TWM624658U - Prediction devices for predicting whether users belong to valuable user groups based on short-term user characteristics - Google Patents

Abstract

A prediction device for predicting whether a user belongs to a valuable user group based on short-term characteristics of a user executes steps as follows: obtaining a grouping result, wherein the grouping result indicates that at least one user belongs to a value user group; obtaining multiple user characteristics of the value user group, and establishing a first grouping model of the value user group based on the user characteristics, wherein at least one of the user characteristics is a user long-term characteristic of the user; extracting short-term user characteristics from the first grouping model, and establishing a second grouping model of the value user group based on the short-term user characteristics; and using the second grouping model to predict whether another one user belongs to the value user group based on short-term characteristics of the other one user.

Description

A predictive device that predicts whether a user belongs to a valuable user group based on the short-term characteristics of the user. set

The present invention relates to a prediction device for predicting whether a user belongs to a certain value user group, and particularly a prediction device for predicting whether a user belongs to a value user group based on short-term characteristics of the user.

In the traditional user management system, it is often necessary to wait until the monthly/quarterly/semi-annual/annual user bills to understand the value contribution of each user to the commodity or service provider (that is, the enterprise that provides the service or commodity). degree. However, users' behaviors are changing rapidly. In the past, the practice of waiting until monthly/quarterly/semi-annual/annual user bills to determine which value user group a user belongs to would make the product or service provider lose sight of the statistical data, resulting in It is impossible to instantly optimize user services for the current user environment and promote opportunities for users to increase consumption.

For example, based on quarterly billing, a user may belong to a high-value group with high purchases, but if you wait until the next quarterly bill to see if the user is still in a high-value group, you may have Churn this user. Simply put, if one can instantly understand The behavior of users has a better chance of maintaining high-value users, and may even turn quasi-high-value users into high-value users as soon as possible.

In order to solve the problems of the prior art, the present invention provides a prediction method for predicting whether a user belongs to a valuable user group based on short-term characteristics of the user. The method is executed in a prediction device, and includes: acquiring a grouping result, wherein the grouping result indicates that at least one user belongs to A value user group; a feature engineering is implemented based on a machine learning algorithm to obtain a plurality of user characteristics of the value user group, and a first grouping model of the value user group is established through the plurality of user characteristics, wherein At least one of the plurality of user features is a long-term feature of the user, and the long-term feature of the user is a feature that can only be obtained after reaching a target time; a plurality of short-term features of the user are extracted from the first clustering model, and the establishing a second grouping model of the valuable user group with short-term characteristics of a plurality of users, wherein the short-term characteristics of the users are a characteristic that can be obtained before the target time arrives; and using the second grouping model according to A plurality of user short-term characteristics of another user predict whether the other user belongs to the value user group.

In this novel embodiment, the plurality of users are grouped according to the plurality of user data obtained by the plurality of users when the target time arrives, so as to obtain the grouping result.

In this novel embodiment, the grouping results are grouped by an RFM model or a machine learning algorithm.

In this novel embodiment, the prediction method further includes: determining a marketing strategy for the other user according to whether the other user is predicted to belong to the valuable user group.

In the novel embodiment of the present invention, a plurality of user features in the valuable user group whose importance is the top few or exceeding a threshold value are selected as the plurality of user features in the valuable user group.

In the new embodiment of the present invention, a plurality of short-term features of a plurality of users whose importance is the top few or exceeding a threshold value in the first grouping model are selected as the plurality of users of the second grouping model short-term characteristics.

In the new embodiment of the present invention, a plurality of short-term features of users whose importance is the top few or exceeding a threshold value in the first grouping model are selected, and a plurality of short-term features of users are extended from the selected short-term features of users. short-term features of users, and the selected short-term features of users and the extended short-term features of users are used as the short-term features of users in the second grouping model.

In this novel embodiment, the prediction method further comprises: monitoring an intersection between the predicted multiple users of the value user group and the multiple users of the value user group of the grouping result over time. to determine whether the marketing strategies of the multiple users are correct.

In this novel embodiment, the prediction method further includes: integrating multiple users of the predicted value user group with the multiple users of the value user group in the grouping result as the value user the plurality of users of the group.

In this novel embodiment, using the second grouping model to predict whether the other user belongs to the valuable user group according to the multiple short-term characteristics of the other user is through a feature comparison, and The feature comparison is achieved through a class of neural network classification or computing a cosine similarity.

According to the purpose of the present invention, the present invention provides a non-volatile storage medium storing a plurality of code codes, which are read by a computer device to execute one of the above prediction methods.

According to the purpose of the present invention, the present invention provides a prediction device for predicting whether a user belongs to a value user group based on the short-term characteristics of the user. , and a plurality of modules operate to perform one of the above prediction methods.

Compared with the prior art, the new model can predict the user's value score (ie, predict whether the user belongs to a certain value user group) through the short-term characteristics of the user, such as daily or single consumption behavior, so it can target the individual user. , so that users can respond to the current service more instantly and comprehensively, without waiting for the final consumption result to respond.

1: Value user grouping system

11: Platform device

11S: User data set

111: Browsing Behavior

112: Interests

113: Device Information

114: Internet address (IP) or device identification

12: User management system

12S: User Data Set

121: Demographics

122:Consumption record

123: Accounting records

124: Interactive Record

13: Open Source Device

13S: User Data Set

131: Climatic Geography

132: Financial Economics

133: Humanities Population

134: Medical Biology

14:Database

15: Prediction device

151: Data processing module

152: Machine Learning Modules

153: Model Deployment Module

154: Marketing Module

21: User short-term feature set

211~21M: short-term characteristics of users

22: User group

231~23N: Value user group

S31~S35: Steps

41: Current RFM clustering results

411: High-value user group

42: Predicted RFM clustering results

421: High-value user group

51: Current value user base

52: Predicted value user base

531~533: Intersection

FIG. 1 is a block diagram of a value user grouping system according to an embodiment of the present invention.

FIG. 2 is a conceptual schematic diagram of grouping valuable users based on short-term characteristics of users according to the novel embodiment of the present invention.

FIG. 3 is a flowchart of a method for predicting value user groups based on short-term characteristics of users according to the novel embodiment of the present invention.

4 is a schematic diagram of an RFM prediction group and a current RFM group established by a method or apparatus for predicting value user groups based on short-term characteristics of users according to a novel embodiment of the present invention.

5A is a schematic diagram of the value user group obtained at the first time and the current value user group obtained by the method or device for predicting value user groups based on the short-term characteristics of users according to the novel embodiment of the present invention.

5B is a schematic diagram of a value user group and a current value user group obtained at a second time by a method or device for predicting value user groups based on short-term characteristics of users according to the novel embodiment of the present invention.

5C is another schematic diagram of the value user group and the current value user group obtained at a second time by the method or device for predicting value user groups based on short-term characteristics of users according to the novel embodiment of the present invention.

The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, and the protection scope of the present invention can be more clearly defined.

This new model uses machine learning methods to collect and analyze user data (including user short-term data, user long-term data and user non-time-related data) to analyze user behavior (corresponding to including user long-term characteristics, user short-term characteristics or non-temporal user data). User characteristics of related characteristics) to establish a first grouping model, and group users based on user data of users through the established first grouping model. Then, the correlation between the user value score and the user's short-term user characteristics is obtained through the first clustering model. This correlation can be understood as verifying whether the user's short-term user characteristics have a significant contribution to the user value score. The second clustering model rebuilt by short-term features has explanatory power and can be used according to the user's short-term features (representing short-term features). user behavior in the future) to predict user value (ie, predict whether a user belongs to a certain value user group, such as a high-value user group). For example, whether the number of times users log on to the web page per week is related to the user's lifetime value, or whether users are more willing to spend more if they stay on the homepage for more than 5 seconds each time they log on the page. Through feature engineering, user behavior analysis is performed from user data (including user short-term data and user long-term data) to extract important and meaningful user features (including user short-term features, user long-term features, and user non-time-related features). For example, divide user data into short-term user data and long-term user data on a daily/weekly/monthly/quarterly/semi-annual/yearly basis. Different business scenarios require different lengths of time. For example, for e-commerce retail, weekly In the long term, every hour is short term, whereas for the financial industry, every three years may be long term and weekly is short term. Simply put, the definition of short-term and long-term between short-term user data and long-term user data will be different for different business situations. In addition, the user's non-time-related data refers to time-independent data, such as the user's residence address, and the user's non-time-related features refer to time-independent features.

In the present invention, users are firstly grouped into valuable user groups by using user data, and each user is assigned to one of a plurality of valuable user groups. In terms of a grouping or classification model (for example, an RFM model (Recency Frequency Monetary Model, that is, a classification model that groups values such as recent consumption, consumption frequency, consumption amount, etc.), users may be grouped into, for example, high-value (VIP) user groups , high consumption user group, frequent purchase user group, recent high active user group or other value user group. Next, establish the first grouping model of a certain value user group (for example, the high value user group who are interested). After that, Perform feature engineering, use the first clustering model to find important and meaningful short-term user characteristics for this value user group, and use these important and meaningful short-term user characteristics to model the second clustering that is only related to the user's short-term behavior Then, based on the second clustering model, according to the short-term user characteristics of the current user, predict the user's mid- and long-term customer value evaluation The second grouping model is used to predict whether the user belongs to the valuable user group of the second grouping model according to the short-term user characteristics of the current user.

The above-mentioned use of the second grouping model to predict whether the user belongs to the valuable user group of the second grouping model according to the short-term characteristics of the current user is carried out through feature comparison. The second grouping model can be a grouping model obtained by a classification algorithm or a regression algorithm. Or neural-like networks trained by other machine learning algorithms. In addition, the feature comparison can also be performed by calculating the correlation between the values of the short-term features of the users and the values of the short-term features of the users in the second clustering model. For example, the cosine similarity can be calculated. Predict whether the user belongs to the value user group of the second grouping model.

Simply put, whether the defined RFM model, the first grouping model or the second grouping model is used, it can be used to score the user value of users. The RFM model and the first grouping model can use long-term user data. Grouping with the long-term characteristics of users. Under the current situation of fierce business competition, the RFM model and the first grouping model may not be able to perform effective and accurate marketing to users until there are long-term user data, so valuable users may be lost. Therefore, the present invention can use the second grouping model to score the user's value in advance according to the user's short-term data (that is, according to the user's short-term characteristics), so as to carry out effective and accurate marketing to the user, increase the consumption amount of the user and effectively avoid valuable User churn.

The meaning of user features that are important and meaningful to valuable user groups can refer to user features whose importance is the top few or whose importance exceeds a certain threshold. For example, the purchase amount is an important and meaningful user feature for high-value user groups. But user characteristics for spending with cash or credit cards may be unimportant and meaningless to high-value user groups. When building a swarming model, various user characteristics can be defined so that when evaluating a swarming model, you can understand whether the swarming model is Whether it has explanatory power (that is, whether the accuracy of the clustering is sufficient). If the explanatory power is not enough, the user features can be further increased. However, usually in a segmentation model, some user characteristics are unimportant and meaningless to the value user group of the segmentation model.

The meaning of using the grouping model to predict users may refer to scoring the user's user value, that is, performing feature comparison to predict whether the user belongs to the value user group of the grouping model. The feature comparison may refer to calculating the similarity between the value of the user feature of the user and the value of the user feature of the value user group of the grouping model, or may refer to the calculated probability result of the similarity. Of course, there are many methods of feature comparison, and the new model is not limited by the probability results calculated by the neural network or the method of calculating cosine similarity, and other methods of feature comparison implemented by support vector machines or decision trees. also within the scope of the present invention.

In this new model, the definition of short-term can refer to the short-term answer that is shorter than the answer you want to know. For example, you want to know which users contribute more at the end of the month. Two weeks or three weeks are short-term compared to one month, that is, a period shorter than the target period can be called short-term, and the target period is called long-term. It should be noted that some features are not related to time, usually not long-term features of users or short-term features of users, but non-time-related features of users. Household address is neither a short-term feature nor a long-term feature of the user. User short-term features specifically refer to features with a relative time concept.

In this new model, user data is usually original data, not user features, unless the database is initially designed to store user features as user data. Average, and store the number of user visits to the store and the average weekly consumption as user data. Features are representative data, which can be understood as input parameters. Meaningful input parameters or variables will affect the final user value score, such as the user's browsing behavior characteristics, which may be that they often click on advertisement content after entering the web page, or like to watch videos, or stay on the web page for no longer than 10 seconds, etc., could be meaningful features.

Common features can be divided into numerical type, category type and time type in terms of type. These three types of features can be converted to each other, for example, the features of the numerical type can be processed and converted into the features of the time type. The characteristic of the numerical type is to use the statistical description method to divide the overall data range into scores. The feature of the category type refers to what category it belongs to, for example, the device information may be a laptop, a mobile phone or a tablet, and which one of the device is a laptop, a mobile phone and a tablet is the feature of the category type. In addition, the features of the numerical type can also be converted into the features of the category type, or the features of the category type can also be weighted according to the weight of each category, so as to be converted into the features of the numerical type. The time-type feature may be the frequency, number of times, or the size of the numerical feature based on weekly/monthly/quarterly calculation methods, etc. There are many calculation methods, and the method described in the present invention is not limited. Simply put, any of the three types of features can be processed to generate features of the other.

The purpose of using user short-term features in this new model is to use short-term feature parameters to predict long-term results. Although user short-term features are related to time, because feature types can be converted to each other, in a broad sense, user short-term features can include the aforementioned numerical types , the features of category type and time type, but the features can be segmented by a shorter time (short-term) than the target time (long-term). The short-term characteristics of users are relative concepts. For example, it is generally possible to wait until the end of the month to know the final expenditure of the user in the current month, but from the user's weekly consumption records, frequency, cycle, number of visits, browsing frequency, clicks and frequency, and whether there is a subscription. Shopping carts or wish lists, as well as the user's past history, etc., have a good chance of predicting the user's purchase history at the end of the month. After identifying the important short-term characteristics of users, an automated process can be established accordingly. In actual use, it is to automatically convert and extract short-term user characteristics from user data.

In the present invention, an extended feature may refer to a feature that does not exist originally, and an extended feature may refer to a feature calculated, converted or extracted using existing features or existing data. Data that is too detailed may be difficult to use directly as a feature. Therefore, the data can be converted to produce. For example, general address data may contain a lot of information, such as the address of the Presidential Office: "106 No. 1, Section 2, Xinsheng South Road, Daan District, Taipei City", If the data is directly used for modeling, the data will be too detailed to be used. At this time, some feature conversion processing will be performed, which may convert the original "Address: No. 1, Section 2, Xinsheng South Road, Daan District, Taipei City, 106". "Postal Code: 106, City: Taipei City, Daan District, Township Urban District, Other Address: No. 1, Section 2, Xinsheng South Road" or "latitude and longitude: 25N/121E" which may obtain rough positioning information through longitude and latitude query. Extended features for model training and processing. Another example: if the existing feature is "login frequency every 28 days = 10", the extended feature of "average login frequency per week = 2" or "average login frequency per biweekly = 4" can be extended. Another example is that the existing or original data is "ID: 0001; 2020-9-1 10:00:05; log in to the homepage; use device device=iphone12", which can be extended with "user ID", "login time", Multiple extended features related to "login times" and "use device", wherein, for example, the feature of "login time" can include multiple extended features calculated in different time periods and periods, and the feature of "use device" can include the feature related to " Device brand” and “device model” related extension features, etc.

Another example is given to illustrate the extended feature. Suppose that in a numerical regression application, the existing features of a model for evaluating a chip's performance include resistance R and current I. If someone skilled in the relevant field proposes that power P is also an important and meaningful feature , at this time, a feature of power P=I ² R can be extended as a feature of a model for evaluating the performance of a chip. Another example: in the context of Internet marketing, the total number of impressions I and the total number of clicks TC are existing features, and the extended feature of CTR=TC/I can be extended to evaluate the effectiveness of advertisements, keywords, etc. Another example: a user data includes "ID: 0001; 2020-9-1 10: 00: 05; log in to the home page", "ID: 0001; 2020-9-1 10: 00: 15; leave the home page", "ID : 0001; 2020-9-5 14: 08: 30; log in to the home page", "ID: 0001; 2020-9-5 14: 08: 30; consumption; consumption amount = 1,000" and "ID: 0001; 2020-9 -5 14:08:50; leave the homepage", according to the visit status recorded in the original data, that is to analyze the time of logging in/leaving the homepage. The number of homepages, the number of homepage logins in September, the number of homepage logins in the first week of September, the frequency of visits during the day, the duration of each login, and the length of the session. In short, in the present invention, similar to the above-mentioned method of feature extension, the existing user data or predefined user features are extended to have important and meaningful features for the grouping model.

Please refer to FIG. 1 , which is a block diagram of a value user grouping system according to a new embodiment of the present invention. The value user grouping system 1 is used in conjunction with the user service system (not shown in the figure), the user service system is used to provide services to users, and the value user grouping system 1 is used to group users according to the user information of the service users. To predict whether the user belongs to a certain value user group, so as to determine the marketing strategy and increase the satisfaction of transactions and users. The services provided above are, for example, online or physical shopping, online or physical transactions, online or physical investments, online or physical subscriptions, or online or physical consultations, etc. That is, the user service system may include online service hosts or physical storefront sales points Terminal (POS), and the present invention is not limited by this. In other words, the valuable user grouping system 1 can group valuable users for users who use online services or physical services.

The value user grouping system 1 includes a platform device 11 , a user management system 12 , a public data source device 13 , a database 14 , and a prediction device 15 for predicting whether a user belongs to a value user group based on short-term characteristics of users, wherein the database 14 is connected to the platform device 11 by signal , a user management system 12 , a public data source device 13 and a prediction device 15 . In addition, the "signal connection" in the present invention refers to a wireless or wired connection through software or hardware, so that signals or information can be transmitted to each other in a plurality of components that are signal-connected.

The platform device 11 is an enterprise platform or other type of platform, which is used to obtain user information of user browsing behavior 111 , interests and preferences 112 , device information 113 , Internet address (IP) or device identification 114 . The user management system 12 is used to obtain user data of the user's demographic data 121 , consumption records 122 , account records 123 and interaction records 124 . The public data source device 13 is used to obtain user data of the user's climate geography 131 , financial economy 132 , human population 133 and medical biology 134 . The multiple user data of the user obtained by the platform device 11, the user management system 12 and the public data source device 13 can be stored in the database 14, and the database 14 can be an unstructured database, but the present invention is not limited to this, and also Possibly a structured repository.

The prediction device 15 is usually realized by means of software and a computer device. For example, the computer device includes an operation unit, a storage unit and a communication unit, wherein the operation unit is electrically connected to the storage unit and the communication unit. The computing unit will read multiple code codes stored in the non-volatile storage medium, and after the multiple code codes are executed, the method of predicting the user's value score based on the user's short-term characteristics of the user (that is, using the user's short-term user characteristics to predict the user's value score) can be executed. feature to predict whether a user belongs to a certain value user group). In addition, the prediction device 15 can also be implemented through a hardware circuit, for example, by designing Verilog or VHDL code to program a Field Programmable Array (FPGA) chip to realize the prediction device 15 of a pure hardware circuit, or, for example, through Fabrication of Application Specific Chips (ASICs) to achieve pure The prediction device 15 of the hardware circuit. All in all, whether the prediction device 15 is implemented using software or hardware is not intended to limit the present invention.

Regardless of whether the prediction device 15 is implemented by software or hardware, it can be divided into multiple modules as shown in FIG. 1 . In FIG. 1, the prediction device 15 includes a data processing module 151, a machine learning module 152, a model deployment module 153, and a marketing module 154. The data processing module 151, the machine learning module 152, the model deployment module 153 and the In addition to being signal-connected to each other, the marketing modules 154 are also signal-connected to the database 14 . Of course, the module division of the prediction device 15 in FIG. 1 is only one way to realize the present invention, and the method in FIG. 1 is not intended to limit the present invention. In addition, the data processing module 151 , the machine learning module 152 , the model deployment module 153 and the marketing module 154 will be described together with the method for predicting value user groups based on the short-term characteristics of users in FIG. 3 of the present invention. I won't go into details first.

Before explaining the details of the prediction device 15 of the present invention and the prediction method of FIG. 3 , please refer to FIG. 2 of the present invention. FIG. 2 is a conceptual schematic diagram of grouping valuable users based on short-term user characteristics according to the present invention. In FIG. 2 , the user data set 12S of the demographic data 121 , the consumption record 122 , the account record 123 and the interaction record 124 among the plurality of user data is used to group the value users of the plurality of users, and each user will be grouped into one of the multiple value user groups 231, 232, . . . , 23N, where N is a positive integer greater than or equal to 2. The grouping method of valuable users can be, for example, the RFM model, or the grouping method using other models, such as grouping by static characteristics (age, gender, occupation, registration time point), or by dynamic characteristics (RFM parameters, membership level or promotion. Status, account balance, and the method of pickup/delivery of the goods used) are grouped, and this new model is not limited by this.

In other words, on the right side of Figure 2, the RFM model can be used to first group multiple users with valuable users, wherein the RFM model must be based on the user data generated by the user data set 12S in the user data before the value users can be classified. grouping. Users of User Profile 12S Most of the data are long-term user data. Therefore, the RFM model needs at least long-term user data to be able to group valuable users. In addition, the user data of the user data set 12S may also include user short-term data and user non-time-related data.

As mentioned above, the purpose of this new model is to use short-term user characteristics that are shorter than the statistical time of users' long-term characteristics to know in advance whether each user of the user group 22 may belong to a certain member of multiple valuable user groups 231-23N. a group. Therefore, through the user data of browsing behavior 111, interests 112, device information 113, Internet address or device identification 114 in the user data set 11S and climate geography 131, financial economy 132, human population in the user data set 13S 133. A user short-term feature set 21 including a plurality of user short-term features 211-21M can be generated from the user profile of the medical biological device 134.

In order to make it possible to use at least a part of the short-term user features 211-21M in the user short-term feature set 21 to predict whether the user belongs to a certain group of the multiple value user groups 231-23N, the present invention uses the value user groups 231-23N interested in (maybe all of them are interested) to establish a corresponding first grouping model, and perform feature engineering on the first grouping model to find out the short-term characteristics of users to establish a second grouping model, so as to use the second The user's short-term characteristics predict whether the user belongs to a certain group of the value user groups 231-23N. For example, for the value user group 231, find out the user characteristics that are important and meaningful to the value user group 231 (including user long-term characteristics, user short-term characteristics and user non-time-related characteristics), and then perform feature engineering to find the first The short-term characteristics of users in the grouping model, and use these short-term characteristics of users to establish a second grouping model of the value user group 231, and then use the second grouping model of the value user group 231 to predict whether the user belongs to Value User Group 231.

Next, please refer to FIG. 1 and FIG. 3 of the present invention. FIG. 3 is a flowchart of a method for predicting value user groups for users based on short-term characteristics of users according to an embodiment of the present invention. First, in step S31, the data processing module 151 is used to obtain a grouping result, wherein the grouping result represents at least one The users included in the value user group. Further, the data processing module 151 can collect user data (including at least long-term user data), and classify the users into corresponding valuable user groups according to the user data. For example, through the RFM model, users can be classified into high-value (VIP) user groups, high-consumption user groups, frequent purchase users, recent high-active users or other valuable user groups based on user profiles. In addition, in step S31, the data processing module 151 may also process the user data to obtain user characteristics (including at least long-term user characteristics), and use a preset grouping model to obtain grouping results based on the user characteristics. Simply put, the method of obtaining the grouping result by the data processing module 151 is not intended to limit the present invention. Even, in other embodiments of the present invention, the grouping may not be performed by the data processing module 151, but the data processing module 151 only receives the grouping result.

In the case of the membership system commonly used in general e-commerce, each user will obtain a user account after registration, and all future behaviors of the user in the user service system will be recorded in the database 14, such as when the user consumes and how much he consumes. Money, time, place and items of consumption, which pages were clicked on before consumption, how many items were put into the shopping cart and how many items were actually purchased, etc. After the data processing module 151 obtains these data and collects statistics at the end of the month, it can obtain the grouping results through various grouping methods (for example, the aforementioned RFM model, but it can also be grouped by product hobbies or activity participation). Taking the RFM model as an example, the RFM model classifies users into high-value (VIP) user groups, high-consumption user groups, frequent purchase users, recent high-active users, or other users according to the total consumption, consumption frequency and the last consumption. value user group.

In step S32, a first grouping model of the valuable user group is established through the data processing module 151 and the machine learning module 152. According to the grouping result, it is possible to know who are the users of a certain valuable user group. The data processing module 151 can process the user data of the users of this valuable user group, and obtain or identify from the user data according to the system preset or artificially set rules. multiple user characteristics (including at least long-term user characteristics, but may also include user short-term characteristics and user non- time-dependent features). Next, the machine learning module 152 establishes a first grouping model of the valuable user group according to the user characteristics obtained by the data processing module 151 .

Optionally, since some of the user characteristics may not be important and meaningful to the value user group, if the user characteristics generated by the data processing module 151 are not known in advance, they will definitely be related to the value user group, then in step S32 , the machine learning module 152 can learn to select only the user characteristics that are important and meaningful to the value user group. Feature engineering can be performed using machine learning algorithms to extract and identify user features that are important and meaningful to value user groups (eg high value user groups). Usually, feature selection can be performed, and the first few user features (or the importance exceeding the threshold value) are selected as important user features of the value user group according to the importance score, and these important user features of the value user group are for the value user group. are meaningful and important user characteristics. Then, the establishment of the first grouping model can be completed by using these important user characteristics.

For example, after knowing which users belong to the high-value user group and which users do not belong to the high-value user group, it is possible to extract meaningful and important user characteristics for the high-value user group. Groups can also extract user features in a similar way. The extracted user characteristics can be, for example, the users of the high-value user group may browse the web page for at least 10 seconds per day, or the users of the high-value user group will enter the homepage of the web page for at least 5 seconds per week, or the high-value user group. Users are accustomed to logging in via mobile phones rather than PCs, etc. Knowing that the above user characteristics are important, the data processing module 151 will process the user data to extract the above user characteristics, and the machine learning module 152 will establish a high-value user group based on these user characteristics. The first clustering model.

The user data of the user is usually original data, which needs to be processed and converted by the data processing module 151 . In this example, the user profile of one of the users in the high-value user group includes "ID: 0001; 2020-9-1 10:00:05; login home page; using device device=iphone12", "ID: 0001; 2020-9-1 10: 00: 15; leave the home page; use device device=iphone12", "ID: 0001; 2020-9-5 14: 08: 30; log in to the home page; use device device=iphone12 ", "ID: 0001; 2020-9-5 14: 08: 30; consumption; consumption amount=1,000; using device device=iphone12" and "ID: 0001; 2020-9-5 14: 08: 50; leave the home page ; using device device=iphone12".

The data processing module 151 can thus automatically extract the characteristics of the user of “ID: 0001” as “weekly login frequency=2”, “average stay time=15”, “weekly homepage stay times=2” and “every Weekly consumption amount = 1,000" and "use device device = iphone12". In this example, "use device device=iphone12" can be regarded as a non-time-related feature of the user. If the statistics are made at the end of the month, then "weekly login frequency=2", "average stay time=15", "weekly homepage" "Number of stays = 2" and "Weekly consumption amount = 1,000" are both short-term characteristics of users. It should be noted that although the long-term characteristics of users are not listed here, in fact, the data processing module 151 can also obtain long-term characteristics of users. For example, "Monthly Login Frequency=9", "Average Stay Time=12", "Monthly Home Page Stays=9", and "Monthly Spending Amount=3,000".

In step S33, based on the first grouping model, the machine learning module 152 performs feature engineering to extract short-term user features, and establishes a second grouping model according to the extracted short-term user features. Please note here that the second clustering model can also use the user non-temporal related features as input parameters, that is, the machine learning module 152 can further extract the user non-temporal related features, and use the extracted user short-term features and user non-temporal related features. Build a second clustering model. If the user characteristics of the first grouping model include long-term user characteristics, short-term user characteristics, and non-time-related characteristics of users, the short-term characteristics of users in the first grouping model can be selected according to their importance (optionally, important and short-term characteristics can also be selected. meaningful user non-time-related features), and then using the extracted user features, a second grouping model can be established. Feature engineering can be implemented through machine learning algorithms, and the present invention is not limited to this.

In some cases, the second clustering model may not have complete explanatory power, because the first clustering model may not have enough short-term features of users, or even no short-term features of users at all. Therefore, in step S33, the data processing module 151 can further process the user data, and extend a plurality of short-term user characteristics according to the user characteristics of the first grouping model, so that the established second grouping model has more explanatory power. For example, if one of the long-term characteristics of users in the first clustering model is "Monthly login frequency>9", the short-term users with "weekly login frequency>2" or "biweekly login frequency>4" can be extended. feature. In addition, in the feature extension part, one of the methods is to draw a radar chart to extend the short-term user features according to the weak items of the user features in the first clustering model, so that the second clustering model has sufficient explanatory power.

For example, if it is found that for the first segmentation model of the high-value user group, the important and meaningful short-term characteristics of users are "weekly login frequency>2", "average dwell time>10" and "weekly consumption amount> 1,000”, then use these short-term user features to build and train a second clustering model. In addition, if there are unimportant and meaningless user features in the first grouping model of the high-value user group as "weekly homepage times = 2" and "use device device = iphone12", these user features will not be used for Train and build a second clustering model to avoid overfitting.

The data processing module 151 can determine the short-term period of the short-term feature according to a short-term period rule, and the short-term period can be used as the observation period of the short-term feature value. The data processing module 151 can adjust the short-term period of the short-term feature according to a short-term period rule. For example, the data processing module 151 can analyze the distribution of the feature value on the time axis as a basis for adjusting the short-term period. For example, the data processing module 151 can perform time series analysis on the long-term feature data set, such as finding out seasonality and other features, so as to apply corresponding short-term period rules. Short-term periods can include calendar periods such as hours, days, weeks, months, quarters, and so on. Short-term time periods can also include changes in calendar periods, such as bi-daily, bi-weekly, etc. The short-term characteristics of users are not necessarily divided into fixed periods. If user behaviors such as browsing, visiting, and consumption are non-linear, a non-fixed period can be used to set short-term periods. For example, when the user's behavior on the platform device 11 is relatively active at the beginning of each month, but the activity decreases with time to the end of the month, if a fixed period is used to segment the feature, the feature value near the end of the month may make the prediction inaccurate. Users are used to spending an amount V ₁ >10,000 yuan in the first week of each month, but the consumption amount decreases until the end of the month. For example, the consumption amount in the second to fourth weeks is V ₂ =2,000, V ₃ =0, V ₄ = 0, if the monthly consumption amount is a long-term feature, and if the short-term feature is set to a fixed period, the consumption amount in the last one or two weeks may predict this user data as a negative example. In order to solve this problem, the short-term period can be set by increasing, decreasing or other changes. For example, the period is not necessarily continuous. For example, a specific time point can be used as the anchor point to observe the period within a window starting from the anchor point, for example, Take the first day of each month as the anchor point to observe the values in the next week, that is, set a window of seven days, set the first week as the first cycle, and set the first to fourth weeks as the second cycle , or even ignore weeks 1 to 4 and only make the first week of the month a period of short-term features; or, set the window to be every Monday through Wednesday and ignore every Thursday through Sunday to make every Monday to Wednesday as a short-term characteristic cycle, etc. Alternatively, the period may be incremented by adding one day to the previous period, for example, in the sequence of period P, P ₁ =1 day, P ₂ =2 day, P ₃ =3 day, P ₄ =4 day...P _n = n days etc. Alternatively, the period of short-term features can be increased in a gradient manner, for example, within one month after membership registration, the short-term period is weekly, within two to six months after registration, the short-term period is every two weeks, etc. Periods can also overlap, such as window sliding to determine the short-term period, if the window is set to 3 days, September 1st to September 3rd is a short-term period, September 2nd to September 4th is a short-term period, 9 The period from March 3 to September 5 is a short-term period. It can be seen from the above description that short-term characteristics are relative to long-term characteristics, so long-term characteristics may also be non-fixed or have corresponding changes with short-term characteristics. The long-term feature may have a period of one month, and within two to six months after registration as a member, the long-term feature may have a period of six months.

Next, after the second grouping model is established, in step S34, the model deployment module 153 deploys the second grouping model, and uses the second grouping model according to the user's user short-term User characteristics predict whether the user belongs to the valuable user group of the second grouping model. Simply put, after the second grouping model is deployed, every time the user logs in, the prediction device 15 can know whether the user belongs to the second group without waiting for the target time (such as a long-term period with long-term characteristics) to arrive. The value user base of the model. Incidentally, the practice of using the second clustering model to predict whether a user belongs to the valuable user group of the second clustering model based on the user's short-term user characteristics can be done through feature comparison, such as neural network-like classification or calculation. Cosine similarity.

After that, after predicting whether the user belongs to the value user group according to the short-term characteristics of the user, in step S35, the marketing module 154 determines the marketing to the user according to the prediction result of whether the user belongs to the value user group of the second grouping model Strategy. For example, if short-term features are extracted based on the user's user data for a week, and the user is predicted to be a high-value user group based on the user's short-term features, the discount coupons or promotions for quasi-high-value users or high-value users can be directly sent to the user. to increase the success rate of transactions or subscriptions. Incidentally, the predicted result of the predicted user as a value user group can be presented in the form of probability, for example, 50%, 70% or 90%, therefore, the marketing module 154 can give different discounts according to the probability value of the predicted result information to users.

For example, multiple user short-term signatures of multiple users are "ID: 0001; weekly login frequency=2; average stay time=15; weekly consumption amount=800", "ID: 0002; weekly login frequency =0.01; Average length of stay=7; Weekly spend=600", "ID: 0003; Weekly login frequency=1; Average length of stay=50; Weekly spend=50,000", "ID: 0004; Weekly Login Frequency=3; Average Stay Time=25; Weekly Spend=1,300" and "ID: 0005; Weekly Login Frequency=10; Average Stay Time=50; Weekly Spend=15,300", and the second point The short-term characteristics of users in the group model are "weekly login frequency>2", "average stay time>10" and "weekly consumption amount>1,000", it can be predicted that users with "ID: 0004" and "ID: 0005" belong to The value user group of the second grouping model.

Please note that the above-mentioned second clustering model can be updated according to the period of short-term features, such as updating according to the shortest period, according to the period of the characteristic with the shortest period among all the features of the second grouping model, or any time, real-time, Or update almost instantaneously, such as every hour or two. For example, even if the user's consumption amount increases because of an event, which can be festive, one-off, temporary, sudden, short-term, etc., the user who organizes the event to increase the user's consumption will be short-term. The features can also be regarded as the user short-term features of the second clustering model, ie, the user short-term features related to the activity are used to supplement the original multiple user short-term features. After the event is over, because the second grouping model is updated at any time, after a period of time after the event is over, the second grouping model will not take the short-term user characteristics that the event promotes users to increase consumption as its short-term user characteristics , that is, the user short-term features related to the activity are removed from the multiple short-term user features. Another approach is to prepare two second grouping models. One of the second grouping models is a second grouping model that does not consider the short-term characteristics of users related to holding activities for normal use, and the other is a second grouping model. It is a second grouping model that considers the short-term characteristics of users related to holding activities as a second grouping model for holding activities. For example, if the event is to promote a certain type of product, the second grouping model used for the event can use the short-term characteristics of the user's interest and the amount purchased during the event every week (assuming the event time is one month) as the first grouping model. User short-term characteristics of a dichotomous model.

In addition, the source and type of the aforementioned user data are not limited by the aforementioned content. The raw data of the user data may also include sensing and transmission to the prediction device through the Internet of Things or other devices. For example, when a user parks in a store, the parking payment opportunity records the user's parking time and number of times, etc. The second clustering model can take the ratio of the user's weekly parking times to the number of visits to the store as its short-term user characteristics. In addition, there may be more than one data source for the user data of the same user. For example, user data obtained through the POS of physical chain stores in different branches to obtain short-term user characteristics. Another example is a chain coffee shop where the user is not at the original workplace because of the transfer. Therefore, the short-term user characteristics of users can be obtained correctly after the user data from the two data sources are integrated, which also indicates the integration of non-temporal characteristics and short-term characteristics. application.

Please refer to FIG. 4 . FIG. 4 is a schematic diagram of the RFM prediction grouping and the current RFM grouping established by the method or apparatus for predicting value user grouping for users based on short-term characteristics of users according to the novel embodiment of the present invention. The RFM prediction grouping can refer to predicting the customers who meet each value customer group in the aforementioned short-term characteristic method for each value customer group based on the grouping result of the RFM model. The high-value user group 411 in the current RFM grouping result 41 after the RFM grouping is performed according to the user data at the target time may be different from the high-value user group 421 in the RFM grouping result 42 predicted according to the prediction method or device of the novel embodiment. Users, users of the high-value user group 421 of the predicted RFM grouping result 42 can be regarded as high-value or quasi-high-value users, and can be directly marketed. That is to say, users in the high-value user group 421 can be directly added to the high-value user group 411, so as to optimize the current RFM grouping result 41 and find potential users with high value as much as possible. To put it simply, the prediction method of this novel embodiment further includes the step of combining the predicted value user group with the current value user group. In addition, the user changes between the high-value user groups 411 and 421 can also be used to predict whether the current marketing strategy is correct, and the description is as follows.

Please refer to FIG. 5A and FIG. 5B at the same time. FIG. 5A is a schematic diagram of the value user group and the current value user group obtained at the first time by the method for predicting the value user group or the device according to the short-term characteristics of the user according to the new embodiment of the present invention, 5B is a schematic diagram of the value user group and the current value user group obtained at a second time by the method or device for predicting value user groups based on short-term characteristics of users according to the novel embodiment of the present invention. At the first time, the intersection 531 between the predicted value user group 52 and the current value user group 51, but at the second time, the predicted price The intersection 532 between the value user group 52 and the current value user group 51 is smaller than the intersection 531, that is, the number of users in the intersection becomes less, and the users in the intersection and the changes in the intersection can be counted on the users in it. Analysis to understand that the original customer (the value user group 51 in the intersection 531) may no longer meet the value user group 51 or is lost. Therefore, it is possible to further understand whether the marketing strategy for the value customer group 51 may be wrong or needs to be adjusted, It leads to the loss of users of the valuable user group, so it can be determined whether the marketing strategy needs to be adjusted.

Please refer to FIG. 5A and FIG. 5C at the same time. FIG. 5C is another example of the value user group and the current value user group obtained at a second time by the method or device for predicting value user groups based on short-term characteristics of users according to the new embodiment of the present invention. Schematic. At the first time, the intersection 531 between the predicted value user group 52 and the current value user group 51, but at the second time, the intersection 533 between the predicted value user group 52 and the current value user group 51 changes. Therefore, it can be known that the marketing strategy may have an effect, so the original marketing strategy can be maintained. In some embodiments, the prediction device 15 may also include or be electrically connected to a display module (not shown), and the display module may display user group graphs similar to those shown in FIGS. 5A-5C according to the data of the user groups 51 and 52 . The transformation is presented on the screen for monitoring the change between the predicted value user group 52 or the change between the current value user group 51 and the predicted value user group 52 . To put it simply, the new type of prediction device uses the short-term characteristics of users to predict the value user group for users. The intersection changes of the value user group and the current value user group at different time points can be monitored to determine whether the marketing strategy is not. Appropriately, that is, the forecasting method of the present novel embodiment further includes the step of monitoring the intersection changes to determine whether the marketing strategy is correct.

In some embodiments, the present invention can implement the modules or methods of the present disclosure by a computer software program, wherein the computer software program can be stored in a non-volatile storage medium, such as Such as flash memory devices, magnetic storage devices, optical discs or other types of storage media. For example, the device of the present invention can use a computer software program to fit a first user data set using a machine learning algorithm to select a user short-term feature set. The user short-term feature set may include at least one short-term feature. The first set of user data may include a set of long-term user characteristics with a first periodicity. The user long-term feature set may contain long-term feature data. The computer software program can perform feature engineering on the user data set to select the user short-term feature set. The feature engineering may include feature importance analysis as a basis for selecting the short-term feature set of the user. The feature engineering may include short-term feature extraction to extract the user's short-term feature set from the user's long-term feature set. The user short-term feature set may include user short-term feature data with a second periodicity. The first periodicity may include a first period, and the second periodicity may include a second period; for the same feature, the second period may be shorter than the first period. The computer software program can process the first user data set according to the user short-term feature set to generate a second user data set. Processing the first set of user data may include deleting features from the first set of user data that are outside the set of short-term features of the user. The computer software program can use the machine learning algorithm to fit the second user data set to generate a user group prediction model. The computer software program can use the user group prediction model to predict the third user data set to generate a first prediction result. The third set of user data may include user data of at least one specific user. The third set of user data may include the set of short-term characteristics of the user. In addition to the prediction result of whether a specific user belongs to a specific group, the first prediction result may further include a fourth user data set. The fourth user data set may include user data predicted by the user group prediction model as positive examples. The computer software program can perform union analysis on the first user data set and the fourth user data set to obtain a first union analysis result. The computer software program can perform data merging on the first user data set and the fourth user data set to obtain a fifth user data set Account data set. After generating the first prediction result, the computer software program may use the user group prediction model to predict the sixth user data set to generate a second prediction result. The sixth user data set may include user data of at least one specific user. The sixth set of user profiles may include the set of short-term characteristics of the user. In addition to the prediction result of whether a specific user belongs to a specific group, the second prediction result may further include a seventh user data set. The seventh user data set may include user data predicted by the user group prediction model as positive examples. The computer software program can perform data merging on the first user data set and the seventh user data set to obtain an eighth user data set. The computer software program can perform a union analysis on the first user data set and the seventh user data set to obtain a second union analysis result. The computer software program can perform an intersection comparison between the first union analysis result and the second union analysis result to generate an intersection comparison result, wherein the intersection comparison result can more truly reflect the situation of a particular group. User short-term characteristics, the computer software program can establish a classification model for judging or predicting whether the user belongs to the specific group according to the intersection comparison result.

In some embodiments, the present invention includes a computer program product that can be used for user data processing, wherein the computer software program can be stored in a non-volatile storage medium, such as a flash memory device, a magnetic storage device, an optical disk or a Other types of storage media, and the computer software program is loaded by the computer to predict and expand the user data set, and the steps include the following: performing feature engineering on the first user data set to make the first user data set. Including user data having a long-term user feature set, the user long-term feature set including features with a first periodicity, the first user data set including user data of a first number of users; using a machine model algorithm for the first user data set The user data set is fitted to generate a user short-term feature set, the user short-term feature set includes features with a second periodicity, wherein the first periodicity of the user long-term feature set has The first period, the second period of the user short-term feature set has a second period, and the first period is longer than the second period; the first user data set is feature selected according to the user short-term feature set to generate a second period a user data set, so that the second user data set contains user data with the short-term characteristics of the user, and the machine learning algorithm is used to fit the second user data set to generate a user cluster prediction model; using the user cluster The prediction model predicts the third user data set to generate a first user cluster prediction result, and the first user cluster prediction result includes a fourth user data set in addition to the prediction result of whether a specific user belongs to a specific ethnic group; and performing data merging on the first user data set and the fourth user data set to generate a fifth user data set, the fifth user data set includes the user data of the second number of users, and the first user data set in the fifth user data set 2. The number of users is higher than the number of first users in the first user data set.

In some embodiments, the user data set may include data of a plurality of different users, the feature data set may include a plurality of feature fields of different data attributes, each of the feature fields may include a plurality of feature values, each of the feature A plurality of eigenvalues of a field can constitute a eigenvalue range. The computer software program can calculate or select a characteristic threshold value according to the characteristic value range of each characteristic field. The data of the aforementioned feature data set may include feature threshold values. Each of the characteristic thresholds can be the maximum value, the minimum value, the mean value, the median, the mode, the percentile, the quartile, the standard deviation, and the like in the characteristic value range of each of the characteristic fields. According to the attributes of different eigenvalues, the computer software program can use different judging rules for the eigenvalues to determine the eigenvalues. For example, the computer software program can determine the correlation between the eigenvalue and the target variable or label fitted by the machine learning algorithm. If the correlation is determined to be a positive correlation, it can calculate or select the minimum value in the range of eigenvalues as the feature Threshold value, and if it is judged that the correlation is negative correlation, the maximum value in the range of eigenvalues can be calculated or selected as the characteristic threshold value. For example, if the target variable of the machine learning algorithm is consumption money Amount, the characteristic is the weekly login frequency, and the minimum value of the characteristic value range is 2. The computer software program can determine that the consumption amount has a positive correlation with the weekly login frequency, and can select the minimum value as the characteristic threshold value. In some instances, if there is a weak correlation between the target variable and the feature, the computer software program may calculate or select the arithmetic mean of the range of feature values as the feature threshold. In some instances, if the feature is a count attribute, the computer software program may calculate or select the mode of the feature value range as the feature threshold value. In some examples, the computer software program may set a threshold coefficient h=0,0.5,1,1.8,2,... and calculate the product of the eigenvalue range statistics and the threshold value coefficient as the eigenthreshold value ; For example, if h=0.5 is set, the characteristic threshold of the aforementioned weekly login frequency is 1. In some examples, the computer software program can receive a characteristic threshold setting command to set the characteristic threshold according to the setting command.

The function of the new model is that it can predict the user's value score (that is, predict whether the user belongs to a certain value user group) through the user's short-term characteristics of the user, such as daily or single consumption behavior, so it can be aimed at individual users. Allow users to respond to the current service more instantly and comprehensively, without having to wait for the final consumption result to respond. Further, through this new model, high-value users can be effectively maintained, and quasi-high-value users can be turned into high-value users as soon as possible, and even the deficiencies of existing customer groups such as existing RFM groups can be compensated, and quasi-high-value users can be regarded as high-value users. (That is, combine the predicted high-value user group with the high-value user group of the current group) to increase the amount and success rate of transactions and subscriptions.

The above description of "the present invention", the descriptions in the various embodiments, modified examples and the contents disclosed in the drawings are only an example for describing the new model described in the claims, and may also include different features or Various implementation combinations of the implementation aspects, so the new type described in the claims is not limited by the content disclosed in the above description, embodiments or drawings of the "new type". this application The description in the first claim is just an example, and the description in the claim can be appropriately changed according to the description in the description, drawings, and the like.

1: Value user grouping system

11: Platform device

111: Browsing Behavior

112: Interests

113: Device Information

114: Internet address (IP) or device identification

12: User management system

121: Demographics

122:Consumption record

123: Accounting records

124: Interactive Record

13: Open Source Device

131: Climatic Geography

132: Financial Economics

133: Humanities Population

134: Medical Biology

14:Database

15: Prediction device

151: Data processing module

152: Machine Learning Modules

153: Model Deployment Module

154: Marketing Module

Claims (10)

A prediction device for predicting whether a user belongs to a value user group based on short-term characteristics of the user, comprising electrically connecting a plurality of hardware circuits to each other, and configuring a plurality of modules, wherein: a data processing module in the plurality of modules group to obtain a grouping result, wherein the grouping result indicates that at least one user belongs to a value user group; a machine learning module in the multiple modules implements a feature engineering based on a machine learning algorithm to obtain the value user multiple user characteristics of the group, and establish a first grouping model of the valuable user group based on the multiple user characteristics, wherein at least one of the multiple user characteristics is a long-term user characteristic, and the user long-term characteristic is A feature that can only be obtained after reaching a target time; the machine learning module extracts a plurality of short-term features of users from the first grouping model, and uses the short-term features of the plurality of users to establish a second feature of the valuable user group a clustering model, wherein the user short-term characteristic is a characteristic that can be obtained before the target time is reached; and a model deployment module of the plurality of modules uses the second clustering model according to another A plurality of user short-term characteristics of a user predict whether the other user belongs to the value user group. The prediction device of claim 1, further comprising grouping the plurality of users according to the plurality of user data obtained by the plurality of users when the target time arrives, so as to obtain the grouping result. The prediction apparatus according to claim 1, wherein a plurality of user characteristics whose importance is the top or exceeds a threshold value in the valuable user group are selected as the plurality of user characteristics of the valuable user group. The prediction device according to claim 1, wherein a plurality of short-term features of users whose importance is the top or exceeds a threshold value in the first clustering model are selected as the multiple short-term features of the second clustering model. short-term characteristics of users. The prediction apparatus according to claim 1, wherein a plurality of short-term features of users whose importance is the top few or exceeding a threshold value in the first grouping model are selected, and a plurality of short-term features of the plurality of users are extended from the selected short-term features A plurality of short-term features of users are obtained, and the selected short-term features of users and the extended short-term features of users are used as the short-term features of users in the second grouping model. The prediction device of claim 1, wherein the prediction device predicts a plurality of users as belonging to the value user group using the second grouping model; and the prediction device monitors the predicted multiplicity of the value user group A temporal change of an intersection between a user and the multiple users of the valuable user group of the grouping result is used to judge whether the marketing strategy of the multiple users is correct. The prediction device of claim 1, wherein the prediction device predicts a plurality of users as belonging to the value user group using the second grouping model; and the prediction device predicts a plurality of users of the value user group The individual users are integrated with the plurality of users of the valuable user group of the grouping result as the plurality of users of the valuable user group. The prediction device of claim 1, wherein the prediction device updates the second grouping model during an event to supplement the plurality of short-term user features with at least one short-term user feature related to an event, and after the event holding period ends, the second grouping model is updated to remove the user short-term feature related to the event from the plurality of short-term user features. The prediction device of claim 1, wherein the prediction device prepares another second clustering model; the prediction device updates the other second clustering model and uses the other second clustering model during an event. A binary group model predicts whether the other user belongs to the value user group; and the prediction device does not update the second group model during the event period. A predictive device that predicts whether a user belongs to a valuable user group based on the short-term characteristics of the user. The device includes a plurality of hardware circuits that are electrically connected to each other. After the prediction device loads a computer program for predicting and expanding the user data set, it configures a plurality of modules that are signal-connected to each other. The modules include A data processing module and a machine learning module cooperate with each other to perform steps including: performing feature engineering on a first set of user data such that the first set of user data includes a set of long-term features of the user. user data, the user long-term feature set includes features with a first periodicity, the first user data set includes user data of a first number of users; a machine model algorithm is used to fit the first user data set to obtain generating a user short-term feature set, the user short-term feature set includes features with a second periodicity, wherein the first periodicity of the user long-term feature set has a first period, and the second periodicity of the user short-term feature set has a second periodicity period, the first period is longer than the second period; the first user data set is selected according to the short-term feature set of the user to generate a second user data set, so that the second user data set contains the short-term features of the user the user data set, use the machine learning algorithm to fit the second user data set to generate a user cluster prediction model; use the user cluster prediction model to predict the third user data set to generate a first user cluster a prediction result, the first user cluster prediction result includes a fourth user data set; and data merging of the first user data set and the fourth user data set to generate a fifth user data set, the fifth user data set User data including the second user data set, the second user data set of the fifth user data set is higher than the first user data set of the first user data set.

Images