Information recommendation model training and information recommendation method, device and computer equipment

Abstract

The present application relates to a method, apparatus, computer device, storage medium, and computer program product for training an information recommendation model. The method comprises: inputting training recommendation object information and training recommendation information into an initial information recommendation model; the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree; performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility; fuses each interaction bias degree and each interaction possibility to obtain a fused recommendation degree corresponding to the training recommendation information; calculates fusion loss based on each interaction label and fusion recommendation degree to obtain fusion loss information; and uses each interaction bias degree to bias-adjust the fusion loss information to obtain target loss information; and trains a first target information recommendation model based on the target loss information. The use of this method can improve the accuracy of information recommendation.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to an information recommendation model training and information recommendation method, apparatus, computer equipment, storage medium and computer program product. Background Art

[0002] With the development of artificial intelligence, multi-objective recommendation technology has emerged. Multi-objective recommendation often involves weighing different business objectives for information. Typically, multiple objectives for the recommended information are evaluated first, and then the evaluation results are fused to obtain a fused evaluation result. Recommendation decisions are made based on this fused evaluation result. For example, when recommending videos, multi-objective evaluations are typically performed based on factors such as click behavior and viewing time. Currently, when fusion is performed, pre-set weights are typically used for weighted fusion. However, using pre-set weights for weighted fusion can easily reduce the accuracy of the fused evaluation results, leading to reduced recommendation accuracy. Summary of the Invention

[0003] Based on this, it is necessary to provide an information recommendation model training and information recommendation method, device, computer equipment, computer-readable storage medium and computer program product that can improve the accuracy of information recommendation in response to the above technical problems.

[0004] In one aspect, the present application provides a method for training an information recommendation model. The method comprises:

[0005] Obtain training samples, which include training recommendation object information, training recommendation information, and various interaction labels;

[0006] The training recommendation object information and the training recommendation information are input into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree. The initial information recommendation model also performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. The initial information recommendation model performs fusion based on each interaction bias degree and each interaction possibility to obtain a fused recommendation degree corresponding to the training recommendation information.

[0007] The fusion loss is calculated based on each interaction label and the degree of fusion recommendation to obtain the fusion loss information, and the fusion loss information is biased and adjusted using each interaction bias degree to obtain the target loss information;

[0008] Based on the target loss information, the initial information recommendation model is reversely updated to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining training samples is returned to execute until the training completion condition is met to obtain the first target information recommendation model.

[0009] On the other hand, the present application also provides an information recommendation model training device. The device includes:

[0010] The sample acquisition module is used to obtain training samples, which include training recommendation object information, training recommendation information and various interaction labels;

[0011] An initial prediction module is used to input the training recommendation object information and the training recommendation information into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. The initial information recommendation model performs fusion based on each interaction bias degree and each interaction possibility to obtain a fused recommendation degree corresponding to the training recommendation information.

[0012] The loss calculation module is used to calculate the fusion loss based on each interaction label and the fusion recommendation degree to obtain the fusion loss information, and use each interaction bias degree to adjust the bias of the fusion loss information to obtain the target loss information;

[0013] The iterative module is used to reversely update the initial information recommendation model based on the target loss information to obtain an updated information recommendation model, use the updated information recommendation model as the initial information recommendation model, and return to the step of obtaining training samples until the training completion condition is met to obtain the first target information recommendation model.

[0014] In another aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are performed:

[0015] Obtain training samples, which include training recommendation object information, training recommendation information, and various interaction labels;

[0016] The training recommendation object information and the training recommendation information are input into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree. The initial information recommendation model also performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. The initial information recommendation model performs fusion based on each interaction bias degree and each interaction possibility to obtain a fused recommendation degree corresponding to the training recommendation information.

[0017] The fusion loss is calculated based on each interaction label and the degree of fusion recommendation to obtain the fusion loss information, and the fusion loss information is biased and adjusted using each interaction bias degree to obtain the target loss information;

[0018] Based on the target loss information, the initial information recommendation model is reversely updated to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining training samples is returned to execute until the training completion condition is met to obtain the first target information recommendation model.

[0019] On the other hand, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0020] Obtain training samples, which include training recommendation object information, training recommendation information, and various interaction labels;

[0021] The training recommendation object information and the training recommendation information are input into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree. The initial information recommendation model also performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. The initial information recommendation model performs fusion based on each interaction bias degree and each interaction possibility to obtain a fused recommendation degree corresponding to the training recommendation information.

[0022] The fusion loss is calculated based on each interaction label and the degree of fusion recommendation to obtain the fusion loss information, and the fusion loss information is biased and adjusted using each interaction bias degree to obtain the target loss information;

[0023] Based on the target loss information, the initial information recommendation model is reversely updated to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining training samples is returned to execute until the training completion condition is met to obtain the first target information recommendation model.

[0024] On the other hand, the present application also provides a computer program product. The computer program product includes a computer program, which, when executed by a processor, implements the following steps:

[0025] Obtain training samples, which include training recommendation object information, training recommendation information, and various interaction labels;

[0026] The training recommendation object information and the training recommendation information are input into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree. The initial information recommendation model also performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. The initial information recommendation model performs fusion based on each interaction bias degree and each interaction possibility to obtain a fused recommendation degree corresponding to the training recommendation information.

[0027] The fusion loss is calculated based on each interaction label and the degree of fusion recommendation to obtain the fusion loss information, and the fusion loss information is biased and adjusted using each interaction bias degree to obtain the target loss information;

[0028] Based on the target loss information, the initial information recommendation model is reversely updated to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining training samples is returned to execute until the training completion condition is met to obtain the first target information recommendation model.

[0029] The above-mentioned information recommendation model training method, device, computer equipment, storage medium and computer program product, by inputting the training recommendation object information and the training recommendation information into the initial information recommendation model, and then the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and uses the training recommendation object information and the training recommendation information to perform interaction prediction to obtain each interaction possibility, and then uses each interaction bias degree and each interaction possibility to fuse to obtain the fused recommendation degree corresponding to the training recommendation information, that is, by fusing the interaction bias degree during fusion, the accuracy of the obtained fused recommendation degree is improved, and then uses each interaction label and the fused recommendation degree to calculate the fusion loss to obtain the fusion loss information, and uses each interaction bias degree to adjust the bias of the fusion loss information to obtain the target loss information, that is, by using the interaction bias degree to adjust the bias of the fusion loss information, the accuracy of the obtained target loss information is improved, and then uses the target loss information to iteratively train the initial information recommendation model to obtain the first target information recommendation model, thereby improving the accuracy of the first target information recommendation model.

[0030] In one aspect, the present application provides an information recommendation method. The method comprises:

[0031] Obtain information to be recommended and information about the person to be recommended;

[0032] Input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain the target recommendation degree, wherein the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses based on each interaction bias degree and each interaction possibility to obtain the fusion recommendation degree corresponding to the training recommendation information; calculates the fusion loss based on each interaction label and the fusion recommendation degree to obtain the fusion loss information, and uses each interaction bias degree to adjust the bias of the fusion loss information to obtain the target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model;

[0033] When the target recommendation degree exceeds the preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the information of the object to be recommended.

[0034] On the other hand, the present application also provides an information recommendation device. The device includes:

[0035] An information acquisition module is used to acquire information to be recommended and information about the object to be recommended;

[0036] A recommendation prediction module is used to input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain a target recommendation degree, wherein the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses based on each interaction bias degree and each interaction possibility to obtain a fusion recommendation degree corresponding to the training recommendation information; calculates a fusion loss based on each interaction label and the fusion recommendation degree to obtain fusion loss information, and uses each interaction bias degree to perform bias adjustment on the fusion loss information to obtain target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model;

[0037] The recommendation module is used to recommend the information to be recommended to the target recommendation object corresponding to the information of the object to be recommended when the target recommendation degree exceeds a preset recommendation degree threshold.

[0038] In another aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are performed:

[0039] Obtain information to be recommended and information about the person to be recommended;

[0040] Input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain the target recommendation degree, wherein the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses based on each interaction bias degree and each interaction possibility to obtain the fusion recommendation degree corresponding to the training recommendation information; calculates the fusion loss based on each interaction label and the fusion recommendation degree to obtain the fusion loss information, and uses each interaction bias degree to adjust the bias of the fusion loss information to obtain the target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model;

[0041] When the target recommendation degree exceeds the preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the information of the object to be recommended.

[0042] On the other hand, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0043] Obtain information to be recommended and information about the person to be recommended;

[0044] Input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain the target recommendation degree, wherein the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses based on each interaction bias degree and each interaction possibility to obtain the fusion recommendation degree corresponding to the training recommendation information; calculates the fusion loss based on each interaction label and the fusion recommendation degree to obtain the fusion loss information, and uses each interaction bias degree to adjust the bias of the fusion loss information to obtain the target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model;

[0045] When the target recommendation degree exceeds the preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the information of the object to be recommended.

[0046] On the other hand, the present application also provides a computer program product. The computer program product includes a computer program, which, when executed by a processor, implements the following steps:

[0047] Obtain information to be recommended and information about the person to be recommended;

[0048] Input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain the target recommendation degree, wherein the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses based on each interaction bias degree and each interaction possibility to obtain the fusion recommendation degree corresponding to the training recommendation information; calculates the fusion loss based on each interaction label and the fusion recommendation degree to obtain the fusion loss information, and uses each interaction bias degree to adjust the bias of the fusion loss information to obtain the target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model;

[0049] When the target recommendation degree exceeds the preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the information of the object to be recommended.

[0050] The above-mentioned information recommendation method, device, computer equipment, storage medium and computer program product input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain the target recommendation degree, and use the training recommendation object information to perform interaction bias prediction to obtain each interaction bias degree, and use the training recommendation object information and the training recommendation information to perform interaction prediction to obtain each interaction possibility, and then use each interaction bias degree and each interaction possibility to fuse to obtain the fused recommendation degree corresponding to the training recommendation information, that is, by fusing the interaction bias degree during fusion, the accuracy of the obtained fused recommendation degree is improved, and then each interaction label and the fused recommendation degree are used to perform fusion loss. Calculation is performed to obtain fusion loss information, and each interaction bias degree is used to perform bias adjustment on the fusion loss information to obtain target loss information, that is, by using the interaction bias degree to perform bias adjustment on the fusion loss information, the accuracy of the obtained target loss information is improved, and then the target loss information is used to iteratively train the initial information recommendation model to obtain a first target information recommendation model, thereby improving the first target information recommendation model, and then the first target information recommendation model is used to recommend information, thereby improving the accuracy of the obtained target recommendation degree, and then when the target recommendation degree exceeds the preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the information of the object to be recommended, thereby improving the accuracy of the information recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a diagram of an application environment of an information recommendation model training method in one embodiment;

[0052] Figure 2 Schematic diagram of a flow chart of a method for training an information recommendation model in one embodiment;

[0053] Figure 3 A schematic diagram of a flow chart for obtaining various interaction possibilities in one embodiment;

[0054] Figure 4 A schematic diagram of the network structure of an interactive prediction network in a specific embodiment;

[0055] Figure 5 A schematic diagram of the network structure of a fusion network in a specific embodiment;

[0056] Figure 6 A schematic diagram of a process for obtaining target loss information in one embodiment;

[0057] Figure 7 Schematic diagram of a process for calculating cross entropy loss in one embodiment;

[0058] Figure 8A schematic diagram of a process for obtaining fusion loss information in one embodiment;

[0059] Figure 9 Schematic diagram of a flow chart of a method for obtaining sample weights in one embodiment;

[0060] Figure 10 A schematic diagram of a process for obtaining a second target information recommendation model in one embodiment;

[0061] Figure 11 1 is a flow chart of an information recommendation method according to an embodiment;

[0062] Figure 12 A schematic diagram of a process for obtaining an information recommendation model in a specific embodiment;

[0063] Figure 13 A schematic diagram of the structure of a multi-objective recommendation model in a specific embodiment;

[0064] Figure 14 for Figure 13 Schematic diagram of a news recommendation page in a specific embodiment;

[0065] Figure 15-A This is a test comparison diagram of the number of clicks on pictures and texts in a specific embodiment;

[0066] Figure 15-B for Figure 15-A A schematic diagram of comparison of test result indicators in a specific embodiment;

[0067] Figure 16 is a structural block diagram of an information recommendation model training device in one embodiment;

[0068] Figure 17 is a structural block diagram of an information recommendation device in one embodiment;

[0069] Figure 18 is a diagram of the internal structure of a computer device in one embodiment;

[0070] Figure 19 FIG. 4 is a diagram showing the internal structure of a computer device in another embodiment. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0072] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0073] The solution provided in the embodiments of the present application involves technologies such as multi-task learning of artificial intelligence, which is specifically illustrated by the following embodiments:

[0074] The information recommendation model training method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 via a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104 or placed on the cloud or other servers. When the server 104 receives the training instruction sent by the terminal 102, it obtains training samples from the data storage system. The training samples include training recommendation object information, training recommendation information and various interaction labels. The server 104 inputs the training recommendation object information and the training recommendation information into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain various interaction bias levels, and performs various interaction predictions based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities. The server 104 performs fusion based on the various interaction bias levels and the various interaction possibilities to obtain a fusion recommendation level corresponding to the training recommendation information. The server 104 calculates fusion loss based on the various interaction labels and the fusion recommendation level to obtain fusion loss information, and uses the various interaction bias levels to bias-adjust the fusion loss information to obtain target loss information. The server 104 reversely updates the initial information recommendation model based on the target loss information to obtain an updated information recommendation model, uses the updated information recommendation model as the initial information recommendation model, and returns to the step of obtaining training samples for execution until the training completion condition is met to obtain the first target information recommendation model. Terminal 102 may include, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart car devices. Portable wearable devices may include smart watches, smart bracelets, and head-mounted devices. Server 104 may be implemented as a standalone server or a server cluster consisting of multiple servers.

[0075] In one embodiment, Figure 2 As shown, a method for training an information recommendation model is provided, which is applied to Figure 1 In the example of the server in the embodiment, it is understood that the method can also be applied to the server, and can also be applied to the system including the terminal and the server, and is implemented through the interaction between the terminal and the server. In this embodiment, the following steps are included:

[0076] Step 204: Acquire training samples, where the training samples include training recommendation object information, training recommendation information, and various interaction tags.

[0077] Among them, the training samples are samples used during training, and there can be multiple training samples. Training recommendation information refers to information that has been recommended to the recommended object in the past, and this information includes but is not limited to videos, images, texts and voices. The recommended object refers to the object that obtains the recommended object information, and this object includes a real object or a virtual object. The real object includes but is not limited to people, animals, etc. A virtual object can refer to an object that uses a virtual image for live broadcasting on the Internet. Training recommended object information refers to information used during training that can characterize the recommended object, and this information includes but is not limited to the basic attribute information, social relationship information, consumption capacity information, and behavior information of the recommended object. The basic attribute information is used to characterize the basic attributes of the recommended object, the social relationship information is used to characterize the social relationship of the recommended object, the consumption capacity information is used to characterize the consumption capacity of the recommended object, and the behavior information is used to characterize the behavior of the recommended object. Interaction refers to the interaction of the recommended object with the recommended information, including but not limited to clicks, reading, interactions, etc. The interaction label refers to the label used during training, and this label is used to characterize the interaction results of the recommended object with the training recommended object information. Different interactions of the recommended object with the training recommended object information have different interaction results, that is, different interaction labels. Different interactions can be different business goals. For example, the interaction label can be a click label, and the business goal is to make the recommended object click on the recommended object information.

[0078] Specifically, the server can obtain training samples from a database. These training samples include training recommendation object information, training recommendation information corresponding to the training recommendation object information, and various interaction tags corresponding to the training recommendation object information. These interaction tags are obtained after obtaining the interaction between the training recommendation object and the training recommendation information. The server can also obtain training samples from a training sample set. The server can also obtain training samples from a service provider that provides data services, or from a business provider. The server can also obtain training samples uploaded by a terminal.

[0079] In step 204, the training recommendation object information and the training recommendation information are input into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. Based on each interaction bias degree and each interaction possibility, they are fused to obtain a fused recommendation degree corresponding to the training recommendation information.

[0080] The initial information recommendation model refers to an information recommendation model with initialized model parameters. This information recommendation model is used to predict information recommendations and is an artificial intelligence model built using a neural network. Model parameter initialization can be random, zero, or Gaussian. The interaction bias represents the degree of bias in the training recommendation subject toward corresponding interactive behaviors, that is, the tendency of the recommendation subject itself. A higher interaction bias indicates a higher likelihood that the recommendation subject will interact with the recommended information. In other words, a higher interaction bias indicates a higher probability of that interaction occurring. For example, older recommendation subjects tend to spend more time reading, while younger recommendation subjects tend to have a higher interaction rate. For each interaction, a corresponding interaction bias is predicted. The interaction likelihood represents the likelihood of the corresponding interaction occurring; a higher interaction likelihood indicates a higher likelihood of the corresponding interaction occurring. A corresponding interaction likelihood is predicted for each interaction. The fused recommendation degree represents the degree to which the training recommendation information is recommended to the training recommendation subject. A higher fused recommendation degree indicates a higher likelihood of the training recommendation information being recommended to the recommendation subject.

[0081] Specifically, the server inputs the training recommendation object information and the training recommendation information into the initial information recommendation model, and the initial information recommendation model performs multi-task learning, that is, uses the training recommendation object information to predict the interaction bias to obtain each interaction bias degree, and then uses the training recommendation object information and the training recommendation information to predict each interaction to obtain each interaction possibility, and finally fuses each interaction bias degree and each interaction possibility to obtain the fused recommendation degree corresponding to the training recommendation information.

[0082] In step 204 , fusion loss is calculated based on each interaction tag and the fusion recommendation degree to obtain fusion loss information, and the fusion loss information is bias-adjusted using each interaction bias degree to obtain target loss information.

[0083] The fusion loss information is used to represent the error between the fusion recommendation level output by training and the actual recommendation result. The target loss information refers to the error between the fusion recommendation level adjusted by the interaction bias and the actual recommendation result.

[0084] Specifically, the server uses a loss function to calculate the error between each interaction label and the fusion recommendation level, obtaining the loss of each learning task during multi-task learning. It then performs a fusion loss calculation to obtain fusion loss information. The server then uses the interaction bias levels to obtain the weights of the training samples, and uses the training sample weights to perform a weighted calculation on the fusion loss information to obtain the target loss information.

[0085] In step 204, the initial information recommendation model is reversely updated based on the target loss information to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining training samples is returned to be executed until the training completion condition is met to obtain the first target information recommendation model.

[0086] The training completion conditions refer to the conditions for completing the training of the initial information recommendation model, which may include reaching the maximum number of training iterations, the training loss reaching the maximum threshold, and the model parameters no longer changing. The updated information recommendation model refers to the information recommendation model after the model parameters are updated. The first target information recommendation model refers to the trained multi-target model used for information recommendation.

[0087] Specifically, the server can first determine whether the training completion condition has been met. If the training completion condition has not been met, the server uses a gradient descent algorithm based on the target loss information to reversely update the model parameters in the initial information recommendation model to obtain an updated information recommendation model. The updated information recommendation model is then used as the initial information recommendation model, and the server returns to the step of obtaining training samples and iterates until the training completion condition is met. The information recommendation model that meets the training completion condition is then used as the first target information recommendation model.

[0088] The above-mentioned information recommendation model training method inputs the training recommendation object information and the training recommendation information into the initial information recommendation model, and then the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and uses the training recommendation object information and the training recommendation information to perform interaction prediction to obtain each interaction possibility, and then uses each interaction bias degree and each interaction possibility to fuse to obtain the fused recommendation degree corresponding to the training recommendation information, that is, by fusing the interaction bias degree during fusion, the accuracy of the obtained fused recommendation degree is improved, and then uses each interaction label and the fused recommendation degree to calculate the fusion loss to obtain the fusion loss information, and uses each interaction bias degree to perform bias adjustment on the fusion loss information to obtain the target loss information, that is, by using the interaction bias degree to perform bias adjustment on the fusion loss information, the accuracy of the obtained target loss information is improved, and then uses the target loss information to iteratively train the initial information recommendation model to obtain the first target information recommendation model, thereby improving the accuracy of the first target information recommendation model.

[0089] In one embodiment, step 204, i.e., inputting the training recommendation object information and the training recommendation information into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fuses the various interaction bias degrees and the various interaction possibilities to obtain a fused recommendation degree corresponding to the training recommendation information, including the following steps:

[0090] The interaction prediction network in the initial information recommendation model predicts the degree of each interaction bias of the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility.

[0091] Among them, the interactive prediction network refers to a deep neural network used for prediction. It is a multi-task learning neural network that can predict the degree of each interactive bias and can also perform individual interaction predictions.

[0092] Specifically, when the server inputs the training recommendation object information and the training recommendation information into the initial information recommendation model, the interaction prediction network in the initial information recommendation model uses the initialized bias prediction parameters to predict the interaction bias degrees of the training recommendation object information to obtain the interaction bias degrees. Simultaneously, the initialized interaction prediction parameters are used to predict the interactions of the training recommendation object information and the training recommendation information to obtain the interaction probabilities. In one embodiment, basic features can be extracted from the training recommendation object information and the training recommendation information to obtain basic features of the training recommendation object and basic features of the training recommendation information. The basic features of the training recommendation object and the basic features of the training recommendation information are then input into the initial information recommendation model. The interaction prediction network uses the basic features of the training recommendation object and the basic features of the training recommendation information to predict the interaction bias degrees and interactions to obtain the interaction bias degrees and interaction probabilities.

[0093] Furthermore, each interaction bias degree and each interaction possibility are fused through the fusion network in the initial information recommendation model to obtain the fusion recommendation degree corresponding to the training recommendation information.

[0094] Among them, the fusion network is a deep neural network used to fuse the degree of interaction bias and the possibility of interaction. The deep neural network can be a DNN (Deep Neural Networks) network, or a CNN (Recurrent Neural Network) network, an RNN (Convolutional Neural Network) network, and the like.

[0095] Specifically, the server inputs each interaction bias degree and each interaction possibility output by the interaction prediction network into the fusion network for fusion, and obtains the fusion recommendation degree output by the fusion network.

[0096] In one embodiment, the interaction prediction network includes an interaction bias degree prediction network and an interaction possibility prediction network;

[0097] like Figure 3 As shown, the interaction prediction network in the initial information recommendation model predicts the degree of interaction bias of the training recommendation object information to obtain the degree of interaction bias, and then performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain the possibility of interaction, including:

[0098] Step 302 : performing feature extraction based on the training recommendation object information and the training recommendation information to obtain training recommendation object features and training recommendation information features, and performing feature combination based on the training recommendation object features and the training recommendation information features to obtain combined features.

[0099] The training recommendation object feature refers to the feature corresponding to the training object during training, and the training recommendation information feature refers to the feature corresponding to the training recommendation information. The combined feature refers to the feature obtained by cross-combining the training recommendation object feature and the training recommendation information feature.

[0100] Specifically, the server may perform feature extraction on the training recommendation object information and the training recommendation information, respectively, to obtain training recommendation object features and training recommendation information features. Feature extraction may be performed using a feature extraction network, which may be a deep neural network for feature extraction. The training recommendation object features and the training recommendation information features are then combined to obtain a combined feature.

[0101] In step 304, the training recommendation object features are input into the interaction bias degree prediction network to extract interaction features to obtain the training recommendation object interaction features, and each interaction bias degree is predicted based on the training recommendation object interaction features to obtain each interaction bias degree.

[0102] The interaction bias prediction network is a deep neural network used to predict interaction bias. The training recommendation object interaction features are features extracted by the neural network in the interaction bias prediction network. The training recommendation object interaction features are representation vectors of the recommended objects during training, which can be low-dimensional representation vectors.

[0103] Specifically, the server inputs the training recommendation object features into the interaction bias prediction network and uses the network parameters extracted from the interaction feature extraction to extract the interaction features, obtaining the training recommendation object interaction features. The server then uses the training recommendation object interaction features to predict the interaction bias levels of each task through each interaction bias prediction task network, obtaining the respective interaction bias levels. In other words, predictions are performed simultaneously for each interaction bias prediction task, resulting in the output of each interaction bias level.

[0104] Step 306: Input the training recommendation object interaction features, combination features, and training recommendation information features into the interaction possibility prediction network to perform predictions on each interaction possibility, thereby obtaining each interaction possibility.

[0105] The interaction likelihood prediction network is a deep neural network used to predict interaction likelihood. The interaction likelihood prediction network and the interaction bias degree prediction network have a shared network.

[0106] Specifically, the server splices the training recommendation object interaction features, combination features and training recommendation information features, and inputs the spliced ​​features into the interaction possibility prediction network. The interaction possibility prediction network further extracts the interaction features to obtain each target interaction feature, and then uses each target interaction feature to perform prediction through each interaction possibility prediction task network to obtain the output of each interaction possibility.

[0107] In a specific embodiment, the interactive prediction network can be a multi-target prediction network using a PLE (a new hierarchical extraction multi-task learning network structure), such as Figure 4Figure 2 shows the network architecture of an interaction prediction network, which is used for news recommendation. This network has multiple target tasks, including click-through, duration, and interaction. The interaction prediction network consists of two layers, each of which is a multi-gated hybrid expert network. This is a common network architecture for multi-target learning. The expert networks are typically DNNs, with multiple expert networks used to extract different features. Gating is used to assign weights to each expert. The training recommendation object features are input into the interaction prediction network. The first layer, the multi-gated hybrid expert network, performs feature extraction to obtain the training recommendation object interaction features. These features are then used to predict the interaction bias of each task using a multi-task network, resulting in the output of each interaction bias. Simultaneously, the training recommendation object interaction features, combined features, and training recommendation information features are concatenated and input into the second layer, the multi-gated hybrid expert network, for feature extraction. The network then performs interaction probability prediction to obtain the corresponding interaction likelihood. Each prediction task generates a corresponding interaction likelihood and bias. That is, the interaction prediction network finally outputs the bias score and interaction score corresponding to the click task, the bias score and interaction score corresponding to the duration task, and the bias score and interaction score corresponding to the interactive task.

[0108] In one embodiment, the fusion network in the initial information recommendation model is used to fuse the interaction bias degrees and the interaction possibilities to obtain the fusion recommendation degree corresponding to the training recommendation information, including the following steps:

[0109] Vectors are spliced ​​based on the degree of each interaction bias and each interaction possibility to obtain a spliced ​​vector; the spliced ​​vector is input into the fusion network for fusion to obtain the fusion recommendation degree.

[0110] The concatenated vector refers to a vector obtained by concatenating the degrees of interaction bias and the possibilities of interaction.

[0111] Specifically, the server sequentially concatenates the interaction bias degrees and the interaction possibilities as elements in a vector to obtain a concatenated vector. The server may use the interaction bias degrees as vector elements in the left half of the concatenated vector, and then use the interaction possibilities as vector elements in the right half of the concatenated vector to obtain the concatenated vector. Alternatively, the server may use the interaction bias degrees as vector elements in the right half of the concatenated vector, and then use the interaction possibilities as vector elements in the left half of the concatenated vector to obtain the concatenated vector.

[0112] In a specific embodiment, Figure 5As shown, a schematic diagram of the network structure of a fusion network is provided. Specifically, the vectors of each interaction bias degree and each interaction possibility are concatenated to obtain a concatenated vector. This concatenated vector is then input into the fusion network, i.e., the DNN network, for fusion, resulting in an output of the fused recommendation degree. Specifically, when recommending news, the predicted bias score and interaction score corresponding to the click task, the bias score and interaction score corresponding to the duration task, and the bias score and interaction score corresponding to the interactive task are concatenated to obtain a concatenated vector. This concatenated vector is then input into the DNN fusion network for fusion, resulting in a recommendation score. This recommendation score is then used to determine whether to recommend the corresponding information.

[0113] In the above embodiment, the initial information recommendation model is established by using the interactive prediction network and the fusion network, and then the initial information recommendation model is trained, thereby improving the accuracy of the obtained fusion recommendation level.

[0114] In one embodiment, Figure 6 As shown, step 206, i.e., calculating the fusion loss based on each interaction tag and the fusion recommendation degree to obtain fusion loss information, and using each interaction bias degree to bias-adjust the fusion loss information to obtain target loss information, includes:

[0115] Step 602 : Calculate interaction loss based on each interaction tag and the fusion recommendation degree to obtain each interaction loss information.

[0116] Among them, the interaction loss information is used to characterize the error between the fusion recommendation degree and the interaction label. The smaller the interaction loss information, the closer the fusion recommendation degree is to the actual interaction result corresponding to the interaction label.

[0117] Specifically, the server can use the cross entropy loss function to calculate the loss between each interaction tag and the fusion recommendation degree, and obtain the interaction loss information corresponding to each interaction tag.

[0118] Step 604: Obtain the interaction weights corresponding to the interaction labels, and perform fusion loss calculation based on the interaction weights and the interaction loss information to obtain fusion loss information.

[0119] The interaction weight is used to represent the importance of the corresponding interaction task. Different interaction tasks have different interaction weights, which can be pre-set and can be adjusted according to needs.

[0120] Specifically, the server can pre-store interaction weights in a database and directly retrieve the interaction weights corresponding to each interaction tag from the database when needed. The server can also obtain the interaction weights corresponding to each interaction tag uploaded by the terminal in real time. The server can also obtain the interaction weights corresponding to each interaction tag from the business side. The interaction weights are then used to weight the corresponding interaction loss information, and the weighted sum is calculated to obtain the fusion loss information.

[0121] Step 606: Determine the sample weight corresponding to the training sample based on each interaction label and each interaction bias degree, and perform bias adjustment on the fusion loss information based on the sample weight to obtain the target loss information.

[0122] The sample weight is used to represent the importance of the training sample relative to the training recommendation object. The higher the sample weight is, the more the interaction tendency represented by the training sample can represent the interaction tendency of the training recommendation object.

[0123] Specifically, the server uses each interaction label and each interaction bias to determine the sample weight corresponding to the training sample. The sample weight is then used to perform a weighted calculation on the fusion loss information to obtain the target loss information corresponding to the training sample.

[0124] In the above embodiment, the fusion loss is calculated using each interaction weight and each interaction loss information to obtain the fusion loss information, and then the sample weights corresponding to the training samples are determined using each interaction label and each interaction bias degree. The fusion loss information is biased and adjusted based on the sample weights to obtain the target loss information, thereby improving the accuracy of the obtained target loss information.

[0125] In one embodiment, interaction loss calculation is performed based on each interaction tag and the fusion recommendation degree to obtain each interaction loss information, including the steps of: using each interaction tag and the fusion recommendation degree to perform cross entropy loss calculation respectively to obtain interaction loss information corresponding to each interaction tag.

[0126] Specifically, the server can use the cross entropy loss function to calculate the cross entropy loss, wherein different cross entropy loss functions can be used for different tasks. When the interaction label is the label of a binary classification task, the binary cross entropy loss function can be used to calculate the loss information of the binary classification task. For example, when the interaction label is the label of a click prediction task, the click prediction task is a binary classification task, and the binary cross entropy loss function can be used to calculate the loss information. When the interaction label is the label of a multi-classification task, the multi-classification cross entropy loss function can be used to calculate the loss information of the binary classification task. For example, when the interaction label is the label of a duration prediction task, the duration prediction task can be a multi-classification task, and the multi-classification cross entropy loss function can be used to calculate the loss information. In one embodiment, the server can also use a linear regression loss function, at least a loss function, etc. to calculate the loss information between the interaction label and the degree of fusion recommendation.

[0127] In the above embodiment, the server calculates the cross entropy loss by using each interaction tag and the fusion recommendation degree to obtain the interaction loss information corresponding to each interaction tag, thereby improving the accuracy of the obtained interaction loss information.

[0128] In one embodiment, each interaction tag includes a click tag, a reading time tag, an interaction tag, and an exposure tag;

[0129] like Figure 7 As shown in the figure, each interaction label is used to calculate the cross entropy loss with the fusion recommendation degree, and the interaction loss information corresponding to each interaction label is obtained, including:

[0130] Step 702: Calculate the cross entropy loss between the click label and the fusion recommendation degree to obtain click prediction loss information corresponding to the click label;

[0131] Step 704: Calculate the cross entropy loss between the reading time label and the fusion recommendation level to obtain reading time prediction loss information corresponding to the reading time label.

[0132] Step 706: Calculate the cross entropy loss between the interaction tag and the fusion recommendation degree to obtain interaction prediction loss information corresponding to the interaction tag;

[0133] Step 708 : Calculate the cross entropy loss between the exposure label and the fusion recommendation degree to obtain exposure prediction loss information corresponding to the exposure label.

[0134] Among them, the click label refers to the label of whether the training recommendation object clicks on the training recommendation information, and the click label is the label of the click task. The reading time label refers to the label of whether the training recommendation object reads the information for more than the preset time, and the reading time label is the label of the reading time task. The interaction label refers to the label of whether the training recommendation object interacts with the training recommendation information, and the interaction label is the label of the interaction task. The interaction can be the training recommendation object commenting, liking, collecting, forwarding, and other interactive operations on the training recommendation information. The exposure label is a label used to characterize whether the training recommendation information has been exposed. Usually, the exposure task has been completed after the information is recommended.

[0135] Specifically, each interactive tag includes but is not limited to a click tag, a reading time tag, an interaction tag, and an exposure tag. The cross-entropy loss function can then be used to calculate the loss information for each tag, wherein the server uses a binary cross-entropy loss function to calculate the cross-entropy loss of the click tag and the fusion recommendation degree, and obtains the click prediction loss information corresponding to the click tag. At the same time, the binary cross-entropy loss function is used to calculate the cross-entropy loss of the reading time tag and the fusion recommendation degree, and obtains the reading time prediction loss information corresponding to the reading time tag. At the same time, the binary cross-entropy loss function is used to calculate the cross-entropy loss of the interaction tag and the fusion recommendation degree, and obtains the interaction prediction loss information corresponding to the interaction tag; at the same time, the binary cross-entropy loss function is used to calculate the cross-entropy loss of the exposure tag and the fusion recommendation degree, and obtains the exposure prediction loss information corresponding to the exposure tag.

[0136] In the above embodiment, the loss information of different interaction tags is calculated by using a loss function, thereby improving the efficiency of the obtained loss information.

[0137] In one embodiment, Figure 8 As shown, step 604 obtains the interaction weight corresponding to each interaction label, and calculates the fusion loss based on each interaction weight and each interaction loss information to obtain the fusion loss information, including:

[0138] Step 802: Obtain the interaction priority corresponding to each interaction tag, and determine the sample type corresponding to the training sample based on the interaction priority.

[0139] The interaction priority represents the importance of each interaction task relative to the information recommendation service. A higher priority indicates a more important interaction task. The interaction priority can be set as needed. For example, interaction tasks can be given the highest priority, followed by duration tasks and click tasks.

[0140] Specifically, the server can obtain the interaction priorities corresponding to each interaction tag from the database, or the interaction priorities corresponding to each interaction tag uploaded by the terminal, and then use the interaction task corresponding to the interaction tag with the highest interaction priority as the sample type corresponding to the training sample. For example, if the interaction task has the highest priority, the sample type corresponding to this training sample is an interaction sample.

[0141] Step 804 : searching for each interaction weight corresponding to the sample type from a preset interaction matrix based on the sample type.

[0142] The preset interaction matrix is ​​a pre-set matrix that stores interaction weights. The rows of this matrix represent sample types, and the columns represent interaction tasks. For example, if there are four sample types and three interaction tasks, the preset interaction matrix is ​​a 4x3 matrix.

[0143] Specifically, the server searches for each interaction weight corresponding to the same sample type from the prediction interaction matrix according to the sample type.

[0144] In a specific embodiment, the preset interaction matrix may be as shown in Table 1 below.

[0145] Table 1 Preset interaction matrix

[0146] W click W duration W Interactive interactive 0 0 1.0 Duration 0 0.7 0.6 Click 0.1 0.9 0 exposure 1.0 0 0

[0147] The interaction weights for interactive samples include a click weight of 0, a duration weight of 1, and an interaction weight of 1. The interaction weights for duration samples include a click weight of 0, a duration weight of 0.7, and an interaction weight of 0.6. The interaction weights for click samples include a click weight of 0.1, a duration weight of 0.9, and an interaction weight of 0. The interaction weights for exposure samples include a click weight of 1, a duration weight of 0, and an interaction weight of 0.

[0148] Step 806: Use each interaction weight to weight each interaction loss information to obtain each weighted loss information, and calculate the sum of the information of each weighted loss information to obtain fusion loss information.

[0149] Among them, the weighted loss information refers to the loss information obtained by weighting the interaction loss information using the interaction weight.

[0150] Specifically, the server performs a weighted calculation on each interaction loss using the corresponding interaction weight. For example, for a click sample, the server calculates the interaction loss of the click task multiplied by a click weight of 0.1 to obtain the weighted loss information for the click task. The server then calculates the interaction loss of the duration task multiplied by a duration weight of 0.9 to obtain the weighted loss information for the duration task. The server then calculates the interaction loss of the interactive task multiplied by an interaction weight of 0 to obtain the weighted loss information for the interactive task. The server then calculates the sum of all weighted loss information to obtain the fused loss information.

[0151] In a specific embodiment, the fusion loss information can be calculated using the following formula (1).

[0152]

[0153] Among them, loss matrix Refers to the fusion loss information. loss task Refers to the interaction loss information corresponding to the interactive task. Refers to the interaction weight corresponding to the interaction task.

[0154] In the above embodiment, the sample type corresponding to the training sample is determined by using the interaction priority, and then the various interaction weights corresponding to the sample type are found from the preset interaction matrix. Finally, the various interaction weights are used to weight the various interaction loss information to obtain the various weighted loss information, and the sum of the information of the various weighted loss information is calculated to obtain the fusion loss information, thereby improving the accuracy of the obtained fusion loss information.

[0155] In one embodiment, Figure 9 As shown, step 606, that is, determining the sample weight corresponding to the training sample based on each interaction label and each interaction bias degree, and bias-adjusting the fusion loss information based on the sample weight to obtain the target loss information, includes:

[0156] Step 902: Obtain the interaction priority corresponding to each interaction tag, and determine the sample type corresponding to the training sample based on the interaction priority.

[0157] The interaction priority represents the relative importance of each interaction task relative to the recommendation. A higher interaction priority indicates a more important interaction task. This interaction priority can be set as needed. For example, interaction tasks can be given the highest priority, followed by duration tasks and click tasks.

[0158] Specifically, the server can obtain the interaction priorities corresponding to each interaction tag from the database, or the interaction priorities corresponding to each interaction tag uploaded by the terminal, and then use the interaction task corresponding to the interaction tag with the highest interaction priority as the sample type corresponding to the training sample. For example, if the interaction task has the highest priority, the sample type corresponding to this training sample is an interaction sample.

[0159] Step 904 : Acquire corresponding training sample sequences based on each interaction tag, and determine a target sample sequence from each training sample sequence based on the sample type.

[0160] Among them, the training sample sequence refers to the sequence obtained by sorting each training sample according to the interaction bias degree corresponding to the training sample. The training samples are sorted using the interaction bias degree of different interaction tasks to obtain the training sample sequence corresponding to each interaction task. The target sample sequence refers to the training sample sequence with the same sample type as the training sample, that is, the target sample sequence is obtained by sorting each training sample according to the interaction bias degree of the interaction task corresponding to the sample type. For example, if the sample type of the training sample is a click sample, the click bias degree of the click task corresponding to each training sample is obtained, and then the training samples are sorted according to the click bias degree of the click task corresponding to each training sample to obtain the training sample sequence.

[0161] Specifically, each interactive task corresponding to an interactive label has a corresponding pre-arranged training sample sequence. This training sample sequence is obtained by sorting each training sample according to the degree of interaction bias corresponding to the training sample, and each training sample is a training sample from the same batch. The server obtains the degree of interaction bias of each training sample corresponding to different interactive tasks according to each interactive label, and then sorts each training sample according to the degree of interaction bias to obtain a training sample sequence. Since different interactive tasks have different degrees of interaction bias, each training sample corresponds to the degree of interaction bias of each interactive task, that is, each training sample can be sorted according to the degree of interaction bias of each interactive task to obtain a training sample sequence corresponding to each interactive task, that is, to obtain each training sample sequence. Then, the interactive task corresponding to the sample type of the training sample is obtained, and the training sample sequence corresponding to the interactive task is used as the target sample sequence, and each training sample sequence includes the training sample.

[0162] Step 906: Determine the sequence position corresponding to the training sample from the target sample sequence. When the sequence position exceeds a preset position threshold, obtain the sample weight corresponding to the training sample as the first target sample weight.

[0163] Step 908: When the sequence position does not exceed the preset position threshold, the sample weight corresponding to the training sample is obtained as the second target sample weight.

[0164] The first target sample weight and the second target sample weight are both pre-set sample weights. The first target sample weight is greater than the second target sample weight. The preset position threshold refers to a pre-set position threshold for determining the sample weight. For example, the preset position threshold may be 20% of the sequence position ranking.

[0165] Specifically, the server determines the sequence position corresponding to the training sample from the target sample sequence, that is, determines the sorting position of the training sample in the target sample sequence, and then compares the sequence position corresponding to the training sample with the preset position threshold. When the sequence position exceeds the preset position threshold, the training sample is ranked at the front. At this time, the first target sample weight is used as the sample weight corresponding to the training sample. When the sequence position exceeds the preset position threshold, it means that the training sample hits the bias of the recommended object. At this time, the training sample can better characterize the bias of the recommended object, and a higher weight is given to the training sample. When the sequence position does not exceed the preset position threshold, the training sample is ranked at the back, which means that the training sample is an ordinary sample. At this time, a lower weight is given to the training sample.

[0166] In a specific embodiment, the fusion loss information can be calculated using the following formula (2).

[0167] loss = w bias loss matrix Formula (2)

[0168] Among them, loss refers to the target loss information, loss matrix Refers to the fusion loss information, w bias It refers to the sample weight corresponding to the training sample. The value of the sample weight can be expressed as follows:

[0169]

[0170] in, If the sequence position of the training sample is in the first 20% of the training sample sequence, the weight of the selected sample is 4. If the sequence position of the training sample is not in the first 20% of the training sample sequence, the weight of the selected sample is 1.

[0171] In the above embodiment, the target sample sequence is determined from each training sample sequence by the sample type, and then the sequence position corresponding to the training sample is determined from the target sample sequence. Finally, the sample weight corresponding to the training sample is determined according to the sequence position, so that the obtained sample weight can enhance the tendency of the training recommendation object itself, thereby improving the model training effect and enabling the target information recommendation model obtained by training to improve the accuracy of information recommendation.

[0172] In one embodiment, step 904, i.e., obtaining corresponding training sample sequences based on each interaction tag, includes the following steps:

[0173] Obtain each training sample and each interaction bias degree corresponding to each training sample; sort each training sample according to each interaction bias degree corresponding to each training sample based on each interaction label, and obtain a training sample sequence corresponding to each interaction label.

[0174] Specifically, the server can obtain each training sample and input each training sample into the initial information recommendation model in sequence to obtain the interaction bias degrees corresponding to each training sample. The server can also directly obtain each training sample and the interaction bias degrees corresponding to each training sample from the database. Then, for each interaction task corresponding to the interaction label, all training samples are sorted according to all the interaction bias degrees corresponding to the interaction task, which can be sorted from large to small, and finally a training sample sequence for the interaction task corresponding to each interaction label is obtained. That is, by using the interaction bias degree to sort each training sample, a training sample sequence corresponding to each interaction label is obtained, thereby improving the accuracy of the obtained training sample sequence.

[0175] In one embodiment, Figure 10 As shown, the information recommendation model training method also includes:

[0176] Step 1002: Obtain a training sample set, where the training sample set includes various training samples.

[0177] In step 1004, each training sample is input into the initial information recommendation model in sequence. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. Based on each interaction bias degree and each interaction possibility, fusion is performed to obtain the fusion recommendation degree corresponding to each training sample.

[0178] The training sample set refers to a collection of training samples, and the training sample set may be a batch of samples used during training.

[0179] Specifically, the server can train the initial information recommendation model in batches, that is, each training sample can be input into the initial information recommendation model in turn, and the initial information recommendation model processes each training sample in turn, that is, the interaction bias prediction of the training recommendation object information in each training sample is performed to obtain the interaction bias degree corresponding to each training sample, and the interaction prediction of the training recommendation object information and the training recommendation information in each training sample is performed to obtain the interaction possibility corresponding to each training sample, and then the interaction bias degree and the interaction possibility are fused to obtain the fusion recommendation degree corresponding to each training sample.

[0180] Step 1006 , calculate the fusion loss based on each interaction label and the fusion recommendation degree to obtain fusion loss information, and use each interaction bias degree to adjust the fusion loss information to obtain the target loss information corresponding to each training sample.

[0181] Specifically, the server calculates the target loss information for each training sample, that is, it calculates the cross entropy loss of each interaction label and the fusion recommendation degree for each training sample in turn, obtains the interaction loss information corresponding to each training sample, obtains the interaction priority corresponding to each interaction label, determines the sample type corresponding to the training sample based on the interaction priority, and searches for the interaction weights corresponding to the sample type from the preset interaction matrix based on the sample type; uses each interaction weight to weight each interaction loss information respectively to obtain each weighted loss information, and calculates the sum of the information of each weighted loss information to obtain the fusion loss information corresponding to each training sample.

[0182] Then, each training sample is sorted according to the interaction task according to the interaction bias degree corresponding to each training sample, and the training sample sequence corresponding to each interaction task is obtained. That is, the same interaction task will output different interaction bias degrees for each training sample through the initial information recommendation model. Each training sample is sorted according to the different interaction bias degrees to obtain the training sample sequence corresponding to the same interaction task. Each interaction task is traversed to obtain all training sample sequences.

[0183] Then, the interaction priority corresponding to each training sample is obtained. Based on the interaction priority, the sample type corresponding to the training sample is determined. Based on the sample type, a target sample sequence is determined from each training sample sequence. The sequence position corresponding to the training sample is determined from the target sample sequence. The sample weight corresponding to each training sample is determined based on the sequence position corresponding to the training sample. That is, when the sequence position of the training sample exceeds the preset position threshold, the sample weight is the first target sample weight. When the sequence position of the training sample does not exceed the preset position threshold, the sample weight is the second target sample weight. The sample weight corresponding to each training sample is then used to perform a weighted calculation on the fusion loss information corresponding to each training sample to obtain the target loss information corresponding to each training sample.

[0184] Step 1008: Perform an average calculation based on the target loss information corresponding to each training sample to obtain model loss information.

[0185] Specifically, the server calculates the sum of the target loss information corresponding to each training sample, and then performs an average calculation to obtain the model loss information during training of the training sample set of this batch.

[0186] Step 1010: reversely update the initial information recommendation model based on the model loss information to obtain an updated information recommendation model, use the updated information recommendation model as the initial information recommendation model, and return to the step of obtaining the training sample set until the training completion condition is met to obtain the second target information recommendation model.

[0187] Specifically, the server uses the model loss information to reversely update the parameters in the initial information recommendation model through the gradient descent algorithm to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining the training sample set is returned to execute until the training completion condition is met to obtain the second target information recommendation model.

[0188] In the above embodiment, by obtaining a training sample set, the initial information recommendation model is trained using the training sample set to obtain model loss information, and then the model loss information is used to update the initial information recommendation model, thereby obtaining a second target information recommendation model, thereby improving the efficiency and accuracy of model training.

[0189] In one embodiment, Figure 11 As shown, an information recommendation method is provided, which is applied to Figure 1 In the example of the server in the embodiment, it is understood that the method can also be applied to the server, and can also be applied to the system including the terminal and the server, and is implemented through the interaction between the terminal and the server. In this embodiment, the following steps are included:

[0190] Step 1102: Obtain information to be recommended and information about the object to be recommended.

[0191] The recommended information refers to information that needs to be determined for recommendation. This information may include, but is not limited to, video, images, text, and voice. The recommended object information refers to information about the object to be recommended. The recommended object refers to the object for which a determination is made regarding whether to recommend the recommended information. This object may be a real object or a virtual object.

[0192] Specifically, the server can obtain the information to be recommended and the information of the recommended object from the database. The server can also obtain the information to be recommended and the information of the recommended object uploaded by the terminal. The server can also obtain the information to be recommended and the information of the recommended object from the business party.

[0193] Step 1104: input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain the target recommendation degree, wherein the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses based on each interaction bias degree and each interaction possibility to obtain the fusion recommendation degree corresponding to the training recommendation information; calculates the fusion loss based on each interaction label and the fusion recommendation degree to obtain the fusion loss information, and uses each interaction bias degree to adjust the bias of the fusion loss information to obtain the target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model.

[0194] The target recommendation degree refers to the possibility of recommending the recommendation information to be trained to the object corresponding to the object information to be recommended. The higher the target recommendation degree, the higher the possibility of recommending the recommendation information to be trained to the object corresponding to the object information to be recommended.

[0195] Specifically, the server uses the recommendation information to be trained and the object information to be recommended as the input of the first target information recommendation model. The first target information recommendation model can be a model trained by any one of the embodiments of the information recommendation model training method described above. Then, the first target information recommendation model can predict the degree of interaction bias of the object information to be recommended through the interaction prediction network to obtain each degree of interaction bias, and perform each interaction prediction based on the object information to be recommended and the information to be recommended to obtain each interaction possibility, and then fuse each degree of interaction bias and each interaction possibility through the fusion network to obtain the target recommendation degree. Among them, each degree of interaction bias has a corresponding interaction possibility, that is, the same interaction task has a corresponding degree of interaction bias and a corresponding interaction possibility. In a specific embodiment, both the degree of interaction bias and the interaction possibility can be expressed as a probability or a score.

[0196] Step 1106: When the target recommendation degree exceeds the preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the object information to be recommended.

[0197] The preset recommendation level threshold refers to a preset recommendation level threshold.

[0198] Specifically, when the target recommendation degree does not exceed the preset recommendation degree threshold, it indicates that the information to be recommended does not need to be recommended to the target recommendation object corresponding to the recommended object information. At this time, a prediction can be made as to whether the next information to be recommended should be recommended. When the target recommendation degree exceeds the preset recommendation degree threshold, it indicates that the information to be recommended can be recommended to the target recommendation object corresponding to the recommended object information. At this time, the server can recommend the information to the user terminal corresponding to the target recommendation object. The user terminal includes but is not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc.

[0199] The above-mentioned information recommendation method, device, computer equipment, storage medium and computer program product input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain the target recommendation degree, and use the training recommendation object information to perform interaction bias prediction to obtain each interaction bias degree, and use the training recommendation object information and the training recommendation information to perform interaction prediction to obtain each interaction possibility, and then use each interaction bias degree and each interaction possibility to fuse to obtain the fused recommendation degree corresponding to the training recommendation information, that is, by fusing the interaction bias degree during fusion, the accuracy of the obtained fused recommendation degree is improved, and then each interaction label and the fused recommendation degree are used to perform fusion loss calculation. Calculation is performed to obtain fusion loss information, and each interaction bias degree is used to perform bias adjustment on the fusion loss information to obtain target loss information, that is, by using the interaction bias degree to perform bias adjustment on the fusion loss information, the accuracy of the obtained target loss information is improved, and then the target loss information is used to iteratively train the initial information recommendation model to obtain a first target information recommendation model, thereby improving the accuracy of the first target information recommendation model, and then the first target information recommendation model is used to recommend information, thereby improving the accuracy of the obtained target recommendation degree, and then when the target recommendation degree exceeds the preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the information of the object to be recommended, thereby improving the accuracy of the information recommendation.

[0200] In a specific embodiment, Figure 12 As shown, a method for training an information recommendation model is provided, which specifically includes the following steps:

[0201] Step 1202: Obtain a training sample set, where the training sample set includes various training samples.

[0202] Step 1204: Each training sample is sequentially input into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree. The initial information recommendation model also performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility. The initial information recommendation model then fuses the interaction bias degrees and the interaction possibilities to obtain a fused recommendation degree corresponding to each training sample.

[0203] Step 1206, calculate the cross entropy loss between each interaction label and the fusion recommendation degree for each training sample, obtain each interaction loss information, obtain the interaction priority corresponding to each interaction label, determine the sample type corresponding to the training sample based on the interaction priority, and search for each interaction weight corresponding to the sample type from the preset interaction matrix based on the sample type.

[0204] Step 1208: Use each interaction weight to weight each interaction loss information to obtain each weighted loss information, and calculate the sum of the information of each weighted loss information to obtain the fusion loss information corresponding to each training sample.

[0205] Step 1210: Sort each training sample according to its corresponding interaction bias to obtain corresponding training sample sequences. Obtain the interaction priority corresponding to each training sample, determine the sample type corresponding to the training sample based on the interaction priority, determine a target sample sequence from each training sample sequence based on the sample type, determine the sequence position corresponding to the training sample from the target sample sequence, and determine the sample weight corresponding to each training sample based on the sequence position corresponding to the training sample.

[0206] Step 1212: Use the sample weight corresponding to each training sample to perform weighted calculation on the fusion loss information corresponding to each training sample to obtain the target loss information corresponding to each training sample. Perform an average calculation based on the target loss information corresponding to each training sample to obtain the model loss information.

[0207] Step 1214: reversely update the initial information recommendation model based on the model loss information to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the process returns to the step of obtaining the training sample set until the training completion condition is met. The initial information recommendation model that meets the training completion condition is used as the final trained information recommendation model.

[0208] In a specific embodiment, the information recommendation method is applied to a news recommendation scenario. Specifically, when the news platform recommends news, it obtains information about the news to be recommended and the object to be recommended, and inputs the information about the news to be recommended and the object to be recommended into a news recommendation model. The information recommendation model is a multi-objective recommendation model, such as Figure 13The figure shows a schematic diagram of the structure of a multi-objective recommendation model. The model inputs the news to be recommended and the object to be recommended into the multi-objective prediction network within the model, yielding the outputs of a click bias score and a click recommendation score corresponding to the click task, a duration bias score and a duration recommendation score corresponding to the duration task, and an interaction bias score and an interaction recommendation score corresponding to the interaction task. The click bias score, click recommendation score, duration bias score, duration recommendation score, interaction bias score, and interaction recommendation score are then concatenated to obtain a score vector. This score vector is then input into a multi-objective fusion network for score fusion, yielding an output recommendation score. When the recommendation score exceeds a pre-set recommendation threshold, the news platform recommends the news to be recommended to the terminal corresponding to the object to be recommended. Alternatively, the model can obtain the recommendation score corresponding to each candidate news item, and then sort the candidate news items from highest to lowest according to their recommendation scores to obtain a candidate news sequence. The top-ranked candidate news items in the candidate news sequence are then selected as the news to be recommended. For example, the top 10 candidate news items are selected as the news to be recommended. Then the selected news to be recommended is sent to the recommended terminal. When the recommended terminal receives the recommended news, it is displayed on the news page, such as Figure 14 As shown, this is a schematic diagram of a news recommendation page, where the recommended news is displayed in the form of image and text information streams.

[0209] In a specific embodiment, the information recommendation method is applied to the news function in the instant messaging application, that is, applied to the news recommendation, and the news recommendation model is tested through AB testing. The test results obtained include: Figure 15-A The test comparison diagram of the number of clicks on the picture and text shown in the figure is as follows Figure 15-B The comparison diagram of the test result indicators is shown in Figure 2. It is obvious that after using the news recommendation model for news recommendation, the number of clicks on pictures and texts and various indicators of Kandian News have been significantly improved. In other words, the news recommendation model of this application has significantly improved the accuracy of news recommendations.

[0210] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0211] Based on the same inventive concept, embodiments of the present application also provide an information recommendation model training device and an information recommendation device for implementing the aforementioned information recommendation model training method. The implementation solution provided by the device is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations of one or more information recommendation model training devices or information recommendation device embodiments provided below can be found in the above-mentioned limitations of the information recommendation model training method and information recommendation method, and will not be repeated here.

[0212] In one embodiment, Figure 16 As shown, an information recommendation model training device 1600 is provided, including: a sample acquisition module 1602, an initial prediction module 1604, a loss calculation module 1606 and an iteration module 1608, wherein:

[0213] The sample acquisition module 1602 is used to acquire training samples, where the training samples include training recommendation object information, training recommendation information, and various interaction tags;

[0214] Initial prediction module 1604 is configured to input the training recommendation object information and the training recommendation information into the initial information recommendation model. The initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fuses the various interaction bias degrees and the various interaction possibilities to obtain a fused recommendation degree corresponding to the training recommendation information.

[0215] The loss calculation module 1606 is used to calculate the fusion loss based on each interaction tag and the fusion recommendation degree to obtain fusion loss information, and use each interaction bias degree to adjust the fusion loss information to obtain target loss information;

[0216] Iterative module 1608 is used to reversely update the initial information recommendation model based on the target loss information to obtain an updated information recommendation model, use the updated information recommendation model as the initial information recommendation model, and return to the step of obtaining training samples until the training completion condition is met to obtain the first target information recommendation model.

[0217] In one embodiment, the initial prediction module 1604 is also used to predict each interaction bias degree of the training recommendation object information through the interaction prediction network in the initial information recommendation model to obtain each interaction bias degree, and to perform each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility; and to fuse each interaction bias degree and each interaction possibility through the fusion network in the initial information recommendation model to obtain the fusion recommendation degree corresponding to the training recommendation information.

[0218] In one embodiment, the interaction prediction network includes an interaction bias degree prediction network and an interaction possibility prediction network; the initial prediction module 1604 is also used to perform feature extraction based on the training recommendation object information and the training recommendation information respectively to obtain training recommendation object features and training recommendation information features, perform feature combination based on the training recommendation object features and the training recommendation information features to obtain combined features; input the training recommendation object features into the interaction bias degree prediction network to extract interaction features to obtain training recommendation object interaction features, perform various interaction bias degree predictions based on the training recommendation object interaction features to obtain various interaction bias degrees; input the training recommendation object interaction features, combined features and training recommendation information features into the interaction possibility prediction network to perform various interaction possibility predictions to obtain various interaction possibilities.

[0219] In one embodiment, the initial prediction module 1604 is further configured to perform vector splicing based on each interaction bias degree and each interaction possibility to obtain a splicing vector; and input the splicing vector into a fusion network for fusion to obtain a fusion recommendation degree.

[0220] In one embodiment, the loss calculation module 1606 is also used to perform interaction loss calculation based on each interaction label and the fusion recommendation degree to obtain each interaction loss information; obtain the interaction weight corresponding to each interaction label, perform fusion loss calculation based on each interaction weight and each interaction loss information to obtain fusion loss information; determine the sample weight corresponding to the training sample based on each interaction label and each interaction bias degree, and perform bias adjustment on the fusion loss information based on the sample weight to obtain the target loss information.

[0221] In one embodiment, the loss calculation module 1606 is further configured to perform cross entropy loss calculation using each interaction tag and the fusion recommendation degree to obtain interaction loss information corresponding to each interaction tag.

[0222] In one embodiment, each interaction tag includes a click tag, a reading time tag, an interaction tag, and an exposure tag;

[0223] The loss calculation module 1606 is also used to calculate the cross-entropy loss between the click label and the fusion recommendation degree to obtain the click prediction loss information corresponding to the click label; calculate the cross-entropy loss between the reading time label and the fusion recommendation degree to obtain the reading time prediction loss information corresponding to the reading time label; calculate the cross-entropy loss between the interaction label and the fusion recommendation degree to obtain the interaction prediction loss information corresponding to the interaction label; calculate the cross-entropy loss between the exposure label and the fusion recommendation degree to obtain the exposure prediction loss information corresponding to the exposure label.

[0224] In one embodiment, the loss calculation module 1606 is also used to obtain the interaction priority corresponding to each interaction label, determine the sample type corresponding to the training sample based on the interaction priority; search for each interaction weight corresponding to the sample type from a preset interaction matrix based on the sample type; use each interaction weight to weight each interaction loss information separately to obtain each weighted loss information, and calculate the sum of the information of each weighted loss information to obtain the fusion loss information.

[0225] In one embodiment, the loss calculation module 1606 is also used to obtain the interaction priority corresponding to each interaction label, and determine the sample type corresponding to the training sample based on the interaction priority; obtain the corresponding training sample sequence based on each interaction label, and determine the target sample sequence from each training sample sequence based on the sample type; determine the sequence position corresponding to the training sample from the target sample sequence, and when the sequence position exceeds a preset position threshold, the sample weight corresponding to the training sample is obtained as the first target sample weight; when the sequence position does not exceed the preset position threshold, the sample weight corresponding to the training sample is obtained as the second target sample weight.

[0226] In one embodiment, the loss calculation module 1606 is also used to obtain each training sample and each interaction bias degree corresponding to each training sample; based on each interaction label, each training sample is sorted according to each interaction bias degree corresponding to each training sample to obtain a training sample sequence corresponding to each interaction label.

[0227] In one embodiment, the information recommendation model training apparatus 1600 further includes:

[0228] The second training module is used to obtain a training sample set, which includes various training samples; each training sample is input into the initial information recommendation model in turn, and the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, and performs various interaction predictions based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fuses based on various interaction bias degrees and various interaction possibilities to obtain the fusion recommendation degrees corresponding to each training sample; fusion loss is calculated based on each interaction label and fusion recommendation degree to obtain fusion loss information, and the fusion loss information is biased and adjusted using each interaction bias degree to obtain the target loss information corresponding to each training sample; average calculation is performed based on the target loss information corresponding to each training sample to obtain model loss information; the initial information recommendation model is reversely updated based on the model loss information to obtain an updated information recommendation model, the updated information recommendation model is used as the initial information recommendation model, and the step of obtaining the training sample set is returned to execute until the training completion condition is met to obtain the second target information recommendation model.

[0229] In one embodiment, Figure 17 As shown, an information recommendation device 1700 is provided, comprising: an information acquisition module 1702, a recommendation prediction module 1704 and a recommendation module 1706, wherein:

[0230] Information acquisition module 1702, used to acquire information to be recommended and information about the object to be recommended;

[0231] The recommendation prediction module 1704 is configured to input the recommendation information to be trained and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain a target recommendation degree, wherein the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses the interaction bias degrees and the interaction possibilities to obtain a fused recommendation degree corresponding to the training recommendation information; calculates fusion loss based on each interaction label and the fused recommendation degree to obtain fusion loss information, and uses each interaction bias degree to adjust the fusion loss information to obtain target loss information; and performs iterative training based on the target loss information to obtain the first target information recommendation model;

[0232] The recommendation module 1706 is configured to recommend the information to be recommended to the target recommendation object corresponding to the object information to be recommended when the target recommendation degree exceeds a preset recommendation degree threshold.

[0233] The information recommendation model training device and the various modules within the information recommendation device may be implemented in whole or in part through software, hardware, or a combination thereof. Each of the modules may be embedded in or independent of a processor within a computer device in hardware form, or may be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0234] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 18 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. 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 computer program in the non-volatile storage medium. The database of the computer device is used to store training sample data or data to be recommended and object data to be recommended, etc. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements an information recommendation model training method or an information recommendation method.

[0235] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 19As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface, the display unit and the input device are connected to the system bus via the input / output interface. 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 and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, it implements an information recommendation model training method or an information recommendation method. The display unit of the computer device is used to form a visually visible image, and can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse, etc.

[0236] Those skilled in the art will understand that Figure 18 Or the structure shown in 19 is only a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0237] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0238] 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-mentioned method embodiments are implemented.

[0239] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0240] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions. Users can refuse or conveniently refuse pushed information, etc.

[0241] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and 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-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may 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). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0242] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0243] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for training an information recommendation model, characterized in that: The method comprises: Acquire a training sample, wherein the training sample includes training recommendation object information, training recommendation information, and various interaction tags; Inputting the training recommendation object information and the training recommendation information into an initial information recommendation model, the initial information recommendation model performing interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, performing various interaction predictions based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fusing the various interaction bias degrees and the various interaction possibilities to obtain a fused recommendation degree corresponding to the training recommendation information; Calculating fusion loss based on the interaction tags and the fusion recommendation degree to obtain fusion loss information, and using the interaction bias degrees to adjust the fusion loss information to obtain target loss information; Based on the target loss information, the initial information recommendation model is reversely updated to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining training samples is returned to execute until the training completion condition is met to obtain a first target information recommendation model.

2. The method according to claim 1, characterized in that The inputting the training recommendation object information and the training recommendation information into the initial information recommendation model, the initial information recommendation model performing interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, performing each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fusing each interaction bias degree and each interaction possibility to obtain a fused recommendation degree corresponding to the training recommendation information, including: Predicting each interaction bias degree of the training recommendation object information through the interaction prediction network in the initial information recommendation model to obtain each interaction bias degree, and performing each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility; The respective interaction bias degrees and the respective interaction possibilities are fused through a fusion network in the initial information recommendation model to obtain a fusion recommendation degree corresponding to the training recommendation information.

3. The method according to claim 2, characterized in that The interaction prediction network includes an interaction bias degree prediction network and an interaction possibility prediction network; The step of predicting each interaction bias degree of the training recommendation object information using the interaction prediction network in the initial information recommendation model to obtain each interaction bias degree, and performing each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, includes: Performing feature extraction based on the training recommendation object information and the training recommendation information to obtain training recommendation object features and training recommendation information features, and performing feature combination based on the training recommendation object features and the training recommendation information features to obtain combined features; Inputting the training recommendation object features into the interaction bias degree prediction network to extract interaction features to obtain the training recommendation object interaction features, and performing predictions on each interaction bias degree based on the training recommendation object interaction features to obtain each interaction bias degree; The training recommendation object interaction feature, the combination feature and the training recommendation information feature are input into the interaction possibility prediction network to perform respective interaction possibility predictions to obtain the respective interaction possibilities.

4. The method according to claim 2, characterized in that The fusing the interaction bias degrees and the interaction possibilities through the fusion network in the initial information recommendation model to obtain the fusion recommendation degree corresponding to the training recommendation information includes: Performing vector splicing based on the respective interaction bias degrees and the respective interaction possibilities to obtain a splicing vector; The splicing vector is input into the fusion network for fusion to obtain the fusion recommendation degree.

5. The method according to claim 1, wherein The calculating of fusion loss based on the interaction tags and the fusion recommendation degree to obtain fusion loss information, and bias-adjusting the fusion loss information using the interaction bias degrees to obtain target loss information includes: Calculating interaction loss based on each interaction tag and the fusion recommendation degree to obtain each interaction loss information; Obtaining interaction weights corresponding to the respective interaction labels, and performing fusion loss calculation based on the respective interaction weights and the respective interaction loss information to obtain fusion loss information; The sample weight corresponding to the training sample is determined based on the interaction labels and the interaction bias degrees, and the fusion loss information is bias-adjusted based on the sample weight to obtain target loss information.

6. The method according to claim 5, characterized in that The performing interaction loss calculation based on the interaction tags and the fusion recommendation degree to obtain the interaction loss information includes: Cross entropy loss calculation is performed using each interaction label and the fusion recommendation degree to obtain interaction loss information corresponding to each interaction label.

7. The method according to claim 6, characterized in that The interactive tags include click tags, reading time tags, interaction tags and exposure tags; The cross entropy loss calculation is performed using each interaction label and the fusion recommendation degree to obtain interaction loss information corresponding to each interaction label, including: Calculating the cross entropy loss between the click label and the fusion recommendation degree to obtain click prediction loss information corresponding to the click label; Calculating the cross entropy loss between the reading time label and the fusion recommendation degree to obtain reading time prediction loss information corresponding to the reading time label; Calculating the cross entropy loss between the interaction tag and the fusion recommendation degree to obtain interaction prediction loss information corresponding to the interaction tag; Calculate the cross entropy loss between the exposure label and the fusion recommendation degree to obtain exposure prediction loss information corresponding to the exposure label.

8. The method according to claim 5, characterized in that The acquiring of the interaction weights corresponding to the respective interaction labels, and performing fusion loss calculation based on the respective interaction weights and the respective interaction loss information to obtain fusion loss information includes: Obtaining interaction priorities corresponding to the respective interaction labels, and determining sample types corresponding to the training samples based on the interaction priorities; Searching for each interaction weight corresponding to the sample type from a preset interaction matrix based on the sample type; The interaction loss information is weighted using the interaction weights to obtain weighted loss information, and the sum of the weighted loss information is calculated to obtain the fusion loss information.

9. The method according to claim 5, characterized in that The determining of the sample weights corresponding to the training samples based on the interaction labels and the interaction bias degrees, and performing bias adjustment on the fusion loss information based on the sample weights to obtain target loss information includes: Obtaining interaction priorities corresponding to the respective interaction labels, and determining sample types corresponding to the training samples based on the interaction priorities; Acquire corresponding training sample sequences based on the respective interaction tags, and determine a target sample sequence from the respective training sample sequences based on the sample type; Determining a sequence position corresponding to the training sample from the target sample sequence, and when the sequence position exceeds a preset position threshold, obtaining a sample weight corresponding to the training sample as a first target sample weight; When the sequence position does not exceed the preset position threshold, the sample weight corresponding to the training sample is obtained as the second target sample weight.

10. The method according to claim 9, characterized in that The acquiring corresponding training sample sequences based on the respective interaction labels includes: Obtain each training sample and each interaction bias degree corresponding to each training sample; The training samples are sorted based on the interaction labels according to the interaction bias degrees corresponding to each training sample, to obtain training sample sequences corresponding to the interaction labels.

11. The method according to claim 9, characterized in that The method further comprises: Acquire a training sample set, wherein the training sample set includes various training samples; Inputting each of the training samples into the initial information recommendation model in sequence, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain each interaction bias degree, and performs interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility, and fuses the interaction bias degrees and the interaction possibilities to obtain a fused recommendation degree corresponding to each of the training samples; Calculating fusion loss based on each interaction label and the fusion recommendation degree to obtain fusion loss information, and using each interaction bias degree to bias-adjust the fusion loss information to obtain target loss information corresponding to each training sample; Performing an average calculation based on the target loss information corresponding to each training sample to obtain model loss information; Based on the model loss information, the initial information recommendation model is reversely updated to obtain an updated information recommendation model. The updated information recommendation model is used as the initial information recommendation model, and the step of obtaining the training sample set is returned to be executed until the training completion condition is met, thereby obtaining a second target information recommendation model.

12. An information recommendation method, characterized in that: The method comprises: Obtain information to be recommended and information about the person to be recommended; Input the information to be recommended and the object information to be recommended into the first target information recommendation model for recommendation prediction to obtain a target recommendation degree, wherein a training sample is obtained, the training sample includes training recommendation object information, training recommendation information and various interaction labels, the training recommendation object information and the training recommendation information are input into the initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, and performs various interaction predictions based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fuses the various interaction bias degrees and the various interaction possibilities to obtain a fused recommendation degree corresponding to the training recommendation information; calculates a fusion loss based on the various interaction labels and the fused recommendation degree to obtain fusion loss information, and uses the various interaction bias degrees to perform bias adjustment on the fused loss information to obtain target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model; When the target recommendation degree exceeds a preset recommendation degree threshold, the information to be recommended is recommended to the target recommendation object corresponding to the information of the object to be recommended.

13. An information recommendation model training device, characterized in that: The device comprises: A sample acquisition module is used to acquire training samples, wherein the training samples include training recommendation object information, training recommendation information and various interaction tags; an initial prediction module, configured to input the training recommendation object information and the training recommendation information into an initial information recommendation model, the initial information recommendation model performing interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, performing various interaction predictions based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fusing the various interaction bias degrees and the various interaction possibilities to obtain a fused recommendation degree corresponding to the training recommendation information; a loss calculation module, configured to calculate fusion loss based on the interaction tags and the fusion recommendation degree to obtain fusion loss information, and to adjust the fusion loss information using the interaction bias degrees to obtain target loss information; An iterative module is used to reversely update the initial information recommendation model based on the target loss information to obtain an updated information recommendation model, use the updated information recommendation model as the initial information recommendation model, and return to the step of obtaining training samples to execute until the training completion condition is met, thereby obtaining a first target information recommendation model.

14. The device according to claim 13, characterized in that The initial prediction module is further configured to predict each interaction bias degree of the training recommendation object information using the interaction prediction network in the initial information recommendation model to obtain each interaction bias degree, and to perform each interaction prediction based on the training recommendation object information and the training recommendation information to obtain each interaction possibility; The respective interaction bias degrees and the respective interaction possibilities are fused through a fusion network in the initial information recommendation model to obtain a fusion recommendation degree corresponding to the training recommendation information.

15. The device according to claim 14, characterized in that The interaction prediction network includes an interaction bias degree prediction network and an interaction possibility prediction network; The initial prediction module is also used to perform feature extraction based on the training recommendation object information and the training recommendation information respectively to obtain training recommendation object features and training recommendation information features, perform feature combination based on the training recommendation object features and the training recommendation information features to obtain combined features; input the training recommendation object features into the interaction bias degree prediction network to extract interaction features to obtain training recommendation object interaction features, perform each interaction bias degree prediction based on the training recommendation object interaction features to obtain each interaction bias degree; input the training recommendation object interaction features, the combined features and the training recommendation information features into the interaction possibility prediction network to perform each interaction possibility prediction to obtain each interaction possibility.

16. The device according to claim 14, characterized in that The initial prediction module is further configured to perform vector splicing based on the respective interaction bias degrees and the respective interaction possibilities to obtain a splicing vector; and input the splicing vector into the fusion network for fusion to obtain the fusion recommendation degree.

17. The device according to claim 13, characterized in that The loss calculation module is further configured to perform interaction loss calculation based on the interaction tags and the fusion recommendation degree to obtain interaction loss information; obtain interaction weights corresponding to the interaction tags, and perform fusion loss calculation based on the interaction weights and the interaction loss information to obtain fusion loss information; The sample weight corresponding to the training sample is determined based on the interaction labels and the interaction bias degrees, and the fusion loss information is bias-adjusted based on the sample weight to obtain target loss information.

18. The device according to claim 17, characterized in that The loss calculation module is further configured to perform cross entropy loss calculation using each interaction label and the fusion recommendation degree to obtain interaction loss information corresponding to each interaction label.

19. The device according to claim 18, characterized in that The interactive tags include click tags, reading time tags, interaction tags and exposure tags; The loss calculation module is further configured to calculate the cross entropy loss between the click label and the fusion recommendation degree to obtain click prediction loss information corresponding to the click label; calculate the cross entropy loss between the reading time label and the fusion recommendation degree to obtain reading time prediction loss information corresponding to the reading time label; Calculating the cross entropy loss between the interaction tag and the fusion recommendation degree to obtain interaction prediction loss information corresponding to the interaction tag; Calculate the cross entropy loss between the exposure label and the fusion recommendation degree to obtain exposure prediction loss information corresponding to the exposure label.

20. The device according to claim 17, wherein The loss calculation module is further configured to obtain the interaction priority corresponding to each interaction label, and determine the sample type corresponding to the training sample based on the interaction priority; Based on the sample type, each interaction weight corresponding to the sample type is searched from a preset interaction matrix; each interaction loss information is weighted using the each interaction weight to obtain each weighted loss information, and the sum of the information of each weighted loss information is calculated to obtain the fusion loss information.

21. The device according to claim 17, characterized in that The loss calculation module is further configured to obtain the interaction priority corresponding to each interaction label, and determine the sample type corresponding to the training sample based on the interaction priority; obtain the corresponding training sample sequence based on each interaction label, and determine the target sample sequence from each training sample sequence based on the sample type; determine the sequence position corresponding to the training sample from the target sample sequence, and when the sequence position exceeds a preset position threshold, obtain the sample weight corresponding to the training sample as the first target sample weight; When the sequence position does not exceed the preset position threshold, the sample weight corresponding to the training sample is obtained as the second target sample weight.

22. The device according to claim 21, characterized in that The loss calculation module is also used to obtain each training sample and each interaction bias degree corresponding to each training sample; based on the each interaction label, the each training sample is sorted according to the each interaction bias degree corresponding to each training sample, to obtain the training sample sequence corresponding to each interaction label.

23. The device according to claim 21, characterized in that The device further comprises: The second training module is used to obtain a training sample set, which includes various training samples; input each training sample into the initial information recommendation model in sequence, and the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, and performs various interaction predictions based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fuses the various interaction bias degrees and the various interaction possibilities to obtain the fusion recommendation degrees corresponding to the various training samples; calculates fusion loss based on the various interaction labels and the fusion recommendation degrees to obtain fusion loss information, and uses the various interaction bias degrees to bias-adjust the fusion loss information to obtain the target loss information corresponding to the various training samples; performs average calculation based on the target loss information corresponding to the various training samples to obtain model loss information; reversely updates the initial information recommendation model based on the model loss information to obtain an updated information recommendation model, uses the updated information recommendation model as the initial information recommendation model, and returns to the step of obtaining the training sample set for execution until the training completion condition is met to obtain the second target information recommendation model.

24. An information recommendation device, characterized in that The device comprises: An information acquisition module is used to acquire information to be recommended and information about the object to be recommended; A recommendation prediction module is used to input the information to be recommended and the object information to be recommended into a first target information recommendation model for recommendation prediction to obtain a target recommendation degree, wherein a training sample is obtained, the training sample includes training recommendation object information, training recommendation information and various interaction labels, the training recommendation object information and the training recommendation information are input into an initial information recommendation model, the initial information recommendation model performs interaction bias prediction based on the training recommendation object information to obtain various interaction bias degrees, and performs various interaction predictions based on the training recommendation object information and the training recommendation information to obtain various interaction possibilities, and fuses the various interaction bias degrees and the various interaction possibilities to obtain a fused recommendation degree corresponding to the training recommendation information; performs fusion loss calculation based on the various interaction labels and the fused recommendation degree to obtain fusion loss information, and uses the various interaction bias degrees to perform bias adjustment on the fused loss information to obtain target loss information; performs iterative training based on the target loss information to obtain the first target information recommendation model; The recommendation module is configured to recommend the information to be recommended to the target recommendation object corresponding to the object information to be recommended when the target recommendation degree exceeds a preset recommendation degree threshold.

25. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 12 are implemented.

26. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

27. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images