Target user determination method and computing device

By building a target user determination model, combining the feature processing layer and the expert network layer, the user's historical behavior data is screened and trained, and the accuracy problem of target user determination in multiple business scenarios is solved, and the precise marketing effect is achieved in multiple scenarios .

CN118967181APending Publication Date: 2024-11-15CHEZHI HULIAN BEIJING SCI & TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410939920.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to accurately determine the target users in multiple business scenarios, resulting in poor marketing strategies, and the method based on user similarity is insufficiently generalized in vertical fields and cannot meet customized marketing needs.

Method used

By building a target user determination model, using the feature processing layer and expert network layer to filter and train user historical behavior data, generate a model that can determine the target user under multiple business categories, and combine common attributes and vertical attributes for multi-objective construction Model, automatically find user weight expression.

Benefits of technology

It realizes accurate and efficient target user determination in multiple business scenarios, improves marketing effects, enhances the focus of user vector expression and generalization capabilities of model, and meets the precise marketing needs in multiple scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967181A_ABST
    Figure CN118967181A_ABST
Patent Text Reader

Abstract

The invention discloses a target user determination method and computing equipment, and the method comprises the steps: screening historical behavior data sets of users according to a plurality of business classifications, and determining a user set comprising a plurality of users; generating user training data according to the historical behavior data of each user in the user set, and obtaining a user training data set about service classification; constructing a target user determination model, wherein the target user determination model comprises a feature processing layer and an expert network layer connected with the feature processing layer; training a target user determination model according to the user training data set; and in response to a received target user determination request, classifying the to-be-classified user set according to the target user determination model, and determining one or more target users under the preset service classification from the to-be-classified user set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of recommendation algorithms, and in particular to a target user determination method and computing device. Background Art

[0002] In various demand scenarios such as business marketing, determining the target population for business marketing is a relatively important issue. By defining the target population for business marketing and implementing various marketing strategies to the target population, the effectiveness of various business marketing activities can be effectively improved.

[0003] In the prior art, labels or classification information are added to different types of people, and users similar to a certain user are determined as target groups based on the same labels or the same classification. This method of mining target users based on user similarity diffusion has a good effect in a single business marketing scenario.

[0004] However, users' interests and usage scenarios on the platform are diverse. User similarity is set according to the diverse user behaviors in various usage scenarios of the platform, so as to establish a target user identification model. After user mining based on such a model, the users finally determined may focus on a single dimension or average generalization of multiple dimensions. This makes it difficult to meet the needs of determining target users for each business scenario and conducting customized marketing activities when the platform includes multiple business scenarios.

[0005] Therefore, a new method of determining target users is needed. Summary of the invention

[0006] To this end, the present invention provides a target user determination method in an effort to solve or at least alleviate the above problems.

[0007] According to a first aspect of the present invention, there is provided a target user determination method, which runs in a computing device, and comprises: screening a user's historical behavior data set according to a plurality of business classifications to determine a user set including a plurality of users; generating user training data according to the historical behavior data of each user in the user set to obtain a user training data set for the business classification; constructing a target user determination model, the target user determination model comprising a feature processing layer and an expert network layer connected to the feature processing layer; training the target user determination model according to the user training data set; and in response to receiving a target user determination request, classifying the user set to be classified according to the target user determination model, and determining one or more target users under each preset business classification from the user set to be classified.

[0008] Optionally, in the method according to the present invention, the feature processing layer includes a feature weight processing module and a feature combination module, a deformable convolution module and a feature weight interaction module connected to the feature weight processing module.

[0009] Optionally, in the method according to the present invention, the expert network layer includes a primary network module and an advanced network module, the primary network module includes a primary shared expert module and one or more primary task expert modules, the advanced network module includes an advanced shared expert module and one or more advanced task expert modules, each primary task expert module corresponds to a advanced task expert module, and the output results are input into the corresponding advanced task expert module.

[0010] Optionally, in the method according to the present invention, training the target user determination model according to the user training data set includes: generating a first training sample according to each user training data in the user data training set to obtain a first training sample set, the first training sample including a first number of first-order feature vectors, each feature vector being used to characterize historical behavior data under a business category.

[0011] Optionally, the method according to the present invention also includes: setting the same first discard probability value for the first-order feature vector in each first training sample; inputting the first training sample set into the target user determination model for training to generate a second discard probability value for each first-order feature vector; generating a second training sample based on a second number of first-order feature vectors whose second discard probability values ​​are less than a preset threshold to obtain a second training sample set; and inputting the second training sample set into the target user determination model for training.

[0012] Optionally, in the method according to the present invention, the training sample set is input into the target user determination model for training, including: inputting the training samples in the training sample set into the feature weight processing module of the feature processing layer, and setting a first weight value for each first-order feature vector according to the feature weight processing module; processing the first-order feature vector according to the feature combination module to generate a first high-order feature vector; processing the first-order feature vector according to the deformable convolution module to generate a second high-order feature vector; processing the first-order feature vector according to the feature weight interaction module to generate a third high-order feature vector; inputting the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector into the expert network layer for training, and generating a second weight value for each first-order feature vector; inputting the training samples in the training sample set into the feature weight processing module of the feature processing layer, setting a second weight value for each first-order feature vector according to the feature weight processing module, and training the target user determination model.

[0013] Optionally, in the method according to the present invention, the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector are input into the expert network layer for training, including: inputting the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector into the primary shared expert module and one or more primary task expert modules; using the outputs of the primary shared expert module and one or more primary task expert modules as the input of the high-level shared expert module; using the outputs of the primary shared expert module and each primary task expert module as the input of the high-level task expert module corresponding to the primary task expert module; determining one or more target task expert modules from one or more high-level task expert modules, and determining one or more training target users under a preset business category from a user set based on the outputs of one or more target task expert modules and the high-level shared expert module; and setting parameters for the expert network layer based on one or more training target users and one or more calibrated target users calibrated in the user set under a predetermined business category.

[0014] Optionally, in the method according to the present invention, the set of users to be classified is classified according to the target user determination model, and one or more target users under each preset business category are determined from the set of users to be classified, including: generating samples to be classified according to user data of the users to be classified in the set of users to be classified, the samples to be classified including a second number of first-order feature vectors; and inputting each sample to be classified into the target user determination model to determine whether each user to be classified is a target user under the preset business category.

[0015] According to a second aspect of the present invention, a computing device is provided, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing the target user determination method according to the present invention.

[0016] According to a third aspect of the present invention, there is provided a computer-readable storage medium storing one or more programs, the one or more programs comprising instructions which, when executed by a computing device, cause the computing device to execute the target user determination method according to the present invention.

[0017] According to a fourth aspect of the present invention, there is provided a computer program product, comprising a computer program / instruction, wherein the computer program / instruction, when executed by a processor, implements the target user determination method of the present invention.

[0018] In the target user determination method of the present invention, a target user determination model is constructed, and when training the target user determination model, the historical behavior data set is first screened according to the business classification to obtain a user data training set, and the user training data set about the business classification is used for training; the training obtains a model capable of target classification, that is, capable of determining target users corresponding to multiple business classifications. The model is used to accurately and efficiently determine the target users under each business classification, so as to guide accurate business marketing according to the target users under each business classification and improve the business marketing effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to achieve the above and related purposes, certain illustrative aspects are described herein in conjunction with the following description and accompanying drawings, which indicate various ways in which the principles disclosed herein can be practiced, and all aspects and their equivalents are intended to fall within the scope of the claimed subject matter. By reading the following detailed description in conjunction with the accompanying drawings, the above and other purposes, features and advantages disclosed by the present invention will become more apparent. Throughout the present disclosure, the same reference numerals generally refer to the same parts or elements.

[0020] Figure 1 A flow chart of a target user determination method 100 according to an exemplary embodiment of the present invention is shown;

[0021] Figure 2a A schematic diagram showing a target user determination model 200 according to an exemplary embodiment of the present invention;

[0022] Figure 2b shows a schematic diagram of a feature processing layer 220 according to an exemplary embodiment of the present invention;

[0023] Figure 2c FIG. 2 is a schematic diagram showing an expert network layer 230 according to an exemplary embodiment of the present invention.

[0024] Figure 3 FIG. 3 is a block diagram of a computing device 300 according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0025] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to enable the scope of the present disclosure to be fully conveyed to those skilled in the art. The same reference numerals generally refer to the same components or elements.

[0026] In various demand scenarios such as business marketing, it is a relatively important issue to determine the target population of business marketing. Each business scenario has different business populations, which can be called vertical scenarios, i.e. vertical categories, segmented tracks, professional fields, etc. Vertical categories mean limited fields with clear boundaries. For example, under the large category of skin care, sensitive skin care is a segmented vertical category with fewer people but clearer needs.

[0027] In the automotive field, different business scenarios or business classifications can also be divided. For example, users on the automotive platform have various business scenarios such as viewing cars, selecting cars, buying cars, and using cars. At the same time, they can also be subdivided. For example, in the car-using business scenario, there may also be different vertical demand preferences such as travel, modification, and refueling. In these diverse vertical marketing scenarios, if the model is trained using a general user portrait and the user's unified behavioral characteristics across the entire platform, the user expressions of the candidate population in the training samples used by the model in different vertical scenarios will be the same, and the behavioral attributes and needs of the vertical field will not be sufficiently characterized, resulting in the target users mined through classification and similarity expansion being too generalized and general. The cost of manually mining users with corresponding labels based on the label-based crowd diffusion method is high, and the similar users mined through labels are not generalized enough. In actual business, they do not meet the needs of the target user level, and the scalability is poor. The crowd diffusion method based on classification models and similarity has good generalization, but the general mining method does not focus enough on specific vertical fields.

[0028] Taking the automotive field as an example, users will accumulate automobile interaction behaviors and portrait attributes of different dimensions such as "viewing, selecting, buying and using" during the use of the automobile platform. Users of the automobile platform not only differ in static portrait attributes such as gender, age, and region, but also have different life cycle stages such as car selection, purchase and use. In addition, users' usage purposes and user behaviors are diverse. For example, for users who want to buy a car, different users will view car information in different ways. Some users tend to search and compare car series for parameter screening, some focus on understanding car series details by browsing videos or graphic explanations, and some users prefer to find preferential information and sales channels. In addition to these main scenarios for car selection and purchase, there are also different vertical needs for car use such as travel, modification, and car maintenance. To tap into these vertical needs, target users should not only use general portrait attributes and behavioral characteristics of the main scenarios, but also focus on the different behavioral performances of users in vertical scenarios and vertical resources. Therefore, when launching marketing activities in these different vertical business areas, it is necessary to train models separately based on different business types.

[0029] When training models for these vertical scenarios, the present invention uses a multi-objective deep network structure, integrates user common attributes and vertical attributes, and generates feature vector expressions of users under multiple vertical field targets. Compared with traditional user vector expression methods, it can flexibly perform training and information transfer based on multiple different targets, enrich the integrity of user vector expression, improve the emphasis on vertical fields, and alleviate the situation where insufficient single-target data leads to insufficient training.

[0030] The target user determination method proposed in the present invention extracts the core vertical scenarios from actual business scenarios for multi-objective modeling, and uses one model to determine the target population in multiple scenarios. In terms of features, it not only uses general user portrait attribute features, user interaction behaviors in different columns of the platform, different car models and series, and different resource types, but also integrates the domain characteristics of specific vertical categories to customize and optimize the model; constructs multi-objective models of different vertical categories to meet the target population mining needs of multiple scenarios; when training the model, it also automatically optimizes the weight expression of user general attributes and vertical category attributes in different vertical category scenarios to improve the model training and usage effects.

[0031] Figure 1 FIG. 1 is a flow chart of a target user determination method 100 according to an exemplary embodiment of the present invention. Figure 1 As shown, firstly execute step 110, filter the historical behavior data set of the user according to one or more business categories, and generate a user training data set about the business categories; including: filter the historical behavior data set of the user according to multiple business categories, and determine a user set including multiple users; generate user training data according to the historical behavior data of each user in the user set, and obtain the user training data set about the business categories.

[0032] According to an embodiment of the present invention, historical behavior data can be obtained from a database, and historical behavior data of each user can be extracted from the database. The historical behavior data of each user can be recorded, counted, and stored in the database in advance, and can be stored in the form of user logs. The present invention does not limit the specific collection and acquisition methods of historical behavior data.

[0033] Each user's historical behavior data specifically includes the data generated by each user's historical behavior under each business category. Each business category includes not only the main scenarios that every user will use, such as the homepage of the car category, or a global activity that all users participate in, but also sub-scenarios, that is, vertical business scenarios that are of interest or disinterest, such as car travel, car modification, etc.

[0034] The historical behavior data of each user includes the same part and the different part. The same part is that the historical behavior data of each user includes the behavior data in the main scenario; the different part is that the historical behavior data of each user has different sub-scenario behavior data. Because different users are interested in different sub-scenario data, the sub-scenario data in the historical behavior data collected and stored by different users are also different.

[0035] The user's historical behavior includes data on the user's interaction with information in corresponding scenarios under different business categories, such as clicks, browsing, collections, and comments, etc. The present invention does not limit the specific behavior type and behavior mode of the user's historical behavior.

[0036] When training the model, not only data from different sub-scenarios but also data from the same main scenario are used for training. This can enhance the generalization ability of the model and use more comprehensive features to improve the robustness of the model.

[0037] According to an embodiment of the present invention, when a user's historical behavior data set is screened according to multiple business classifications to determine a user set of one or more users, the business classification used refers to the business classification for which a target user needs to be determined. If a target user under a certain business classification needs to be determined, the user's historical behavior data is screened according to the business classification so as to train a model that can identify the target user under the business classification.

[0038] Screening the user's historical behavior data according to each business category includes: judging whether the user has more historical behaviors under the business category according to the user's historical behavior data, and specifically setting the judging criteria, such as browsing the information of the business category more than the preset number of times per month, collecting the information under the business category more than the preset number of times, etc., and when it is judged that the user has more historical behaviors under the business category according to the judging criteria, the user is added to the user set as a positive sample; after selecting the positive sample, randomly select some users from other users as negative samples and add them to the user set. The obtained user set includes positive samples and negative samples, and both the positive samples and the negative samples include one or more users.

[0039] After obtaining the user set, the historical behavior data of each user in the user set is extracted to generate user training data. According to one embodiment of the present invention, the historical behavior data of the user is extracted, including the behavior data in the main scenario and the behavior data in the sub-scenario. When generating the user training data based on the historical behavior data, the user's portrait attribute data is also obtained and added to the user training data. The user's portrait attribute data is attribute data that each user has, and is static data. The user's historical behavior data is dynamic data. The portrait attribute data includes the user's gender, age, region, etc. The present invention does not limit the type of data specifically included in the user's portrait attribute data.

[0040] According to one embodiment of the present invention, by extracting static data and dynamic data, common data and characteristic data of users, and setting negative samples and positive samples based on the above data, using such samples to train the model can improve the generalization ability of the model and improve the effect of the model in determining the target users.

[0041] Then, step 120 is executed to construct a target user determination model, where the target user determination model includes a feature processing layer and an expert network layer connected to the feature processing layer.

[0042] Figure 2a FIG. 2 is a schematic diagram of a target user determination model 200 according to an exemplary embodiment of the present invention. Figure 2a As shown, the target user determination model 200 includes a model input layer 210, which is responsible for receiving input feature vectors; it also includes a feature processing layer 220, an expert network layer 230 and a model output layer 240 responsible for outputting model operation results.

[0043] Figure 2b FIG. 2 shows a schematic diagram of a feature processing layer 220 according to an exemplary embodiment of the present invention. Figure 2b As shown, the feature processing layer 220 includes a feature weight processing module 221 and a feature combination module 222 , a deformable convolution module 223 and a feature weight interaction module 224 connected to the feature weight processing module 221 .

[0044] The expert network layer includes a primary network module and a high-level network module. The primary network module includes a primary shared expert module and one or more primary task expert modules, and the high-level network module includes a high-level shared expert module and one or more high-level task expert modules, and each primary task expert module corresponds to a high-level task expert module. The present invention does not limit the number of primary task expert modules included in the primary network module and the number of high-level task expert modules included in the high-level network module. Each primary task expert module corresponds to a high-level task expert module, which means that the primary task expert module will send the input result to the corresponding high-level task expert module.

[0045] Figure 2c FIG. 2 shows a schematic diagram of an expert network layer 230 according to an exemplary embodiment of the present invention. Figure 2c As shown, the expert network layer 230 includes a primary network module 231 and a high-level network module 235. The primary network module 231 includes a primary shared expert module 233 and primary task expert modules 232 and 234; the high-level network module 235 includes a high-level shared expert module 237 and high-level task expert modules 236 and 238. The primary task expert module 232 corresponds to the high-level task expert module 236, and the primary task expert module 232 sends the output result to the high-level task expert module 236. The primary task expert module 234 corresponds to the high-level task expert 238, and the primary task expert module 234 sends the output result to the high-level task expert 238.

[0046] Then, step 130 is executed to train the target user determination model according to the user training data set.

[0047] According to one embodiment of the present invention, training a target user determination model based on a user training data set includes: when the user training data set is input into a model output layer, a first training sample is generated based on each user training data in the user data training set to obtain a first training sample set, the first training sample includes a first number of first-order feature vectors, each feature vector is used to characterize historical behavior data under a business category.

[0048] According to an embodiment of the present invention, when generating the first training sample according to each user training data of the user data training set, each data representing the user information in the user training data is mapped to a feature vector, and the feature vector is a first-order feature vector because it does not intersect with other feature vectors. The data representing the user information is mapped to the feature vector, and the embedding method can be used for mapping, and the present invention does not limit the specific method of mapping the data to the feature vector.

[0049] Each first training sample obtained includes a first number of first-order feature vectors. Since the user data contains too much data of business category information, the feature vector generated according to the extracted user data contains a large number of scattered, long-tailed low-frequency features. Directly inputting these feature vectors into the model for training will result in too much calculation during training, which is too time-consuming and not conducive to improving training efficiency. Therefore, the present invention adopts a feature ranking algorithm (Dropout Feature Ranking) to perform feature ranking on the first number of first-order feature vectors, including the steps of: setting the same first drop probability value for the first-order feature vectors in each first training sample; the first drop probability value can be set to 0.5, and the drop probability value represents the probability of the first-order feature vector appearing in the neural network; then the first training sample set is input into the target user determination model for training, so that each first-order feature vector has a certain probability of appearing in the neural network, and regularization is added at the same time, so that the probability of low-importance features appearing in the neural network is low, and a second drop probability value for each first-order feature vector is generated after training, which can represent the feature importance of each first-order feature vector. Regularization is a technique that starts from reducing the complexity of the model. Its essence is to control the number of features learned by the model to minimize it, thereby preventing the introduction of sampling errors in the training set during the training process. By combining these strategies, the model can be better trained to improve accuracy while ensuring good generalization ability. When setting the first discard probability value for each first-order feature vector, a Loss penalty term can also be added to avoid the situation where the discard rate of all features is very low.

[0050] A second training sample is generated based on a second number of first-order feature vectors whose second discard probability value is less than a preset threshold, and a second training sample set is obtained. According to one embodiment of the present invention, the preset threshold can be set to 0.5 or 0.3, and the present invention does not limit the specific setting method of the preset threshold. Finally, the second training sample set is input into the target user determination model for training. When training the target user determination model, the model can be repeatedly trained according to the above method.

[0051] According to one embodiment of the present invention, a training sample set is input into a target user determination model for training, including: inputting training samples in the training sample set into a feature weight processing module of a feature processing layer, and setting a first weight value for each first-order feature vector according to the feature weight processing module.

[0052] According to one embodiment of the present invention, the feature weight processing module can be specifically implemented as a SENet module, and the present invention does not limit the specific implementation method of the feature processing module.

[0053] The feature processing module is used to automatically obtain the importance of each feature through learning, enhance useful features according to their importance, and suppress features that have little effect on the current task. The feature processing module sets different weights for different scenes. Specifically, it can set a higher weight for the first-order feature vector of the main scene and a lower weight for the sub-scenes. This can be achieved through the loss function:

[0054]

[0055] in The value is 0 or 1, indicating whether the i-th sample belongs to the sample space of the k-th scene.

[0056] According to one embodiment of the present invention, the feature processing module performs feature compression on each training sample through a Squeeze operation, a feature channel represented by a real number; then an Excitation operation is performed to generate a weight for each feature channel through the parameter w, i.e., a first weight value; then a Reweight operation is performed to represent the importance of each feature channel after feature selection with the weight of the output of Excitation, and then weighted to the previous feature channel by channel by multiplication to complete the recalibration of the original feature in the channel dimension. By using the feature processing module to perform feature selection on features of different business types, the negative impact of unreliable low-frequency features is weakened based on different vertical field goals, and the role of reliable low-frequency features and important medium and high-frequency features is strengthened.

[0057] Subsequently, the first-order feature vector is processed according to the feature combination module to generate a first high-order feature vector. According to an embodiment of the present invention, the feature combination module can be specifically implemented as an xDeepFM module.

[0058] The feature combination module uses the Hadamard product and other methods to achieve explicit feature interaction on the first-order feature vector to obtain the first high-order feature vector. And when the feature combination module generates the first high-order feature vector, the time consumed is vector-wise, which can reduce the time of generating the first high-order feature vector, and as the number of first-order feature vectors increases, the time consumed increases linearly.

[0059] Subsequently, the first-order feature vector is processed according to the deformable convolution module to generate a second high-order feature vector. According to an embodiment of the present invention, the deformable convolution module can be specifically implemented as a DCN-V2 module.

[0060] The deformable convolution module achieves maximum extraction of feature points by adding a sampling position offset when using the convolution kernel. The generated second-order feature vector can maximize the fusion of the features to be extracted, and obtain a second-order feature vector that better expresses the first-order feature vector information.

[0061] Subsequently, the first-order feature vector is processed according to the feature weight interaction module to generate a third high-order feature vector. According to an embodiment of the present invention, the feature weight interaction module can be specifically implemented as a FiBiNET module.

[0062] The feature weight interaction module includes a SENET layer, which upgrades the obtained first-order feature vector to a feature including feature importance, and obtains the connection between features based on the bilinear interaction method; the generated third-order feature includes not only feature importance but also the degree of connection between features.

[0063] Subsequently, the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector are input into the expert network layer for training to generate a second weight value for each first-order feature vector;

[0064] Generate a second weight value, that is, the weight of each first-order feature vector learned by the feature processing module during training, and then continue training based on the second feature value: input the training samples in the training sample set into the feature weight processing module of the feature processing layer, set a second weight value for each first-order feature vector according to the feature weight processing module, and train the target user determination model.

[0065] The present invention generates a first high-order feature vector, a second high-order feature vector and a third high-order feature vector by using different high-order feature generation modules, and can integrate the module characteristics of each module, maximize the extraction of high-order explicit features and high-order implicit features in user data and capture marginal features in user data. Considering the weights of different features, the generated high-order feature vector can better characterize the characteristics of each sample, which is convenient for classifying users according to the high-order feature vectors and determining target users.

[0066] According to one embodiment of the present invention, the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector are input into the expert network layer for training, including: inputting the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector into a primary shared expert module and one or more primary task expert modules; using the outputs of the primary shared expert module and one or more primary task expert modules as the inputs of the high-level shared expert module; using the outputs of the primary shared expert module and each primary task expert module as the inputs of the high-level task expert module corresponding to the primary task expert module; determining one or more target task expert modules from one or more high-level task expert modules, and determining one or more training target users under a preset business category from a user set based on the outputs of the one or more target task expert modules and the high-level shared expert module; and setting parameters for the expert network layer based on the one or more training target users and one or more calibrated target users calibrated in the user set under the preset business category.

[0067] According to one embodiment of the present invention, the input of the primary shared expert module and each primary task expert module includes a vector of the first high-order feature vector, the second high-order feature vector and the third high-order feature vector. However, the input of the high-level shared expert module is the output of all modules in the primary network module, and the input of the high-level task expert module is the output of the primary shared expert module and the input of the corresponding primary task expert module.

[0068] There are two major problems in the general field of multi-task learning. One is the negative transfer phenomenon. Different tasks of the multi-objective model, especially those with less data, can improve their effects with the help of information transfer. However, when the correlation between the two tasks is weak or very complex, negative transfer often occurs, that is, the effect is very poor after sharing the information embedding. The other is the seesaw phenomenon, where the improvement of the effect of one task is achieved by damaging the effect of another task. For the first problem, by setting the primary network module before the advanced network module, the different expert modules and the shared module interact first, and then enter the advanced network module, negative transfer can be avoided. For the second problem, the expert network layer retains the ability to share features and avoids the interference of harmful parameters by setting up shared expert modules and task expert modules, thereby avoiding the seesaw phenomenon.

[0069] According to one embodiment of the present invention, in order to determine the target users under one or more business categories according to the target user determination model, the expert network layer can be set, including setting one or more high-level task expert modules in one or more high-level task expert modules to be responsible for determining the target users under a business category, and using one or more high-level task expert modules as exclusive expert modules for implementing corresponding tasks under the business category. Each time the target user under a business category is determined, the output of one or more pre-set target task expert modules and high-level shared expert modules is obtained to determine the target user. Through the above configuration, the expert network layer can not only have a good effect when determining the target users under multiple business categories, but also avoid the negative migration phenomenon and accurately determine the target users under each business category.

[0070] According to an embodiment of the present invention, the expert network layer may be specifically implemented as a PLE module.

[0071] According to one embodiment of the present invention, a target user of a business category determined by the target user determination model is a user who intends to buy a car. At the same time, there are other related vertical scenarios on the automobile platform. For example, in addition to buying a car, users may also have vertical needs such as refueling, maintenance, and travel related to the car. These different scenarios can be combined with the basic user portrait of the user on the automobile platform, the behavioral attributes of the general scenario, and the behavioral characteristics of each vertical scenario for multi-scenario modeling. In this way, the user behavior of each scenario can be combined to estimate the user as a whole, and the core attributes of this scenario can be focused on while migrating other scenario information.

[0072] In sample processing, the seed user of each scene is the positive sample of the scene, and other users can be randomly sampled to form negative samples. The user's own static and dynamic portrait features, the user's browsing, clicking, and following of different vertical scenes and columns are embedded at the bottom of the model, and then the features are crossed and selected respectively through the feature processing layer. Usually, there is a correlation between the user's car buying behavior and the car viewing and car use behavior, so the expert network layer extracts the relevant behaviors of different business scenarios and the unique behaviors of the current vertical category. When performing multi-scenario joint training, the loss calculation of each scene needs to merge the samples with the same space and ignore the samples that are not in the sample space of this scene, that is, different scenes still use the user's samples in their respective sample spaces. Finally, based on the predicted probability values ​​of different scene targets, the target users similar to the seed users are determined.

[0073] Finally, step 140 is executed, in response to receiving the target user determination request, the user set to be classified is classified according to the target user determination model, and one or more target users under each preset business classification are determined from the user set to be classified, including the steps of: generating samples to be classified according to user data of the users to be classified in the user classification set, the samples to be classified including a second number of first-order feature vectors; inputting each sample to be classified into the target user determination model to determine whether each user to be classified is a target user under the preset business classification. When performing one or more target users in the user set to be classified, only the second number of first-order feature vectors are collected for each user, thereby reducing the number and dimension of feature vectors input to the model and improving the efficiency of user classification.

[0074] According to an embodiment of the present invention, multiple preset business classifications can be pre-constructed, and by determining whether each user to be classified is a target user under one of the preset business classifications, the users in the set of users to be classified are classified. In the classification result, the user to be classified may not be a target user under any of the preset business classifications, for example, the user to be classified is a new user and has no relevant historical behavior data.

[0075] In the target user determination method of the present invention, a target user determination model is constructed, and when training the target user determination model, the historical behavior data set is first screened according to the business classification to obtain a user data training set, and the user training data set about the business classification is used for training; the training obtains a model capable of target classification, that is, a model capable of determining target users corresponding to multiple business classifications. The model is used to accurately and efficiently determine the target users under each business classification, so as to guide precise business marketing according to the target users under each business classification and improve the business marketing effect. The present invention deeply explores the interactive connections between different vertical categories, strengthens the target population's ability to express diversified interests, and makes the directional expansion of users' preferences in different vertical categories more accurate, and has a certain self-optimization ability. The model is used to support the formulation of marketing activity strategies within the platform, define user groups, conduct precise marketing and refined operations, etc., and has a better promotion effect.

[0076] A target user determination method provided by the present invention is suitable for being executed in a computing device. Figure 3 1 shows a block diagram of a computing device according to an exemplary embodiment of the present invention. In a basic configuration, computing device 300 includes at least one processing unit 320 and system memory 310. According to one aspect, depending on the configuration and type of computing device, system memory 310 includes, but is not limited to, volatile storage (e.g., random access memory), non-volatile storage (e.g., read-only memory), flash memory, or any combination of such memories. According to one aspect, system memory 310 includes operating system 311.

[0077] According to one aspect, operating system 311, for example, is suitable for controlling the operation of computing device 300. Furthermore, examples may be practiced in conjunction with graphics libraries, other operating systems, or any other application programs, and are not limited to any particular application or system. Figure 3 This basic configuration is illustrated in FIG by those components within dashed line 315. According to one aspect, computing device 300 has additional features or functionality. For example, according to one aspect, computing device 300 includes additional data storage devices (removable and / or non-removable), such as magnetic disks, optical disks, or tapes.

[0078] As stated above, according to one aspect, the program module 312 is stored in the system memory 310. According to one aspect, the program module 312 can be implemented as one or more computer program products, and the present application does not limit the type of computer program products, for example, it can include: email, word processing application, spreadsheet application, database application, slide show application, painting or computer-aided application, web browser, etc. In some embodiments according to the present application, the computer program / instructions related to the target user determination method 100 are packaged as a computer program product, and when these computer programs / instructions are executed by the processor (i.e., the processing unit 320), the target user determination method 100 according to the present application is implemented.

[0079] According to one aspect, the examples may be practiced on a circuit comprising discrete electronic components, a packaged or integrated electronic chip containing logic gates, a circuit utilizing a microprocessor, or a single chip containing electronic components or a microprocessor. Figure 3 Each or many components shown in can be integrated in a system on chip (SOC) on a single integrated circuit to practice examples. According to one aspect, such a SOC device may include one or more processing units, a graphics unit, a communication unit, a system virtualization unit, and various application functions, all of which are integrated (or "burned") into a chip substrate as a single integrated circuit. When operated via SOC, the functions described in this article can be operated via a dedicated logic integrated with other components of the computing device 800 on a single integrated circuit (chip). Embodiments of the present invention can also be practiced using other technologies capable of performing logical operations (such as AND, OR, and NOT), including but not limited to mechanical, optical, fluid, and quantum technologies. In addition, embodiments of the present invention can be practiced in a general-purpose computer or in any other circuit or system.

[0080] According to one aspect, the computing device 300 may also have one or more input devices 331, such as a keyboard, a mouse, a pen, a voice input device, a touch input device, etc. Output devices 332 may also be included, such as a display, a speaker, a printer, etc. The aforementioned devices are examples and other devices may also be used. The computing device 300 may include one or more communication connections 333 that allow communication with other computing devices 340. Examples of suitable communication connections 333 include, but are not limited to: RF transmitters, receivers and / or transceiver circuits; Universal Serial Bus (USB), parallel and / or serial ports. The computing device 300 can be communicatively connected to other computing devices 340 via the communication connection 333.

[0081] The embodiment of the present invention also provides a non-transitory readable storage medium, which stores instructions, and the instructions are used to make the computing device perform a method according to an embodiment of the present invention. The readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be a computer-readable instruction, a data structure, a module of a program, or other data. Examples of readable storage media include, but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, tape disk storage or other magnetic storage device or any other non-transitory readable storage medium.

[0082] According to one aspect, communication media is implemented by computer readable instructions, data structures, program modules, or other data in a modulated data signal (e.g., a carrier wave or other transport mechanism), and includes any information delivery media. According to one aspect, the term "modulated data signal" describes a signal that has one or more characteristics set or changed in a manner that encodes information in the signal. By way of example and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, radio frequency (RF), infrared, and other wireless media.

[0083] It should be noted that, although the computing device shown above only includes a processing unit 320, a system memory 310, an input device 331, an output device 332, and a communication connection 333, in a specific implementation, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the device may only include components necessary for implementing the embodiments of the present specification, and need not include all the components shown in the figure.

[0084] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0085] The embodiment of the present invention also includes: A11. A computer program product, comprising a computer program / instructions, wherein the computer program / instructions, when executed by a processor, implements a method as described in any one of A1-A8.

[0086] Similarly, it should be understood that in order to streamline the present disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof.

[0087] In the case where the program code is executed on a programmable computer, the computing device generally includes a processor, a storage medium readable by the processor (including volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device. The memory is configured to store the program code; the processor is configured to execute the target user determination method of the present invention according to the instructions in the program code stored in the memory.

[0088] By way of example and not limitation, computer readable media include computer storage media and communication media. Computer readable media include computer storage media and communication media. Computer storage media stores information such as computer readable instructions, data structures, program modules or other data. Communication media generally embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transmission mechanism, and includes any information delivery medium. Any combination of the above is also included within the scope of computer readable media.

[0089] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and instructional purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art. The disclosure of the present invention is illustrative, rather than restrictive, with respect to the scope of the present invention.

Claims

1. A method for determining a target user, executed in a computing device, the method comprising: Screening the user's historical behavior data set according to multiple business categories to determine a user set including multiple users; Generate user training data based on the historical behavior data of each user in the user set to obtain a user training data set for business classification; Constructing a target user determination model, wherein the target user determination model includes a feature processing layer and an expert network layer connected to the feature processing layer; Training the target user determination model according to the user training data set; In response to receiving the target user determination request, the set of users to be classified is classified according to the target user determination model, and one or more target users under each preset service category are determined from the set of users to be classified.

2. The method of claim 1, wherein: The feature processing layer includes a feature weight processing module and a feature combination module, a deformable convolution module and a feature weight interaction module connected to the feature weight processing module.

3. The method of claim 1, wherein: The expert network layer includes a primary network module and an advanced network module. The primary network module includes a primary shared expert module and one or more primary task expert modules. The advanced network module includes an advanced shared expert module and one or more advanced task expert modules. Each primary task expert module corresponds to a high-level task expert module, and the output result is input into the corresponding high-level task expert module.

4. The method according to any one of claims 1 to 3, wherein: The training of the target user determination model according to the user training data set includes: A first training sample is generated according to each user training data of the user data training set to obtain a first training sample set, wherein the first training sample includes a first number of first-order feature vectors, each feature vector being used to characterize historical behavior data under a business classification.

5. The method of claim 4, further comprising: Setting the same first discard probability value for the first-order feature vector in each first training sample; Inputting the first training sample set into the target user determination model for training to generate a second discard probability value for each first-order feature vector; Generate a second training sample according to a second number of first-order feature vectors whose second discard probability value is less than a preset threshold, to obtain a second training sample set; The second training sample set is input into the target user determination model for training.

6. The method of claim 5, wherein: Input the training sample set into the target user determination model for training, including: Inputting the training samples in the training sample set into the feature weight processing module of the feature processing layer, and setting a first weight value for each first-order feature vector according to the feature weight processing module; Processing the first-order feature vector according to the feature combination module to generate a first high-order feature vector; Processing the first-order feature vector according to the deformable convolution module to generate a second high-order feature vector; Processing the first-order feature vector according to the feature weight interaction module to generate a third high-order feature vector; Inputting the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector into the expert network layer for training to generate a second weight value for each first-order feature vector; The training samples in the training sample set are input into the feature weight processing module of the feature processing layer, and a second weight value is set for each first-order feature vector according to the feature weight processing module to train the target user determination model.

7. The method of claim 6, wherein: The step of inputting the first high-order feature vector, the second high-order feature vector, and the third high-order feature vector into the expert network layer for training includes: Inputting the first high-order feature vector, the second high-order feature vector and the third high-order feature vector into the primary shared expert module and one or more primary task expert modules; Using the outputs of the primary shared expert module and one or more primary task expert modules as inputs to the high-level shared expert module; Using the output of the primary shared expert module and each primary task expert module as the input of the high-level task expert module corresponding to the primary task expert module; Determine one or more target task expert modules from one or more high-level task expert modules, and determine one or more training target users under a preset business category from a user set according to outputs of the one or more target task expert modules and the high-level shared expert module; Parameters of the expert network layer are set according to one or more training target users and one or more calibrated target users calibrated in the user set under a predetermined service category.

8. The method according to any one of claims 1 to 7, wherein: The step of classifying the set of users to be classified according to the target user determination model and determining one or more target users under each preset service category from the set of users to be classified includes: Generate a sample to be classified according to user data of a user to be classified in the set of users to be classified, wherein the sample to be classified includes a second number of first-order feature vectors; Each sample to be classified is input into the target user determination model to determine whether each user to be classified is a target user under a preset business classification.

9. A computing device comprising: one or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in a memory and configured to be executed by one or more processors, and the one or more programs include instructions for executing the method according to any one of claims 1-8.

10. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions which, when executed by a computing device, cause the computing device to perform the method according to any one of claims 1-8.