Method, device, electronic device and storage medium for determining account activity information
By extracting and mapping the historical data of the user account, and using a multi-layer perceptron model to fuse active information in adjacent time periods, the problem of inconsistent LTV prediction in the prior art is solved and the prediction accuracy is improved.
Patent Information
- Application Number
- CN202210344139.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-03-31
AI Technical Summary
When predicting the life cycle value (LTV) of user accounts, the prior art ignores the relationship between different time periods, resulting in short-term LTV being greater than long-term LTV, which is unreasonable.
By extracting and mapping the historical data of the user account, using a multi-layer perceptron model combined with monotonic gain information, and integrating active information from adjacent time periods, we can determine the active information of the user account in each time period, and improve prediction accuracy .
It effectively solves the problem of inconsistent prediction of short-term and long-term active value, improves the accuracy of account active information, and ensures that the prediction results are closer to the actual situation.
Smart Images

Figure CN114819288B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of Internet technology, and in particular to a method, device, electronic device and storage medium for determining account activity information. Background Art
[0002] In the field of Internet technology, a lot of data is long-tailed, which can also be called uneven distribution. For example, the Life Time Value (LTV) of a user account on an application can reflect the activity of the user account in the application. However, the LTV of each user account is unevenly distributed, that is, the active behaviors of a small number of user accounts account for the majority of the total active behaviors, while the active behaviors of most user accounts account for a small part of the total active behaviors. Application providers usually make corresponding decisions based on the LTV of user accounts, such as placing advertisements, providing incentives, and issuing dividends. Therefore, how to accurately determine the LTV of user accounts is a problem that needs to be solved.
[0003] In related technologies, models are usually built for each time period, such as three-month LTV, six-month LTV, one-year LTV, and ten-year LTV, to estimate the future LTV of each user account. However, this approach ignores the correlation between LTVs in different time periods, which can lead to the prediction that the short-term LTV of the same user account is greater than the long-term LTV, such as three-month LTV is greater than one-year LTV, which is obviously unreasonable. Summary of the invention
[0004] The present disclosure provides a method, device, electronic device and storage medium for determining account activity information, which can determine the account activity information of a user account in each time period based on the sequential relationship between two adjacent time periods, thereby improving the accuracy of the predicted account activity information. The technical solution of the present disclosure is as follows:
[0005] According to one aspect of an embodiment of the present disclosure, a method for determining account activity information is provided, including:
[0006] Extracting features from account data of the determined user account within a historical period to obtain account features, wherein the account data includes behavior data and attribute data;
[0007] Mapping the account features to obtain a plurality of first activity information of a plurality of time periods, wherein the first activity information is a predicted activity of the user account in a corresponding time period, a later time period in the plurality of time periods includes a previous time period, and an end date of the plurality of time periods is later than an end date of the historical time period;
[0008] Based on first activity information in a first time period and first activity information in a second time period, account activity information of the user account in the second time period is obtained, the first time period and the second time period are adjacent in order in the multiple time periods, and the second time period includes the first time period, and the account activity information is used to indicate the activity of the user account in the second time period.
[0009] In some embodiments, obtaining the account activity information of the user account in the second time period based on the first activity information in the first time period and the first activity information in the second time period includes:
[0010] Based on a multi-layer perceptron, mapping the first activity information of the first time period to obtain monotonic gain information of the first time period, where the monotonic gain information is used to indicate the influence of the active behavior of the user account in the first time period on the second time period;
[0011] The first activity information in the second time period and the monotonic gain information in the first time period are fused to obtain the account activity information of the user account in the second time period.
[0012] By fusing the first active information in the first time period with the first active information in the second time period, the first active information in the first time period can gain an advantage in the second time period, thereby having a positive impact on the active information of the user account in the second time period and improving the accuracy of the prediction.
[0013] In some embodiments, extracting features from account data of the user account to be determined within a historical time period to obtain account features includes:
[0014] Embedding the account data of the determined user account within the historical time period to obtain an account embedding feature;
[0015] Perform feature crossover on the account embedding feature to obtain the account feature.
[0016] By embedding the sample account data, the discrete data in the sample account data can be converted into dense embedding, making the discrete data continuous for subsequent processing, and then performing feature cross-feature on the account embedding features to improve the accuracy of subsequent predictions.
[0017] In some embodiments, mapping the account features to obtain a plurality of first activity information in a plurality of time periods includes:
[0018] For any time period, the account features are mapped based on the multi-layer perceptron corresponding to the time period to obtain the first active information of the time period.
[0019] By setting up multi-layer perceptrons for different sample time periods, we can model each uneven distribution. That is, the data in each sample time period is unevenly distributed. Based on the corresponding multi-layer perceptron, we can predict the activity of the user account in the corresponding time period and obtain the activity of the user account in multiple sample time periods, thereby improving the prediction efficiency.
[0020] In some embodiments, the method for determining account activity information is implemented based on an account activity model, which is used to predict the account activity information of the user account in multiple future time periods based on the account data of the user account in a period of time.
[0021] By implementing the method for determining account activity information based on an account activity model, the efficiency of determining account activity information is improved.
[0022] In some embodiments, the method for determining account activity information further includes:
[0023] Based on the account activity model of the i-th iteration, feature extraction is performed on the sample account data of the sample account in the first sample time period to obtain sample account features, where the sample account data includes behavior data and attribute data, and i is a positive integer;
[0024] Based on the account activity model of the i-th iteration, the sample account features are mapped to obtain a plurality of second activity information of a plurality of sample time periods, wherein the second activity information is the predicted activity of the sample account in the corresponding sample time period, a sample time period in a later order among the plurality of sample time periods includes a sample time period in a previous order, and an end date of the plurality of sample time periods is later than an end date of the first sample time period;
[0025] Based on the account activity model of the i-th iteration, the multiple second activity information are merged to obtain multiple third activity information of the sample account in the multiple sample time periods, where the third activity information is merged based on the second activity information of two adjacent sample time periods;
[0026] The account activity model of the i-th iteration is trained based on the multiple third activity information and the multiple fourth activity information, wherein the multiple fourth activity information is used to represent the actual activity of the sample account in the multiple sample time periods.
[0027] By mapping the account features of the user account in the first sample time period based on the account activity model, it is possible to predict multiple active information of the user account in multiple sample time periods, and then determine the account active information of the user account in each sample time period based on the sequential relationship between two adjacent sample time periods, thereby improving the accuracy of the predicted third active information.
[0028] In some embodiments, the account activity model based on the i-th iteration performs feature extraction on the sample account data of the sample account in the first sample time period to obtain sample account features, including:
[0029] Based on the first network in the account activity model of the i-th iteration, embedding processing is performed on the sample account data of the sample account in the first sample time period to obtain sample embedding features, and the first network is used for feature embedding;
[0030] Based on the second network in the account activity model of the i-th iteration, feature crossover is performed on the sample embedding features to obtain the sample account features, and the second network is used for feature crossover.
[0031] By embedding the sample account data based on the first network, the discrete data in the sample account data can be converted into Dense Embedding, making the discrete data continuous for subsequent processing. Then, based on the second network, bit-wise feature crossover is performed on the data output by the first network to improve the accuracy of subsequent predictions.
[0032] In some embodiments, the account activity model based on the i-th iteration is used to map the sample account features to obtain multiple second activity information in multiple sample time periods, including:
[0033] For any sample time period, based on the third network corresponding to the sample time period in the account activity model of the i-th iteration, the account features are mapped to obtain the second activity information of the sample time period. The third network is used for feature mapping and includes a multi-layer perceptron.
[0034] By setting up the third network for different sample time periods, modeling of each uneven distribution is achieved, that is, the data in each sample time period is unevenly distributed. Based on the corresponding third network, the activity of the sample account in the corresponding sample time period can be predicted, which improves the prediction range and thus improves the prediction efficiency.
[0035] In some embodiments, the merging of the plurality of second activity information to obtain the third activity information of the sample account in the plurality of sample time periods includes:
[0036] The second active information of the second sample time period and the second active information of the third sample time period are merged to obtain the third active information of the sample account in the third sample time period, the second sample time period is adjacent to the third sample time period in the order of the multiple sample time periods, and the third sample time period includes the second sample time period.
[0037] By fusing the second active information of the second sample time period with the second active information of the third sample time period, the second active information of the second sample time period can gain the third sample time period, thereby having a positive impact on the active information of the sample account in the third sample time period and improving the accuracy of the prediction.
[0038] In some embodiments, the second activity information of the second sample time period and the second activity information of the third sample time period are merged to obtain the third activity information of the sample account in the third sample time period, including:
[0039] Based on the fourth network corresponding to the second sample time period in the account activity model of the i-th iteration, the second activity information of the second sample time period is mapped to obtain the monotonic gain information of the second sample time period, where the monotonic gain information is used to indicate the influence of the activity of the sample account in the second sample time period on the third sample time period;
[0040] Based on the fifth network corresponding to the third sample time period in the account activity model of the i-th iteration, the second activity information of the third sample time period and the monotonic gain information of the second time period are fused to obtain the third activity information of the sample account in the third sample time period.
[0041] By mapping the second active information of the second sample time period through the fourth network corresponding to the second sample time period, the active information of the previous sample time period can be converted into the monotonic gain information of the subsequent sample time period, and then the monotonic gain information is fused with the active information of the subsequent sample time period based on the fifth network, so as to improve the accuracy of the predicted third active information.
[0042] In some embodiments, the method for determining account activity information further includes:
[0043] For any sample time period, based on the fifth network corresponding to the sample time period in the account activity model of the i-th iteration, the activity values of multiple sample accounts in the sample time period are bucketed to obtain multiple active buckets, and the number of sample accounts in each active bucket is within the target number range. The fifth network is used to bucket the user accounts;
[0044] Get the mean of the active value range corresponding to each active bucket;
[0045] Based on the average values of the activity buckets to which the multiple sample accounts belong, fourth activity information of the multiple sample accounts within the sample time period is determined.
[0046] By bucketing multiple sample accounts, the labels of the sample accounts are converted from the original active values to the mean values of the active buckets to which they belong, thereby converting the regression problem into a more easily solved classification problem. That is, the sample accounts are classified into different active buckets, which can not only alleviate the problem of uneven distribution of sample accounts, but also improve the learning efficiency of the model.
[0047] In some embodiments, the training of the account activity model of the i-th iteration based on the plurality of third activity information and the plurality of fourth activity information includes:
[0048] Determine, based on the plurality of third activity information and the plurality of fourth activity information, a plurality of first losses in the plurality of sample time periods, the first losses being used to represent a difference between a predicted activity and an actual activity;
[0049] Based on the plurality of third activity information, determining at least one second loss, where the second loss is used to calibrate a sequential relationship between two pieces of third activity information in two adjacent sample time periods;
[0050] Based on the multiple first losses and the at least one second loss, adjust the parameters of the account activity model of the i-th iteration.
[0051] By calculating the first loss, the parameters of the model can be adjusted based on the difference between the prediction results and the label results to achieve supervised training and improve the model training efficiency and the accuracy of model classification. By calculating the second loss, the sequential relationship between the third active information corresponding to two adjacent sample time periods can be measured. Based on the second loss, the incorrect sequential relationship can be penalized to correct the error, avoid unreasonable situations, and improve the accuracy of the prediction results output by the model.
[0052] According to another aspect of an embodiment of the present disclosure, a device for determining account activity information is provided, including:
[0053] A first extraction unit is configured to extract features from account data of a user account to be determined within a historical time period to obtain account features, wherein the account data includes behavior data and attribute data;
[0054] a first mapping unit configured to map the account features to obtain a plurality of first activity information of a plurality of time periods, wherein the first activity information is a predicted activity of the user account in a corresponding time period, a time period in a later order among the plurality of time periods includes a time period in a previous order, and an end date of the plurality of time periods is later than an end date of the historical time period;
[0055] The first fusion unit is configured to obtain account activity information of the user account in the second time period based on first activity information in a first time period and first activity information in a second time period, the first time period and the second time period are adjacent in order in the multiple time periods, and the second time period includes the first time period, and the account activity information is used to indicate the activity of the user account in the second time period.
[0056] In some embodiments, the first fusion unit is configured to map the first activity information of the first time period based on a multi-layer perceptron to obtain monotonic gain information of the first time period, where the monotonic gain information is used to indicate the impact of the active behavior of the user account in the first time period on the second time period; and fuse the first activity information of the second time period with the monotonic gain information of the first time period to obtain the account activity information of the user account in the second time period.
[0057] In some embodiments, the first extraction unit is configured to perform embedding processing on the account data of the determined user account within the historical time period to obtain account embedding features; and perform feature cross-feature processing on the account embedding features to obtain the account features.
[0058] In some embodiments, the first mapping unit is configured to map the account features for any time period based on the multi-layer perceptron corresponding to the time period to obtain the first activity information for the time period.
[0059] In some embodiments, the device for determining account activity information executes steps based on an account activity model, and the account activity model is used to predict the account activity information of the user account in multiple time periods in the future based on the account data of the user account in a period of time.
[0060] In some embodiments, the means for determining account activity information further comprises:
[0061] The second extraction unit is configured to extract features from sample account data of the sample account in the first sample time period based on the account activity model of the i-th iteration to obtain sample account features, wherein the sample account data includes behavior data and attribute data, and i is a positive integer;
[0062] a second mapping unit, configured to map the sample account features based on the account activity model of the i-th iteration to obtain a plurality of second activity information of a plurality of sample time periods, wherein the second activity information is the predicted activity of the sample account in the corresponding sample time period, a sample time period in a later order among the plurality of sample time periods includes a sample time period in a previous order, and an end date of the plurality of sample time periods is later than an end date of the first sample time period;
[0063] A second fusion unit is configured to fuse the plurality of second activity information based on the account activity model of the i-th iteration to obtain a plurality of third activity information of the sample account in the plurality of sample time periods, wherein the third activity information is obtained by fusing the second activity information of two adjacent sample time periods;
[0064] The model training unit is configured to train the account activity model of the i-th iteration based on the multiple third activity information and the multiple fourth activity information, wherein the multiple fourth activity information is used to represent the actual activity of the sample account in the multiple sample time periods.
[0065] In some embodiments, the second extraction unit is configured to, based on the first network in the account activity model of the i-th iteration, embed the sample account data of the sample account within the first sample time period to obtain sample embedding features, and the first network is used for feature embedding; based on the second network in the account activity model of the i-th iteration, perform feature crossover on the sample embedding features to obtain the sample account features, and the second network is used for feature crossover.
[0066] In some embodiments, the second mapping unit is configured to map the account features for any sample time period based on the third network corresponding to the sample time period in the account activity model of the i-th iteration to obtain second activity information for the sample time period, and the third network is used for feature mapping, including a multi-layer perceptron.
[0067] In some embodiments, the second fusion unit is configured to fuse the second activity information of the second sample time period and the second activity information of the third sample time period to obtain the third activity information of the sample account in the third sample time period, the second sample time period is adjacent to the third sample time period in the order of the multiple sample time periods, and the third sample time period includes the second sample time period.
[0068] In some embodiments, the second fusion unit is configured to map the second activity information of the second sample time period based on the fourth network corresponding to the second sample time period in the account activity model of the i-th iteration, and obtain the monotonic gain information of the second sample time period, wherein the monotonic gain information is used to indicate the influence of the active behavior of the sample account in the second sample time period on the third sample time period; and to fuse the second activity information of the third sample time period and the monotonic gain information of the second time period based on the fifth network corresponding to the third sample time period in the account activity model of the i-th iteration, and obtain the third activity information of the sample account in the third sample time period.
[0069] In some embodiments, the means for determining account activity information further comprises:
[0070] A bucketing unit is configured to bucket the activity values of multiple sample accounts in the sample time period based on the fifth network corresponding to the sample time period in the account activity model of the i-th iteration, to obtain multiple active buckets, wherein the number of sample accounts in each active bucket is within a target number range, and the fifth network is used to bucket the user accounts;
[0071] The bucketing unit is further configured to obtain a mean value of the activity value range corresponding to each active bucket;
[0072] The bucketing unit is further configured to determine fourth activity information of the multiple sample accounts within the sample time period based on the average of the active buckets to which the multiple sample accounts belong.
[0073] In some embodiments, the model training unit is configured to determine multiple first losses for the multiple sample time periods based on the multiple third activity information and the multiple fourth activity information, the first losses being used to represent the difference between the predicted activity and the actual activity; determine at least one second loss based on the multiple third activity information, the second loss being used to calibrate the sequential relationship between two third activity information in two adjacent sample time periods; and adjust the parameters of the account activity model of the i-th iteration based on the multiple first losses and the at least one second loss.
[0074] According to another aspect of an embodiment of the present disclosure, an electronic device is provided, the electronic device comprising:
[0075] one or more processors;
[0076] A memory for storing program codes executable by the processor;
[0077] The processor is configured to execute the program code to implement the above method for determining account activity information.
[0078] According to another aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When a program code in the computer-readable storage medium is executed by a processor of an electronic device, the electronic device can execute the above-mentioned method for determining account activity information.
[0079] According to another aspect of an embodiment of the present disclosure, a computer program product is provided, including a computer program, which implements the above-mentioned method for determining account activity information when executed by a processor.
[0080] The solution for determining account activity information provided by the embodiments of the present disclosure can predict multiple active information of the user account in multiple time periods in the future by mapping the account features of the user account in a first time period, and then determine the account activity information of the user account in each time period based on the sequential relationship between two adjacent time periods, thereby improving the accuracy of the predicted account activity information.
[0081] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS
[0082] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.
[0083] Figure 1 The present invention is a schematic diagram of an implementation environment of a method for determining account activity information according to an exemplary embodiment.
[0084] Figure 2 The present invention is a flowchart of a method for determining account activity information according to an exemplary embodiment.
[0085] Figure 3 The present invention is a flowchart of another method for determining account activity information according to an exemplary embodiment.
[0086] Figure 4 The figure is a schematic diagram of the structure of an account activity model according to an exemplary embodiment.
[0087] Figure 5 The present invention is a block diagram showing a device for determining account activity information according to an exemplary embodiment.
[0088] Figure 6 The invention is a block diagram showing another apparatus for determining account activity information according to an exemplary embodiment.
[0089] Figure 7It is a block diagram of a terminal according to an exemplary embodiment.
[0090] Figure 8 It is a block diagram of a server according to an exemplary embodiment. DETAILED DESCRIPTION
[0091] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.
[0092] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.
[0093] The account data and other information involved in this disclosure are information authorized by the user or fully authorized by all parties.
[0094] The following are the terms used in the present disclosure.
[0095] Multilayer Perceptron (MLP) is a feedforward artificial neural network model that maps multiple input data sets to a single output data set.
[0096] Feature crossover is a method of synthesizing features, which can perform good nonlinear feature fitting on multi-dimensional feature data sets. Feature crossover includes vector-wise feature crossover and element-wise feature crossover. Among them, element-wise means that the elements in the embedded vector are represented by the term bit.
[0097] One-hot encoding, also known as one-bit effective encoding, mainly uses a bit state register to encode multiple states. Each state has its own independent register bit, and only one bit is valid at any time. One-hot encoding uses 0 and 1 to represent some parameters and uses a multi-bit state register to encode multiple states.
[0098] Rectified Linear Unit (ReLU), also known as the rectified linear unit, is an activation function commonly used in artificial neural networks, usually referring to nonlinear functions represented by ramp functions and their variants.
[0099] The scheme for determining account activity information provided by the present disclosure is executed by an electronic device, and the electronic device can be provided as a terminal or a server. When the electronic device is provided as a terminal, the scheme is implemented by the terminal; when it is provided as a server, the scheme is implemented by the server, and the server can receive the account data authorized by the terminal, and the server determines the account activity information of the user account; or the server and the terminal interact to implement the scheme, and the terminal sends a request carrying the user account to the server, and the server determines the multiple account activity information of the user account in multiple time periods in the future, and then feeds back the multiple account activity information to the terminal, and the terminal outputs the multiple account activity information.
[0100] Figure 1 FIG. 1 is a schematic diagram of an implementation environment of a method for determining account activity information according to an exemplary embodiment. Figure 1 , the implementation environment specifically includes: a terminal 101 and a server 102.
[0101] The terminal 101 is at least one of a smart phone, a smart watch, a desktop computer, a laptop, an MP3 player, an MP4 player, and a laptop computer. An application is installed on the terminal 101, and a user can log in to the application through the terminal 101 to obtain the service provided by the application. The terminal 101 can be connected to the server 102 through a wireless network or a wired network, and then send a request carrying a user account to be determined to the server 102. Among them, the terminal 101 generally refers to one of a plurality of terminals, and the present embodiment is illustrated by the terminal 101. Those skilled in the art know that the number of the above-mentioned terminals can be more or less. For example, the above-mentioned terminals are several, or the above-mentioned terminals are dozens or hundreds, or more. The embodiment of the present disclosure does not limit the number of terminals and device types.
[0102] The server 102 is at least one of a single server, multiple servers, a cloud computing platform, and a virtualization center. The server 102 can receive a request sent by the terminal 101, and the request carries a user account to be determined. The server 102 determines the account activity information of the user account in multiple time periods in the future based on the solution for determining account activity information provided by the present disclosure. In some embodiments, the number of the above servers can be more or less, and the embodiments of the present disclosure are not limited to this. Of course, the server 102 can also include other functional servers, such as database servers, query servers, and statistical servers, etc., in order to provide more comprehensive and diversified services.
[0103] Figure 2 is a flow chart showing a method for determining account activity information according to an exemplary embodiment. Figure 2 As shown, the method for determining account activity information is performed by an electronic device and includes the following steps:
[0104] In step S201, feature extraction is performed on the account data of the user account to be determined within a historical time period to obtain account features.
[0105] In the embodiment of the present disclosure, the account data includes the behavior data and attribute data of the user account. Among them, the behavior data is related to the application scenario of the method. For example, when the method is applied to multimedia applications, the behavior data includes the viewing behavior, collection behavior and like behavior of the user account; when the method is applied to game applications, the behavior data includes the login behavior, game behavior and consumption behavior of the user account. Attribute data includes the user portrait of the user account, such as gender, age, occupation and the device used to log in to the user account. The historical time period is the past 3 days, the past 7 days or the past 10 days, etc., and the embodiment of the present disclosure does not limit this. The account data is obtained by the electronic device from a database query based on the user account, and the database is used to store the account data of multiple user accounts; or, the account data is carried by the request received by the electronic device, and the request carries the user account and the account data of the user account in the historical time period.
[0106] In some embodiments, the electronic device can extract features by embedding the account data and then performing feature cross-linking. The embodiments of the present disclosure do not limit the feature processing method.
[0107] In step S202, the account characteristics are mapped to obtain a plurality of first activity information in a plurality of time periods.
[0108] In the embodiment of the present disclosure, by mapping the account features according to the time periods, the first active information of each time period can be obtained, and the first active information is the predicted activity of the user account in the corresponding time period. The multiple first active information corresponds to the multiple time periods one by one, that is, different time periods correspond to different first active information. The time period in the multiple time periods in the latter order includes the time period in the former order, and the end date of the multiple time periods is later than the end date of the above-mentioned historical time period. The start date of the multiple time periods is the same, and the time length is different. In some embodiments, the start date of the multiple time periods is the same, and the start date is within the historical time period, that is, the start date of the multiple time periods is the same as the start date of the historical time period, or the start date of the multiple time periods is the same as the end date of the historical time period, or the start date of the multiple time periods is between the start date and the end date of the historical time period. In some embodiments, the start date of the multiple time periods is the same, and the start date of the multiple time periods is later than the end date of the historical time period, and the embodiment of the present disclosure does not limit this.
[0109] For example, taking the case where the start dates of multiple time periods are the same as the start date of a historical time period, assuming that the current time is January 8, and the historical time period is the past 7 days, the start date of the historical time period is January 1. The start dates of the multiple time periods are the same as the historical time period, that is, the start dates of the multiple time periods are all January 1, and the lengths of the multiple time periods are 30 days, 60 days, 90 days, and 180 days, respectively, that is, the end dates of the multiple time periods are all later than the end date of the historical time period.
[0110] In step S203, the first activity information in the first time period and the first activity information in the second time period are merged to obtain the account activity information of the user account in the second time period.
[0111] In the embodiment of the present disclosure, the first time period is any time period among the multiple time periods except the time period at the end of the sequence, the first time period is adjacent to the second time period in the multiple time periods, and the second time period includes the first time period. Since the time period in the multiple time periods in the later sequence includes the time period in the earlier sequence, the first active information of the time period in the earlier sequence will affect the time period in the later sequence. Therefore, the first active information of the first time period is merged with the first active information of the second time period, and the account active information obtained by the fusion is used as the final active information of the user account in the second time period, and the account active information is used to indicate the activity of the user account in the second time period.
[0112] The solution for determining account activity information provided by the embodiments of the present disclosure can predict multiple active information of the user account in multiple time periods in the future by mapping the account features of the user account in historical time periods, and then determine the account activity information of the user account in each time period based on the sequential relationship between two adjacent time periods, thereby improving the accuracy of the predicted account activity information.
[0113] Above Figure 2 The following is a basic flow of a solution for determining account activity information provided by the present disclosure. The solution provided by the present disclosure is further described below based on an application scenario. In this application scenario, prediction of the number of active days of a user in multiple time periods in the future is used as an example for explanation. Figure 3 FIG. 1 is a flowchart of another method for determining account activity information according to an exemplary embodiment. Figure 3 , the method comprising:
[0114] In the disclosed embodiment, the scheme for determining the account activity information is implemented based on an account activity model, and the account activity model is used to predict the account activity information of the user account in multiple time periods in the future based on the account data of the user account in a period of time. The following takes the i-th iteration as an example, and introduces the model training process based on steps S301 to S304, where i is a positive integer. After the model training is completed, the process of implementing the scheme for determining the account activity information based on the account activity model obtained through training is described based on steps 305 to S307.
[0115] In step S301, the server extracts features of the sample account data of the sample account in the first sample time period based on the account activity model of the i-th iteration to obtain sample account features.
[0116] In an embodiment of the present disclosure, the server can obtain the behavior data and attribute data of multiple sample accounts within a first sample time period. Among them, the sample account is a user account, and the behavior data includes at least one of the user account's login behavior, viewing behavior, publishing behavior, collection behavior, and like behavior. The attribute data includes a user portrait of the user account, such as at least one of gender, age, and device information. Among them, at least one refers to one or more. The length of the first sample time period is 3 days, 7 days, or 10 days, etc., and the embodiment of the present disclosure does not limit this. The sample account data of the sample account in the first time period is obtained by the server from a database query based on the sample account, and the database is used to store account data of multiple user accounts.
[0117] For example, taking the time length of the first sample time period as 7 days, the server obtains the behavioral data and user portrait of the sample account within these 7 days. Among them, the behavioral data is expressed in the form of a sequence, and the sequence includes 7 elements, one element corresponding to one day. For example, the "active time" of the sample account within these 7 days is expressed as a sequence of length 7: [34, 23, 56, 12, 14, 67, 87], which represents the number of minutes that the sample account is active every day. For another example, the "number of likes" of the sample account within these 7 days is expressed as another sequence of length 7: [12, 11, 23, 6, 29, 35], which represents the number of likes of the sample account every day. Similarly, the browsing behavior, collection behavior, and consumption behavior of the sample account are also expressed in the form of a sequence, and are no longer listed one by one. Of course, the server can also obtain unprocessed behavioral data and then process the behavioral data into the form of the above sequence, which is not limited in the embodiment of the present disclosure.
[0118] The server can perform multiple rounds of iterations based on the sample account data of the multiple sample accounts. For the i-th iteration, if i=1, it means the first iteration, and the account activity model of the i-th iteration is the initial model; if i>1, it means it is not the first iteration, and the account activity model of the i-th iteration is the model with parameters adjusted after the end of the i-1-th iteration.
[0119] The following is an introduction to the structure of the account activity model.
[0120] Figure 4 is a schematic diagram of a structure of an account activity model according to an exemplary embodiment. Figure 4 , the account activity model includes a first network, a second network, M third networks, M-1 fourth networks and M fifth networks. Wherein, M is a positive integer, indicating the number of time periods. The first network is used for feature embedding. The second network is used for feature crossover. The third network is used for feature mapping, and the N third networks have the same structure but different parameters. The fourth network is used to transform the output of one fifth network into the input of another fifth network. The fifth network is used to output the predicted activity information of the sample account in the corresponding time period.
[0121] In some embodiments, in the i-th iteration, the server inputs the sample account data of the sample account in the first sample time period into the account activity model of the i-th iteration, and the server can embed the sample account data based on the account activity model of the i-th iteration and then perform feature crossover to obtain the sample account features. The server extracts features from the sample account data of the sample account in the first sample time period based on the account activity model of the i-th iteration to obtain the sample account features, including: the server embeds the sample account data of the sample account in the first sample time period based on the first network in the account activity model of the i-th iteration to obtain sample embedding features, and the first network is used for feature embedding. Then, the server performs feature crossover on the sample embedding features based on the second network in the account activity model of the i-th iteration to obtain the sample account features, and the second network is used for feature crossover, and the feature crossover method is bit-wise feature crossover. Wherein, the second network includes a multi-layer perceptron. By embedding the sample account data based on the first network, the discrete data in the sample account data can be converted into Dense Embedding, making the discrete data continuous for subsequent processing. Then, based on the second network, bit-wise feature crossover is performed on the data output by the first network to improve the accuracy of subsequent predictions.
[0122] For example, the server performs one-hot encoding on discrete features such as gender, ID (Identity document, account number) and device model in the sample data based on the first network, and then converts the one-hot encoding result into Dense Embedding based on the preset embedding table to obtain sample embedding features. Among them, each category corresponds to an embedding table. The server inputs the Dense Embedding output by the first network into the second network, which includes a three-layer MLP. The server performs bit-wise feature crossover on the input Dense Embedding based on multi-layer perception to obtain sample account features. Among them, the output of each layer of the network in the second network is connected to a nonlinear activation function ReLU.
[0123] In step S302, the server maps the sample account features based on the account activity model of the i-th iteration to obtain multiple pieces of second activity information in multiple sample time periods.
[0124] In an embodiment of the present disclosure, the server can map the sample account features respectively based on the multiple third networks corresponding to the multiple sample time periods in the account activity model of the i-th iteration, and obtain multiple second active information of the multiple sample time periods. The second active information is the predicted activity of the sample account in the corresponding sample time period, and the multiple second active information corresponds to the multiple sample time periods one by one, and the multiple third networks correspond to the multiple sample time periods one by one. The sample time periods in the multiple sample time periods in the latter order include the sample time periods in the former order, and the end dates of the multiple sample time periods are later than the end date of the first sample time period. The start dates of the multiple sample time periods are the same, but the time lengths are different. In some embodiments, the start dates of the multiple sample time periods are the same, and the start date is within the first sample time period, or the start dates of the multiple sample time periods are later than the end date of the first sample time period. See step S202 for details, which will not be repeated here.
[0125] In some embodiments, since the structures of the multiple third networks in the account activity model are the same, the parameters are different. Take any sample time period as an example for explanation. The server maps the sample account features based on the account activity model of the i-th iteration to obtain multiple second active information of multiple sample time periods, including: for any sample time period, the server can map the account features based on the third network corresponding to the sample time period in the account activity model of the i-th iteration, and obtain the second active information of the sample time period. The third network is used for feature mapping, including a multi-layer perceptron, and the output of each layer of the network in the third network is connected to a nonlinear activation function ReLU. By setting up the third network for different sample time periods, modeling of each uneven distribution is achieved, that is, the data in each sample time period is unevenly distributed. Based on the corresponding third network, the activity of the sample account in the corresponding sample time period can be predicted, which improves the prediction range and thus improves the prediction efficiency.
[0126] In step S303, the server fuses multiple second activity information based on the account activity model of the i-th iteration to obtain multiple third activity information of the sample account in multiple sample time periods.
[0127] In the embodiment of the present disclosure, the third activity information is obtained by fusing the second activity information of two adjacent sample time periods, that is, the server can fuse the second activity information of two adjacent sample time periods to obtain the third activity information of the later sample time period.
[0128] In some embodiments, the second sample time period is any sample time period other than the sample time period at the end of the sequence in the multiple sample time periods, the second sample time period is adjacent to the third sample time period in the sequence of the multiple sample time periods, and the third sample time period includes the second sample time period. The server fuses multiple second active information based on the account activity model of the i-th iteration to obtain the third active information of the sample account in the multiple sample time periods, including: the server fuses the second active information of the second sample time period with the second active information of the third sample time period to obtain the third active information of the sample account in the third sample time period. By fusing the second active information of the second sample time period with the second active information of the third sample time period, the second active information of the second sample time period can gain the third sample time period, thereby having a positive impact on the active information of the sample account in the third sample time period, thereby improving the accuracy of the prediction.
[0129] For example, the length of the second sample time period is 30 days, and the length of the third sample time period is 60 days. Since the start date of the third sample time period is the same as that of the second sample time period, that is, the third sample time period includes the second sample time period, the active information of the sample account in the second sample time period has a positive impact on the active information of the sample account in the third sample time period. Therefore, the server gains the third sample time period by fusing the second active information of the second sample time period with the second active information of the third sample time period, thereby obtaining the third active information.
[0130] In some embodiments, the server can realize the fusion of the second active information of the second sample time period and the second active information of the third sample time period based on the fourth network and the fifth network in the account activity model of the i-th iteration. The server can map the second active information of the second sample time period based on the fourth network corresponding to the second sample time period in the account activity model of the i-th iteration, and obtain the monotonic gain information of the second sample time period, which is used to indicate the influence of the active behavior of the sample account in the second sample time period on the third sample time period. Then, the server can fuse the second active information of the third sample time period and the monotonic gain information of the second time period based on the fifth network corresponding to the third sample time period in the account activity model of the i-th iteration, and obtain the third active information of the sample account in the third sample time period. It should be noted that the fourth network includes two layers of MLPs whose network parameters are all non-negative, and can connect the fifth networks corresponding to adjacent sample time periods in series. The fourth network is a monotonic information unit (Mono Unit), which is used to realize the second active information output by the fifth network corresponding to the sample time period in the previous order, and the gain of the third active information output by the fifth network corresponding to the sample time period in the subsequent order. The fifth network is the last layer of the third network, and the fifth network can be used to output the result of the third network, that is, to output the second active information. By mapping the second active information of the second sample time period through the fourth network corresponding to the second sample time period, the active information of the sample time period in the previous order can be converted into the monotonic gain information of the sample time period in the subsequent order, and then the monotonic gain information is fused with the active information of the sample time period in the subsequent order based on the fifth network, so as to improve the accuracy of the predicted third active information.
[0131] For example, the start dates of multiple sample time periods are the same as the start date of the first sample time period, and the lengths of the multiple sample time periods are 30 days, 60 days, 90 days, 180 days, and 365 days, respectively. For example, if the second sample time period is a sample time period with a length of 30 days, the third sample time period is a sample time period with a length of 60 days. Figure 4 As shown, the fifth network corresponding to the second sample time period inputs the second gain information a1 output by the third network into the fourth network corresponding to the second sample time period, and the fourth network fits the second gain information to obtain monotonic gain information b, that is, b=f(a1), where f( ) represents the fitting function in the fourth network. Then, the fifth network corresponding to the third sample time period fuses the second gain information a2 corresponding to the third sample time period and the monotonic gain information b to obtain the third gain information c corresponding to the third sample time, that is, C=a2+b.
[0132] It should be noted that the fifth network in the account activity model is used to calculate the loss in addition to outputting the results of the third network. Before calculating the loss, the server can determine the actual active values of the sample accounts corresponding to different sample time periods based on the fifth network in a bucketing manner. Since the model learning is difficult due to the small number of samples at the tail, all the sample accounts at the tail are placed in one bucket, and the sample accounts at the head are also placed in one bucket. The width of the bucket limits the value range of the active value. This method is called variable width bucketing, that is, the width of each bucket is different. Buckets of different widths can keep the number of sample accounts in each bucket within the target number range, that is, the number of samples in each bucket is basically the same. For any sample time period, the server buckets the active values of multiple sample accounts in the sample time period based on the fifth network corresponding to the sample time period in the account activity model of the i-th iteration, and obtains multiple active buckets. The number of sample accounts in each active bucket is within the target number range. The fifth network is used to bucket the user account. Then, the server can obtain the mean of the active value range corresponding to each active bucket. Finally, the server determines the fourth activity information of the multiple sample accounts within the sample time period based on the mean of the active buckets to which the multiple sample accounts belong. By bucketing the multiple sample accounts, the labels of the sample accounts are converted from the original activity values to the mean of the active buckets to which they belong, thereby converting the regression problem into a more easily solved classification problem, that is, classifying the sample accounts into different active buckets, which can not only alleviate the problem of uneven distribution of sample accounts, but also improve the learning efficiency of the model.
[0133] For example, taking a sample time period of 30 days as an example, there are 5 active buckets, and the activity value ranges of the 5 active buckets are: 1, 2-5, 6-12, 13-18, 19-30, that is, the sample account with an activity value in the range of 19-30 falls into the 5th sub-bucket. At this time, all the information passing through the active bucket represents the fourth activity information of the sample account belonging to the active bucket, that is, the fourth activity information of the sample account in the 5th active bucket is (19+30) / 2=24.5.
[0134] In step S304, the server trains the account activity model of the i-th iteration based on the plurality of third activity information and the plurality of fourth activity information.
[0135] In the embodiment of the present disclosure, the multiple fourth activity information are used to represent the actual activity of the sample account in the multiple sample time periods, that is, the fourth activity information represents the activity value corresponding to the active bucket to which the sample account belongs in the corresponding time period.
[0136] In some embodiments, the server can adjust the parameters of the account activity model based on the difference between multiple third activity information and multiple fourth activity information. The step of the server training the account activity model of the i-th iteration based on multiple third activity information and multiple fourth activity information includes: the server determines multiple first losses for multiple sample time periods based on multiple third activity information and multiple fourth activity information, and the first loss is used to represent the difference between the predicted activity and the actual activity. Then the server adjusts the parameters of the account activity model of the i-th iteration based on the multiple first losses and the at least one second loss. The server can determine at least one second loss based on multiple third activity information, and the second loss is used to calibrate the order relationship between two third activity information in two adjacent sample time periods. Among them, the first loss is obtained by calculating the ordinal regression loss, and the second loss is obtained by calculating the ordinal calibration loss. By calculating the first loss, the parameters of the model can be adjusted based on the difference between the prediction results and the label results to achieve supervised training and improve the model training efficiency and the accuracy of model classification. By calculating the second loss, the sequential relationship between the third active information corresponding to two adjacent sample time periods can be measured. Based on the second loss, the incorrect sequential relationship can be penalized to correct the error, avoid unreasonable situations, and improve the accuracy of the prediction results output by the model.
[0137] In some embodiments, for any sample time period, there are multiple active buckets in the sample time period. In order to measure the ranking relationship between the corresponding activity values of different active buckets, the bucket ranking is modeled by the ordered regression loss, that is, the first loss mentioned above. The first loss of the sample time period can be calculated by the following formulas (1)-(4).
[0138]
[0139]
[0140]
[0141]
[0142] Wherein, L(X; θ) represents the first loss; X represents the sample account feature of the sample account; θ represents the model parameter; N represents the number of sample accounts; Ψ(X, θ) represents the ordered loss of each sample account; K represents the total number of active buckets in the current sample time period; k represents the activity value corresponding to the kth active bucket; lx represents the real activity value of the sample account; It indicates the probability that the activity value of the predicted sample account x is greater than the activity value corresponding to the kth active bucket; e indicates the exponential function; y( ) indicates the input of the activation function.
[0143] In some embodiments, in order to avoid unreasonable size relationship between the third active information corresponding to two adjacent sample time periods, the ordered relationship of the third active information corresponding to each sample time period is calibrated by an ordered calibration loss, that is, the second loss mentioned above. The second formula can be calculated by the following formula (5).
[0144]
[0145] Among them, l oc represents the second loss; N represents the number of sample accounts; O represents the number of sample time periods, and o=1 represents the first sample time period in the order of multiple sample time periods; x represents the sample account; D represents the sample data set, which includes multiple sample accounts; max() represents the maximum value function; represents the estimated value corresponding to the o-1th time period, that is, the third active information corresponding to the o-1th time period; represents the estimated value corresponding to the o-th time period, that is, represents the third active information corresponding to the o-th time period.
[0146] For example, the lengths of multiple sample time periods are 30 days, 60 days, and 90 days, respectively, that is, modeling LTV30, LTV60, and LTV90. Theoretically, LTV90>LTV60 is greater than LTV30. If the model predicts LTV30>=LTV60, or LTV60>=LTV90, it is unreasonable. The corresponding formula (5) above will obtain a loss corresponding to the sample time period, which can penalize the model so that the model learns the sequential relationship between the outputs of different sample time periods.
[0147] It should be noted that in order to make the model training steps described in the above steps 301-304 easier to understand, the model training process is further introduced below. The account activity model provided by the embodiment of the present disclosure is an efficient multi-label ordered dependency monotonic network model framework (Order-Dependency Monotonic Neural Network, referred to as ODMN). The training process of the account activity model includes two parts: forward calculation and back propagation. During the forward calculation process, the processed sample account data passes through the Embedding Layer (embedding layer) & Share Dense Layer (shared dense layer) and multi-layer nonlinear transformation to output the hidden layer vector. Among them, the Embedding Layer is Figure 4 The first network in the Shared Dense Layer is Figure 4 The hidden layer vector is input into the Tower network of each subtask, namely Figure 4 The third network in . Each Tower network models an unbalanced distribution, which corresponds to a sample time period, and there is an ordered dependency between these unbalanced distributions. Correspondingly, these Towers model loworder, mid order, and high order from left to right. It should be noted that low, mid, and high are explained as three levels. When there are more than three sample time periods, other methods can be used to represent the unbalanced distributions corresponding to different sample time periods. For example, in the scenario of estimating the number of active days (Life Time, LT) of the sample account in the future, there are 5 Tower networks, which are used to estimate LT30, LT60, LT90, LT180, and LT365 respectively. The final Output Layer of each Tower network will output a logit (logistic regression) value. Among them, the Output Layer is Figure 4 The fifth network in the tower network, the logit is used to calculate the first loss and the second loss, and is also used to output the predicted second active value. The second active value output by the Output Layer will be input into the Mono Unit (monotonic information unit), which will convert the second active value into monotonic gain information as one of the inputs of the Output Layer of the next Tower network. The Mono Unit is Figure 4 The fourth network in the network. Then, the OutputLayer of each Tower network will use the above-mentioned sequential regression loss and ordered calibration loss to train the model, which will not be repeated here. During the back-propagation process, the stochastic gradient descent algorithm is used to minimize the loss in the forward calculation, which is the sum of the first loss and the second loss. The parameters of the model and the parameters of the embedding layer are updated based on this loss.
[0148] It should be noted that the embodiment of the present application is explained by taking the training of the account activity model as an example. The above steps provided in the embodiment of the present application can also use other unevenly distributed data to train models of different applications, such as using the shopping behavior of the user account to train the account shopping model to predict the shopping situation of the user account; or using the gaming behavior of the account to train the account gaming model to predict the gaming situation of the user account. The embodiment of the present application does not limit this.
[0149] The following steps S305 to S307 describe the application process of the model. After the account activity model is trained to convergence, the server receives a request from the user account based on the terminal, and the server obtains account data based on the request, such as the user account's recent historical behavior data, user portrait, etc. The server inputs the account data into the account activity model, and the account activity model determines the account activity information corresponding to the user account in each time period based on the account data.
[0150] In step S305, the server extracts features from the account data of the determined user account within a historical time period based on the account activity model to obtain account features.
[0151] In the embodiment of the present disclosure, the account data includes behavior data and attribute data of the user account in a historical time period.
[0152] For example, the current time is January 8, and the historical time period is from January 1 to January 7, then the server can obtain the behavior data and user account portrait of the user account within the above 7 days.
[0153] In some embodiments, the server can extract features from the account data of the input model based on the first network and the second network in the account activity model to obtain account features. The server embeds the account data of the user account to be determined in the historical time period to obtain account embedding features, and then performs feature crossover on the account embedding features to obtain account features. By embedding the sample account data, the discrete data in the sample account data can be converted into dense embedding, making the discrete data continuous for subsequent processing, and then performing feature crossover on the account embedding features to improve the accuracy of subsequent predictions.
[0154] In step S306, the server maps the account characteristics based on the account activity model to obtain multiple first activity information in multiple time periods.
[0155] In the disclosed embodiment, the multiple time periods correspond one-to-one to the multiple third networks in the account activity model, the time period in the latter order among the multiple time periods includes the time period in the former order, and the time period in the first order among the multiple time periods includes the historical time period, and the first active information is the activity of the user account in the corresponding time period predicted by the account activity model. Among them, the third network includes a multi-layer perceptron, and for any time period, based on the multi-layer perceptron corresponding to the time period, the account features are mapped to obtain the first active information of the time period. By setting up the third network for different sample time periods, modeling of each uneven distribution is achieved, that is, the data in each sample time period is unevenly distributed, and based on the corresponding third network, the activity of the user account in the corresponding time period can be predicted, which improves the prediction range and thus improves the prediction efficiency.
[0156] In step S307, the server merges the first activity information in the first time period and the first activity information in the second time period based on the account activity model to obtain the account activity information of the user account in the second time period.
[0157] In an embodiment of the present disclosure, the first time period and the second time period are adjacent in order in the multiple time periods, and the second time period includes the first time period, and the account activity information is used to indicate the activity of the user account in the second time period.
[0158] For example, the first time period is from January 1 to January 30, and the second time period is from January 1 to February 29, that is, the time length of the first time period is 30 days, and the time length of the second time period is 60 days.
[0159] In some embodiments, the server fuses the above-mentioned active information based on the fourth network and the fifth network in the account activity model. The server maps the first active information of the first time period to obtain the monotonic gain information of the first time period, and the monotonic gain information is used to indicate the impact of the active behavior of the user account in the first time period on the second time period. Then the server fuses the first active information of the second time period with the monotonic gain information of the first time period to obtain the account active information of the user account in the second time period. By fusing the first active information of the first time period with the first active information of the second time period, the first active information of the first time period can gain the second time period, thereby having a positive impact on the active information of the user account in the second time period and improving the accuracy of the prediction.
[0160] The solution provided by the embodiment of the present disclosure can predict multiple active information of the user account in multiple time periods in the future by mapping the account features of the user account in the first time period based on multiple first networks corresponding to multiple time periods, and then determine the account active information of the user account in each time period based on the sequential relationship between two adjacent time periods, thereby improving the accuracy of the predicted account active information. Moreover, it has been verified that the account active information output by the account active model is very close to the real account active information, and alleviates the phenomenon that the predicted short-term active value is greater than the predicted long-term active value, and has achieved good results in the scenario of predicting the LTV of the user account.
[0161] All the above optional technical solutions can be arbitrarily combined to form optional embodiments of the present disclosure, and will not be described in detail here.
[0162] Figure 5 is a block diagram of a device for determining account activity information according to an exemplary embodiment. Figure 5 The device includes: a first extraction unit 501, a first mapping unit 502 and a first fusion unit 503.
[0163] The first extraction unit 501 is configured to extract features from account data of a user account to be determined within a historical period of time to obtain account features, where the account data includes behavior data and attribute data;
[0164] A first mapping unit 502 is configured to map the account feature to obtain a plurality of first activity information of a plurality of time periods, wherein the first activity information is a predicted activity of the user account in a corresponding time period, a later time period in the plurality of time periods includes a previous time period, and an end date of the plurality of time periods is later than an end date of the historical time period;
[0165] The first fusion unit 503 is configured to obtain account activity information of the user account in the second time period based on the first activity information in the first time period and the first activity information in the second time period, the first time period is adjacent to the second time period in the order of the multiple time periods, and the second time period includes the first time period, and the account activity information is used to indicate the activity of the user account in the second time period.
[0166] The device provided by the embodiment of the present disclosure can predict multiple active information of the user account in multiple time periods in the future by mapping the account features of the user account in the first time period based on multiple first networks corresponding to multiple time periods, and then determine the account active information of the user account in each time period based on the sequential relationship between two adjacent time periods, thereby improving the accuracy of the predicted account active information.
[0167] In some embodiments, the first fusion unit 503 is configured to map the first activity information of the first time period based on a multi-layer perceptron to obtain monotonic gain information of the first time period, where the monotonic gain information is used to indicate the impact of the active behavior of the user account in the first time period on the second time period; and to fuse the first activity information of the second time period with the monotonic gain information of the first time period to obtain the account activity information of the user account in the second time period.
[0168] In some embodiments, the first extraction unit 501 is configured to perform embedding processing on the account data of the determined user account within the historical time period to obtain an account embedding feature; and perform feature cross-processing on the account embedding feature to obtain the account feature.
[0169] In some embodiments, the first mapping unit 502 is configured to map the account features for any time period based on a multi-layer perceptron corresponding to the time period to obtain the first activity information for the time period.
[0170] In some embodiments, the device for determining account activity information executes steps based on an account activity model, and the account activity model is used to predict the account activity information of the user account in multiple future time periods based on the account data of the user account in a period of time.
[0171] In some embodiments, Figure 6 is a block diagram of another apparatus for determining account activity information according to an exemplary embodiment. Figure 6 , the device for determining account activity information also includes:
[0172] The second extraction unit 504 is configured to extract features from sample account data of the sample account in the first sample time period based on the account activity model of the i-th iteration to obtain sample account features, where the sample account data includes behavior data and attribute data, and i is a positive integer;
[0173] The second mapping unit 505 is configured to map the sample account features based on the account activity model of the i-th iteration to obtain a plurality of second activity information of a plurality of sample time periods, wherein the second activity information is the predicted activity of the sample account in the corresponding sample time period, wherein the sample time period in the latter order of the plurality of sample time periods includes the sample time period in the former order, and the end date of the plurality of sample time periods is later than the end date of the first sample time period;
[0174] The second fusion unit 506 is configured to fuse the plurality of second activity information based on the account activity model of the i-th iteration to obtain a plurality of third activity information of the sample account in the plurality of sample time periods, where the third activity information is fused based on the second activity information of two adjacent sample time periods;
[0175] The model training unit 507 is configured to train the account activity model of the i-th iteration based on the multiple third activity information and the multiple fourth activity information, where the multiple fourth activity information is used to represent the actual activity of the sample account in the multiple sample time periods.
[0176] In some embodiments, the second extraction unit 504 is configured to perform embedding processing on the sample account data of the sample account within the first sample time period based on the first network in the account activity model of the i-th iteration to obtain sample embedding features, and the first network is used for feature embedding; based on the second network in the account activity model of the i-th iteration, perform feature crossover on the sample embedding features to obtain the sample account features, and the second network is used for feature crossover.
[0177] In some embodiments, the second mapping unit 505 is configured to map the account features for any sample time period based on the third network corresponding to the sample time period in the account activity model of the i-th iteration to obtain the second activity information of the sample time period, and the third network is used for feature mapping, including a multi-layer perceptron.
[0178] In some embodiments, the second fusion unit 506 is configured to fuse the second activity information of the second sample time period and the second activity information of the third sample time period to obtain the third activity information of the sample account in the third sample time period, and the second sample time period is adjacent to the third sample time period in the order of the multiple sample time periods, and the third sample time period includes the second sample time period.
[0179] In some embodiments, the second fusion unit 506 is configured to map the second activity information of the second sample time period based on the fourth network corresponding to the second sample time period in the account activity model of the i-th iteration, and obtain the monotonic gain information of the second sample time period, where the monotonic gain information is used to indicate the influence of the active behavior of the sample account in the second sample time period on the third sample time period; based on the fifth network corresponding to the third sample time period in the account activity model of the i-th iteration, fuse the second activity information of the third sample time period and the monotonic gain information of the second time period to obtain the third activity information of the sample account in the third sample time period.
[0180] In some embodiments, see Figure 6As shown, the device for determining account activity information also includes:
[0181] The bucketing unit 508 is configured to bucket the activity values of multiple sample accounts in the sample time period based on the fifth network corresponding to the sample time period in the account activity model of the i-th iteration, to obtain multiple active buckets, wherein the number of sample accounts in each active bucket is within a target number range, and the fifth network is used to bucket the user accounts;
[0182] The bucketing unit 508 is further configured to obtain a mean value of the activity value range corresponding to each active bucket;
[0183] The bucketing unit 508 is further configured to determine fourth activity information of the multiple sample accounts within the sample time period based on the average of the activity buckets to which the multiple sample accounts belong.
[0184] In some embodiments, the model training unit 507 is configured to determine multiple first losses for the multiple sample time periods based on the multiple third activity information and the multiple fourth activity information, the first loss being used to represent the difference between the predicted activity and the actual activity; determine at least one second loss based on the multiple third activity information, the second loss being used to calibrate the sequential relationship between two third activity information in two adjacent sample time periods; and adjust the parameters of the account activity model of the i-th iteration based on the multiple first losses and the at least one second loss.
[0185] It should be noted that the device for determining the account activity information provided in the above embodiment only uses the division of the above functional units as an example when determining the account activity information. In actual applications, the above functions can be assigned to different functional units as needed, that is, the internal structure of the electronic device is divided into different functional units to complete all or part of the functions described above. In addition, the device for determining the account activity information provided in the above embodiment and the method embodiment for determining the account activity information belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.
[0186] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.
[0187] When an electronic device is provided as a terminal, Figure 7700 is a block diagram of a terminal 700 according to an exemplary embodiment. The terminal 700 may be a smart phone, a tablet computer, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), a laptop computer or a desktop computer. The terminal 700 may also be called a user device, a portable terminal, a laptop terminal, a desktop terminal or other names.
[0188] Typically, the terminal 700 includes a processor 701 and a memory 702 .
[0189] The processor 701 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 701 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 701 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 701 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 701 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.
[0190] The memory 702 may include one or more computer-readable storage media, which may be non-transitory. The memory 702 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 702 is used to store at least one program code, which is used to be executed by the processor 701 to implement the method for determining account activity information provided in the method embodiment of the present disclosure.
[0191] In some embodiments, the terminal 700 may further optionally include: a peripheral device interface 703 and at least one peripheral device. The processor 701, the memory 702 and the peripheral device interface 703 may be connected via a bus or a signal line. Each peripheral device may be connected to the peripheral device interface 703 via a bus, a signal line or a circuit board. Specifically, the peripheral device includes: at least one of a display screen 704, an audio circuit 705 and a power supply 706.
[0192] The peripheral device interface 703 may be used to connect at least one peripheral device related to I / O (Input / Output) to the processor 701 and the memory 702. In some embodiments, the processor 701, the memory 702, and the peripheral device interface 703 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 701, the memory 702, and the peripheral device interface 703 may be implemented on a separate chip or circuit board, which is not limited in this embodiment.
[0193] The display screen 704 is used to display the UI (User Interface). The UI may include graphics, text, icons, videos, and any combination thereof. When the display screen 704 is a touch display screen, the display screen 704 also has the ability to collect touch signals on the surface or above the surface of the display screen 704. The touch signal can be input to the processor 701 as a control signal for processing. At this time, the display screen 704 can also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, the display screen 704 can be one, and the front panel of the terminal 700 is set; in other embodiments, the display screen 704 can be at least two, which are respectively set on different surfaces of the terminal 700 or are folded; in some other embodiments, the display screen 704 can be a flexible display screen, which is set on the curved surface or folded surface of the terminal 700. Even, the display screen 704 can also be set as a non-rectangular irregular shape, that is, a special-shaped screen. The display screen 704 can be made of materials such as LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode, organic light-emitting diode).
[0194] The audio circuit 705 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, and convert the sound waves into electrical signals and input them into the processor 701 for processing, or input them into the radio frequency circuit 704 to achieve voice communication. For the purpose of stereo acquisition or noise reduction, there may be multiple microphones, which are respectively arranged at different parts of the terminal 700. The microphone may also be an array microphone or an omnidirectional acquisition microphone. The speaker is used to convert the electrical signal from the processor 701 or the radio frequency circuit 704 into sound waves. The speaker may be a traditional film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert the electrical signal into sound waves audible to humans, but also convert the electrical signal into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio circuit 705 may also include a headphone jack.
[0195] The power supply 706 is used to power various components in the terminal 700. The power supply 706 can be an alternating current, a direct current, a disposable battery, or a rechargeable battery. When the power supply 706 includes a rechargeable battery, the rechargeable battery can support wired charging or wireless charging. The rechargeable battery can also be used to support fast charging technology.
[0196] Those skilled in the art will understand that Figure 7 The structure shown in the figure does not constitute a limitation on the terminal 700, and the terminal 700 may include more or less components than those shown in the figure, or combine some components, or adopt a different component arrangement.
[0197] When the electronic device is provided as a server, Figure 8 It is a block diagram of a server 800 according to an exemplary embodiment. The server 800 may have relatively large differences due to different configurations or performances, and may include one or more processors (Central Processing Units, CPU) 801 and one or more memories 802, wherein the memory 802 stores at least one program code, and the at least one program code is loaded and executed by the processor 801 to implement the method for determining account activity information provided by the above-mentioned various method embodiments. Of course, the server may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server 800 may also include other components for implementing device functions, which will not be repeated here.
[0198] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 702 or a memory 802 including instructions, and the instructions can be executed by the processor 701 of the terminal 700 or the processor 801 of the server 800 to complete the above method. Alternatively, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.
[0199] In an exemplary embodiment, a computer program product is also provided, including a computer program, which implements the above-mentioned method for determining account activity information when executed by a processor.
[0200] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.
[0201] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.
Claims
1. A method for determining account activity information, characterized in that: The method comprises: Extracting features from account data of the determined user account within a historical period to obtain account features, wherein the account data includes behavior data and attribute data; Mapping the account features to obtain a plurality of first activity information of a plurality of time periods, wherein the first activity information is a predicted activity of the user account in a corresponding time period, a later time period in the plurality of time periods includes a previous time period, and an end date of the plurality of time periods is later than an end date of the historical time period; Based on a multi-layer perceptron, mapping is performed on first activity information of a first time period to obtain monotonic gain information of the first time period, the first activity information of the first time period is used to gain a second time period, the monotonic gain information is used to indicate an influence of the active behavior of the user account in the first time period on the second time period, the first time period and the second time period are adjacent in order in the multiple time periods, and the second time period includes the first time period; The first activity information of the second time period and the monotonic gain information of the first time period are fused to obtain the account activity information of the user account in the second time period, where the account activity information is used to indicate the activity of the user account in the second time period.
2. The method for determining account activity information according to claim 1, characterized in that: The feature extraction of the account data of the determined user account in the historical time period to obtain the account features includes: Embedding the account data of the determined user account within the historical time period to obtain an account embedding feature; Perform feature crossover on the account embedding feature to obtain the account feature.
3. The method for determining account activity information according to claim 1, characterized in that: Mapping the account characteristics to obtain a plurality of first active information in a plurality of time periods includes: For any time period, the account features are mapped based on the multi-layer perceptron corresponding to the time period to obtain the first active information of the time period.
4. The method for determining account activity information according to any one of claims 1 to 3, characterized in that: The method for determining account activity information is implemented based on an account activity model, and the account activity model is used to predict the account activity information of the user account in multiple future time periods based on the account data of the user account in a period of time.
5. The method for determining account activity information according to claim 4, characterized in that: The method for determining account activity information further includes: Based on the account activity model of the i-th iteration, feature extraction is performed on the sample account data of the sample account in the first sample time period to obtain sample account features, where the sample account data includes behavior data and attribute data, and i is a positive integer; Based on the account activity model of the i-th iteration, the sample account features are mapped to obtain a plurality of second activity information of a plurality of sample time periods, wherein the second activity information is the predicted activity of the sample account in the corresponding sample time period, a sample time period in a later order among the plurality of sample time periods includes a sample time period in a previous order, and an end date of the plurality of sample time periods is later than an end date of the first sample time period; Based on the account activity model of the i-th iteration, the multiple second activity information are merged to obtain multiple third activity information of the sample account in the multiple sample time periods, where the third activity information is merged based on the second activity information of two adjacent sample time periods; The account activity model of the i-th iteration is trained based on the multiple third activity information and the multiple fourth activity information, wherein the multiple fourth activity information is used to represent the actual activity of the sample account in the multiple sample time periods.
6. The method for determining account activity information according to claim 5, characterized in that: The account activity model based on the i-th iteration performs feature extraction on the sample account data of the sample account in the first sample time period to obtain sample account features, including: Based on the first network in the account activity model of the i-th iteration, embedding the sample account data of the sample account in the first sample time period to obtain sample embedding features, wherein the first network is used for feature embedding; Based on the second network in the account activity model of the i-th iteration, feature crossover is performed on the sample embedding features to obtain the sample account features, and the second network is used for feature crossover.
7. The method for determining account activity information according to claim 5, characterized in that: The account activity model based on the i-th iteration is used to map the sample account features to obtain a plurality of second activity information in a plurality of sample time periods, including: For any sample time period, based on the third network corresponding to the sample time period in the account activity model of the i-th iteration, the account features are mapped to obtain second activity information of the sample time period, and the third network is used for feature mapping, including a multi-layer perceptron.
8. The method for determining account activity information according to claim 5, characterized in that: The fusing the plurality of second active information to obtain a plurality of third active information of the sample account in the plurality of sample time periods includes: The second activity information of the second sample time period and the second activity information of the third sample time period are merged to obtain the third activity information of the sample account in the third sample time period, the second sample time period and the third sample time period are adjacent in order among the multiple sample time periods, and the third sample time period includes the second sample time period.
9. The method for determining account activity information according to claim 8, characterized in that: The fusing the second activity information of the second sample time period and the second activity information of the third sample time period to obtain the third activity information of the sample account in the third sample time period includes: Based on the fourth network corresponding to the second sample time period in the account activity model of the i-th iteration, mapping the second activity information of the second sample time period to obtain monotonic gain information of the second sample time period, wherein the monotonic gain information is used to indicate the influence of the active behavior of the sample account in the second sample time period on the third sample time period; Based on the fifth network corresponding to the third sample time period in the account activity model of the i-th iteration, the second activity information of the third sample time period and the monotonic gain information of the second time period are fused to obtain the third activity information of the sample account in the third sample time period.
10. The method for determining account activity information according to claim 5, characterized in that: The method for determining account activity information further includes: For any sample time period, based on the fifth network corresponding to the sample time period in the account activity model of the i-th iteration, the activity values of multiple sample accounts in the sample time period are bucketed to obtain multiple active buckets, and the number of sample accounts in each active bucket is within a target number range, and the fifth network is used to bucket the user accounts; Get the mean of the active value range corresponding to each active bucket; Based on the average values of the activity buckets to which the multiple sample accounts belong, fourth activity information of the multiple sample accounts within the sample time period is determined.
11. The method for determining account activity information according to claim 5, characterized in that: The training of the account activity model of the i-th iteration based on the plurality of third activity information and the plurality of fourth activity information includes: Determine, based on the plurality of third activity information and the plurality of fourth activity information, a plurality of first losses for the plurality of sample time periods, the first losses being used to represent a difference between a predicted activity and an actual activity; Based on the plurality of third activity information, determining at least one second loss, where the second loss is used to calibrate a sequential relationship between two pieces of third activity information in two adjacent sample time periods; Based on the multiple first losses and the at least one second loss, adjust the parameters of the account activity model of the i-th iteration.
12. A device for determining account activity information, characterized in that: The device comprises: A first extraction unit is configured to extract features from account data of a user account to be determined within a historical time period to obtain account features, wherein the account data includes behavior data and attribute data; a first mapping unit configured to map the account features to obtain a plurality of first activity information of a plurality of time periods, wherein the first activity information is a predicted activity of the user account in a corresponding time period, a time period in a later order among the plurality of time periods includes a time period in a previous order, and an end date of the plurality of time periods is later than an end date of the historical time period; The first fusion unit is configured to map the first activity information of the first time period based on a multi-layer perceptron to obtain monotonic gain information of the first time period, the first activity information of the first time period is used to gain the second time period, the monotonic gain information is used to indicate the influence of the active behavior of the user account in the first time period on the second time period, the first time period and the second time period are adjacent in order among the multiple time periods, and the second time period includes the first time period; the first activity information of the second time period and the monotonic gain information of the first time period are fused to obtain the account activity information of the user account in the second time period, and the account activity information is used to indicate the activity of the user account in the second time period.
13. The device for determining account activity information according to claim 12, characterized in that: The first extraction unit is configured to perform embedding processing on the account data of the determined user account within the historical time period to obtain an account embedding feature; Perform feature crossover on the account embedding feature to obtain the account feature.
14. The device for determining account activity information according to claim 12, characterized in that: The first mapping unit is configured to map the account features for any time period based on the multi-layer perceptron corresponding to the time period to obtain the first activity information of the time period.
15. The device for determining account activity information according to any one of claims 12 to 14, characterized in that: The device for determining account activity information executes steps based on an account activity model, and the account activity model is used to predict the account activity information of the user account in multiple future time periods based on the account data of the user account in a period of time.
16. The device for determining account activity information according to claim 15, characterized in that: The device for determining account activity information further includes: The second extraction unit is configured to extract features from sample account data of the sample account in the first sample time period based on the account activity model of the i-th iteration to obtain sample account features, wherein the sample account data includes behavior data and attribute data, and i is a positive integer; a second mapping unit, configured to map the sample account features based on the account activity model of the i-th iteration to obtain a plurality of second activity information of a plurality of sample time periods, wherein the second activity information is the predicted activity of the sample account in the corresponding sample time period, a sample time period in a later order among the plurality of sample time periods includes a sample time period in a previous order, and an end date of the plurality of sample time periods is later than an end date of the first sample time period; A second fusion unit is configured to fuse the plurality of second activity information based on the account activity model of the i-th iteration to obtain a plurality of third activity information of the sample account in the plurality of sample time periods, wherein the third activity information is obtained by fusing the second activity information of two adjacent sample time periods; The model training unit is configured to train the account activity model of the i-th iteration based on the multiple third activity information and the multiple fourth activity information, wherein the multiple fourth activity information is used to represent the actual activity of the sample account in the multiple sample time periods.
17. The device for determining account activity information according to claim 16, characterized in that: The second extraction unit is configured to, based on the first network in the account activity model of the i-th iteration, embed the sample account data of the sample account within the first sample time period to obtain sample embedding features, and the first network is used for feature embedding; based on the second network in the account activity model of the i-th iteration, perform feature crossover on the sample embedding features to obtain the sample account features, and the second network is used for feature crossover.
18. The device for determining account activity information according to claim 16, characterized in that: The second mapping unit is configured to map the account features for any sample time period based on the third network corresponding to the sample time period in the account activity model of the i-th iteration to obtain second activity information of the sample time period, and the third network is used for feature mapping, including a multi-layer perceptron.
19. The device for determining account activity information according to claim 16, characterized in that: The second fusion unit is configured to fuse the second activity information of the second sample time period and the second activity information of the third sample time period to obtain the third activity information of the sample account in the third sample time period, the second sample time period and the third sample time period are adjacent in order in the multiple sample time periods, and the third sample time period includes the second sample time period.
20. The device for determining account activity information according to claim 19, characterized in that: The second fusion unit is configured to map the second activity information of the second sample time period based on the fourth network corresponding to the second sample time period in the account activity model of the i-th iteration, and obtain the monotonic gain information of the second sample time period, where the monotonic gain information is used to indicate the influence of the active behavior of the sample account in the second sample time period on the third sample time period; and to fuse the second activity information of the third sample time period and the monotonic gain information of the second time period based on the fifth network corresponding to the third sample time period in the account activity model of the i-th iteration, and obtain the third activity information of the sample account in the third sample time period.
21. The device for determining account activity information according to claim 16, characterized in that: The device for determining account activity information further includes: A bucketing unit is configured to bucket the activity values of multiple sample accounts in the sample time period based on the fifth network corresponding to the sample time period in the account activity model of the i-th iteration, to obtain multiple active buckets, wherein the number of sample accounts in each active bucket is within a target number range, and the fifth network is used to bucket the user accounts; The bucketing unit is further configured to obtain a mean value of the activity value range corresponding to each active bucket; The bucketing unit is further configured to determine fourth activity information of the multiple sample accounts within the sample time period based on the average of the active buckets to which the multiple sample accounts belong.
22. The device for determining account activity information according to claim 16, characterized in that: The model training unit is configured to determine a plurality of first losses for the plurality of sample time periods based on the plurality of third activity information and the plurality of fourth activity information, wherein the first losses are used to represent a difference between a predicted activity and an actual activity; Based on the plurality of third activity information, determining at least one second loss, where the second loss is used to calibrate a sequential relationship between two pieces of third activity information in two adjacent sample time periods; Based on the multiple first losses and the at least one second loss, adjust the parameters of the account activity model of the i-th iteration.
23. An electronic device, characterized in that: The electronic device comprises: one or more processors; a memory for storing program code executable by the processor; The processor is configured to execute the program code to implement the method for determining account activity information as described in any one of claims 1 to 11.
24. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method for determining account activity information as described in any one of claims 1 to 11.
25. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for determining account activity information according to any one of claims 1 to 11 is implemented.
