Information push model processing method and device, computer equipment and storage medium

Abstract

The invention relates to an information push model processing method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring first information, and determining a plurality of pieces of second information belonging to the same information category as the first information; determining a plurality of receiver users of the second information, and determining a user embedding vector of each receiver user of the second information; determining an information embedding vector of each piece of second information, and performing model pre-training according to each information embedding vector and each user embedding vector to obtain a pre-trained information pushing model; determining a receiver user of the first information, and determining a user embedding vector of the receiver user of the first information; determining an information embedding vector of the first information; and based on each user embedding vector and the information embedding vector of the first information, retraining the pre-trained information pushing model to obtain a trained information pushing model. By adopting the method, the audience expansion accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technologies, and particularly to a method, apparatus, computer device, computer-readable storage medium, and computer program product for processing an information push model. Background Art

[0002] With the development of computer technologies, information push technologies have emerged, through which information can be pushed to different users. In information push, it is often necessary to perform Audience Expansion based on historical push data. Audience Expansion aims to expand the scale of the target audience of information to increase the opportunities for information exposure and conversion. Audience Expansion often recalls users who have clicked on the pushed information through a model, screens out similar groups of people based on the users' data, and uses the screened-out groups of people as the recipients of the information.

[0003] However, for some information with few push times or few target audiences, the sample data used in model training is less, resulting in a weak Audience Expansion ability of the trained model. Summary of the Invention

[0004] Based on this, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for processing an information push model that can improve accuracy in view of the above technical problems.

[0005] In a first aspect, this application provides a method for processing an information push model. The method includes:

[0006] Obtain a first piece of information, and determine a plurality of second pieces of information that belong to the same information category as the first piece of information;

[0007] Determine the recipient users of the plurality of second pieces of information respectively, and determine the user embedding vectors corresponding to the recipient users of each second piece of information;

[0008] Determine the information embedding vector of each second piece of information, and perform model pre-training according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model;

[0009] Determine the recipient user of the first piece of information, and determine the user embedding vector corresponding to the recipient user of the first piece of information;

[0010] Determine the information embedding vector of the first piece of information;

[0011] Retrain the pre-trained information push model based on the user embedding vector corresponding to the recipient user of each of the second pieces of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, to obtain a trained information push model.

[0012] In a second aspect, the present application also provides a processing device for an information push model. The device includes:

[0013] An information acquisition module, configured to acquire a first piece of information and determine a plurality of second pieces of information belonging to the same information category as the first piece of information;

[0014] A first determination module, configured to determine the recipient user of each of the plurality of second pieces of information and determine the user embedding vector corresponding to the recipient user of each of the second pieces of information;

[0015] A pre-training module, configured to determine the information embedding vector of each of the second pieces of information, and perform model pre-training according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model;

[0016] A second determination module, further configured to determine the recipient user of the first piece of information and determine the user embedding vector corresponding to the recipient user of the first piece of information;

[0017] A retraining module, configured to determine the information embedding vector of the first piece of information; retrain the pre-trained information push model based on the user embedding vector corresponding to the recipient user of each of the second pieces of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, to obtain a trained information push model.

[0018] In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0019] Acquire a first piece of information and determine a plurality of second pieces of information belonging to the same information category as the first piece of information;

[0020] Determine the recipient user of each of the plurality of second pieces of information and determine the user embedding vector corresponding to the recipient user of each of the second pieces of information;

[0021] Determine the information embedding vector of each of the second pieces of information, and perform model pre-training according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model;

[0022] Determine the recipient user of the first information, and determine the user embedding vector corresponding to the recipient user of the first information;

[0023] Determine the information embedding vector of the first information;

[0024] Based on the user embedding vector corresponding to the recipient user of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, retrain the pre-trained information push model to obtain a trained information push model.

[0025] Fourthly, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the following steps are implemented:

[0026] Obtain the first information, and determine a plurality of second information belonging to the same information category as the first information;

[0027] Determine the recipient user of each of the plurality of second information, and determine the user embedding vector corresponding to the recipient user of each second information;

[0028] Determine the information embedding vector of each of the second information, and perform model pre-training according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model;

[0029] Determine the recipient user of the first information, and determine the user embedding vector corresponding to the recipient user of the first information;

[0030] Determine the information embedding vector of the first information;

[0031] Based on the user embedding vector corresponding to the recipient user of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, retrain the pre-trained information push model to obtain a trained information push model.

[0032] Fifthly, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0033] Obtain the first information, and determine a plurality of second information belonging to the same information category as the first information;

[0034] Determine the recipient user of each of the plurality of second information, and determine the user embedding vector corresponding to the recipient user of each second information;

[0035] Determine the information embedding vector of each of the second pieces of information, and perform model pre-training based on each information embedding vector and each user embedding vector to obtain a pre-trained information push model;

[0036] Determine the recipient user of the first piece of information, and determine the user embedding vector corresponding to the recipient user of the first piece of information;

[0037] Determine the information embedding vector of the first piece of information;

[0038] Based on the user embedding vectors corresponding to the recipient users of each of the second pieces of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, retrain the pre-trained information push model to obtain a trained information push model.

[0039] The above-mentioned processing method, device, computer device, storage medium, and computer program product of the information push model, by obtaining the first piece of information and determining a plurality of second pieces of information belonging to the same information category as the first piece of information, and determining the recipient users of the plurality of second pieces of information respectively, can obtain the historical push data of other similar information as a supplement when the historical push data of the first piece of information is insufficient, so as to obtain more training samples. Determine the user embedding vector corresponding to the recipient user of each second piece of information, determine the information embedding vector of each second piece of information, and perform model pre-training based on each information embedding vector and each user embedding vector, so that all the same user embedding vectors are shared in each pre-training, enabling the model to obtain the commonalities of the recipient users of similar information during training. Obtain a pre-trained information push model, determine the recipient user of the first piece of information, and determine the user embedding vector corresponding to the recipient user of the first piece of information, determine the information embedding vector of the first piece of information, and perform targeted retraining on the pre-trained information push model based on the user embedding vectors corresponding to the recipient users of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, so that the trained information push model can not only obtain the ability to extract the commonalities between other users and the historical recipient users of the first piece of information, but also obtain the ability to extract the characteristics related to the first piece of information, thereby obtaining an information push model dedicated to expanding the recipient users of the first piece of information, realizing the effective expansion of the target audience of the first piece of information, and helping to improve the exposure rate and conversion rate of the first piece of information. Brief Description of the Drawings

[0040] Figure 1 It is an application environment diagram of the processing method of the information push model in an embodiment;

[0041] Figure 2 It is a flowchart of the processing method of the information push model in an embodiment;

[0042] Figure 3 Schematic diagram for audience expansion of the trained information push model in one embodiment;

[0043] Figure 4 Flow schematic diagram of the processing method of the information push model in another embodiment;

[0044] Figure 5 Schematic diagram of the structure of the feature selection gate in one embodiment;

[0045] Figure 6 Schematic diagram of the processing process of the feature selection gate in one embodiment;

[0046] Figure 7 Flow schematic diagram of the steps for obtaining the hidden layer feature representation corresponding to the second information in one embodiment;

[0047] Figure 8 Flow schematic diagram of the steps for obtaining the trained information push model in one embodiment;

[0048] Figure 9 Schematic diagram of the structure of the adaptation layer in one embodiment;

[0049] Figure 10 Schematic diagram of calculating the retraining loss in the retraining stage in one embodiment;

[0050] Figure 11 Schematic diagram for audience expansion of the trained information push model in another embodiment;

[0051] Figure 12 Schematic diagram of the architecture of the information push model in another embodiment;

[0052] Figure 13 Schematic diagram of the replaceable structure of the information push model in one embodiment;

[0053] Figure 14 Structure block diagram of the processing device of the information push model in one embodiment;

[0054] Figure 15 Internal structure diagram of a computer device in one embodiment. Detailed implementation manners

[0055] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

[0056] The embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, assisted driving, etc. For example, it can be applied to the field of artificial intelligence (AI) technology. Among them, artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable the machines to have the functions of perception, reasoning, and decision-making. The solution provided by the embodiments of the present application relates to a processing method for an information push model in artificial intelligence, which will be specifically described in the following embodiments.

[0057] The processing method for the information push model provided by the embodiments of the present application can be applied to an application environment as Figure 1 shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed in the cloud or on other servers. Both the terminal 102 and the server 104 can independently execute the processing method for the information push model provided in the embodiments of the present application. The terminal 102 and the server 104 can also be used in cooperation to execute the processing method for the information push model provided in the embodiments of the present application. When the terminal 102 and the server 104 are used in cooperation to execute the processing method for the information push model provided in the embodiments of the present application, the terminal 102 obtains the first information, and determines a plurality of second information belonging to the same information category as the first information from the server 104. The terminal 102 determines the recipient users of each of the plurality of second information, and determines the user embedding vector corresponding to the recipient user of each second information. The terminal 102 determines the information embedding vector of each second information, performs model pre-training based on each information embedding vector and each user embedding vector, and obtains a pre-trained information push model. The terminal 102 determines the recipient user of the first information, and determines the user embedding vector corresponding to the recipient user of the first information. The terminal 102 determines the information embedding vector of the first information, and retrains the pre-trained information push model based on the user embedding vector corresponding to the recipient user of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, and obtains a trained information push model. The trained information push model can run on the terminal 102 or the server 104.

[0058] Among them, the terminal 102 can be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0059] In one embodiment, as Figure 2 shown, a processing method for an information push model is provided. Taking this method applied to Figure 1 the computer device in Figure 1 (the computer device can be the terminal or the server in

[0060] Step S202: Obtain the first information and determine multiple second information belonging to the same information category as the first information.

[0061] Among them, the first information and the multiple second information are training samples for model training. The first information and the second information belong to the same information category.

[0062] The information category refers to the classification of information. Different information can be divided into different information categories according to the field. Multiple information under the same field can be further divided into different information categories. For example, in the financial field, it can be divided into information for recommending financial products, information for recommending credit cards, etc., but not limited to this. Different information for recommending credit cards is different, but these information belong to the same information category.

[0063] Specifically, the computer device obtains the first information and determines the information category of the first information. The computer device obtains multiple second information belonging to the same information category as the first information.

[0064] In this embodiment, the computer device obtains the first information, determines the historical push platform of the first information, and obtains multiple second information belonging to the same information category as the first information from the historical push platform.

[0065] Step S204: Determine the recipient users of the multiple second information respectively, and determine the user embedding vector corresponding to the recipient user of each second information.

[0066] Among them, the recipient user of the second information refers to the user who receives the second information.

[0067] Specifically, the computer device can determine the recipient user of each second piece of information among the multiple second pieces of information. Further, the computer device can determine the recipient users of the multiple second pieces of information within a historical time period. Here, the historical time period refers to a past time period, which can be measured in units such as days, weeks, months, etc. For example, starting from the current moment, the past 24 hours can be used as a historical time period, or a month can be used as a historical time period, etc., but not limited to this.

[0068] The computer device obtains the recipient feature information of each recipient user, determines the embedding vector corresponding to each recipient feature information, and obtains the user embedding vector corresponding to each recipient user.

[0069] In this embodiment, the first piece of information and the multiple second pieces of information are both information pushed in the past, that is, the first piece of information and the multiple second pieces of information are both information that has been pushed. The recipient user of the first piece of information refers to the user who has received the first piece of information. The recipient user of the second piece of information refers to the user who has received the second piece of information.

[0070] The recipient user of the first piece of information can be a user who has received and viewed the first piece of information. The recipient user of the second piece of information can be a user who has received and viewed the second piece of information.

[0071] Step S206: Determine the information embedding vector of each second piece of information, and perform model pre-training based on each information embedding vector and each user embedding vector to obtain a pre-trained information push model.

[0072] Specifically, the computer device determines the information embedding vector of each second piece of information. Further, the computer device determines the modal information corresponding to each second piece of information in at least one preset modality, and determines the modal embedding vector of the modal information. The modal embedding vector of the modal information corresponding to the second piece of information is used as the information embedding vector of the second piece of information.

[0073] Here, the preset modality refers to the visual presentation method of the second piece of information, which can specifically be an image, text, audio, or video. The modal information refers to the information of the second piece of information in a certain preset modality. For example, the modal information of the second piece of information corresponding to an image refers to the image in the second piece of information. The modal information of the second piece of information corresponding to text refers to the text information in the second piece of information. The modal information of the second piece of information corresponding to audio refers to the audio information in the second piece of information. The modal information of the second piece of information corresponding to video refers to the video information in the second piece of information.

[0074] The computer device performs model pre-training on the information push model to be trained according to the information embedding vector of each second piece of information and the user embedding vector corresponding to the recipient user of each second piece of information, so as to adjust the weight parameters of the information push model to be trained, and stop when the pre-training stop condition is met, and obtain a pre-trained information push model.

[0075] In one embodiment, the information push model to be trained has first-class weight parameters and second-class weight parameters. The computer device keeps the second-class weight parameters of the information push model to be trained unchanged, and pre-trains the information push model to be trained according to the information embedding vectors of each second piece of information and each user embedding vector, so as to adjust the first-class weight parameters and obtain a pre-trained information push model. The pre-trained information push model has the adjusted first-class weight parameters and the unchanged second-class weight parameters.

[0076] In one embodiment, the computer device performs model pre-training according to each information embedding vector and each user embedding vector, and calculates the pre-training loss of the model pre-training. The weight parameters of the information push model to be trained are adjusted based on the pre-training loss. Further, the first-class weight parameters of the information push model to be trained are adjusted based on the pre-training loss.

[0077] Step S208: Determine the recipient user of the first piece of information, and determine the user embedding vector corresponding to the recipient user of the first piece of information.

[0078] Herein, the recipient user of the first piece of information refers to the user who receives the first piece of information.

[0079] Specifically, the computer device can determine the recipient user of the first piece of information. Further, the computer device can determine the recipient user of the first piece of information within a historical time period.

[0080] The computer device obtains the recipient feature information of each recipient user of the first piece of information, determines the embedding vector corresponding to each recipient feature information, and obtains the user embedding vector corresponding to each recipient user of the first piece of information.

[0081] In one embodiment, the computer device can determine the recipient user of the first piece of information within the first historical time period, and determine the recipient user of each second piece of information within the second historical time period. For example, determine the recipient user corresponding to December 12th for the first piece of information, and determine the recipient user corresponding to December 10th for each second piece of information.

[0082] Step S210: Determine the information embedding vector of the first piece of information.

[0083] Specifically, the computer device determines the information embedding vector of the first piece of information. Further, the computer device determines the modal information corresponding to the first piece of information for at least one preset modality, and determines the modal embedding vector of the modal information. The modal embedding vector of the modal information corresponding to the first piece of information is used as the information embedding vector of the first piece of information.

[0084] Among them, the preset modality refers to the visual presentation method of the first information, which can specifically be an image, text, audio, or video. The modality information refers to the information of the first information in a certain preset modality. For example, the modality information corresponding to the image of the first information refers to the image in the first information. The modality information corresponding to the text of the first information refers to the text information in the first information. The modality information corresponding to the audio of the first information refers to the audio information in the first information. The modality information corresponding to the video of the first information refers to the video information in the first information.

[0085] Step S212: Based on the user embedding vectors corresponding to the recipient users of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, retrain the pre-trained information push model to obtain a trained information push model.

[0086] Among them, retraining means retraining the information push model obtained through pre-training. The trained information push model is used to expand the audience of the recipient user of the first information to obtain more recipient users, as Figure 3 shown.

[0087] In this embodiment, pre-training can be offline training, and retraining can be online training.

[0088] Specifically, based on the user embedding vectors corresponding to the recipient users of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, retrain the pre-trained information push model to adjust the weight parameters of the pre-trained information push model and obtain a trained information push model. Among them, the weight parameters adjusted in model pre-training are different from or can be the same as those adjusted in retraining.

[0089] For example, the weight parameters include the first type of weight parameters and the second type of weight parameters, and the first type of weight parameters is different from the second type of weight parameters. Adjust the first type of weight parameters of the information push model during model pre-training, and adjust the second type of weight parameters of the information push model during model retraining.

[0090] When the weight parameters adjusted in model pre-training are the same as those adjusted in retraining, it can be that during retraining, the weight parameters obtained after pre-training adjustment are further adjusted.

[0091] In one embodiment, the computer retrains a pre-trained information push model based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, and calculates the retraining loss of the model retraining. The weight parameters of the pre-trained information push model are adjusted based on the retraining loss. Further, the second type of weight parameters of the pre-trained information push model are adjusted based on the retraining loss.

[0092] In this embodiment, the computer device obtains the recipient feature information of each recipient user of each second piece of information, and determines the user embedding vector corresponding to each recipient user. A shared feature set is generated based on the user embedding vectors corresponding to each recipient user, and the shared feature set includes the recipient features of each recipient user of each second piece of information and the corresponding user embedding vectors. As Figure 4 shown, the pre-trained information push model is obtained by using the user embedding vectors in the shared feature set and the information embedding vectors of each second piece of information. The trained information push model is obtained by using the user embedding vectors in the shared feature set and the information embedding vector of the first piece of information for retraining.

[0093] In the above information push model processing method, obtaining the first piece of information, determining a plurality of second pieces of information belonging to the same information category as the first piece of information, and determining the recipient users of the plurality of second pieces of information can obtain the historical push data of other similar information as a supplement when the historical push data of the first piece of information is insufficient, so as to obtain more training samples. Determining the user embedding vector corresponding to the recipient user of each second piece of information, determining the information embedding vector of each second piece of information, and performing model pre-training according to each information embedding vector and each user embedding vector, so that all the same user embedding vectors are shared in each pre-training, enabling the model to obtain the commonalities of the recipient users extracting similar information during training. Obtaining the pre-trained information push model, determining the recipient user of the first piece of information, and determining the user embedding vector corresponding to the recipient user of the first piece of information, determining the information embedding vector of the first piece of information, and performing targeted retraining on the pre-trained information push model based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, so that the trained information push model can not only obtain the ability to extract the commonalities between other users and the historical recipient users of the first piece of information, but also obtain the ability to extract the characteristics related to the first piece of information, thereby obtaining an information push model dedicated to expanding the recipient users of the first piece of information, realizing the effective expansion of the target audience of the first piece of information, and helping to improve the exposure rate and conversion rate of the first piece of information.

[0094] In one embodiment, model pre-training is performed based on each information embedding vector and each user embedding vector to obtain a pre-trained information push model, including: performing model pre-training based on each information embedding vector and each user embedding vector to obtain a pre-trained information push model and the hidden layer feature representation corresponding to each second information;

[0095] Based on the user embedding vector corresponding to the recipient user of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, the pre-trained information push model is re-trained to obtain a trained information push model, including: based on the user embedding vector corresponding to the recipient user of each second information, the user embedding vector corresponding to the recipient user of the first information, the information embedding vector of the first information, and each hidden layer feature representation, the pre-trained information push model is re-trained to obtain a trained information push model.

[0096] Among them, the hidden layer feature representation corresponding to the second information refers to the feature representation of the second information output by the pre-trained information push model.

[0097] The pre-trained information push model includes multiple layers. The hidden layer feature representation corresponding to the second information is the hidden layer feature representation output by the target layer in the pre-trained information push model. For example, the pre-trained information push model includes 3 layers, the second layer is used as the target layer, and the hidden layer feature representation is output by the second layer.

[0098] The hidden layer feature representation corresponding to one second information includes the hidden layer feature representations corresponding to each recipient user of one second information. The hidden layer feature representation corresponding to the recipient user represents the correlation between the information embedding vector of the second information and the user embedding vector of the recipient user. For example, if the second information A has 3 recipient users, then the hidden layer feature representation corresponding to the second information A includes the hidden layer feature representations corresponding to these 3 recipient users respectively.

[0099] Specifically, the computer device determines the information embedding vector of the second information it targets. The computer device performs model pre-training based on the information embedding vector of the second information it targets and the user embedding vector corresponding to each recipient user of the second information to obtain a pre-trained information push model, and the hidden layer feature representation of the second information output by the pre-trained information push model. In the same processing manner, the hidden layer feature representations corresponding to each second information can be obtained.

[0100] In this embodiment, the computer device performs model pre-training based on the information embedding vector of the second information it targets and the user embedding vector corresponding to each recipient user of the second information to obtain a pre-trained information push model, and the hidden layer feature representation of the second information output by the target layer in the pre-trained information push model.

[0101] The computer device can splice the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the hidden layer feature representation to obtain a spliced feature representation. Based on the spliced feature representation and the information embedding vector of the first piece of information, the pre-trained information push model is retrained and stopped when the retraining stop condition is reached, and the trained information push model is obtained.

[0102] In this embodiment, the model is pre-trained according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model and the hidden layer feature representation corresponding to each second piece of information. The hidden layer feature representation is the feature representation closest to the model output result, and this hidden layer feature representation is used as an input feature in the model retraining to guide the information push model to learn the commonalities when pushing each similar second piece of information to its respective recipient user. Based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, the information embedding vector of the first piece of information, and each hidden layer feature representation, retraining the pre-trained information push model can guide the information push model to extract the commonalities between the first piece of information and the second piece of information, and between the recipient users of the second piece of information, thereby improving the audience expansion ability of the information push model.

[0103] In one embodiment, pre-training the model according to each information embedding vector and each user embedding vector to obtain the hidden layer feature representation corresponding to each second piece of information includes:

[0104] For each second piece of information, determine the feature weight corresponding to the targeted second piece of information, and the feature weights corresponding to the other second pieces of information among the multiple second pieces of information except the targeted second piece of information; according to the information embedding vector of the targeted second piece of information, the feature weight corresponding to the targeted second piece of information, the user embedding vector corresponding to the recipient user of the targeted second piece of information, and the feature weights corresponding to the other second pieces of information and the user embedding vectors corresponding to the recipient users of the other second pieces of information, pre-train the model to obtain the hidden layer feature representation corresponding to the targeted second piece of information.

[0105] Among them, the feature weight corresponding to the second piece of information refers to the feature weight assigned to the recipient user of the second piece of information, which is also equivalent to the feature weight assigned to the user embedding vector corresponding to the recipient user of the second piece of information. The feature weight corresponding to the second piece of information is used to adjust the user embedding vector corresponding to the recipient user of the second piece of information. For example, multiply the feature weight by the user embedding vector to obtain a pre-trained user embedding vector.

[0106] Specifically, for each second piece of information, the computer device determines the feature weight corresponding to the second piece of information, determines the other second pieces of information among the multiple second pieces of information except the second piece of information targeted, and determines the feature weight corresponding to each of the other second pieces of information.

[0107] In this embodiment, the feature weight corresponding to the targeted second piece of information is greater than the feature weights corresponding to each of the other second pieces of information.

[0108] In this embodiment, the information push model to be trained includes a feature selection gate, which is used to adjust the weights of different second pieces of information to retain the differences between different second pieces of information.

[0109] The information push model to be trained processes one second piece of information each time, that is, one second piece of information serves as a pre-training task. In one pre-training task, the information embedding vector of one second piece of information and the user embedding vectors of the recipient users of all second pieces of information are input into the information push model to be trained, and the hidden layer feature representation corresponding to the one second piece of information is obtained.

[0110] The feature selection gate is used to assign a larger feature weight to the recipient user of the one second piece of information and a smaller feature weight to the recipient users of the other second pieces of information. The feature weight is used to reflect the importance of the user embedding vector of the recipient user in the current pre-training task.

[0111] For example, there are multiple second pieces of information A, B, and C. The current pre-training task is to perform pre-training using the second piece of information B, that is, a larger feature weight is assigned to the recipient user of the second piece of information B, and smaller feature weights are assigned to the recipient users of the second pieces of information A and C.

[0112] As Figure 5 shown, it is a schematic structural diagram of the feature selection gate in an embodiment. The processing process of the feature selection gate selectiongate is as Figure 6 shown, including:

[0113] Input embedding: , N is the batch size

[0114] Truncate the input gradient to prevent the learning of the feature gate from having a negative impact on the network:

[0115]

[0116] Mean pooling to reduce the number of parameters:

[0117] After passing through the first layer, it outputs a K-dimensional MLP network: =

[0118] The feature weights are obtained through the MLP network with the second-layer output being D:

[0119] =

[0120] Based on the weights, adjust the weights of each feature of the input embedding:

[0121] Automatically determine the importance of the features of each recipient user under the pre-training task through the feature selection gating mechanism.

[0122] The computer device can adjust the user embedding vector corresponding to the recipient user of the second information targeted according to the feature weights corresponding to the second information targeted, to obtain the pre-trained user embedding vector. The computer device can adjust the user embedding vectors corresponding to the recipient users of other second information according to the feature weights corresponding to the respective other second information, to obtain the pre-trained user embedding vectors corresponding to the recipient users of each other second information.

[0123] The computer device performs model pre-training according to the information embedding vector of the second information targeted, the pre-trained user embedding vector corresponding to the recipient user of the second information targeted, and the pre-trained user embedding vectors corresponding to the recipient users of each other second information, to obtain the hidden layer feature representation corresponding to the second information targeted.

[0124] In the same processing manner, the hidden layer feature representations corresponding to each second information can be obtained.

[0125] In this embodiment, for each second information, determining the feature weights corresponding to the second information targeted and the feature weights corresponding to the respective other second information among the multiple second information other than the second information targeted can adjust the feature weights of different second information to adapt to different pre-training tasks, so as to be able to retain the differences between different pre-training tasks based on the feature weights. Performing model pre-training according to the information embedding vector of the second information targeted, the feature weights corresponding to the second information targeted, the user embedding vector corresponding to the recipient user of the second information targeted, and the feature weights corresponding to the respective other second information and the user embedding vectors corresponding to the recipient users of the other second information enables different pre-training tasks to share the same all user embedding vectors, enabling the model to extract the commonalities of different pre-training tasks during training while still being able to retain the differences between each pre-training task.

[0126] In one embodiment, such as Figure 7As shown, based on the information embedding vector for the second information, the feature weight corresponding to the second information, the user embedding vector corresponding to the recipient user of the second information, and the feature weights corresponding to other second information and the user embedding vectors corresponding to the recipient users of other second information, model pre-training is performed to obtain the hidden layer feature representation corresponding to the second information, including:

[0127] Step S702: Determine the pre-trained user embedding vector of the recipient user of the second information according to the feature weight corresponding to the second information and the user embedding vector corresponding to the recipient user of the second information.

[0128] Specifically, the computer device can adjust the user embedding vector corresponding to the recipient user of the second information according to the feature weight corresponding to the second information to obtain the pre-trained user embedding vector.

[0129] In this embodiment, the computer device can multiply the feature weight corresponding to the second information and the user embedding vector corresponding to the recipient user of the second information to obtain the pre-trained user embedding vector corresponding to the recipient user of the second information.

[0130] Step S704: Determine the pre-trained user embedding vectors of the recipient users of other second information according to the feature weights corresponding to other second information and the user embedding vectors corresponding to the recipient users of other second information.

[0131] Specifically, the computer device can adjust the user embedding vectors corresponding to the recipient users of other second information according to the feature weights corresponding to other second information to obtain the pre-trained user embedding vectors corresponding to the recipient users of each other second information.

[0132] In this embodiment, for each other second information, the computer device multiplies the feature weight corresponding to the other second information and the user embedding vector corresponding to the recipient user of the other second information to obtain the pre-trained user embedding vector corresponding to the recipient user of the other second information.

[0133] Step S706: Perform model pre-training according to the information embedding vector of the second information, the pre-trained user embedding vector of the recipient user of the second information, and the pre-trained user embedding vectors of the recipient users of other second information to obtain the hidden layer feature representation corresponding to the second information.

[0134] Specifically, the computer device performs model pre-training based on the information embedding vector for the second information, the pre-trained user embedding vector corresponding to the recipient user of the second information, and the pre-trained user embedding vectors corresponding to the recipient users of each other second information, so as to adjust the weight parameters of the information push model to be trained, and stop when the pre-training stop condition is met, obtaining the pre-trained information push model and the hidden layer feature representation corresponding to the second information.

[0135] In this embodiment, the feature weight corresponding to the second information is greater than the feature weights corresponding to other second information, so as to increase the importance of the features of the recipient user of the second information in this pre-training task and reduce the importance of the features of the recipient users of other second information in this pre-training task.

[0136] In this embodiment, according to the feature weight corresponding to the second information and the user embedding vector corresponding to the recipient user of the second information, the pre-trained user embedding vector of the recipient user of the second information is determined, which can adjust the importance of the features of the recipient user of the second information in the pre-training task based on the feature weight in the pre-training task of the second information. According to the feature weights corresponding to each other second information and the user embedding vectors corresponding to the recipient users of the other second information, the pre-trained user embedding vectors corresponding to the recipient users of the other second information are determined, which can adjust the importance of the features of the recipient users of the other second information in the pre-training task based on the feature weight in the pre-training task of the second information. Model pre-training is performed based on the information embedding vector of the second information, the pre-trained user embedding vector of the recipient user of the second information, and the pre-trained user embedding vectors corresponding to the recipient users of the other second information, so that the obtained hidden layer feature representation corresponding to the second information can include the feature commonalities of each recipient user and the feature differences of the recipient user of the second information.

[0137] In one embodiment, as Figure 8 shown, performing model pre-training according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model includes:

[0138] Step S802, perform model pre-training on the information push model to be trained according to each information embedding vector and each user embedding vector to adjust the first type of weight parameters of the information push model to be trained, and obtain a pre-trained information push model; the information push model to be trained has the first type of weight parameters and the second type of weight parameters, and the pre-trained information push model has the adjusted first type of weight parameters and the second type of weight parameters.

[0139] Among them, the information push model to be trained has first - type weight parameters and second - type weight parameters. The pre - trained information push model has adjusted first - type weight parameters and second - type weight parameters. The trained information push model has adjusted first - type weight parameters and adjusted second - type weight parameters.

[0140] Specifically, the computer device fixes the second - type weight parameters of the information push model to be trained to keep the second - type weight parameters unchanged during the pre - training process. The computer device inputs each information embedding vector and each user embedding vector into the information push model to be trained to perform model pre - training on the information push model to be trained and calculates the pre - training loss. Based on the pre - training loss, the first - type weight parameters of the information push model to be trained are adjusted and the pre - training continues until it stops when the pre - training stops adjusting, obtaining the pre - trained information push model. This pre - trained information push model has adjusted first - type weight parameters and unchanged second - type weight parameters.

[0141] In this embodiment, based on the pre - training loss, the first - type weight parameters of the information push model to be trained are adjusted and the pre - training continues until it stops when the pre - training stops adjusting, obtaining the pre - trained information push model and the hidden - layer feature representations corresponding to each second piece of information.

[0142] Based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, the pre - trained information push model is retrained to obtain the trained information push model, including steps S804 - S806:

[0143] Step S804, keep the adjusted first - type weight parameters unchanged.

[0144] Specifically, the computer device fixes the adjusted first - type weight parameters of the pre - trained information push model to keep the adjusted first - type weight parameters unchanged during the retraining process.

[0145] Step S806, based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, the pre - trained information push model is retrained to adjust the second - type weight parameters of the pre - trained information push model, obtaining the trained information push model.

[0146] Specifically, the computer device inputs the user embedding vector corresponding to the recipient user of each second message, the user embedding vector corresponding to the recipient user of the first message, and the message embedding vector of the first message into the pre-trained message push model to retrain the pre-trained message push model and calculate the retraining loss of the model retraining. Based on the retraining loss, the second type of weight parameters of the pre-trained message push model are adjusted and retraining continues until it stops when the retraining stops adjusting, and a trained message push model is obtained. The trained message push model has the adjusted first type of weight parameters and the adjusted second type of weight parameters.

[0147] In this embodiment, the computer device retrains the pre-trained message push model based on the hidden layer feature representation corresponding to each second message, the user embedding vector corresponding to the recipient user of each second message, the user embedding vector corresponding to the recipient user of the first message, and the message embedding vector of the first message, so as to adjust the second type of weight parameters of the pre-trained message push model and obtain a trained message push model.

[0148] The message push model includes multiple adaptation layers, and the structure of the adaptation layer is as Figure 9 shown. Among them, the first type of weight parameters are pre-trained parameters, such as and b in the figure, and are only adjusted in the pre-training stage. The first type of weight parameters are fine-tuning parameters, such as and in the figure, and are adjusted in the retraining stage.

[0149] In this embodiment, according to each message embedding vector and each user embedding vector, the message push model to be trained is pre-trained to adjust the first type of weight parameters of the message push model to be trained, and a pre-trained message push model is obtained. Keeping the adjusted first type of weight parameters unchanged can effectively avoid contaminating the pre-trained knowledge in the retraining stage and retain the original pre-trained information. Based on the user embedding vector corresponding to the recipient user of each second message, the user embedding vector corresponding to the recipient user of the first message, and the message embedding vector of the first message, the pre-trained message push model is retrained to adjust the second type of weight parameters of the pre-trained message push model, so that the output in the retraining stage is closer to the true output of the first message, so that the trained message push model can not only obtain the commonality between other users and the already pushed users, but also extract the characteristics related to the first message to be pushed, so as to expand more recipient users for the first message and improve the conversion rate of the first message.

[0150] In this embodiment, the number of the second type of weight parameters is less than the number of the first type of weight parameters, so that the retraining nodes only need a small amount of training data to perform relatively sufficient fine-tuning, expanding the application scenarios.

[0151] In one embodiment, according to each information embedding vector and each user embedding vector, pre-train the information push model to be trained to adjust the first type of weight parameters of the information push model to be trained, and obtain a pre-trained information push model, including:

[0152] Through the information push model to be trained, for each second piece of information, according to the information embedding vector of the second piece of information and the user embedding vectors corresponding to the recipient users of each second piece of information, determine the pre-trained push scores for pushing the second piece of information to each recipient user; determine the pre-trained push labels for the second piece of information for each recipient user; based on the difference between the pre-trained push scores and the pre-trained push labels, adjust the first type of weight parameters of the information push model to be trained, and obtain a pre-trained information push model.

[0153] Among them, the pre-trained push score refers to the predicted score for pushing the second piece of information to the recipient user calculated by the information push model to be trained during the pre-training stage. This pre-trained push score represents the probability that the model predicts that the second piece of information is pushed to the recipient user.

[0154] Specifically, for each second piece of information, the computer device inputs the information embedding vector of the second piece of information and the user embedding vectors corresponding to the recipient users of all second pieces of information into the information push model to be trained. The information push model to be trained calculates the pre-trained push scores for pushing the second piece of information to each recipient user according to the information embedding vector of the second piece of information and the user embedding vectors corresponding to the recipient users of all second pieces of information. In the same processing manner, the information push model to be trained can output the pre-trained push scores for pushing each second piece of information to each recipient user.

[0155] The computer device obtains the pre-trained push labels for the second piece of information for each recipient user. The pre-trained push labels represent the real push scores, that is, they represent whether the second piece of information is pushed to a certain recipient user. The computer device determines the difference between the pre-trained push scores for pushing the second piece of information to each recipient user and the pre-trained push labels, and adjusts the first type of weight parameters of the information push model to be trained based on each difference, and obtains a pre-trained information push model.

[0156] In this embodiment, through the information push model to be trained, for each second piece of information, according to the information embedding vector of the second piece of information targeted and the user embedding vector corresponding to the recipient user of each second piece of information, determine the pre-training push score for pushing the targeted second piece of information to each recipient user, and determine the pre-training push label for the targeted second piece of information for each recipient user. Then, based on the difference between the pre-training push score and the pre-training push label, adjust the first type of weight parameters of the information push model to be trained to preliminarily train the information push model and obtain the pre-trained information push model.

[0157] In one embodiment, based on the difference between the pre-training push score and the pre-training push label, adjusting the first type of weight parameters of the information push model to be trained to obtain the pre-trained information push model includes:

[0158] Through the information push model to be trained, based on the pre-training push score and the pre-training push label, determine the pre-training loss; based on the pre-training loss, adjust the first type of weight parameters of the information push model to be trained to obtain the pre-trained information push model.

[0159] Specifically, the computer device calculates the pre-training loss through the information push model to be trained based on each pre-training push score and each pre-training push label. The computer device obtains a pre-training loss threshold, and based on the pre-training loss and the pre-training loss threshold, adjusts the first type of weight parameters of the information push model to be trained to obtain the pre-trained information push model.

[0160] In this embodiment, when the pre-training loss is greater than the pre-training loss threshold, the computer device adjusts the first type of weight parameters of the information push model to be trained and continues training after the adjustment until the pre-training loss is less than or equal to the pre-training loss threshold and then stops, obtaining the pre-trained information push model. Take the first type of weight parameters corresponding to when the pre-training loss is less than or equal to the pre-training loss threshold as the adjusted first type of weight parameters. The pre-trained information push model has the adjusted first type of weight parameters.

[0161] In this embodiment, through the information push model to be trained, based on the pre-training push score and the pre-training push label, determine the pre-training loss, and based on the pre-training loss, adjust the first type of weight parameters of the information push model to be trained to gradually reduce the loss during pre-training, thereby obtaining the pre-trained information push model.

[0162] In one embodiment, based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, the pre-trained information push model is retrained to adjust the second type of weight parameters of the pre-trained information push model, and a trained information push model is obtained, including:

[0163] Through the pre-trained information push model, according to the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, determine the retraining push score of the first piece of information pushed to each recipient user for the first piece of information;

[0164] Determine the retraining push label of the first piece of information for each recipient user; based on the difference between the retraining push score and the retraining push label, adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model.

[0165] Among them, the retraining push score refers to the predicted score of the first piece of information pushed to the recipient user calculated by the information push model to be trained in the retraining stage. This retraining push score represents the probability that the first piece of information is predicted to be pushed to the recipient user by the model.

[0166] Specifically, the computer device inputs the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information into the pre-trained information push model. Through the pre-trained information push model, according to the user embedding vectors and the information embedding vector of the first piece of information, determine the retraining push score of the first piece of information pushed to each recipient user for the first piece of information.

[0167] The computer device obtains the retraining push label of the second piece of information for each recipient user. The retraining push label represents the true push score, that is, it represents whether the first piece of information is pushed to a certain recipient user.

[0168] The computer device determines the difference between each retraining push score and the corresponding retraining push label. Based on these differences, adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model.

[0169] In this embodiment, through a pre-trained information push model, according to the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, determine the re-training push score for pushing the first piece of information to each recipient user for the first piece of information, and determine the re-training push label for the first piece of information for each recipient user; based on the difference between the re-training push score and the re-training push label, adjust the second type of weight parameters of the pre-trained information push model to perform targeted training on the preliminarily trained information push model, and generate an information push model dedicated to pushing the first piece of information, so that the information push model can specifically expand the recipient users of the first piece of information to push the first piece of information to more users and improve the conversion rate of the information.

[0170] In one embodiment, based on the difference between the re-training push score and the re-training push label, adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model, including:

[0171] Through the pre-trained information push model, based on the re-training push score and the re-training push label, determine the re-training loss; based on the re-training loss, adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model.

[0172] Specifically, the computer device calculates the re-training loss through the re-trained information push model based on the re-training push score and the re-training push label. The computer device obtains a re-training loss threshold, and based on the re-training loss and the re-training loss threshold, adjusts the second type of weight parameters of the information push model to be trained to obtain a re-trained information push model.

[0173] In this embodiment, when the re-training loss is greater than the re-training loss threshold, the computer device adjusts the second type of weight parameters of the information push model to be trained and continues to train after adjustment until the re-training loss is less than or equal to the re-training loss threshold, and then stops the re-trained information push model. Take the second type of weight parameters corresponding to when the re-training loss is less than or equal to the re-training loss threshold as the adjusted second type of weight parameters. The re-trained information push model has the adjusted first type of weight parameters and the adjusted second type of weight parameters.

[0174] As Figure 10 shown, since the information of each pre-training task has different importance for the first piece of information, adaptively adjust the importance of different pre-training tasks through the Pre Information Transfer Module, and its main structure is implemented using the attention mechanism:

[0175]

[0176] Among them, is the feature weight of each second piece of information, and the calculation formula is as follows:

[0177]

[0178] ,

[0179]

[0180]

[0181] is the scaling factor, which is used to control the dimension for easy network learning, and k is the dimension of the current input.

[0182] The final input in the retraining stage is the concatenation of the above two parts, that is:

[0183] in p ut t = [ I nfo t || sEM D t ] ∈ R N*(D + d2)

[0184] Input into the multi-layer MLP network exclusive to retraining to obtain the final output . Then, the loss can be calculated based on the true label label of the first piece of information, and updated based on the gradient:

[0185]

[0186] In this embodiment, through the pre-trained information push model, based on the retraining push score and the retraining push label, the loss generated by the model in the retraining stage is determined. Based on the retraining loss, the second type of weight parameters of the pre-trained information push model are adjusted, so that the loss in the retraining stage gradually decreases, and the trained information push model is obtained. Moreover, through retraining, the pre-trained information push model can be trained specifically to generate an information push model dedicated to expanding the recipient users of the first piece of information, which can effectively improve the accuracy of model prediction.

[0187] In one embodiment, model pre-training is performed based on each information embedding vector and each user embedding vector to obtain a pre-trained information push model, including: performing model pre-training based on each information embedding vector and each user embedding vector to obtain a pre-trained information push model, and the target user embedding vector corresponding to the recipient user of each second piece of information;

[0188] Determining the user embedding vector corresponding to the recipient user of the first piece of information includes: based on the target user embedding vector corresponding to the recipient user of each second piece of information, determining the target user embedding vector corresponding to the recipient user of the first piece of information;

[0189] Based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, retraining the pre-trained information push model to obtain a trained information push model, including:

[0190] Based on the target user embedding vector corresponding to the recipient user of each second piece of information, the target user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, retraining the pre-trained information push model to obtain a trained information push model.

[0191] Specifically, the information push model to be trained has shared weight parameters, first-type weight parameters, and second-type weight parameters. The shared weight parameters are the weight parameters used to determine the target user embedding vector of the recipient user.

[0192] The computer device performs model pre-training according to each information embedding vector and each user embedding vector to adjust the shared parameters and the first-type weight parameters of the pre-trained information push model. When the pre-training stops, a pre-trained information push model is obtained. This pre-trained information push model has adjusted shared parameters and adjusted first-type weight parameters.

[0193] Based on the adjusted shared parameters, the pre-trained information push model adjusts the user embedding vector corresponding to each recipient user to obtain the target user embedding vector corresponding to each recipient user.

[0194] The computer device determines the user embedding vector corresponding to the recipient user of the first piece of information, and based on the adjusted shared parameters, adjusts the user embedding vector corresponding to the recipient user of the first piece of information to obtain the target user embedding vector corresponding to the recipient user of the first piece of information.

[0195] In this embodiment, the computer device can obtain the recipient feature information corresponding to the recipient user of the first piece of information, as well as the recipient feature information corresponding to the recipient user of each second piece of information. Based on these recipient feature information, the target user embedding vector corresponding to the recipient user of the first piece of information is determined from the target user embedding vectors corresponding to the recipient users of each second piece of information.

[0196] The computer device retrains the pre-trained information push model based on the target user embedding vectors corresponding to the recipient users of each second piece of information, the target user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, and obtains a trained information push model.

[0197] In this embodiment, model pre-training is performed according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model, as well as the target user embedding vectors corresponding to the recipient users of each second piece of information. Thus, the target user embedding vectors of each recipient user can be obtained based on pre-training, and the target user embedding vectors are used as the input for retraining. Combining with the information embedding vector of the first piece of information, targeted training is performed on the pre-trained information push model, so that the trained information push model can accurately expand more recipient users of the first piece of information.

[0198] In one embodiment, the method further includes:

[0199] Determine the user embedding vectors of multiple candidate users through the trained information push model;

[0200] Through the trained information push model, based on the user embedding vectors of multiple candidate users and the information embedding vector of the first piece of information, determine the push score of the first piece of information pushed to each candidate user; through the trained information push model, based on the push score, screen out the recipient users of the first piece of information from multiple candidate users, and push the first piece of information to the screened recipient users.

[0201] Specifically, the computer device obtains the feature information of multiple candidate users through the trained information push model, converts each feature information into a corresponding embedding vector, and obtains the user embedding vector of each candidate user.

[0202] The terminal determines the push score of the first piece of information pushed to each candidate user through the trained information push model based on the user embedding vectors of multiple candidate users and the information embedding vector of the first piece of information. The terminal screens out the recipient users of the first piece of information from multiple candidate users through the trained information push model based on the push score, and pushes the first piece of information to the screened recipient users.

[0203] In this embodiment, the trained information push model determines the push scores for pushing the first piece of information to each candidate user based on the user embedding vectors of the respective candidate users and the information embedding vector of the first piece of information, including:

[0204] The trained information push model determines the push scores for pushing the first piece of information to each candidate user based on the user embedding vector of the recipient user of the first piece of information, the user embedding vectors of the respective candidate users, and the information embedding vector of the first piece of information.

[0205] As Figure 11 shown, the trained information push model determines multiple candidate users and determines the user embedding vectors of the respective candidate users, and screens new recipient users from the multiple candidate users based on the user embedding vector corresponding to the historical recipient user of the first piece of information, the user embedding vectors of the respective candidate users, and the information embedding vector of the first piece of information.

[0206] In this embodiment, pushing the first piece of information to the screened recipient users includes:

[0207] The trained information push model obtains the historical push channels of the first piece of information; the trained information push model pushes the first piece of information to the screened recipient users according to the historical push channels.

[0208] In this embodiment, pushing the first piece of information to the screened recipient users includes:

[0209] Determine the push channel of the first piece of information, and the trained information push model pushes the first piece of information to the screened recipient users according to the push channel.

[0210] In this embodiment, the trained information push model determines the user embedding vectors of the respective candidate users, so as to calculate the push scores for pushing the first piece of information to each candidate user based on the user embedding vectors of the respective candidate users and the information embedding vector of the first piece of information, and thus screen the recipient users of the first piece of information from the multiple candidate users based on the push scores, quickly and accurately realizing the expansion of the recipient users of the first piece of information. Push the first piece of information to the screened recipient users to improve the exposure rate and conversion rate of the first piece of information.

[0211] In this embodiment, the information push model scores the predicted click probabilities of all users and recalls the top N users with high click probabilities according to the scores, so as to realize the expansion of the audience for the delivery of the first piece of information.

[0212] In one embodiment, a processing method for an information push model is provided, which is applied to a computer device and includes:

[0213] Obtain the first information and determine multiple second information belonging to the same information category as the first information;

[0214] Determine the recipient users of the multiple second information respectively, and determine the user embedding vectors corresponding to the recipient users of each second information;

[0215] For each second information, determine the feature weight corresponding to the targeted second information, and the feature weights corresponding to the other second information except the targeted second information among the multiple second information respectively;

[0216] According to the feature weight corresponding to the targeted second information and the user embedding vector corresponding to the recipient user of the targeted second information, determine the pre-trained user embedding vector of the recipient user of the targeted second information;

[0217] According to the feature weights corresponding to the other second information respectively and the user embedding vectors corresponding to the recipient users of the other second information, determine the pre-trained user embedding vectors corresponding to the recipient users of the other second information;

[0218] Keep the second type of weight parameters of the information push model to be trained unchanged, and through the information push model to be trained, for each second information, according to the information embedding vector of the targeted second information, the pre-trained user embedding vector of the recipient user of the targeted second information, and the pre-trained user embedding vectors corresponding to the recipient users of the other second information, determine the pre-trained push scores for the targeted second information to be pushed to each recipient user;

[0219] Determine the pre-trained push labels for the targeted second information for each recipient user, and based on the pre-trained push scores and the pre-trained push labels, determine the pre-trained loss.

[0220] Based on the pre-trained loss, adjust the first type of weight parameters of the information push model to be trained to obtain the pre-trained information push model, the hidden layer feature representations corresponding to each second information, and the target user embedding vectors corresponding to the recipient users of each second information.

[0221] Determine the recipient user of the first information, and determine the target user embedding vector corresponding to the recipient user of the first information;

[0222] Determine the information embedding vector of the first information;

[0223] Keep the adjusted first type of weight parameters unchanged, and through the pre-trained information push model, according to the target user embedding vectors corresponding to the recipient users of each second information, the target user embedding vector corresponding to the recipient user of the first information, the information embedding vector of the first information, and each hidden layer feature representation, determine the re-training push scores for the first information to be pushed to each recipient user for the first information.

[0224] Determine the retraining push tags of the first piece of information for each recipient user. Based on the retraining push scores and retraining push tags, determine the retraining loss through a pre-trained information push model.

[0225] Based on the retraining loss, adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model; the trained information push model has the adjusted first type of weight parameters and the adjusted second type of weight parameters.

[0226] Through the trained information push model, determine the user embedding vectors of multiple candidate users respectively;

[0227] Through the trained information push model, based on the user embedding vectors of multiple candidate users respectively and the information embedding vector of the first piece of information, determine the push scores of the first piece of information pushed to each candidate user;

[0228] Through the trained information push model, based on the push scores, screen out the recipient users of the first piece of information from multiple candidate users, and push the first piece of information to the screened recipient users.

[0229] In one embodiment, a processing method for an information push model is provided, which is applied to the advertising placement scenario of credit cards. In this application scenario, the 5 largest-scale credit card advertising placements in history are used as offline pre-training tasks, and the advertising placement of the target credit card is used as an online training task. The information push model is a pre-training-fine-tuning AMDN model. Through this model, the advertising click probabilities of all users are scored, and the top N users are recalled according to the scores, and the advertisements of the target credit card are placed on multiple channels, effectively realizing the audience expansion of the advertisements of the target credit card. The architecture diagram of the pre-training-fine-tuning AMDN model is as Figure 12 shown. The dotted part in the figure represents that the model parameters are only updated during offline pre-training, the solid part represents that the model parameters are updated during online retraining, and the rest represents the mixed state, where a part of the network parameters are updated during pre-training and another part are updated during retraining. This model mainly includes 3 parts:

[0230] Multi-domain feature representation and selection module: Represent the input features as embeddings, and adjust the weights of each dimension of features through a feature selection gate to adapt to different offline pre-training tasks, and obtain the final feature representation for downstream use. The multi-domain feature representation and selection module may include a first determination module and a second determination module.

[0231] Pre-training Information Extraction and Utilization Module, i.e., Pre-training Module: Different offline tasks use independent parameters to avoid mutual influence. For each layer of each task, the network parameters are split into two parts, the first type of weight parameters and the second type of weight parameters, which are updated in the offline stage and the online stage respectively by referring to LoRA (Low-Rank Adaptation of Large Language Models).

[0232] Target Task Module, i.e., Re-training Module: The shared embedding of the pre-training task undergoes target task feature selection and is concatenated with the information transferred from the pre-training task and transformed by the Pre Information Transfer Module, and then input into the multi-layer MLP network with exclusive parameters for the target task for scoring the target task.

[0233] In this embodiment, there are M offline pre-training tasks and 1 online target task. Each pre-training task corresponds to a second piece of information, and the target task corresponds to the first piece of information. The following symbols are defined:

[0234] : The set of all features. The union of the features of each task needs to be taken to ensure the consistency of the input of each task, with a total of D dimensions, where:

[0235] : It represents that there are multiple pre-training tasks, and it is the identifier indicating which pre-training task each sample information belongs to. The sample information is the second piece of information.

[0236] : It represents the label of the pre-training task. p indicates which pre-training task the current sample information belongs to.

[0237] : It represents the target task label, that is, the re-training push label corresponding to the first piece of information.

[0238] : The simplified representation of a single-layer feedforward neural network, , is the input of this layer of neural network. Both W and b are learnable network parameters, is the activation function of the neural network such as ReLU. For example:

[0239]

[0240] This function represents transforming a matrix with the original dimension of 512*256 into an output of 512*64 through a single-layer MLP, using the ReLU activation function.

[0241] : A simplified representation of the neural network embedding layer that maps sparse id features into fixed-length learnable vectors of dimension d. Given the original input features , then .

[0242] : A simplified representation of the mean pooling layer of a neural network. If D features are input, each feature being d-dimensional, after mean pooling, the result will be . The principle is to calculate the mean for all dimensions of each d-dimensional vector, thus obtaining a floating-point mean result.

[0243] : Gradient truncation, which prevents the gradient update of nodes and the nodes before it. That is, during backpropagation of the gradient, the gradient of this node and its directly associated previous nodes are not updated. This operation does not change the dimension of the input nodes.

[0244] : A simplified representation of a single-layer LoRA layer, where is the input to this layer of the network, is the activation function of the neural network, such as ReLU, etc. Different from the MLP layer , LoRA can support achieving the effect of pre-training - fine-tuning with a small number of parameters. For each layer, in addition to the pre-trained and , a low-rank bypass is added to obtain the updated network parameters . The expressions for the updated W and b are:

[0245]

[0246]

[0247] where . Therefore, through LoRA, the effect of full-scale fine-tuning can be achieved with a very small number of parameters, and it is also more beneficial to retain the pre-training information and avoid forgetting. It should be noted that A needs to be initialized with a random Gaussian distribution, and B and need to be initialized with a 0 matrix to ensure that the bypass matrix is still a 0 matrix at the beginning of training. For example:

[0248]

[0249]

[0250]

[0251] It represents that the matrix of the original dimension 512*256 is transformed into an output of 512*64 through a single-layer LoRA and activated by the ReLU activation function.

[0252] Processing of the multi-domain feature characterization and selection module: A total of M pieces of second information are used. For the convenience of information sharing and migration, the union of the historical data of these M pieces of second information has dimensions, and the original input features can be characterized as , where N is the batch size. First, these original sparse features need to be transformed into dense embeddings. With the help of the neural network embedding layer, they are represented as fixed-length learnable vectors. Each feature is represented as a d-dimensional fixed-length vector. After representation, X can generate the corresponding EMB:

[0253]

[0254] To ensure that each task can distinguish the importance of different features during training, a feature selection gate is introduced for each task, and the weights of the features of the sample information used are adjusted according to the characteristics of each task. The importance of each feature in this pre-training task is automatically determined through the feature selection gating mechanism. The calculation process of the feature selection gate selection gate from input to output is as follows:

[0255] Input embedding: , where N is the batch size

[0256] Truncate the input gradient to prevent the learning of the feature gate from having a negative impact on the network:

[0257]

[0258] Mean pooling to reduce the number of parameters:

[0259] The output after the first layer is a K-dimensional MLP network: =

[0260] The output after the second layer is a D-dimensional MLP network to obtain the feature weights:

[0261] =

[0262] Adjust the weights of each feature of the input embedding based on the weights:

[0263] For the M pre-training tasks of the AMDN model, a shared embedding bottom layer is used , before entering the unique parameter network of each task, it passes through the selection gate exclusive to each task to obtain the outputs of M pre-training tasks. . At this time, the input of each pre-training task has selected the original features to be more suitable for the current pre-training task.

[0264] To reduce the number of parameters, each group of features can be further reduced in dimension by mean pooling, that is: ;

[0265] Among them, , the processed result is still denoted as .

[0266] Processing of the pre-training information extraction and utilization module: The basic composition structure of the pre-training information extraction and utilization module is the LoRA unit. To fully retain the differences of each pre-training task, a dedicated multi-layer LoRA network is constructed for each pre-training task, where the pre-training parameters and b are only trained using the samples of each pre-training task itself in the offline stage. The fine-tuning parameters and are then trained using the samples of the target task in the online stage. This processing has the following effects:

[0267] The original pre-training parameters and b are no longer modified and updated, avoiding the contamination of pre-training knowledge by the target task and retaining the original pre-training knowledge. The fine-tuning parameters and are trained based on the data of the target task, making the output of the pre-training network closer to the target task. Since and the parameter scale of is much smaller than the pre-training parameters and b, only a small amount of target task data is needed for relatively sufficient fine-tuning, expanding the application scenario.

[0268] The input of a certain pre-training task p is , and for a binary classification task, its output is . Assuming that the network structure is three layers and the output dimensions are d1*d2*1 respectively, then the output of the first layer can be expressed as: = .

[0269] The output of the second layer can be expressed as: = .

[0270] The output of the last layer can be expressed as: = .

[0271] It should be noted that the sigmoid activation function is used in the last layer of the binary classification task.

[0272] Through the above operations, the outputs of M pre-training tasks can be obtained , and these outputs will be used to calculate the loss with the true labels of the corresponding tasks. Assume is the number of samples of the p-th pre-training task, then the output of each pre-training task will also have entries. Therefore .

[0273] For the binary classification task, cross-entropy loss can be used. Since the training samples contain samples of M tasks at the same time, the loss needs to be extended to M tasks. The final loss form is as follows. It should be noted that this loss will only take effect during offline training. That is, the following formula is the pre-training loss function:

[0274]

[0275] At the same time, retain the output of the second-to-last layer of the network unique to each pre-training task, that is , and these outputs will be used in the target task module.

[0276] The processing of the target task module is as follows:

[0277] The target task module is retrained in the online stage. It receives the output of the pre-training module and the shared embedding as inputs, and is trained only using the sample data of the target task through an MLP network with exclusive parameters. When predicting, each candidate user of the target task is scored, and the audience is expanded according to the score. The sample data of the target task is the first information. The input of the target task module includes two parts:

[0278] The first part of the input uses the embedding shared with the pre-training task, denoted as EMD. After passing through a selection gate similar to the pre-training task and performing avg pooling, the input of the target task is obtained .

[0279] The second part of the input comes from the pre-training information extraction and utilization module, which is the output of the second-to-last layer of the tower network unique to each pre-training task , where .

[0280] Since the information of each pre-training task varies in importance to the target task, a Pre Information Transfer Module is designed to adaptively adjust the importance of different pre-training tasks. Its main structure is implemented using the attention mechanism:

[0281]

[0282] Among them, is the feature weight of each second piece of information, and the calculation formula is as follows:

[0283]

[0284] ,

[0285]

[0286]

[0287] is the scaling coefficient, used to control the dimension for easy network learning, and k is the dimension of the current input.

[0288] The final input in the retraining stage is the concatenation of the above two parts, that is:

[0289] in p ut t = [ I nfo t || sEM D t ] ∈ R N*(D + d2)

[0290] Input into the multi-layer MLP network exclusive to retraining to obtain the final output . Then, the loss can be calculated based on the true label label of the first piece of information and updated based on the gradient:

[0291]

[0292] It should be noted that This only takes effect during online retraining.

[0293] In this embodiment, the offline pre-training combined with online fine-tuning architecture is adopted. The sample information of the historically similar audience expansion task is introduced for pre-training in the offline stage, and fine-tuned in the online stage based on the sample information of the target audience expansion task. The pre-training task network and the target task network have both shared parameters and their own unique parameter parts, which can avoid the forgetting of pre-training knowledge as much as possible while retaining the differences between tasks.

[0294] The pre-training network in the offline stage draws on the LoRA structure and only updates the pre-trained parameter part without updating the low-rank parameter part. Different pre-training tasks share the underlying embedding to share the common features among tasks, and at the same time introduce a feature selection gate for each pre-training task to individualize the weights of features to retain the differences between tasks.

[0295] In the online stage, the target task uses the same feature embedding as the pre-training network and adjusts the feature weights through the feature selection gate, retaining the differences of the target task while utilizing the pre-training knowledge. The target task uses an MLP network with independent parameters instead of reusing the parameters of the pre-training network, fitting the target task to the maximum extent and avoiding contaminating the pre-training knowledge.

[0296] To make the pre-training task closer to the target task, in the online stage, the low-rank parameter part of the LoRA layer of the pre-training network is fine-tuned based on the target task samples, but the pre-trained parameter part is not changed, so that the pre-training network is closer to the target task while avoiding contaminating the pre-training knowledge when the target task samples are too few or differ greatly from the pre-training task.

[0297] When training the target task network, in addition to reusing the embedding information of the pre-training network, the output vector of the penultimate layer (non-output layer) of each pre-training task network is also passed to the target task as high-order information. Since the importance of different pre-training tasks for the target task is different, the attention mechanism is used to adaptively determine the weights of the information of each pre-training task, compensating for the defect of limited knowledge transfer caused by only sharing the embedding between the pre-training task and the downstream task.

[0298] In this embodiment, as Figure 13 shown, the main network of the pre-training task can use a multi-expert structure such as MMoE to improve the fitting ability. The feature selection gate can use other feature selection methods, and the attention mechanism in the Pre InformationTransfer Module can be replaced with other types of attention, such as additive attention. In addition, the main model structure of the target task module can also use other arbitrary model structures such as DeepFM in addition to MLP.

[0299] In the audience expansion task of credit card advertisement push scenarios, when conducting offline evaluation, three metrics are used: AUC (Area Under ROC), P@K%, and R@K%. Among them, P@K% and R@K% represent the accuracy and recall rate of the top K% users among all users. Their calculation methods are as follows:

[0300]

[0301] Among them, represents the set of true target users, represents the set of top K% predicted users with the highest scores. Online evaluation uses the advertisement click-through rate CTR (Click-Through Rate) metric, which represents the probability that an advertisement is clicked and is usually presented in the form of a percentage. The more accurate the expanded population is and the more significant the interest in the target advertisement, the higher the CTR. CTR can be calculated using the following formula:

[0302] CTR = Number of clicks / Number of impressions × 100%

[0303] Among them, the number of clicks refers to the number of times a user actually clicks on an advertisement after seeing it, and the number of impressions refers to the number of times an advertisement is shown to a user. The test results are shown in the following table:

[0304]

[0305] Among them, MLP is an End2End multi-layer perceptron model trained only with the sample data of the target task. MLP+pretrain first pre-trains MLP based on the pre-training task samples and then fine-tunes it based on the sample information of the target task. Pinterest constructs a common embedding for each user based on the pre-training task and target task samples, calculates the similarity between the overall population and the seed population based on the embedding, and the target users with higher similarity are more likely to be target users. MetaHeac obtains good initial parameters based on the meta-learning pre-training model and then fine-tunes it based on the sample information of the target task. From the comparison in the table, it can be seen that the information push model in this embodiment has better performance than other models both offline and online.

[0306] It should be understood that although the steps in the flowcharts involved in the above embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0307] Based on the same inventive concept, an embodiment of the present application further provides a processing device for an information push model for implementing the processing method of the information push model involved above. The implementation solution provided by this device to solve problems is similar to the implementation solution described in the above method. Therefore, the specific limitations in one or more embodiments of the processing device for the information push model provided below can refer to the limitations on the processing method of the information push model in the above text, and will not be elaborated here.

[0308] In one embodiment, as Figure 14 shown, a processing device 1400 for an information push model is provided, including: an information acquisition module 1402, a first determination module 1404, a pre-training module 1406, a second determination module 1408, and a re-training module 1410, where:

[0309] The information acquisition module 1402 is configured to acquire first information and determine a plurality of second information belonging to the same information category as the first information.

[0310] The first determination module 1404 is configured to determine the recipient users of the plurality of second information respectively, and determine the user embedding vectors corresponding to the recipient users of each second information.

[0311] The pre-training module 1406 is configured to determine the information embedding vectors of each second information, perform model pre-training according to each information embedding vector and each user embedding vector, and obtain a pre-trained information push model.

[0312] The second determination module 1408 is further configured to determine the recipient user of the first information and determine the user embedding vector corresponding to the recipient user of the first information.

[0313] A retraining module 1410 is configured to determine an information embedding vector of the first information; and retrain a pre-trained information push model based on the user embedding vectors corresponding to the recipient users of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, so as to obtain a trained information push model.

[0314] In this embodiment, the first information is obtained, and multiple second information belonging to the same information category as the first information is determined. The recipient users of the multiple second information are determined, which can obtain the historical push data of other similar information as a supplement when the historical push data of the first information is insufficient, so as to obtain more training samples. The user embedding vectors corresponding to the recipient users of each second information are determined, the information embedding vectors of each second information are determined, and model pre-training is performed according to each information embedding vector and each user embedding vector, so that all the same user embedding vectors are shared in each pre-training, enabling the model to obtain the commonalities of the recipient users of similar information during training. The pre-trained information push model is obtained, the recipient user of the first information is determined, the user embedding vector corresponding to the recipient user of the first information is determined, the information embedding vector of the first information is determined, and the pre-trained information push model is retrained specifically based on the user embedding vectors corresponding to the recipient users of each second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, so that the trained information push model can not only obtain the ability to extract the commonalities between other users and the historical recipient users of the first information, but also obtain the ability to extract the characteristics related to the first information, thereby obtaining an information push model dedicated to expanding the recipient users of the first information, realizing the effective expansion of the target audience of the first information, and helping to improve the exposure rate and conversion rate of the first information.

[0315] In one embodiment, the pre-training module 1406 is further configured to perform model pre-training according to each information embedding vector and each user embedding vector, so as to obtain a pre-trained information push model and the hidden layer feature representations corresponding to each second information;

[0316] The retraining module 1410 is further configured to retrain the pre-trained information push model based on the user embedding vectors corresponding to the recipient users of each second information, the user embedding vector corresponding to the recipient user of the first information, the information embedding vector of the first information, and each hidden layer feature representation, so as to obtain a trained information push model.

[0317] In this embodiment, model pre-training is performed according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model and the hidden layer feature representation corresponding to each second information. This hidden layer feature representation is the feature representation closest to the model output result, and this hidden layer feature representation is used as an input feature in the model re-training to guide the information push model to learn the commonalities that exist when pushing each similar second information to its respective recipient user. Based on the user embedding vector corresponding to the recipient user of each second information, the user embedding vector corresponding to the recipient user of the first information, the information embedding vector of the first information, and each hidden layer feature representation, the pre-trained information push model is re-trained, which can guide the information push model to extract the commonalities between the first information and the second information, and between the recipient users of the second information and the recipient users of the second information, thereby improving the audience expansion ability of the information push model.

[0318] In one embodiment, the pre-training module 1406 is further configured to, for each second information, determine the feature weight corresponding to the targeted second information, and the feature weights corresponding to the other second information except the targeted second information among the multiple second information; perform model pre-training according to the information embedding vector of the targeted second information, the feature weight corresponding to the targeted second information, the user embedding vector corresponding to the recipient user of the targeted second information, and the feature weights corresponding to the other second information and the user embedding vectors corresponding to the recipient users of the other second information, to obtain the hidden layer feature representation corresponding to the targeted second information.

[0319] In this embodiment, for each second information, determining the feature weight corresponding to the targeted second information and the feature weights corresponding to the other second information except the targeted second information among the multiple second information can adjust the feature weights of different second information to adapt to different pre-training tasks, so as to be able to retain the differences between different pre-training tasks based on the feature weights. Performing model pre-training according to the information embedding vector of the targeted second information, the feature weight corresponding to the targeted second information, the user embedding vector corresponding to the recipient user of the targeted second information, and the feature weights corresponding to the other second information and the user embedding vectors corresponding to the recipient users of the other second information enables different pre-training tasks to share the same all user embedding vectors, enabling the model to extract the commonalities of different pre-training tasks during training, while also being able to retain the differences between each pre-training task.

[0320] In one embodiment, the pre-training module 1406 is further configured to determine a pre-trained user embedding vector of the recipient user of the second information to be targeted according to the feature weight corresponding to the second information to be targeted and the user embedding vector corresponding to the recipient user of the second information to be targeted; determine pre-trained user embedding vectors of the recipient users of other second information according to the feature weights corresponding to the respective other second information and the user embedding vectors corresponding to the recipient users of the other second information; perform model pre-training according to the information embedding vector of the second information to be targeted, the pre-trained user embedding vector of the recipient user of the second information to be targeted, and the pre-trained user embedding vectors of the recipient users of the other second information, so as to obtain a hidden layer feature representation corresponding to the second information to be targeted.

[0321] In this embodiment, determining a pre-trained user embedding vector of the recipient user of the second information to be targeted according to the feature weight corresponding to the second information to be targeted and the user embedding vector corresponding to the recipient user of the second information to be targeted can, in the pre-training task of the second information to be targeted, adjust the importance degree of the features of the recipient user of the second information to be targeted in the pre-training task based on the feature weight. Determining pre-trained user embedding vectors of the recipient users of other second information according to the feature weights corresponding to the respective other second information and the user embedding vectors corresponding to the recipient users of the other second information can, in the pre-training task of the second information to be targeted, adjust the importance degree of the features of the recipient users of the other second information in the pre-training task based on the feature weight. Performing model pre-training according to the information embedding vector of the second information to be targeted, the pre-trained user embedding vector of the recipient user of the second information to be targeted, and the pre-trained user embedding vectors of the recipient users of the other second information enables the obtained hidden layer feature representation corresponding to the second information to be targeted to include the feature commonalities of each recipient user and the feature differences of the recipient user of the second information to be targeted.

[0322] In one embodiment, the pre-training module 1406 is further configured to perform model pre-training on the information push model to be trained according to each information embedding vector and each user embedding vector, so as to adjust the first type of weight parameters of the information push model to be trained and obtain a pre-trained information push model; the information push model to be trained has the first type of weight parameters and the second type of weight parameters, and the pre-trained information push model has the adjusted first type of weight parameters and the second type of weight parameters;

[0323] The retraining module 1410 is further configured to keep the adjusted first type of weight parameters unchanged; and retrain the pre-trained information push model based on the user embedding vectors corresponding to the recipient users of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, so as to adjust the second type of weight parameters of the pre-trained information push model, and obtain a trained information push model.

[0324] In this embodiment, according to each information embedding vector and each user embedding vector, the information push model to be trained is pre-trained to adjust the first type of weight parameters of the information push model to be trained, and a pre-trained information push model is obtained. Keeping the adjusted first type of weight parameters unchanged can effectively avoid contaminating the pre-trained knowledge in the retraining stage and retain the original pre-trained information. Based on the user embedding vectors corresponding to the recipient users of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, the pre-trained information push model is retrained to adjust the second type of weight parameters of the pre-trained information push model, so that the output in the retraining stage is closer to the true output of the first piece of information. Therefore, the trained information push model can not only obtain the commonality between other users and the already pushed users, but also extract the characteristics related to the first piece of information to be pushed, so as to expand more recipient users for the first piece of information and improve the conversion rate of the first piece of information.

[0325] In one embodiment, the pre-training module 1406 is further configured to, for each second piece of information, through the information push model to be trained, determine the pre-training push scores for pushing the second piece of information to each recipient user according to the information embedding vector of the second piece of information and the user embedding vectors corresponding to the recipient users of each second piece of information; determine the pre-training push labels for the second piece of information for each recipient user; and adjust the first type of weight parameters of the information push model to be trained based on the difference between the pre-training push scores and the pre-training push labels, so as to obtain a pre-trained information push model.

[0326] In this embodiment, through the information push model to be trained, for each second piece of information, the pre-training push scores for pushing the second piece of information to each recipient user are determined according to the information embedding vector of the second piece of information and the user embedding vectors corresponding to the recipient users of each second piece of information, and the pre-training push labels for the second piece of information for each recipient user are determined. Therefore, based on the difference between the pre-training push scores and the pre-training push labels, the first type of weight parameters of the information push model to be trained are adjusted to preliminarily train the information push model and obtain a pre-trained information push model.

[0327] In one embodiment, the pre-training module 1406 is further configured to, through the information push model to be trained, determine a pre-training loss based on the pre-training push score and the pre-training push label; and based on the pre-training loss, adjust the first type of weight parameters of the information push model to be trained to obtain a pre-trained information push model.

[0328] In this embodiment, through the information push model to be trained, a pre-training loss is determined based on the pre-training push score and the pre-training push label, and based on the pre-training loss, the first type of weight parameters of the information push model to be trained are adjusted, so that the loss is gradually reduced during pre-training, thereby obtaining a pre-trained information push model.

[0329] In one embodiment, the re-training module 1410 is further configured to, through the pre-trained information push model, determine a re-training push score for pushing the first information to each recipient user for the first information according to the user embedding vector corresponding to each recipient user of the second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information; determine a re-training push label for the first information for each recipient user; and based on the difference between the re-training push score and the re-training push label, adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model.

[0330] In this embodiment, through the pre-trained information push model, a re-training push score for pushing the first information to each recipient user for the first information is determined according to the user embedding vector corresponding to each recipient user of the second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, and a re-training push label for the first information for each recipient user is determined; and based on the difference between the re-training push score and the re-training push label, the second type of weight parameters of the pre-trained information push model are adjusted to perform targeted training on the preliminarily trained information push model to generate an information push model dedicated to pushing the first information, so that the information push model can specifically expand the recipient users of the first information to push the first information to more users and improve the conversion rate of the information.

[0331] In one embodiment, the re-training module 1410 is further configured to, through the pre-trained information push model, determine a re-training loss based on the re-training push score and the re-training push label; and based on the re-training loss, adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model.

[0332] In this embodiment, through a pre-trained information push model, based on the retraining push scores and retraining push labels, the loss generated by the model in the retraining stage is determined. Based on the retraining loss, the second type of weight parameters of the pre-trained information push model are adjusted so that the loss in the retraining stage gradually decreases, and an information push model with training completed is obtained. Moreover, through retraining, the information push model obtained through pre-training can be trained specifically to generate an information push model dedicated to expanding the recipient users of the first information, which can effectively improve the accuracy of model prediction.

[0333] In one embodiment, the pre-training module 1406 is further configured to perform model pre-training according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model, and the target user embedding vector corresponding to the recipient user of each second information.

[0334] The second determination module 1408 is further configured to determine the target user embedding vector corresponding to the recipient user of the first information based on the target user embedding vector corresponding to the recipient user of each second information.

[0335] The retraining module 1410 is further configured to perform retraining on the pre-trained information push model based on the target user embedding vector corresponding to the recipient user of each second information, the target user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information to obtain an information push model with training completed.

[0336] In this embodiment, model pre-training is performed according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model and the target user embedding vector corresponding to the recipient user of each second information, so that the target user embedding vector of each recipient user can be obtained based on pre-training, and this target user embedding vector is used as the input of retraining. Combining with the information embedding vector of the first information, targeted training is performed on the pre-trained information push model, so that the information push model with training completed can accurately expand more recipient users of the first information.

[0337] In one embodiment, the device further includes:

[0338] A push module, configured to determine the user embedding vector of each of multiple candidate users through the information push model with training completed; determine the push score of the first information pushed to each candidate user based on the user embedding vector of each of the multiple candidate users and the information embedding vector of the first information through the information push model with training completed; screen out the recipient users of the first information from the multiple candidate users based on the push scores through the information push model with training completed, and push the first information to the screened recipient users.

[0339] In this embodiment, through the information push model completed by training, the user embedding vectors of multiple candidate users are determined, and based on the user embedding vectors of multiple candidate users and the information embedding vector of the first information, the push scores of the first information pushed to each candidate user are calculated. Thus, based on the push scores, the recipient users of the first information are screened out from multiple candidate users, quickly and accurately realizing the expansion of the recipient users of the first information. The first information is pushed to the screened recipient users to increase the exposure rate and conversion rate of the first information.

[0340] Each module in the processing device of the above information push model can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0341] In one embodiment, a computer device is provided. The computer device can be a server or a terminal. Taking the server as an example, its internal structure diagram can be as Figure 15 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the processing data of the information push model. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a processing method of an information push model.

[0342] Those skilled in the art can understand that Figure 15 the structure shown in

[0343] is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0344] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0345] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0346] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0347] For the pushed information, the user can refuse or conveniently decline the push.

[0348] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0349] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0350] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A processing method for an information push model, characterized in that The method includes: Obtain first information and determine multiple second information belonging to the same information category as the first information; Determine the recipient users of each of the multiple second information, and determine the user embedding vectors corresponding to the recipient users of each of the second information; Determine the information embedding vector of each of the second information, and perform model pre-training according to each of the information embedding vectors and each of the user embedding vectors to obtain a pre-trained information push model; Determine the recipient user of the first information, and determine the user embedding vector corresponding to the recipient user of the first information; Determine the information embedding vector of the first information; Based on the user embedding vectors corresponding to the recipient users of each of the second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, re-train the pre-trained information push model to obtain a trained information push model.

2. The method according to claim 1, wherein The performing model pre-training according to each of the information embedding vectors and each of the user embedding vectors to obtain a pre-trained information push model includes: Perform model pre-training according to each of the information embedding vectors and each of the user embedding vectors to obtain a pre-trained information push model and the hidden layer feature representations corresponding to each of the second information; The re-training the pre-trained information push model based on the user embedding vectors corresponding to the recipient users of each of the second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information to obtain a trained information push model includes: Based on the user embedding vectors corresponding to the recipient users of each of the second information, the user embedding vector corresponding to the recipient user of the first information, the information embedding vector of the first information, and each of the hidden layer feature representations, re-train the pre-trained information push model to obtain a trained information push model.

3. The method according to claim 2, characterized in that, The performing model pre-training according to each of the information embedding vectors and each of the user embedding vectors to obtain the hidden layer feature representations corresponding to each of the second information includes: For each of the second information, determine the feature weight corresponding to the targeted second information, and the feature weights corresponding to the other second information except the targeted second information among the multiple second information; Perform model pre-training according to the information embedding vector of the targeted second information, the feature weight corresponding to the targeted second information, the user embedding vector corresponding to the recipient user of the targeted second information, and the feature weights corresponding to the other second information and the user embedding vectors corresponding to the recipient users of the other second information to obtain the hidden layer feature representation corresponding to the targeted second information.

4. The method according to claim 3, wherein The performing model pre-training according to the information embedding vector of the targeted second information, the feature weight corresponding to the targeted second information, the user embedding vector corresponding to the recipient user of the targeted second information, and the feature weights corresponding to the other second information and the user embedding vectors corresponding to the recipient users of the other second information to obtain the hidden layer feature representation corresponding to the targeted second information includes: Determine the pre-trained user embedding vector of the recipient user of the second information to be targeted based on the feature weight corresponding to the second information to be targeted and the user embedding vector corresponding to the recipient user of the second information to be targeted; Determine the pre-trained user embedding vector of the recipient user corresponding to each of the other second information based on the feature weight corresponding to each of the other second information and the user embedding vector corresponding to the recipient user of each of the other second information; Perform model pre-training based on the information embedding vector of the second information to be targeted, the pre-trained user embedding vector of the recipient user of the second information to be targeted, and the pre-trained user embedding vector of the recipient user corresponding to each of the other second information, to obtain the hidden layer feature representation corresponding to the second information to be targeted.

5. The method according to claim 1, characterized in that The performing model pre-training based on each of the information embedding vectors and each of the user embedding vectors to obtain a pre-trained information push model includes: Perform model pre-training on the information push model to be trained based on each of the information embedding vectors and each of the user embedding vectors, to adjust the first type of weight parameters of the information push model to be trained, and obtain a pre-trained information push model; the information push model to be trained has the first type of weight parameters and the second type of weight parameters, and the pre-trained information push model has the adjusted first type of weight parameters and the second type of weight parameters; The re-training the pre-trained information push model based on the user embedding vector corresponding to the recipient user of each of the second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information to obtain a trained information push model includes: Keep the adjusted first type of weight parameters unchanged; Re-train the pre-trained information push model based on the user embedding vector corresponding to the recipient user of each of the second information, the user embedding vector corresponding to the recipient user of the first information, and the information embedding vector of the first information, to adjust the second type of weight parameters of the pre-trained information push model, and obtain a trained information push model.

6. The method according to claim 5, characterized in that, The performing model pre-training on the information push model to be trained based on each of the information embedding vectors and each of the user embedding vectors, to adjust the first type of weight parameters of the information push model to be trained, and obtain a pre-trained information push model includes: Through the information push model to be trained, for each of the second information, determine the pre-trained push score of the second information to be targeted pushed to each of the recipient users according to the information embedding vector of the second information to be targeted and the user embedding vector corresponding to the recipient user of each of the second information; Determine the pre-trained push label of the second information to be targeted for each of the recipient users; Based on the difference between the pre-trained push score and the pre-trained push label, adjust the first type of weight parameters of the information push model to be trained, and obtain a pre-trained information push model.

7. The method according to claim 6, wherein Adjusting the first type of weight parameters of the information push model to be trained based on the difference between the pre-trained push score and the pre-trained push label to obtain a pre-trained information push model includes: Determining a pre-trained loss through the information push model to be trained based on the pre-trained push score and the pre-trained push label; Adjusting the first type of weight parameters of the information push model to be trained based on the pre-trained loss to obtain a pre-trained information push model.

8. The method according to claim 5, wherein Retraining the pre-trained information push model based on the user embedding vectors corresponding to the recipient users of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information to adjust the second type of weight parameters of the pre-trained information push model to obtain a trained information push model includes: Determining, through the pre-trained information push model, a retraining push score for pushing the first piece of information to each recipient user for the first piece of information based on the user embedding vector corresponding to the recipient user of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information; Determining a retraining push label for the first piece of information for each recipient user; Adjusting the second type of weight parameters of the pre-trained information push model based on the difference between the retraining push score and the retraining push label to obtain a trained information push model.

9. The method according to claim 8, wherein Adjusting the second type of weight parameters of the pre-trained information push model based on the difference between the retraining push score and the retraining push label to obtain a trained information push model includes: Determining a retraining loss through the pre-trained information push model based on the retraining push score and the retraining push label; Adjusting the second type of weight parameters of the pre-trained information push model based on the retraining loss to obtain a trained information push model.

10. The method according to claim 1, wherein Pre-training the model according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model includes: Pre-training the model according to each information embedding vector and each user embedding vector to obtain a pre-trained information push model and the target user embedding vectors corresponding to the recipient users of each second piece of information; Determining the user embedding vector corresponding to the recipient user of the first piece of information includes: Determining the target user embedding vector corresponding to the recipient user of the first piece of information based on the target user embedding vectors corresponding to the recipient users of each second piece of information; Retraining the pre-trained information push model based on the user embedding vectors corresponding to the recipient users of each second piece of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information to obtain a trained information push model includes: Retrain the pre-trained information push model based on the target user embedding vector corresponding to the recipient user of each of the second pieces of information, the target user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, to obtain a trained information push model.

11. The method according to any one of claims 1 to 10, characterized in that, The method further includes: Determine the user embedding vector of each of multiple candidate users through the trained information push model; Determine the push score of the first piece of information pushed to each of the candidate users based on the user embedding vector of each of the multiple candidate users and the information embedding vector of the first piece of information through the trained information push model; Screen out the recipient user of the first piece of information from the multiple candidate users based on the push score through the trained information push model, and push the first piece of information to the screened recipient user.

12. A processing device for an information push model, characterized in that, The device includes: An information acquisition module, configured to acquire a first piece of information and determine multiple second pieces of information belonging to the same information category as the first piece of information; A first determination module, configured to determine the recipient user of each of the multiple second pieces of information and determine the user embedding vector corresponding to the recipient user of each of the second pieces of information; A pre-training module, configured to determine the information embedding vector of each of the second pieces of information, perform model pre-training according to each information embedding vector and each user embedding vector, to obtain a pre-trained information push model; A second determination module, further configured to determine the recipient user of the first piece of information and determine the user embedding vector corresponding to the recipient user of the first piece of information; A retraining module, configured to determine the information embedding vector of the first piece of information; retrain the pre-trained information push model based on the user embedding vector corresponding to the recipient user of each of the second pieces of information, the user embedding vector corresponding to the recipient user of the first piece of information, and the information embedding vector of the first piece of information, to obtain a trained information push model.

13. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

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

15. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 11.

Images