Information pushing method and device, equipment and storage medium

Abstract

The embodiment of the invention discloses an information push method and device, equipment and a storage medium, and the method comprises the steps: firstly obtaining a first feature of candidate push information belonging to a first information category, the first information category being one of a plurality of second information categories; then obtaining category features of a plurality of second information categories obtained by pre-training; extracting a second feature of the object attribute of the target object in a different second information category; fusing the plurality of second features and the plurality of category features to obtain object features of the target object adaptive to the plurality of second information categories; and when the similarity between the object feature and the first feature is not less than a preset similarity threshold, pushing the candidate push information to the target object. According to the embodiment of the invention, the association between the object feature and the first feature of the candidate push information is constructed through the affiliation between the first information category and the second information category, so that the candidate push information can be pushed to the target object more accurately.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to an information push method, apparatus, device and storage medium. Background Technology

[0002] In related technologies, information push platforms generally push information to target objects based on a dual-tower model, see [link to relevant documentation]. Figure 1 As shown, the dual-tower model includes an object tower model and an information tower model. The object tower model extracts features based on the object attributes of the target object, outputting object attribute features. The information tower model extracts features based on the information to be pushed, outputting information features. Finally, the similarity between the object attribute features and the information features is calculated to determine how to push the information to the target object.

[0003] However, the current method cannot guarantee the accuracy of the similarity between the object attribute features output by the object tower model and the information features output by the information tower model, which may easily lead to the dual-tower model failing to accurately push information to the target object. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail in this application. This overview is not intended to limit the scope of the claims.

[0005] This application provides an information push method, apparatus, device, and storage medium that can push information to target objects more accurately.

[0006] On the one hand, embodiments of this application provide an information push method, including:

[0007] Obtain the first feature of the candidate push information, wherein the candidate push information belongs to the first information category;

[0008] Obtain category features of multiple second information categories obtained from pre-training, wherein the first information category is one of the multiple second information categories;

[0009] The object attributes of the target object are extracted based on each of the second information categories to obtain the second features of the object attributes under different second information categories;

[0010] By fusing multiple second features and multiple category features, the target object is obtained as an object feature that adapts to multiple second information categories;

[0011] Calculate the first similarity between the object feature and the first feature. If the first similarity is not less than a preset similarity threshold, push the candidate push information to the target object.

[0012] On the other hand, embodiments of this application provide an information push device, including:

[0013] The first feature acquisition unit is used to acquire the first feature of the candidate push information, wherein the candidate push information belongs to the first information category;

[0014] The category feature acquisition unit is used to acquire category features of multiple second information categories obtained through pre-training, wherein the first information category is one of the multiple second information categories;

[0015] The second feature extraction unit is used to extract features of the object attributes of the target object based on each of the second information categories, so as to obtain the second features of the object attributes under different second information categories.

[0016] The feature fusion unit is used to perform feature fusion on multiple second features and multiple category features to obtain object features of the target object that are adapted to multiple second information categories;

[0017] An information push unit is used to calculate a first similarity between the object feature and the first feature, and when the first similarity is not less than a preset similarity threshold, push the candidate push information to the target object.

[0018] Optionally, the feature fusion unit is specifically used for:

[0019] Attention is calculated for each second feature and the corresponding category feature to obtain the attention features of the object attribute under different second information categories;

[0020] The multiple attention features are weighted and summed to obtain the object features of the target object that are adapted to the second information category.

[0021] Optionally, the feature fusion unit is specifically used for:

[0022] For each of the second features, the second feature is used as a key vector and a value vector, respectively, and the corresponding category feature is used as a query vector. Attention is calculated to obtain the attention vector corresponding to each of the second features.

[0023] Based on the attention vectors corresponding to all the second features, the attention features of the object attributes under different second information categories are obtained.

[0024] Optionally, the feature fusion unit is specifically used for:

[0025] Obtain the fusion weights corresponding to each attention feature;

[0026] Multiple target weights are determined from all the fusion weights, and multiple corresponding target attention features are determined from all the attention features based on the multiple target weights;

[0027] The target attention features and their corresponding target weights are weighted and summed to obtain the object features of the target object that are adapted to the second information category.

[0028] Optionally, the step of feature fusion of the multiple second features and the multiple category features is performed by the object feature extraction model; the feature fusion unit is specifically used for:

[0029] Obtain the hidden layer parameters of the object feature extraction model;

[0030] Based on the hidden layer parameters and each attention feature, the initial weight corresponding to each attention feature is calculated;

[0031] The initial weights corresponding to each attention feature are normalized to obtain the fusion weights corresponding to each attention feature.

[0032] Optionally, the hidden layer parameters include a first hidden layer coefficient and a second hidden layer coefficient; the feature fusion unit is specifically used for:

[0033] For each attention feature, the product of the attention feature and the first hidden layer coefficient is added to the second hidden layer coefficient to obtain the initial weight corresponding to each attention feature.

[0034] Optionally, the steps of extracting features from the object attributes of the target object based on each of the second information categories, and the steps of fusing multiple second features and multiple category features, are both performed by the object feature extraction model; the step of extracting features from the candidate push information is performed by the push information feature extraction model.

[0035] The information push device further includes a category feature training unit, which is used for:

[0036] Obtain the initial features for each of the second information categories;

[0037] Obtain training samples and training labels. The training samples include object samples and push information samples. The information category to which the push information samples belong is one of multiple second information categories.

[0038] The push information feature extraction model is invoked to extract features from the push information sample, thereby obtaining the first sample features;

[0039] The object feature extraction model is invoked to extract features from the object sample based on each of the second information categories, thereby obtaining the second sample features of the object sample under different second information categories.

[0040] The object feature extraction model is invoked to perform feature fusion on multiple second sample features and multiple initial features to obtain object sample features of the object sample that are adapted to multiple second information categories;

[0041] Calculate the second similarity between the object sample features and the first sample features;

[0042] A first loss value is calculated based on the initial features of multiple second information categories;

[0043] The second loss value is calculated based on the first sample features and the initial features of multiple second information categories;

[0044] The third loss value is calculated based on the second similarity and the training labels;

[0045] Based on the first loss value, the second loss value, and the third loss value, the initial features of the multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model are adjusted.

[0046] The adjusted initial features of the multiple second information categories are used as the category features of the multiple second information categories.

[0047] Optionally, the category feature training unit is specifically used for:

[0048] Obtain the first hyperparameter;

[0049] The first parameter is determined based on the quantity of the second information category;

[0050] Subtract the first hyperparameter from the initial feature of each second information category to obtain the first difference of each second information category;

[0051] Calculate the product of the initial features of each pair of the second information categories and the first difference to obtain multiple first products;

[0052] By performing a positive operation on each of the multiple first products, multiple first results are obtained;

[0053] Based on the first parameter and multiple first results, a first loss value is calculated.

[0054] Optionally, the category feature training unit is specifically used for:

[0055] Determine the second information category to which the push information sample belongs;

[0056] Calculate the first distance between the first sample feature and the initial feature of the second information category to which the push information sample belongs;

[0057] The second loss value is calculated based on the first distance.

[0058] Optionally, the category feature training unit is specifically used for:

[0059] Based on the second similarity and the training labels, the positive loss when the model predicts correctly and the negative loss when the model predicts incorrectly are calculated.

[0060] The third loss value is calculated based on the positive loss and the negative loss.

[0061] Optionally, the category feature training unit is specifically used for:

[0062] Obtain the second hyperparameter;

[0063] The corrected loss value is obtained by multiplying the second hyperparameter by the sum of the first loss value and the second loss value.

[0064] The total loss value is calculated based on the corrected loss value and the third loss value.

[0065] Based on the overall loss value, the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model are adjusted.

[0066] On the other hand, embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described image classification method or model training method.

[0067] On the other hand, embodiments of this application also provide a computer-readable storage medium storing a computer program, which is executed by a processor to implement the above-described image classification method or model training method.

[0068] On the other hand, embodiments of this application also provide a computer program product, which includes a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the image classification method or model training method described above.

[0069] In the information push method provided in this application embodiment, a first feature of candidate push information is first obtained, wherein the candidate push information belongs to a first information category; then, category features of multiple pre-trained second information categories are obtained, wherein the first information category is one of the multiple second information categories; then, feature extraction based on each second information category is performed on the object attributes of the target object to obtain second features of the object attributes under different second information categories; next, feature fusion is performed on the multiple second features and multiple category features to obtain object features of the target object adapted to multiple second information categories; at this time, a first similarity between the object features and the first features is calculated, and when the first similarity is not less than a preset similarity threshold, the candidate push information is pushed to the target object. Since the second features are obtained by feature extraction based on each second information category on the object attributes of the target object, the second features can characterize the bias of the object attributes of the target object towards each second information category. Next, by fusing multiple second features and multiple category features to obtain object features, the object attributes and the second information category factor can be organically combined. This allows the object features to highlight the semantics under multiple second information categories, thereby enabling them to characterize the target object's degree of interest in different second information categories. Furthermore, since the first information category is one of multiple second information categories, the object features obtained by feature fusion based on second features and category features can have a certain correlation with the first features of the candidate push information. This improves the accuracy of the first similarity between the object features and the first features, thus enabling more accurate push of the candidate push information to the target object.

[0070] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. Attached Figure Description

[0071] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0072] Figure 1 This is a structural diagram of a dual-tower model in related technologies;

[0073] Figure 2 A schematic diagram illustrating an optional implementation environment provided for an embodiment of this application;

[0074] Figure 3 A schematic diagram of another optional implementation environment provided for the embodiments of this application;

[0075] Figure 4 This is a flowchart of an information push method provided in an embodiment of this application;

[0076] Figure 5 This is an execution architecture diagram of an information push method provided in this application during the inference phase;

[0077] Figure 6 It shows Figure 5 The processing principle of the feature extraction model on each side of the neural network architecture;

[0078] Figure 7 This is an execution architecture diagram of an information push method provided in this application embodiment during the training phase;

[0079] Figure 8 This is a schematic diagram of an information push device provided in an embodiment of this application;

[0080] Figure 9 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0081] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0082] In related technologies, the method of pushing information to target objects based on the dual-tower model cannot guarantee the accuracy of the similarity between the object attribute features output by the object tower model and the information features output by the information tower model. This can easily lead to the dual-tower model failing to accurately push information to the target object.

[0083] To more accurately push information to target objects, embodiments of this application provide an information push method, an information push device, an electronic device, a computer-readable storage medium, and a computer program product. In this method, a first feature of candidate push information is first obtained, where the candidate push information belongs to a first information category; then, category features of multiple pre-trained second information categories are obtained, where the first information category is one of the multiple second information categories; next, feature extraction is performed on the object attributes of the target object based on each second information category to obtain second features of the object attributes under different second information categories; then, feature fusion is performed on the multiple second features and multiple category features to obtain object features of the target object adapted to multiple second information categories; at this point, a first similarity between the object features and the first features is calculated; when the first similarity is not less than a preset similarity threshold, the candidate push information is pushed to the target object. Since the second features are obtained by extracting features from the object attributes of the target object based on each second information category, the second features can characterize the bias of the target object's object attributes towards each second information category. Next, by fusing multiple second features and multiple category features to obtain object features, the object attributes and the second information category factor can be organically combined. This allows the object features to highlight the semantics under multiple second information categories, thereby enabling them to characterize the target object's degree of interest in different second information categories. Furthermore, since the first information category is one of multiple second information categories, the object features obtained by feature fusion based on second features and category features can have a certain correlation with the first features of the candidate push information. This improves the accuracy of the first similarity between the object features and the first features, thus enabling more accurate push of the candidate push information to the target object.

[0084] Figure 2 This is a schematic diagram of an implementation environment provided in an embodiment of this application. See also... Figure 2 The environment of this embodiment may include a first server 201, wherein the first server 201 may be a node in the blockchain, and this embodiment does not specifically limit it.

[0085] The first server 201 can be an independent 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.

[0086] See Figure 2In one application scenario, the first server 201 can first obtain the first feature of the candidate push information, which belongs to the first information category. Then, it can obtain the category features of multiple second information categories obtained through pre-training, where the first information category is one of the multiple second information categories. Then, it can extract the object attributes of the target object based on the features of each second information category to obtain the second features of the object attributes under different second information categories. Then, it can perform feature fusion on the multiple second features and multiple category features to obtain the object features of the target object that are adapted to multiple second information categories. Then, it can calculate the first similarity between the object features and the first features. When the first similarity is not less than a preset similarity threshold, the candidate push information is pushed to the target object.

[0087] Figure 3 This is a schematic diagram of another implementation environment provided in this application. See also... Figure 3 The environment of this embodiment may include a server cluster 301, which includes a second server 302, a third server 303, and a fourth server 304. The second server 302, the third server 303, and the fourth server 304 can be directly or indirectly connected through wired or wireless communication. The second server 302, the third server 303, and the fourth server 304 can all be nodes in the blockchain, but this embodiment does not specifically limit them.

[0088] The second server 302, the third server 303, and the fourth server 304 can all be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Specifically, the second server 302 can acquire the features of the push information and send these features to the fourth server 304; the third server 303 can process the object attributes of the target object to obtain object features that fit multiple second information categories and send these object features to the fourth server 304; the fourth server 304 can calculate the similarity between the object features and the features of the push information, and if the calculated similarity is not less than a preset similarity threshold, the push information is pushed to the target object.

[0089] See Figure 3In one application scenario, when server cluster 301 needs to perform an information push task, for the second server 302, the second server 302 can obtain the first feature of the candidate push information, which belongs to the first information category. At this time, the second server 302 can send the obtained first feature to the fourth server 304. For the third server 303, the third server 303 can first obtain the category features of multiple pre-trained second information categories, where the first information category is one of the multiple second information categories. Then, it performs feature extraction on the object attributes of the target object based on each second information category to obtain the second features of the object attributes under different second information categories. Then, it performs feature fusion on the multiple second features and multiple category features to obtain the object features of the target object that are adapted to multiple second information categories, and sends the obtained object features to the fourth server 304. For the fourth server 304, after receiving the first feature and the object feature, the fourth server 304 can calculate the first similarity between the object feature and the first feature. When the first similarity is not less than a preset similarity threshold, the candidate push information is pushed to the target object.

[0090] In one embodiment, a pre-defined distributed coordination mechanism within the server cluster 301 can control the execution order of the second server 302 acquiring the first feature and the third server 303 extracting features from the object attributes. The two can be executed synchronously or asynchronously, and no specific limitation is made here.

[0091] It should be noted that in various specific embodiments of this application, when processing data related to the characteristics of the target object (e.g., a user's attribute information or a set of attribute information) is required, the target object's permission or consent will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application need to obtain the target object's attribute information, separate permission or consent from the target object will be obtained through pop-ups or redirection to a confirmation page. Only after obtaining the target object's separate permission or consent will the necessary target object-related data for the normal operation of the embodiments of this application be obtained.

[0092] Figure 4 This is a flowchart illustrating an information push method provided in an embodiment of this application. This information push method can be executed by a single server or by a server cluster. In this embodiment, the method is described using an example of execution by a single server. See also... Figure 4 The information push method may include steps 410 to 430.

[0093] Step 410: Obtain the first feature of the candidate push information, which belongs to the first information category;

[0094] Step 420: Obtain the category features of multiple second information categories obtained from pre-training, wherein the first information category is one of the multiple second information categories;

[0095] Step 430: Extract features from the object attributes of the target object based on each second information category to obtain the second features of the object attributes under different second information categories;

[0096] Step 440: Perform feature fusion on multiple second features and multiple category features to obtain object features of the target object that are adapted to multiple second information categories;

[0097] Step 450: Calculate the first similarity between the object feature and the first feature. If the first similarity is not less than the preset similarity threshold, push the candidate push information to the target object.

[0098] In one embodiment, push information refers to information actively sent by an information push platform to a target object. Here, the information push platform refers to a technical platform used to send personalized or relevant information, notifications, messages, etc., to the target object in real time or periodically. Push information can be updates, notifications, advertisements, music, videos, etc., and is not specifically limited here. Candidate push information refers to push information selected by the information push platform from multiple push information stored therein, and whose suitability for being pushed to the target object is determined by the information push platform. For example, assuming the information push platform stores push information A, B, and C, and push information A is selected by the information push platform, and then it is determined whether push information A can be pushed to the target object, push information A is considered candidate push information. The information push platform can select candidate push information from multiple stored push information through selection methods such as random selection or selection according to preset rules, and is not specifically limited here.

[0099] In one embodiment, the first feature refers to the vector obtained by extracting features from candidate push information using a pre-trained push information feature extraction model. Here, the push information feature extraction model refers to a neural network model used to extract features from push information. The push information feature extraction model can be trained from one or more neural network models, such as a Convolutional Neural Network (CNN) or a Transformer model, and is not specifically limited here.

[0100] In one embodiment, before obtaining the first feature of the candidate push information, the first feature of the candidate push information can be obtained by extracting features from the candidate push information using a pre-trained push information feature extraction model and then stored. During the process of obtaining the first feature of the candidate push information, the first feature of the candidate push information can be called from among the features of multiple stored push information based on the candidate push information. By pre-extracting features from the push information, it is not necessary to extract features of the candidate push information in real time during the information push process to the target object, thereby improving the efficiency of the information push process. Furthermore, since the first feature extracted by the same push information feature extraction model for the same candidate push information is unique, when pushing the candidate push information to different objects, the push information feature extraction model does not need to repeatedly extract features of the same candidate push information, which helps save computational resources.

[0101] In one embodiment, in the process of obtaining the first feature of the candidate push information, the candidate push information can be obtained first, and then the candidate push information can be extracted using a pre-trained push information feature extraction model to obtain the first feature of the candidate push information.

[0102] In one embodiment, an information category refers to an information category used for classifying push information, and a second information category refers to a specific information category obtained by classifying push information according to the attributes of each push information stored by the information push platform. For example, suppose there are candidate push information in the form of advertisements, namely "fan advertisement" and "children's clothing advertisement". "Fan advertisement" is classified into "Electrical category" and "children's clothing advertisement" is classified into "Clothing category". Then, "Electrical category" and "Clothing category" are the second information categories. For the candidate push information "fan advertisement", "Electrical category" is the first information category to which "fan advertisement" belongs; for the candidate push information "children's clothing advertisement", "Clothing category" is the first information category to which "children's clothing advertisement" belongs. For another example, suppose there are candidate push notifications that are specifically in the form of advertisements, namely "fan advertisement" and "car advertisement". "Fan advertisement" is classified under "furniture industry" and "car advertisement" is classified under "vehicle industry". In this case, "furniture industry" and "vehicle industry" are the second information categories. However, for the candidate push notification "fan advertisement", "furniture industry" is the first information category to which "fan advertisement" belongs; and for the candidate push notification "car advertisement", "vehicle industry" is the first information category to which "car advertisement" belongs.

[0103] In one embodiment, the multiple second information categories preset by the information push platform may have a hierarchical relationship (one second information category is a subordinate category of another second information category), or they may only have a same-level relationship. The specific relationship is not limited here.

[0104] In one embodiment, category features refer to information attributes used to characterize the second information category. Pre-trained category features refer to the features of the second information category obtained after updating the features of the second information category during the training phase based on the loss value, which can be used in the inference phase. Specifically, when the process of feature extraction of object attributes is in the training phase, the loss value is determined based on the features of the pushed information samples, and the features of the second information category are adjusted accordingly.

[0105] In one embodiment, when there is a hierarchical relationship between multiple second information categories preset by the information push platform, in the process of obtaining the category features of multiple pre-trained second information categories, the hierarchical information of the second information categories can be obtained first. The hierarchical information of the second information categories refers to the information used to represent the specific level of the second information categories used in the process of extracting the object attributes of the target object based on the features of each second information category. Then, the target level is determined based on the hierarchical information of the second information categories, and the category features of multiple pre-trained second information categories in the target level are obtained. For example, suppose the second information category includes "car", "sedan", "clothes", and "top", and the candidate push message is "car maintenance discount notification". Here, "car" is the parent second information category of "sedan", and "clothes" is the parent second information category of "top". In the process of obtaining category features, the hierarchical information of the second information category is obtained first. Suppose the hierarchical information of the second information category specifically identifies the specific level of the second information category used this time as the lower level (that is, the level where "sedan" and "top" are located). Then, based on the hierarchical information of the second information category, the target level can be determined as the lower level. Then, the pre-trained category features of "sedan" and "top" are obtained. At this time, the feature extraction process of object attributes will be based on the category features of "sedan" and "top". Since lower-level hierarchies can more specifically represent the content of higher-level second information categories, the accuracy of feature extraction for lower-level second information categories can be higher than that for higher-level second information categories. Conversely, when extracting features for higher-level second information categories, the efficiency of feature extraction for higher-level second information categories can be higher than that for lower-level second information categories. Therefore, in the process of extracting object attributes, selecting multiple levels of second information categories through the second information category hierarchy information can control the accuracy and efficiency of the extracted object attribute features. This can effectively improve the efficiency of extracting second features during the inference stage while also maintaining the accuracy of the second features.

[0106] In one embodiment, the object attributes of the target object refer to the characteristics that the target object possesses at multiple levels. Specifically, the object attributes can be behavioral characteristics, interests and hobbies, health information, geographical location information, psychological state, etc., without being specifically limited here.

[0107] In one embodiment, feature extraction of the object attributes of the target object based on each second information category refers to the process of extracting features of the object attributes of the target object for each second information category to obtain the second features under that second information category. In this process, the number of second features corresponds to the number of second information categories.

[0108] In one embodiment, feature extraction based on each second information category during the inference phase can be achieved by pre-setting a category weight value for each second information category. Then, by extracting the features of the object attributes and multiplying the features of the object attributes by a category weight value, feature extraction of the object attributes of the target object for a second information category can be realized.

[0109] In one embodiment, each second information category is pre-set with a category convolution kernel for feature extraction. The object attributes of the target object are extracted for a second information category by using the category convolution kernel corresponding to a second information category.

[0110] In one embodiment, feature fusion of multiple second features and multiple category features refers to first fusing each of the multiple second features with its corresponding category feature, and then fusing the multiple features obtained from the previous fusion again to obtain the object features. The feature fusion can be feature fusion for all second information categories, or feature fusion for multiple features within all second information categories; the specific method is not limited here.

[0111] In one embodiment, object features refer to object attribute features that carry information representing the target object's degree of interest in multiple second information categories. Adapting to multiple second information categories means that the object features can represent the target object's degree of interest in multiple second information categories. For example, suppose the second features of object attributes under second information categories A, B, C, and D correspond to A1, B1, C1, and D1 respectively, and the category features of second information categories A, B, C, and D correspond to A2, B2, C2, and D2 respectively. Further suppose that feature fusion is performed on all second information categories. Then, in the process of feature fusion of multiple second features and multiple category features, each second feature can first be fused with its corresponding category feature to obtain features A1A2, B1B2, C1C2, and D1D2. Then, features A1A2, B1B2, C1C2, and D1D2 are fused together to obtain object features A1A2B1B2C1C2D1D2. For each second feature, the fusion of the second feature and the corresponding category feature can be by concatenating the second feature with the corresponding category feature, or by multiplying the second feature with the corresponding category feature, etc., and the specific method is not limited here.

[0112] In one embodiment, the first similarity can be obtained by calculating one of the specific data such as cosine similarity, Euclidean distance, and Manhattan distance between the object feature and the first feature, without being limited to any specific data here.

[0113] In one embodiment, after calculating the first similarity between the object feature and the first feature, the first similarity can be compared with a preset similarity threshold. If the first similarity is not less than the preset similarity threshold, the candidate push information is pushed to the target object.

[0114] In one embodiment, after comparing the first similarity with a preset similarity threshold, if the first similarity is less than the preset similarity threshold, the process of pushing candidate push information to the target object is canceled.

[0115] In one embodiment, during the feature fusion of multiple second features and multiple category features to obtain object features of a target object adapted to multiple second information categories, attention calculation can be performed on each second feature and its corresponding category feature to obtain attention features of the object attribute under different second information categories. Then, the multiple attention features are weighted and summed to obtain the object features of the target object adapted to multiple second information categories. Since the second feature has a bias towards a particular second information category, by performing attention calculation on each second feature and its corresponding category feature, the category feature can be transformed into a standard for the target object's interest in the second information category. This allows each attention feature to represent the emphasis of the object attribute on its corresponding second information category, thus enabling the attention feature to characterize the target object's interest in a particular second information category. Furthermore, the weighted summation of multiple attention features allows the interest in a particular second information category represented by each attention feature to be abstracted, thereby enabling the object feature to represent the interest in multiple second information categories. In addition, since the first information category is one of multiple second information categories, and the first information category is the second information category to which the candidate push information belongs, the attention calculation process can be associated with the first feature, thereby effectively utilizing the relevant attribute information of the candidate push information.

[0116] In one embodiment, attention computation refers to the process of performing attention-based feature fusion on the second feature and category feature corresponding to each second information category to obtain attention features. Here, attention features refer to information used to characterize the degree of interest of a target object in push information belonging to a second information category.

[0117] In one embodiment, the weighted summation of multiple attention features can be a weighted summation of all attention features, or a weighted summation of multiple attention features among all attention features; the specific method is not limited here. The multiple attention features can be randomly selected or selected according to a preset rule; the specific method is not limited here.

[0118] In one embodiment, during the attention calculation process for each second feature and its corresponding category feature, an attention weight can be calculated for each second information category. Then, during the weighted summation of multiple attention features, the attention weight for each second information category is first compared with a preset weight threshold. Next, the attention features of the second information categories corresponding to attention weights with values ​​greater than or equal to the weight threshold are weighted and summed. By comparing the attention weight for each second information category with the preset weight threshold, the second information categories that the target object is relatively interested in among multiple second information categories can be determined. Therefore, the attention features under these more interesting second information categories can be weighted and summed. This allows the object features to emphasize the target object's interest in the second information categories that are relatively interesting among multiple second information categories. Consequently, the information push process can pay more attention to the impact of the interest in the more interesting second information categories on information push, which is beneficial for more accurately pushing information to the target object.

[0119] In one embodiment, in the process of performing attention calculation on each second feature and its corresponding category feature to obtain the attention features of object attributes under different second information categories, specifically, for each second feature, the second feature can be used as a key vector and a value vector, and the corresponding category feature can be used as a query vector for attention calculation to obtain the attention vector corresponding to each second feature. Then, based on the attention vectors corresponding to all second features, the attention features of object attributes under different second information categories are obtained. By using the second features as key vectors and value vectors and the corresponding category features as query vectors for attention calculation, the focus of the second features on the information in the category features can be determined, and the information that is of particular interest in the category features can be highlighted. Since the second feature can characterize the bias of object attributes towards a second information category, the attention information highlighted by the attention vector can characterize the degree of bias of the target object towards a second information category, thereby better representing the degree of interest of the target object in a second information category. Furthermore, since the first information category is one of multiple second information categories, and the first information category is the second information category to which the candidate push information belongs, the attention vector corresponding to the first information category among multiple attention vectors can represent the target object's interest in the first information category. This allows for the construction of the association between object attributes and candidate push information, which is beneficial for strengthening the perception of candidate push information.

[0120] In one embodiment, using the second feature as both the key vector and the value vector, and using the corresponding category feature as the query vector, refers to using the second feature as the focus standard for the information included in the corresponding category feature, and using the second feature as the weighted representation of the information that focuses on the second feature among the information included in the corresponding category feature.

[0121] In one embodiment, the second feature is used as the key vector and value vector respectively, and the corresponding category feature is used as the query vector. Attention is calculated to obtain the attention vector corresponding to the second feature. This can be achieved by the following formula (1).

[0122]

[0123] in, It can be used to represent the attention vector corresponding to the i-th second feature, where i can be used to represent the order of the second information categories. It should be noted that there is a one-to-one correspondence between multiple second features and multiple second information categories.

[0124] σ() can be used to represent a normalization function, which can be a softmax function, a sigmoid function, etc., without being limited here;

[0125] Q can be used to query vectors, where Q can be substituted with the category feature of the second information category corresponding to the i-th second feature, and the category feature is a 1×d feature vector;

[0126] K can be used to represent the key vector, where K can be substituted with the i-th second feature, the second feature is an n×d feature vector, and n can be used to represent the number of second features;

[0127] K T It can be used to represent the transpose of the key vector K;

[0128] V can be used to represent a value vector, where V can be substituted with the i-th second feature.

[0129] d can be used to represent the dimension of a vector;

[0130] d k It can be used to represent the vector dimension of the i-th second feature.

[0131] In one embodiment, since features can be represented by vectors, and attention features are calculated based on the second features for category features, in the process of obtaining the attention features of object attributes under different second information categories according to the attention vectors corresponding to all second features, the attention vector corresponding to each second feature can be output as the attention feature under the second information category corresponding to the category feature targeted by the second feature.

[0132] In one embodiment, in the process of weighted summation of multiple attention features to obtain object features of a target object suitable for multiple second information categories, the fusion weight corresponding to each attention feature can be obtained first. Then, multiple target weights are determined from all fusion weights, and multiple target attention features are determined from all attention features based on the multiple target weights. The multiple target attention features and their respective target weights are weighted summation to obtain object features of the target object suitable for multiple second information categories. Since the fusion weight can represent the importance of its corresponding attention feature during the fusion process, by determining the target weight through the fusion weight and then determining the target attention feature based on the target weight, multiple more important second information categories can be identified from all second information categories. Thus, the attention features under these more important second information categories can be weighted summation, which allows the object features to focus on representing the target object's interest in important and interesting second information categories among multiple second information categories, which is beneficial for more accurately pushing information to the target object.

[0133] In one embodiment, the fusion weight refers to the weight used to adjust the importance of an attention feature relative to other attention features during the attention feature fusion process, thereby indicating the importance of the target object to the second information category corresponding to that attention feature in other second information categories. The fusion weight can be a preset value, or it can be obtained by weight calculation using relevant data (such as category features, second features, etc.), etc., and is not specifically limited here.

[0134] In one embodiment, before determining multiple target weights from all fused weights, a target weight threshold can be preset. Then, in the process of determining multiple target weights from all fused weights, each fused weight can be compared with the target weight threshold, and the fused weights that are greater than or equal to the target weight threshold are determined as target weights.

[0135] In one embodiment, during the process of determining multiple target weights from all fusion weights, the fusion weights can be sorted according to their specific values. Then, the fusion weight with the largest value among multiple preset numbers is determined as the target weight. The sorting of all fusion weights can be from largest to smallest or from smallest to largest; no specific limitation is made here.

[0136] In one embodiment, determining the corresponding multiple target attention features among all attention features based on multiple target weights refers to the process of determining the attention feature corresponding to each target weight as the target attention feature. Here, the target attention feature refers to the attention feature corresponding to a target weight among all attention features.

[0137] In one embodiment, the process of weighted summing of multiple target attention features and their corresponding target weights to obtain object features suitable for multiple second information categories refers to first multiplying each target attention feature and its corresponding target weight, and then summing the multiplication results of each target attention feature to obtain object features that can represent the degree of interest of the target object in the second information categories corresponding to the multiple target weights. Here, the multiplication result corresponding to the target attention feature is a feature that has been multiplied by the target weights.

[0138] In one embodiment, the target attention features and their corresponding target weights are weighted and summed to obtain the target object features that are adapted to multiple second information categories, which can be represented by the following formula (2).

[0139]

[0140] Among them, U I It can be used to represent object characteristics;

[0141] t can be used to represent the order of target attention features;

[0142] G can be used to represent the number of fusion weights, where the number of fusion weights is the same as the number of second information categories;

[0143] M can be used to represent the number of target weights, where M can be less than or equal to the number of second information categories;

[0144] G to M can be used to represent determining M target weights from G fusion weights;

[0145] It can be used to represent the attention feature of the t-th target;

[0146] H t It can be used to represent the target weight corresponding to the t-th target attention feature.

[0147] In one embodiment, the step of feature fusion of multiple second features and multiple category features is performed by an object feature extraction model. Specifically, in obtaining the fusion weight corresponding to each attention feature, the hidden layer parameters of the object feature extraction model can be obtained first. Then, based on the hidden layer parameters and each attention feature, the initial weight corresponding to each attention feature is calculated. Finally, the initial weight corresponding to each attention feature is normalized to obtain the fusion weight corresponding to each attention feature. By calculating the initial weights using the hidden layer parameters and attention features, the importance of the second information category can be determined based on the target object's interest in the second information category. This improves the semantic salience of the fusion weights on the attention features, thereby increasing the semantic distance between various target attention features and improving the accuracy of the semantic expression of object features.

[0148] In one embodiment, the object feature extraction model refers to a neural network model used to extract features from the object attributes of a target object based on various second information categories, obtain second features of the object attributes under different second information categories, and fuse multiple second features and multiple category features to obtain object features of the target object adapted to multiple second information categories. The object feature extraction model can be trained using one of the specific neural networks such as Deep & Cross Network (DCN) or CNN; no specific limitation is made here.

[0149] In one embodiment, hidden layer parameters refer to parameters used to adjust the importance of an attention feature relative to the attention feature during the attention feature fusion process. Hidden layer parameters may include one or more coefficients. Furthermore, hidden layer parameters can be a specific numerical value or a vector. Additionally, hidden layer parameters can be preset parameters set individually for each second information category, or preset parameters shared by multiple second information categories, etc., and are not specifically limited here.

[0150] In one embodiment, calculating the initial weight corresponding to each attention feature based on the hidden layer parameters and each attention feature refers to the process of calculating the initial weight for each attention feature based on the hidden layer parameters and the attention feature itself. The initial weight is a vector-like data set.

[0151] In one embodiment, the normalization function used in normalizing the initial weights corresponding to each attention feature can be a softmax function, a sigmoid function, etc., and is not specifically limited here. Alternatively, the initial weights corresponding to multiple attention features can be normalized using the same normalization function, or each attention feature can be normalized using a pre-defined normalization function, and so on, and is not specifically limited here.

[0152] In one embodiment, the hidden layer parameters may include a first hidden layer coefficient and a second hidden layer coefficient. In the process of calculating the initial weight corresponding to each attention feature based on the hidden layer parameters and each attention feature, specifically, for each attention feature, the product of the attention feature and the first hidden layer coefficient can be added to the second hidden layer coefficient to obtain the initial weight corresponding to each attention feature.

[0153] In one embodiment, adding the product of the attention feature and the first hidden layer coefficient to the second hidden layer coefficient means first multiplying the attention feature by the first hidden layer coefficient to obtain the product of the attention feature and the first hidden layer coefficient, and then adding this product to the second hidden layer coefficient. Here, both the first and second hidden layer coefficients can be either a specific numerical value or a preset vector; no specific limitation is made here.

[0154] In one embodiment, the process of normalizing the initial weights corresponding to each attention feature obtained based on the first hidden layer coefficient and the second hidden layer coefficient to obtain the fusion weights corresponding to each attention feature can be represented by the following formula (3).

[0155]

[0156] Among them, H iIt can be used to represent the fusion weight corresponding to the i-th attention feature. It should be noted that multiple attention features correspond one-to-one with multiple second information categories.

[0157] σ() can be used to represent a normalization function;

[0158] W can be used to represent the first hidden layer coefficient;

[0159] b can be used to represent the second hidden layer coefficient;

[0160] It can be used to represent the attention feature corresponding to the i-th second information category;

[0161] It can be used to represent the initial weight corresponding to the i-th second information category.

[0162] In one embodiment, the steps of extracting features from the object attributes of the target object based on each second information category, and the steps of fusing multiple second features and multiple category features, can both be executed by the object feature extraction model. The step of obtaining the first feature of the candidate push information can be executed by the push information feature extraction model. Therefore, in the process of generating category features for multiple second information categories, specifically, the initial features of each second information category can be obtained first. Then, training samples and training labels are obtained. The training samples include object samples and push information samples. The information category to which the push information sample belongs is one of multiple second information categories. Then, the push information feature extraction model is called to extract features from the push information sample to obtain the first sample features. Then, the object feature extraction model is called to extract features from the object sample based on each second information category to obtain the object sample features in different second information categories. The system first uses the second sample features under two information categories. Then, it calls the object feature extraction model to fuse multiple second sample features and multiple initial features to obtain object sample features adapted to multiple second information categories. Next, it calculates the second similarity between the object sample features and the first sample features. Then, it calculates the first loss value based on the initial features of multiple second information categories. Then, it calculates the second loss value based on the first sample features and the initial features of multiple second information categories. Then, it calculates the third loss value based on the second similarity and training labels. Then, it adjusts the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model based on the first, second, and third loss values. Finally, it uses the adjusted initial features of multiple second information categories as the category features of multiple second information categories. In this training process, firstly, since the information category to which the push information sample belongs is one of multiple second information categories, adjusting the initial features can enhance the connection between the object feature extraction model and the push information feature extraction model. Secondly, adjusting the parameters of the object feature extraction model can improve the accuracy of the target object's interest in multiple second information categories as expressed by the object features output by the object feature extraction model. Thirdly, since there is a subordinate relationship between the information category to which the push information sample belongs and multiple second information categories, adjusting the parameters of the push information feature extraction model can enable the first feature output by the push information feature extraction model to form a semantic expression for the information category it belongs to, thereby enhancing the correlation between the first feature and the category features of the information category it belongs to.By enhancing the connection between the object feature extraction model and the push information feature extraction model, improving the accuracy of the target object's interest in multiple second information categories as expressed by the object features, and enhancing the correlation between the first feature and the category features of the information category, the accuracy of the similarity between the object features and the first feature can be improved, which is conducive to improving the accuracy of pushing candidate push information to the target object.

[0163] In one embodiment, the initial features refer to the information attributes used to characterize the second information category during the training phase. At the beginning of the first training round, the initial features can be randomly initialized features or specific initial features that have a certain semantic distance from the initial features of other second information categories. It is important to note that during the training phase, the initial features of the second information category do not have corresponding label features set for calculating the loss value of the initial features.

[0164] In one embodiment, the training label refers to the probability of a push notification sample being a positive example relative to an object sample. Here, a push notification sample refers to a specific push notification as a sample, and an object sample refers to a specific object as a sample. A positive example means that the push notification sample belongs to the push notifications that the object sample is interested in. When the object sample is not interested in the push notification sample, the push notification sample is identified as a negative example relative to the object sample (i.e., the push notification sample belongs to the push notifications that the object sample is not interested in). For example, if the label probability of an object sample being interested in a push notification sample is 0.8, then the training label of the push notification sample is 0.8.

[0165] In one embodiment, calling the object feature extraction model to extract features from the object sample based on each second information category, and obtaining the second sample features of the object sample under different second information categories, refers to the process of extracting features from the object attributes of the object sample for each second information category to obtain the second sample features under that second information category. In other words, apart from the different object attributes input to the object feature extraction model, the process of extracting the second sample features of the object sample under different second information categories is the same as the process of extracting the second features of the target object under different second information categories in the inference stage. This process can be referred to the process in the above-described inference stage embodiment, and will not be elaborated here.

[0166] In one embodiment, invoking an object feature extraction model to fuse multiple second sample features and multiple initial features to obtain object sample features adapted to multiple second information categories refers to the process of first fusing each second sample feature with its corresponding initial feature, and then fusing the resulting multiple features to obtain object sample features. In other words, during feature fusion, except that the second features differ from the second sample features and the category features differ from the initial features, the process of fusing multiple second sample features and multiple initial features is the same as the process of fusing multiple second features and multiple category features in the inference stage. This process can be referred to the process in the inference stage described above, and will not be elaborated here. Here, object sample features refer to object sample attribute features carrying information representing the object sample's degree of interest in each second information category.

[0167] In one embodiment, the process of calculating the second similarity between the object sample features and the first sample features is the same as the process of calculating the first similarity between the object features and the first features, except that the object sample features are different from the object features and the first sample features are different from the first features. This process can refer to the process in the above-described inference stage embodiment, and will not be repeated here.

[0168] In one embodiment, the first loss value refers to the loss value of the semantic distance between the initial features of multiple second information categories. The larger the first loss value, the smaller the semantic distance between the initial features of multiple second information categories, that is, the less obvious the semantic difference between multiple second information categories. Conversely, the smaller the first loss value, the larger the semantic distance between the initial features of multiple second information categories, that is, the more obvious the semantic difference between multiple second information categories.

[0169] In one embodiment, the second loss value refers to the loss value of the clustering result of the first sample features centered on the initial features of the second information category to which they belong. The larger the second loss value, the more difficult it is for the first sample features to cluster with the initial features of the second information category to which they belong, and the less obvious the bias of the object attributes of the object samples for each second information category. Conversely, the smaller the second loss value, the easier it is for the first sample features to cluster with the initial features of the second information category to which they belong, and the more obvious the bias of the object attributes of the object samples for each second information category to which they belong.

[0170] In one embodiment, the third loss value refers to the loss value between the second similarity and the training label. The larger the third loss value, the easier it is to push push information that is not of interest to the target sample. Conversely, the smaller the third loss value, the easier it is to push push information that is of interest to the target sample.

[0171] In one embodiment, during the process of adjusting the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model based on the first loss value, the specific steps involve first adjusting the model parameters of the object feature extraction model and the initial features of multiple second information categories based on the first, second, and third loss values, and then adjusting the model parameters of the push information feature extraction model based on the adjusted initial features of multiple second information categories. The initial features and the model parameters of the object feature extraction model can be adjusted based on one or more of the first, second, and third loss values; no specific limitation is made here. By first adjusting the model parameters of the object feature extraction model and the initial features of multiple second information categories, the differences in the degree of interest in each second information category represented by the object features can be made more apparent. Furthermore, since the model parameters of the push information feature extraction model are based on the adjusted initial features of multiple second information categories, the correlation between the push information feature extraction model and the object feature extraction model can be improved.

[0172] In one embodiment, in the process of calculating the first loss value based on the initial features of multiple second information categories, a first hyperparameter can be obtained first. Then, the first parameter is determined based on the number of second information categories. Next, the first hyperparameter is subtracted from the initial features of each second information category to obtain the first difference for each second information category. Then, the product of the initial features and the first differences of each pair of second information categories is calculated to obtain multiple first products. Then, the positive operation is performed on each of the multiple first products to obtain multiple first results. Finally, the first loss value is calculated based on the first parameter and the multiple first results. By subtracting the first hyperparameter from the initial features of each second information category and calculating the product of the initial features and the first differences of each pair of second information categories, the difference between the initial features of each second information category and the initial features of any other second information category can be better expressed through the first products. This allows the first loss value to more accurately express the semantic distance loss between each second information category and any other second information category, which is beneficial for more accurate adjustment of the initial features.

[0173] In one embodiment, the first hyperparameter can be a preset vector matrix or a preset specific value, which is not limited here.

[0174] In one embodiment, the process of determining the first parameter based on the number of the second information category can be carried out by the following formula (4).

[0175]

[0176] Here, E can be used to represent the first parameter;

[0177] K can be used to represent the number of the second information category.

[0178] In one embodiment, calculating the product of the initial features and the first difference of each pair of second information categories to obtain multiple first products refers to the process of multiplying the product of the first difference of each second information category with the initial features of the next second information category to obtain the first product corresponding to that second information category.

[0179] In one embodiment, the process of calculating the first loss value can be carried out by the following formula (5).

[0180]

[0181] Among them, L i2i It can be used to represent the first loss value;

[0182] k can be used to represent the number of the second information category;

[0183] i and j can be used to indicate the order of the second information category;

[0184] I i It can be used to represent the initial features of the i-th second information category;

[0185] I j It can be used to represent the initial features of the j-th second information category;

[0186] μ can be used to represent the first hyperparameter;

[0187] (I j -μ) can be used to represent the first difference corresponding to the j-th second information category;

[0188] I i ·(I j -μ) can be used to represent the first product corresponding to the j-th second information category;

[0189] [] + This can be used to represent the positive operation. Note that when the first product is less than 0, [] + The result of the operation is 0;

[0190] [I i ·(I j -μ)] + It can be used to represent the first result;

[0191] E can be used to represent the first parameter.

[0192] In one embodiment, during the process of calculating the second loss value based on the first sample features and the initial features of multiple second information categories, the second information category to which the push information sample belongs can be determined first. Then, a first distance is calculated between the first sample features and the initial features of the second information category to which the push information sample belongs. Finally, the second loss value is calculated based on the first distance. By calculating the first distance between the first sample features and the initial features of the second information category to which the push information sample belongs, the first distance can accurately represent the convergence of the first sample features towards the initial features of the second information category to which the push information sample belongs. Thus, calculating the second loss value using the first distance can more accurately represent the loss between the first sample features and the initial features of the second information category to which the push information sample belongs. This is beneficial for accurately adjusting the spatial distribution of the first sample features and the initial features of the second information category to which the push information sample belongs, thereby promoting the mutual learning process between the initial features and the inference information feature extraction model through the second loss value.

[0193] In one embodiment, the push information sample includes an information category identifier. Specifically, in determining the second information category to which the push information sample belongs, the second information category can be queried among multiple second information categories based on the information category identifier. Here, the information category identifier refers to a marker used to identify the information category to which the push information sample belongs.

[0194] In one embodiment, the first distance can be one of various characteristic spatial geometric distances such as Euclidean distance and Manhattan distance, and no specific limitation is made here.

[0195] In one embodiment, the process of calculating the second loss value can be carried out by the following formula (6).

[0196]

[0197] Among them, L a2i It can be used to represent the second loss value;

[0198] A a The first sample feature can be used to represent a sample of push information. It should be noted that the first sample feature is a matrix.

[0199] I[A a This can be used as an initial feature to represent the second information category to which a push notification sample belongs;

[0200] A can be used to represent the number of rows in the matrix representing the features of the first sample;

[0201] 'a' can be used to represent the order of rows in a matrix;

[0202] ||||2 can be used to represent the L2 norm;

[0203] It can be used to represent the first distance.

[0204] In one embodiment, in calculating the third loss value based on the second similarity and training labels, the positive loss when the model predicts correctly and the negative loss when the model predicts incorrectly can be calculated first based on the second similarity and training labels. Then, the third loss value is calculated based on the positive and negative losses. Calculating the third loss value using both positive and negative losses allows it to intuitively reflect the distance between the probability distribution predicted by the model and the true probability distribution. This reflects the gap between the model's classification boundary for push notification samples and the correct boundary represented by the training labels. This helps the model learn the correct classification boundary more quickly, and thus more accurately adjust the object feature extraction model and the push notification feature extraction model.

[0205] In one embodiment, the positive loss refers to the loss value between the second similarity and the training label, and the negative loss refers to the loss value between the probability that the push information sample is not a positive example relative to the object sample and the corresponding sample probability. For example, suppose the training label corresponding to a push information sample is 0.8 and the second similarity is 0.7. Then the probability that the push information sample is not a positive example relative to the object sample is 0.3 (i.e., 1-0.7), and the sample probability that the push information sample is not a positive example relative to the object sample is 0.2 (i.e., 1-0.8). Therefore, the positive loss for this push information sample is the loss value between 0.8 (i.e., the training label) and 0.7 (i.e., the second similarity), and the negative loss is the loss value between 0.2 and 0.3. It should be noted that the positive loss is not obtained by subtracting the second similarity from the training label, and the negative loss is not obtained by subtracting the probability that the push information sample is not a positive example relative to the object sample and the corresponding sample probability.

[0206] In one embodiment, calculating the third loss value based on positive and negative losses refers to the process of calculating the current third loss value by superimposing the positive and negative losses of the current sample pair on the positive and negative losses of multiple sample pairs whose second similarity has already been calculated. Here, a sample pair refers to a combination of sample data consisting of an object sample, a push notification sample corresponding to the object sample, and the training label of the push notification sample. For example, assuming there are two sample pairs, when the second similarity of the first sample pair is calculated, the current third loss value is calculated based on the positive and negative losses of the first sample pair. When the second sample pair is calculated, the current third loss value is calculated by superimposing the positive and negative losses of the second sample pair on the existing third loss value (that is, calculating the current third loss value based on the positive and negative losses of the first and second sample pairs).

[0207] In one embodiment, the process of calculating the third loss value based on the positive loss and the negative loss can be carried out by the following formula (7).

[0208]

[0209] Among them, L main It can be used to represent the third loss value;

[0210] M can represent the number of sample pairs for which the second similarity has been calculated;

[0211] MB can be used to represent the total number of sample pairs with second similarity that need to be calculated during the training phase;

[0212] m can be used to represent the order in which sample information is pushed;

[0213] y m It can be used to represent the second similarity corresponding to the m-th push message sample;

[0214] It can be used to represent the training label corresponding to the m-th push message sample;

[0215] It can be used to represent the positive loss corresponding to the m-th push information sample. It should be noted that when the m-th push information sample belongs to a sample pair for which the second similarity has not been calculated, the positive loss is 0.

[0216] It can be used to represent the negative loss corresponding to the m-th push information sample. It should be noted that when the m-th push information sample belongs to a sample pair for which the second similarity has not been calculated, the negative loss is 0.

[0217] In one embodiment, during the process of adjusting the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model based on the first loss value, the second loss value, and the third loss value, specifically, the second hyperparameter can be obtained first. Then, the second hyperparameter is multiplied by the sum of the first loss value and the second loss value to obtain the corrected loss value. Then, the overall loss value is calculated based on the corrected loss value and the third loss value. Finally, the model parameters of the initial features of multiple second information categories, the push information feature extraction model, and the object feature extraction model are adjusted based on the overall loss value.

[0218] In one embodiment, multiplying the second hyperparameter by the sum of the first loss value and the second loss value means first summing the first loss value and the second loss value, and then multiplying the sum by the second hyperparameter.

[0219] In one embodiment, the process of calculating the total loss value based on the second hyperparameter, the first loss value, the second loss value, and the third loss value can be represented by the following formula (8).

[0220] L = L main +π(L i2i +L a2i (8)

[0221] Where L can be used to represent the total loss value;

[0222] L main It can be used to represent the third loss value;

[0223] π can be used to represent a second hyperparameter, which is a parameter used to adjust the loss contribution of the first loss value and the second loss value.

[0224] L i2i It can be used to represent the first loss value;

[0225] L a2i It can be used to represent the second loss value.

[0226] The following is a specific example to illustrate the reasoning phase of this application: Figure 5 The neural network architecture used in the information push method of this embodiment is shown. Figure 6 It shows Figure 5 The processing principles of the feature extraction models on each side of the neural network architecture. See [link / reference]. Figure 5 and Figure 6For the object feature extraction model on the object side (i.e., the side that inputs object attributes), the object feature extraction model is equipped with a biased network for each second information category (here, the biased network refers to a neural network used to extract features from the object attributes of the target object to obtain the second feature of the object attribute under a second information category), and can call the category features of multiple second information categories. The category features of multiple second information categories can form a semantic vector space. Since the first information category to which the candidate push information belongs after passing through the push information feature extraction model on the information side (i.e., the side that inputs push information) belongs is one of multiple second information categories, in the process of calculating the first similarity, the object feature extraction model on the object side and the push information feature extraction model on the information side can build a connection through this semantic vector space. Taking the push notification feature extraction model's pre-processing of stored push notification features as an example, when determining whether to push a candidate push notification to a target object, firstly, the first feature obtained after processing by the push notification feature extraction model is retrieved from the candidate push notification. Next, the object attributes of the target object are input into the object feature extraction model. This model can extract biased features based on each second information category using a biased network for each category, obtaining second features of the object attributes under different second information categories. Then, the second features of the object attributes under different second information categories are fused with the semantic vector space to obtain attention features of the object attributes under different second information categories. Finally, multiple attention features are weighted and summed to obtain object features of the target object adapted to multiple second information categories. Then, the first similarity between the object features and the first feature is calculated. If the first similarity is not less than a preset similarity threshold, the candidate push notification is pushed to the target object; if the first similarity is less than the preset similarity threshold, the push notification is canceled.

[0227] The following is a specific example to illustrate the training phase of this application: Figure 7 It shows Figure 7 The architecture during the training phase is shown in the image. Figure 7For the object feature extraction model on the object side, the model has an initial bias network for each second information category and can call initial features from multiple second information categories. These initial features can form an initial semantic vector space. Taking the push notification feature extraction model pre-extracting features from stored push notifications as an example, when it is necessary to determine whether to push a push notification sample to an object sample, firstly, the first sample features obtained after processing by the push notification feature extraction model are called. Then, the object attributes of the object sample are input into the object feature extraction model. The model can extract bias features based on each second information category through the bias network of each second information category, obtaining second sample features of the object attributes under different second information categories. Then, the second sample features of the object attributes under different second information categories are fused with the initial semantic vector space to obtain attention sample features of the object attributes under different second information categories. Finally, multiple attention sample features are weighted and summed to obtain object sample features of the object sample adapted to multiple second information categories. Then, the second similarity between the object sample features and the first sample features is calculated. If the second similarity is not less than a preset similarity threshold, the push information sample is pushed to the object sample; if the second similarity is less than the preset similarity threshold, the push information sample is canceled from being pushed to the object sample. At this point, the adjustment process begins. Based on the initial features of multiple second information categories, a first loss value is calculated. Then, based on the first sample features and the initial features of multiple second information categories, a second loss value is calculated. Then, based on the second similarity and training labels, a third loss value is calculated. Then, based on the first loss value, the second loss value, and the third loss value, the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model are adjusted. Finally, the adjusted initial features of multiple second information categories are used as the category features of multiple second information categories. During this adjustment process, by adjusting the initial features, the differences between the attention sample features output by the semantic vector space gradually become apparent. This allows the object attributes of the target object to accurately reflect the bias of each second information category through the semantic vector space. Furthermore, since the information category to which the push information sample belongs is one of the second information categories, when the initial features are adjusted, the features of the push information can be clustered towards the features of its second information category. In this way, the association between the object feature extraction model on the object side and the push information feature extraction model on the information side, constructed through this semantic vector space, can be gradually strengthened during the training phase. This allows the semantic vector space to improve the accuracy of pushing the candidate push information to the target object during the inference phase.

[0228] The following examples illustrate the application scenarios of the embodiments of this application.

[0229] It should be noted that the information push method provided in this application embodiment can be applied to different application scenarios such as video recommendation for video scenarios, private radio music push for music scenarios, and commercial information push for cloud social application scenarios. The following description will take video recommendation for video scenarios, private radio music push for music scenarios, and commercial information push for cloud social application scenarios as examples.

[0230] Scene 1

[0231] The information push method provided in this application can be applied to video recommendation in video scenarios. For example, for some video software, after being opened and run by a target user, the video software interface can recommend videos stored on the video software vendor's server to the target user. If the recommended videos match the target user's interests, it can help extend the time the target user spends using the video software, thereby increasing the target user's stickiness to the video software. Therefore, the information push method provided in this application can be used to push videos to the target user.

[0232] Assume the video software vendor uses a single server, and the server pre-stores multiple push videos and the first feature of each push video. When the video software is triggered and opened by the target object, the server first obtains the first feature of the candidate push video, where the candidate push video belongs to a first video category. Next, it obtains category features of multiple pre-trained second video categories, where the first video category is one of the multiple second video categories. Then, it extracts features from the target object's attributes based on each second video category, obtaining second features of the object attributes under different second video categories. Next, it fuses the multiple second features and multiple category features to obtain object features of the target object adapted to multiple second video categories. Finally, it calculates the first similarity between the object features and the first features. If the first similarity is not less than a preset similarity threshold, the candidate push video is pushed to the target object.

[0233] Scene 2

[0234] The information push method provided in this application can be applied to music push for private radio settings in music scenarios. For example, some music software has a private radio function. When a target user runs the music software and enters the private radio interface, this function allows the music software to play music inferred to be of interest to the target user one by one. For these music software programs, if the music played to the target user matches the target user's interests, it is beneficial to extend the time the target user uses the music software, thereby increasing the target user's user stickiness to the music software. Therefore, the information push method provided in this application can be used to push music to the target user based on the music style that the target user is interested in.

[0235] Assume the music software vendor uses a single server, and the server pre-stores multiple recommended music tracks and a first feature for each track. When the private radio function is enabled, the server first obtains the first feature of the candidate recommended music tracks, indicating that the tracks belong to a first music style category. Next, it obtains category features for multiple pre-trained second music style categories, where the first music style category is one of several second music style categories. Then, it extracts features from the target object's attributes based on each second music style category, obtaining second features for the object attributes under different second music style categories. Finally, it fuses these second features and category features to obtain object features for the target object that are adapted to multiple second music style categories. Finally, it calculates the first similarity between the object features and the first features. If the first similarity is not less than a preset similarity threshold, the candidate recommended music track is pushed to the target object.

[0236] Scene 3

[0237] The information push method provided in this application can be applied to commercial information push in cloud social application scenarios. For example, some cloud social application vendors offer commercial information push services, allowing other vendors to promote commercial information on their cloud social applications. However, users of cloud social applications are generally interested in commercial information from specific industries. When randomly pushing commercial information to users, there is a certain probability that commercial information from industries the user is not interested in will be pushed to them. This may negatively impact the effectiveness of the commercial information push. Therefore, the information push method provided in this application can push commercial information to the target audience based on the industries the target audience is interested in.

[0238] Assuming the cloud social application vendor uses a single server, and the server pre-stores multiple commercial push notifications and the first feature of each commercial push notification, when a target object runs the cloud social application, the server first obtains the first feature of the candidate commercial push notification, where the candidate commercial push notification belongs to a first industry category. Next, it obtains category features from multiple pre-trained second industry categories, where the first industry category is one of the multiple second industry categories. Then, it extracts features from the target object's attributes based on each second industry category, obtaining second features of the object attributes under different second industry categories. Finally, it fuses the multiple second features and multiple category features to obtain object features of the target object adapted to multiple second industry categories. Then, it calculates the first similarity between the object features and the first features. If the first similarity is not less than a preset similarity threshold, the candidate commercial push notification is pushed to the target object.

[0239] Reference Figure 8 This application also discloses an information push device 800, which can implement the information push method in the preceding embodiments. The information push device 800 may include:

[0240] The first feature acquisition unit 810 can be used to acquire the first feature of the candidate push information, which belongs to the first information category.

[0241] The category feature acquisition unit 820 can be used to acquire category features of multiple second information categories obtained through pre-training, wherein the first information category is one of the multiple second information categories;

[0242] The second feature extraction unit 830 can be used to extract features of the object attributes of the target object based on each second information category, and obtain the second features of the object attributes under different second information categories.

[0243] The feature fusion unit 840 can be used to fuse multiple second features and multiple category features to obtain object features of the target object that are adapted to multiple second information categories.

[0244] The information push unit 850 can be used to calculate the first similarity between the object features and the first feature. When the first similarity is not less than a preset similarity threshold, the candidate push information is pushed to the target object.

[0245] In one embodiment, the feature fusion unit 840 may specifically be used for:

[0246] Attention is calculated for each second feature and its corresponding category feature to obtain the attention features of object attributes under different second information categories.

[0247] By weighted summation of multiple attention features, object features that fit multiple second information categories for the target object are obtained.

[0248] In one embodiment, the feature fusion unit 840 may specifically be used for:

[0249] For each second feature, the second feature is used as the key vector and value vector respectively, and the corresponding category feature is used as the query vector. Attention is calculated to obtain the attention vector corresponding to each second feature.

[0250] Based on the attention vectors corresponding to all the second features, the attention features of object attributes under different second information categories are obtained.

[0251] In one embodiment, the feature fusion unit 840 may specifically be used for:

[0252] Obtain the fusion weights corresponding to each attention feature;

[0253] Multiple target weights are determined from all fusion weights, and multiple target attention features are determined from all attention features based on the multiple target weights;

[0254] The target attention features and their corresponding target weights are weighted and summed to obtain the object features of the target object that are adapted to multiple second information categories.

[0255] In one embodiment, the step of feature fusion of multiple second features and multiple category features is performed by an object feature extraction model; the feature fusion unit 840 can specifically be used for:

[0256] Obtain the hidden layer parameters of the object feature extraction model;

[0257] Based on the hidden layer parameters and each attention feature, the initial weights corresponding to each attention feature are calculated.

[0258] The initial weights corresponding to each attention feature are normalized to obtain the fusion weights corresponding to each attention feature.

[0259] In one embodiment, the hidden layer parameters may include a first hidden layer coefficient and a second hidden layer coefficient; the feature fusion unit 840 may specifically be used for:

[0260] For each attention feature, the product of the attention feature and the first hidden layer coefficient is added to the second hidden layer coefficient to obtain the initial weight corresponding to each attention feature.

[0261] In one embodiment, the steps of extracting features from the object attributes of the target object based on each second information category, and the steps of fusing multiple second features and multiple category features, are both performed by the object feature extraction model; the step of extracting features from the candidate push information is performed by the push information feature extraction model.

[0262] The information push device also includes a category feature training unit, which can be used for:

[0263] Obtain the initial features for each second information category;

[0264] Obtain training samples and training labels. The training samples include object samples and push information samples. The information category to which the push information samples belong is one of multiple second information categories.

[0265] The push notification feature extraction model is invoked to extract features from the push notification sample, thus obtaining the first sample features.

[0266] The object feature extraction model is invoked to extract features from the object sample based on each second information category, thereby obtaining the second sample features of the object sample under different second information categories.

[0267] The object feature extraction model is invoked to fuse multiple second sample features and multiple initial features to obtain object sample features that are adapted to multiple second information categories.

[0268] Calculate the second similarity between the object sample features and the first sample features;

[0269] The first loss value is calculated based on the initial features of multiple second information categories;

[0270] The second loss value is calculated based on the features of the first sample and the initial features of multiple second information categories;

[0271] The third loss value is calculated based on the second similarity and the training labels;

[0272] Based on the first loss value, the second loss value, and the third loss value, the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model are adjusted.

[0273] The adjusted initial features of multiple second information categories will be used as category features for multiple second information categories.

[0274] In one embodiment, the category feature training unit can specifically be used for:

[0275] Obtain the first hyperparameter;

[0276] The first parameter is determined based on the quantity of the second information category;

[0277] Subtract the first hyperparameter from the initial features of each second information category to obtain the first difference of each second information category;

[0278] Calculate the product of the initial features and the first difference of each pair of second information categories to obtain multiple first products;

[0279] By performing positive operations on multiple first products, multiple first results are obtained;

[0280] The first loss value is calculated based on the first parameter and multiple first results.

[0281] In one embodiment, the category feature training unit can specifically be used for:

[0282] Determine the second information category to which the push notification sample belongs;

[0283] Calculate the first distance between the features of the first sample and the initial features of the second information category to which the push information sample belongs;

[0284] The second loss value is calculated based on the first distance.

[0285] In one embodiment, the category feature training unit can specifically be used for:

[0286] Based on the second similarity and training labels, the positive loss when the model predicts correctly and the negative loss when the model predicts incorrectly are calculated.

[0287] The third loss value is calculated based on the positive and negative losses.

[0288] In one embodiment, the category feature training unit can specifically be used for:

[0289] Obtain the second hyperparameter;

[0290] The corrected loss value is obtained by multiplying the second hyperparameter by the sum of the first and second loss values.

[0291] The total loss value is calculated based on the corrected loss value and the third loss value;

[0292] Based on the overall loss value, the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model are adjusted.

[0293] It should be noted that since the information push device 800 of this embodiment can implement the information processing method for the communication group as in the previous embodiment, the information push device 800 of this embodiment and the information push method of the previous embodiment have the same technical principle and the same beneficial effect. In order to avoid repetition, it will not be described again here.

[0294] Reference Figure 9 This application also discloses an electronic device 900, which includes:

[0295] At least one processor 910;

[0296] At least one memory 920 is used to store at least one program;

[0297] When at least one program is executed by at least one processor 910, the information push method described above is implemented.

[0298] This application also discloses a computer-readable storage medium storing a processor-executable computer program, which, when executed by a processor, is used to implement the information push method described above.

[0299] This application also discloses a computer program product, including a computer program or computer instructions, which are stored in a computer-readable storage medium. The processor of the electronic device reads the computer program or computer instructions from the computer-readable storage medium and executes the computer program or computer instructions, causing the electronic device to perform the information push method described above.

[0300] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatuses.

[0301] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0302] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or units, and may be electrical, mechanical, or other forms.

[0303] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0304] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0305] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0306] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0307] The above provides a detailed description of the preferred embodiments of this application. However, this application is not limited to the above-described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. An information push method, characterized in that, include: Obtain the first feature of the candidate push information, wherein the candidate push information belongs to the first information category; Obtain category features of multiple second information categories obtained from pre-training, wherein the first information category is one of the multiple second information categories; The object attributes of the target object are extracted based on each of the second information categories to obtain the second features of the object attributes under different second information categories; By fusing multiple second features and multiple category features, the target object is obtained as an object feature that adapts to multiple second information categories; Calculate the first similarity between the object feature and the first feature. If the first similarity is not less than a preset similarity threshold, push the candidate push information to the target object.

2. The method according to claim 1, characterized in that, The feature fusion of multiple second features and multiple category features to obtain object features of the target object that are adapted to multiple second information categories includes: Attention is calculated for each second feature and the corresponding category feature to obtain the attention features of the object attribute under different second information categories; The multiple attention features are weighted and summed to obtain the object features of the target object that are adapted to the second information category.

3. The method according to claim 2, characterized in that, The step of performing attention calculation on each second feature and the corresponding category feature to obtain the attention features of the object attribute under different second information categories includes: For each of the second features, the second feature is used as a key vector and a value vector, respectively, and the corresponding category feature is used as a query vector. Attention is calculated to obtain the attention vector corresponding to each of the second features. Based on the attention vectors corresponding to all the second features, the attention features of the object attributes under different second information categories are obtained.

4. The method according to claim 2, characterized in that, The step of weighted summing of multiple attention features to obtain object features of the target object that are adapted to multiple second information categories includes: Obtain the fusion weights corresponding to each attention feature; Multiple target weights are determined from all the fusion weights, and multiple corresponding target attention features are determined from all the attention features based on the multiple target weights; The target attention features and their corresponding target weights are weighted and summed to obtain the object features of the target object that are adapted to the second information category.

5. The method according to claim 4, characterized in that, The step of feature fusion of multiple second features and multiple category features is performed by the object feature extraction model; The step of obtaining the fusion weights corresponding to each attention feature includes: Obtain the hidden layer parameters of the object feature extraction model; Based on the hidden layer parameters and each attention feature, the initial weight corresponding to each attention feature is calculated; The initial weights corresponding to each attention feature are normalized to obtain the fusion weights corresponding to each attention feature.

6. The method according to claim 5, characterized in that, The hidden layer parameters include a first hidden layer coefficient and a second hidden layer coefficient; The step of calculating the initial weight corresponding to each attention feature based on the hidden layer parameters and each attention feature includes: For each attention feature, the product of the attention feature and the first hidden layer coefficient is added to the second hidden layer coefficient to obtain the initial weight corresponding to each attention feature.

7. The method according to claim 1, characterized in that, The steps of extracting features from the object attributes of the target object based on each of the second information categories, and the steps of fusing multiple second features and multiple category features, are both performed by the object feature extraction model; the step of obtaining the first feature of the candidate push information is performed by the push information feature extraction model. The category features of multiple second information categories are obtained through training using the following steps: Obtain the initial features for each of the second information categories; Obtain training samples and training labels. The training samples include object samples and push information samples. The information category to which the push information samples belong is one of multiple second information categories. The push information feature extraction model is invoked to extract features from the push information sample, thereby obtaining the first sample features; The object feature extraction model is invoked to extract features from the object sample based on each of the second information categories, thereby obtaining the second sample features of the object sample under different second information categories. The object feature extraction model is invoked to perform feature fusion on multiple second sample features and multiple initial features to obtain object sample features of the object sample that are adapted to multiple second information categories; Calculate the second similarity between the object sample features and the first sample features; A first loss value is calculated based on the initial features of multiple second information categories; The second loss value is calculated based on the first sample features and the initial features of multiple second information categories; The third loss value is calculated based on the second similarity and the training labels; Based on the first loss value, the second loss value, and the third loss value, the initial features of the multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model are adjusted. The adjusted initial features of the multiple second information categories are used as the category features of the multiple second information categories.

8. The method according to claim 7, characterized in that, The step of calculating the first loss value based on the initial features of multiple second information categories includes: Obtain the first hyperparameter; The first parameter is determined based on the quantity of the second information category; Subtract the first hyperparameter from the initial feature of each second information category to obtain the first difference of each second information category; Calculate the product of the initial features of each pair of the second information categories and the first difference to obtain multiple first products; By performing a positive operation on each of the multiple first products, multiple first results are obtained; Based on the first parameter and multiple first results, a first loss value is calculated.

9. The method according to claim 7, characterized in that, The step of calculating the second loss value based on the first sample features and the initial features of multiple second information categories includes: Determine the second information category to which the push information sample belongs; Calculate the first distance between the first sample feature and the initial feature of the second information category to which the push information sample belongs; The second loss value is calculated based on the first distance.

10. The method according to claim 7, characterized in that, The step of calculating the third loss value based on the second similarity and the training labels includes: Based on the second similarity and the training labels, the positive loss when the model predicts correctly and the negative loss when the model predicts incorrectly are calculated. The third loss value is calculated based on the positive loss and the negative loss.

11. The method according to claim 7, characterized in that, The step of adjusting the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model based on the first loss value, the second loss value, and the third loss value includes: Obtain the second hyperparameter; The corrected loss value is obtained by multiplying the second hyperparameter by the sum of the first loss value and the second loss value. The total loss value is calculated based on the corrected loss value and the third loss value. Based on the overall loss value, the initial features of multiple second information categories, the model parameters of the push information feature extraction model, and the model parameters of the object feature extraction model are adjusted.

12. An information push device, characterized in that, include: The first feature acquisition unit is used to acquire the first feature of the candidate push information, wherein the candidate push information belongs to the first information category; The category feature acquisition unit is used to acquire category features of multiple second information categories obtained through pre-training, wherein the first information category is one of the multiple second information categories; The second feature extraction unit is used to extract features of the object attributes of the target object based on each of the second information categories, so as to obtain the second features of the object attributes under different second information categories. The feature fusion unit is used to perform feature fusion on multiple second features and multiple category features to obtain object features of the target object that are adapted to multiple second information categories; An information push unit is used to calculate a first similarity between the object feature and the first feature, and when the first similarity is not less than a preset similarity threshold, push the candidate push information to the target object.

13. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; The information push method as described in any one of claims 1 to 11 is implemented when at least one of the programs is executed by at least one of the processors.

14. A computer-readable storage medium, characterized in that, It stores a processor-executable computer program, which, when executed by the processor, is used to implement the information push method as described in any one of claims 1 to 11.

15. A computer program product, comprising a computer program or computer instructions, characterized in that, The computer program or the computer instructions are stored in a computer-readable storage medium. The processor of the electronic device reads the computer program or the computer instructions from the computer-readable storage medium and executes the computer program or the computer instructions, causing the electronic device to perform the information push method as described in any one of claims 1 to 11.

Images