Method and system for constructing screening model and method and system for determining potential users

Abstract

The method for constructing the potential user screening model according to one aspect of the invention comprises the following steps: training a gradient boosting decision tree model by using a first training data set to obtain a trained gradient boosting decision tree model, wherein the first training data set is determined based on a plurality of structural features of a user and a user tag; determining partial structured features in the plurality of structured features of the user through a trained gradient boosting decision tree model; training the reordering model by using a second training data set to obtain a trained reordering model, wherein the second training data set is determined at least based on the partial structured features of the user, the non-structured features of the user and the user tag; and constructing a joint loss function at least based on the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained reordering model, and updating the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model by using the joint loss function to construct a potential user screening model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology, and more particularly to a method and system for constructing a potential user screening model, a method and system for determining a potential user, a computer readable storage medium and a computer program product for implementing the above method. BACKGROUND

[0002] With the development of Internet technology, enterprises and organizations have established more and more online business platforms to meet the needs of users for online services. These platforms have accumulated a large amount of user data, such as including user's basic information, transaction records, behavior data, etc. Generally, enterprises and organizations need to process massive user data to screen potential users from a large number of users, so as to improve marketing effect and user conversion rate.

[0003] Traditional potential user screening methods mainly rely on simple rule engines, statistical analysis or machine learning models, which can help screen potential users to some extent, but have some limitations. For example, machine learning models mainly rely on user features such as age, gender, consumption level, membership level, ignore the potential value of user features such as user comments, complaint recordings, and have the problems of important user features being covered up and semantic understanding being missing, which affect the final screening effect. For example, although deep learning models can process diversified user features, they cannot distinguish important information in user features and give sufficient attention, but only simply concatenate user features in vector dimensions, resulting in poor model training effect in the case of insufficient user samples and data distribution deviation, which affects the final screening effect. In addition, deep learning models need to consume a large amount of computing resources and processing time for training and reasoning, which is difficult to meet the rapid iteration demand in actual business. SUMMARY

[0004] In order to solve or at least alleviate one or more of the above problems, the following technical solutions are provided.

[0005] According to a first aspect of the present application, there is provided a method for constructing a potential user screening model, the method comprising the following steps: training a gradient boosting decision tree model to obtain a trained gradient boosting decision tree model using a first training dataset, the first training dataset being determined based on a plurality of structured features of a user and a user label; determining, via the trained gradient boosting decision tree model, a portion of the plurality of structured features of the user; training a re-ranking model to obtain a trained re-ranking model using a second training dataset, the second training dataset being determined based on at least the portion of the plurality of structured features of the user, an unstructured feature of the user, and the user label; and constructing a joint loss function based on at least a predicted value of the trained gradient boosting decision tree model and a predicted value of the trained re-ranking model, and updating model parameters of the trained gradient boosting decision tree model and model parameters of the trained re-ranking model using the joint loss function to construct the potential user screening model.

[0006] According to an embodiment of the present application, the method for constructing a potential user screening model, wherein the first training dataset is determined by pre-processing the plurality of structured features of the user, the pre-processing the plurality of structured features of the user comprises: determining numerical features and non-numerical features in the plurality of structured features of the user; detecting outliers in the numerical features and replacing the detected outliers; and selectively performing class merging processing on the non-numerical features to obtain one or more classes of non-numerical features, and performing encoding processing on the one or more classes of non-numerical features.

[0007] According to an embodiment of the present application or any of the above embodiments, the method for constructing a potential user screening model, wherein the second training dataset is determined by pre-processing the unstructured feature of the user, the pre-processing the unstructured feature of the user comprises: performing one or more of de-duplication processing, noise reduction processing, ranking processing, and classification processing on the unstructured feature of the user.

[0008] According to an embodiment of the present application or any of the above embodiments, the method for constructing a potential user screening model, wherein the user label is used to identify whether the user belongs to potential users or not.

[0009] In an embodiment of the method for constructing a potential user screening model according to the present disclosure or any of the above embodiments, training the gradient boosting decision tree model using the first training dataset to obtain the trained gradient boosting decision tree model comprises: generating a plurality of hyperparameter combinations based on a plurality of hyperparameters of the gradient boosting decision tree model to be trained and a discrete value range of each hyperparameter; training the gradient boosting decision tree model using the first training dataset and traversing model performance corresponding to each hyperparameter combination in the plurality of hyperparameter combinations using a grid search algorithm to determine an optimal hyperparameter combination in the plurality of hyperparameter combinations; and obtaining the trained gradient boosting decision tree model based on at least the optimal hyperparameter combination.

[0010] In an embodiment of the method for constructing a potential user screening model according to the present disclosure or any of the above embodiments, the plurality of hyperparameters comprises at least two of the following: a learning rate, a number of trees, a depth of a tree, a minimum weight required for a leaf node, a regularization parameter, a percentage of training data used to train each tree, and a percentage of structured features used to train each tree.

[0011] In an embodiment of the method for constructing a potential user screening model according to the present disclosure or any of the above embodiments, determining, via the trained gradient boosting decision tree model, part of the plurality of structured features of the user comprises: obtaining, via the trained gradient boosting decision tree model, a feature importance of each structured feature of the user; and selecting, based on the feature importance of each structured feature of the user, part of the plurality of structured features of the user.

[0012] In an embodiment of the method for constructing a potential user screening model according to the present disclosure or any of the above embodiments, the feature importance of each structured feature of the user is used to indicate a degree of contribution of the structured feature to a prediction of the trained gradient boosting decision tree model.

[0013] In an embodiment of the method for constructing a potential user screening model according to the present disclosure or any of the above embodiments, the second training dataset is determined by: generating enhanced text feature information of the user based on a combination of part of the plurality of structured features of the user and the unstructured feature of the user; combining, using a separator, the potential user screening rule and the enhanced text feature information of the user to generate combined enhanced text feature information; and determining the second training dataset based on the combined enhanced text feature information and the user label.

[0014] According to the method for constructing a potential user screening model of any one of the preceding embodiments, wherein constructing a joint loss function based at least on the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained reordering model comprises: determining a weighted predicted value based on a weighted sum of the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained reordering model; and constructing the joint loss function based at least on a difference between the weighted predicted value and the user label.

[0015] According to the method for constructing a potential user screening model of any one of the preceding embodiments, wherein the weight distribution of the weighted sum of the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained reordering model is determined based on an importance of the structured feature relative to the unstructured feature.

[0016] According to the method for constructing a potential user screening model of any one of the preceding embodiments, wherein updating the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model using the joint loss function to construct the potential user screening model comprises: determining a gradient of the joint loss function by back propagation; updating the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model according to the gradient of the joint loss function using a gradient descent algorithm, so that the joint loss function is minimized; and constructing the potential user screening model based on the updated model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model.

[0017] According to the method for constructing a potential user screening model of any one of the preceding embodiments, wherein the gradient boosting decision tree model is an extreme gradient boosting decision tree model, and the reordering model is a semantic vector-based reordering model.

[0018] According to the method for constructing a potential user screening model of any one of the preceding embodiments, wherein the model parameters of the trained gradient boosting decision tree model comprise weight parameters of leaf nodes and split parameters of trees, and the model parameters of the trained reordering model comprise layer weights and layer biases.

[0019] According to the method for constructing a potential user screening model of any one of the preceding embodiments, wherein the gradient boosting decision tree model is used to determine a candidate set of potential users from a plurality of users, and the reordering model is used to determine a target potential user from the candidate set of potential users.

[0020] According to a second aspect of the present application, there is provided a system for constructing a potential user screening model, the system comprising: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, execution of the computer program causing the following operations: training a gradient boosting decision tree model using a first training dataset to obtain a trained gradient boosting decision tree model, wherein the first training dataset comprises a plurality of structured features of a user and a user label; determining, via the trained gradient boosting decision tree model, part of the plurality of structured features of the user; training a re-ranking model using a second training dataset to obtain a trained re-ranking model, wherein the second training dataset is determined based at least on the part of the plurality of structured features of the user, unstructured features of a user, and the user label; and constructing a joint loss function based at least on a predicted value of the trained gradient boosting decision tree model and a predicted value of the trained re-ranking model, updating model parameters of the trained gradient boosting decision tree model and model parameters of the trained re-ranking model using the joint loss function to construct the potential user screening model.

[0021] According to a third aspect of the present application, there is provided a computer readable storage medium comprising instructions which, when executed, perform the steps of the method for constructing a potential user screening model according to the first aspect of the present application.

[0022] According to a fourth aspect of the present application, there is provided a computer program product comprising instructions which, when executed by a processor, implement the steps of the method for constructing a potential user screening model according to the first aspect of the present application.

[0023] According to a fifth aspect of the present application, there is provided a method for determining a potential user, the method comprising the following steps: obtaining user features of a plurality of users, the user features comprising a plurality of structured features of a user and unstructured features of a user; inputting the user features into a potential user screening model to obtain a target potential user, wherein the potential user screening model is constructed according to the method for constructing a potential user screening model according to the first aspect of the present application.

[0024] According to a sixth aspect of the present application, a system for determining potential users is provided, the system comprising: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, the running of the computer program causing the following operations: obtaining user features of a plurality of users, the user features comprising a plurality of structured features of the users and unstructured features of the users; inputting the user features into a potential user screening model to obtain target potential users, wherein the potential user screening model is constructed according to the method for constructing a potential user screening model of the first aspect of the present application.

[0025] According to a seventh aspect of the present application, a computer-readable storage medium is provided, comprising instructions which, when executed, perform the steps of the method for determining potential users according to the fifth aspect of the present application.

[0026] According to an eighth aspect of the present application, a computer program product is provided, comprising instructions which, when executed by a processor, implement the steps of the method for determining potential users according to the fifth aspect of the present application.

[0027] According to the scheme for constructing a potential user screening model of one or more embodiments of the present application, the first training data set comprising a plurality of structured features of users and user labels is used to train a gradient boosting decision tree model, and the second training data set determined based at least on part of the structured features of the users, unstructured features of the users and user labels is used to train a reordering model, a joint loss function is constructed based at least on the predicted values of the trained gradient boosting decision tree model and the predicted values of the trained reordering model, the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model are updated using the joint loss function to construct the potential user screening model, through the synergy of the gradient boosting decision tree model and the reordering model, multi-modal data collaborative mining based on the structured features and unstructured features of the users is realized, important user features in the structured and unstructured features can be mined even when the sample data quality is uneven, the accuracy of user feature extraction is improved, thereby improving the prediction effect of the potential user screening model, the construction of the joint loss function makes it only necessary to fit the residual between the predicted values and the actual values, thereby improving the construction efficiency of the potential user screening model. The potential user screening model constructed according to the scheme for constructing a potential user screening model of one or more embodiments of the present application adopts a phased prediction framework, uses the gradient boosting decision tree model to determine a candidate set of potential users from a plurality of users, and uses the reordering model to determine target potential users from the candidate set of potential users, which significantly improves the screening efficiency while ensuring the screening accuracy, and can meet the rapid response demand under a large-scale user. BRIEF DESCRIPTION OF DRAWINGS

[0028] The above and / or other aspects and advantages of the present application will become more apparent by describing in detail various aspects of the present application with reference to the attached drawings, in which like reference numerals refer to like elements, and wherein:

[0029] Figure 1 A flowchart of a method for building a potential user screening model according to one or more embodiments of the present application is shown.

[0030] Figure 2 A process diagram for building a potential user screening model according to one or more embodiments of the present application is shown.

[0031] Figure 3 A block diagram of a system for building a potential user screening model according to one or more embodiments of the present application is shown.

[0032] Figure 4 A flowchart of a method for determining a potential user according to one or more embodiments of the present application is shown.

[0033] Figure 5 A block diagram of a system for determining a potential user according to one or more embodiments of the present application is shown. DETAILED DESCRIPTION

[0034] The present application will now be described more fully with reference to the accompanying drawings, in which exemplary embodiments of the application are shown. The present application may, however, be embodied in different forms, and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and fully convey the scope of the application to those skilled in the art.

[0035] In this specification, terms such as "comprise" and "include" indicate that the technical solution of the present application does not exclude the presence of other elements and steps not directly or explicitly stated in the specification and claims.

[0036] Unless specifically stated, terms such as "first" and "second" are not used to denote order or importance, but are used to distinguish one element from another.

[0037] In the context of this application, a user's structured characteristics refer to user characteristics with clear definitions and fixed formats, typically represented in the form of data tables for easy statistical analysis and processing. For example, a user's structured characteristics may include their age, gender, transaction amount, membership level, application activity, etc. A user's unstructured characteristics refer to user characteristics without clear definitions and fixed formats, typically represented in non-data table formats, such as text, images, audio, video, etc. For example, a user's unstructured characteristics may include user comment data (e.g., comment text, images), behavioral data (e.g., user browsing paths and behavior records on a website), and click data (e.g., application click sequences, dwell time), etc.

[0038] In the following, various exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings.

[0039] Figure 1 A flowchart illustrating a method for constructing a potential user screening model according to one or more embodiments of this application is shown.

[0040] like Figure 1 As shown, in step S101, the gradient boosting decision tree model is trained using the first training dataset to obtain the trained gradient boosting decision tree model. The first training dataset is determined based on multiple structured features of the user and user labels.

[0041] Optionally, multiple structured features of a user can be obtained and preprocessed. The preprocessed structured features and corresponding user labels are then used as the first training dataset. In one embodiment, user labels can be used to identify whether a user is a potential user or not. For example, a user label can be implemented as (0,1), where 1 indicates that the user is a potential user and 0 indicates that the user is not a potential user.

[0042] In one embodiment, preprocessing multiple structured features of a user may include: identifying numerical and non-numerical features among the multiple structured features of the user; detecting outliers in the numerical features and replacing the detected outliers; selectively performing category merging on the non-numerical features to obtain one or more categories of non-numerical features; and encoding the one or more categories of non-numerical features, such as using One-Hot encoding. In one embodiment, the first training dataset may include preprocessed numerical features, preprocessed non-numerical features, and corresponding user labels, wherein the preprocessed non-numerical features may include multiple vectors that have been encoded and bundled together.

[0043] As an example, the quartile method can be utilized to detect outliers in the numerical type features and replace the detected outliers with a judgment threshold. For example, the quartile method can be utilized to detect the daily transaction amount of a user, and determine the daily transaction amount of more than 10,000 yuan as an outlier and replace the outlier with more than 10,000 yuan. As an example, low-frequency category merging can be performed on the non-numerical type features to obtain one or more categories of non-numerical type features, i.e., merging multiple non-numerical type features with lower frequency into one category of non-numerical type features. For example, by performing low-frequency category merging on the user's location, the user's location can be merged into Beijing, Shanghai, Shenzhen, the United States, Europe, etc.

[0044] Optionally, in step S101, a plurality of hyperparameter combinations can be generated based on a plurality of hyperparameters of the gradient boosting decision tree model to be trained and a discrete value range of each hyperparameter, the gradient boosting decision tree model is trained using the first training data set and the model performance (e.g., accuracy, recall, etc.) corresponding to each hyperparameter combination in the plurality of hyperparameter combinations is traversed using a grid search algorithm, the optimal hyperparameter combination in the plurality of hyperparameter combinations is determined, and the trained gradient boosting decision tree model is obtained based on at least the optimal hyperparameter combination. Optionally, the hyperparameters of the gradient boosting decision tree model to be trained can include learning rate, number of trees, depth of tree, minimum weight required for leaf node, regularization parameter, percentage of training data used to train each tree, percentage of structured features used to train each tree, etc. In an embodiment, the gradient boosting decision tree model can be implemented as an eXtreme Gradient Boosting (XGBoost) model.

[0045] In step S103, part of the structured features in the plurality of structured features of the user is determined via the trained gradient boosting decision tree model.

[0046] Optionally, in step S103, the feature importance of each structured feature of the user can be obtained via the trained gradient boosting decision tree model, and part of the structured features of the user is selected based on the feature importance of each structured feature of the user, for example, part of the important structured features is selected from the plurality of structured features of the user based on the comparison of the feature importance of each structured feature with a feature importance threshold. Optionally, the feature importance of a structured feature can be used to indicate the degree of contribution of the structured feature to the prediction of the trained gradient boosting decision tree model.

[0047] In one embodiment, the gain of each structured feature of the user, i.e., the sum of the improvement of the objective function in all trees represented by the structured feature, can be determined via the trained gradient boosting decision tree model to represent the feature importance of each structured feature by using the gain of each structured feature, and the higher the gain, the more important the structured feature. In one embodiment, the coverage of each structured feature of the user, i.e., the total number of samples covered in all trees represented by the structured feature, can be determined via the trained gradient boosting decision tree model to represent the feature importance of each structured feature by using the coverage of each structured feature, and the higher the coverage, the more important the structured feature. In one embodiment, the frequency of each structured feature of the user, i.e., the number of times used in all trees represented by the structured feature, can be determined via the trained gradient boosting decision tree model to represent the feature importance of each structured feature by using the frequency of each structured feature, and the higher the frequency, the more important the structured feature. Illustratively, the age and gender of the user can be selected as part of the structured features from the age, gender, transaction amount, membership level, application activity of the user based on the gain of each structured feature. By determining part of the structured features of the user via the trained gradient boosting decision tree model, important features of the structured features of the user can be extracted to improve the training speed and prediction performance of the subsequent model.

[0048] In step S105, the reordering model is trained to obtain a trained reordering model using a second training data set determined based on at least part of the structured features of the user, the unstructured features of the user and the user label.

[0049] Optionally, the unstructured features of the user can be obtained and preprocessed, and the second training data set can be determined based on at least part of the structured features of the user, the preprocessed unstructured features of the user and the corresponding user label determined in step S103. In one embodiment, preprocessing the unstructured features of the user can include performing de-duplication processing, noise reduction processing, sorting processing, classification processing, etc. on the unstructured features of the user. Illustratively, the user click sequence (e.g., a click sequence of sequentially clicking the homepage, my, assets, balance of the application), the stay duration, the comment text, etc. can be preprocessed to eliminate invalid or duplicate records to generate the preprocessed unstructured features of the user, such as "user A tends to choose high-performance electronic products, and recent comments show a high degree of attention to sound quality and noise reduction function". It should be noted that the preprocessed unstructured features of the user can be understood as the unstructured features of the user described in natural language, which converts the user preferences, behavior patterns, interaction features, etc. implied in the original user data into user text portraits through semantic expression, thereby improving the prediction performance of the subsequent model.

[0050] In an embodiment, the enhanced text feature information of the user can be generated based on a combination of part of the plurality of structured features of the user and the unstructured feature of the user, the potential user screening rule and the enhanced text feature information of the user are combined by using the separator to generate the combined enhanced text feature information, and the combined enhanced text feature information and the user label are taken as the second training data set. As an example, the selected part of the plurality of structured features can be the age and gender of the user, such as “middle-aged” and “male”, the unstructured feature of the user can be “frequently use electronic type consumer coupons”, and the enhanced text feature information of the user generated by the combination can be “middle-aged male frequently use electronic type consumer coupons”. Exemplarily, the combined enhanced text feature information can be represented as [query, SEP, document], wherein query represents the potential user screening rule, SEP represents the separator, and document represents the enhanced text feature information “middle-aged male frequently use electronic type consumer coupons” of the user. For example, the potential user screening rule can be that the user is expected to participate in a certain marketing activity, the user is expected to consume in a certain consumption scene, the user is expected to redeem electronic coupons, etc. In an embodiment, the reordering model can be implemented as a semantic vector-based reordering model, such as a BGE (BAAI General Embedding)-Reranker model.

[0051] In step S107, a joint loss function is constructed based on at least the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained reordering model, and the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model are updated by using the joint loss function to construct the potential user screening model.

[0052] Optionally, in step S107, a weighted predicted value can be determined based on a weighted sum of the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained reordering model, and a joint loss function is constructed based on at least the difference between the weighted predicted value and the user label. The construction of the joint loss function makes it only necessary to fit the residual between the predicted value and the actual value, reduces the learning difficulty of the model, and improves the learning efficiency of the model.

[0053] In an embodiment, the weighted predicted value p pred :

[0054] p pred =αp XGBoost +(1-α)p BGE-Reranker Formula (1)

[0055] wherein p XGBoosta prediction value of the trained gradient boosting decision tree model, p BGE-Reranker a prediction value of the trained re-ranking model, a represents a weighting weight. In one embodiment, the weighting weight can be determined based on the importance of structured features relative to unstructured features.

[0056] In one embodiment, the joint loss function l can be constructed by the following formula (2):

[0057]

[0058] wherein p BGE-Reranker a prediction value of the trained re-ranking model, a represents a weighting weight, y represents a user label, p pred represents a weighted prediction value.

[0059] Optionally, in step S107, the gradient of the joint loss function can be determined by back propagation, the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model are updated according to the gradient of the joint loss function using a gradient descent algorithm, so that the joint loss function is minimized, and the potential user screening model is constructed based on the updated model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model.

[0060] In one embodiment, the model parameters of the gradient boosting decision tree model can include weight parameters of leaf nodes, splitting parameters of trees, etc., and the model parameters of the re-ranking model can include layer weights, layer biases, etc. In one embodiment, the gradient boosting decision tree model can be used to determine a candidate set of potential users from a plurality of users, and the re-ranking model can be used to determine a target potential user from the candidate set of potential users. By adopting a staged prediction framework, the candidate set of potential users is determined from a plurality of users by the gradient boosting decision tree model, and then the target potential user is determined from the candidate set of potential users by the re-ranking model, which significantly improves the screening efficiency while ensuring the screening accuracy, and can meet the demand for rapid response under large-scale users.

[0061] According to the method for constructing a potential user screening model of one or more embodiments of the present application, a structured feature modeling + semantic vector enhancement + joint loss function optimization implementation system is proposed. First, the gradient boosting decision tree model is trained using multiple structured features of the user, and important structured features in the multiple structured features of the user are determined via the trained gradient boosting decision tree model. Then, the important structured features are fused with the unstructured features of the user to obtain enhanced text feature information of the user, such as constructing [query, SEP, document] format data to form a semantic vector, and the training of the re-ranking model is completed using the semantic vector. Finally, the optimization of the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model is completed by constructing a joint loss function, so that only the residual between the predicted value and the actual value needs to be fitted, the learning difficulty of the model is reduced, and the learning efficiency of the model is improved.

[0062] According to the method for constructing a potential user screening model of one or more embodiments of the present application, the gradient boosting decision tree model is trained using a first training data set including multiple structured features of the user and a user label, and the re-ranking model is trained using a second training data set determined based on at least part of the structured features of the user, the unstructured features of the user, and the user label. A joint loss function is constructed based on at least the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained re-ranking model, and the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model are updated using the joint loss function to construct a potential user screening model. Through the synergistic effect of the gradient boosting decision tree model and the re-ranking model, multi-modal data collaborative mining based on structured features and unstructured features of the user is realized, so that important user features in the structured and unstructured features can be mined even when the sample data quality is uneven, the accuracy of user feature extraction is improved, and thus the prediction effect of the potential user screening model is improved. The construction of the joint loss function makes it necessary to only fit the residual between the predicted value and the actual value, thereby improving the construction efficiency of the potential user screening model.

[0063] Figure 2 A process schematic diagram for constructing a potential user screening model according to one or more embodiments of the present application is shown.

[0064] As Figure 2As shown in FIG. 2, the process 200 for building a potential user screening model includes training of a gradient boosting decision tree model using a first training dataset, training of a re-ranking model using a second training dataset, and optimization of model parameters of the trained gradient boosting decision tree model and model parameters of the trained re-ranking model using a joint loss function, where the second training dataset is built based on outputs of the trained gradient boosting decision tree model, and the joint loss function is built based on at least predicted values of the trained gradient boosting decision tree model and predicted values of the trained re-ranking model. The following describes each stage in detail. Figure 2 The implementation process of each stage is further described.

[0065] In one embodiment, a plurality of hyperparameter combinations can be generated based on a plurality of hyperparameters of the gradient boosting decision tree model to be trained and a discrete value range of each hyperparameter, the gradient boosting decision tree model is trained using the first training dataset and a grid search algorithm is used to traverse model performance corresponding to each hyperparameter combination in the plurality of hyperparameter combinations, the optimal hyperparameter combination in the plurality of hyperparameter combinations is determined, and the trained gradient boosting decision tree model is obtained based on at least the optimal hyperparameter combination. Optionally, the hyperparameters of the gradient boosting decision tree model to be trained can include learning rate, number of trees, depth of trees, minimum weight and of leaf nodes, regularization parameter, percentage of training data used to train each tree, percentage of structured features used to train each tree, etc. In one embodiment, the gradient boosting decision tree model can be implemented as an XGBoost model.

[0066] In one embodiment, during the training of the gradient boosting decision tree model using the first training dataset, model performance can be evaluated after each iteration using a test set, evaluation indicators can include accuracy, precision, recall, etc., and when the evaluation indicators improve by a small amount (e.g., less than 0.01), the model training can be stopped to prevent overfitting. Optionally, the first training dataset can include a plurality of structured features of preprocessed users and corresponding user labels, where the user labels are used to identify whether the user belongs to potential users or not.

[0067] In one embodiment, after obtaining the trained gradient boosting decision tree model, the feature importance of each structured feature of the user can be obtained via the trained gradient boosting decision tree model, and part of the important structured features are selected from the plurality of structured features of the user based on the feature importance of each structured feature of the user for the construction of the second training data set. In one embodiment, the enhanced text feature information of the user can be generated based on the combination of part of the structured features in the plurality of structured features of the user and the unstructured features of the user, the combined enhanced text feature information is generated by combining the potential user screening rule with the enhanced text feature information of the user using a separator, and the combined enhanced text feature information and the user label are used as the second training data set. As an example, part of the structured features selected from the plurality of structured features can be the age and gender of the user, such as "middle-aged" and "male", the unstructured features of the user can be "frequently use electronic type consumer coupons", and the enhanced text feature information of the user generated by the combination of the two can be "middle-aged male frequently use electronic type consumer coupons". Exemplarily, the combined enhanced text feature information can be represented as [query, SEP, document], wherein query represents the potential user screening rule, SEP represents the separator, and document represents the enhanced text feature information of the user "middle-aged male frequently use electronic type consumer coupons". In one embodiment, the reordering model can be implemented as a semantic vector-based reordering model, such as a BGE-Reranker model.

[0068] In one embodiment, after obtaining the trained reordering model, the weighted prediction value can be determined based on the weighted sum of the prediction value of the trained gradient boosting decision tree model and the prediction value of the trained reordering model, and the joint loss function is constructed based on at least the difference between the weighted prediction value and the user label. The construction of the joint loss function makes it only necessary to fit the residual between the predicted value and the actual value, reduces the learning difficulty of the model, and improves the learning efficiency of the model.

[0069] In one embodiment, the weighted prediction value can be calculated by the above formula (1), and the joint loss function can be constructed by the above formula (2). After constructing the joint loss function, the gradient of the joint loss function can be determined by back propagation, and the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model are optimized according to the gradient of the joint loss function using the gradient descent algorithm, so that the joint loss function is minimized, thereby constructing the potential user screening model based on the optimized model parameters of the trained gradient boosting decision tree model and the model parameters of the trained reordering model. In one embodiment, the model parameters of the gradient boosting decision tree model can include weight parameters of leaf nodes, tree splitting parameters, etc., and the model parameters of the reordering model can include layer weights, layer biases, etc.

[0070] Figure 3 A block diagram of a system for constructing a potential user screening model is shown in accordance with one or more embodiments of the present application.

[0071] As shown in Figure 3 The system 300 for constructing a potential user screening model includes a memory 310, a processor 320, and a computer program 330 stored on the memory 310 and executable on the processor 320. The processor 320 executes the computer program 330 to implement a method for constructing a potential user screening model in accordance with an aspect of the present application.

[0072] Figure 4 A flowchart of a method for determining a potential user is shown in accordance with one or more embodiments of the present application.

[0073] As shown in Figure 4 In step S401, user features of a plurality of users are obtained, the user features including a plurality of structured features and unstructured features of the users.

[0074] In step S402, the user features of the plurality of users are input to a potential user screening model to obtain a target potential user, wherein the potential user screening model is constructed according to the method for constructing a potential user screening model in accordance with one or more embodiments of the present application.

[0075] Optionally, in step S402, a candidate set of potential users can be determined from the plurality of users based on the plurality of structured features of the users by a gradient boosting decision tree model in the potential user screening model, and a part of the structured features of the users can be determined, and the target potential user can be determined from the candidate set of potential users based on the unstructured features and the part of the structured features of the users by a reordering model in the potential user screening model.

[0076] In one embodiment, a dynamic threshold (such as a probability value of 0.8) can be set to filter out a part of the users (e.g., with a probability value lower than 0.8) from the plurality of users to obtain the candidate set of potential users (e.g., about 5% of the original amount of users). By utilizing the efficient decision path of the gradient boosting decision tree model, a preliminary screening of tens of millions of users can be performed in minutes, significantly improving the screening efficiency of the target potential user. In one embodiment, a predicted probability of each candidate potential user in the candidate set of potential users can be determined based on the unstructured features and the part of the structured features of the users, and the predicted probabilities can be sorted, and the target potential user can be determined based on the sorted predicted probabilities. Exemplarily, one or more candidate potential users in the top 10% of the sorted predicted probabilities can be determined as the target potential user.

[0077] The method for determining potential users according to one or more embodiments of the present application adopts a staged prediction framework, determines a candidate set of potential users from a plurality of users by using a gradient boosting decision tree model, and determines a target potential user from the candidate set of potential users by using a re-ranking model, thereby significantly improving screening efficiency while ensuring screening accuracy and meeting the demand for rapid response under a large-scale user scenario. As an example, Table 1 below shows the user screening effect achieved by a potential user screening model constructed according to one or more embodiments of the present application:

[0078] Table 1

[0079] User filtering rules Region Reach Conversion rate Increase over control Rule A Region 1-15 37 million 2％ 3x Rule B Region 16 4 million 27％ 5x Rule C Region 17 34 thousand 3％ 5x

[0080] It should be noted that the number of reached users in Table 1 refers to the number of target potential users screened out by the potential user screening model constructed according to one or more embodiments of the present application through multiple experiments, the conversion rate refers to the conversion rate of the target potential users converted into actual users, and the improvement compared with the control group refers to the improvement multiple of the conversion rate achieved by the potential user screening model constructed according to one or more embodiments of the present application compared with the traditional screening method. By way of example, the user screening rules shown in Table 1 can include old-for-new promotion, consumption coupon promotion, card management, etc.

[0081] Figure 5 A block diagram of a system for determining potential users according to one or more embodiments of the present application is shown.

[0082] As shown in Figure 5 The system 500 for determining potential users includes a memory 510, a processor 520, and a computer program 530 stored on the memory 510 and executable on the processor 520. The processor 520 executes the computer program 530 to implement the method for determining potential users according to an aspect of the present application.

[0083] The present application can also be implemented as a computer readable storage medium including instructions that, when executed, perform the method for constructing a potential user screening model according to an aspect of the present application. In addition, the present application can also be implemented as a computer readable storage medium including instructions that, when executed, perform the method for determining potential users according to an aspect of the present application.

[0084] Where applicable, various embodiments provided by the present application can be implemented using hardware, software, or combinations of hardware and software. Also, where applicable, the various hardware components and / or software components set forth herein can be combined into a composite component including software, hardware, and / or both, where applicable. It will depend on the particular application, specific design constraints imposed, and / or other factors, as is properly tailored by those of ordinary skill having the benefit of the present disclosure. What is described herein is not intended as an exhaustive description of the various ways in which a given unit of hardware and / or software might be implemented. As such, the following description is intended to provide an appropriate basis for the understanding of the various embodiments provided by the present application.

[0085] Software, such as program code and / or data, according to the present application, can be stored on one or more computer storage media. It is also contemplated that software identified herein can be implemented using one or more general purpose or special purpose computers and / or computer systems, networked and / or otherwise. Where applicable, the order of the blocks of the processes can be changed, combined into composite blocks, and / or separated into sub-blocks to provide features described herein.

[0086] The embodiments and examples set forth herein are provided to best explain and provide examples of embodiments of the present application and to enable one of ordinary skill in the art to practice and use the present application. It is intended that the embodiments and examples contained herein be considered in a descriptive sense only and not limiting.

Claims

1. A method for building a potential user screening model, characterized in that, The method comprises the following steps: training a gradient boosting decision tree model using a first training dataset to obtain a trained gradient boosting decision tree model, the first training dataset being determined based on a plurality of structured features of a user and a user label; determining, via the trained gradient boosting decision tree model, part of the plurality of structured features of the user; training a re-ranking model using a second training dataset to obtain a trained re-ranking model, the second training dataset being determined based on at least part of the plurality of structured features of the user, unstructured features of the user and the user label; and constructing a joint loss function based on at least predicted values of the trained gradient boosting decision tree model and predicted values of the trained re-ranking model, and updating model parameters of the trained gradient boosting decision tree model and model parameters of the trained re-ranking model using the joint loss function to construct the potential user screening model.

2. The method of claim 1, wherein the first training dataset is determined at least by pre-processing the plurality of structured features of the user, the pre-processing the plurality of structured features of the user comprising: determining numerical features and non-numerical features in the plurality of structured features of the user; detecting outliers in the numerical features and replacing the detected outliers; and selectively performing class merging processing on the non-numerical features to obtain one or more classes of non-numerical features, and performing encoding processing on the one or more classes of non-numerical features.

3. The method of claim 1, wherein the second training dataset is determined at least by pre-processing the unstructured features of the user, the pre-processing the unstructured features of the user comprising: performing one or more of de-duplication processing, noise reduction processing, ranking processing, classification processing on the unstructured features of the user.

4. The method of claim 1, wherein the user label is used to identify whether the user belongs to potential users or not.

5. The method of claim 1, wherein training a gradient boosting decision tree model using a first training dataset to obtain a trained gradient boosting decision tree model comprises: generating a plurality of hyperparameter combinations based on a plurality of hyperparameters of the gradient boosting decision tree model to be trained and discrete value ranges of each hyperparameter; training the gradient boosting decision tree model using the first training dataset and traversing model performances corresponding to each hyperparameter combination in the plurality of hyperparameter combinations using a grid search algorithm to determine an optimal hyperparameter combination in the plurality of hyperparameter combinations; and obtaining the trained gradient boosting decision tree model based on at least the optimal hyperparameter combination.

6. The method of claim 5, wherein the plurality of hyperparameters comprises at least two of the following: learning rate, number of trees, depth of trees, minimum weight required for leaf nodes, regularization parameter, percentage of training data used to train each tree, percentage of structured features used to train each tree. ​ ​ 7. The method of claim 1, wherein determining, via the trained gradient boosting decision tree model, the subset of structured features from the plurality of structured features of the user comprises: obtaining feature importance of each structured feature of the user via the trained gradient boosting decision tree model; and selecting the subset of structured features from the plurality of structured features of the user based on the feature importance of each structured feature of the user.

8. The method of claim 7, wherein the feature importance of each structured feature of the user is indicative of a degree of contribution of the structured feature to a prediction of the trained gradient boosting decision tree model.

9. The method of claim 1, wherein the second training dataset is determined by: generating enhanced textual feature information of the user based on a combination of the subset of structured features from the plurality of structured features of the user and the unstructured feature of the user; combining the potential user screening rule with the enhanced textual feature information of the user using a delimiter to generate combined enhanced textual feature information; and determining the second training dataset based on the combined enhanced textual feature information and the user label.

10. The method of claim 1, wherein constructing a joint loss function based on at least a predicted value of the trained gradient boosting decision tree model and a predicted value of the trained re-ranking model comprises: determining a weighted predicted value based on a weighted sum of the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained re-ranking model; and constructing the joint loss function based on at least a difference between the weighted predicted value and the user label.

11. The method of claim 10, wherein a weight assignment of the weighted sum of the predicted value of the trained gradient boosting decision tree model and the predicted value of the trained re-ranking model is determined based on an importance of the structured feature relative to the unstructured feature.

12. The method of claim 1, wherein updating model parameters of the trained gradient boosting decision tree model and model parameters of the trained re-ranking model using the joint loss function to construct the potential user screening model comprises: determining a gradient of the joint loss function by backpropagation; updating the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model according to the gradient of the joint loss function using a gradient descent algorithm such that the joint loss function is minimized; and constructing the potential user screening model based on the updated model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model.

13. The method of claim 1, wherein the gradient boosting decision tree model is an extreme gradient boosting decision tree model and the re-ranking model is a semantic vector based re-ranking model.

14. The method of claim 1, wherein the model parameters of the trained gradient boosting decision tree model comprise weight parameters of leaf nodes, split parameters of trees, and the model parameters of the trained re-ranking model comprise layer weights and layer biases. 15.The method of claim 1, wherein the gradient boosting decision tree model is used to determine a candidate set of potential users from a plurality of users, and the re-ranking model is used to determine a target potential user from the candidate set of potential users.

16. A system for building a potential user screening model, the system comprising: The system comprises: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, the running of the computer program causing the following operations: training a gradient boosting decision tree model using a first training dataset to obtain a trained gradient boosting decision tree model, the first training dataset being determined based on a plurality of structured features of a user and a user label; determining, via the trained gradient boosting decision tree model, a part of the plurality of structured features of the user; training a re-ranking model using a second training dataset to obtain a trained re-ranking model, the second training dataset being determined based on at least the part of the plurality of structured features of the user, unstructured features of the user, and the user label; and constructing a joint loss function based on at least a predicted value of the trained gradient boosting decision tree model and a predicted value of the trained re-ranking model, and updating model parameters of the trained gradient boosting decision tree model and model parameters of the trained re-ranking model using the joint loss function to construct the potential user screening model. 17.The system of claim 16, wherein the first training dataset is determined at least by pre-processing the plurality of structured features of the user, the pre-processing the plurality of structured features of the user comprising: determining numerical features and non-numerical features in the plurality of structured features of the user; detecting outliers in the numerical features and replacing the detected outliers; and selectively performing class merging processing on the non-numerical features to obtain one or more classes of non-numerical features, and performing encoding processing on the one or more classes of non-numerical features. 18.The system of claim 16, wherein the second training dataset is determined at least by pre-processing the unstructured features of the user, the pre-processing the unstructured features of the user comprising: performing one or more of de-duplication processing, noise reduction processing, ranking processing, and classification processing on the unstructured features of the user. 19.The system of claim 16, wherein the user label is used to identify whether the user belongs to a potential user or not. 20.The system of claim 16, wherein the running of the computer program causing training a gradient boosting decision tree model using a first training dataset to obtain a trained gradient boosting decision tree model comprises: generating a plurality of hyperparameter combinations based on a plurality of hyperparameters of the gradient boosting decision tree model to be trained and a discrete value range of each hyperparameter; training the gradient boosting decision tree model using the first training dataset and traversing model performance corresponding to each hyperparameter combination in the plurality of hyperparameter combinations using a grid search algorithm to determine an optimal hyperparameter combination in the plurality of hyperparameter combinations; and ​ and obtaining the trained gradient boosting decision tree model based at least on the optimal combination of hyperparameters.

21. The system of claim 20, wherein the plurality of hyperparameters comprises at least two of: a learning rate, a number of trees, a depth of trees, a minimum weight sum required at a leaf node, a regularization parameter, a percentage of training data used to train each tree, a percentage of structured features used to train each tree.

22. The system of claim 16, wherein execution of the computer program causes determining, via the trained gradient boosting decision tree model, a subset of structured features from the plurality of structured features of the user comprises: obtaining feature importances of each structured feature of a user via the trained gradient boosting decision tree model; and selecting the subset of structured features from the plurality of structured features of the user based on the feature importances of each structured feature of the user.

23. The system of claim 22, wherein the feature importance of each structured feature of the user is used to indicate a degree of contribution of the structured feature to a prediction of the trained gradient boosting decision tree model.

24. The system of claim 16, wherein execution of the computer program causes the second training dataset to be determined by: generating enhanced textual feature information of the user based on a combination of the subset of structured features from the plurality of structured features of the user and the unstructured features of the user; combining the latent user screening rule with the enhanced textual feature information of the user using a delimiter to generate combined enhanced textual feature information; and determining the second training dataset based on the combined enhanced textual feature information and the user label.

25. The system of claim 16, wherein execution of the computer program causes constructing a joint loss function based at least on a prediction value of the trained gradient boosting decision tree model and a prediction value of the trained re-ranking model comprises: determining a weighted prediction value based on a weighted sum of the prediction value of the trained gradient boosting decision tree model and the prediction value of the trained re-ranking model; and constructing the joint loss function based at least on a difference between the weighted prediction value and the user label.

26. The system of claim 25, wherein a weight assignment of the weighted sum of the prediction value of the trained gradient boosting decision tree model and the prediction value of the trained re-ranking model is determined based on an importance of the structured features relative to the unstructured features.

27. The system of claim 16, wherein execution of the computer program causes updating model parameters of the trained gradient boosting decision tree model and model parameters of the trained re-ranking model using the joint loss function to construct the latent user screening model comprises: determining a gradient of the joint loss function by backpropagation; updating the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model according to the gradient of the joint loss function using a gradient descent algorithm such that the joint loss function is minimized; and constructing the potential user screening model based on the model parameters of the trained gradient boosting decision tree model and the model parameters of the trained re-ranking model. 28.The system of claim 16, wherein the gradient boosting decision tree model is an extreme gradient boosting decision tree model and the re-ranking model is a semantic vector based re-ranking model. 29.The system of claim 16, wherein the model parameters of the trained gradient boosting decision tree model comprise weight parameters of leaf nodes and split parameters of trees, and the model parameters of the trained re-ranking model comprise layer weights and layer biases. 30.The system of claim 16, wherein the gradient boosting decision tree model is used to determine a candidate set of potential users from a plurality of users, and the re-ranking model is used to determine a target potential user from the candidate set of potential users.

31. A computer readable storage medium, characterized in that, The computer storage medium comprises instructions that, when executed, perform the method for constructing a potential user screening model according to any one of claims 1-15.

32. A computer program product, characterised in that, The computer program product comprises instructions that, when executed by a processor, implement the method for constructing a potential user screening model according to any one of claims 1-15.

33. A method for determining a potential user, the method comprising: The method comprises the following steps: obtaining user features of a plurality of users, the user features comprising a plurality of structured features of a user and unstructured features of the user; inputting the user features into a potential user screening model to obtain a target potential user, wherein the potential user screening model is constructed according to the method for constructing a potential user screening model according to any one of claims 1-15. 34.The method of claim 33, wherein inputting the user features into a potential user screening model to obtain a target potential user comprises: determining, by a gradient boosting decision tree model in the potential user screening model, a candidate set of potential users from a plurality of users based on the plurality of structured features of the user and determining part of the plurality of structured features of the user; and determining, by a re-ranking model in the potential user screening model, a target potential user from the candidate set of potential users based on the unstructured features of the user and the part of the plurality of structured features. 35.The method of claim 34, wherein determining, by a re-ranking model in the potential user screening model, a target potential user from the candidate set of potential users based on the unstructured features of the user and the part of the plurality of structured features comprises: determining a predicted probability of each candidate potential user in the candidate set of potential users based on the unstructured features of the user and the part of the plurality of structured features and ranking the predicted probabilities; and determining the target potential user based on the ranked predicted probabilities.

36. A system for determining a potential user, the system comprising: The system comprises: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, execution of the computer program causing the following operations: obtaining user features of a plurality of users, the user features comprising a plurality of structured features of a user and unstructured features of the user; inputting the user features to a potential user screening model to obtain a target potential user, the potential user screening model being constructed according to the method for constructing a potential user screening model of any one of claims 1-15.

37. The system of claim 36, wherein the execution of the computer program causes inputting the user features to a potential user screening model to obtain a target potential user comprises: determining, by a gradient boosting decision tree model in the potential user screening model, a candidate set of potential users from a plurality of users based on the plurality of structured features of the user and determining a subset of structured features from the plurality of structured features of the user; and determining, by a re-ranking model in the potential user screening model, a target potential user from the candidate set of potential users based on the unstructured features of the user and the subset of structured features.

38. The system of claim 37, wherein the execution of the computer program causes determining, by the re-ranking model in the potential user screening model, a target potential user from the candidate set of potential users based on the unstructured features of the user and the subset of structured features comprises: determining a predicted probability of each candidate potential user in the candidate set of potential users based on the unstructured features of the user and the subset of structured features and ranking the predicted probabilities; and determining the target potential user based on the ranked predicted probabilities. the computer storage medium comprises instructions that, when executed, perform the method for determining a potential user of any one of claims 33-35. the computer program product comprises instructions that, when executed by a processor, implement the method for determining a potential user of any one of claims 33-35.

39. A computer-readable storage medium, characterized in that, the computer program product comprises instructions that, when executed by a processor, implement the method for determining a potential user of any one of claims 33-35.

40. A computer program product, characterised in that, ​

Images