the promotional attribute features and promotional interaction features of the corresponding promotional information item. The promotional attribute features of each promotional information item are obtained by aggregating the corresponding static attribute features and dynamic attribute features. Then, the server can calculate the matching degree between each promotional feature and the object attribute features, thereby obtaining the push result for each promotional information item targeting the target object identifier. The push result may include the result of not pushing some promotional information and pushing some promotional information. At this point, the server can push and display all the promotional information items that can be pushed to the requesting terminal corresponding to the target object identifier.

[0179] In the above embodiments, by obtaining the promotion features corresponding to different promotion information, and then calculating the matching degree between each promotion feature and the object attribute features, the push result of each promotion information to the target object can be determined. In this way, the push results of multiple different promotion information can be quickly determined and pushed, thereby improving the efficiency of promotion information push.

[0180] In some embodiments, the promotional information push request is triggered by the dynamic content display page corresponding to the identifier of the object to be pushed;

[0181] S208, Determine the push result of the current promotional information for the target audience based on the matching degree, including the following steps:

[0182] Based on the matching degree, the current promotional information is pushed to the dynamic content display page corresponding to the target audience. The dynamic content display page is used to display dynamic content and the pushed promotional information.

[0183] Dynamic content display pages refer to pages used to display dynamic content, which is content displayed in a streaming format on the page. For example, dynamic content can be an information stream arranged chronologically on the page, allowing users to access the latest or most interesting content. It can also be an information stream arranged in the order of content publication. Furthermore, it can be an information stream ordered according to a pre-defined order, which can be a custom order or an order predefined by the information provider. Dynamic content can also be personalized content pushed based on data such as the user's interests, behaviors, and preferences. Dynamic content takes various forms, including but not limited to text, images, videos, audio, and other media. It may originate from various sources, such as news media, social media, and brand promotions. Dynamic content can also be published by websites, applications, social media platforms, search engines, and other digital media. Dynamic content can exist and be visualized on different platforms, i.e., it can be displayed on mobile applications, websites, or social media platforms. For example, dynamic content can be social media messages displayed on a social media page of a social media platform, and continuously updated dynamic messages that can be viewed through the social media page. Dynamic content can be various news items, articles, etc., displayed on the information page of a news app. Dynamic content can be real-time, such as content posted in real-time on social media platforms. Dynamic content can also be pre-set, such as pre-defined dynamic content displayed within a news app.

[0184] In this embodiment, the promotional information push method of this application is applied to a dynamic content display page. When an object browses the dynamic content on the dynamic content display page, it can trigger a promotional information push request. At this time, the server corresponding to the dynamic content display page receives the promotional information push request and responds to the request by filtering the promotional information to be pushed according to the steps in any of the above embodiments of the promotional information push method, and then pushes the promotional information to the dynamic content display page. At this time, the dynamic content display page receives the promotional information and can display it according to a pre-set display format. For example, it can be displayed together with the dynamic content in the form of an information stream, or it can be displayed on the dynamic content display page in a non-information stream format such as a pop-up window or banner.

[0185] In the above embodiments, by obtaining the dynamic content display page corresponding to the target object identifier and triggering the promotion information push request, the current promotion information is pushed to the dynamic content display page based on the matching degree. Then, the promotion information can be pushed to the dynamic content display page according to the push result. Finally, the dynamic content display page can display the dynamic content and the pushed promotion information at the same time, so that the target can browse the dynamic content and the pushed promotion information at the same time, which improves the push effect of the promotion information and increases the possibility of conversion of the promotion information, thereby saving push resources.

[0186] In a specific embodiment, such as Figure 10 The diagram illustrates a flowchart of a method for pushing promotional information, executed by a computer device, which can be a server or a terminal. In this embodiment, a distributed data node is used as the server for illustration. Specifically:

[0187] S1002, obtain the promotion information push request. The promotion information push request carries the identifier of the object to be pushed. Obtain the object attributes corresponding to the identifier of the object to be pushed, and obtain the static and dynamic attributes corresponding to the current promotion information.

[0188] S1004. Input the object attributes, static attributes, and dynamic attributes into the promotion information push model. Obtain the numerical attributes and categorical attributes corresponding to the identifier of the object to be pushed through the object network in the promotion information push model. Map the categorical attributes through a preset embedding lookup table to obtain categorical features. Then, concatenate the categorical features with the numerical attributes to obtain the object attribute features.

[0189] S1006, by embedding and representing static attributes through the feature extraction network in the promotion information network of the promotion information push model, static attribute features are obtained; content attributes are embedded and represented to obtain content attribute features; and promotion vendor status attributes are embedded and represented to obtain promotion vendor status features. The content attribute features and promotion vendor status features are concatenated to obtain the current dynamic attribute features.

[0190] S1008: Obtain the historical dynamic attribute features corresponding to the current promotional information, and calculate the feature distance between the current dynamic attribute features and the historical dynamic attribute features. If the feature distance is less than a preset distance threshold, then the current dynamic attribute features are taken as the target dynamic attribute features corresponding to the current promotional information;

[0191] S1010, if the feature distance is greater than or equal to the preset distance threshold, then obtain the updated dynamic attribute feature corresponding to the current dynamic attribute feature, and use the updated dynamic attribute feature as the target dynamic attribute feature corresponding to the current promotion information.

[0192] Among them, the feature distance between the updated dynamic attribute features and the historical dynamic attribute features is less than a preset distance threshold.

[0193] S1012, the static attribute features are weighted by the aggregation operation network based on the static aggregation weight to obtain the static weighted features, and the target dynamic attribute features are weighted by the dynamic aggregation weight to obtain the dynamic weighted features. The feature fusion of the static weighted features and the dynamic weighted features is calculated to obtain the generalized attribute features.

[0194] Among them, the aggregation operation network is associated with the identifier of the object to be pushed.

[0195] S1014. Obtain the promotion interaction features corresponding to the current promotion information through the fusion computing network, perform linear mapping on the promotion attribute features through the linear mapping network to obtain linear mapping features, and fuse the linear mapping features with the promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

[0196] Among them, the linear mapping feature and the promotion interaction feature have the same feature dimension.

[0197] S1016, the matching degree between the object attribute features and the current promotion features is calculated by the push prediction network in the promotion information push model, and the push result of the current promotion information for the target object identifier is determined based on the matching degree.

[0198] In the above embodiments, by using the aggregation operation network in the promotion information push model to aggregate static attribute features and dynamic attribute features, and further fusing them with the promotion interaction features corresponding to the current promotion information, the current promotion features can be obtained, thus enabling the obtained current promotion features to accurately represent the current promotion information. Furthermore, by aggregating through an aggregation network associated with the identifier of the target object, the aggregated promotion attribute features can better reflect the importance of the current promotion information to the identifier of the target object, further improving the accuracy of the obtained current promotion features. Moreover, by calculating the feature distance between the current dynamic attribute features and historical dynamic attribute features, the current dynamic attribute features are limited to a preset distance threshold, thereby keeping the changes in dynamic attributes stable and avoiding excessive changes in dynamic attributes, thus ensuring the accuracy of the obtained dynamic attribute features. Then, the linear mapping feature is fused with the promotion interaction feature to obtain the current promotion feature corresponding to the current promotion information. That is, even when the promotion interaction feature corresponding to the new promotion information is empty, the linear mapping feature and the promotion interaction feature can still be fused to obtain the current promotion feature, ensuring that the obtained current promotion feature has a feature value. Then, the matching degree between the current promotion feature and the object attribute feature is calculated to determine the push result of the current promotion information for the target object identifier, which improves the accuracy of the push result and thus improves the accuracy of the push and reduces the waste of push resources.

[0199] In a specific embodiment, such as Figure 11The diagram illustrates the timing of promotional information push notifications. Specifically: The vendor sending the promotional information can send a push command containing the current promotional information to the server via their terminal. The server retrieves the current promotional information and extracts its corresponding promotional features. This involves extracting the static and dynamic attributes of the current promotional information, aggregating these static and dynamic attributes to obtain the promotional attribute features, and then fusing these features with the promotional interaction features to arrive at the current promotional features. The server then saves these current promotional features and returns a notification indicating that the push settings are complete to the vendor's terminal. Next, the server receives a promotional information push request sent by the user via their terminal. This request carries the identifier of the target audience. At this point, the server inputs the object attributes corresponding to the target object identifier into the promotion information push model to predict the push of promotional information, thus obtaining the various promotional messages to be pushed. Specifically, the promotion information push model extracts the features of the object attributes, obtains the object attribute features, and sequentially retrieves the current promotional feature to be predicted from all saved promotional features. It then calculates the matching degree between the object attribute features and the current promotional feature, and determines the push result of the current promotional message for the target object identifier based on the matching degree. Finally, the server obtains the push result of the various promotional messages to be pushed and returns them to the user terminal. The user terminal receives and displays the promotional information pushed by the server. In other words, by using the promotion information push model for push prediction, the server can improve the efficiency of obtaining the various promotional messages to be pushed, thereby enhancing the user experience.

[0200] In a specific embodiment, such as Figure 12As shown, a schematic diagram of a model architecture for real-time push of promotional information is provided. Specifically, the server obtains the push request for promotional information in real time. At this time, the server obtains the real-time object attributes and inputs the object attributes into the push model for promotional information. The push model for promotional information extracts the embedding representation features of the object attributes through the embedding representation layer to obtain the embedding representation features. The embedding representation features are then input into a deep neural network for deep feature extraction to obtain the object attribute features. The deep neural network can be a DNN network. Through the fully connected layer and the ReLU (Linear rectification function) activation function layer inside the DNN network, the deep representation of the features can be extracted. The ReLU activation function helps to increase the nonlinear expression capability of the model. At this time, the push model for promotional information can sequentially search for the promotional features associated with the promotional feature index from the promotional feature index database to obtain the current promotional feature. For example, the nearest neighbor retrieval method can be used to calculate the promotional feature closest to it, and then filter out the top N promotional information that best matches it. Then, the dot product operation result between the current promotional feature and the object attribute feature is calculated to obtain the matching degree. The matching degree can be calculated using the formula (3) shown below.

[0201] Formula (3)

[0202] in, It refers to the attribute characteristics of an object. It refers to the sigmoid function. "y" refers to the promotion feature. "y" refers to the matching degree. If the matching degree is greater than a preset threshold, the current promotion information corresponding to the current promotion feature is used as the promotion information to be pushed to the requesting terminal. Specifically, the server pre-extracts the promotion features of each promotion information, constructs an index for each promotion feature, and associates and saves the index and promotion feature in the database. For example, ... Figure 13The diagram shows the architecture for extracting promotional information features, specifically the network architecture of the promotion network. Specifically: The server inputs the dynamic attributes, static attributes, and promotional interaction data corresponding to the promotional information into the embedding representation layer of the promotion network for embedding representation mapping, obtaining dynamic attribute features, static attribute features, and promotional interaction features. Then, the dynamic and static attribute features are input into an aggregation operation network for aggregation operations, resulting in aggregated features. At this point, the feature dimension of the aggregated features is different from the feature dimension of the promotional interaction features; that is, the aggregated features are input into a linear mapping network for linear mapping using linear weights. This linear mapping network can be a single-layer neural network or a multi-layer neural network, resulting in linearly mapped features. The feature dimension of these linearly mapped features is the same as the dimension of the promotional interaction features. The server then calculates the sum of the linearly mapped features and the promotional interaction features, i.e., performs element-wise addition, obtaining the summed features. Finally, the summed features are input into a deep neural network for deep feature extraction, yielding the output promotional features. All extracted features are normalized using a normalization algorithm, such as L2 norm normalization.

[0203] The advertising push model can then be further validated by comparing the traditional dual-tower model with the dual-tower model that directly concatenates attribute features and interaction features. The validation comparison results are shown in Table 1 below.

[0204] Table 1. Model Validation Comparison Results

[0205]

[0206] Here, AUC refers to Area Under the ROC Curve, an important metric for evaluating the performance of a binary classification model. Recall10@1 is a metric for evaluating information push performance, representing the proportion of the top N scores given by the current model to the top K true rankings of all samples. For example, if the true ranking of 3 samples is [a, b, c, d], and the push model scores [b, c, d, a], then Recall10@2 is 0.5 because the top 3 scores are [b, c, d], while the top 2 true rankings are [a, b]. The [b, c, d] group contains 'b', which accounts for 50% of the top 2 true rankings. Similarly, if the push model scores [b, a, c, d], and the top 3 rankings [b, a, c] contain both 'a' and 'b' from the true rankings, then Recall10@2 is 1. As can be clearly seen from Table 1, this application significantly improves the AUC by 2.19% and the recall10@1 by 3.35% compared to the traditional dual-tower model. Compared to the directly spliced ​​model, this application also significantly improves the AUC by 1.56% and the recall10@1 by 2.40%.

[0207] In one specific embodiment, this promotional information push method can be applied to social media platforms for advertising push. Specifically: when a user browses social dynamic content through a social page on a social media platform, an advertising push request can be triggered. At this time, the social media platform's server receives the advertising push request, responds by obtaining user attribute information corresponding to the user identifier, and then inputs the user attribute information and each ad index to be pushed into a pre-deployed promotional information push model. The promotional information push model predicts the push results for the ads corresponding to each pre-saved ad index to be pushed, obtains the matching degree between the user and each ad, and selects a preset number of ads as the advertising information to be pushed according to the matching degree. The ads corresponding to each ad index to be pushed can include previously pushed ads and new ads uploaded by advertising companies. Then the server pushes the selected preset number of ad information to the user's social media platform, where the pushed advertising information is displayed. The user can then view the pushed advertising information through the social media platform. That is, by using the promotional information push model for advertising push, the problem of low advertising push accuracy during the cold start of new ads is solved, thereby improving the accuracy of advertising push and saving advertising push resources.

[0208] In one specific embodiment, this promotional information push method can be applied to news push on a news information platform. Specifically: when a user displays news on the news feed page of the news information platform and needs to obtain the latest news, they can trigger a news push request. At this time, the news information platform's server receives the news push request, obtains user attribute information based on the user identifier carried in the news push request, and then inputs the user attribute information and the index of each news item to be pushed into a pre-trained promotional information push model to predict the push result, obtaining the matching degree between each news item to be pushed and the user. Each news item to be pushed can include historically pushed news and real-time news that has not been pushed before. Then, multiple news items to be pushed are selected sequentially according to the matching degree as the news to be pushed to the user's request end. The server can also further filter the multiple news items to be pushed using a pre-set fine ranking model to obtain the most relevant preset number of news items, and then push the preset number of news items to the user's request end. The request end receives the pushed news and displays the news in the form of an information stream on the news feed page. This news can be a brief news item. When the user browses a brief news item that interests them, they can click on the brief news item to view the news details page. This involves using a pre-trained promotional information push model to predict the matching degree between each news item to be pushed and the user, then filtering out news items relevant to the user based on the matching degree, and finally pushing the relevant news items. This solves the problem of low news push accuracy during the cold start of real-time news, thereby improving the accuracy of news push and saving news push resources.

[0209] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0210] Based on the same inventive concept, this application also provides a promotional information push device for implementing the aforementioned promotional information push method. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more promotional information push device embodiments provided below can be found in the limitations of the promotional information push method described above, and will not be repeated here.

[0211] In one embodiment, such as Figure 14 As shown, a promotional information push device 1400 is provided, including: a request module 1402, an object feature acquisition module 1404, a promotional feature acquisition module 1406, and a push confirmation module 1408, wherein:

[0212] Request module 1402 is used to obtain a promotional information push request, which carries the identifier of the object to be pushed to.

[0213] The object feature acquisition module 1404 is used to acquire the object attribute features corresponding to the identifier of the object to be pushed.

[0214] The promotion feature acquisition module 1406 is used to acquire the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight is associated with the identifier of the object to be pushed.

[0215] The push determination module 1408 is used to calculate the degree of matching between the object attribute features and the current promotion features, and to determine the push result of the current promotion information for the target object identifier based on the degree of matching.

[0216] In one embodiment, the promotion feature acquisition module 1406 is further configured to acquire the static attributes corresponding to the current promotion information, and extract the features of the static attributes to obtain static attribute features; acquire the dynamic attributes corresponding to the current promotion information, and extract the features of the dynamic attributes to obtain dynamic attribute features; acquire the aggregation weight associated with the identifier of the object to be pushed, and perform aggregation operation on the static attribute features and dynamic attribute features based on the aggregation weight to obtain promotion attribute features; acquire the promotion interaction features corresponding to the current promotion information, and perform fusion operation on the promotion attribute features and promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

[0217] In one embodiment, the aggregation weight includes static aggregation weight and dynamic aggregation weight; the promotion feature acquisition module 1406 is further configured to perform weighted operation on static attribute features based on static aggregation weight to obtain static weighted features; perform weighted operation on dynamic attribute features based on dynamic aggregation weight to obtain dynamic weighted features; and calculate the feature fusion of static weighted features and dynamic weighted features to obtain promotion attribute features.

[0218] In one embodiment, the promotion feature acquisition module 1406 is further configured to acquire the promotion information push model associated with the identifier of the object to be pushed; use the network parameters of the aggregation operation network in the promotion information push model as the aggregation weight associated with the identifier of the object to be pushed; and perform aggregation operation on the static attribute features and dynamic attribute features based on the aggregation weight to obtain the promotion attribute features.

[0219] In one embodiment, the promotion feature acquisition module 1406 is further configured to perform linear mapping on the promotion attribute features through a linear mapping network to obtain linear mapping features, and match the feature dimensions of the linear mapping features with the promotion interaction features; and fuse the linear mapping features with the promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

[0220] In one embodiment, the dynamic attributes include the content attributes and the promotion vendor status attributes corresponding to the current promotion information; the promotion feature acquisition module 1406 is also used to embed and represent the static attributes to obtain static attribute features; embed and represent the content attributes to obtain content attribute features, and embed and represent the promotion vendor status attributes to obtain promotion vendor status features; and concatenate the content attribute features and the promotion vendor status features to obtain dynamic attribute features.

[0221] In one embodiment, the object feature acquisition module 1404 is further configured to acquire the numerical attribute and categorical attribute corresponding to the identifier of the object to be pushed; to perform embedding representation mapping on the categorical attribute through a preset embedding lookup table to obtain the categorical feature; and to concatenate the categorical feature with the numerical attribute to obtain the object attribute feature.

[0222] In one embodiment, the promotion feature acquisition module 1406 is further configured to acquire the static attribute features corresponding to the current promotion information, and acquire the current dynamic attribute features and historical dynamic attribute features corresponding to the current promotion information; calculate the offset between the current dynamic attribute features and the historical dynamic attribute features, and determine the target dynamic attribute features corresponding to the current promotion information based on the offset and the current dynamic attribute features; acquire the aggregation weight associated with the identifier of the object to be pushed, and perform aggregation operation on the static attribute features and the target dynamic attribute features based on the aggregation weight to obtain the target promotion attribute features; acquire the promotion interaction features corresponding to the current promotion information, and perform fusion operation on the target promotion attribute features and the promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

[0223] In one embodiment, the offset magnitude includes the feature distance;

[0224] The promotion feature acquisition module 1406 is also used to calculate the feature distance between the current dynamic attribute feature and the historical dynamic attribute feature; if the feature distance is less than a preset distance threshold, the current dynamic attribute feature is used as the target dynamic attribute feature corresponding to the current promotion information; if the feature distance is greater than or equal to the preset distance threshold, the updated dynamic attribute feature corresponding to the current dynamic attribute feature is acquired, and the updated dynamic attribute feature is used as the target dynamic attribute feature corresponding to the current promotion information, and the feature distance between the updated dynamic attribute feature and the historical dynamic attribute feature is less than the preset distance threshold.

[0225] In one embodiment, the promotional information push device 1400 further includes:

[0226] The model prediction module is used to obtain the object attributes corresponding to the target object identifier and the static and dynamic attributes corresponding to the current promotion information; input the object attributes, static attributes, and dynamic attributes into the promotion information push model; extract the object attribute features of the object attributes through the object network in the promotion information push model; obtain the promotion interaction features corresponding to the current promotion information through the promotion information network in the promotion information push model, extract the static attribute features of static attributes and the dynamic attribute features of dynamic attributes, and obtain the aggregation weight associated with the target object identifier. Based on the aggregation weight, the static attribute features and dynamic attribute features are aggregated to obtain the promotion attribute features, and the promotion attribute features and promotion interaction features are fused to obtain the current promotion features corresponding to the current promotion information; calculate the matching degree between the object attribute features and the current promotion features through the push prediction network in the promotion information push model, and determine the push result of the current promotion information for the target object identifier based on the matching degree.

[0227] In one embodiment, the promotion information network includes: a feature extraction network, an aggregation operation network associated with the identifiers of the objects to be pushed, and a fusion operation network;

[0228] The model prediction module is also used to obtain the promotion interaction features corresponding to the current promotion information through the feature extraction network, and extract the static attribute features of static attributes and the dynamic attribute features of dynamic attributes; the static attribute features and dynamic attribute features are aggregated and operated through the aggregation operation network to obtain the promotion attribute features; the promotion attribute features and promotion interaction features are fused and operated through the fusion operation network to obtain the current promotion features corresponding to the current promotion information.

[0229] In one embodiment, the promotional information push device 1400 further includes:

[0230] The model training module is used to obtain a pre-trained promotional information push model, which is obtained by training the initial promotional information push model using training data and training labels corresponding to the training object identifiers. It then obtains the training data and training labels corresponding to the identifiers to be pushed to; inputs the training data corresponding to the identifiers to be pushed to the pre-trained promotional information push model to obtain the training push result of the output promotional information for the identifiers to be pushed to; updates the aggregation operation network in the pre-trained promotional information push model based on the training push result and training labels to obtain the updated aggregation operation network corresponding to the identifiers to be pushed to; and obtains the updated promotional information push model based on the updated aggregation operation network. Finally, it uses the updated promotional information push model as the pre-trained promotional information push model and iteratively executes the steps of obtaining the training data and training labels corresponding to the identifiers to be pushed to until the training completion condition is met, resulting in a promotional information push model including the aggregation operation network associated with the identifiers to be pushed to.

[0231] In one embodiment, the promotion feature acquisition module 1406 is further configured to acquire the current promotion index corresponding to the current promotion information; search for the corresponding promotion feature in the promotion feature database based on the current promotion index to obtain the current promotion feature corresponding to the current promotion information; the promotion index and the corresponding promotion feature are pre-stored in the promotion feature database.

[0232] In one embodiment, the current promotion information includes at least two: the promotion feature acquisition module 1406 is further configured to acquire the current promotion features corresponding to at least two pieces of current promotion information respectively;

[0233] The push determination module 1408 is also used to calculate the degree of matching between the object attribute features and each current promotion feature, and to determine the push results of at least two current promotion information for the object identifier to be pushed based on each degree of matching.

[0234] In one embodiment, the promotional information push request is triggered by the dynamic content display page corresponding to the identifier of the object to be pushed; the push determination module 1408 is also used to determine the push result of the current promotional information to the dynamic content display page corresponding to the identifier of the object to be pushed based on the matching degree, and the dynamic content display page is used to display dynamic content and the pushed promotional information.

[0235] Each module in the aforementioned promotional information push device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0236] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 15 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores promotional information, promotional features of each promotional information, and indexes of each promotional feature. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a promotional information push method.

[0237] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 16As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a method for pushing promotional information. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0238] Those skilled in the art will understand that Figure 15 or Figure 16 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0239] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0240] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0241] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0242] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all 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.

[0243] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, etc., and are not limited to these.

[0244] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0245] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for pushing promotional information, characterized in that, The method includes: Obtain a promotional information push request, wherein the promotional information push request carries the identifier of the target object to be pushed to; Obtain the object attribute features corresponding to the identifier of the object to be pushed; Obtain the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight of the aggregation is associated with the identifier of the object to be pushed. Calculate the degree of matching between the object attribute features and the current promotion features, and determine the push result of the current promotion information for the object identifier to be pushed based on the degree of matching.

2. The method according to claim 1, characterized in that, The step of obtaining the current promotion features corresponding to the current promotion information includes: Obtain the static attributes corresponding to the current promotion information, and extract the features of the static attributes to obtain static attribute features; Obtain the dynamic attributes corresponding to the current promotion information, and extract the features of the dynamic attributes to obtain dynamic attribute features; Obtain the aggregation weight associated with the identifier of the object to be pushed, and perform aggregation operation on the static attribute features and the dynamic attribute features based on the aggregation weight to obtain the promotion attribute features; Obtain the promotion interaction features corresponding to the current promotion information, and perform a fusion operation on the promotion attribute features and the promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

3. The method according to claim 2, characterized in that, The aggregation weights include static aggregation weights and dynamic aggregation weights; The aggregation operation on the static attribute features and the dynamic attribute features based on the aggregation weight to obtain the generalized attribute features includes: The static attribute features are weighted based on the static aggregation weights to obtain static weighted features; The dynamic attribute features are weighted based on the dynamic aggregation weights to obtain dynamic weighted features; The feature fusion of the static weighted features and the dynamic weighted features is calculated to obtain the generalized attribute features.

4. The method according to claim 1, characterized in that, The determination of the promotion attribute features includes: Obtain the promotion information push model associated with the identifier of the object to be pushed; The network parameters of the aggregation operation network in the promotion information push model are used as the aggregation weights associated with the identifier of the object to be pushed; Based on the aggregation weight, the static attribute features and the dynamic attribute features are aggregated to obtain the generalized attribute features.

5. The method according to claim 1, characterized in that, The determination of the current promotion features includes: The promotion attribute features are linearly mapped through a linear mapping network to obtain linear mapping features, and the linear mapping features are matched with the feature dimensions of the promotion interaction features. The linear mapping feature and the promotion interaction feature are fused to obtain the current promotion feature corresponding to the current promotion information.

6. The method according to claim 1, characterized in that, The dynamic attributes include the content attributes and the status attributes of the advertising provider corresponding to the current promotion information; The determination of the static and dynamic attribute features includes: The static attributes are embedded and represented to obtain static attribute features; The content attributes are embedded and represented to obtain content attribute features, and the promotion vendor status attributes are embedded and represented to obtain promotion vendor status features. The dynamic attribute features are obtained by concatenating the content attribute features and the promotion vendor status features.

7. The method according to claim 1, characterized in that, The step of obtaining the object attribute features corresponding to the identifier of the object to be pushed includes: Obtain the numeric and categorical attributes corresponding to the identifier of the object to be pushed; The categorical attribute is embedded and mapped using a preset embedding lookup table to obtain the categorical feature, and the categorical feature is concatenated with the numerical attribute to obtain the object attribute feature.

8. The method according to claim 1, characterized in that, The step of obtaining the current promotion features corresponding to the current promotion information includes: Obtain the static attribute features corresponding to the current promotion information, and obtain the current dynamic attribute features and historical dynamic attribute features corresponding to the current promotion information; Calculate the offset between the current dynamic attribute feature and the historical dynamic attribute feature, and determine the target dynamic attribute feature corresponding to the current promotion information based on the offset and the current dynamic attribute feature; Obtain the aggregate weight associated with the identifier of the object to be pushed, and perform an aggregation operation on the static attribute features and the target dynamic attribute features based on the aggregate weight to obtain the target promotion attribute features; Obtain the promotion interaction features corresponding to the current promotion information, and perform a fusion operation on the target promotion attribute features and the promotion interaction features to obtain the current promotion features corresponding to the current promotion information.

9. The method according to claim 8, characterized in that, The offset magnitude includes the characteristic distance; The step of calculating the offset between the current dynamic attribute feature and the historical dynamic attribute feature, and determining the target dynamic attribute feature corresponding to the current promotion information based on the offset and the current dynamic attribute feature, includes: Calculate the feature distance between the current dynamic attribute feature and the historical dynamic attribute feature; If the feature distance is less than a preset distance threshold, then the current dynamic attribute feature is taken as the target dynamic attribute feature corresponding to the current promotion information; If the feature distance is greater than or equal to the preset distance threshold, then the updated dynamic attribute feature corresponding to the current dynamic attribute feature is obtained, and the updated dynamic attribute feature is used as the target dynamic attribute feature corresponding to the current promotion information. The feature distance between the updated dynamic attribute feature and the historical dynamic attribute feature is less than the preset distance threshold.

10. The method according to claim 1, characterized in that, The method further includes: Obtain the object attributes corresponding to the identifier of the object to be pushed, and obtain the static and dynamic attributes corresponding to the current promotion information; The object attributes, the static attributes, and the dynamic attributes are input into the promotional information push model; The object attribute features of the object attributes are extracted through the object network in the promotion information push model; The promotion interaction features corresponding to the current promotion information are obtained through the promotion information network in the promotion information push model, and the static attribute features of the static attributes and the dynamic attribute features of the dynamic attributes are extracted. The aggregation weight associated with the identifier of the object to be pushed is obtained. The static attribute features and the dynamic attribute features are aggregated based on the aggregation weight to obtain the promotion attribute features. The promotion attribute features and the promotion interaction features are fused to obtain the current promotion features corresponding to the current promotion information. The matching degree between the object attribute features and the current promotion features is calculated by the push prediction network in the promotion information push model, and the push result of the current promotion information for the object to be pushed is determined based on the matching degree.

11. The method according to claim 10, characterized in that, The promotion information network includes: a feature extraction network, an aggregation operation network associated with the identifier of the object to be pushed, and a fusion operation network; The process of obtaining the promotional interaction features corresponding to the current promotional information through the promotional information network in the promotional information push model, extracting the static attribute features of the static attributes and the dynamic attribute features of the dynamic attributes, obtaining the aggregation weight associated with the identifier of the object to be pushed, performing aggregation operations on the static attribute features and the dynamic attribute features based on the aggregation weight to obtain promotional attribute features, and performing fusion operations on the promotional attribute features and the promotional interaction features to obtain the current promotional features corresponding to the current promotional information includes: The feature extraction network is used to obtain the promotion interaction features corresponding to the current promotion information, and the static attribute features of the static attributes and the dynamic attribute features of the dynamic attributes are extracted. The static attribute features and the dynamic attribute features are aggregated and processed by the aggregation operation network to obtain the generalized attribute features; The promotion attribute features and the promotion interaction features are fused and calculated using the fusion computing network to obtain the current promotion features corresponding to the current promotion information.

12. The method according to claim 10, characterized in that, The training of the promotional information push model includes the following steps: A pre-trained promotional information push model is obtained by training the initial promotional information push model using training data and training labels corresponding to the training object identifier. Obtain the training data and training labels corresponding to the identifier of the object to be pushed; The training data corresponding to the identifier of the object to be pushed is input into the pre-trained promotion information push model to obtain the training push result of the training promotion information for the identifier of the object to be pushed. Based on the training push results and the training labels, the aggregation operation network in the pre-trained promotion information push model is updated to obtain the updated aggregation operation network corresponding to the identifier of the object to be pushed, and the updated promotion information push model is obtained based on the updated aggregation operation network. The updated promotional information push model is used as a pre-trained promotional information push model, and the steps of obtaining the training data corresponding to the identifier of the object to be pushed and the training labels corresponding to the training data are iteratively executed until the training completion condition is met, thus obtaining a promotional information push model including the aggregation operation network associated with the identifier of the object to be pushed.

13. The method according to claim 1, characterized in that, The step of obtaining the current promotion features corresponding to the current promotion information includes: Obtain the current promotion index corresponding to the current promotion information; Based on the current promotion index, the corresponding promotion feature is searched in the promotion feature database to obtain the current promotion feature corresponding to the current promotion information. The promotion feature database pre-stores the promotion index and the corresponding promotion feature.

14. The method according to claim 1, characterized in that, The current promotional information includes at least two: The step of obtaining the current promotion features corresponding to the current promotion information includes: Obtain the current promotion features corresponding to the at least two current promotion information items respectively; The step of calculating the matching degree between the object attribute features and the current promotion features, and determining the push result of the current promotion information for the target object identifier based on the matching degree, includes: Calculate the degree of matching between the object attribute features and each current promotion feature, and determine the push results of the at least two current promotion information for the object identifier to be pushed based on each degree of matching.

15. The method according to claim 1, characterized in that, The promotional information push request is triggered through the dynamic content display page corresponding to the identifier of the target object; The step of determining the push result of the current promotional information for the target object identifier based on the matching degree includes: Based on the matching degree, the current promotional information is pushed to the dynamic content display page corresponding to the target object identifier. The dynamic content display page is used to display dynamic content and the pushed promotional information.

16. A promotional information push device, characterized in that, The device includes: The request module is used to obtain a promotional information push request, wherein the promotional information push request carries the identifier of the object to be pushed; The object feature acquisition module is used to acquire the object attribute features corresponding to the identifier of the object to be pushed; The promotion feature acquisition module is used to acquire the current promotion features corresponding to the current promotion information. The current promotion features are obtained by fusing the promotion attribute features and promotion interaction features corresponding to the current promotion information. The promotion attribute features are obtained by aggregating the static attribute features and dynamic attribute features corresponding to the current promotion information. The static attribute features and dynamic attribute features are obtained by extracting features from the static attributes and dynamic attributes corresponding to the current promotion information, respectively. The aggregation weight of the aggregation is associated with the identifier of the object to be pushed. The push determination module is used to calculate the matching degree between the object attribute features and the current promotion features, and determine the push result of the current promotion information for the object identifier to be pushed based on the matching degree.

17. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 15.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 15.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 15.

Images