Method, apparatus and medium for pushing media content

Abstract

The application discloses a media content pushing method, device, equipment and medium, relates to the technical field of Internet, and in particular to the technical field of multimedia. The method comprises the following steps: inputting object information of an obtained target object and content information of candidate media content into a matching model and a probability estimation model respectively, so as to obtain a matching value of the target object and the candidate media content and an initial probability of a preset operation of the target object; the preset operation is an operation of the target object on the expression interest degree of the candidate media content; and attention allocation processing is performed on the initial probability of the preset operation based on a fusion probability of the preset operation obtained by fusion processing on the initial probability and the matching value, so as to obtain a target probability of the preset operation; the target probability and the matching value are subjected to fusion processing to determine an interest score of the target object on the candidate media content, and if the interest score is greater than a preset threshold, the candidate media content is pushed to the target object. The method can reduce the resource consumption and waste of a network platform.

Description

Technical Field

[0001] This application generally relates to the field of Internet technology, specifically to the field of multimedia technology, and in particular to a method, apparatus, device, and medium for pushing media content. Background Technology

[0002] With the development and popularization of internet technology, the amount of media content online has also increased dramatically. Targeted delivery of media content that may interest a target audience has become a key focus for many online platforms. Furthermore, if the level of interest in media content can be anticipated before delivery, targeted content delivery can be achieved.

[0003] In existing technologies, information about the object to be predicted and the media content is input into a prediction model to predict the probability (e.g., click-through rate) of the object performing an action on the media content. This probability is then used to assess the object's interest in the media content and to push relevant media content. However, due to the diversity and complexity of media content, the relevant information suffers from sparsity, leading to low accuracy in probability prediction. This makes it easy for media content to be pushed based on predicted probabilities to result in ineffective pushes, wasting the network platform's resources. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide a method, apparatus, device and medium for pushing media content, which can improve the effectiveness of media content pushing, thereby reducing the overhead and waste of network platform resources.

[0005] Firstly, this application provides a method for pushing media content. The method includes: acquiring object information of a target object and content information of candidate media content; inputting the object information and content information into a matching model to obtain a matching value between the target object and the candidate media content; the matching value is used to characterize the degree of matching between the target object and the candidate media content; inputting the object information and content information into a probability prediction model to obtain an initial probability of the target object performing a preset operation; the preset operation is an operation in which the target object expresses interest in the candidate media content; fusing the initial probability and the matching value to obtain a fused probability of the preset operation, and performing attention allocation processing on the initial probability of the preset operation based on the fused probability of the preset operation to obtain a target probability of the preset operation; fusing the target probability and the matching value to determine the target object's interest score in the candidate media content, and pushing the candidate media content to the target object if the interest score is greater than a preset threshold.

[0006] Secondly, this application provides a media content push device, which includes: an acquisition module for acquiring object information of a target object and content information of candidate media content; a matching module for inputting the object information and content information into a matching model to obtain a matching value between the target object and the candidate media content; the matching value is used to characterize the degree of matching between the target object and the candidate media content; a probability prediction module for inputting the object information and content information into a probability prediction model to obtain an initial probability of the target object performing a preset operation; the preset operation is an operation in which the target object expresses interest in the candidate media content; a first prediction module for fusing the initial probability and the matching value to obtain a fused probability of the preset operation, and performing attention allocation processing on the initial probability of the preset operation based on the fused probability of the preset operation to obtain a target probability of the preset operation; and a second prediction module for fusing the target probability and the matching value to determine the target object's interest score in the candidate media content, and pushing the candidate media content to the target object when the interest score is greater than a preset threshold.

[0007] In one embodiment, the matching value includes a matching score and a matching probability; the matching module is specifically used for...

[0008] The object information and content information are input into the matching model, and feature transformation processing is performed on the object information and content information respectively to obtain object features and content features.

[0009] The object features are compressed to obtain the first transformation feature, and the content features are compressed to obtain the second transformation feature.

[0010] Determine the similarity between the first and second transformation features, and determine the matching score based on the similarity.

[0011] The matching probability is obtained by performing feature compression and normalization on the matching score.

[0012] In one embodiment, the preset operation includes a first operation, a second operation, and a third operation; the first operation is a preceding operation to the second operation; the second operation is a preceding operation to the third operation, and the degree of interest represented by the first operation, the second operation, and the third operation is from low to high.

[0013] In one embodiment, the probability prediction module is specifically used for,

[0014] By inputting the object information and content information into the probability prediction model, the initial probability of the first operation, the source probability of the second operation, and the source probability of the third operation are obtained.

[0015] The initial probability of the second operation is obtained based on the initial probability of the first operation and the source probability of the second operation.

[0016] The initial probability of the third operation is obtained based on the initial probability of the second operation and the source probability of the third operation.

[0017] In one embodiment, the probability prediction module is specifically used to fuse the initial probability of the first operation and the source probability of the second operation to obtain the initial probability of the second operation.

[0018] In one embodiment, the probability prediction module is specifically used to fuse the initial probability of the second operation and the source probability of the third operation to obtain the initial probability of the third operation.

[0019] In one embodiment, the first prediction module is specifically used for,

[0020] The matching probability is fused with the initial probability of the first operation to obtain the fused probability of the first operation.

[0021] The matching probability is fused with the initial probability of the second operation to obtain the fused probability of the second operation.

[0022] The matching probability is fused with the initial probability of the third operation to obtain the fused probability of the third operation.

[0023] In one embodiment, the probability prediction module is specifically used for,

[0024] The object information and content information are input into the probability prediction model, and feature transformation processing is performed on the object information and content information respectively to obtain object features and content features.

[0025] The object features and content features are concatenated to obtain the first concatenated feature.

[0026] Based on the first splicing feature, determine the initial weight of the preset operation.

[0027] The initial weights of the preset operation are fused with the first concatenated feature to obtain the fused feature of the preset operation.

[0028] Based on the fusion characteristics of the preset operation, the initial probability of the preset operation is obtained.

[0029] In one embodiment, the probability prediction module is specifically used for,

[0030] Based on the first splicing feature, the preset operation is weighted to obtain at least one weighted result of the preset operation.

[0031] The results of at least one weight allocation are fused to obtain the initial weights for the preset operation.

[0032] In one embodiment, the acquisition module is further configured to acquire training samples, which include object information of multiple objects, content information of multiple media content, and the actual matching value between each object and each media content.

[0033] The matching module is used to input training samples into the initial matching model to obtain the predicted matching value between each object and each media content.

[0034] The matching module is also used to iteratively train the initial matching model based on the first loss of the actual matching value and the predicted matching value until the first loss converges to obtain the matching model.

[0035] In one embodiment, the probability prediction module is further configured to input training samples into the initial probability prediction model to obtain the initial probability predicted by the training samples.

[0036] The first prediction module is used to obtain the target probability of each object performing a preset operation on each media content based on the initial prediction probability and the prediction matching value.

[0037] The first prediction module is also used to iteratively train the initial matching model and the initial probability prediction model based on the second loss between the predicted target probability and the actual target probability, until the second loss converges to obtain the matching model and the probability prediction model.

[0038] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in embodiments of this application.

[0039] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in embodiments of this application.

[0040] Fifthly, embodiments of this application provide a computer program product including instructions that, when executed, cause the method described in embodiments of this application to be performed.

[0041] The media content push method, apparatus, device, and medium proposed in this application determine the target probability of an object performing a preset operation on media content by combining the degree of matching between the object and the media content with the initial probability of the object performing a preset operation on the media content. This addresses the problem in existing technologies where the sparsity of relevant information in media content leads to low accuracy in probability prediction, thus affecting the evaluation results of interest level. Specifically, the object information of the target object and the content information of the candidate media content are input into a matching model and a probability prediction model, respectively. The matching model obtains a matching value representing the degree of matching between the target object and the candidate media content, and the probability prediction model obtains the initial probability of the target object performing a preset operation, where the preset operation is the operation by which the target object expresses interest in the candidate media content. Furthermore, the initial probability and the matching value are fused to obtain the fused probability of the preset operation, and attention allocation processing is performed on the initial probability of the preset operation based on the fused probability of the preset operation to obtain the target probability of the preset operation. In this way, by incorporating the matching value between the target object and the candidate media content into the prediction process of the final probability of the object performing related operations on the media content, the diversity of parameters is improved. This effectively solves the problem of low accuracy in probability prediction caused by the sparsity of relevant information in the media content, thereby improving the accuracy of the evaluation results of interest level.

[0042] Furthermore, considering that existing technologies rely solely on the probability of related operations to evaluate an object's interest in media content, the evaluation elements are too simplistic and may result in a significant discrepancy between the actual interest level and the actual content being pushed, leading to unsatisfactory media content delivery. Therefore, this application combines target probabilities and matching values ​​through a fusion process to comprehensively determine the object's interest score in candidate media content. The interest score obtained through this method more accurately represents the target object's actual interest in the candidate media content. Based on this, pushing candidate media content to the target object according to the interest score can filter out media content that better matches the object's interests, improve the accuracy of personalized push notifications, and effectively utilize the push resources of the network platform, reducing the overhead and waste of push resources.

[0043] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0044] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0045] Figure 1 A schematic diagram of the structure of a media content push system provided in an embodiment of this application;

[0046] Figure 2 A flowchart illustrating a method for pushing media content provided in an embodiment of this application;

[0047] Figure 3 This is a schematic diagram of the structure of a target probability prediction model provided in an embodiment of this application;

[0048] Figure 4 This is a schematic diagram of the structure of a matching model provided in an embodiment of this application;

[0049] Figure 5 This is a schematic diagram of the structure of a probability prediction model provided in an embodiment of this application;

[0050] Figure 6 A schematic diagram of another probability prediction model provided in the embodiments of this application;

[0051] Figure 7 This is a schematic diagram illustrating the acquisition of fusion probability in an embodiment of this application.

[0052] Figure 8 A schematic diagram illustrating an initial probability acquisition process provided in an embodiment of this application;

[0053] Figure 9 A schematic diagram of another probability prediction model provided in the embodiments of this application;

[0054] Figure 10 A schematic diagram illustrating the training of a matching model provided in an embodiment of this application;

[0055] Figure 11 A schematic diagram illustrating the training of a matching model and a probability prediction model provided in an embodiment of this application;

[0056] Figure 12 A training diagram illustrating another matching model and probability prediction model provided in an embodiment of this application;

[0057] Figure 13 A schematic diagram of the structure of a media content push device provided in an embodiment of this application;

[0058] Figure 14 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0059] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0060] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0061] The following explains the terms used in the embodiments of this application:

[0062] 1. Exposure of media content

[0063] Exposure to media content refers to the act of viewing media content.

[0064] 2. Clicking on media content

[0065] Clicking on media content refers to the action of an object clicking on the exposed media content.

[0066] 3. Superficial conversion operations of media content

[0067] Shallow conversion operations of media content refer to the process where, after an object clicks on an item or link contained in the media content, it is redirected to the corresponding page and performs a corresponding operation. These operations include: application download, application activation, and form registration, etc.

[0068] 4. Deep transformation of media content

[0069] Deep conversion operations of media content refer to actions performed after an audience clicks on items or links within the media content, navigates to a corresponding page, and performs actions such as making payments, downloading applications, and ensuring application retention the following day. Deep conversion operations are follow-up actions to shallow conversion operations of the same media content.

[0070] To better deliver media content to target audiences, it's possible to anticipate their level of interest in the content. Current technology involves training a prediction model by collecting information about the target audience, the media content, and related actions as training data. This model then predicts the probability (e.g., click-through rate) of the target audience's actions related to the media content. Based on this probability, the model assesses the target audience's interest in the media content and delivers the appropriate content. However, the diversity and complexity of media content lead to sparsity issues in this information, resulting in low accuracy in probability prediction. This makes probability-based content delivery prone to ineffective delivery, wasting network platform resources.

[0071] Based on this, embodiments of this application provide a method, apparatus, device, and medium for pushing media content. This method integrates the matching value between an object and candidate media content into the prediction process of the final probability of an object performing a related operation on the media content. This enriches the training and prediction data of the model, effectively alleviating the sparsity problem of relevant information in the media content and improving the accuracy of predicting the final probability of preset operations. Furthermore, based on the final probability of the related operation, the matching value between the object and candidate media content is combined to comprehensively determine the object's interest score for the related operation. Thus, pushing candidate media content to the target object based on the interest score can filter out media content that better matches the object's interests, improving the accuracy of personalized push and effectively utilizing the push resources of the network platform, reducing the overhead and waste of push resources.

[0072] Figure 1 This is a schematic diagram illustrating the structure of a media content push system provided in an embodiment of this application. The media content push method provided in this embodiment can be applied to the media content push system 100. (Reference) Figure 1 The media content delivery system 100 includes one or more user devices 101 and a server 102.

[0073] The user equipment 101 may be, but is not limited to, devices such as personal computers, platform computers, smartphones, and vehicle terminals; this embodiment does not limit this. It should be noted that, although... Figure 1 Only one user equipment 101 is depicted, but those skilled in the art will understand that this application can support any number of user equipments.

[0074] Server 102 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0075] In one implementation, both user equipment 101 and server 102 can be used independently to execute the media content push method provided in the embodiments of this application.

[0076] For example, server 102 or user device 101 acquires training samples and trains the initial matching model and initial probability prediction model based on the acquired training samples to obtain the matching model and probability prediction model. Then, server 102 or user device 101 acquires object information of the target object and content information of the candidate media content, and determines the target probability of the target object performing a preset operation on the candidate media content through the matching model and probability prediction model; furthermore, based on the target probability and matching value, it determines the target object's interest score on the candidate media content, and based on the comparison result of the interest score and a preset threshold, it determines whether to push the candidate media content to the target object.

[0077] In this application embodiment, obtaining the object information of the target object requires obtaining the consent or permission of the target object, and also needs to comply with laws, regulations and relevant rules and standards.

[0078] The following will combine Figure 1 The technical solutions of this application and how they solve the aforementioned technical problems will be described in detail with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0079] like Figure 2 As shown in the embodiment of this application, a method for pushing media content is provided, which can be applied to... Figure 1 The method specifically includes the following steps: (shown in the user equipment 101 and / or server 102)

[0080] 201. Obtain the object information of the target object and the content information of the candidate media content, input the object information and content information into the matching model, and obtain the matching value between the target object and the candidate media content; the matching value is used to characterize the degree of matching between the target object and the candidate media content.

[0081] Optionally, object information may include, but is not limited to, basic attribute information and behavioral interest information. Basic attribute information describes the basic situation and essential attributes of the target object; for example, describing the object's gender, age, etc. Behavioral interest information may include, but is not limited to, one or more of the following: consumption level and object context information. For any object to be pushed to (i.e., the target object), the object context information may include, but is not limited to, the time context information, location context information, and mood context information of the object's access to historical media content, which is crucial for improving the accuracy of push notifications. The time context information of the object's access to historical media content can reflect the object's recent and long-term behavior, thereby reflecting changes in the object's recent interests, ensuring the continuity of object interest prediction during the media content evaluation process. Content information may include, but is not limited to, the media content's identifier, the media content's main identifier, the media content's category, the media content's textual semantic information, the media content's image information, and the contextual features of the object's request.

[0082] For example, in the system described above, the server can provide an internet platform for objects, allowing objects to browse various media content provided by the internet platform through user devices. Here, the media content can be various forms of media content provided by the server through the internet platform, such as advertisements; advertisements can be product advertisements or service advertisements, etc.

[0083] In one possible implementation, after obtaining the object information of the target object and the content information of the candidate media content, feature transformation processing is performed on the object information and content information to obtain object features and content features. Then, the object features and content features are input into the matching model to obtain the matching value between the target object and the candidate media content.

[0084] Optionally, feature transformation processing of object information and content information may specifically include: encoding object information and content information to obtain object features and content features. The encoding process may be one-hot encoding.

[0085] Furthermore, since the object features and content features obtained through One-Hot encoding may contain high-dimensional features, especially the identifier information in object information and content information, the feature dimension after One-Hot encoding is very high, and high-dimensional features are not convenient for subsequent processing; therefore, high-dimensional features can be converted into low-dimensional features after One-Hot encoding.

[0086] Optionally, the encoding process for object information and content information may specifically include: encoding the object information and content information using one-hot encoding, and obtaining initial object features and initial content features based on the encoding results; performing feature mapping processing on the initial object features and initial content features to compress the feature dimensions and obtain object features and content features.

[0087] In one possible implementation, feature mapping can be performed on the initial features of the object and the initial features of the content based on the embedding layer, converting both into low-dimensional features. Specifically, taking the initial features of the content as an example, the initial features of the content are input into a hash function to obtain the output of the hash function. This output is used as a key to look up the value corresponding to the key in the embedding table, and the low-dimensional feature corresponding to this value is output. This low-dimensional feature is then used as the content feature. This completes the feature mapping process for the initial features of the content. The feature mapping process for the initial features of the object can refer to the feature mapping process for the initial features of the content described above, and will not be repeated here.

[0088] It's important to note that during the training of the matching model, the low-dimensional features output from the model's embedding layer are stored as values ​​in a lookup table. This facilitates the subsequent direct conversion of high-dimensional features into low-dimensional features. Specifically, before training the matching model, all stored features in this lookup table can be initialized. During model training, these features are updated based on the backpropagation gradient. The more times the model is trained, the higher the accuracy of converting high-dimensional features into low-dimensional features.

[0089] Understandably, a matching model is a neural network model used to determine the degree of matching between an object and candidate media content. After inputting object features and content features into the matching model, it performs non-linear cross-processing on different dimensions of the object and content features respectively. This allows it to obtain more information on non-linear and combined features from both the object and content features, resulting in a first transformed feature corresponding to the object features and a second transformed feature corresponding to the content features. Then, based on the first and second transformed features, the matching value between the object and the candidate media content is determined.

[0090] 202. Input the object information and content information into the probability prediction model to obtain the initial probability of the target object performing the preset operation.

[0091] The preset operation refers to the action taken by the target audience to express their interest in the candidate media content. Further, the preset operation includes at least one of a click operation and a conversion operation; wherein the click operation and the conversion operation represent different levels of interest in the candidate media content by the target audience. The conversion operation is a further action taken on the media content after the click operation; therefore, it can be understood that the level of interest represented by the conversion operation is higher than that represented by the click operation. Specifically, the conversion operation includes at least one of a shallow conversion operation and a deep conversion operation.

[0092] It should be noted that the specific parameters included in the object information and content information input into the matching model can be the same as or different from the specific parameters included in the object information and content information input into the probability prediction model. For example, the object information input into the matching model may include the parameter of age; however, the object information input into the probability prediction model may not include the parameter of age. The specifics can be adjusted according to the actual situation, and this application embodiment does not impose any limitations on this.

[0093] In another possible implementation, after obtaining the object information of the target object and the content information of the candidate media content, feature transformation processing is performed on the object information and content information to obtain object features and content features. Then, the object features and content features are input into the probability prediction model to obtain the initial probability of the target object performing the preset operation.

[0094] In one implementation, the preset operation may include at least one of a first operation, a second operation, and a third operation. The first operation occurs before the second operation, and the second operation occurs based on the occurrence of the first operation; the second operation occurs before the third operation, and the third operation occurs based on the occurrence of the second operation. Therefore, the first operation is a preceding operation to the second operation, and the second operation is a preceding operation to the third operation; that is, the second operation is a following operation to the first operation, and the third operation is a subsequent operation to the second operation.

[0095] For example, taking media content as advertising, the content of the advertisement might be pushing a specific application (AAP). Assume the first action is clicking the advertisement, the second action is downloading the app pushed in the advertisement, and the third action is paying for the app. That is, in the interaction between the user and the advertisement, the user needs to click the advertisement (first action) before downloading (second action) the corresponding app in the advertisement, and then pay for the app (third action). Therefore, the first action is the initial action (such as clicking), the second action can be understood as a shallow conversion action, and the third action can be understood as a deep conversion action.

[0096] Specifically, the second action can be the action taken by other users immediately after clicking into the ad content, such as activation, login, or registration, which can be called a shallow conversion action; the third action can be the action that other users may take after performing the second action, such as payment or next-day retention, which can be called a deep conversion action. The specific form of the action depends on the actual scenario and is not limited here.

[0097] Based on the above, the initial probability of the preset operation may include at least one of the initial probability of the first operation, the initial probability of the second operation, and the initial probability of the third operation.

[0098] 203. The initial probability and the matching value are fused to obtain the fused probability of the preset operation. Based on the fused probability of the preset operation, attention is allocated to the initial probability of the preset operation to obtain the target probability of the preset operation.

[0099] The process of fusing the initial probability and the matching value to obtain the fusion probability of the preset operation can be achieved by adding the data of the same dimension in the matching value and the initial probability of the preset operation, or by concatenating the matching value and the initial probability of the preset operation.

[0100] In practical applications, since the preset operation includes at least one of the first, second, and third operations, when the preset operation includes one of them, taking the first operation as an example, it can specifically be: determining the target probability of the target object performing the first operation on the candidate media content based on the initial probability of the first operation and the matching value. When the preset operation includes two of the operations, taking the preset operation including the first and second operations as an example, determining the target probability of the target object performing the preset operation on the candidate media content based on the initial probability of the first operation and the matching value specifically includes: determining the target probability of the target object performing the first operation on the candidate media content based on the initial probability of the first operation and the matching value; and determining the target probability of the target object performing the second operation on the candidate media content based on the initial probability of the second operation and the matching value.

[0101] As an example, refer to Figure 3 Taking a preset operation that includes a first operation and a second operation as an example, the initial probability of the first operation is fused with the matching value to obtain the fused probability of the first operation; the initial probability of the second operation is fused with the matching value to obtain the fused probability of the second operation; attention allocation processing is performed on the initial probability of the first operation based on the fused probability of the first operation and the fused probability of the second operation to obtain the target probability of the first operation; and attention allocation processing is performed on the initial probability of the first operation based on the fused probability of the first operation and the fused probability of the second operation to obtain the target probability of the second operation.

[0102] In one implementation, an attention layer is used to assign attention to the initial probability of each preset operation based on the fusion probability of all preset operations, resulting in a target probability for each preset operation. This target probability is essentially used to assign an attention weight to the initial probability of each preset operation. This attention weight for each preset operation reflects the importance of its initial probability among all preset operations.

[0103] In a preferred embodiment, the attention weight corresponding to the initial probability of the t-th preset operation is calculated according to the following formula:

[0104]

[0105] Where, α t The logit represents the attention weight corresponding to the initial probability of the t-th preset operation. t Represents the fusion probability of the t-th preset operation; logit p The fusion probability of the second operation is indicated; tanh represents the tangent function. W t b represents the model parameters in the attention layer; z represents the state vector in the attention layer.

[0106] Furthermore, assuming the preset operations include a first operation, a second operation, and a third operation, the attention weights corresponding to the initial probabilities of each operation can be determined according to the following formula:

[0107] α t =z t tanh(W1*Logit1+W2*Logit2+W3*Logit3+b)

[0108] Where Logit1 represents the fusion probability of the first operation; Logit2 represents the fusion probability of the second operation; and Logit3 represents the fusion probability of the third operation.

[0109] Then, for the weight α t Normalization:

[0110]

[0111] Finally, the target probability of the t-th preset operation is logit t This represents the initial probability of each preset operation at the t-th position.

[0112] It is understandable that, based on the above formula, when there is only one preset operation, the weight assigned to the initial probability of the preset operation in step 203 is 1. Therefore, for the case where there is only one preset operation, step 203 can be skipped.

[0113] 204. The target probability and matching value are fused to determine the target object's interest score in the candidate media content. If the interest score is greater than a preset threshold, the candidate media content is pushed to the target object.

[0114] In a preferred embodiment, when the matching value includes a matching score and a matching probability, the target probability and the matching value are fused to determine the target object's interest score in the candidate media content. This process includes the following two implementation methods:

[0115] Method 1: Determine the conversion score of the target object to the candidate media content based on the target probability, and then sum the conversion score and the matching score to obtain the target object's interest score to the candidate media content.

[0116] Method 2: Weight the target probability and the matching probability to obtain the interest probability, and determine the interest score based on the interest probability.

[0117] Specifically, determining the conversion score of the target object for candidate media content based on the target probability includes: the conversion score can be the product of the bidding coefficient of the preset operation corresponding to the media content and the target probability.

[0118] In practical applications, since different types of media content can correspond to different calculation methods, the process of determining the target audience's interest score in candidate media content can be based on calling the corresponding calculation method for the candidate media content; then, the target probability of the preset operation can be calculated according to the calculation method to obtain the conversion score.

[0119] For example, for regular advertising, it is generally only necessary to consider the target probability of the first action (such as the click-through rate for a click action), the target probability of the second action (such as the shallow conversion rate for a shallow conversion action), and the bid coefficient for the second action. Therefore, the conversion score can be determined according to the following formula:

[0120] conv score =ecvr1*ecvr2*bid1

[0121] Where, conv score represents the conversion score; ecvr1 represents the target probability of the first operation; ecvr2 represents the target probability of the second operation; bid1 represents the bid coefficient of the second operation.

[0122] For important ads, it may be necessary to consider the target probability of the first action (e.g., click-through rate for a click), the target probability of the second action (e.g., shallow conversion rate for a shallow conversion), the target probability of the third action (e.g., deep conversion rate for a deep conversion), and the overall bid coefficient. Therefore, the conversion score can be determined using the following formula:

[0123] conv score =ecvr1*ecvr2*ecvr3*bid

[0124] Where ecvr3 represents the target probability of the third operation; bid represents the overall bid coefficient.

[0125] For example, the overall bid coefficient can be a weighted value of at least two of the bid coefficients of the first operation, the second operation, and the third operation, or it can be the value of one of the bid coefficients of the three operations.

[0126] For multi-target ads, it may be necessary to consider the choice of bidding coefficient. For example, if the number of targets indicated by the content push target is greater than the preset value, the target probability of the second operation and the target probability of the third operation should be compared to determine the corresponding calculation method.

[0127] In one possible scenario, for multi-target advertising, when ecvr1 > ecvr2,

[0128] conv score =ecvr1*ecvr2*bid1

[0129] otherwise:

[0130] conv score =ecvr1*ecvr3*bid2

[0131] Where ecvr3 represents the target probability of the third operation; bid2 represents the bid coefficient of the third operation.

[0132] In one possible implementation, to address the situation where multiple candidate media contents may correspond to the same object, the media content to be pushed to the target object can be selected from the multiple candidate media contents based on the interest score. Specifically, the multiple candidate media contents are sorted according to their interest scores, and the candidate media contents with interest scores greater than a preset threshold are selected as the media content to be pushed to the target object. Then, they are pushed in the order of sorting.

[0133] In another possible scenario, considering that the media content pushed to the same target audience may be similar, a sorting strategy can be implemented. First, candidate media content with an interest score greater than a preset threshold is sorted to determine a first sequence. Then, historical push information corresponding to the target audience is obtained, and a sorting strategy is determined based on this information. The first sequence is then adjusted according to the push rules to obtain a second sequence. The target media content is then determined from the second sequence. For example, based on the candidate media content output during the sorting stage, the media content is re-sorted according to its interest score, combined with sorting strategies such as diversity, frequency control, category control, and specific result weighting. The media content with the highest score is then obtained and pushed to the target audience.

[0134] It should be noted that the preset threshold can be set according to the actual situation, and this application embodiment does not limit it.

[0135] The media content push method proposed in this application addresses the problem in existing technologies where the sparsity of information in media content leads to low accuracy in probability prediction, thus affecting the evaluation of interest levels. Specifically, the object information of the target object and the content information of candidate media content are input into a matching model and a probability prediction model, respectively. The matching model obtains a matching value characterizing the degree of matching between the target object and the candidate media content, while the probability prediction model obtains the initial probability of the target object performing a preset operation, where the preset operation is the action by which the target object expresses interest in the candidate media content. Furthermore, the initial probability and the matching value are fused to obtain a fused probability of the preset operation, and attention allocation is applied to the initial probability of the preset operation based on this fused probability to obtain the target probability of the preset operation. In this way, by incorporating the matching value between the target object and the candidate media content into the prediction process of the final probability of the object performing related operations on the media content, the diversity of parameters is improved. This effectively solves the problem of low accuracy in probability prediction caused by the sparsity of relevant information in the media content, thereby improving the accuracy of the evaluation results of interest level.

[0136] Furthermore, considering that existing technologies rely solely on the probability of related operations to evaluate an object's interest in media content, the evaluation elements are too simplistic and may result in a significant discrepancy between the actual interest level and the actual content being pushed, leading to unsatisfactory media content delivery. Therefore, this application combines target probabilities and matching values ​​through a fusion process to comprehensively determine the object's interest score in candidate media content. The interest score obtained through this method more accurately represents the target object's actual interest in the candidate media content. Based on this, pushing candidate media content to the target object according to the interest score can filter out media content that better matches the object's interests, improve the accuracy of personalized push notifications, and effectively utilize the push resources of the network platform, reducing the overhead and waste of push resources.

[0137] In one embodiment, to combine the matching value with the initial probability and with the target probability, the matching model can specifically determine the similarity between the target object and the candidate media content to determine the matching score in the matching value, and generate the matching probability based on the matching score. In one implementation, the matching value includes a matching score and a matching probability. Inputting object information and content information into the matching model to obtain the matching value between the target object and the candidate media content includes: inputting object information and content information into the matching model; performing feature transformation processing on the object information and content information respectively to obtain object features and content features; performing feature compression processing on the object features to obtain a first transformed feature, and performing feature compression processing on the content features to obtain a second transformed feature; determining the similarity between the first transformed feature and the second transformed feature, and determining the matching score based on the similarity; and performing feature compression and normalization processing on the matching score to obtain the matching probability.

[0138] Optional, refer to Figure 4 In the matching model, one-hot encoding is used to encode object information and content information to obtain initial object features and initial content features. Feature mapping (i.e., embedding) is then performed on the initial object features and initial content features to obtain object features and content features.

[0139] Optional, refer to Figure 4 The first transformation feature can be performed on the object features based on the first MLP network in the multilayer perceptron (MLP) layer of the matching model; the second transformation feature can be performed on the content features based on the second MLP network in the MLP layer of the matching model.

[0140] It should be noted that the first transformation feature compression process for object features specifically involves performing nonlinear cross-processing on different dimensions of the object features to obtain more information on the nonlinear and combined features of the object features, thereby achieving feature compression and obtaining the first transformation feature corresponding to the object features. The compression process for content features is similar and will not be elaborated here. Furthermore, the feature compression process for both object and content features can also be understood as feature extraction from the object and content features to obtain the first and second transformation features.

[0141] In practical applications, the similarity between the first transformed feature and the second transformed feature can be determined based on algorithms such as Euclidean distance, cosine similarity, and cosine distance.

[0142] In one preferred option, refer to Figure 4 The first transformed feature and the second transformed feature are subjected to a dot product operation to determine the cosine similarity. Then, the similarity between the first transformed feature and the second transformed feature is determined based on the cosine similarity, and a matching score is determined based on the similarity. Here, the similarity can be directly determined as the matching score, or the product of the similarity and a preset value can be determined as the matching score; this embodiment of the application does not limit this approach.

[0143] Furthermore, after obtaining the matching score, it can be done through... Figure 4 The third MLP network shown performs feature compression and normalization on the matching scores to obtain the matching probability.

[0144] In one possible implementation, to ensure data consistency when fusing the initial probability and the matching value, the initial probability can be fused based on the matching probability in the matching value, and the target probability of each preset operation can be determined based on the fused probabilities of all preset operations. Therefore, fusing the initial probability and the matching value to obtain the fused probability of the preset operation includes: fusing the matching probability and the initial probability of the preset operation to obtain the fused probability of the preset operation. In this way, by fusing the matching probability and the initial probability of the preset operation, not only is the data consistency ensured during the fusion process, but the addition of the matching probability also increases the parameter basis for predicting the target probability, improving parameter diversity and thus ensuring the accuracy of the target probability prediction for the preset operation.

[0145] Furthermore, the matching probability and the initial probability of the preset operation can be fused by adding the data of the same dimension from the matching probability and the initial probability of the preset operation to obtain the fused probability of the preset operation. In this case, the matching probability, the initial probability of the preset operation, and the fused probability of the preset operation have the same dimension.

[0146] For example, if the matching probability is (al, b1) and the initial probability of the preset operation is (a2, b2), then the fusion probability of the preset operation can be (a1+a2, b1+b2).

[0147] Furthermore, the above-described fusion method requires that the dimensions of the matching probability and the initial probability of the preset operation be the same. When the dimensions of the matching probability and the initial probability of the preset operation are different, one probability dimension can be used as the standard dimension, and the other probability dimension can be transformed to become the standard dimension. For example, the dimension of the matching probability can be used as the standard dimension, and the dimension of the initial probability of the preset operation can be transformed to make the transformed dimension of the initial probability of the preset operation the same as the dimension of the matching probability. Alternatively, a standard dimension can be set, and the dimension of the matching probability can be transformed to become the standard dimension; and the dimension of the initial probability of the preset operation can also be transformed to become the standard dimension.

[0148] In another implementation, the matching probability and the initial probability of the preset operation are concatenated to obtain the fusion probability of the preset operation.

[0149] For example, if the matching probability is (a1, b1) and the initial probability of the preset operation is (a2, b2), then the fusion probability of the preset operation can be (a1, b1, a2, b2). Therefore, the dimension of the fusion probability of the preset operation is the sum of the dimension of the matching probability and the dimension of the initial probability of the preset operation.

[0150] In this embodiment, object information and content information are input into the matching model. Feature transformation processing is performed on the object information and content information respectively to obtain object features and content features. Feature compression processing is then applied to the object features to obtain first transformed features, and feature compression processing is applied to the content features to obtain second transformed features. The similarity between the first and second transformed features is determined, and a matching score is determined based on the similarity. Since the matching value needs to be combined with the initial probability and the target probability respectively, feature compression and normalization processing is performed on the matching score to obtain the matching probability. In this way, by adjusting the data format contained in the matching value, it is convenient to combine the subsequent matching value with the initial probability and with the target probability.

[0151] In one embodiment, the preset operations include a first operation, a second operation, and a third operation; the first operation is a preceding operation to the second operation; the second operation is a preceding operation to the third operation. The degree of interest represented by the first operation, the second operation, and the third operation ranges from low to high.

[0152] The first operation can be a click, while the second and third operations are conversion operations. The second operation can be a shallow conversion operation, and the third operation can be a deep conversion operation.

[0153] Furthermore, considering the sparsity problem of relevant information in media content, the initial probabilities of preceding operations can be incorporated into the prediction of the initial probabilities of subsequent operations. Therefore, referring to... Figure 5 The object information and content information are input into the probability prediction model to obtain the initial probability of the target object performing a preset operation, including: inputting the object information and content information into the probability prediction model to obtain the initial probability of the first operation, the source probability of the second operation, and the source probability of the third operation; obtaining the initial probability of the second operation based on the initial probability of the first operation and the source probability of the second operation; and obtaining the initial probability of the third operation based on the initial probability of the second operation and the source probability of the third operation.

[0154] The initial probability of the first operation, the source probability of the second operation, and the source probability of the third operation are obtained based on the same data processing order but different model parameters.

[0155] Specifically, refer to Figure 6 The object information and content information are input into the probability prediction model. Feature transformation is performed on the object information and content information respectively to obtain object features and content features. The object features and content features are then concatenated to obtain a first concatenated feature. This first concatenated feature is then input into at least one gating network to determine the initial weights of the first, second, and third operations. The initial weights of the first operations are then fused with the first concatenated feature to obtain a first fused feature; the initial weights of the second operations are fused with the first concatenated feature to obtain a second fused feature; and the initial weights of the third operations are fused with the first concatenated feature to obtain a third fused feature. Subsequently, the first fused feature is input into a fourth MLP to obtain the initial probability of the first operation; the second fused feature is input into a fifth MLP to obtain the source probability of the second operation; and the third fused feature is input into a sixth MLP to obtain the source probability of the third operation. Feature fusion here can be achieved by multiplying the initial weights of each operation with the first concatenated feature.

[0156] In this embodiment, by using parameter passing between preset operations on media content to input the probability of the preceding action into the prediction of the subsequent action, the inaccuracy of the prediction caused by the data sparsity of the subsequent operation is avoided, the accuracy of the initial probability is improved, and the interest score obtained later is more consistent with the actual interest of the object in the candidate media content, so as to improve the utilization rate of the network platform's push resources and reduce the overhead and waste of push resources.

[0157] In one embodiment, obtaining the initial probability of the second operation based on the initial probability of the first operation and the source probability of the second operation includes: fusing the initial probability of the first operation and the source probability of the second operation to obtain the initial probability of the second operation.

[0158] Specifically, refer to Figure 6 The initial probability of the first operation and the source probability of the second operation are concatenated in the first connection layer to obtain the first connection probability. The first connection probability and the initial probability of the first operation are then input into the seventh MLP to obtain the initial probability of the second operation.

[0159] In one implementation, the initial probability of the first operation and the source probability of the second operation are fused. This can be achieved by adding the data of the same dimension from the initial probability of the first operation and the source probability of the second operation to obtain the first connection probability. The initial probability of the first operation, the source probability of the second operation, and the first connection probability all have the same dimension.

[0160] For example, if the initial probability of the first operation is (cl, d1) and the source probability of the second operation is (c2, d2), then the first connection probability can be (c1+c2, d1+d2).

[0161] In another implementation, the initial probability of the first operation and the source probability of the second operation are concatenated to obtain the first connection probability.

[0162] For example, if the initial probability of the first operation is (c1, d1) and the initial probability of the preset operation is (c2, d2), then the fusion probability of the preset operation can be (c1, d1, c2, d2). Therefore, the dimension of the first connection probability is the sum of the dimension of the initial probability of the first operation and the dimension of the source probability of the second operation.

[0163] In this embodiment, since the initial probability of the first operation is input into the prediction of the initial probability of the second operation by means of parameter passing between the first operation and the second operation, the inaccuracy of the prediction of the second operation due to data sparsity is avoided, and the accuracy of the initial probability of the second operation is improved.

[0164] In one embodiment, obtaining the initial probability of the third operation based on the initial probability of the second operation and the source probability of the third operation includes: fusing the initial probability of the second operation and the source probability of the third operation to obtain the initial probability of the third operation.

[0165] Specifically, refer to Figure 6The initial probability of the second operation and the source probability of the third operation are concatenated in the second connection layer to obtain the second connection probability. The second connection probability and the initial probability of the second operation are then input into the eighth MLP to obtain the initial probability of the third operation.

[0166] It should be noted that the embodiment of fusing the initial probability of the second operation and the source probability of the third operation to obtain the initial probability of the third operation can refer to the relevant embodiment of fusing the initial probability of the first operation and the source probability of the second operation described above, and will not be repeated here.

[0167] In this embodiment, since the initial probability of the second operation is input into the prediction of the initial probability of the third operation by means of parameter passing between the second operation and the third operation, the inaccuracy of the prediction caused by the data sparsity of the second operation is avoided, and the accuracy of the initial probability of the second operation is improved.

[0168] In one embodiment, reference is made to Figure 7 The initial probability and the matching value are fused to obtain the fused probability of the preset operation, including: fusing the matching probability with the initial probability of the first operation to obtain the fused probability of the first operation; fusing the matching probability with the initial probability of the second operation to obtain the fused probability of the second operation; and fusing the matching probability with the initial probability of the third operation to obtain the fused probability of the third operation.

[0169] Specifically, the matching probability is concatenated with the initial probability of the first operation to obtain the fusion probability of the first operation; the matching probability is concatenated with the initial probability of the second operation to obtain the fusion probability of the second operation; and the matching probability is concatenated with the initial probability of the third operation to obtain the fusion probability of the third operation.

[0170] The above is just one way to fuse the matching probability with the initial probability of each operation. Other fusion methods are also possible. For example, the matching probability can be multiplied or added to the initial probability of each operation, and the result can be used as the fused probability of each operation.

[0171] In another possible implementation, if the preset operation includes one or two of the first, second and third operations, the target probability of the target object performing the preset operation on the candidate media content can be determined in the same way as described above to determine the fusion probability of each operation.

[0172] In this embodiment, matching data of object information and content information are incorporated into the initial probability of the first operation, the initial probability of the second operation, and the initial probability of the third operation. This increases the parameter features in the prediction process of the fusion probability of each operation, thereby improving the prediction accuracy of the target probability of each operation.

[0173] In one embodiment, reference is made to Figure 8 By inputting object information and content information into the probability prediction model, the initial probability of the target object performing a preset operation is obtained, including:

[0174] 801. Input the object information and content information into the probability prediction model, and perform feature transformation processing on the object information and content information respectively to obtain object features and content features.

[0175] 802. Perform a splicing operation on the object features and content features to obtain the first spliced ​​feature.

[0176] 803. Determine the initial weight of the preset operation based on the first splicing feature.

[0177] 804. The initial weights of the preset operation are fused with the first concatenation feature to obtain the fused feature of the preset operation.

[0178] 805. Based on the fusion characteristics of the preset operation, obtain the initial probability of the preset operation.

[0179] It should be noted that the object information and content information are processed separately for feature transformation in the probability prediction model to obtain object features and content features. For the relevant embodiments of the feature transformation processing of object information and content information in the matching model in step 201 above, it will not be repeated here.

[0180] In one implementation, the first splicing feature is input into at least one gating network to determine the initial weight of the preset operation, and the initial weight of the preset operation is fused with the first splicing feature to obtain the fused feature of the preset operation.

[0181] Furthermore, for each gating network, the first concatenated feature is input into the gating network, and feature extraction is performed on the first concatenated feature to obtain the first feature. The first features of the preset operations in each gating network are then fused to obtain the second feature, which is then normalized to obtain the initial weights of the preset operations. It should be noted that different gating networks extract different features from the first concatenated feature. Each preset operation corresponds to a gating network with the same structure but different gating parameters. That is, different gating networks correspond to different gating parameters, and based on these different gating parameters, different first features are determined for the preset operations.

[0182] In a preferred embodiment, the first features of the preset operations in each gating network are added together to obtain the second feature.

[0183] Reference Figure 9 Taking a preset operation including a first operation, a second operation, and a third operation as an example, the first concatenation features are input into gating networks 1, 2, and 3, respectively. The first features determined in gating network 1 are (p11, p12, p13); the first features determined in gating network 2 are (p21, p22, p23); and the first features determined in gating network 3 are (p31, p32, p33). Assuming that the initial weight of the first operation is obtained based on the first features of gating networks 1 and 2, then the second feature of the first operation, p10, is p11 + p21; then, the initial weight of the first operation is normalized to obtain the initial weight of the first operation. The initial weight of the second operation is obtained based on the first features of gating networks 1 and 2, then the second feature of the first operation, p20, is p12 + p22; then, the initial weight of the second operation is normalized to obtain the initial weight of the second operation. The initial weights of the third operation are obtained based on the first features of gated network 1, gated network 2 and gated network 3. Then, the second feature of the third operation is p30 = p13 + p23 + p33. After that, the initial weights of the third operation are normalized to obtain the initial weights of the third operation.

[0184] Specifically, gating network 1 extracts features from the first splicing features according to the first gating parameter to obtain the first features of each operation; gating network 2 extracts features from the first splicing features according to the second gating parameter to obtain the first features of each operation; and gating network 3 extracts features from the first splicing features according to the third gating parameter to obtain the first features of each operation.

[0185] Optionally, the initial weight of the preset operation can be multiplied with the first concatenation feature to obtain the fusion feature of the preset operation.

[0186] In this embodiment, by concatenating object features and content features, a first concatenated feature is obtained that associates the two features. The initial weight of the preset operation can be determined based on the first concatenated feature. The initial weight is then fused with the first concatenated feature to obtain the fused feature of the preset operation. This allows data in the first concatenated feature that is not related to the initial probability prediction of the preset operation to be filtered out, so that the obtained initial probability is more consistent with the actual situation of the target object performing the preset operation on the candidate media content.

[0187] In one embodiment, determining the initial weight of a preset operation based on a first splicing feature includes: performing weight allocation processing on the preset operation based on the first splicing feature to obtain at least one weight allocation result of the preset operation; and performing fusion processing on the at least one weight allocation result to obtain the initial weight of the preset operation.

[0188] Specifically, the first splicing feature is input into at least one gating network to perform weight allocation processing on the preset operation, thereby obtaining at least one weight allocation result for the preset operation. In practical applications, the number of gating networks can be adjusted according to requirements; this application embodiment does not impose any limitation on the specific number of gating networks.

[0189] Furthermore, the gating network controls how much of the initial concatenated features of the input should be retained and how much should be discarded. In this embodiment, for each gating network, a corresponding weight can be assigned to each preset operation to obtain the weight allocation result for each preset operation. The weight allocation results for each preset operation are then weighted to obtain the initial weight for each preset operation.

[0190] Optionally, the process of fusing at least one weight allocation result to obtain the initial weight of the preset operation specifically includes: adding at least one weight allocation result together and normalizing the result of the addition to obtain the initial weight of the preset operation.

[0191] In this embodiment, the weight allocation process can capture important information about the preset operation from the first splicing feature, and combine the different weight allocation results for the preset operation to determine the initial weight of the preset operation, thereby improving the accuracy of the initial probability prediction of the preset operation.

[0192] In one embodiment, a training process for a matching model is also provided: obtaining training samples, which include object information of multiple objects, content information of multiple media content, and the actual matching value between each object and each media content; inputting the training samples into an initial matching model to obtain the predicted matching value between each object and each media content; iteratively training the initial matching model based on a first loss of the actual matching value and the predicted matching value until the first loss converges to obtain the matching model.

[0193] In one implementation, when the true matching value includes the true matching probability, the training samples are input into the initial matching model to obtain the predicted matching probability of each object with each media content; based on the first loss of the true matching probability and the predicted matching probability, the initial matching model is iteratively trained until the first loss converges to obtain the matching model.

[0194] For example, refer to Figure 10In the initial matching model, one-hot encoding is used to encode object and content information, resulting in initial object and content features. Feature mapping (embbedding) is then applied to these initial features to obtain object and content features. First, a first transformed feature is performed on the object features using the first MLP network in the multilayer perceptron layer of the initial matching model; and second, a second transformed feature is performed on the content features using the second MLP network in the matching model's MLP layer. The similarity between the first and second transformed features is then determined, and a predicted matching score is calculated based on this similarity. After obtaining the predicted matching score, a third MLP network performs feature compression and normalization on the matching score to obtain the predicted matching probability. Finally, a first loss is determined based on the predicted and true matching probabilities, and the initial matching model is iteratively trained using this first loss until it converges to obtain the final matching model.

[0195] In another implementation, when the true matching value includes the true matching score, the training samples are input into the initial matching model to obtain the predicted matching score between each object and each media content; the initial matching model is iteratively trained based on the first loss of the true matching score and the predicted matching score until the first loss converges to obtain the matching model.

[0196] In another implementation, when the true matching value includes the true matching score and the true matching probability, the training samples are input into the initial matching model to obtain the predicted matching score between each object and each media content; based on the first loss of the true matching score and the predicted matching score, as well as the first loss of the true matching probability and the predicted matching probability, the initial matching model is iteratively trained until the first loss of the true matching score and the predicted matching score, as well as the first loss of the true matching probability and the predicted matching probability, all converge to obtain the matching model.

[0197] In one possible scenario, to facilitate subsequent feature matching using a matching model, the object information and media content information in the training samples can be converted into features and then matched for dimensionality matching, i.e., feature mapping (embbedding). Specifically, firstly, initial object features corresponding to the object information and initial content features corresponding to the content information need to be generated; then, feature transformation processing is performed on the initial object features and initial content features to obtain object features (object embedding) and content features (content embedding), where both object features and content features correspond to the target dimension; then, a hash function is applied to each object embedding and content embedding to obtain the hash of each object embedding and content embedding; the hash of each embedding is used as the key, and the corresponding embedding is used as the value, stored in an embeddingtable (lookup table) so that the object embedding can be called based on the initial object features and the content embedding can be called based on the initial content features during the prediction process. During the initial matching model training process, the embedding in the lookup table can be updated based on the backpropagation gradient. The more times the model is trained, the higher the accuracy of converting high-dimensional feature vectors into low-dimensional feature vectors. When performing embedding processing on the initial object features corresponding to object information and the initial content information corresponding to content information, a shared embedding table can be used.

[0198] In this embodiment, by inputting training samples into the initial matching model, the predicted matching values ​​of each object and each media content are obtained; based on the first loss of the true matching value and the predicted matching value, the initial matching model is iteratively trained until the first loss converges to obtain the matching model, so as to obtain the matching value used to characterize the degree of matching between the target object and the candidate media content through the matching model.

[0199] In one embodiment, reference is made to Figure 11 The training samples also include the true target probabilities of each object performing preset operations on each media content; the method further includes: inputting the training samples into the initial probability prediction model to obtain the initial probability predicted by the training samples; obtaining the predicted target probabilities of each object performing preset operations on each media content based on the predicted initial probabilities and the predicted matching values; iteratively training the initial matching model and the initial probability prediction model based on the second loss between the predicted target probabilities and the true target probabilities until the second loss converges to obtain the matching model and the probability prediction model.

[0200] Specifically, the second loss is the negative log-likelihood loss. It can be calculated using the following formula:

[0201] L(x, H) = -log(P(ectr|ecvr))

[0202] Where L(x, H) represents the second loss; ectr represents the predicted target probability; and ecvr represents the true target probability.

[0203] For example, refer to Figure 12 Taking a preset operation including a first operation, a second operation, and a third operation as an example, the training samples are input into the initial probability prediction model to obtain the initial probabilities of the first operation, the second operation, and the third operation predicted by the training samples. Based on the initial probabilities of the first operation, the second operation, and the third operation, and the predicted matching value, the target probabilities of each object predicting each media content for the first operation, the second operation, and the third operation are obtained. The loss for the first operation is determined based on the initial probabilities of the first operation; the loss for the second operation is determined based on the initial probabilities of the second operation; the loss for the third operation is determined based on the initial probabilities of the third operation; and a second loss is determined based on the losses of the first, second, and third operations. The initial matching model and the initial probability prediction model are iteratively trained based on the second loss until the second loss converges to obtain the matching model and the probability prediction model.

[0204] Specifically, the second loss is the negative log-likelihood loss. It can be calculated using the following formula:

[0205] L(x,H)=-{log[P(ectr1|ecvr1)]+log[P(ectr2|ecvr2)]+log[P(ectr3|ecvr3)]}

[0206] Where L(x, h) represents the second loss; ectr1 represents the predicted target probability of the first operation; ecvr1 represents the actual target probability of the first operation; ectr2 represents the predicted target probability of the second operation; ecvr2 represents the actual target probability of the second operation; ectr3 represents the predicted target probability of the third operation; and ecvr3 represents the actual target probability of the third operation.

[0207] In this embodiment, by inputting training samples into the initial matching model and the initial probability prediction model, the predicted matching value and the predicted initial probability of each object and each media content are obtained, and the predicted target probability is determined based on the predicted matching value and the predicted initial probability. Based on the second loss of the true target probability and the predicted target probability, the initial matching model and the initial probability prediction model are iteratively trained until the second loss converges to obtain the matching model and the probability prediction model, so that the matching value used to characterize the matching degree between the target object and the candidate media content can be obtained through the matching model, and the initial probability of the preset operation can be obtained through the probability prediction model.

[0208] It should be noted that although the operation of the method of the present invention is described in a specific order in the accompanying drawings, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed in order to achieve the desired result.

[0209] Figure 13 This is a block diagram of a media content delivery device according to an embodiment of this application.

[0210] like Figure 13 As shown, the media content push device includes: an acquisition module 1301, a matching module 1302, a probability estimation module 1303, a first prediction module 1304, and a second prediction module 1305.

[0211] The acquisition module 1301 is used to acquire object information of the target object and content information of candidate media content.

[0212] The matching module 1302 is used to input object information and content information into the matching model to obtain the matching value between the target object and the candidate media content; the matching value is used to characterize the degree of matching between the target object and the candidate media content.

[0213] The probability prediction module 1303 is used to input object information and content information into the probability prediction model to obtain the initial probability of the target object performing a preset operation; the preset operation is the operation by which the target object expresses interest in the candidate media content.

[0214] The first prediction module 1304 is used to fuse the initial probability and the matching value to obtain the fused probability of the preset operation, and to perform attention allocation processing on the initial probability of the preset operation based on the fused probability of the preset operation to obtain the target probability of the preset operation.

[0215] The second prediction module 1305 is used to fuse the target probability and the matching value to determine the target object's interest score for candidate media content, and push candidate media content to the target object if the interest score is greater than a preset threshold.

[0216] In one embodiment, the matching value includes a matching score and a matching probability. The matching module 1302 is specifically used for...

[0217] The object information and content information are input into the matching model, and feature transformation processing is performed on the object information and content information respectively to obtain object features and content features.

[0218] The object features are compressed to obtain the first transformation feature, and the content features are compressed to obtain the second transformation feature.

[0219] Determine the similarity between the first and second transformation features, and determine the matching score based on the similarity.

[0220] The matching probability is obtained by performing feature compression and normalization on the matching score.

[0221] In one embodiment, the preset operation includes a first operation, a second operation, and a third operation; the first operation is a preceding operation to the second operation; the second operation is a preceding operation to the third operation, and the degree of interest represented by the first operation, the second operation, and the third operation is from low to high.

[0222] In one embodiment, the probability prediction module 1303 is specifically used for,

[0223] By inputting the object information and content information into the probability prediction model, the initial probability of the first operation, the source probability of the second operation, and the source probability of the third operation are obtained.

[0224] The initial probability of the second operation is obtained based on the initial probability of the first operation and the source probability of the second operation.

[0225] The initial probability of the third operation is obtained based on the initial probability of the second operation and the source probability of the third operation.

[0226] In one embodiment, the probability estimation module 1303 is specifically used to fuse the initial probability of the first operation and the source probability of the second operation to obtain the initial probability of the second operation.

[0227] In one embodiment, the probability prediction module 1303 is specifically used to fuse the initial probability of the second operation and the source probability of the third operation to obtain the initial probability of the third operation.

[0228] In one embodiment, the first prediction module 1304 is specifically used for,

[0229] The matching probability is fused with the initial probability of the first operation to obtain the fused probability of the first operation.

[0230] The matching probability is fused with the initial probability of the second operation to obtain the fused probability of the second operation.

[0231] The matching probability is fused with the initial probability of the third operation to obtain the fused probability of the third operation.

[0232] In one embodiment, the probability prediction module 1303 is specifically used for,

[0233] The object information and content information are input into the probability prediction model, and feature transformation processing is performed on the object information and content information respectively to obtain object features and content features.

[0234] The object features and content features are concatenated to obtain the first concatenated feature.

[0235] Based on the first splicing feature, determine the initial weight of the preset operation.

[0236] The initial weights of the preset operation are fused with the first concatenated feature to obtain the fused feature of the preset operation.

[0237] Based on the fusion characteristics of the preset operation, the initial probability of the preset operation is obtained.

[0238] In one embodiment, the probability prediction module 1303 is specifically used for,

[0239] Based on the first splicing feature, the preset operation is weighted to obtain at least one weighted result of the preset operation.

[0240] The results of at least one weight allocation are fused to obtain the initial weights for the preset operation.

[0241] In one embodiment, the acquisition module 1301 is further configured to acquire training samples, which include object information of multiple objects, content information of multiple media content, and the actual matching value between each object and each media content.

[0242] The matching module 1302 is used to input training samples into the initial matching model to obtain the predicted matching value between each object and each media content.

[0243] The matching module 1302 is also used to iteratively train the initial matching model based on the first loss of the true matching value and the predicted matching value until the first loss converges to obtain the matching model.

[0244] In one embodiment, the probability prediction module 1303 is further configured to input training samples into the initial probability prediction model to obtain the initial probability predicted by the training samples.

[0245] The first prediction module 1304 is used to obtain the target probability of each object performing a preset operation on each media content based on the initial prediction probability and the prediction matching value.

[0246] The first prediction module 1304 is also used to iteratively train the initial matching model and the initial probability prediction model based on the second loss between the predicted target probability and the true target probability, until the second loss converges to obtain the matching model and the probability prediction model.

[0247] The media content push device proposed in this application addresses the problem in existing technologies where the sparsity of information in media content leads to low accuracy in probability prediction, thus affecting the evaluation of interest levels. Specifically, the object information of the target object and the content information of candidate media content are input into a matching model and a probability prediction model, respectively. The matching model obtains a matching value characterizing the degree of matching between the target object and the candidate media content, while the probability prediction model obtains the initial probability of the target object performing a preset operation, where the preset operation is the operation by which the target object expresses interest in the candidate media content. Furthermore, the initial probability and the matching value are fused to obtain a fused probability of the preset operation, and attention allocation is applied to the initial probability of the preset operation based on this fused probability to obtain the target probability of the preset operation. In this way, by incorporating the matching value between the target object and the candidate media content into the prediction process of the final probability of the object performing related operations on the media content, the diversity of parameters is improved. This effectively solves the problem of low accuracy in probability prediction caused by the sparsity of relevant information in the media content, thereby improving the accuracy of the evaluation results of interest level.

[0248] Furthermore, considering that existing technologies rely solely on the probability of related operations to evaluate an object's interest in media content, the evaluation elements are too simplistic and may result in a significant discrepancy between the actual interest level and the actual content being pushed, leading to unsatisfactory media content delivery. Therefore, the push device in this application combines target probabilities and matching values ​​through a fusion process to comprehensively determine the object's interest score for candidate media content. The interest score obtained in this way better represents the target object's actual interest in the candidate media content. Based on this, pushing candidate media content to the target object based on the interest score can filter out media content that better matches the object's interests, improve the accuracy of personalized push, and effectively utilize the push resources of the network platform, reducing the overhead and waste of push resources.

[0249] It should be understood that the units recorded in the media content delivery device and the references Figure 2 The steps in the described method correspond to each other. Therefore, the operations and features described above for the method are also applicable to the media content delivery device and the units contained therein, and will not be repeated here. The media content delivery device can be pre-implemented in the browser or other security applications of the computer device, or it can be loaded into the browser or other security applications of the computer device through download or other means. The corresponding units in the media content delivery device can cooperate with the units in the computer device to implement the solution of the embodiments of this application.

[0250] The division of modules or units mentioned in the detailed description above is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0251] It should be noted that for details not disclosed in the media content push device of this application embodiment, please refer to the details disclosed in the above embodiments of this application, which will not be repeated here.

[0252] The following is for reference. Figure 14 , Figure 14 A schematic diagram of a computer device suitable for implementing embodiments of this application is shown, such as... Figure 14 As shown, the computer system 1400 includes a central processing unit (CPU) 1401, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 1402 or programs loaded from storage section 1408 into random access memory (RAM) 1403. The RAM 1403 also stores various programs and data required for the system's operating instructions. The CPU 1401, ROM 1402, and RAM 1403 are interconnected via a bus 1404. An input / output (I / O) interface 1405 is also connected to the bus 1404.

[0253] The following components are connected to I / O interface 1405: an input section 1406 including a keyboard, mouse, etc.; an output section 1407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1408 including a hard disk, etc.; and a communication section 1409 including a network interface card such as a LAN card, modem, etc. The communication section 1409 performs communication processing via a network such as the Internet. A drive 1410 is also connected to I / O interface 1405 as needed. Removable media 1411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1410 as needed so that computer programs read from them can be installed into storage section 1408 as needed.

[0254] Specifically, according to embodiments of this application, the flowchart above refers to... Figure 2 The described process can be implemented as a computer software program. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. In such an embodiment, the computer program contains program code for performing the methods shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 1409, and / or installed from removable medium 1411. When the computer program is executed by central processing unit (CPU) 1401, it performs the functions defined in the system of this application.

[0255] It should be noted that the computer-readable medium shown in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0256] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operational instructions of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two connected blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified functions or operational instructions, or using a combination of dedicated hardware and computer instructions.

[0257] The units or modules described in the embodiments of this application can be implemented in software or hardware. The described units or modules can also be housed in a processor; for example, a processor can be described as including a violator detection unit, a multimodal detection unit, and a recognition unit. The names of these units or modules do not necessarily constitute a limitation on the unit or module itself.

[0258] On the other hand, this application also provides a computer-readable storage medium, which may be included in the computer device described in the above embodiments, or may exist independently and not assembled into the computer device. The aforementioned computer-readable storage medium stores one or more programs that, when used by one or more processors, execute the media content push method described in this application. For example, it may execute... Figure 2 The steps of the media content delivery method are shown.

[0259] This application provides a computer program product including instructions that, when executed, cause the method described in this application to be performed. For example, it can execute... Figure 2 The steps of the media content delivery method are shown.

[0260] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for pushing media content, characterized in that, include: Obtain object information of the target object and content information of the candidate media content, input the object information and the content information into the matching model, and obtain the matching value between the target object and the candidate media content; The matching value is used to characterize the degree of matching between the target object and the candidate media content; The object information and the content information are input into the probability prediction model to obtain the initial probability of the target object performing a preset operation. The preset operation is the operation by which the target object expresses its degree of interest in the candidate media content; The initial probability and the matching value are fused to obtain the fused probability of the preset operation, and attention allocation is performed on the initial probability of the preset operation based on the fused probability of the preset operation to obtain the target probability of the preset operation. The target probability and the matching value are fused to determine the target object's interest score in the candidate media content. If the interest score is greater than a preset threshold, the candidate media content is pushed to the target object.

2. The method for pushing media content according to claim 1, characterized in that, The matching value includes a matching score and a matching probability. The step of inputting the object information and the content information into the matching model to obtain the matching value between the target object and the candidate media content includes: The object information and the content information are input into the matching model, and feature transformation processing is performed on the object information and the content information respectively to obtain object features and content features; The object features are subjected to feature compression processing to obtain a first transformation feature, and the content features are subjected to feature compression processing to obtain a second transformation feature; Determine the similarity between the first transformation feature and the second transformation feature, and determine the matching score based on the similarity; The matching probability is obtained by performing feature compression and normalization on the matching score.

3. The method for pushing media content according to claim 2, characterized in that, The preset operation includes a first operation, a second operation, and a third operation. The first operation is a preceding operation to the second operation, and the second operation is a preceding operation to the third operation. The degree of interest represented by the first operation, the second operation, and the third operation is from low to high.

4. The method for pushing media content according to claim 3, characterized in that, The step of inputting the object information and the content information into the probability prediction model to obtain the initial probability that the target object will perform a preset operation on the candidate media content includes: The object information and the content information are input into the probability prediction model to obtain the initial probability of the first operation, the source probability of the second operation, and the source probability of the third operation. The initial probability of the second operation is obtained based on the initial probability of the first operation and the source probability of the second operation. The initial probability of the third operation is obtained based on the initial probability of the second operation and the source probability of the third operation.

5. The method for pushing media content according to claim 4, characterized in that, The step of obtaining the initial probability of the second operation based on the initial probability of the first operation and the source probability of the second operation includes: The initial probability of the first operation and the source probability of the second operation are fused to obtain the initial probability of the second operation.

6. The method for pushing media content according to claim 4, characterized in that, The step of obtaining the initial probability of the third operation based on the initial probability of the second operation and the source probability of the third operation includes: The initial probability of the second operation and the source probability of the third operation are fused to obtain the initial probability of the third operation.

7. The method for pushing media content according to claim 4 or 5, characterized in that, The process of fusing the initial probability and the matching value to obtain the fused probability of the preset operation includes: The matching probability is fused with the initial probability of the first operation to obtain the fused probability of the first operation; The matching probability is fused with the initial probability of the second operation to obtain the fused probability of the second operation; The matching probability is fused with the initial probability of the third operation to obtain the fused probability of the third operation.

8. The method for pushing media content according to claim 1, characterized in that, The step of inputting the object information and the content information into the probability prediction model to obtain the initial probability of the target object performing a preset operation includes: The object information and the content information are input into the probability prediction model, and feature transformation processing is performed on the object information and the content information respectively to obtain object features and content features; The object features and the content features are concatenated to obtain the first concatenated feature; Based on the first splicing feature, determine the initial weight of the preset operation; The initial weight of the preset operation is fused with the first concatenation feature to obtain the fused feature of the preset operation; The initial probability of the preset operation is obtained based on the fusion characteristics of the preset operation.

9. The method for pushing media content according to claim 8, characterized in that, Determining the initial weight of the preset operation based on the first splicing feature includes: Based on the first splicing feature, the preset operation is weighted to obtain at least one weighting result of the preset operation. The at least one weight allocation result is fused to obtain the initial weight of the preset operation.

10. The method for pushing media content according to any one of claims 1-9, characterized in that, The method further includes: Obtain training samples, which include object information of multiple objects, content information of multiple media contents, and the actual matching values ​​between each object and each media content; The training samples are input into the initial matching model to obtain the predicted matching value between each object and each media content; The initial matching model is iteratively trained based on the first loss of the actual matching value and the predicted matching value until the first loss converges to obtain the matching model.

11. The method for pushing media content according to claim 10, characterized in that, The training samples also include the true target probability of each object performing the preset operation on each media content; the method further includes: The training samples are input into the initial probability prediction model to obtain the initial probability predicted by the training samples; Based on the predicted initial probability and the predicted matching value, the predicted target probability of each object performing the preset operation on each media content is obtained. Based on the second loss between the predicted target probability and the actual target probability, the initial matching model and the initial probability prediction model are iteratively trained until the second loss converges to obtain the matching model and the probability prediction model.

12. A media content delivery device, characterized in that, include: The acquisition module is used to acquire object information of the target object and content information of the candidate media content; The matching module is used to input the object information and the content information into the matching model to obtain the matching value between the target object and the candidate media content; The matching value is used to characterize the degree of matching between the target object and the candidate media content; The probability prediction module is used to input the object information and the content information into the probability prediction model to obtain the initial probability of the target object performing a preset operation. The preset operation is the operation by which the target object expresses its degree of interest in the candidate media content; The first prediction module is used to fuse the initial probability and the matching value to obtain the fused probability of the preset operation, and to perform attention allocation processing on the initial probability of the preset operation based on the fused probability of the preset operation to obtain the target probability of the preset operation. The second prediction module is used to fuse the target probability and the matching value to determine the target object's interest score in the candidate media content, and push the candidate media content to the target object if the interest score is greater than a preset threshold.

13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, The computer program product includes instructions that, when executed, cause the method as described in any one of claims 1 to 11 to be performed.

Images