Method for constructing dynamic expansion quantity prediction model, prediction method and device

Abstract

The application provides a dynamic capacity expansion prediction model construction method, a prediction method and a device. The method comprises the following steps: obtaining first sample data and second sample data of users in a training user set; constructing a Wide&Deep model; training the model by using the sample data; and taking the trained Wide&Deep model as a dynamic capacity expansion prediction model. The application can improve the accuracy of dynamic capacity expansion.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method for constructing, predicting, and apparatus for a dynamic quantitative prediction model. Background Technology

[0002] In some application scenarios, it is necessary to expand the user base based on a seed user group, such as finding similar users from a pre-set user pool as the target audience. Patent document 1 (CN 112508609 A) discloses a user expansion prediction method, which predicts user expansion based on user profile features and user operation behavior features performed on different data sources on the terminal, and can obtain relatively accurate prediction results. However, user expansion can also be carried out from other perspectives. Summary of the Invention

[0003] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0004] This invention provides a method for constructing a dynamic expansion model, the method comprising the following steps:

[0005] S100, Obtain first sample data and second sample data of users in the training user set; wherein, the training user set includes a positive sample training set and a negative sample training set, the positive sample training set is composed of the target population, and the negative sample training set is composed of people who do not belong to the target population; the first sample data includes a list of apps corresponding to the user, and the second sample data includes user information corresponding to the user; the user information includes at least user ID and user tag;

[0006] S200, Construct a Wide&Deep model; the Wide&Deep model is associated with an APP vector database and a user vector database. Each row of data in the APP vector database includes the corresponding APP ID and the corresponding APP vector, and each row of data in the user vector database includes the corresponding user information and the corresponding user vector.

[0007] S300, input the first sample data in the current batch of sample data into the Deep deep learning model in the Wide&Deep model, and input the second sample data into the Wide linear model in the Wide&Deep model to obtain the output result of the Wide&Deep model;

[0008] S400: Based on the output results and sample labels of all samples in the current batch, obtain the current loss function value and determine whether the current loss function value meets the preset model training termination condition. If it meets the condition, execute S600; otherwise, execute S500.

[0009] S500, based on the current loss function value, adjust the network parameters of the current initial dynamic expansion prediction model and the feature values ​​in the APP vector corresponding to the first sample data of the current batch to update the APP vector database, and use the sample data of the next batch as the sample data of the current batch, and execute S300.

[0010] S600 yields a trained Wide&Deep model, which serves as a dynamic expansion prediction model.

[0011] Another embodiment of the present invention provides a dynamic quantitative prediction method, comprising the following steps:

[0012] Obtain the data to be predicted, which includes corresponding user information and a list of apps;

[0013] The data to be predicted is input into a dynamic expansion prediction model for expanding the target population to obtain the corresponding prediction result; wherein, the dynamic expansion prediction model is constructed by the method described in any one of claims 1 to 6;

[0014] Based on the prediction results, determine whether the corresponding user belongs to the target group.

[0015] Another embodiment of the present invention provides a dynamic quantitative prediction device, comprising:

[0016] The data acquisition module is used to acquire the data to be predicted; the data to be predicted includes corresponding user information and an APP list.

[0017] The prediction module is used to input the data to be predicted into a dynamic expansion prediction model for expanding the target population, and to obtain the corresponding prediction result; wherein, the dynamic expansion prediction model is constructed by the method described in any one of claims 1 to 6;

[0018] The judgment module is used to determine whether a user belongs to the target group based on the prediction results.

[0019] The present invention has at least the following beneficial effects:

[0020] The dynamic scaling method provided in this invention expands the target population based on the user's installed APP, user ID, and user tags using a trained Wide&Deep model. Since the APP vector is dynamically adjusted based on the training results during the training process of the Wide&Deep model, the influence of features that have a significant impact on the training results is made more prominent, thereby improving the prediction accuracy. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating the method for constructing a dynamic quantitative prediction model provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Figure 1 A flowchart illustrating the method for constructing a dynamic quantitative prediction model provided in an embodiment of the present invention.

[0025] This invention provides a method for constructing a dynamic expansion model, such as... Figure 1 As shown, the method may include the following steps:

[0026] S100, Obtain first sample data and second sample data of users in the training user set; wherein, the training user set includes a positive sample training set and a negative sample training set, the positive sample training set is composed of the target population, and the negative sample training set is composed of people who do not belong to the target population; the first sample data includes a list of apps corresponding to the user, and the second sample data includes user information corresponding to the user; the user information includes at least user ID and user tag.

[0027] In this embodiment of the invention, the target audience can be a seed audience, provided by the demand side, such as people who prefer video content or games. The first sample data in the positive sample training set can be determined based on the user ID provided by the demand side. The user ID can be encrypted phone number, device ID, IMEI, IDFA, or other information that can uniquely associate a user. The user tags in the positive sample training set can be determined based on the user ID provided by the demand side. In this embodiment of the invention, user tags can be the user's own attribute information, such as the user's age, gender, city of residence, and mobile phone model.

[0028] In this embodiment of the invention, user tags can effectively reflect a user's actual needs, preferences, and the information they might be interested in. For example, users of different age groups often have significantly different interests; younger users tend to be more interested in fashion topics, while older users tend to be more interested in health-related topics. Furthermore, a user's education level can often reflect their income level to some extent; users with lower education levels tend to have lower incomes, while those with higher education levels tend to have higher incomes. Users with lower incomes may be more interested in product promotions, while users with higher incomes may be more interested in information about high-end products.

[0029] In this embodiment of the invention, the app list can effectively reflect a user's actual needs, preferences, and the information they might be interested in to a certain extent. For example, if a user's app list contains many game apps, it indicates that the user is more interested in game information; if a user's app list contains many shopping apps, it indicates that the user is more interested in shopping information, and so on.

[0030] In practical applications, each user ID, its corresponding list of apps, and user tags are stored in a pre-defined database. In this embodiment of the invention, the negative sample training set can be obtained by randomly selecting a set number of data points from the pre-defined database.

[0031] In this embodiment of the invention, the ratio of positive samples to negative samples can be determined based on the actual situation, for example, it can be set to 3:5, etc.

[0032] S200, Construct a Wide&Deep model; the Wide&Deep model is associated with an APP vector database and a user vector database. Each row of data in the APP vector database includes the corresponding APP ID and the corresponding APP vector, and each row of data in the user vector database includes the corresponding user information and the corresponding user vector.

[0033] In this embodiment of the invention, a key feature of the Wide&Deep model is its simultaneous ability to remember and generalize. The "memory" capability refers to the ability to infer data with strong correlations to historical data. This invention uses neural networks to capture high-order features that distinguish target and non-target users, thereby improving generalization ability. Generalization ability refers to the ability to predict new data that has almost never appeared before, based on historical data and the transfer of data correlations. In this embodiment, due to the high feature dimensionality, important features may become less significant during model training. Therefore, the "memory" capability of the wide side is crucial for remembering important features. For example, by analyzing the types of apps a user has previously installed and utilizing an existing user database, it can determine what other app types the user might be interested in, and then recommend these to users with dynamic expansion needs. These "other" app types may be types that have never appeared before.

[0034] In this embodiment of the invention, APP vectors characterize APP sequence information, reflecting data related to user activity, including APP package name, name, installation package size, number of times the user uses the APP within different periods, login duration, downloads or registrations, etc. This data can be obtained based on an APP vector generation model, and the specific generation method can be existing technology. For example, a graph network can be constructed based on the relationships between sample APPs, and then input into an initial graph network model for training, resulting in a trained graph network model as the APP vector generation model. In this embodiment of the invention, the relationships between APPs can be determined from the perspectives of package name, name, whether they exist on the same device, whether they belong to the same category, and whether they are adjacent in the same APP list.

[0035] In this embodiment of the invention, user vectors are used to characterize a user's preference information for apps. For example, they can be used to predict which apps a user will install in the future based on the user's current app list. User vectors can be obtained based on a user vector generation model, and the specific generation method can be existing technology, such as based on attention mechanisms.

[0036] S300, input the first sample data from the current batch of sample data into the Deep deep learning model in the Wide&Deep model, and input the second sample data into the Wide linear model in the Wide&Deep model to obtain the output result of the Wide&Deep model.

[0037] In this embodiment of the invention, the second sample data is input into the Wide linear model to comprehensively analyze, using this data and the memory capacity of the Wide model, which apps (or types of apps) users are typically interested in. The first sample data is input into the Deep learning model to comprehensively analyze, using this data and the generalization ability of the Deep learning model, which apps (or types of apps) users might also be interested in. In other words, this specification considers that the aforementioned list of apps possesses the generalization conditions necessary for the Deep learning model to perform generalization analysis.

[0038] Furthermore, in S300, the Deep linear model is used to perform the following operations:

[0039] S301, Based on the input first sample data, obtain the corresponding APP vector from the APP vector database.

[0040] S302, perform average pooling on the obtained APP vector to obtain the first feature.

[0041] In one embodiment of the present invention, for example, the obtained APP vector is an N*M dimensional vector. For the j-th feature, the feature after average pooling is (F 1j +F 2j +…+F rj +…+F nj ) / n, F rj Let be the j-th feature in the r-th row, where j takes values ​​from 1 to M and r takes values ​​from 1 to N.

[0042] S303 enables multiple neural networks to extract features from the first feature to obtain the second feature.

[0043] Those skilled in the art will know that the specific methods for obtaining the first and second features can be adopted using existing methods.

[0044] Furthermore, in S300, the Wide linear model is used to perform the following operations:

[0045] S310, Based on the input second sample data, obtain the corresponding user vector from the user vector database;

[0046] S311, perform feature extraction on the obtained user vector to obtain the third feature.

[0047] Those skilled in the art will know that the specific method for obtaining the third feature can be any existing method.

[0048] Furthermore, in S300, the output is obtained by fusing the second and third features. The specific fusion method can employ existing techniques, such as weighted summation.

[0049] S400: Based on the output results and sample labels of all samples in the current batch, obtain the current loss function value and determine whether the current loss function value meets the preset model training termination condition. If it does, execute S600; otherwise, execute S500.

[0050] In this embodiment of the invention, the loss function is not specifically limited. The preset model training termination condition can be set based on actual needs. In a non-limiting illustrative embodiment, the model training termination condition may include the loss function being less than a preset first loss threshold or the loss function being less than a preset second loss threshold, and remaining unchanged, wherein the first loss threshold is less than the second loss threshold.

[0051] S500, based on the current loss function value, adjust the network parameters of the current initial dynamic expansion prediction model and the feature values ​​in the APP vector corresponding to the first sample data of the current batch to update the APP vector database, and use the sample data of the next batch as the sample data of the current batch, and execute S300.

[0052] Because the APP vector is dynamically adjusted based on the training results during the training of the Wide&Deep model, the influence of features that have a significant impact on the training results is made more prominent, thereby improving the prediction accuracy.

[0053] S600 yields a trained Wide&Deep model, which serves as a dynamic expansion prediction model.

[0054] Furthermore, in another embodiment of the present invention, S302 may specifically include:

[0055] S3021, the obtained APP vectors are divided into m vector categories, where m is the number of APP categories determined based on the APP list corresponding to the positive sample training set.

[0056] In one embodiment of the present invention, the APP category corresponding to the positive sample training set can be provided by the provider of the target population.

[0057] In another embodiment of the present invention, the APP categories corresponding to the positive sample training set can be obtained by clustering the APPs corresponding to the positive sample training set. Existing clustering methods can be used for this specific clustering.

[0058] S3022, perform average pooling on the APP vector corresponding to each vector category to obtain the corresponding average pooling feature.

[0059] S3023, the m average pooling features are concatenated to obtain the concatenated feature, which is then used as the first feature.

[0060] The technical effect of S3021 to S3023 is that by performing average pooling on the APP vectors of different categories, the feature distinction becomes more obvious, and the features that have a greater impact on the target audience are further affected, thus making the prediction results more accurate.

[0061] Another embodiment of the present invention provides a dynamic quantitative prediction method, comprising the following steps:

[0062] Obtain the data to be predicted, which includes the corresponding user information and APP list.

[0063] In this embodiment of the invention, the data to be predicted can be obtained based on a preset user database. The preset user database stores historically accumulated user information and a corresponding list of apps.

[0064] The data to be predicted is input into a dynamic expansion prediction model for expanding the target population to obtain the corresponding prediction results; wherein, the dynamic expansion prediction model is constructed using the method described in the foregoing embodiment.

[0065] Based on the prediction results, determine whether the corresponding user belongs to the target group.

[0066] If the prediction result is greater than the set threshold, it means that the user belongs to the target group; otherwise, it does not.

[0067] Another embodiment of the present invention also provides a dynamic quantitative prediction device, comprising:

[0068] The data acquisition module is used to acquire the data to be predicted; the data to be predicted includes corresponding user information and an APP list.

[0069] The prediction module is used to input the data to be predicted into a dynamic expansion prediction model for expanding the target population, and to obtain the corresponding prediction results; wherein, the dynamic expansion prediction model is constructed by the aforementioned method.

[0070] The judgment module is used to determine whether a user belongs to the target group based on the prediction results.

[0071] If the prediction result is greater than the set threshold, it means that the user belongs to the target group; otherwise, it does not.

[0072] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a method in the method embodiments, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the method provided in the above embodiments.

[0073] Embodiments of the present invention also provide an electronic device, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0074] Embodiments of the present invention also provide a computer program product including program code, which, when the program product is run on an electronic device, causes the electronic device to perform the steps of the methods described above in various exemplary embodiments of the present invention.

[0075] While specific embodiments of the invention have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. Those skilled in the art should also understand that various modifications can be made to the embodiments without departing from the scope and spirit of the invention. The scope of this invention is defined by the appended claims.

Claims

1. A method for constructing a dynamic expansion model, characterized in that, The method includes the following steps: S100, Obtain first sample data and second sample data of users in the training user set; wherein, the training user set includes a positive sample training set and a negative sample training set, the positive sample training set is composed of the target population, and the negative sample training set is composed of people who do not belong to the target population; the first sample data includes a list of apps corresponding to the user, and the second sample data includes user information corresponding to the user; the user information includes at least user ID and user tag; S200, Construct a Wide & Deep model; the Wide & Deep model is associated with an APP vector database and a user vector database. Each row of data in the APP vector database includes the corresponding APP ID and the corresponding APP vector, and each row of data in the user vector database includes the corresponding user information and the corresponding user vector. S300, input the first sample data in the current batch of sample data into the Deep deep learning model in the Wide & Deep model, and input the second sample data into the Wide linear model in the Wide & Deep model to obtain the output result of the Wide & Deep model; S400: Based on the output results and sample labels of all samples in the current batch, obtain the current loss function value and determine whether the current loss function value meets the preset model training termination condition. If it meets the condition, execute S600; otherwise, execute S500. S500, based on the current loss function value, adjust the network parameters of the current initial dynamic expansion prediction model and the feature values ​​in the APP vector corresponding to the first sample data of the current batch to update the APP vector database, and use the sample data of the next batch as the sample data of the current batch, and execute S300. S600, obtain the trained Wide & Deep model, as the dynamic expansion prediction model; In S300, the Deep learning model is used to perform the following operations: S301, Based on the input first sample data, obtain the corresponding APP vector from the APP vector database; S302, Perform average pooling on the obtained APP vector to obtain the first feature; S303 enables multiple neural networks to extract features from the first feature to obtain the second feature; In S300, the Wide linear model is used to perform the following operations: S310, Based on the input second sample data, obtain the corresponding user vector from the user vector database; S311, Perform feature extraction on the obtained user vector to obtain the third feature; In S300, the output is obtained by fusing the second and third features; S302 specifically includes: S3021, the obtained APP vectors are divided into m vector categories, where m is the number of APP categories determined based on the APP list corresponding to the positive sample training set; S3022, Perform average pooling on the APP vector corresponding to each vector category to obtain the corresponding average pooling feature; S3023, the m average pooling features are concatenated to obtain the concatenated feature, which is then used as the first feature.

2. The method according to claim 1, characterized in that, The app categories corresponding to the positive sample training set are provided by the target audience provider, or they are obtained by clustering the apps corresponding to the positive sample training set.

3. The method according to claim 2, characterized in that, The target audience is the seed audience provided by the demand side.

4. A dynamic quantitative prediction method, characterized in that, Includes the following steps: Obtain the data to be predicted, which includes corresponding user information and a list of apps; The data to be predicted is input into a dynamic expansion prediction model for expanding the target population to obtain the corresponding prediction result; wherein, the dynamic expansion prediction model is constructed by the method described in any one of claims 1 to 3; Based on the prediction results, determine whether the corresponding user belongs to the target group.

5. A dynamic quantitative prediction device, characterized in that, include: The data acquisition module is used to acquire the data to be predicted. The data to be predicted includes the corresponding user information and a list of apps; The prediction module is used to input the data to be predicted into a dynamic expansion prediction model for expanding the target population, and to obtain the corresponding prediction result; wherein, the dynamic expansion prediction model is constructed by the method described in any one of claims 1 to 3; The judgment module is used to determine whether a user belongs to the target group based on the prediction results.

Images