Parameter prediction method and device, and storage medium

Abstract

The present disclosure relates to a parameter prediction method and device, and a storage medium. Historical feature data corresponding to a target object generated in a preset historical time period is obtained. A target multi-task prediction model is used to predict a demand parameter of at least one target object of a target user at a preset future time according to the historical feature data, the demand parameter representing the interest degree of the target user in the target object. In the process of pre-training the target multi-task prediction model, at least two first tasks of a plurality of tasks corresponding to the target multi-task prediction model are used to train a prediction model corresponding to another first task according to the positive sample label of one first task and the positive sample label of another first task, and the two first tasks include two tasks of the plurality of tasks in which the demand parameters influence each other.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of machine learning, and particularly relates to a parameter prediction method and device and storage medium. BACKGROUND

[0002] In the field of advertisement recommendation, the CTR (Click-Through Rate) and the CVR (Conversion Rate) of an advertisement are two key indicators for measuring the traffic of the advertisement, wherein the CTR is used to measure the attractiveness of the advertisement and the interest degree of the user to the advertisement, and the CVR is used to measure the proportion of users triggering a certain specific behavior in a specific time period, for example, the CVR can reflect the promotion effect of the advertisement on the user's purchase, download and other behaviors. Therefore, accurately predicting the CTR and the CVR of the advertisement has important guiding significance for recommending the advertisement to the target user.

[0003] In the related art, the prediction of the click rate and the conversion rate of the advertisement can be realized through a deep learning model, and the addition of the deep learning model also improves the effect of the advertisement recommendation system in predicting the click rate and other parameters of the user. SUMMARY

[0004] To overcome the problems in the related art, the present disclosure provides a parameter prediction method, device and storage medium.

[0005] According to a first aspect of an embodiment of the present disclosure, a parameter prediction method is provided, comprising:

[0006] obtaining historical feature data corresponding to a target object generated in a preset historical time period;

[0007] predicting, by a target multi-task prediction model according to the historical feature data, a demand parameter of at least one target object of a target user at a preset future time, the demand parameter representing the interest degree of the target user to the target object;

[0008] In the process of pre-training the target multi-task prediction model, at least two first tasks in a plurality of tasks corresponding to the target multi-task prediction model are used to train a prediction model corresponding to another first task according to the positive sample label of one first task and the positive sample label of another first task, and the two first tasks include two tasks in the plurality of tasks that influence each other in terms of the demand parameter.

[0009] Optionally, the target multi-task prediction model is pre-trained by the following method:

[0010] obtaining a plurality of preset sample data corresponding to the target object, each of the preset sample data comprising input sample data and label data, the label data comprising positive sample labels or negative sample labels;

[0011] inputting the plurality of input sample data into a preset multi-task prediction model to be trained to obtain model output data corresponding to each task respectively;

[0012] for each of the at least two first tasks, determining a first single-task loss function corresponding to the first task according to the positive sample label of the first task, the positive sample label of another first task, and the model output data of the first task; for each of the other tasks in the plurality of tasks except the at least two first tasks, determining a second single-task loss function corresponding to the other task according to the positive sample label corresponding to the other task and the model output data of the other task;

[0013] determining a global loss function according to the first single-task loss function corresponding to each first task respectively and the second single-task loss function corresponding to each other task respectively;

[0014] training the preset multi-task prediction model through the global loss function to obtain the target multi-task prediction model.

[0015] Optionally, the training of the preset multi-task prediction model through the global loss function to obtain the target multi-task prediction model comprises:

[0016] calculating a first loss value of the preset multi-task prediction model through the global loss function according to the positive sample label, the negative sample label, and the model output data corresponding to each task respectively;

[0017] training the preset multi-task prediction model according to the first loss value to obtain the target multi-task prediction model.

[0018] Optionally, the preset multi-task prediction model comprises a feature extraction layer and a plurality of tower models, each of the tower models is connected with the feature extraction layer, and the tower model corresponds to the task one by one; the method further comprises:

[0019] for each of the tower models, fixing model parameters of other tower models except the tower model in the preset multi-task prediction model, and calculating a second loss value of a single-task prediction model through a preset single-task loss function according to the model output data of the second task corresponding to the tower model and the label data corresponding to the second task;

[0020] After model training on the single-task prediction model according to the second loss value, a to-be-determined multi-task prediction model is obtained;

[0021] After model training on the preset multi-task prediction model through the global loss function, the target multi-task prediction model is obtained.

[0022] After model training on the to-be-determined multi-task prediction model through the global loss function, the target multi-task prediction model is obtained.

[0023] Optionally, the historical feature data includes multiple pieces of data, the target users include multiple target users, and the prediction of the demand parameter of at least one target object of a target user at a preset future time by the target multi-task prediction model according to the historical feature data includes:

[0024] After inputting the multiple pieces of historical feature data into the target multi-task prediction model, the demand parameter of at least one target object of each target user at the preset future time is output.

[0025] Optionally, the target object includes an advertisement, and the demand parameter includes a click rate and / or a conversion rate.

[0026] Optionally, each piece of historical feature data includes user feature data, target object feature data, and statistical feature data, the user feature data includes user behavior data, the target multi-task prediction model includes a feature extraction model and multiple tower models, each tower model is connected with the feature extraction model, and each tower model corresponds to one task, and the feature extraction model includes a DCN model, an attention mechanism model connected with the DCN model, and a feature combination model connected with the DCN model and the attention mechanism model, respectively.

[0027] After inputting the multiple pieces of historical feature data into the target multi-task prediction model, the demand parameter of at least one target object of each target user at the preset future time is output.

[0028] After inputting other user feature data in the user feature data except the user behavior data, the target object feature data, and the statistical feature data into the DCN model, and inputting the user behavior data into the attention mechanism model, target feature data is obtained by performing feature fusion on first feature data output by the DCN model and second feature data output by the attention mechanism model through the feature combination model.

[0029] For each task, the demand parameter corresponding to the task is output by the target tower model corresponding to the task according to the target feature data.

[0030] According to a second aspect of the embodiments of the present disclosure, a parameter prediction apparatus is provided, comprising:

[0031] an acquisition module configured to acquire historical feature data corresponding to a target object generated in a preset historical time period;

[0032] a prediction module configured to predict, by a target multi-task prediction model, a demand parameter of at least one target object of a target user at a preset future time according to the historical feature data, the demand parameter representing an interest degree of the target user to the target object;

[0033] In the process of pre-training the target multi-task prediction model, in at least two first tasks of a plurality of tasks corresponding to the target multi-task prediction model, a positive sample label of one first task and a positive sample label of another first task are used to train a prediction model corresponding to the another first task, and the two first tasks include two tasks of the plurality of tasks whose demand parameters influence each other.

[0034] According to a third aspect of the embodiments of the present disclosure, a parameter prediction apparatus is provided, comprising:

[0035] a processor;

[0036] a memory for storing processor-executable instructions;

[0037] The processor is configured to implement the steps of the method according to the first aspect of the present disclosure.

[0038] According to a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the steps of the parameter prediction method according to the first aspect of the present disclosure.

[0039] The technical solutions provided by the embodiments of the present disclosure can have the following beneficial effects: in the process of training a multi-task prediction model, when two tasks of a plurality of tasks whose demand parameters influence each other are both positive samples, a prediction model corresponding to another task is trained based on a positive sample label of one task and a positive sample label of the another task, so that the positive sample labels of the two tasks can be distilled from each other, thereby realizing that different tasks can learn label information from each other, and improving the overall prediction effect of the multi-task model.

[0040] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which are incorporated herein and constitute part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, further serve to explain the principles of the present disclosure.

[0042] Figure 1 is a flow chart of a parameter prediction method according to an example embodiment.

[0043] Figure 2 is a structural schematic diagram of a target multi-task prediction model according to an example embodiment.

[0044] Figure 3 is a flow chart of a training method of a multi-task prediction model according to an example embodiment.

[0045] Figure 4 is a flow chart of another model training method according to an example embodiment. Figure 3

[0046] Figure 5 is a block diagram of a parameter prediction device according to an example embodiment.

[0047] Figure 6 is a block diagram of another parameter prediction device according to an example embodiment. Figure 5

[0048] Figure 7 is a block diagram of a device for parameter prediction according to an example embodiment. DETAILED DESCRIPTION

[0049] The example embodiments will be described in detail herein with reference to the accompanying drawings. In the following description, unless otherwise indicated, like numbers in the different drawings represent similar or analogous elements. The following description of example embodiments is not representative of all embodiments consistent with the present disclosure. Rather, it is merely an example of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0050] It should be noted that all actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the corresponding data protection regulations and policies of the country where the device is located, and with the authorization given by the owner of the corresponding device.

[0051] ​​The present disclosure is mainly applied to the prediction of demand parameters such as click rate or conversion rate of advertisements in an advertisement recommendation scenario. Taking click rate prediction as an example, in recent years, multi-task learning (MTL) has brought significant effect improvement to the CTR prediction field, but the CTR prediction field usually needs to solve multiple tasks, such as user click prediction, user download prediction, advertisement ranking, user behavior analysis, etc. Traditional CTR prediction models usually only focus on a single task, and cannot fully utilize the multiple information in the data, resulting in unsatisfactory model effect. That is, although a single CTR prediction model can improve the effect of CTR prediction, the prediction effect of the CVR task has a bottleneck, and the industry can solve this problem through a multi-task model, but the multi-task model does not well consider the relationship between different tasks, for example, taking clicks and downloads as examples, there is a strong correlation between clicks and downloads. Therefore, how to fully utilize the associated information between different tasks to improve the prediction effect of each task is a problem to be solved.

[0052] To solve the above problems, the present disclosure provides a parameter prediction method, device and storage medium. The specific embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0053] Figure 1 is a flowchart of a parameter prediction method according to an exemplary embodiment, as shown in Figure 1 The method comprises the following steps:

[0054] In step S101, historical feature data corresponding to a target object generated in a preset historical time period is obtained.

[0055] The target object may, for example, include an advertisement and can include multiple advertisements, and the historical feature data includes data of any operation related to the target object performed by a user (which can be multiple users) in a preset historical time period, for example, data generated by the user clicking an advertisement, or downloading or purchasing a service corresponding to the advertisement in the preset historical time period.

[0056] That is, the historical feature data includes multiple data, and each piece of historical feature data includes user feature data, target object feature data and statistical feature data, and the user feature data includes user behavior data.

[0057] For example, the user feature data can include the city where the user is located, the gender of the user, the model currently used by the user, user behavior data, etc., and the user behavior data can include, for example, user click, download of certain advertisements or product behavior data. Taking the target object as an advertisement, the target object feature data is the advertisement feature data, which can include the application name corresponding to the advertisement, the language of the advertisement, etc. The statistical feature data can include, for example, the click rate of the advertisement, the click rate of different advertisements corresponding to different advertisement positions, etc.

[0058] In addition, the historical feature data includes discrete features (such as user behavior feature descriptions, the city where the user is located, the gender of the user, etc.) and numerical features (such as the click rate of a certain advertisement), and after obtaining the historical feature data, the discrete features can be subjected to embedding vectorization operation, and the numerical features can be subjected to 0-1 normalization operation, and an input feature matrix can be constructed based on the operated historical features, which can be used as model input data in the online prediction stage, or as a training set and a test set in the model training stage.

[0059] In step S102, a target multi-task prediction model is used to predict a demand parameter of at least one target object for a target user at a preset future time, the demand parameter representing the interest degree of the target user for the target object.

[0060] The demand parameter can include, for example, a click rate and / or a conversion rate. Taking the target object as an advertisement, the click rate can be understood as the probability of the target user clicking the advertisement, and the conversion rate can be understood as the probability of the target user downloading or purchasing the goods corresponding to the advertisement. Generally, the higher the click rate and / or conversion rate, the more interested the target user is in the advertisement.

[0061] In addition, the target multi-task prediction model can simultaneously predict one or more of the tasks of the click rate of the advertisement, the conversion rate of the advertisement, and the ranking of the advertisement.

[0062] In this step, after inputting a plurality of pieces of historical feature data into the target multi-task prediction model, the demand parameter of at least one target object for each target user at the preset future time can be output.

[0063] For example, historical feature data corresponding to a plurality of advertisements generated within T days can be obtained, and then the historical feature data generated within T days is used to predict the probability of a user clicking and / or downloading a certain advertisement on T+1 day, i.e., to predict the CTR of the user, by using the target multi-task prediction model.

[0064] It should be noted that in the process of pre-training the target multi-task prediction model, at least two first tasks of the plurality of tasks corresponding to the target multi-task prediction model can be trained according to the positive sample label of one first task and the positive sample label of another first task. The two first tasks include two tasks of the plurality of tasks in which the demand parameters influence each other.

[0065] For example, assuming that the target multi-task prediction model can be used to predict two tasks of click rate and conversion rate, it can be understood that when a user downloads a service corresponding to an advertisement, it usually means that the user has clicked on the advertisement. Similarly, after a user clicks on an advertisement, it is also possible to further download the service corresponding to the advertisement. Therefore, for an advertisement, the two tasks of click rate and conversion rate influence each other, and therefore the two first tasks described above can include click rate and conversion rate.

[0066] Taking at least two first tasks of click rate prediction and conversion rate prediction as an example, the positive sample corresponding to the click rate prediction includes sample data of a user having clicked on an advertisement, i.e., the positive sample label of the first task is a binary hard label (such as 0 or 1) representing that a certain advertisement has been clicked. The positive sample corresponding to the conversion rate prediction includes sample data of a user having downloaded or purchased a service corresponding to an advertisement, i.e., the positive sample label of the other first task is a binary hard label (such as 0 or 1) representing that a certain advertisement has been downloaded or purchased. In this way, in the process of training the prediction model corresponding to the click rate prediction task, the positive sample label corresponding to the click rate prediction task and the positive sample label corresponding to the conversion rate prediction task can be used to train the prediction model corresponding to the click rate prediction task. Similarly, in the process of training the prediction model corresponding to the conversion rate prediction task, the positive sample label corresponding to the conversion rate prediction task and the positive sample label corresponding to the click rate prediction task can be used to train the prediction model corresponding to the conversion rate prediction task.

[0067] For example, Figure 2 is a structural schematic diagram of a target multi-task prediction model according to an example embodiment, wherein Figure 2 The target multi-task prediction model shown in FIG. 1 can correspond to A task and B task, that is, the target multi-task prediction model can be used to predict the demand parameter corresponding to A task and can also be used to predict the demand parameter corresponding to B task. Assuming that A task is a click rate prediction task and B task is a conversion rate prediction task, the prediction model corresponding to the click rate prediction task includes Figure 2 In addition to the tower model B and the B task output shown in FIG. 1, the other model parts also include Figure 2Other model parts shown in addition to tower model A and A task output, here only in conjunction with Figure 2 The prediction model is exemplified, and the disclosure is not limited thereto.

[0068] Using the above method, in the process of training the multi-task prediction model, when the demand parameters of multiple tasks mutually affect two tasks at the same time as positive samples, the positive sample label of one task and the positive sample label of another task are used to train the prediction model corresponding to the other task, so that the positive sample labels of the two tasks can learn from each other, thereby realizing that different tasks can learn label information from each other, and improving the overall prediction effect of the multi-task model.

[0069] The training process of the target multi-task prediction model is described below.

[0070] Figure 3 is a flow chart of a multi-task prediction model training method according to an exemplary embodiment, as shown in Figure 3 As shown, the target multi-task prediction model can be pre-trained through the following steps:

[0071] In step S301, a plurality of preset sample data corresponding to the target object are obtained, each of the preset sample data including input sample data and label data, the label data including positive sample labels or negative sample labels.

[0072] Among them, for each preset sample data, the sample label of this preset sample data can include multiple labels, the number of labels being the same as the number of tasks of the multi-task model to be trained, and the labels corresponding to the tasks one by one.

[0073] For example, the target multi-task prediction model to be trained at present is a two-task prediction model that can predict the click rate and conversion rate of an advertisement at the same time, then for each preset sample data, the preset sample data includes user feature data, advertisement feature data, statistical features of the advertisement, and first label data and second label data, wherein the first label data is used to indicate whether the user in the preset sample data clicks the advertisement in the preset sample data, and the second label data is used to indicate whether the user in the preset sample data downloads the service corresponding to the advertisement in the preset sample data. Further, the first label data can include positive sample labels and negative sample labels, and the second label data can also include positive sample labels and negative sample labels.

[0074] In addition, the plurality of preset sample data may be, for example, 25 million pieces of advertisement user data generated in the past 7 days; after the plurality of preset sample data is obtained, the plurality of preset sample data may be divided into a training set and a test set, and the selection ratio may be, for example, 0.9:0.1. Then, the hyperparameters may be set as follows: the batch size is 512, and the learning rate is 1e-4. The overall GAUC (Group Area Under the Curve) value of the test set is used as a monitoring index to guide the gradient descent direction of the training, and finally the multi-task prediction model with the best performance is obtained. The selection of the training sample data and the setting of the hyperparameters are only illustrative, and the present disclosure is not limited in this regard.

[0075] In step S302, after the plurality of input sample data is input into the preset multi-task prediction model to be trained, model output data corresponding to each task is obtained.

[0076] In step S303, for each of the at least two first tasks, a first single-task loss function corresponding to the first task is determined according to the positive sample label of the first task, the positive sample label of another first task, and the model output data of the first task; for each other task in the plurality of tasks except the at least two first tasks, a second single-task loss function corresponding to the other task is determined according to the positive sample label corresponding to the other task and the model output data of the other task.

[0077] For ease of description, the two first tasks are denoted as A task and B task, respectively. In this step, a first single-task loss function corresponding to the A task may be determined according to the positive sample label of the A task, the positive sample label of the B task, and the model output data of the A task, and a first single-task loss function corresponding to the B task may be determined according to the positive sample label of the B task, the positive sample label of the A task, and the model output data of the B task.

[0078] For example, the first single-task loss function corresponding to the A task may be determined by the following formula (1):

[0079] (1)

[0080] wherein, , ;

[0081] the first single-task loss function corresponding to the A task is denoted as L A, the knowledge distillation loss function of the A task is denoted as L A KD, the original loss function of the A task is denoted as L A ori, the knowledge distillation loss function of the A task is denoted as L A KD a corresponding preset coefficient; represents a cross-entropy calculation, which can be represented as , represents model output data, represents a sample label, and i represents the i-th sample. represents a label corresponding to the A task, represents model output data corresponding to the A task, represents a sigmoid function, which can transform the output to the range [0, 1]; represents the output of the B task before the sigmoid function is input, as shown in Figure 2 , the output of the B task can guide the model training of the prediction model corresponding to the A task based on the multi-task knowledge distillation, represents the output of the A task before the sigmoid function is input. is a hyperparameter for adjusting the probability distribution, and the adjusted probability distribution is smoother and easier to guide learning.

[0082] The first single-task loss function corresponding to the B task can be determined by the following formula (2):

[0083] (2)

[0084] wherein, , ;

[0085] represents the first single-task loss function corresponding to the B task, represents the knowledge distillation loss function of the B task, represents the original loss function of the B task, represents the knowledge distillation loss function of the B task a corresponding preset coefficient, represents a label corresponding to the B task, represents model output data corresponding to the B task; represents the output of the A task before the sigmoid function is input, as shown in Figure 2 , the output of the A task can guide the model training of the prediction model corresponding to the B task based on the multi-task knowledge distillation, represents the output of the B task before the sigmoid function is input.

[0086] It should be noted that the A task and the B task are any two tasks of the plurality of tasks corresponding to the target multi-task prediction model that influence each other, for example, the A task is an advertisement click rate prediction task, and the B task is an advertisement conversion rate prediction task. In addition to the first task that influences each other, the plurality of tasks corresponding to the target multi-task prediction model can also include other tasks, but the other tasks are independent of each other, for example, the other tasks are represented as a C task, and the second single-task loss function corresponding to the C task can be obtained based on the following formula (3) according to the positive sample label corresponding to the C task and the model output data of the C task:

[0087] (3)

[0088] wherein, the second single-task loss function corresponding to the C task is represented as L C, cross-entropy calculation is represented as H, the label corresponding to the C task is represented as y C, the model output data corresponding to the C task is represented as x C. The above example is only illustrative, and the present disclosure does not limit this.

[0089] In step S304, the global loss function is determined according to the first single-task loss function corresponding to each first task and the second single-task loss function corresponding to each other task.

[0090] In a possible application scenario, the target multi-task prediction model only corresponds to a plurality of first tasks, for example, the A task and the B task in the above example, and then the global loss function can be determined according to the first single-task loss function corresponding to each first task by the following formula (4):

[0091] (4)

[0092] wherein, the global loss function is represented as L, the first single-task loss function of the A task is represented as L A, the first single-task loss function of the B task is represented as L B.

[0093] In another possible application scenario, the target multi-task prediction model corresponds to at least two first tasks and at least one other task, for example, the A task and the B task are two first tasks, and the C task is another task other than the first task in the above example, and then the global loss function can be determined by the following formula (5):

[0094] (5)

[0095] wherein, the global loss function is represented as L. a first single-task loss function representing the A task, a first single-task loss function representing the B task, a second single-task loss function corresponding to the C task.

[0096] In step S305, after the preset multi-task prediction model is trained by the global loss function, the target multi-task prediction model is obtained.

[0097] In this step, the first loss value of the preset multi-task prediction model can be calculated by the global loss function according to the positive sample label, the negative sample label and the model output data corresponding to each task; and after the preset multi-task prediction model is trained according to the first loss value, the target multi-task prediction model is obtained.

[0098] Taking the A task and the B task corresponding to the target multi-task prediction model as an example, the global loss function corresponding to the A task and the B task can be determined as formula (4) by formula (1), (2) and (4); the first loss value can be obtained by substituting the positive sample label, the negative sample label and the model output data of each task into formula (4); and then the target multi-task prediction model can be obtained by training the preset multi-task prediction model based on the first loss value using the back propagation algorithm.

[0099] It should be noted that after the model training is completed, the trained model can be tested using a test set, the overall GAUC value of the test set is used as a monitoring index to guide the gradient descent direction of the training, so as to obtain the target multi-task prediction model with the best performance.

[0100] By using the above model training method, when the A task and the B task are both positive samples (for example, clicks and conversions), the soft label of the A task is used as the distillation soft label to guide the model learning of the B task, and vice versa, the soft label of the B task is used as the distillation soft label to guide the model learning of the A task, so that the information sharing between different tasks is achieved, and the multi-task prediction effect of the model is improved, that is, the prediction effect of the A task and the B task can be improved at the same time.

[0101] In another possible implementation of the present disclosure, in order to improve the efficiency and accuracy of model training, the prediction models corresponding to different tasks can be pre-trained respectively, and then the multi-task prediction models after pre-training are further trained.

[0102] In a possible implementation, the preset multi-task prediction model includes a feature extraction layer and a plurality of tower models, each of the tower models is connected with the feature extraction layer, and the tower model corresponds to one task, as shown in Figure 2 ​Figure 2 The multi-task prediction model shown can make predictions on the A task and the B task, Figure 2 The multi-task prediction model shown includes tower model A and tower model B, where model A corresponds to the A task, and model B corresponds to the B task. In addition, Figure 2 The DCN model, the attention mechanism model, and the feature combination model in the feature extraction layer are used to extract features from the input data.

[0103] Figure 4 is according to Figure 3 Another model training method flowchart is shown in the embodiment shown, as Figure 4 The training method further includes:

[0104] In step S306, for each tower model, the model parameters of other tower models in the preset multi-task prediction model except the tower model are fixed, and a second loss value of a single-task prediction model is calculated by a preset single-task loss function according to the model output data of the second task corresponding to the tower model and the label data corresponding to the second task.

[0105] In step S307, after the single-task prediction model is trained according to the second loss value, a pending multi-task prediction model is obtained.

[0106] For example, continuing with Figure 2 , the prediction model corresponding to the A task and the prediction model corresponding to the B task can be preheated and trained respectively. First, in the process of training the prediction model corresponding to the A task, the model parameters of tower model B can be fixed, and then the A task output data corresponding to the A task (i.e., the model output data of the A task corresponding to tower model A) and the label data of the A task are used to calculate a second loss value of the A task prediction model (i.e., the single-task prediction model corresponding to the A task, such as Figure 2 except tower model B) by the original loss function corresponding to the A task (i.e., the preset single-task loss function corresponding to the A task, and the specific symbol explanation can be referred to the description above) and then the A task prediction model is trained according to the second loss value. After that, the prediction model corresponding to the B task can be preheated and trained, the model parameters of tower model A can be fixed, and then the B task output data corresponding to the B task (i.e., the model output data of the B task corresponding to tower model B) and the label data of the B task are used to calculate a second loss value of the B task prediction model (i.e., the single-task prediction model corresponding to the B task, such as Figure 2 except tower model A) by the original loss function corresponding to the B task (i.e., the preset single-task loss function corresponding to the B task, and the specific symbol explanation can be referred to the description above).The second loss value of the model part other than the tower model A is obtained, and then the model training is performed on the B task prediction model according to the second loss value. After that, the A task prediction model and the B task prediction model after the preheating training can be combined to obtain the to-be-determined multi-task prediction model, and the above example is only illustrative, and the present disclosure is not limited in this regard.

[0107] In this way, after the model training of the to-be-determined multi-task prediction model through the global loss function in the process of step S305, the target multi-task prediction model can be obtained. The specific training process of the to-be-determined multi-task prediction model to obtain the target multi-task prediction model can be referred to in the present disclosure Figure 3 The corresponding embodiments are not described here again.

[0108] It should be further pointed out that the traditional knowledge distillation method needs to train a teacher model first, and then use the soft label of the teacher model to guide the training of a student model. The training time of the model is relatively long, and the training efficiency is relatively low. However, in the training process of the multi-task prediction model of the present disclosure, task A and task B are mutually distilled and learned, that is, the prediction model corresponding to each task can be trained at the same time. Compared with the traditional knowledge distillation method, the training efficiency can be improved. In addition, compared with the traditional training of the multi-task prediction model, the training index GAUC of the multi-task prediction model provided in the present disclosure is significantly improved.

[0109] The specific implementation process of parameter prediction based on the trained target multi-task prediction model is described below.

[0110] The target multi-task prediction model in the present disclosure includes a feature extraction model and a plurality of tower models, as shown in Figure 2 The feature extraction model can include a DCN model (such as DCN-V2), an attention mechanism model connected with the DCN model, and a feature combination model connected with the DCN model and the attention mechanism model, respectively. Each tower model is connected with the feature extraction model, and each tower model corresponds to the task one by one. In this way, the other user feature data in the user feature data except the user behavior data, the target object feature data (such as the advertisement feature in Figure 2 The first feature data output by the DCN model and the second feature data output by the attention mechanism model are fused through the feature combination model to obtain target feature data after the user behavior data is input into the attention mechanism model. For each task, the target tower model corresponding to the task is used to output the demand parameter corresponding to the task according to the target feature data. For example Figure 2As shown, for the A task, the feature output by the feature combination model is input to the tower model A, and the predicted A task output result is obtained through the tower model A; for the B task, the feature output by the feature combination model is input to the tower model B, and the predicted B task output result is obtained through the tower model B.

[0111] Figure 5 is a block diagram of a parameter prediction device 500 according to an exemplary embodiment. Referring to Figure 5 The device comprises:

[0112] An acquisition module 501 is configured to acquire historical feature data corresponding to a target object generated in a preset historical time period;

[0113] A prediction module 502 is configured to predict, by a target multi-task prediction model, a demand parameter of at least one target object of a target user at a preset future time according to the historical feature data, the demand parameter representing an interest degree of the target user to the target object.

[0114] In the process of pre-training the target multi-task prediction model, in at least two first tasks of a plurality of tasks corresponding to the target multi-task prediction model, a prediction model corresponding to another first task is trained according to a positive sample label of one first task and a positive sample label of another first task, and the two first tasks include two tasks of the plurality of tasks whose demand parameters influence each other.

[0115] Optionally, Figure 6 is a block diagram of another parameter prediction device according to Figure 5 As shown in the embodiment, the device further comprises a model training module 503. Figure 6

[0116] The model training module 503 is configured to pre-train a target multi-task prediction model by the following manner:

[0117] A plurality of preset sample data corresponding to the target object are acquired, each of the preset sample data comprising input sample data and label data, the label data comprising a positive sample label or a negative sample label;

[0118] After inputting a plurality of input sample data into a preset multi-task prediction model to be trained, model output data corresponding to each task is acquired;

[0119] ​For each of the at least two first tasks, a first single-task loss function corresponding to the first task is determined according to the positive sample label of the first task, the positive sample label of another first task, and the model output data of the first task; for each other task in the plurality of tasks except the at least two first tasks, a second single-task loss function corresponding to the other task is determined according to the positive sample label corresponding to the other task and the model output data of the other task;

[0120] A global loss function is determined according to the first single-task loss function corresponding to each first task and the second single-task loss function corresponding to each other task;

[0121] After the preset multi-task prediction model is trained by the global loss function, the target multi-task prediction model is obtained.

[0122] Optionally, the model training module 503 is configured to calculate a first loss value of the preset multi-task prediction model by the global loss function according to the positive sample label, the negative sample label and the model output data corresponding to each task; and obtain the target multi-task prediction model after the preset multi-task prediction model is trained according to the first loss value.

[0123] Optionally, the preset multi-task prediction model comprises a feature extraction layer and a plurality of tower models, each of the tower models is connected with the feature extraction layer, and the tower model corresponds to one of the tasks;

[0124] The model training module 503 is configured to, for each of the tower models, fix the model parameters of other tower models except the tower model in the preset multi-task prediction model, calculate a second loss value of a single-task prediction model by a preset single-task loss function according to the model output data of the second task corresponding to the tower model and the label data corresponding to the second task, obtain a to-be-determined multi-task prediction model after the single-task prediction model is trained according to the second loss value, and obtain the target multi-task prediction model after the to-be-determined multi-task prediction model is trained by the global loss function.

[0125] Optionally, the historical feature data comprises a plurality of data, the target user comprises a plurality of target users, and the prediction module 502 is configured to input the plurality of historical feature data into the target multi-task prediction model, and output the demand parameter of at least one target object of each target user at the preset future time.

[0126] Optionally, the target object comprises an advertisement, and the demand parameter comprises a click rate and / or a conversion rate.

[0127] Optionally, each piece of the historical feature data comprises user feature data, target object feature data and statistical feature data, the user feature data comprises user behavior data, the target multi-task prediction model comprises a feature extraction model and a plurality of tower models, each of the tower models is connected with the feature extraction model, and each of the tower models corresponds to the task one by one; the feature extraction model comprises a DCN model, an attention mechanism model connected with the DCN model, and a feature combination model connected with the DCN model and the attention mechanism model respectively.

[0128] The prediction module 502 is configured to input other user feature data in the user feature data except the user behavior data, the target object feature data and the statistical feature data into the DCN model, input the user behavior data into the attention mechanism model, obtain target feature data by performing feature fusion on first feature data output by the DCN model and second feature data output by the attention mechanism model through the feature combination model, and output the demand parameter corresponding to each task through the target tower model corresponding to the task according to the target feature data.

[0129] As to the apparatus in the above-described embodiments, the specific manners in which various modules perform operations have been described in detail in the embodiments of the method, and thus will not be described in detail here.

[0130] The present disclosure also provides a computer readable storage medium having stored thereon computer program instructions, which, when executed by a processor, implement the steps of the parameter prediction method provided by the present disclosure.

[0131] Figure 7 is a block diagram of an apparatus 700 for parameter prediction according to an exemplary embodiment. For example, the apparatus 700 can be 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, and the like.

[0132] Referring to Figure 7 , the apparatus 700 can include one or more of the following components: a processing component 702, a memory 704, a power supply component 706, a multimedia component 708, an audio component 710, an input / output (I / O) interface 712, a sensor component 714, and a communication component 716.

[0133] The processing component 702 generally controls the overall operations of the device 700, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 702 can include one or more processors 720 to execute instructions delivered from the memory 704 to complete all or part of the parameter prediction method described above. In addition, the processing component 702 can include one or more modules to facilitate interaction between the processing component 702 and other components. For example, the processing component 702 can include a multimedia module to facilitate the interaction between the multimedia component 708 and the processing component 702.

[0134] The memory 704 is configured to store various types of data to support operations of the device 700. Examples of these data include instructions for any application or method operating on the device 700, contact data, phonebook data, messages, pictures, videos, and the like. The memory 704 can be implemented by any type of volatile or non-volatile storage devices 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 storage, flash memory, magnetic disk or optical disk.

[0135] The power component 706 provides power to the various components of the device 700. The power component 706 can include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 700.

[0136] The multimedia component 708 includes a screen providing an output interface between the device 700 and a user. In some embodiments, the screen can include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes the touch panel, the screen can be implemented as a touch screen to receive an input signal from a user. The touch panel includes one or more touch sensors to sense a touch, a slide, and a gesture on the touch panel. The touch sensors can not only sense a boundary of a touching or sliding action, but also detect duration and pressure related to the touching or sliding action. In some embodiments, the multimedia component 708 includes a front camera and / or a back camera. The front camera and / or the back camera can receive external multimedia data when the device 700 is in an operating mode, such as a shooting mode or a video mode. Each of the front and back cameras can be a fixed optical lens system or have a focal length and optical zoom capability.

[0137] The audio component 710 is configured to output and / or input audio signals. For example, the audio component 710 includes a microphone (MIC) that is configured to receive an external audio signal when the device 700 is in an operation mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signal can be further stored in the memory 704 or transmitted via the communication component 716. In some embodiments, the audio component 710 also includes a speaker for outputting audio signals.

[0138] The input / output interface 712 provides an interface between the processing component 702 and peripheral interface modules, which can be a keypad, a click wheel, buttons, and the like. The buttons can include, but are not limited to, a home button, a volume button, a start button, and a lock button.

[0139] The sensor component 714 includes one or more sensors for providing status assessments of various aspects of the device 700. For example, the sensor component 714 can detect an open / closed position of the device 700, relative positioning of components, such as a display and a keypad of the device 700, a change of position of the device 700 or a component of the device 700, presence or absence of user contact with the device 700, changes in orientation or acceleration / deceleration

[0140] The communication component 716 is configured to facilitate wired or wireless communication between the device 700 and other devices. The device 700 can access a wireless network based on a corresponding communication standard, such as WiFi, 2G, or 3G, or a combination thereof. In an example embodiment, the communication component 716 receives broadcast signals or broadcast-related information from external broadcast management systems via a broadcast channel. In an example embodiment, the communication component 716 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.

[0141] In an exemplary embodiment, the apparatus 700 can be implemented using 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, micro-controllers, microprocessors or other electronic devices, to perform the parameter prediction method described above.

[0142] In an exemplary embodiment, a non-transitory computer readable storage medium including instructions, such as the memory 704 including instructions, is also provided, which can be executed by the processor 720 of the apparatus 700 to perform the parameter prediction method described above. 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 disc, and an optical data storage device, etc.

[0143] In another exemplary embodiment, a computer program product is also provided, which contains a computer program executable by a programmable apparatus, the computer program having code sections for performing the parameter prediction method described above when executed by the programmable apparatus.

[0144] Other embodiments of the present disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the present disclosure. The present disclosure is intended to cover any variations, uses or adaptive changes of the present disclosure following the general principles thereof and including the general and specific knowledge in the art not disclosed in the specification. The specification and examples are to be regarded as illustrative only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0145] It should be understood that the present disclosure is not limited to the precise structures described and shown in the drawings, and that various modifications and changes can be made without departing from the scope of the present disclosure. The scope of the present disclosure is limited only by the claims that follow.

Claims

1. A method of parameter prediction, characterized by, The method comprises: obtaining historical feature data corresponding to a target object generated in a preset historical time period; the target object comprises an advertisement, and the historical feature data comprises operation data of a user related to the advertisement in the preset historical time period; predicting, by a target multi-task prediction model according to the historical feature data, a demand parameter of at least one target object of a target user at a preset future time, the demand parameter representing the interest degree of the target user in the target object; the demand parameter comprises a click rate and / or a conversion rate; in the process of pre-training the target multi-task prediction model, at least two first tasks in a plurality of tasks corresponding to the target multi-task prediction model, the prediction model corresponding to another first task is trained according to the positive sample label of one first task and the positive sample label of another first task, and the two first tasks include two tasks in which the demand parameters influence each other in the plurality of tasks.

2. The method of claim 1, wherein, The target multi-task prediction model is pre-trained in the following way: obtaining a plurality of preset sample data corresponding to the target object, each of the preset sample data comprising input sample data and label data, the label data comprising a positive sample label or a negative sample label; after inputting a plurality of input sample data into a preset multi-task prediction model to be trained, obtaining model output data corresponding to each task respectively; for each first task in the at least two first tasks, determining a first single-task loss function corresponding to the first task according to the positive sample label of the first task, the positive sample label of another first task, and the model output data of the first task; for each other task in the plurality of tasks except the at least two first tasks, determining a second single-task loss function corresponding to the other task according to the positive sample label corresponding to the other task and the model output data of the other task; determining a global loss function according to the first single-task loss function corresponding to each first task and the second single-task loss function corresponding to each other task respectively; after model training of the preset multi-task prediction model by the global loss function, obtaining the target multi-task prediction model.

3. The method of claim 2, wherein, After model training of the preset multi-task prediction model by the global loss function, the target multi-task prediction model is obtained, which comprises: calculating a first loss value of the preset multi-task prediction model by the global loss function according to the positive sample label, the negative sample label and the model output data corresponding to each task respectively; after model training of the preset multi-task prediction model according to the first loss value, obtaining the target multi-task prediction model.

4. The method of claim 2, wherein, The preset multi-task prediction model comprises a feature extraction layer and a plurality of tower models, each of the tower models is connected with the feature extraction layer, and the tower model corresponds to one task; the method further comprises: For each tower model, fixing model parameters of other tower models in the preset multi-task prediction model except the tower model, calculating a second loss value of a single-task prediction model through a preset single-task loss function according to model output data of a second task corresponding to the tower model and label data corresponding to the second task; After model training of the single-task prediction model according to the second loss value, a to-be-determined multi-task prediction model is obtained; After model training of the preset multi-task prediction model through the global loss function, the target multi-task prediction model is obtained. After model training of the to-be-determined multi-task prediction model through the global loss function, the target multi-task prediction model is obtained.

5. The method according to any one of claims 1 to 4, characterized in that, The historical feature data includes multiple pieces of data, the target user includes multiple target users, and the demand parameter of the target user for at least one target object at a preset future time is predicted by the target multi-task prediction model according to the historical feature data, including: After inputting the multiple pieces of historical feature data into the target multi-task prediction model, the demand parameter of each target user for at least one target object at the preset future time is output.

6. The method of claim 5, wherein, Each piece of historical feature data includes user feature data, target object feature data and statistical feature data, the user feature data includes user behavior data, the target multi-task prediction model includes a feature extraction model and multiple tower models, each tower model is connected with the feature extraction model, and each tower model corresponds to one task; the feature extraction model includes a DCN model, an attention mechanism model connected with the DCN model, and a feature combination model connected with the DCN model and the attention mechanism model respectively; After inputting the multiple pieces of historical feature data into the target multi-task prediction model, the demand parameter of each target user for at least one target object at the preset future time is output. After inputting other user feature data in the user feature data except the user behavior data, the target object feature data and the statistical feature data into the DCN model, and inputting the user behavior data into the attention mechanism model, target feature data is obtained by feature fusion of first feature data output by the DCN model and second feature data output by the attention mechanism model through the feature combination model; For each task, the target feature data is input into the target tower model corresponding to the task to output the demand parameter corresponding to the task.

7. A parameter prediction device characterized by comprising: including: The acquisition module is configured to acquire historical feature data corresponding to a target object generated in a preset historical time period; The target object includes an advertisement, and the historical feature data includes operation data related to the advertisement performed by a user in the preset historical time period; A prediction module is configured to predict, by a target multi-task prediction model, a demand parameter of at least one target object by a target user at a preset future time according to the historical feature data, the demand parameter representing an interest degree of the target user to the target object; the demand parameter includes a click rate and / or a conversion rate. In the process of pre-training the target multi-task prediction model, in at least two first tasks of a plurality of tasks corresponding to the target multi-task prediction model, a prediction model corresponding to another first task is trained according to a positive sample label of one first task and a positive sample label of another first task, and the two first tasks include two tasks of the plurality of tasks whose demand parameters influence each other.

8. A parameter prediction apparatus characterized by comprising: Comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to perform the steps of the parameter prediction method of any one of claims 1-6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The program instructions are executed by the processor to implement the steps of the method of any one of claims 1-6.

Images