CN115249081A

CN115249081A - Object type prediction method and device, computer equipment and storage medium

Info

Publication number: CN115249081A
Application number: CN202110462219.3A
Authority: CN
Inventors: 林岳
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-04-27
Filing date: 2021-04-27
Publication date: 2022-10-28

Abstract

The embodiment of the application discloses a method and a device for predicting object types, computer equipment and a storage medium; according to the method and the device, time series data of the target object in the historical time interval can be acquired; performing regression processing on the time series data to obtain regression characteristic information of the time series data; calculating data distribution characteristic information of the time sequence data in each time dimension based on data distribution of the time sequence data in at least one time dimension; calculating time sequence correlation characteristic information of the time sequence data based on the time sequence correlation relationship of the time sequence data in the historical time interval; and predicting the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information. The scheme can accurately predict the object type of an object based on time-series data of the object.

Description

Object type prediction method and device, computer equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method and an apparatus for predicting an object type, a computer device, and a storage medium.

Background

Time series data is data collected at different times, can be used to describe the time-varying phenomenon, and can reflect the time-varying state or degree of a certain thing or phenomenon. During the application use process, different time series data are correspondingly generated by different objects, so that the objects can be classified based on the time series data of the objects to determine the object types to which the objects belong. And the type prediction of the object is not only beneficial to the application to better know the object distribution, but also beneficial to the application to provide better service.

In the research and practice process of the related art, the inventor of the present application finds that the prediction of the type of the object based on the time series data is currently realized by a rule discrimination method, but the rule discrimination method has low efficiency and large prediction error, and cannot be further optimized in practical application, so that the method for predicting the type of the object based on the time series data still needs to be improved.

Disclosure of Invention

The embodiment of the application provides a method and a device for predicting an object type, a computer device and a storage medium, which can accurately predict the object type of an object based on time series data of the object.

The embodiment of the application provides a method for predicting an object type, which comprises the following steps:

acquiring time sequence data of a target object in a historical time interval;

performing regression processing on the time series data to obtain regression characteristic information of the time series data;

calculating data distribution characteristic information of the time sequence data in each time dimension based on data distribution of the time sequence data in at least one time dimension;

calculating time sequence correlation characteristic information of the time sequence data based on the time sequence correlation relationship of the time sequence data in the historical time interval;

and predicting the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information.

Correspondingly, an embodiment of the present application further provides an apparatus for predicting an object type, including:

the data acquisition unit is used for acquiring time series data of the target object in a historical time interval;

the regression processing unit is used for carrying out regression processing on the time series data to obtain regression characteristic information of the time series data;

a first calculation unit, configured to calculate data distribution characteristic information of the time series data in each time dimension based on data distribution of the time series data in at least one time dimension;

the second calculation unit is used for calculating time sequence correlation characteristic information of the time sequence data based on the time sequence correlation relationship of the time sequence data in the historical time interval;

and the type prediction unit is used for predicting the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information.

In one embodiment, the regression processing unit includes:

a coefficient determination subunit configured to determine a regression coefficient for the time-series data, wherein the regression coefficient characterizes a time-series correlation of the time-series data;

a model determination subunit configured to determine, based on the regression coefficient, a regression model required for performing regression processing on the time-series data;

and the regression processing subunit is used for performing regression processing on the time series data through the regression model to obtain regression characteristic information of the time series data.

In an embodiment, the coefficient determining subunit is configured to:

performing data conversion on the time sequence data to obtain a stable time sequence corresponding to the time sequence data; determining a set of candidate regression coefficients for the time series data based on the stationary time series, wherein the set of candidate regression coefficients comprises at least one set of candidate regression coefficients, each set of candidate regression coefficients corresponding to a regression model; performing model evaluation on the regression models corresponding to each group of candidate regression coefficients to obtain an evaluation result; determining a target regression coefficient from the candidate regression coefficient set based on the evaluation result, the target regression coefficient being a regression coefficient of the time-series data.

In one embodiment, the first computing unit includes:

the first obtaining subunit is configured to obtain a trained model corresponding to each time dimension;

and the first calculating subunit is configured to calculate, based on the data distribution of the time series data in each time dimension, data distribution feature information of the time series data in the time dimension through a trained model corresponding to the time dimension.

In an embodiment, the time dimension includes a trending time dimension, and the trained model includes a trend prediction model corresponding to the trending time dimension; the first obtaining subunit is configured to:

acquiring a sample data set required by model training; determining model type information and model parameter information; determining a trend prediction model to be trained based on the model type information and the model parameter information; and performing model training on the trend prediction model to be trained through the sample data set to obtain the trained trend prediction model.

In an embodiment, the time dimension includes at least one mutation time dimension, and the trained model includes a mutation prediction model corresponding to each mutation time dimension; the first computing subunit is configured to:

determining time window information of the catastrophe time dimension, wherein the time window information represents the duration of a time window corresponding to the catastrophe time dimension; calculating, by the mutation prediction model, a mutation distribution characteristic of the time-series data within the time window based on a data distribution of the time-series data over the mutation time dimension; and performing characteristic combination on the mutation distribution characteristic information to obtain data distribution characteristic information of the time series data on the mutation time dimension.

In an embodiment, the time dimension includes at least one mutability time dimension, the time dimension includes a periodicity time dimension, and the trained model includes a periodicity prediction model corresponding to the periodicity time dimension; the first computing subunit is configured to:

decomposing the time sequence data through the periodic prediction model based on the data distribution of the time sequence data on the periodic time dimension to obtain the periodic distribution characteristics of the time sequence data; and smoothing the periodic distribution characteristic to obtain data distribution characteristic information of the time series data on the periodic time dimension.

In one embodiment, the second computing unit includes:

the second acquisition subunit is used for acquiring a trained time sequence correlation prediction model, wherein the time sequence correlation prediction model is used for calculating time sequence correlation characteristic information of time sequence data;

the relation analysis subunit is used for analyzing the time sequence incidence relation of the time sequence data in the historical time interval through the time sequence incidence prediction model;

and the second calculating subunit is used for calculating the time sequence related characteristic information of the time sequence data based on the analysis result.

In one embodiment, the type prediction unit includes:

the characteristic fusion subunit is used for carrying out characteristic fusion on the regression characteristic information, the data distribution characteristic information corresponding to each time dimension and the time sequence correlation characteristic information to obtain fused characteristics;

and the type prediction subunit is used for predicting the object type of the target object in a preset time interval based on the fused features.

In one embodiment, the feature fusion subunit is configured to:

performing feature fusion on the data distribution features corresponding to the time dimensions to obtain fused data distribution feature information; respectively determining the weights of the regression feature information, the time sequence correlation feature information and the fused data distribution feature information; and performing feature fusion on the regression feature information, the time sequence correlation feature information and the fused data distribution feature information based on the weight to obtain fused features.

Accordingly, the present application also provides a storage medium having a computer program stored thereon, where the computer program is executed by a processor to implement the steps of the object type prediction method as shown in the present application.

Accordingly, the present application further provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the object type prediction method according to the present application.

According to the method and the device, time series data of the target object in the historical time interval can be acquired; performing regression processing on the time series data to obtain regression characteristic information of the time series data; calculating data distribution characteristic information of the time-series data in each time dimension based on data distribution of the time-series data in at least one time dimension; calculating time sequence correlation characteristic information of the time sequence data based on the time sequence correlation relationship of the time sequence data in the historical time interval; and predicting the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information.

According to the scheme, regression characteristic information can be obtained by performing regression processing on time series data of the target object, and specifically, the regression characteristic information can be used for predicting the object type of the target object in a short period; predicting the object type of the target object in a long term from a plurality of time dimensions by calculating data distribution characteristic information of the time-series data of the target object on the plurality of time dimensions; by calculating time sequence correlation characteristic information of the time sequence data of the target object, the type of the object to which the target object belongs can be predicted through a neural network model. Furthermore, the scheme can perform feature fusion on the obtained feature information, so that the short-term prediction result, the long-term prediction result and the neural network prediction result of the target object are fused, and the object type of the target object is comprehensively predicted based on the fusion result. In this way, the scheme avoids the situation that the index required for predicting the object type needs to be manually adjusted in the rule-based judging process, and can accurately and efficiently predict the object type of the object based on the time series data of the object by a data mining-based method.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

Fig. 1 is a scene schematic diagram of a prediction method for an object type according to an embodiment of the present application;

FIG. 2 is a flowchart of a method for predicting object types according to an embodiment of the present disclosure;

FIG. 3 is a schematic diagram of a model structure of a prediction method for an object type according to an embodiment of the present disclosure;

FIG. 4 is a schematic diagram of another model structure of a prediction method for object types according to an embodiment of the present disclosure;

FIG. 5 is a schematic diagram of another model structure of a prediction method for object types according to an embodiment of the present disclosure;

FIG. 6 is a schematic diagram of another model structure of a prediction method for object types according to an embodiment of the present disclosure;

FIG. 7 is a schematic diagram of another model structure of a prediction method for object types according to an embodiment of the present disclosure;

FIG. 8 is a schematic diagram of another model structure of a prediction method for object types according to an embodiment of the present disclosure;

FIG. 9 is a schematic diagram of another model structure of a prediction method for object types according to an embodiment of the present disclosure;

FIG. 10 is a schematic flowchart of another method for predicting object types according to an embodiment of the present disclosure;

fig. 11 is a schematic structural diagram of an object type prediction apparatus according to an embodiment of the present application;

fig. 12 is another schematic structural diagram of an object type prediction apparatus according to an embodiment of the present application;

fig. 13 is another schematic structural diagram of an object type prediction apparatus according to an embodiment of the present application;

fig. 14 is another schematic structural diagram of an object type prediction apparatus provided in an embodiment of the present application;

fig. 15 is another schematic structural diagram of an object type prediction apparatus according to an embodiment of the present application;

fig. 16 is a schematic structural diagram of a computer device provided in an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

The embodiment of the application provides a method and a device for predicting an object type. Specifically, the embodiment of the application provides an object type prediction device suitable for computer equipment. The computer device may be a terminal or a server, and the terminal may be a mobile phone, a tablet computer, a notebook computer, and the like. The server may be a single server or a server cluster composed of a plurality of servers.

The embodiment of the application takes a prediction device of an object type as an example, and introduces a prediction method of the object type.

Referring to fig. 1, the server 10 may acquire time-series data of the target object within the history time interval, for example, the server 10 may acquire time-series data of the target object within the history time interval by receiving the time-series data transmitted by the terminal 20.

Further, the server 10 may perform regression processing on the time series data to obtain regression feature information of the time series data; calculating data distribution characteristic information of the time sequence data in each time dimension based on the data distribution of the time sequence data in at least one time dimension; and calculating time-series correlation characteristic information of the time-series data based on the time-series correlation of the time-series data in the historical time interval. Moreover, the server 10 may predict the object type of the target object within the preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time-series correlation feature information obtained through calculation.

Alternatively, the server 10 may send the prediction result to the terminal 20, so that the terminal 20 may further perform operations such as application management and product operation of the service based on the prediction result.

The following are detailed below. It should be noted that the following description of the embodiments is not intended to limit the preferred order of the embodiments.

With the research and progress of artificial intelligence technology, the artificial intelligence technology is developed and applied in a plurality of fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, unmanned aerial vehicles, robots, smart medical care, smart customer service, and the like.

Among them, artificial Intelligence (AI) is a theory, method, technique and application system that simulates, extends and expands human Intelligence using a digital computer or a machine controlled by a digital computer, senses the environment, acquires knowledge and uses the knowledge to obtain the best result. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the implementation method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence base technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.

The method for predicting the object type relates to the technologies such as artificial intelligence machine learning and the like, and can be executed by a terminal or a server or by the combination of the terminal and the server; in the embodiment of the present application, a prediction method of an object type is performed by a server as an example, specifically, performed by a prediction apparatus of an object type integrated in a server, as shown in fig. 2, a specific flow of the prediction method of an object type may be as follows:

101. and acquiring time sequence data of the target object in the historical time interval.

The object is a usage entity of the application or service, for example, the object may include a usage client of the application or service, such as an individual user, an organization, and the like.

In an embodiment, the target object may be an object of an object type to be predicted, for example, the target object may be a user of a user type to be predicted, that is, a target user; as another example, the target object may be an organization of the tissue type to be predicted, i.e., a target tissue; and so on.

The time sequence is a sequence formed by arranging the numerical values of the same statistical index according to the occurrence time sequence. The main purpose of time series analysis is to predict the future based on existing historical data. Time series data is data collected at different times for situations where the described phenomenon varies over time. Such data reflects the state or extent of change of an object, phenomenon, etc. over time.

In one embodiment, the target object may be a target user, and whether the target user of the game application a belongs to a high quality user may be predicted by the method described herein, for example, a user more prone to pay in the game application a may be regarded as a high quality user, and a user not prone to pay in the game application a may be regarded as a low quality user. The time series data for the target user may be the amount of money the target user spends in the gaming application each day for some 30 days in the past.

In another embodiment, the target object may be a target user, and whether the target user of the financial application a belongs to a good user may be predicted by the method described herein, and particularly, for a financial application, it is often necessary to assess user qualification to determine what risk level service is provided to the user, such as providing a higher risk financial service, such as a loan, for a good user and providing a low risk financial service or not providing a risk financial service for a bad user. The time series data of the target user may include financial behavior information of the target user in the financial application each day for the past half of the year, for example, the financial behavior information may include an amount of loan, an amount of spending, and the like.

The historical time interval is a past time range, for example, the historical time interval may be within the last 7 days, within 30 days of a past month, and the like. It should be noted that the historical time interval may be a continuous time interval or a discontinuous time interval.

The method for acquiring the time-series data of the target object in the historical time interval can be various, for example, a terminal can run a related application, and the terminal can send the time-series data of the target object in the historical time interval to the server, so that the server can acquire the time-series data and predict the object type of the target object based on the time-series data subsequently; for another example, the server may send a request to the database to obtain time-series data of the target object in the historical time interval through the request; and so on.

102. And performing regression processing on the time series data to obtain regression characteristic information of the time series data.

Regression is a statistical analysis method for determining the interdependent quantitative relationships between two or more variables. The regression processing can be carried out in various ways, for example, the regression processing can be divided into univariate regression and multivariate regression analysis according to the number of variables involved; according to the dependent variable, the method can be divided into simple regression analysis and multiple regression analysis; according to the relationship type between independent variable and dependent variable, the method can be divided into linear regression analysis and nonlinear analysis; the method can be divided into linear regression (dependent variable is continuous variable) and logistic regression (dependent variable is logical variable) according to whether the dependent variable is continuous or not; and so on.

The regression feature information is feature information obtained after regression processing.

For example, after a regression model required for performing regression processing on the time series data is determined, regression processing may be performed on the time series data through the regression model to obtain regression feature information, where the regression feature information may be used to predict an object type to which a target object belongs in a short period. Specifically, the step of performing regression processing on the time series data to obtain regression feature information of the time series data may include:

determining a regression coefficient for the time series data, wherein the regression coefficient characterizes a time-series correlation of the time series data;

determining a regression model required for performing regression processing on the time series data based on the regression coefficient;

and performing regression processing on the time series data through a regression model to obtain regression characteristic information of the time series data.

The regression coefficient is a correlation coefficient required for performing regression processing on the time-series data, and in an embodiment, the regression coefficient may include a correlation coefficient of the time-series data, such as an autocorrelation coefficient, a partial autocorrelation coefficient, and the like. In particular, the correlation coefficient measure refers to the degree of interaction between two different events with each other; the autocorrelation coefficient measures the degree of correlation of the same event between two different periods, and vividly measures the influence of the past behavior of the object on the current object; the partial autocorrelation coefficients are used for removing the influence of some variables and then the autocorrelation coefficients are examined.

The time sequence correlation refers to the correlation degree of the variables corresponding to the statistical indexes in the time sequence at different periods.

There may be various methods for determining the regression coefficient of the time-series data, for example, in an embodiment, the regression coefficient to be determined may include an autocorrelation coefficient and a partial autocorrelation coefficient of the time-series data, and specifically, the step "determining the regression coefficient of the time-series data" may include:

performing data conversion on the time series data to obtain a stable time series corresponding to the time series data;

determining a set of candidate regression coefficients for the time series data based on the stationary time series, wherein the set of candidate regression coefficients comprises at least one set of candidate regression coefficients, each set of candidate regression coefficients corresponding to one regression model;

performing model evaluation on the regression models corresponding to each group of candidate regression coefficients to obtain an evaluation result;

based on the evaluation result, a target regression coefficient is determined from the candidate regression coefficient set, and the target regression coefficient is taken as the regression coefficient of the time-series data.

The time sequence is a sequence formed by arranging numerical values of a certain statistical index of a certain phenomenon on different times according to a time sequence. For a time series, if the mean value has no systematic variation (no trend), the variance has no systematic variation, and the periodic variation is strictly eliminated, the time series can be considered to be stationary, that is, the time series can be considered to be stationary time series.

Wherein, the data conversion refers to a data processing procedure of converting the original time series data into a corresponding smooth time series.

For example, the time series data may be subjected to data mapping to observe whether the time series data is a stationary time series, and if the time series data is not a stationary time series, the time series data may be subjected to a d-order difference operation to convert the time series data into a stationary time series. In particular, a difference operation is also called a difference function, which is a concept in mathematics. It may map the primitive function f (x) to f (x + a) -f (x + b). The differential operation corresponds to a differential operation, and is an important concept in the calculus, and in practical application, the differential operation may include a forward differential and an inverse differential.

After obtaining a stationary time series corresponding to the time series data, a set of candidate regression coefficients for the time series data may be determined based on the stationary time series, where the set of candidate regression coefficients includes at least one set of candidate regression coefficients, and each set of candidate regression coefficients corresponds to a regression model.

In an embodiment, the autocorrelation coefficient and the partial autocorrelation coefficient of the stationary time series corresponding to the time series data may be obtained respectively, and a candidate rank p and an order q may be obtained by analyzing the autocorrelation map and the partial autocorrelation map, and then a candidate regression coefficient set of the time series data may be determined according to the candidate rank p and the order q. For example, for a certain time series data, by analyzing the autocorrelation graph and partial autocorrelation graph of the time series data, it can be determined that the autocorrelation graph shows that three orders of lag exceed the confidence boundary, and the correlation graph shows that the partial autocorrelation coefficients at 1 to 7 orders of lag exceed the confidence boundary, so that the set of candidate regression coefficients of the time series data can be determined as { (0,1), (7,0), (7,1) }, which includes three sets of candidate regression coefficients, each set of candidate regression coefficients has the format (p, q), and each set of candidate regression coefficients corresponds to one regression model.

As an example, each set of candidate regression coefficients may correspond to a differential Integrated Moving Average Autoregressive model (ARIMA), wherein the ARIMA model is one of the time series prediction analysis methods. Specifically, in ARIMA (p, d, q), AR is "autoregressive", and p is the number of autoregressive terms; MA is "moving average", q is the number of terms of the moving average, and d is the number of differences (order) made to make it a stationary sequence. Thus, for a set of candidate regression coefficients { (0,1), (7,0), (7,1) }, the following regression model can be determined: ARMA (0,1), ARMA (7,0) and ARMA (7,1).

Further, model evaluation may be performed on the regression model corresponding to each set of candidate regression coefficients, so that the target regression coefficient can be determined from the set of candidate regression coefficients based on the evaluation result.

For example, for an ARIMA model, the quality of the obtained p and q parameter values can be judged by adopting an Akaike Information Criterion (AIC) or a Bayesian Information Criterion (BIC). The AIC information criterion is established on the basis of the concept of entropy, and the complexity of an estimated model and the goodness of the model fitting data can be balanced; the BIC information criterion is that under incomplete information, partial unknown states are estimated by subjective probability, then occurrence probability is corrected by Bayes, and finally an optimal decision is made by using an expected value and correction probability;

it should be noted that, in practical applications, the model evaluation may be performed by combining the AIC information criterion with the BIC information criterion, and other information criteria for other evaluations may also be introduced.

In one embodiment, for the following regression models ARMA (0,1) corresponding to the candidate regression coefficient set { (0,1), (7,0), (7,1) }, ARMA (7,0) and ARMA (7,1) find that both AIC and BIC of the model ARMA (7,0) are minimum, it can be determined (7,0) as the target regression coefficient, and the target regression coefficient can be used as the regression coefficient of the time-series data.

After the regression coefficient of the time series data is determined, the regression model corresponding to the regression coefficient may be determined as the regression model required for performing the regression processing on the time series data, and the time series data may be subjected to the regression processing by the determined regression model, and the output result of the regression model may be used as the regression feature information of the time series data. For example, for the following regression models ARMA (0,1), ARMA (7,0) and ARMA (7,1) corresponding to the candidate regression coefficient set { (0,1), (7,0), (7,1) }, it can be determined that the model ARMA (7,0) is a regression model required for the regression processing of time-series data. Further, regression characteristic information of the time series data can be obtained by subjecting the time series data to regression processing by a model ARMA (7,0).

103. And calculating data distribution characteristic information of the time series data in each time dimension based on the data distribution of the time series data in at least one time dimension.

The time dimension is a feature type to which the time sequence data belong is divided from the time perspective, so that feature information of the time sequence data on different time dimensions can be obtained by analyzing the time sequence data from different time dimensions.

For example, the time dimension can include a trending time dimension, a periodic time dimension, and a catastrophe time dimension, among others. Specifically, analyzing the time series data from the trending time dimension can analyze how the whole time series data grows, and predict how the whole time series data will grow in the future, namely, the whole growth trend in the future; analyzing the time series data from a periodic time dimension, the time series data can be analyzed for periodic trends, such as monthly trends, quarterly trends, yearly trends, and the like; by analyzing the time series data from the time dimension of the mutation, the time series data can be analyzed for the trend of the change in some mutational time, for example, the trend of the change of the time series data in holidays or some large events.

The data distribution characteristic information is characteristic information describing a data distribution of the time-series data in a time dimension, for example, a data distribution characteristic of the time-series data in a trending time dimension, and may be characteristic information describing a data distribution of the time-series data in a trending time dimension, for example, the characteristic information may be related information indicating how the entire time-series data grows and how the data will grow in a future time. For another example, the data distribution characteristic of the time series data in the periodic time dimension may be characteristic information describing the data distribution of the time series data in the periodic time dimension, for example, the characteristic information may be related information characterizing a periodic trend of the time series data, and the periodic trend may be, for example, a monthly trend, a quarterly trend, a yearly trend, or the like; for another example, the data distribution characteristic of the time series data in the mutation time dimension may be characteristic information describing the data distribution of the time series data in the mutation time dimension, for example, the characteristic information may be related information characterizing a trend of the time series data in some mutation times, and the mutation times may be holidays or times of some large events, for example.

The expression form of the data distribution characteristic information may be various, for example, a numerical value obtained after calculation; as another example, the function may be a function obtained by fitting time-series data, and the like.

Since the data distribution characteristic information of the time series data in each time dimension is calculated by considering the data distribution of the time series data in a plurality of time dimensions in the present application, a corresponding model can be trained for each time dimension, so that the data distribution characteristic information of the time series data in the corresponding time dimension can be calculated by the trained model corresponding to each time dimension. Specifically, the step "calculating data distribution characteristic information of the time-series data in each time dimension based on the data distribution of the time-series data in at least one time dimension" may include:

acquiring a trained model corresponding to each time dimension;

and calculating data distribution characteristic information of the time sequence data in each time dimension through a trained model corresponding to the time dimension based on the data distribution of the time sequence data in the time dimension.

In one embodiment, referring to fig. 3, the time series data may be analyzed in four time dimensions: a trending time dimension, a periodic time dimension, a mutating time dimension, and a random time dimension. The trained model corresponding to the trend time dimension can be represented by g (t) in fig. 3, and is used for fitting the aperiodic change in the time sequence and judging that the time sequence is in an ascending or descending trend; the model after training corresponding to the periodic time dimension can be represented by p (t) in fig. 3, and is used for fitting the periodic variation in the time series, such as variation trend of each week, each month, each season, and the like; the trained model corresponding to the catastrophe time dimension can be identified by h (t) in fig. 3, and is used for fitting the influence of the potential holidays with non-fixed periods and the change points on the predicted value, such as holidays, application operation activity days and the like; the trained model for the random time dimension can be represented by ε (t) in FIG. 3 to fit unpredicted random fluctuations, e.g., ε (t) can be subject to a Gaussian distribution.

Further, in this embodiment, the data distribution characteristic information of the time series data in each time dimension may be calculated through a trained model corresponding to the time dimension based on the data distribution of the time series data in the time dimension. Specifically, the time series data may be input to each trained model as the independent variable t, an output value of each trained model is obtained, and the output value is used as data distribution characteristic information of the time series data in a corresponding time dimension.

In an embodiment, the trained model corresponding to the trending time dimension may be used as a trend prediction model, and specifically, the step "obtaining the trained model corresponding to each time dimension" may include:

acquiring a sample data set required by model training;

determining model type information and model parameter information;

determining a trend prediction model to be trained based on the model type information and the model parameter information;

and performing model training on the trend prediction model to be trained through the sample data set to obtain the trained trend prediction model.

The sample data set is a set of sample data, and the sample data is composed of time series data of the sample object in a historical time interval. Since the sample is a part of individual observation or investigation and is the whole of the study object as a whole, the sample object is a part of observation or adjustment. For example, the method described in the present application may be applied to predict the user type to which the target user of the game application a belongs, and then the sample object of the game application a may be a sample user of the game application a, that is, a part of users of the game application a; as another example, the target application may be application B that provides a service to a customer, where the customer may include an individual user, an organization, etc., and the sample object for application B may be a sample customer of application B, i.e., a partial customer of application B.

The model type information is the relevant information of the model type required to be determined when describing and establishing the trend prediction model, so that the established trend prediction model is the model under the model type. For example, when analyzing time series data from a trending time dimension, two types of models may be provided for modeling: non-linear growth models and linear growth models.

The model parameter information is relevant information of model parameters required to be determined when the trend prediction model is established, so that the established trend prediction model is established based on the model parameters. For example, if the model type is a logistic regression model in the category of nonlinear growth models, the model parameter information may include growth maximum information, growth rate information, offset information, and the like; for another example, if the model type is a segmented linear model in the linear growth model category, the model parameter information may include growth rate information, offset information, segmentation point information, and the like.

After the model type information and the model parameter information are determined, the trend prediction model to be trained can be determined based on the determined model type information and the determined model parameter information. Further, model training can be performed on the trend prediction model to be trained through the sample data set, so that the trained trend prediction model is obtained.

For example, for a segmented linear model in a linear growth model category, the model hyper-parameters to be tuned may include the number of segmentation points, and specifically, the larger the value of the number of segmentation points is, the stronger the fitting degree of the trained trend prediction model to the sample data set is, but the risk of overfitting is also increased, so that the hyper-parameters in the model may be tuned by the sample data set to perform model training, thereby obtaining the trained trend prediction model.

It should be noted that, in the above example, the time dimension is taken as the trend time dimension, and the step of "obtaining the trained model corresponding to the time dimension" is explained, in practical application, when the time dimension is other conditions, for example, when the time dimension is a periodic time dimension, a catastrophe time dimension, or a random time dimension, the step of "obtaining the trained model corresponding to the time dimension" may refer to a mode when the time dimension is a periodic time dimension.

In an embodiment, the time dimension may include at least one mutagenic time dimension, for example, the mutagenic time dimension is a time series data analyzed from holidays or application activity days, so that different holidays or application activity days may correspond to different mutagenic time dimensions, and thus, a trained model corresponding to each mutagenic time dimension may be used as a mutation prediction model corresponding to the mutagenic dimension, and in particular, the step "calculating data distribution characteristic information of the time series data in the time dimension based on the data distribution of the time series data in each time dimension through the trained model corresponding to the time dimension" may include:

determining time window information of a catastrophe time dimension, wherein the time window information represents the duration of a time window corresponding to the catastrophe time dimension;

calculating the mutation distribution characteristics of the time sequence data in a time window through a mutation prediction model based on the data distribution of the time sequence data in the mutation time dimension;

and performing characteristic combination on the mutation distribution characteristic information to obtain data distribution characteristic information of the time series data on the mutation time dimension.

The time window information of the catastrophe time dimension represents the duration of the time window corresponding to the catastrophe time dimension, that is, the time window information is used for setting a time window for each catastrophe prediction model so as to fit the influence of holidays or application operation activity days on a target object in the real world, which often has a corresponding window period. For example, in a scene of user consumption, taking a valentine's day as an example, the influence of the valentine's day on the user consumption has a window period, for example, the first few days and the last few days of the valentine's day.

Further, a mutation distribution characteristic of the time series data within a corresponding time window can be calculated through a mutation prediction model corresponding to the mutation time dimension based on the data distribution of the time series data in the mutation time dimension.

In an embodiment, the catastrophe time event corresponding to the catastrophe dimension may be taken as an example for explanation, and the catastrophe prediction model may calculate the catastrophe distribution characteristics of the time series data within the time window by calculating the influence of the holiday on the time series data within the time window, specifically, since in practical application, the holiday may have a great influence on the time series, the influence calculated here is the influence of the holiday in the time window on the final predicted value of the time series data within the preset time interval.

As an example, the mutation prediction model may set the influence in the same window period to the same value, and for another example, the mutation prediction model may calculate the influence on each day in the window period by following a normal distribution; and so on. For example, for a certain holiday, which may be counted in days, and the time window set for it may be 7 days, the mutagenicity distribution characteristics of the time series data for each of the 7 days of the window period may be calculated by the mutation prediction model, respectively, for example, the influence of each day within the window period may be set to the same value; as another example, the effect on each day during the window period can be set following a normal distribution; and so on. After calculating the distribution characteristics of the mutation of the time sequence data in the time window period every day, the time characteristic of the mutation of the time sequence data in the time window can be determined.

After the mutation distribution characteristics of the time series data in the time window are obtained, the data distribution characteristic information of the time series data in the mutation time dimension can be determined by performing characteristic combination on the mutation distribution characteristic information.

For example, when the expression form of the mutation distribution characteristic information is a specific numerical value, the mutation distribution characteristics may be combined by adding, multiplying, or performing a weighted calculation on the mutation distribution characteristics.

It should be noted that, in this embodiment, in consideration of differences between time windows and influence degrees of different mutational events, for example, different holidays or application operation activity days, different mutational prediction models may be set for different mutational time dimensions, so that influences of different mutational events at different time points may be used as independent models.

In an embodiment, a trained model corresponding to a periodic time dimension may be used as the periodic prediction model, and specifically, the step "calculating data distribution characteristic information of the time series data in each time dimension through the trained model corresponding to the time dimension based on the data distribution of the time series data in the time dimension" may include:

decomposing the time sequence data through a periodic prediction model based on the data distribution of the time sequence data in a periodic time dimension to obtain the periodic distribution characteristics of the time sequence data;

and smoothing the periodic distribution characteristic to obtain data distribution characteristic information of the time series data on the periodic time dimension.

Since the time series data may include periodic variation trends of various periodic types, for example, periodic variation trends of each week, each month, each season, etc., the periodic distribution characteristics of the time series data in the periodic time dimension may be obtained by decomposing the time series data.

There are various ways to decompose the time series data, for example, since any periodic function can be represented by an infinite series composed of a sine function and a cosine function, in an embodiment, the time series data can be decomposed by a fourier series, and specifically, the corresponding periodic property can be fitted by the fourier series to obtain a fitting function of the time series data, for example, the generated fitting function can be s ' (t), where s ' (t) can receive a past or future time point t ' as an argument and output a corresponding s ' (t ') value as a predicted value. In this embodiment, the fitting function may be characterized as a periodic distribution of the time series data.

The smoothing processing is used for adjusting the smoothness of the features, after the periodic distribution features of the time series data are obtained, the smoothing processing can be performed on the periodic distribution features, and the processed data distribution features are used as expressions of the data distribution features of the time series data in the periodic time dimension to calculate the data distribution features of the time series data in the periodic time dimension in a preset time interval.

In an embodiment, the periodic distribution characteristic of the time-series data may be s '(t), and s' (t) may be smoothed by the following equation to obtain processed s (t), and s (t) is taken as data distribution characteristic information of the time-series data in the periodic time dimension: s (t) = s' (t) β, where β to Normal (0, σ).

As an example, in practical applications, the relevant steps in step 103 may be performed by a prophet model, where the prophet is an open-source time series framework that can be used to process time series data.

104. And calculating time-series correlation characteristic information of the time-series data based on the time-series correlation of the time-series data in the historical time interval.

And the time sequence incidence relation of the time sequence data in the historical time interval is used for describing the incidence relation of the time sequence data among different time nodes of the historical time interval. And the time sequence correlation characteristic information is the characteristic information calculated based on the time sequence correlation of the time sequence data.

The time-series correlation characteristic information may be calculated based on the time-series correlation, for example, by a machine learning model, such as a long short-Term Memory artificial neural network (LSTM).

Machine Learning (ML) is a multi-domain cross subject, and relates to multiple subjects such as probability theory, statistics, approximation theory, convex analysis and algorithm complexity theory. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

The LSTM is a time-cycle neural network, see fig. 4, the model on the right side in fig. 4 is a model structure of the LSTM model, and the model on the left side is a model structure of a Recurrent Neural Network (RNN), and since the LSTM implements analysis of a time sequence association relationship of time sequence data in a historical time interval through a gate structure, the LSTM can be used to solve a long-term dependence problem existing in a general Recurrent neural network.

Specifically, x in FIG. 4 ^t For the entry of time-series data in the current time state, h ^t-1 Indicating the input received from the previous node, y ^t Is the output at the current node state, and h ^t Is an output that is passed to the next node. Having only one delivery state h compared to RNN ^t LSTM has two transmission states: c. C ^t And h ^t Wherein for c passed down ^t Change very slowly, c of the usual output ^t Is c from the last state ^t-1 Plus some value, and h ^t There will often be a large difference between different nodes.

Since the LSTM can perform better in a longer sequence than a general RNN, specifically, the step of "calculating the time-series correlation characteristic information of the time-series data based on the time-series correlation relationship of the time-series data in the historical time interval" may include:

acquiring a trained time sequence correlation prediction model, wherein the time sequence correlation prediction model is used for calculating time sequence correlation characteristic information of time sequence data;

analyzing a time sequence incidence relation of the time sequence data in a historical time interval through a time sequence incidence prediction model;

based on the analysis result, time-series correlation characteristic information of the time-series data is calculated.

In an embodiment, the trained LSTM model may be used as a time-series correlation prediction model required for calculating time-series correlation characteristic information, and further, the time-series correlation of the time-series data in the historical time interval may be analyzed through the LSTM model.

For example, the trained LSTM model may be a chain structure with repeated neural network modules, and the repeated neural network modules in the LSTM model may refer to fig. 5, in which a gate structure for processing time series data may be included, and in particular, the gate structure may implement selective extraction of information. For example, referring to fig. 6, the basic gate structure may be implemented by a sigmoid neural layer and a point-by-point multiplication operation, specifically, each element output by the sigmoid layer is a real number between 0 and 1, which may be used to characterize the weight for information extraction, for example, 0 means no information is extracted, and 1 means all information is extracted. In the LSTM model, the time-series correlations of time series data over historical time intervals can be resolved based on the basic gate structure.

As an example, the LSTM model may include the following three composite gate structures determined based on the basic gate structure: a forgetting gate structure, an input gate structure, and an output gate structure. Wherein a forgetting gate structure can be used to derive time-series data x ^t Can be used to selectively filter partial information, an input gate structure can be used to filter partial information from time series data x ^t To selectively extract portions of the information, the output gate structure may be used to determine the information output by the neural network module. In this way, the time sequence correlation of the time sequence data in the historical time interval can be analyzed through the composite gate structures.

In particular, the forgetting gate structure can read h ^t-1 And x ^t And outputting a value between 0 and 1 to determine the right to retain timing informationHeavy, e.g., 1 means that all information is retained, i.e., no information is filtered; 0 means that no information is retained, i.e. all information is filtered. In one embodiment, the implementation mechanism of the forgetting gate structure can refer to the following formula: f. of ^t ＝σ(W ^f ·[h ^t-1 ，x ^t ]+b ^f ) Where σ denotes a sigmoid function, W ^f And b ^f The post-training parameters are indicated. Referring to fig. 7, an illustration of the structure of a forgetting gate can be highlighted on the basis of fig. 5.

In particular, the input gate structure may be used to derive time series data x from ^t For example, the implementation mechanism of the input gate structure may refer to the following equation:

i ^t ＝σ(W ⁱ ·[h ^t-1 ，x ^t ]+b ⁱ )

where σ denotes a sigmoid function, W ⁱ 、W ^C 、b ⁱ And b ^C The post-training parameters are indicated. Referring to fig. 8, a schematic of the input gate structure is highlighted on the basis of fig. 5.

In particular, the output gate mechanism may be used to determine the information output by the neural network module, for example, the implementation mechanism of the output gate structure may refer to the following equation:

where σ denotes a sigmoid function, W ^o And b ^o The post-training parameters are indicated. Referring to fig. 9, a schematic of the output gate structure is highlighted on the basis of fig. 5.

In this way, the time-series data x can be analyzed by the LSTM model ^t The time sequence correlation relationship in the historical time interval, because the LSTM model after training will be based on the time sequence data x ^t Analysis result of (2) to generateAnd the final output result is obtained, so that the final output result of the trained LSTM model can be used as the time sequence related characteristic information of the time sequence data.

It should be noted that the present application does not limit the execution sequence among step 102, step 103, and step 104, and step 102, step 103, and step 104 may or may not be performed simultaneously, and the present application does not limit this.

105. And predicting the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information.

For example, when the target object is a target user, in a scenario of a financial application, the user may be divided into two types, one type is a poor user who strictly rejects and does not provide a risk service (hereinafter referred to as "strict user"), and the other type is a non-strict good user, so in this example, the user type may include a good user type and a poor user type; for another example, in the context of a game application, users who are more inclined to pay in the game application may be considered high-quality users, and users who are not inclined to pay in the game application may be considered low-quality users, and thus, in this example, the user types may include a high-quality user type and a low-quality user type; and so on.

The preset time interval is a time interval in which the type of the target object is to be predicted, and may be, for example, 3 months in the future. It should be noted that the preset time interval may be a continuous time interval or a discontinuous time interval.

Based on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time sequence correlation feature information, there are various ways to predict the object type of the target object in the preset time interval, for example, feature fusion may be performed on the feature information to obtain a fused feature, and the object type of the target object is determined based on the fused feature. Specifically, the step of predicting the object type of the target object in the preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time-series correlation feature information may include:

performing feature fusion on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information to obtain fused features;

and predicting the object type of the target object in a preset time interval based on the fused features.

The feature fusion refers to processing the plurality of feature information to obtain a fused feature referring to the plurality of feature information. The feature fusion method may be various, and for example, the feature information may be added, multiplied, or weighted. Specifically, the step of performing feature fusion on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time sequence correlation feature information to obtain fused features may include:

performing feature fusion on the data distribution features corresponding to each time dimension to obtain fused data distribution feature information;

respectively determining the weights of the regression feature information, the time sequence correlation feature information and the fused data distribution feature information;

and performing feature fusion on the regression feature information, the time sequence correlation feature information and the fused data distribution feature information based on the weight to obtain fused features.

Similarly, there may be a plurality of ways to perform feature fusion on the data distribution features corresponding to each time dimension, for example, referring to fig. 3, in an embodiment, the data distribution features corresponding to each time dimension are added to perform feature fusion on the data distribution features, and a result obtained by the addition is used as fused data distribution feature information.

In another embodiment, the data distribution characteristics corresponding to each time dimension may be multiplied to perform characteristic fusion, and the result obtained by the multiplication is used as fused data distribution characteristic information.

In another embodiment, the data distribution characteristics corresponding to each time dimension may be subjected to weighting processing to perform characteristic fusion, and a result obtained by the weighting processing is used as fused data distribution characteristic information, where the weight may be set based on a business requirement or may be adjusted by designing a weight design model.

Further, feature fusion may be performed on the regression feature information, the time-series associated feature information, and the fused data distribution feature information, for example, by means of weighting processing, specifically, weights of the regression feature information, the time-series associated feature information, and the fused data distribution feature information may be determined respectively; and performing feature fusion on the regression feature information, the time sequence association feature information and the fused data distribution feature information based on the weight to obtain a fused feature, wherein the weight can be set based on business requirements and can also be adjusted by designing a weight design model.

As an example, the feature information may be represented by specific numerical values, where the regression feature information, the time-series correlation feature information, and the fused data distribution feature information of the time-series data may be 5.1,4.5,6.2, respectively, and the corresponding weight allocation may be 0.2,0.5,0.3, respectively, so that the fused feature is 5.1 × 0.2+4.5 × 0.5+6.2 × 0.3=5.13.

After the fused features are obtained, the object type of the target object in the preset time interval can be predicted further based on the fused features.

For example, a feature threshold may be set, so that the object type to which the target object belongs within a preset time interval may be determined by comparing the fused feature with the feature threshold; for another example, the fused features of the target object and the fused features of the full-scale object are sorted, and the object type of the target object in the preset time interval is determined based on the sorting result; and so on.

As can be seen from the above, the present embodiment can acquire time series data of the target object in the historical time interval; performing regression processing on the time series data to obtain regression characteristic information of the time series data; calculating data distribution characteristic information of the time-series data in each time dimension based on data distribution of the time-series data in at least one time dimension; calculating time sequence correlation characteristic information of the time sequence data based on the time sequence correlation relationship of the time sequence data in the historical time interval; and predicting the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information.

According to the scheme, regression characteristic information can be obtained by performing regression processing on time series data of the target object, and specifically, the regression characteristic information can be used for predicting the object type of the target object in a short period; predicting the object type of the target object in a long term from a plurality of time dimensions by calculating data distribution characteristic information of the time-series data of the target object on the plurality of time dimensions; by calculating time sequence correlation characteristic information of the time sequence data of the target object, the type of the object to which the target object belongs can be predicted through a neural network model. Furthermore, the scheme can perform feature fusion on the obtained feature information, so that the short-term prediction result, the long-term prediction result and the neural network prediction result of the target object are fused, and the object type of the target object is comprehensively predicted based on the fusion result. In this way, the scheme avoids the situation that the index required for predicting the object type needs to be manually adjusted in the process of judging based on the rule, and can accurately and efficiently predict the object type of the object based on the time series data of the object by the method based on the data mining.

The method described in the above examples is further described in detail below by way of example.

In this embodiment, a description will be given by taking an example in which a prediction apparatus of an object type is integrated in a server and a terminal, and the server may be a single server or a server cluster composed of a plurality of servers; the terminal can be a mobile phone, a tablet computer, a notebook computer and other equipment.

As shown in fig. 10, a method for predicting an object type includes the following specific steps:

201. the terminal sends the time series data of the target object in the historical time interval to the server.

In an embodiment, the object type prediction method described in the present application may be applied to an application scenario of resource oriented delivery, and specifically, in an actual operation process of a product, the product is limited by the consideration of delivering resources or reducing user harassment, and some specific qualified user groups need to be directionally delivered, and on the premise of limited resources, maximization of a delivery-output ratio is achieved, and at this time, a part of high-value user groups often needs to be locked for directional operation, and user directional screening needs to be performed with the help of a total declaration period value of a user for operation.

Where the total life cycle value (LTV), meaning the lifetime value of a client, is the sum of all economic returns that a company receives from all the interactions of the user. The method is generally applied to the field of marketing, is used for measuring the value of an enterprise client to the enterprise, and is determined as an important reference index for whether the enterprise can obtain high profit.

Thus, the LTV index may be determined, in particular, the definition of LTV is different for each traffic scenario, generally set as the number/amount of times a certain behavior occurs n days in the future. Further, the terminal may collect LTV index values of the target user within a historical time interval, generate time-series data of the target user within the historical time interval, and transmit the time-series data to the server.

202. The server acquires the time sequence data sent by the terminal.

203. And the server performs regression processing on the time series data to obtain regression characteristic information of the time series data.

In one embodiment, the server may determine a regression coefficient for the time series data, wherein the regression coefficient characterizes a time-series correlation of the time series data. The server may further determine a regression model required for performing regression processing on the time series data based on the determined regression coefficient, and perform regression processing on the time series data through the regression model to obtain regression feature information of the time series data. As an example, the regression model may include an ARIMA model.

204. The server calculates data distribution characteristic information of the time sequence data in each time dimension based on the data distribution of the time sequence data in at least one time dimension.

In an embodiment, the server may obtain a trained model corresponding to each time dimension, and calculate data distribution characteristic information of the time series data in each time dimension through the trained model corresponding to each time dimension based on data distribution of the time series data in each time dimension. By way of example, the time dimension may include a trending time dimension, a mutating time dimension, a periodic time dimension, a random time dimension, and the like.

205. The server calculates time-series correlation characteristic information of the time-series data based on the time-series correlation of the time-series data in the historical time interval.

In an embodiment, the server may obtain a trained time series correlation prediction model, where the time series correlation prediction model is used to calculate time series correlation characteristic information of the time series data, and for example, an LSTM model may be used as the time series correlation prediction model. And the server can analyze the time sequence correlation relationship of the time sequence data in the historical time interval through the time sequence correlation prediction model, so that the server can calculate the time sequence correlation characteristic information of the time sequence data based on the analysis result.

As an example, the server may split the sample data set into a training set and a test set, for example, the first 2/3 part of the sample data set may be used as the training set and the last 1/3 part may be used as the test set in time order. Further, the server may perform normalization processing of 0 to 1 on the sample to obtain a normalized sample, and then perform model training on the LSTM through the normalized sample. Optionally, the server may perform model evaluation on the trained LSTM, specifically, the server may calculate a mean square error of a final test, perform inverse normalization on a predicted value obtained by model prediction to reduce the predicted value to a value of itself, and perform model evaluation on the trained LSTM based on the value.

206. And the server predicts the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information.

In an embodiment, the server may perform feature fusion on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time sequence association feature information to obtain a fused feature, and specifically, the server may perform feature fusion on the data distribution feature corresponding to each time dimension to obtain fused data distribution feature information, and determine weights of the regression feature information, the time sequence association feature information, and the fused data distribution feature information, respectively, so that the regression feature information, the time sequence association feature information, and the fused data distribution feature information may be subjected to feature fusion based on the weights to obtain the fused feature. Further, the server may predict an object type of the target object within a preset time interval based on the fused features.

The feature fusion refers to processing the plurality of feature information to obtain a fused feature referring to the plurality of feature information. The feature fusion method may be various, and for example, the feature information may be added, multiplied, or weighted.

207. And the server sends the prediction result to the terminal.

208. And the terminal acquires the prediction result sent by the server and develops the service based on the prediction result.

In an embodiment, the terminal may obtain the type prediction result of the target user sent by the server, sort the type prediction result of the target user and the type prediction results of all users, and preferentially operate a user group with a top ranking.

The regression feature information can be obtained by performing regression processing on the time series data of the target object, and specifically, the regression feature information can be used for predicting the object type to which the target object belongs in a short period; predicting the object type of the target object in a long term from a plurality of time dimensions by calculating data distribution characteristic information of the time-series data of the target object on the plurality of time dimensions; by calculating time series correlation characteristic information of the time series data of the target object, the type of the object to which the target object belongs can be predicted through a neural network model. Further, the embodiment of the application can perform feature fusion on the obtained feature information, so that the short-term prediction result, the long-term prediction result and the neural network prediction result of the target object are fused, and the object type to which the target object belongs is comprehensively predicted based on the fusion result. In this way, the embodiment of the present application avoids the situation that the index required for predicting the object type needs to be manually adjusted in the rule-based determination process, and the embodiment of the present application can accurately and efficiently predict the object type to which the target object belongs based on the time-series data of the target object by using the data mining-based method.

In addition, in the user operation process, the embodiment of the application predicts the user LTV through an automatic data method, avoids errors of empiric meaning and artificial judgment of a product manager, helps a product and a service team to improve operation efficiency and drive product growth, and therefore plays an important role in refined operation and growth of products.

In order to better implement the method, correspondingly, the embodiment of the application also provides an object type prediction device, wherein the object type prediction device can be integrated in a server or a terminal. The server can be a single server or a server cluster consisting of a plurality of servers; the terminal can be a mobile phone, a tablet computer, a notebook computer and other equipment.

For example, as shown in fig. 11, the object type prediction apparatus may include a data acquisition unit 301, a regression processing unit 302, a first calculation unit 303, a second calculation unit 304, and a type prediction unit 305, as follows:

a data acquisition unit 301 configured to acquire time-series data of the target object in a history time interval;

a regression processing unit 302, configured to perform regression processing on the time series data to obtain regression feature information of the time series data;

a first calculating unit 303, configured to calculate data distribution characteristic information of the time series data in each time dimension based on data distribution of the time series data in at least one time dimension;

a second calculating unit 304, configured to calculate time-series correlation characteristic information of the time-series data based on a time-series correlation relationship of the time-series data in the historical time interval;

a type prediction unit 305, configured to predict an object type of the target object within a preset time interval based on the regression feature information, the data distribution feature information corresponding to each of the time dimensions, and the time-series correlation feature information.

In an embodiment, referring to fig. 12, the regression processing unit 302 may include:

a coefficient determination subunit 3021 configured to determine a regression coefficient of the time-series data, wherein the regression coefficient represents a time-series correlation of the time-series data;

a model determination subunit 3022 that may be configured to determine a regression model required for performing regression processing on the time-series data based on the regression coefficient;

the regression processing subunit 3023 may be configured to perform regression processing on the time-series data through the regression model to obtain regression feature information of the time-series data.

In an embodiment, the coefficient determining subunit 3021 may be configured to:

In an embodiment, referring to fig. 13, the first calculating unit 303 may include:

a first obtaining subunit 3031, configured to obtain a trained model corresponding to each time dimension;

the first calculating subunit 3032 may be configured to calculate, based on the data distribution of the time-series data in each time dimension, data distribution feature information of the time-series data in the time dimension through a trained model corresponding to the time dimension.

In an embodiment, the time dimension includes a trending time dimension, and the trained model includes a trend prediction model corresponding to the trending time dimension; the first obtaining subunit 3031 may specifically be configured to:

In an embodiment, the time dimension includes at least one mutation time dimension, and the trained model includes a mutation prediction model corresponding to each mutation time dimension; the first calculating subunit 3032 may specifically be configured to:

determining time window information of the mutation time dimension, wherein the time window information represents the duration of a time window corresponding to the mutation time dimension; calculating, by the mutation prediction model, a mutation distribution characteristic of the time-series data within the time window based on a data distribution of the time-series data over the mutation time dimension; and performing characteristic combination on the mutation distribution characteristic information to obtain data distribution characteristic information of the time series data on the mutation time dimension.

In an embodiment, the time dimension includes at least one mutability time dimension, the time dimension includes a periodicity time dimension, and the trained model includes a periodicity prediction model corresponding to the periodicity time dimension; the first calculating subunit 3032 may specifically be configured to:

In an embodiment, referring to fig. 14, the second calculating unit 304 may include:

a second obtaining subunit 3041, configured to obtain a trained time sequence correlation prediction model, where the time sequence correlation prediction model is used to calculate time sequence correlation characteristic information of time sequence data;

a relationship analysis subunit 3042, configured to analyze, through the time-series correlation prediction model, a time-series correlation of the time-series data in a historical time interval;

the second calculating subunit 3043 may be configured to calculate, based on the analysis result, time-series relevant feature information of the time-series data.

In an embodiment, referring to fig. 15, the type prediction unit 305 may include:

the feature fusion subunit 3051 is configured to perform feature fusion on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time-sequence associated feature information to obtain a fused feature;

the type prediction subunit 3052 may be configured to predict, based on the fused feature, an object type of the target object within a preset time interval.

In an embodiment, the feature fusion subunit 3051 may be specifically configured to:

In a specific implementation, the above units may be implemented as independent entities, or may be combined arbitrarily to be implemented as the same or several entities, and the specific implementation of the above units may refer to the foregoing method embodiments, which are not described herein again.

As can be seen from the above, in the object type prediction apparatus of the present embodiment, the data acquisition unit 301 acquires time-series data of the target object within the history time interval; performing regression processing on the time series data by a regression processing unit 302 to obtain regression feature information of the time series data; calculating, by the first calculation unit 303, data distribution characteristic information of the time-series data in each time dimension based on a data distribution of the time-series data in at least one time dimension; calculating, by a second calculation unit 304, time-series correlation characteristic information of the time-series data based on a time-series correlation of the time-series data within the historical time interval; the type prediction unit 305 predicts the object type of the target object within a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time-series correlation feature information.

According to the scheme, regression characteristic information can be obtained by performing regression processing on time series data of the target object, and specifically, the regression characteristic information can be used for predicting the object type to which the target object belongs in a short period; predicting the object type of the target object in a long term from a plurality of time dimensions by calculating data distribution characteristic information of the time-series data of the target object on the plurality of time dimensions; by calculating time series correlation characteristic information of the time series data of the target object, the type of the object to which the target object belongs can be predicted through a neural network model. Furthermore, the scheme can perform feature fusion on the obtained feature information, so that the short-term prediction result, the long-term prediction result and the neural network prediction result of the target object are fused, and the object type of the target object is comprehensively predicted based on the fusion result. In this way, the scheme avoids the situation that the index required for predicting the object type needs to be manually adjusted in the process of judging based on the rule, and can accurately and efficiently predict the object type of the object based on the time series data of the object by the method based on the data mining.

In addition, an embodiment of the present application further provides a computer device, where the computer device may be a server or a terminal, and as shown in fig. 16, a schematic structural diagram of the computer device according to the embodiment of the present application is shown, specifically:

the computer device may include components such as a memory 401 including one or more computer-readable storage media, a processor 402 including one or more processing cores, and a power supply 403. Those skilled in the art will appreciate that the computer device configuration illustrated in FIG. 16 does not constitute a limitation of computer devices, and may include more or fewer components than those illustrated, or some components may be combined, or a different arrangement of components. Wherein:

the memory 401 may be used to store software programs and modules, and the processor 402 executes various functional applications and data processing by operating the software programs and modules stored in the memory 401. The memory 401 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function (such as a sound playing function, an image playing function, etc.), and the like; the storage data area may store data (such as audio data, a phonebook, etc.) created according to the use of the computer device, etc. Further, the memory 401 may include high speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device. Accordingly, the memory 401 may also include a memory controller to provide the processor 402 and the input unit 603 access to the memory 401.

The processor 402 is a control center of the computer device, connects various parts of the entire mobile phone by using various interfaces and lines, and performs various functions of the computer device and processes data by operating or executing software programs and/or modules stored in the memory 401 and calling data stored in the memory 401, thereby integrally monitoring the mobile phone. Optionally, processor 402 may include one or more processing cores; preferably, the processor 402 may integrate an application processor, which primarily handles operating systems, user interfaces, applications, etc., and a modem processor, which primarily handles wireless communications. It will be appreciated that the modem processor described above may not be integrated into the processor 402.

The computer device also includes a power supply 403 (e.g., a battery) for powering the various components, which may preferably be logically coupled to the processor 402 via a power management system to manage charging, discharging, and power consumption management functions via the power management system. The power supply 403 may also include any component of one or more dc or ac power sources, recharging systems, power failure detection circuitry, power converters or inverters, power status indicators, and the like.

Although not shown, the computer device may further include a camera, a bluetooth module, etc., which will not be described herein. Specifically, in this embodiment, the processor 402 in the computer device loads the executable file corresponding to the process of one or more application programs into the memory 401 according to the following instructions, and the processor 402 runs the application programs stored in the memory 401, so as to implement various functions as follows:

acquiring time sequence data of a target object in a historical time interval; performing regression processing on the time series data to obtain regression characteristic information of the time series data; calculating data distribution characteristic information of the time-series data in each time dimension based on data distribution of the time-series data in at least one time dimension; calculating time sequence correlation characteristic information of the time sequence data based on the time sequence correlation relationship of the time sequence data in the historical time interval; and predicting the object type of the target object in a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension and the time sequence correlation feature information.

The above operations can be implemented in the foregoing embodiments, and are not described in detail herein.

As can be seen from the above, the computer device of this embodiment may obtain regression feature information by performing regression processing on the time-series data of the target object, and specifically, the regression feature information may be used to predict the object type to which the target object belongs in a short period of time; predicting the object type of the target object in a long term from a plurality of time dimensions by calculating data distribution characteristic information of the time-series data of the target object on the plurality of time dimensions; by calculating time sequence correlation characteristic information of the time sequence data of the target object, the type of the object to which the target object belongs can be predicted through a neural network model. Further, the computer device of this embodiment can perform feature fusion on the obtained feature information, thereby implementing fusion on the short-term prediction result, the long-term prediction result, and the neural network prediction result of the target object, and comprehensively predicting the object type to which the target object belongs based on the fusion result. In this way, the computer device according to this embodiment avoids the situation that the index required for predicting the object type needs to be manually adjusted in the rule-based determination process, and the computer device according to this embodiment can accurately and efficiently predict the object type to which the target object belongs based on the time-series data of the target object by using the data mining-based method.

It will be understood by those skilled in the art that all or part of the steps of the methods of the above embodiments may be performed by instructions or by associated hardware controlled by the instructions, which may be stored in a computer readable storage medium and loaded and executed by a processor.

To this end, embodiments of the present application provide a storage medium, in which a plurality of instructions are stored, where the instructions can be loaded by a processor to execute steps in any one of the object type prediction methods provided in the embodiments of the present application. For example, the instructions may perform the steps of:

Wherein the storage medium may include: read Only Memory (ROM), random Access Memory (RAM), magnetic or optical disks, and the like.

Since the instructions stored in the storage medium may execute the steps in the method for predicting an object type provided in the embodiment of the present application, beneficial effects that can be achieved by the method for predicting an object type provided in the embodiment of the present application may be achieved, for which details are shown in the foregoing embodiment and are not described herein again.

According to an aspect of the application, a computer program product or computer program is provided, comprising computer instructions, the computer instructions being stored in a computer readable storage medium. The computer instructions are read by a processor of a computer device from a computer-readable storage medium, and the computer instructions are executed by the processor to cause the computer device to perform the methods provided in the various alternative implementations of the predictive aspect of object types described above.

The object type prediction method, apparatus, computer device and storage medium provided in the embodiments of the present application are described in detail above, and specific examples are applied herein to explain the principles and embodiments of the present application, and the description of the above embodiments is only used to help understand the method and its core ideas of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present application.

Claims

1. A method for predicting an object type, comprising:

acquiring time sequence data of a target object in a historical time interval;

2. The method according to claim 1, wherein performing regression processing on the time-series data to obtain regression feature information of the time-series data includes:

and performing regression processing on the time series data through the regression model to obtain regression characteristic information of the time series data.

3. The method of predicting a type of an object according to claim 2, wherein determining regression coefficients of the time-series data comprises:

performing data conversion on the time sequence data to obtain a stable time sequence corresponding to the time sequence data;

determining a set of candidate regression coefficients for the time series data based on the stationary time series, wherein the set of candidate regression coefficients comprises at least one set of candidate regression coefficients, each set of candidate regression coefficients corresponding to a regression model;

determining a target regression coefficient from the candidate regression coefficient set based on the evaluation result, the target regression coefficient being a regression coefficient of the time-series data.

4. The method for predicting the object type according to claim 1, wherein calculating data distribution characteristic information of the time-series data in each time dimension based on the data distribution of the time-series data in at least one time dimension comprises:

acquiring a trained model corresponding to each time dimension;

and calculating data distribution characteristic information of the time sequence data in the time dimension through a trained model corresponding to the time dimension based on the data distribution of the time sequence data in each time dimension.

5. The method according to claim 4, wherein the time dimension comprises a trend time dimension, and the trained model comprises a trend prediction model corresponding to the trend time dimension;

obtaining a trained model corresponding to each time dimension, including:

acquiring a sample data set required by model training;

determining model type information and model parameter information;

6. The method according to claim 4, wherein the time dimension comprises at least one mutation time dimension, and the trained model comprises a mutation prediction model corresponding to each mutation time dimension;

calculating data distribution characteristic information of the time sequence data in each time dimension through a trained model corresponding to the time dimension based on the data distribution of the time sequence data in each time dimension, wherein the data distribution characteristic information comprises:

determining time window information of the catastrophe time dimension, wherein the time window information represents the duration of a time window corresponding to the catastrophe time dimension;

calculating, by the mutation prediction model, a mutation distribution characteristic of the time-series data within the time window based on a data distribution of the time-series data over the mutation time dimension;

7. The method according to claim 4, wherein the time dimension comprises a periodic time dimension, and the trained model comprises a periodic prediction model corresponding to the periodic time dimension;

decomposing the time sequence data through the periodic prediction model based on the data distribution of the time sequence data on the periodic time dimension to obtain the periodic distribution characteristics of the time sequence data;

8. The method for predicting the object type according to claim 1, wherein calculating the time-series correlation characteristic information of the time-series data based on the time-series correlation of the time-series data in the historical time interval comprises:

analyzing the time sequence incidence relation of the time sequence data in a historical time interval through the time sequence incidence prediction model;

and calculating time-series related characteristic information of the time-series data based on the analysis result.

9. The method according to claim 1, wherein predicting the object type of the target object within a preset time interval based on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time-series correlation feature information comprises:

10. The method for predicting the object type according to claim 9, wherein performing feature fusion on the regression feature information, the data distribution feature information corresponding to each time dimension, and the time-series correlation feature information to obtain a fused feature includes:

performing feature fusion on the data distribution features corresponding to the time dimensions to obtain fused data distribution feature information;

11. An apparatus for predicting an object type, comprising:

a first calculation unit, configured to calculate data distribution characteristic information of the time-series data in each time dimension based on data distribution of the time-series data in at least one time dimension;

12. An electronic device comprising a memory and a processor; the memory stores an application program, and the processor is configured to execute the application program in the memory to perform the operation of the object type prediction method according to any one of claims 1 to 10.

13. A storage medium storing a plurality of instructions adapted to be loaded by a processor to perform the steps of the object type prediction method according to any one of claims 1 to 10.