Information recommendation method and device and storage medium

By using the intention prediction model end-to-end training of teacher models in information recommendation, the problem of time and complexity of model update is solved, the accuracy and training efficiency of recommendations are improved, and the information recommendation effect is achieved quickly.

CN119027208APending Publication Date: 2024-11-26BEIJING XIAOMI MOBILE SOFTWARE CO LTD +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202310614755.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-26
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In the information recommendation scenario, the two-stage training strategy of the model in the prior art causes the model to be updated for a long time, increase the complexity of the online operation, and it is difficult to improve the accuracy of the recommendation.

Method used

Adopt the intention prediction model based on end-to-end training of the teacher model. By obtaining user information, the user feature vector matrix is ​​constructed and inputted into the pre-trained intention prediction model to determine whether to send recommendation information. There is no need to pre-train the teacher model and direct guidance. Training of intention prediction models.

Benefits of technology

The prediction accuracy of the intention prediction model is improved, the training time of the intention prediction model and the operation complexity of the model are reduced, and the rapid and accurate information recommendation is achieved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027208A_ABST
    Figure CN119027208A_ABST
Patent Text Reader

Abstract

The invention relates to an information recommendation method and device and a storage medium, and the method comprises the steps: obtaining user information of a target user, the user information comprising at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity; constructing a user feature vector matrix corresponding to the user information; the user feature vector matrix is input into a pre-trained intention prediction model to obtain an intention prediction label of the target user, the intention prediction model is a multi-task model obtained based on teacher model end-to-end training, and the intention prediction label comprises a prediction label corresponding to each prediction task in the intention prediction model; the teacher model corresponds to a target prediction task in the intention prediction model; and according to the intention prediction tag, determining whether to send recommendation information corresponding to the target commodity to the target user. The prediction accuracy of the intention prediction model can be ensured, and the training time of the intention prediction model is effectively shortened.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and in particular to an information recommendation method, device and storage medium. Background Art

[0002] With the development of artificial intelligence technology, machine learning and artificial intelligence technologies are becoming more and more common in people's daily lives. In related technologies, a knowledge distillation model training method is proposed. However, in search, recommendation and other scenarios, the model is often faced with online operations. At this time, if a two-stage training strategy is adopted, when updating the model, it is necessary to wait for the teacher model to be updated before the student model can be updated. The whole process is time-consuming and increases the complexity of online operations. Therefore, in the information recommendation scenario, how to improve the accuracy of information recommendation while reducing the training time is a difficult problem currently faced. Summary of the invention

[0003] In order to overcome the problems existing in the related art, the present disclosure provides an information recommendation method, device and storage medium.

[0004] According to a first aspect of an embodiment of the present disclosure, there is provided an information recommendation method, comprising:

[0005] Acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity;

[0006] Constructing a user feature vector matrix corresponding to the user information;

[0007] Inputting the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training based on a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model;

[0008] Determine whether to send recommendation information corresponding to the target product to the target user according to the intention prediction label.

[0009] Optionally, the intention prediction model includes a cross module, a plurality of Expert modules, and a Gate weight layer corresponding to each task and a Tower layer corresponding to each task;

[0010] Inputting the user feature vector matrix into the pre-trained intention prediction model to obtain the intention prediction label of the target user includes:

[0011] Inputting the user feature vector matrix into the crossover module to obtain a first feature vector, where the first feature vector is a feature vector that crosses high and low dimensions;

[0012] Inputting the first feature vector into the multiple Expert modules respectively to obtain multiple second feature vectors;

[0013] Input the plurality of second feature vectors into the Gate weight layer corresponding to each task respectively, so that each Gate weight layer performs weighted fusion on the plurality of second feature vectors according to the corresponding task to obtain a third feature vector corresponding to each task;

[0014] Each of the third feature vectors is input into the corresponding Tower layer to obtain the intention prediction label corresponding to each task.

[0015] Optionally, the training of the intention prediction model includes:

[0016] Acquire sample user information and sample label information of the sample user, wherein the sample label information includes a pre-labeled sample label corresponding to each task of the intention prediction model, and the sample user information includes at least one of the following: sample portrait information, sample attribute information of the sample product, and sample interaction information between the sample user and the sample product;

[0017] Constructing a sample feature vector matrix according to the sample portrait information, the sample attribute information and the sample interaction information;

[0018] Inputting the sample feature vector matrix into the teacher model to obtain a first teacher label corresponding to the sample user;

[0019] Inputting the sample feature vector matrix into the initial intention prediction model to obtain the sample prediction label corresponding to the sample user;

[0020] Determine a first loss corresponding to each prediction task in the initial intention prediction model according to the sample prediction label and the sample label information; determine a second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label; determine a third loss according to the first sample label corresponding to the target prediction task in the sample label information and the first teacher label;

[0021] updating the parameters corresponding to each prediction task in the initial intention prediction model according to the first loss; updating the parameters corresponding to the target prediction task in the initial intention prediction model according to the second loss; and updating the teacher model according to the third loss;

[0022] Repeat the step of obtaining sample user information and sample label information of sample users, to the step of updating the teacher model according to the third loss, until the preset conditions are met, and the trained intention prediction model is obtained.

[0023] Optionally, the teacher model further includes a calibration unit, and the method further includes:

[0024] Inputting the first predicted probability corresponding to the first teacher label into the calibration unit to obtain a calibrated second teacher label;

[0025] Determine a fourth loss according to the second teacher label and the first sample label;

[0026] updating the calibration unit according to the fourth loss;

[0027] The determining of the second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label includes:

[0028] The second loss is determined according to the first predicted label and the second teacher label.

[0029] Optionally, updating the calibration unit according to the fourth loss includes:

[0030] The calibration unit is updated according to the fourth loss using an adaptive moment estimation Adam optimizer.

[0031] Optionally, the teacher model further includes an error correction unit, and the method further includes:

[0032] inputting a second predicted probability corresponding to the second teacher label into the error correction unit;

[0033] When the confidence of the second teacher label is greater than or equal to a preset threshold, the error correction unit outputs the second teacher label;

[0034] When the confidence of the second teacher label is less than the preset threshold, the error correction unit outputs a preset third teacher label;

[0035] Wherein, the preset threshold is determined according to the first sample label;

[0036] The determining of the second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label includes:

[0037] The second loss is determined according to the first predicted label and the second teacher label or the third teacher label.

[0038] Optionally, updating the parameters corresponding to each prediction task in the initial intention prediction model according to the first loss includes:

[0039] Monitor the AUC index corresponding to each prediction task in each update of the initial intention prediction model;

[0040] The preset conditions include:

[0041] After the parameters corresponding to each prediction task in the initial intention prediction model are updated a preset number of times, the current AUC index corresponding to each prediction task in the initial intention prediction model is greater than the AUC index before the preset number of times.

[0042] According to a second aspect of an embodiment of the present disclosure, there is provided an information recommendation device, including:

[0043] An acquisition module is configured to acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity;

[0044] A construction module, configured to construct a user feature vector matrix corresponding to the user information;

[0045] A prediction module is configured to input the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training based on a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model;

[0046] The recommendation module is configured to determine whether to send recommendation information corresponding to the target product to the target user based on the intention prediction tag.

[0047] According to a third aspect of an embodiment of the present disclosure, there is provided an information recommendation device, including:

[0048] processor;

[0049] a memory for storing processor-executable instructions;

[0050] Wherein, the processor is configured to:

[0051] Acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity;

[0052] Constructing a user feature vector matrix corresponding to the user information;

[0053] Inputting the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training based on a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model;

[0054] Determine whether to send recommendation information corresponding to the target product to the target user according to the intention prediction label.

[0055] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the program instructions are executed by a processor, the steps of any one of the methods described in the first aspect of the present disclosure are implemented.

[0056] The technical solution provided by the embodiments of the present disclosure may include the following beneficial effects: by obtaining the user information of the target user and constructing a user feature vector matrix corresponding to the user information, and then inputting the user feature vector matrix into the intention prediction model obtained by end-to-end training based on the teacher model to obtain the intention prediction label corresponding to the target user, and determining whether to send the target product recommendation information to the target user based on the intention prediction label, there is no need to pre-train the teacher model and then guide the training of the intention prediction model based on the trained teacher model. This not only ensures the prediction accuracy of the intention prediction model and can accurately determine whether to send the target product recommendation information to the user, but also effectively reduces the training time of the intention prediction model and reduces the operational complexity of the model online.

[0057] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0059] Figure 1 The figure is a flowchart of an information recommendation method according to an exemplary embodiment.

[0060] Figure 2 The present invention is a flowchart of a method for training an intention prediction model according to an exemplary embodiment.

[0061] Figure 3 It is a structural diagram of end-to-end training of an intention prediction model according to an exemplary embodiment.

[0062] Figure 4 It is a schematic diagram of loss function design according to an exemplary embodiment.

[0063] Figure 5 The invention is a block diagram of an information recommendation device according to an exemplary embodiment.

[0064] Figure 6 The invention is a block diagram of a device for information recommendation according to an exemplary embodiment. DETAILED DESCRIPTION

[0065] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0066] It should be noted that all actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the relevant data protection laws and policies of the country where the device is located and with the authorization given by the owner of the corresponding device.

[0067] Figure 1 is a flowchart of an information recommendation method according to an exemplary embodiment. The execution subject of the method may be, for example, a server or any other electronic device with information processing capabilities, such as Figure 1 As shown, the method comprises the following steps:

[0068] S101: Obtain user information of a target user.

[0069] The user information includes at least one of the following: user portrait information, product attribute information of the target product, and interaction information between the target user and the target product.

[0070] For example, user portrait information may include the user's age, gender, city, etc. The target product may be, for example, a certain mobile phone, or any product. Taking a mobile phone as an example, the product attribute information of the target product may include, for example, the model, system version, number of SIM cards, etc. of the mobile phone. The interaction information between the user and the target product may include, for example, the user's usage data in various applications, the user's behavior records on the target mall page, purchase records, and other information.

[0071] It is understandable that the specific content included in the user information can be determined based on actual needs or information actually obtained, and the embodiments of the present disclosure do not specifically limit this.

[0072] S102: Construct a user feature vector matrix corresponding to the user information.

[0073] In some embodiments, the user information may be first cleaned to exclude default or erroneous data in the user information, and feature extracted to process it into the format required by the subsequent model. Optionally, the user information may be categorized and discretized to construct a feature vector matrix corresponding to the user information.

[0074] S103: Input the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user.

[0075] Among them, the intention prediction model is a multi-task model obtained through end-to-end training of the teacher model, the intention prediction label includes the prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to the target prediction task in the intention prediction model.

[0076] It can be understood that the intention prediction model is the student model of the teacher model. In the end-to-end training process based on the teacher model, the training data can be input into the teacher model and the intention prediction model respectively to obtain prediction results, so as to update the teacher model and the intention prediction model at the same time, and the knowledge of the teacher model can be transferred to the intention prediction model, without the need to teach the intention prediction model based on the pre-trained teacher model.

[0077] For example, the multi-task model may include three subtasks: overall purchase, offline purchase, and offline specific model purchase, but is not limited thereto. For example, it may also include online purchase, online specific model purchase, and other tasks. The prediction labels corresponding to the overall purchase and offline purchase subtasks may be hard labels, such as binary classification labels. For example, the label corresponding to the overall purchase task may be 0 or 1. When the label is 0, it can be represented that the user will not purchase the phone, and when the label is 1, it can be represented that the user will purchase the phone; the prediction label corresponding to the offline specific model purchase subtask may be a soft label. For example, the prediction label of the task corresponding to the target user may be [0, 0.1]; [1, 0.9]. At this time, the label may represent that the user has a probability of 0.1 not to purchase a specific model of mobile phone offline, and a probability of 0.9 to purchase a specific model of mobile phone offline. The target prediction task corresponding to the teacher model may be, for example, offline specific model purchase, and accordingly, the label output by the teacher model may also be a soft label. In other examples, the multi-task model may also include other models. For example, if the target product is a certain type of clothing, the multi-task model may include tasks such as clothing purchase, online store purchase, and offline purchase of clothing of a specific color. The disclosed embodiment does not limit what tasks the multi-task model specifically includes.

[0078] It is understandable that the main function of the teacher model is to learn the corresponding task prediction probability based on the training samples, and then guide the multi-task model, that is, the intention prediction model to learn the soft targets related to the teacher network, thereby transferring knowledge to the multi-task model, thereby improving the prediction accuracy of the target prediction task in the multi-task model and achieving a common improvement in the performance of multiple tasks.

[0079] S104: Determine whether to send recommendation information corresponding to the target product to the target user based on the intention prediction tag.

[0080] For example, whether to send the recommendation information corresponding to the target product to the target user can be determined based on whether the prediction tags corresponding to each task in the intention prediction tag meet the preset conditions. For example, the preset condition can be that the tag corresponding to the overall purchase task is 1, and the prediction tag corresponding to the offline specific model purchase subtask represents that the probability of the user purchasing the specific model offline is greater than 0.5; or, the preset condition can be that the tag corresponding to the overall purchase task is 1, and the offline purchase task is also 1, and so on. Those skilled in the art can calibrate the preset condition based on actual needs or recommendation needs. The embodiments of the present disclosure do not limit the specific scheme for determining whether to send the recommendation information corresponding to the target product to the target user.

[0081] It can be understood that in step S101, the target user may include multiple users, and correspondingly, the user information corresponding to the multiple users can be obtained respectively, and the user feature vector matrix can be constructed based on the user information of each user respectively, and the user feature vector matrix corresponding to each user can be input into the trained intention prediction model respectively to obtain the intention prediction label corresponding to each user, and based on the intention prediction label corresponding to each user, it can be determined whether to send recommendation information to each user.

[0082] In the disclosed embodiment, the user information of the target user is obtained and a user feature vector matrix corresponding to the user information is constructed, and then the user feature vector matrix is ​​input into the intention prediction model obtained through end-to-end training of the teacher model to obtain the intention prediction label corresponding to the target user, and based on the intention prediction label, it is determined whether to send the recommendation information of the target product to the target user. There is no need to pre-train the teacher model and then guide the training of the intention prediction model based on the trained teacher model. This not only ensures the prediction accuracy of the intention prediction model and accurately determines whether to send the recommendation information of the target product to the user, but also effectively reduces the training time of the intention prediction model and reduces the operational complexity of the model online.

[0083] In an exemplary scenario, after the business party obtains a population package including multiple user information, it can send new product SMS to the users in the population package. After receiving the SMS, the user can click on the SMS link to obtain detailed information about the new product and ultimately decide whether to make an appointment or purchase the new mobile phone. To a certain extent, the more accurate the predicted intended population, the higher the SMS click and conversion rate brought to the business party. Therefore, the accuracy of the predicted intention prediction label is largely related to the revenue of the SMS delivery business scenario. In the disclosed embodiment, the accuracy of the intention prediction label is concerned, specifically including the accuracy of the overall purchasing population, the accuracy of the offline purchasing population, and the accuracy of offline specific model purchases.

[0084] In some optional embodiments, the intention prediction model includes a cross module, a plurality of Expert modules, and a Gate weight layer corresponding to each task and a Tower layer corresponding to each task. In some examples, the cross module can be a Cross module or a CrossNet cross module, a plurality of Expert modules can be referred to as Experts modules, and the Gate weight layer can also be referred to as a Gate layer.

[0085] Optionally, the user feature vector matrix is ​​input into a pre-trained intention prediction model to obtain the intention prediction label of the target user, including:

[0086] The user feature vector matrix is ​​input into the cross module to obtain a first feature vector, which is a feature vector of high and low dimensional cross; the first feature vector is respectively input into the multiple Expert modules to obtain multiple second feature vectors; the multiple second feature vectors are respectively input into the Gate weight layer corresponding to each task, so that each Gate weight layer performs weighted fusion on the multiple second feature vectors according to the corresponding task to obtain a third feature vector corresponding to each task; each of the third feature vectors is input into the corresponding Tower layer to obtain an intention prediction label corresponding to each task.

[0087] It is understandable that there is obviously a certain correlation between the various tasks in the intention prediction model. The use of multiple Expert modules can effectively extract the underlying features shared by the various tasks.

[0088] In the above embodiment, different from the general multi-task model that directly learns the input feature vector, the intention prediction model in the embodiment of the present disclosure takes into account that the user feature vector matrix of the target user has a large dimension, so a cross module is added to the input layer of the model to automatically extract feature information, obtain high- and low-dimensional cross features, and achieve the purpose of dimensionality reduction at the same time, thereby reducing the complexity of the model and accelerating the model training and convergence speed. For example, the first feature vector output by the cross module can be expressed as shown in the following formula: input =[e1,e2,...,e n ].

[0089] In the above embodiment, since the cross module has extracted low-order and high-order features of the input data, each Expert module can adopt a DNN (Deep Neural Networks) structure to perform deeper feature mining. In the Expert module, the DNN structure can be implemented using a single fully connected layer, thereby enabling the Expert module to achieve a mapping of the dimension of the user feature matrix from (k+n) dimensions to m dimensions.

[0090] For example, the specific implementation of each Expert module is as follows: i (x) = W i *x input +b i , where i is the number of the Expert module, ranging from [1, N]. For example, if N = 12, W i With b i They represent the weight coefficients in Experts respectively.

[0091] Furthermore, in order to capture the differences between different tasks, a gating network, namely, a Gate weight layer, is designed separately for each task in the above embodiment, and different weights of the Expert model are obtained for different tasks, thereby realizing the selective use of the common features output by the Expert module. The Gate weight layer realizes the selective use of Experts by different output weights, thereby making it easier for the model to capture the differences between tasks.

[0092] For example, the Gate weight layer can be composed of a fully connected layer and a softmax activation function, and the input is multiple second feature vectors output by multiple Expert modules. The feature output can be expressed as: g k (x) = softmax(W gk *x input ). Where k is the number of the task, for example, k = 3, W gk is the weight coefficient of the fully connected layer in the Gate layer, g k (x) is the expert weight combination learned by the model. The expert features are weighted and fused through the Gate layer, and finally the feature representation f of each task is output. k (x), and then based on the activation function, we get g k (x), which is the third eigenvector corresponding to each task. The gating networks corresponding to different tasks can learn different Experts combination modes, so the model can more easily capture the correlation and differences between subtasks.

[0093] Taking the task of purchasing a mobile phone as an example, based on the third eigenvector corresponding to each task, the Tower layer is added to predict the probability of the user's purchase tendency. In this example, three subtasks are designed: overall purchase, offline purchase, and offline purchase of specific models. Each task is connected to the Tower layer to predict the probability of each task. Among them, the Tower layer structure is implemented by a single fully connected layer, and finally the sigmoid activation function is used to output the predicted label of each task.

[0094] Figure 2 is a flowchart of a method for training an intention prediction model according to an exemplary embodiment. The execution subject of the method may be, for example, a server. The execution subject of the method may be Figure 1 The execution entities of the information recommendation method shown in the figure are the same or different, and the embodiment of the present disclosure does not limit this. For example, the first server may execute Figure 2 The steps shown in the figure obtain the trained intention prediction model, and send the trained intention prediction model to the second server so that the second server executes Figure 1 The steps shown.

[0095] like Figure 2 As shown, the training method of the intention prediction model includes the following steps:

[0096] S201. Obtain sample user information and sample label information of sample users.

[0097] Among them, the sample label information includes the pre-labeled sample labels corresponding to each task of the intention prediction model, and the sample user information includes at least one of the following: sample portrait information, sample attribute information of sample products, and sample interaction information between the sample user and the sample product.

[0098] It is understandable that the sample label information may be label information pre-labeled based on real situations. For example, user information corresponding to a user who has purchased a certain model of mobile phone is obtained as sample user information, and the user is labeled with corresponding label information.

[0099] In some possible implementations, the sample user may include multiple sample users, and the ratio of positive and negative samples in the multiple sample users may be 1: 1. For example, a user purchase sample within 20 days may be obtained, and all positive samples therein, that is, sample user information of sample users who purchased a certain model of mobile phone, may be retained, and random sampling may be performed in the negative sample to obtain positive and negative samples with a ratio of 1: 1.

[0100] S202. Construct a sample feature vector matrix according to the sample portrait information, the sample attribute information and the sample interaction information.

[0101] The optional implementation method of constructing the sample feature vector matrix in step S202 can be found in Figure 1 The optional implementation of step S102 and other related parts in the present disclosure will not be repeated here.

[0102] S203: Input the sample feature vector matrix into the teacher model to obtain a first teacher label corresponding to the sample user.

[0103] In some embodiments, the teacher model may include a DNN module, a Tower layer and a sigmoid function. For example, the DNN module may be a deep neural network, which may be similar to the Expert module in the above embodiment. In the teacher model, the DNN module may be used to extract high-order features of the input information, wherein the DNN module may be implemented by a three-layer fully connected layer, and finally the Tower layer and the sigmoid function are input to obtain the prediction label of the corresponding target prediction task, i.e., the first teacher label.

[0104] S204: Input the sample feature vector matrix into the initial intention prediction model to obtain the sample prediction label corresponding to the sample user.

[0105] It can be understood that the initial intention prediction model can be an untrained intention prediction model, or an intention prediction model that has undergone preliminary training. The structure of the initial intention prediction model is the same as the trained intention prediction model. After training, the trained intention prediction model can be obtained.

[0106] S205. Determine the first loss corresponding to each prediction task in the initial intention prediction model according to the sample prediction label and the sample label information; determine the second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label; determine the third loss according to the first sample label corresponding to the target prediction task in the sample label information and the first teacher label.

[0107] In some embodiments, both the first loss and the third loss may be cross entropy losses, that is, loss values ​​calculated based on the cross entropy loss function. The second loss may be, for example, a distillation loss, which may represent the loss when the teacher model transfers knowledge to the intention prediction model. Optionally, the calculation formula of the cross entropy loss function may be as follows: Where N is the number of samples, y i represents the true label of the training sample, that is, the label in the sample prediction label, p i represents the predicted probability of the model.

[0108] For example, the first loss may include the L calculated by comparing the hard label of the overall purchase predicted by the initial intention prediction model with the corresponding sample label in the sample label information. buy , the L calculated by the hard label of offline purchase predicted by the initial intention prediction model and the corresponding sample label in the sample label information offline And, the hard label of offline specific model purchase predicted by the initial intention prediction model and the corresponding sample label L in the sample label information student In addition, the second loss can be expressed as L kd , the third loss can be expressed as L teacher .

[0109] S206. Update the parameters corresponding to each prediction task in the initial intention prediction model according to the first loss; update the parameters corresponding to the target prediction task in the initial intention prediction model according to the second loss; and update the teacher model according to the third loss.

[0110] For example, updating the parameters corresponding to each prediction task may include, for example, updating each Expert module, and at least one of the Gate weight layer and the Tower layer corresponding to each prediction task. Updating the teacher model may include, for example, updating the DNN module and / or the Tower layer.

[0111] In some embodiments, the step of obtaining sample user information and sample label information of the sample user in step S201 is repeated, and the step of updating the teacher model according to the third loss in step S206 is repeated until the preset conditions are met to obtain the trained intention prediction model.

[0112] In the disclosed embodiment, the sample feature vector matrix corresponding to the sample user is input into the teacher model and the initial intention prediction model respectively, and the first teacher label output by the teacher model and the sample prediction label corresponding to each task in the initial intention prediction model are obtained, and then the first loss, the second loss and the third loss are determined, and then the teacher model and the initial intention prediction model are finally trained to obtain the intention prediction model, which can effectively ensure the accuracy of the trained intention prediction model, so that the trained intention prediction model can accurately determine whether to send the recommendation information of the target product to the user, and an end-to-end training method is adopted, without the need to train the teacher model first, which effectively reduces the training time of the intention prediction model.

[0113] In some optional embodiments, the teacher model further includes a calibration unit, and the method further includes:

[0114] Inputting the first predicted probability corresponding to the first teacher label into the calibration unit to obtain a calibrated second teacher label;

[0115] Determine a fourth loss according to the second teacher label and the first sample label;

[0116] updating the calibration unit according to the fourth loss;

[0117] The determining the second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label includes:

[0118] The second loss is determined according to the first predicted label and the second teacher label.

[0119] For example, the fourth loss can be expressed as L cal , the fourth loss can be a cross entropy loss.

[0120] It is understandable that since the performance of the teacher model is not stable in the early stage of training, directly sending the first teacher label output by it as a soft target to the student module may interfere with the learning of the student module. Therefore, this scheme introduces a calibration unit to learn unbiased probability estimation.

[0121] For example, the calibration unit can design a linear layer and a sigmoid activation function to transform the output value of the teacher module. The specific implementation method is shown in the following formula: cal =w cal ·r teacher +b cal ;

[0122] Among them, r teacher represents the logit value output by the teacher module, that is, the first predicted probability corresponding to the first teacher label, r cal represents the calibrated logit value, y cal It represents the output probability value after activation by the sigmoid function, which ranges from 0 to 1. It can be understood that the logit value can be a feature representation that has not been activated by the activation function.

[0123] Optionally, to ensure that the calibrated probability is as consistent as possible with the label corresponding to the target prediction task, this scheme adds a calibration loss L cal , which is the fourth loss, to update the parameters w of the calibration unit cal With b cal ,The calibration loss is implemented using the cross entropy loss function.

[0124] By adopting the above scheme, a calibration unit is designed to transform the output value of the teacher model, and the second loss is calculated based on the calibrated second teacher label to update the parameters corresponding to the target prediction task in the initial intention prediction model, thereby further improving the accuracy of the trained intention prediction model. In addition, the gradient back propagation of the calibration unit is cut off, effectively avoiding the interference of the student module on the teacher module during the distillation process.

[0125] In some optional embodiments, updating the calibration unit according to the fourth loss includes:

[0126] The calibration unit is updated according to the fourth loss using an Adam (Adaptive Moment Estimation) optimizer.

[0127] Among them, the Adam optimizer can calculate the adaptive learning rate of each parameter. This method not only stores the exponentially decaying average of the previous squared gradient, but also maintains the exponentially decaying average of the previous gradient.

[0128] Using the above scheme, compared with other adaptive learning rate algorithms, the Adam optimizer converges faster and has more effective learning effects. It can also correct problems existing in other optimization techniques, such as learning rate disappearance, slow convergence, or high variance parameter updates leading to large fluctuations in the loss function. It effectively ensures the accuracy of the calibration unit and effectively avoids the problem that the performance of the teacher module is still unstable in the early stage of training, resulting in inaccurate intention prediction models obtained through training.

[0129] In some embodiments, the teacher model further includes an error correction unit, and the method further includes:

[0130] inputting a second predicted probability corresponding to the second teacher label into the error correction unit;

[0131] When the confidence of the second teacher label is greater than or equal to a preset threshold, the error correction unit outputs the second teacher label;

[0132] When the confidence of the second teacher label is less than the preset threshold, the error correction unit outputs a preset third teacher label;

[0133] Wherein, the preset threshold is determined according to the first sample label;

[0134] The determining of the second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label includes:

[0135] The second loss is determined according to the first predicted label and the second teacher label or the third teacher label.

[0136] Optionally, a fault tolerance threshold m can be set in the error correction unit to restrict and correct the inaccurate structure predicted by the teacher module. The implementation method is shown in the following formula: cal ←[y]·max{[y]·r cal ,m}. Among them, r cal is the calibrated logit value of the teacher module, which is the second predicted probability mentioned above. [y] is an indicator function. When y is 1, that is, when y is a positive sample, [y] = 1, otherwise, [y] = 0. m is the fault tolerance threshold, which is a hyperparameter. For example, the value of m can be set to 0. When the teacher module is correct enough, that is, the predicted probability is at least the preset threshold σ(m), the output of the teacher module will not be affected. If the prediction is wrong, the output value of the corrected teacher model is m, which is the third teacher label.

[0137] For example, when y is a positive sample, the corresponding error correction module is r cal The value of is as large as possible. On the contrary, when y is a negative sample, the corresponding r calThe value of is as small as possible. That is to say, the preset threshold σ(m) may be determined based on y, for example, σ(m) may be positively correlated with y.

[0138] By adopting the above scheme, the error correction unit can be used to correct the learning errors of the teacher model, eliminate the inaccurate predictions of the teacher module in the early stage of training to assume convergence, and further improve the quality of knowledge in the later stage to improve the performance of the trained intention prediction model. In addition, the gradient back propagation of the error correction unit is cut off, which effectively avoids the interference of the student module on the teacher module during the distillation process.

[0139] In some optional embodiments, updating the parameters corresponding to each prediction task in the initial intention prediction model according to the first loss includes:

[0140] Monitor the AUC (Area Under Curve) indicator corresponding to each prediction task in the initial intention prediction model each time it is updated;

[0141] The preset conditions include:

[0142] After the parameters corresponding to each prediction task in the initial intention prediction model are updated a preset number of times, the current AUC index corresponding to each prediction task in the initial intention prediction model is greater than the AUC index before the preset number of times.

[0143] Among them, the AUC index represents the area under the ROC (receiver operating characteristic curve) curve. The AUC index is an evaluation index for binary classification models in machine learning. The AUC index can reflect the model's ability to rank samples. Its statistical significance is the probability that a positive sample is randomly selected from all positive samples and a negative sample is randomly selected from all negative samples, and the current score makes the positive sample rank ahead of the negative sample.

[0144] Furthermore, the drawing of the ROC curve can include the following steps: sorting each test sample from large to small according to the probability value of the test sample belonging to the positive sample; from high to low, taking the "Score" value as the threshold, when the probability of the test sample belonging to the positive sample is greater than or equal to this threshold, we consider it to be a positive sample, otherwise it is a negative sample; each time a different threshold is selected, we can get a set of FPR (False positive rate) and TPR (True positive rate), that is, a point on the ROC curve.

[0145] When the threshold is set to 1 and 0, we can get the points (0,0) and (1,1) on the ROC curve respectively. Connecting these (FPR, TPR) pairs, we get the ROC curve. The more threshold values ​​we have, the smoother the ROC curve becomes. AUC is the area under the ROC curve. The closer the AUC is to 1, the better the classifier performance is, which means the higher the prediction accuracy of the corresponding task is.

[0146] In addition, the preset number of times may be a pre-calibrated update round, for example, the preset number of times may be, for example, epoch=5. After each pair of initial prediction models are updated 5 times, the AUC of the model in the current round on the test set can be obtained, and the AUC in the current update round is compared with the AUC value of the previous round (the initial AUC is set to 0). If the AUC value of the current round is the highest, it is considered that the model obtained by the current round of training has better performance, and the model parameters of the current round are saved as the best model, that is, the trained intention prediction model; otherwise, continue training for the next round until the trained intention prediction model is obtained.

[0147] In some other optional embodiments, except L cal The Adam optimizer can also be used to update model parameters for losses other than the loss, and the offline purchase prediction model with the best performance can be finally obtained by monitoring the AUC indicators of the model on the three tasks.

[0148] Based on the above overall inventive concept, the present disclosure provides the following embodiments: Figure 3 A schematic diagram of the structure of an end-to-end training of an intention prediction model as shown in Figure 4 A schematic diagram of loss function design is shown in FIG. Figure 4 The loss function design diagram shown can be based on Figure 3 The intention prediction model shown is designed with the structure of the teacher model.

[0149] Reference Figure 3 As shown, the intention prediction model includes an input layer, a cross module, multiple Expert modules, and three Gate layers and three Tower layers corresponding to the first to third tasks respectively, and the third task of the intention prediction model corresponds to the teacher model. The teacher model includes a DNN module, a Tower layer, a calibration unit, and an error correction unit.

[0150] For example, the first task may be a task of purchasing a mobile phone, the second task may be a task of purchasing a mobile phone offline, and the third task may be a task of purchasing a specific mobile phone model offline. The predicted label output by the Tower layer corresponding to the first task may be, for example, a binary classification label corresponding to the task of purchasing a mobile phone, the predicted label output by the Tower layer corresponding to the second task may be, for example, a binary classification label corresponding to the task of purchasing a mobile phone offline, and the soft label output by the Tower layer corresponding to the third task.

[0151] Reference Figure 3 The output of the Tower layer in the teacher model can be the first teacher label in the above embodiment, the output of the calibration unit can be the second teacher label, and the output of the error correction unit can be the third teacher label. In other embodiments, the output of the error correction unit can also be the second teacher label.

[0152] It can be understood that the embodiment of the present disclosure is based on the multi-task model structure in the related technology. On this basis, a knowledge distillation scheme is introduced, and an attempt is made to start from an end-to-end training strategy, design a calibration unit and an error correction mechanism to refine the knowledge of the teacher module, so as to provide accurate soft label information to the student module, that is, the prediction module corresponding to the third task.

[0153] also, Figure 3 This is only an exemplary implementation. In other possible implementations, other tasks may be set, and the embodiments of the present disclosure are not limited to this.

[0154] Reference Figure 4 , where softmax can be an activation function to obtain the final prediction label. The loss function of the model can be calculated as follows:

[0155] L total =L buy +L offline +αL student +βL kd +L cal +L teacher Among them, L buy , L offline , L student is the first loss corresponding to the three tasks, representing the loss function of overall purchase, offline purchase and offline specific model purchase respectively. All three are set as cross entropy loss function. The teacher module loss function corresponding to the offline specific model purchase task consists of three parts, namely L teacher , L cal And the knowledge distillation loss L kd , where L teacher Responsible for updating the parameters of the teacher module, which is the cross entropy loss function; L calResponsible for updating the parameters of the calibration unit, which is also the cross entropy loss function; distillation loss L kd It is responsible for transferring the knowledge of the teacher module to the multi-task module, which ultimately affects the parameter learning of the multi-task module. kd The expression of is as follows: Among them, p i T =softmax(r i / T),q i T =softmax(z i / T), p i T It represents the value of the softmax output on the i-th category of the error correction unit in the calibration branch under the condition of T. This task is a binary classification task, so the value of i is 0 or 1; q i T represents the value of the softmax output of the offline specific model module in the multi-task model under the condition of T on the i-th category, r i represents the logit value output by the error correction unit, i.e., r in the above embodiment cal , z i Represents the logit value of the output of the offline specific model module in the multi-task model, and M represents the total number of samples. Among them, α and β represent the loss function weights of the hard label learning and soft label learning of the offline specific model purchase task, respectively. In this module, both values ​​are set to 1. It is worth noting that the above-mentioned under the condition of T can specifically refer to under the condition of the temperature coefficient T. For example, the temperature coefficient T can be a temperature hyperparameter of softmax, and the temperature coefficient T can be less than or equal to 1. When the temperature coefficient T = 0, the softmax is a traditional softmax. The temperature coefficient T can be used to adjust the smoothness of the softmax output.

[0156] In addition, in terms of model evaluation, in some embodiments, accuracy can be used as a performance evaluation indicator of the model. Taking the accuracy of offline purchasing population as an example, the specific meaning is the proportion of real offline purchasing users among the top 100w populations predicted by the model. The model evaluation data uses a full sample of 1 day for prediction, a total of 4.4kw data, and the best performing intention prediction model is predicted on the evaluation data to obtain the predicted probability value of each task; then for each task, the top 100w populations with the highest prediction probability are taken as the output population package, and the final model outputs the top 100w populations with stronger purchasing intentions, the top 100w offline purchasing populations, and the top 100w offline specific model purchasing populations; the accuracy index of these three groups of people is evaluated separately, and the best performing intention prediction model is obtained based on the comparison of accuracy indexes.

[0157] Figure 5 is a block diagram of an information recommendation device according to an exemplary embodiment. Figure 5 As shown, the information recommendation device 50 includes:

[0158] The acquisition module 51 is configured to acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity;

[0159] A construction module 52 is configured to construct a user feature vector matrix corresponding to the user information;

[0160] The prediction module 53 is configured to input the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training of a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model;

[0161] The recommendation module 54 is configured to determine whether to send recommendation information corresponding to the target product to the target user according to the intention prediction tag.

[0162] Optionally, the intention prediction model includes a cross module, a plurality of Expert modules, and a Gate weight layer corresponding to each task and a Tower layer corresponding to each task;

[0163] The prediction module 53 is configured to:

[0164] Inputting the user feature vector matrix into the crossover module to obtain a first feature vector, where the first feature vector is a feature vector that crosses high and low dimensions;

[0165] Inputting the first feature vector into the multiple Expert modules respectively to obtain multiple second feature vectors;

[0166] Input the plurality of second feature vectors into the Gate weight layer corresponding to each task respectively, so that each Gate weight layer performs weighted fusion on the plurality of second feature vectors according to the corresponding task to obtain a third feature vector corresponding to each task;

[0167] Each of the third feature vectors is input into the corresponding Tower layer to obtain the intention prediction label corresponding to each task.

[0168] Optionally, the information recommendation device 50 includes:

[0169] A first acquisition module is configured to acquire sample user information and sample label information of a sample user, wherein the sample label information includes a pre-labeled sample label corresponding to each task of the intention prediction model, and the sample user information includes at least one of the following: sample portrait information, sample attribute information of a sample commodity, and sample interaction information between the sample user and the sample commodity;

[0170] A first construction module is configured to construct a sample feature vector matrix according to the sample portrait information, the sample attribute information and the sample interaction information;

[0171] A first prediction module is configured to input the sample feature vector matrix into the teacher model to obtain a first teacher label corresponding to the sample user;

[0172] A second prediction module is configured to input the sample feature vector matrix into an initial intention prediction model to obtain a sample prediction label corresponding to the sample user;

[0173] The first determination module is configured to determine a first loss corresponding to each prediction task in the initial intention prediction model according to the sample prediction label and the sample label information; determine a second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label; determine a third loss according to the first sample label corresponding to the target prediction task in the sample label information and the first teacher label;

[0174] A first updating module is configured to update the parameters corresponding to each prediction task in the initial intention prediction model according to the first loss; update the parameters corresponding to the target prediction task in the initial intention prediction model according to the second loss; and update the teacher model according to the third loss;

[0175] The second updating module is configured to repeat the step of obtaining sample user information and sample label information of the sample user, to the step of updating the teacher model according to the third loss, until the preset conditions are met, and the trained intention prediction model is obtained.

[0176] Optionally, the teacher model further includes a calibration unit, and the information recommendation device 50 includes:

[0177] a calibration module, configured to input the first predicted probability corresponding to the first teacher label into the calibration unit to obtain a calibrated second teacher label;

[0178] A second determination module is configured to determine a fourth loss according to the second teacher label and the first sample label;

[0179] a third updating module, configured to update the calibration unit according to the fourth loss;

[0180] The first determination module is configured to:

[0181] The second loss is determined according to the first predicted label and the second teacher label.

[0182] Optionally, the third update module is configured to:

[0183] The calibration unit is updated according to the fourth loss using an adaptive moment estimation Adam optimizer.

[0184] Optionally, the teacher model further includes an error correction unit, and the information recommendation device 50 includes:

[0185] an error correction module, configured to input a second prediction probability corresponding to the second teacher label into the error correction unit;

[0186] A first output module, configured to cause the error correction unit to output the second teacher label when the confidence level of the second teacher label is greater than or equal to a preset threshold;

[0187] A second output module is configured to cause the error correction unit to output a preset third teacher label when the confidence of the second teacher label is less than the preset threshold;

[0188] Wherein, the preset threshold is determined according to the first sample label;

[0189] The first determination module is configured to:

[0190] The second loss is determined according to the first predicted label and the second teacher label or the third teacher label.

[0191] Optionally, the first update module is configured to:

[0192] Monitor the AUC index corresponding to each prediction task in each update of the initial intention prediction model;

[0193] The preset conditions include:

[0194] After the parameters corresponding to each prediction task in the initial intention prediction model are updated a preset number of times, the current AUC index corresponding to each prediction task in the initial intention prediction model is greater than the AUC index before the preset number of times.

[0195] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0196] The present disclosure also provides a computer-readable storage medium having computer program instructions stored thereon, which implement the steps of the information recommendation method provided by the present disclosure when the program instructions are executed by a processor.

[0197] In an embodiment of the present disclosure, there is also provided an information recommendation device, including:

[0198] a processor; a memory for storing instructions executable by the processor;

[0199] Wherein, the processor is configured to:

[0200] Acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity;

[0201] Constructing a user feature vector matrix corresponding to the user information;

[0202] Inputting the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training based on a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model;

[0203] Determine whether to send recommendation information corresponding to the target product to the target user according to the intention prediction label.

[0204] Figure 6 6 is a block diagram of an apparatus 600 for information recommendation according to an exemplary embodiment. For example, the apparatus 600 may be a server, a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0205] Reference Figure 6 , the device 600 may include one or more of the following components: a processing component 602 , a memory 604 , a power component 606 , a multimedia component 608 , an audio component 610 , an input / output interface 612 , a sensor component 614 , and a communication component 616 .

[0206] The processing component 602 generally controls the overall operation of the device 600, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 602 may include one or more processors 620 to execute instructions to complete all or part of the steps of the above-mentioned method. In addition, the processing component 602 may include one or more modules to facilitate the interaction between the processing component 602 and other components. For example, the processing component 602 may include a multimedia module to facilitate the interaction between the multimedia component 608 and the processing component 602.

[0207] The memory 604 is configured to store various types of data to support operations on the device 600. Examples of such data include instructions for any application or method operating on the device 600, contact data, phone book data, messages, pictures, videos, etc. The memory 604 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0208] The power supply component 606 provides power to the various components of the device 600. The power supply component 606 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 600.

[0209] The multimedia component 608 includes a screen that provides an output interface between the device 600 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 608 includes a front camera and / or a rear camera. When the device 600 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

[0210] The audio component 610 is configured to output and / or input audio signals. For example, the audio component 610 includes a microphone (MIC), and when the device 600 is in an operating mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 604 or sent via the communication component 616. In some embodiments, the audio component 610 also includes a speaker for outputting audio signals.

[0211] The input / output interface 612 provides an interface between the processing component 602 and the peripheral interface modules, which may be keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: a home button, a volume button, a start button, and a lock button.

[0212] The sensor assembly 614 includes one or more sensors for providing various aspects of status assessment for the device 600. For example, the sensor assembly 614 can detect the open / closed state of the device 600, the relative positioning of components, such as the display and keypad of the device 600, and the sensor assembly 614 can also detect the position change of the device 600 or a component of the device 600, the presence or absence of user contact with the device 600, the orientation or acceleration / deceleration of the device 600, and the temperature change of the device 600. The sensor assembly 614 may include a proximity sensor configured to detect the presence of a nearby object without any physical contact. The sensor assembly 614 may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 614 may also include an accelerometer, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0213] The communication component 616 is configured to facilitate wired or wireless communication between the device 600 and other devices. The device 600 can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 616 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 616 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0214] In an exemplary embodiment, the device 600 can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above-mentioned information recommendation method.

[0215] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 604 including instructions, and the instructions can be executed by a processor 620 of the device 600 to complete the above-mentioned information recommendation method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0216] In addition to being an independent electronic device, the above-mentioned device can also be a part of an independent electronic device. For example, in one embodiment, the device can be an integrated circuit (IC) or a chip, wherein the integrated circuit can be an IC or a collection of multiple ICs; the chip can include but is not limited to the following types: GPU (Graphics Processing Unit), CPU (Central Processing Unit), FPGA (Field Programmable Gate Array), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), SOC (System on Chip, SoC), etc. The above-mentioned integrated circuit or chip can be used to execute executable instructions (or codes) to implement the above-mentioned information recommendation method. The executable instructions can be stored in the integrated circuit or chip, or can be obtained from other devices or equipment, for example, the integrated circuit or chip includes a processor, a memory, and an interface for communicating with other devices. The executable instruction may be stored in the memory, and when the executable instruction is executed by the processor, the above-mentioned information recommendation method may be implemented; alternatively, the integrated circuit or chip may receive the executable instruction through the interface and transmit it to the processor for execution, so as to implement the above-mentioned information recommendation method.

[0217] The device 600 can operate based on an operating system stored in the memory 804, such as Windows Server 200. TM , MacOS X TM , UnixTM , Linux TM , FreeBSD TM or similar.

[0218] In another exemplary embodiment, a computer program product is further provided. The computer program product includes a computer program that can be executed by a programmable device. The computer program has a code portion for executing the above information recommendation method when executed by the programmable device.

[0219] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the present disclosure. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0220] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. An information recommendation method, characterized in that: include: Acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity; Constructing a user feature vector matrix corresponding to the user information; Inputting the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training based on a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model; Determine whether to send recommendation information corresponding to the target product to the target user according to the intention prediction label.

2. The method according to claim 1, characterized in that The intention prediction model includes a cross module, a plurality of Expert modules, and a Gate weight layer corresponding to each task and a Tower layer corresponding to each task; Inputting the user feature vector matrix into the pre-trained intention prediction model to obtain the intention prediction label of the target user includes: Inputting the user feature vector matrix into the crossover module to obtain a first feature vector, where the first feature vector is a feature vector that crosses high and low dimensions; Inputting the first feature vector into the multiple Expert modules respectively to obtain multiple second feature vectors; Input the plurality of second feature vectors into the Gate weight layer corresponding to each task respectively, so that each Gate weight layer performs weighted fusion on the plurality of second feature vectors according to the corresponding task to obtain a third feature vector corresponding to each task; Each of the third feature vectors is input into the corresponding Tower layer to obtain the intention prediction label corresponding to each task.

3. The method according to claim 1, characterized in that The training of the intention prediction model includes: Acquire sample user information and sample label information of the sample user, wherein the sample label information includes a pre-labeled sample label corresponding to each task of the intention prediction model, and the sample user information includes at least one of the following: sample portrait information, sample attribute information of the sample product, and sample interaction information between the sample user and the sample product; Constructing a sample feature vector matrix according to the sample portrait information, the sample attribute information and the sample interaction information; Inputting the sample feature vector matrix into the teacher model to obtain a first teacher label corresponding to the sample user; Inputting the sample feature vector matrix into the initial intention prediction model to obtain the sample prediction label corresponding to the sample user; Determine a first loss corresponding to each prediction task in the initial intention prediction model according to the sample prediction label and the sample label information; determine a second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label; determine a third loss according to the first sample label corresponding to the target prediction task in the sample label information and the first teacher label; updating the parameters corresponding to each prediction task in the initial intention prediction model according to the first loss; updating the parameters corresponding to the target prediction task in the initial intention prediction model according to the second loss; and updating the teacher model according to the third loss; Repeat the step of obtaining sample user information and sample label information of sample users, to the step of updating the teacher model according to the third loss, until the preset conditions are met, and the trained intention prediction model is obtained.

4. The method according to claim 3, characterized in that The teacher model further includes a calibration unit, and the method further includes: Inputting the first predicted probability corresponding to the first teacher label into the calibration unit to obtain a calibrated second teacher label; Determine a fourth loss according to the second teacher label and the first sample label; updating the calibration unit according to the fourth loss; The determining the second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label includes: The second loss is determined according to the first predicted label and the second teacher label.

5. The method according to claim 4, characterized in that The updating of the calibration unit according to the fourth loss comprises: The calibration unit is updated according to the fourth loss using an adaptive moment estimation Adam optimizer.

6. The method according to claim 4, characterized in that The teacher model also includes an error correction unit, and the method further includes: inputting a second predicted probability corresponding to the second teacher label into the error correction unit; When the confidence of the second teacher label is greater than or equal to a preset threshold, the error correction unit outputs the second teacher label; When the confidence of the second teacher label is less than the preset threshold, the error correction unit outputs a preset third teacher label; Wherein, the preset threshold is determined according to the first sample label; The determining of the second loss according to the first prediction label corresponding to the target prediction task in the sample prediction label and the first teacher label includes: The second loss is determined according to the first predicted label and the second teacher label or the third teacher label.

7. The method according to any one of claims 3 to 6, characterized in that: The updating of the parameters corresponding to each prediction task in the initial intention prediction model according to the first loss includes: Monitor the AUC index corresponding to each prediction task in each update of the initial intention prediction model; The preset conditions include: After the parameters corresponding to each prediction task in the initial intention prediction model are updated a preset number of times, the current AUC index corresponding to each prediction task in the initial intention prediction model is greater than the AUC index before the preset number of times.

8. An information recommendation device, characterized in that: include: An acquisition module is configured to acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity; A construction module, configured to construct a user feature vector matrix corresponding to the user information; A prediction module is configured to input the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training based on a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model; The recommendation module is configured to determine whether to send recommendation information corresponding to the target product to the target user based on the intention prediction tag.

9. An information recommendation device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to: Acquire user information of a target user, wherein the user information includes at least one of the following: user portrait information, commodity attribute information of a target commodity, and interaction information between the target user and the target commodity; Constructing a user feature vector matrix corresponding to the user information; Inputting the user feature vector matrix into a pre-trained intention prediction model to obtain an intention prediction label of the target user, wherein the intention prediction model is a multi-task model obtained by end-to-end training based on a teacher model, the intention prediction label includes a prediction label corresponding to each prediction task in the intention prediction model, and the teacher model corresponds to a target prediction task in the intention prediction model; Determine whether to send recommendation information corresponding to the target product to the target user according to the intention prediction label.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Citation Information

Cited By

  • Method and device for testing bearing rigidity of aero-engine, storage medium and equipment

    CN116839840A