Customer loss early warning method, device, equipment, medium and program product

Abstract

The invention provides a customer loss early warning method, device and equipment, a medium and a program product, relates to the technical field of artificial intelligence, and can be applied to the field of financial science and technology. The method comprises the following steps: under the condition that authorization of a target customer to use related data thereof is obtained, obtaining data of the target customer in each of W continuous time periods ending the current moment to obtain a target sequence; standardizing the target sequence to obtain an original standard sequence; the original standard sequence is input into a target time sequence behavior anomaly detection model, a reconstruction sequence output by the target time sequence behavior anomaly detection model is obtained, and the target time sequence behavior anomaly detection model is a time sequence behavior anomaly detection model independently constructed for a target customer; comparing the original standard sequence with the reconstructed sequence to obtain a deviation degree; and when the deviation degree meets an abnormal alarm condition, triggering loss early warning of the target customer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and can be used in the field of financial technology, and more particularly to a customer loss early warning method, device, equipment, medium and program product. BACKGROUND

[0002] For the early warning scheme of customer loss, many are based on fixed rule engine monitoring, in which the pre-set static standard triggers an abnormal early warning when the customer's related data reaches the relevant conditions in the static standard.

[0003] However, this scheme has defects: first, the fixed rule engine cannot adaptively match the internal fluctuation characteristics of different customers, and is prone to produce a large number of false alarms or miss real early risk signals; second, the data utilization dimension of this scheme is extremely limited, and its judgment basis is mainly limited to the short-term percentage change of a few single indicators, ignoring deeper information reflecting the health of the customer-enterprise relationship and customer stability such as environment, behavior inertia and consumption patterns, resulting in a certain waste of information assets; third, this scheme is essentially a kind of lagging event response mechanism, and lacks the ability to capture the gradual trend of customer behavior patterns. SUMMARY

[0004] The present application provides a customer loss early warning method, device, equipment, medium and program product, which can at least partially solve one or more of the above problems.

[0005] The first aspect of the embodiment of the present application provides a customer loss early warning method. The method comprises: obtaining the data of a target customer in each period of W consecutive periods ending at the current time to obtain a target sequence, under the condition that the target customer authorizes the use of its related data, wherein W is an integer greater than 1; performing standardization processing on the target sequence to obtain an original standard sequence; inputting the original standard sequence into a target time series behavior anomaly detection model and obtaining a reconstructed sequence output by the target time series behavior anomaly detection model, wherein the target time series behavior anomaly detection model is a time series behavior anomaly detection model independently constructed for the target customer, wherein different time series behavior anomaly detection models are constructed for different customers; comparing the original standard sequence and the reconstructed sequence to obtain a deviation degree; and triggering the loss early warning of the target customer when the deviation degree meets the abnormal alarm condition.

[0006] According to an embodiment of the present application, the time series anomaly detection model independently constructed for each customer is constructed based on a deep learning architecture of a long short-term memory network autoencoder, wherein the target training data set of the target time series behavior anomaly detection model is from the historical data of the target customer, wherein each input sample in the target training data set includes a historical sequence composed of data of the target customer in each time period of consecutive W time periods; and the output sample corresponding to each input sample includes the historical sequence.

[0007] According to an embodiment of the present application, the comparison of the original standardized sequence and the reconstructed sequence to obtain the deviation degree includes: calculating the mean square error between the original standardized sequence and the reconstructed sequence, wherein the mean square error represents the deviation degree.

[0008] According to an embodiment of the present application, the standardization processing of the target sequence to obtain the original standardized sequence includes: normalizing the data in the target sequence based on the maximum value and the minimum value of the historical data of the target customer to obtain the original standardized sequence.

[0009] According to an embodiment of the present application, when the deviation degree meets the abnormal alarm condition, triggering the churn warning of the target customer includes: adding the deviation degree to the feature vector of the target customer to obtain a target feature vector; inputting the target feature vector into the trained dynamic risk scoring model to obtain a target risk score of the output of the dynamic risk scoring model; and when the target risk score is higher than a preset target warning threshold, triggering the churn warning of the target customer.

[0010] According to an embodiment of the present application, W is configured to include a plurality of selectable values, wherein a plurality of deviation degrees are obtained corresponding to the plurality of selectable values; and adding the deviation degree to the feature vector of the target customer to obtain the target feature vector includes: adding the plurality of deviation degrees to the feature vector of the target customer to obtain the target feature vector.

[0011] According to an embodiment of the present application, the method further includes: obtaining a feature vector of the target customer, wherein the feature vector of the target customer includes an n-dimensional feature, and the n-dimensional feature is obtained through feature fusion processing, wherein n is an integer greater than or equal to 1. The feature fusion processing includes: in the case that the authorization of the customer to use the historical data thereof is obtained, obtaining the historical data of m customers, m being an integer greater than 1; extracting h-dimensional feature data of each customer based on the historical data of each customer, wherein h is an integer greater than or equal to n; constructing a Pearson correlation coefficient matrix between the h-dimensional features of the m customers based on the h-dimensional feature data of the m customers; and based on the Pearson correlation coefficient matrix, merging any pair of features whose absolute value of the Pearson correlation coefficient exceeds a preset correlation threshold into one feature to finally obtain the n-dimensional feature.

[0012] According to an embodiment of the present application, the step of merging any pair of features whose absolute value of Pearson correlation coefficient exceeds a preset correlation threshold into one feature comprises: obtaining a pair of to-be-merged features whose absolute value of Pearson correlation coefficient exceeds a preset correlation threshold; obtaining the correlation between each to-be-merged feature and a risk score output by the dynamic risk scoring model; retaining the feature with higher correlation in the pair of to-be-merged features and eliminating the other feature.

[0013] According to an embodiment of the present application, before triggering the churn warning of the target customer when the target risk score is higher than a preset target warning threshold, the method further comprises: obtaining a target level to which the target customer belongs; and obtaining a target warning threshold corresponding to the target level, wherein the warning thresholds corresponding to different levels are different.

[0014] In a second aspect, the present application provides a customer churn warning device. The customer churn warning device comprises a time-series behavior anomaly detection module. The time-series behavior anomaly detection module is deployed with different time-series behavior anomaly detection models independently constructed for different customers. The time-series behavior anomaly detection module comprises a serialization unit, a model prediction unit, and a reconstruction error calculation and behavior anomaly score generation unit.

[0015] The serialization unit is configured to, in the case that authorization of the target customer to use its related data is obtained, obtain data of the target customer in each of the last W time periods ending at the current time to obtain a target sequence, and perform standardization processing on the target sequence to obtain an original standard sequence, wherein W is an integer greater than 1.

[0016] The model prediction unit is configured to input the original standard sequence into the target time-series behavior anomaly detection model, and obtain a reconstruction sequence output by the target time-series behavior anomaly detection model, wherein the target time-series behavior anomaly detection model is a time-series behavior anomaly detection model independently constructed for the target customer.

[0017] The reconstruction error calculation and behavior anomaly score generation unit is configured to compare the original standard sequence and the reconstruction sequence to obtain a deviation degree, and trigger a churn warning of the target customer when the deviation degree meets an anomaly warning condition.

[0018] In a third aspect, the present application provides an electronic device, comprising: one or more processors; a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the above method.

[0019] In a fourth aspect, the present application further provides a computer-readable storage medium having stored thereon a computer program or instructions, wherein the computer program or instructions are executed by a processor to implement the steps of the above method.

[0020] A fifth aspect of the embodiments of the present application further provides a computer program product comprising computer programs or instructions, which, when executed by a processor, implement the steps of the above method. BRIEF DESCRIPTION OF DRAWINGS

[0021] The above and other objects, features and advantages of the present application will become more apparent from the following description when taken in conjunction with the accompanying drawings, in which:

[0022] Figure 1 A flow chart of a customer churn early warning method according to an embodiment of the present application is schematically shown;

[0023] Figure 2 A flow chart of a customer churn early warning method according to another embodiment of the present application is schematically shown;

[0024] Figure 3 A system architecture diagram of a customer churn early warning method to which embodiments of the present application can be applied is schematically shown;

[0025] Figure 4 An execution flow chart of a data preprocessing module in an embodiment of the present application is schematically shown;

[0026] Figure 5 An execution flow chart of a multi-dimensional feature engineering module in an embodiment of the present application is schematically shown;

[0027] Figure 6 An execution flow chart of an adaptive hierarchical early warning module in an embodiment of the present application is schematically shown; and

[0028] Figure 7 A block diagram of an electronic device suitable for implementing a customer churn early warning method according to embodiments of the present application is schematically shown. DETAILED DESCRIPTION

[0029] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. It is to be understood, however, that the description is merely exemplary and is intended to provide a thorough understanding of the embodiments of the present application. The following description is not intended to limit the scope of the present application, and the embodiments of the present application can be implemented in various ways. Furthermore, in the following description, descriptions of well-known functions and constructions are omitted to avoid obscuring the concept of the present application.

[0030] The terms used herein are merely used to describe specific embodiments and are not intended to limit the present application. The terms "include" and "have" used herein indicate the presence of the stated features, steps, operations, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, or components.

[0031] All terms used herein, including technical and scientific terms, have the meanings commonly understood by one of ordinary skill in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0032] In the technical solutions of the present application, the user information (including but not limited to user personal information, user image information, user equipment information such as location information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards, necessary security measures are taken, do not violate public order and good customs, and corresponding operation portals are provided for users to choose authorization or refusal.

[0033] In the scenario of making automated decisions with personal information, the method, device and system provided by the embodiment of the present application all provide corresponding operation portals for users to choose to agree or refuse the automated decision result; if the user chooses to refuse, the expert decision process is entered.

[0034] Figure 1 A flowchart of a customer churn early warning method according to an embodiment of the present application is schematically shown.

[0035] As shown in Figure 1 The customer churn early warning method of this embodiment can include operations S110-S150.

[0036] In operation S110, in the case where the authorization of the target customer to use its related data is obtained, the data of the target customer in each of the W consecutive time periods ending at the current time is obtained to obtain a target sequence, where W is an integer greater than 1. The W time periods are, for example, 3 months, 6 months or 12 months.

[0037] In operation S120, the target sequence is standardized to obtain an original standard sequence.

[0038] In operation S130, the original standard sequence is input into a target time series behavior anomaly detection model, and a reconstructed sequence output by the target time series behavior anomaly detection model is obtained. The reconstructed sequence is a sequence of the same length as the original standard sequence.

[0039] The target time series behavior anomaly detection model is a time series behavior anomaly detection model independently constructed for the target customer, such as that constructed using the historical data of the target customer. Different time series behavior anomaly detection models can be independently constructed for different customers.

[0040] In operation S140, the original standard sequence and the reconstructed sequence are compared to obtain a deviation degree.

[0041] In operation S150, when the deviation degree satisfies an abnormal alarm condition, a loss warning of the target customer is triggered. For example, in an embodiment, a mean square error between the original standardized sequence and the reconstructed sequence can be calculated, and the mean square error can be used to represent the deviation degree. Then, it can be determined whether the mean square error is greater than a preset deviation threshold. If yes, a warning signal is sent out.

[0042] Embodiments of the present application can independently construct a time series behavior anomaly detection model for each customer. The model of each customer is learned by learning the "normal" behavior fluctuation pattern of the customer, and the current behavior model of the customer is analyzed to predict whether the user behavior deviates from the normal fluctuation pattern. In this way, early subtle deviation of the customer behavior pattern can be captured, so as to realize early and accurate risk warning.

[0043] Figure 2 A flowchart of a customer loss warning method according to another embodiment of the present application is schematically shown.

[0044] As shown in Figure 2 According to the customer loss warning method of this embodiment, in addition to operations S110-S140, operations S250-S270 can also be included. Operations S250-S270 can be regarded as a specific embodiment of operation S150.

[0045] First, the deviation degree is obtained through operations S110-S140.

[0046] Next, in operation S250, the deviation degree is added to the feature vector of the target customer to obtain a target feature vector. The feature vector of the target customer is obtained by collecting multi-dimensional data of the target customer and then through feature engineering. This deviation degree can be added to the feature vector of the target customer as a new feature for evaluating whether the customer behavior is abnormal, and more judgment basis is provided for the prediction of the subsequent dynamic risk scoring model.

[0047] In an embodiment, W is configured to include multiple selectable values, and according to each selectable value, a corresponding deviation degree can be obtained through operations S110-S140. For example, the deviation degrees corresponding to short-term (such as 3 months), medium-term (such as 6 months) and long-term (such as 12 months) of the customer can be obtained respectively. Then, in operation S250, the multiple deviation degrees can be added to the target feature vector.

[0048] In operation S260, the target feature vector is input to the trained dynamic risk scoring model, and a target risk score of an output of the dynamic risk scoring model is obtained. The dynamic risk scoring model is trained by using historical data of a large number of customers, and is used to predict a customer churn probability.

[0049] In operation S270, when the target risk score is higher than a preset target early warning threshold, a customer churn early warning of the target customer is triggered.

[0050] The embodiment of the present application can combine the result of predicting whether the user behavior deviates from the normal fluctuation mode into the judgment of the dynamic risk scoring model, and further improve the accuracy of predicting the customer churn probability.

[0051] Figure 3 A system architecture diagram of a customer churn early warning method to which the embodiment of the present application can be applied is schematically shown. The following introduces the customer churn early warning method and device determined by the embodiment of the present application in combination with Figure 3 The customer churn early warning method and device determined by the embodiment of the present application are introduced in detail. It should be noted that the following introduction takes a bank customer as an example, which is only an example. The customer churn early warning method and device determined by the embodiment of the present application can be used in the field of financial technology, and can also be used in any field other than the field of financial technology, and the application of the present application is not limited.

[0052] Reference Figure 3 The embodiment provides a customer churn risk prediction system 300. The system 300 can deeply integrate customer multi-dimensional static and dynamic data, capture subtle changes of customer behavior patterns by using a time series behavior anomaly detection model, construct a dynamic risk scoring model by using an ensemble learning algorithm, and can combine customer level classification to perform adaptive hierarchical early warning. The target of the embodiment is to change the customer retention strategy from passive "after-the-event response" to active "pre-event prediction and intervention", so as to improve the accuracy, timeliness and effectiveness of resource allocation of risk identification.

[0053] According to the embodiment of the present application, the system 300 can realize accurate, dynamic and quantitative prediction of customer churn risk through the collaborative work of the data acquisition and preprocessing module 301, the multi-dimensional feature engineering module 302, the time series behavior anomaly detection module 303, the dynamic risk scoring model module 304 and the adaptive hierarchical early warning module 305. The following introduces each module in detail.

[0054] (I) Data acquisition and preprocessing module 301

[0055] According to an embodiment of the present application, the data collection and preprocessing module 301 is the basic data processing hub of the entire system 300, and its core task is to build a unified, standardized and information-rich customer-level time series dataset, providing input for the subsequent multi-dimensional feature engineering module 302, time series behavior anomaly detection module 303 and dynamic risk scoring model module 304.

[0056] Figure 4 The execution flowchart of the data preprocessing module 301 in an embodiment of the present application is schematically shown. The execution flow of the data collection and preprocessing module 301 can include an automatic and configurable data processing pipeline, which can achieve the accuracy, robustness and efficiency of the conversion process from heterogeneous data sources to standard feature matrices.

[0057] In combination Figure 4 , the execution flow of the data collection and preprocessing module 301 can include the following steps: S41 extracting raw data, S42 missing value processing, S43 anomaly value detection and processing, and S44 data standardization / normalization processing.

[0058] In step S41, raw data is extracted, which can specifically include: data source definition and interface access, data extraction and uniform identifier alignment, and customer-level time series data wide table construction.

[0059] The purpose of data source definition and interface access is to establish a stable data path between the system 300 and various internal business systems. The system administrator can explicitly define the required data sources in the configuration interface. For each data source, the system can configure its access method, authentication information and target data table or view. In some embodiments of the present application, the establishment of data interface can adopt a standardized technical solution. For a relational database-based system, a Java Database Connectivity (JDBC) driver can be used for connection, and the configuration content can include database IP address, port, service name, username and encrypted password, etc. For modern systems that provide Application Programming Interface (API) services, the API endpoint URL, HTTP request method and API key for authentication, etc. can be configured. Through this step, the system 300 obtains the ability to programmatically and authorizedly access customer data scattered in various corners.

[0060] The purpose of data extraction and unique identifier alignment is to extract raw data from the accessed data sources according to a preset scheduling strategy, and to associate the data from different sources through a globally unique customer identifier. The system 300 can internally build a configurable scheduling engine, for example, the data extraction task can be performed at the business low peak period (such as the early morning period) of each natural day to obtain the full amount of the latest data before.

[0061] When extracting raw data, the data of the customer within a certain time span can be specified from the internal business systems (such as the first system, the second system, and the third system). The obtained data can include multi-dimensional data, such as customer basic data, transaction records of customer service (such as cooperation with a product or participation in a transaction), and communication data with the customer.

[0062] After extracting the raw data, data alignment is performed. The key to data alignment is the unique identifier of each customer, which can be the core hub that connects the scattered data from the above-mentioned various internal business systems. After data extraction, the system 300 can perform internal association operations to determine the attribution of the customer through the unique identifier of the customer. For records that may be missing the customer identifier in some internal business systems (such as the first system), the system 300 can start a fuzzy matching and repair logic based on multiple fields such as customer name and organization code to maximize the integrity of the data.

[0063] After extracting the raw data, a customer-level time series data wide table can be constructed.

[0064] The raw data is mostly in a serial manner (such as transaction data), which is not suitable for direct input into a machine learning model, and needs to be aggregated and converted. For example, a structured data wide table with "customer-time" as the double primary key can be constructed.

[0065] In an embodiment, each row of the wide table can represent a complete state snapshot of a customer at a specific time unit (such as "month"). The specific implementation process can be as follows: first, create an empty data frame with the customer unique identifier and month (such as format 'YYYY-MM') as the joint primary key. Then, the raw records extracted from each internal business system (such as the first system, the second system, and the third system) can be grouped and aggregated according to the customer unique identifier and month, and the aggregation forms the statistical indicators of each system, such as the monthly transaction total amount and the total number of transactions of the transaction system. Next, these monthly aggregated indicators and the customer's profile information (such as company size, industry) are connected to the aforementioned empty data frame through the customer unique number, thereby forming a structured customer-level monthly data wide table.

[0066] At step S42, missing value processing, which can specifically involve systematic data cleaning and missing value filling.

[0067] Systematic data cleaning refers to, after obtaining the structured data wide table, performing data cleaning based on the data wide table to process various quality problems of the data generated in the generation, transmission, and storage process, including systematic processing of missing values and data formats.

[0068] Regarding missing value filling, a rule-based and multi-strategy filling method can be used to ensure the rationality and accuracy of the filling. A specific implementation logic is: traversing each field to be processed in the data wide table, and automatically selecting the optimal filling strategy according to the type and missing rate of the field. Several optional filling strategies are listed here.

[0069] Strategy one (missing of categorical variables): for enterprise customers, when it comes to missing of features such as company size and industry classification / text type, a preset independent category identifier "unknown" can be uniformly replaced. This can preserve the potential implications of "information missing" itself and let the subsequent model learn its impact by itself.

[0070] Strategy two (missing of low-proportion numerical variables): for numerical features such as transaction amount, first calculate the missing value proportion in the current processing batch. If the proportion is lower than a configurable threshold (default is 5%), start statistical filling. Considering the skewness distribution characteristics of financial data, the median can be used for filling, rather than the mean, because the median is not sensitive to extreme values and is more robust.

[0071] Strategy three (missing of time series numerical variables): for core time series features such as monthly total amount, if a month's data is missing, the context can be checked. If there are valid data points before and after the missing point, linear interpolation can be used for filling. The calculation formula is:

[0072] ,

[0073] Where is the value to be filled, is the position of time point t. If the missing occurs at the head or tail of the sequence, forward filling or backward filling strategy can be used to fill with the latest valid observation.

[0074] Strategy four (high proportion of numerical variable missing): If the missing rate of a numerical feature is higher than a threshold, simple statistical imputation may introduce large bias. In this case, model-based imputation method can be initiated. This method takes the column to be imputed as the target variable and all other columns as the predictor variables, and builds a linear model for the column. The missing values are then filled with the predicted values from the model. This process is iteratively performed on all columns with missing values until convergence, so that the correlation between variables is used to make more accurate estimates.

[0075] Regarding data format unification, a strict type conversion and formatting process can be performed. For example, ensure that all date fields are uniformly converted to YYYY-MM-DD format; all numerical fields are converted to 64-bit floating-point numbers to ensure precision; and all categorical fields are converted to string type. This can avoid errors caused by inconsistent data types in subsequent calculations.

[0076] In step S43, anomaly value detection and processing, specifically multi-dimensional anomaly value detection and repair.

[0077] Anomaly values, i.e. data points that are significantly different from the rest of the observations in the dataset, can seriously affect the effectiveness of model training. In the embodiments of the present application, a multi-dimensional, double-filtering mechanism combining statistical methods and business rules is used to identify and handle anomaly values.

[0078] The double-filtering mechanism can include: first filtering: statistical distribution-based anomaly detection; and second filtering: business logic rule-based verification.

[0079] The first filtering, statistical distribution-based anomaly detection, is as follows. For approximately normally distributed numerical features, the standard score method is used. For each data point in the feature column , calculate its standard score:

[0080]

[0081] where and are the mean and standard deviation of the column, respectively.

[0082] Data points with absolute values greater than a preset threshold (default 3) are marked as potential anomaly values.

[0083] For skew distribution features commonly found in financial data (such as amount), a more robust interquartile range method can be used. First, calculate the first quartile (Q1) and the third quartile (Q3) of the feature, and calculate the interquartile range IQR = Q3 - Q1. Then, define the anomaly value judgment boundary as:

[0084] ,

[0085] Here, k is a configurable sensitivity coefficient (default value is 1.5). Any data point falling outside this range will be marked as a potential outlier.

[0086] The second layer of filtering involves validation based on business logic rules, as detailed below. Building upon statistical filtering, a series of hard-coded business rules can be applied for secondary validation to capture statistically possible but business-unlikely erroneous data. For example, the total amount must be greater than or equal to zero; another example is that the number of payroll payments must be a positive integer and should not exceed the maximum number of days in a month (e.g., 31 days). Any data point that violates these business rules will be directly marked as an outlier.

[0087] For all data points ultimately marked as outliers, two configurable remediation strategies are available. One is the capping method, which replaces outliers exceeding the upper boundary with the upper boundary value and outliers below the lower boundary with the lower boundary value. For example, when using the interquartile range method, any outlier greater than the upper boundary is replaced with the upper boundary value. The values ​​were all reset to Alternatively, it can be treated as a missing value and the missing value filling logic in step S42 can be invoked for processing.

[0088] In step S44, data standardization / normalization is performed. This is the final step in data preprocessing, aiming to mathematically transform the cleaned and repaired data to make it more suitable as input for subsequent machine learning models. For features with severely positively skewed distributions (e.g., amounts, bank balances), a logarithmic transformation can be used to make the data distribution closer to a normal distribution, thereby improving the performance of linear or distance-based models. Furthermore, to eliminate the influence of differences in units and numerical ranges between different features, all numerical features can be standardized using the standard score method, transforming all features to the same scale with a mean of 0 and a standard deviation of 1, to avoid the model biasing towards features with larger numerical ranges when calculating weights. For specific downstream models, configuration for min-max normalization can also be supported, linearly scaling the data to the [0,1] interval.

[0089] (II) Multidimensional Feature Engineering Module 302

[0090] In a specific embodiment of the present invention, the multidimensional feature engineering module 302 is the core bridge connecting the original data and the prediction model. Its fundamental task is to upgrade the client-level time series data wide table generated by the data acquisition and preprocessing module 301 and transform it into a feature vector space with strong predictive ability and high information density.

[0091] The multidimensional feature engineering module 302 can systematically and creatively generate derived features that can characterize customer status, behavioral patterns, relationship depth, and changing trends from multiple dimensions through a series of preset mathematical transformations and aggregation operations based on financial business logic and data science principles. The output of the multidimensional feature engineering module 302 can be the final feature matrix, which can be transmitted to the dynamic risk scoring model module 304.

[0092] Figure 5 The schematic diagram illustrates the execution flowchart of the multi-dimensional feature engineering module 302 in one embodiment of the present invention. For example... Figure 5 As shown, the multidimensional feature engineering module 302 can include a set of features construction, encoding, calculation and selection processes. In one embodiment, the process can be divided into four main steps: step S51 static attribute feature extraction, step S52 temporal dynamic feature extraction, step S53 relation feature extraction, and step S54 feature fusion and filtering.

[0093] Step S51: Encoding and Quantization of Static Attribute Features. This step aims to process the non-numerical, categorical static information in customer profiles, converting it into a numerical format that can be recognized and processed by machine learning models. These features remain unchanged over a long period of time, forming the cornerstone of customer profiles.

[0094] Taking corporate clients served by banks as an example, for unordered multi-category variables such as company size (its value range can be "large", "medium", "small", "micro"), contracting institution level (its value range can be "head office level", "branch level", "sub-branch level"), and industry classification code, one-hot encoding technology can be used for processing. For binary Boolean features such as whether a client is confidential, the system directly maps them to integers 0 and 1, such as 1 representing "yes" and 0 representing "no", so that they can directly participate in subsequent mathematical operations.

[0095] In step S52, time-series dynamic feature extraction is performed. This step is the core and key of feature engineering, aiming to extract deep-seated patterns of dynamic evolution from customer time-series data (such as monthly series data) to capture customer behavior patterns at different time scales.

[0096] In one embodiment, a multi-window calculation mechanism is introduced, which can preset a set of time window lengths W. For example, when the time period is in months, W={3,6,12}, representing short-term, medium-term, and long-term observation windows, respectively. All the time-series dynamic features described below will be calculated independently for each preset window length W. Since W has a set of values, the feature dimension can be multiplied, greatly enriching the information that the model can perceive.

[0097] set up For any numerical time-series variable (such as total monthly amount) observed in month t, in one embodiment, some or all of the following five categories of features (2.1 to 2.5) can be calculated.

[0098] 2.1 Statistical descriptive features: These features aim to capture the scale and concentration of a customer's business within a specific time window.

[0099] 2.1.1, Rolling Mean. The calculation formula is:

[0100]

[0101] This feature reflects the average business level of customers within window W.

[0102] 2.1.2, Rolling Standard Deviation. The calculation formula is as follows:

[0103] ,

[0104] This feature quantifies the dispersion or instability of a customer's business scale.

[0105] 2.1.3 Other rolling statistics. For example, the maximum, minimum, median, and 25th and 75th percentiles within the rolling window can be calculated to more comprehensively describe the distribution of the data.

[0106] 2.2 Trend Change Features: These features aim to quantify the direction and rate of customer business development.

[0107] 2.2.1, Slope of Rolling Window Linear Regression. This can be used for time-series data points within a window W. for Construct a univariate linear regression model:

[0108] Where, the independent variable i is the time step, and the dependent variable is... These are business indicator values. The system calculates and extracts the slope of the regression model. As a core trend feature, the slope is calculated using the following formula:

[0109]

[0110] A significantly negative slope is a signal of business decline for enterprise customers.

[0111] 2.2.2 The ratio of the current value to the rolling average is calculated as follows:

[0112] ;

[0113] This ratio reflects the current point in time where the business performance is positioned relative to its recent historical average; a value greater than 1 indicates above average, and a value less than 1 indicates below average.

[0114] 2.3 Business fluctuation characteristics: These characteristics can measure the relative stability of a bank's corporate clients' business and exclude the interference of economies of scale.

[0115] 2.3.1 Rolling coefficient of variation: The calculation formula is as follows

[0116] ,

[0117] The coefficient of variation is a dimensionless indicator of relative volatility that makes customers of different business sizes comparable.

[0118] 2.4 Frequency and Regularity Characteristics: These characteristics mainly describe the regularity of salary payment behavior.

[0119] Let's take a bank's corporate clients as an example. This feature can be used to assess the health of a corporate client's business operations. Let's assume... This refers to the date in month t when a customer processes a transaction (such as placing an order, making a transaction, or receiving a salary) through the bank.

[0120] 2.4.1 The daily average and standard deviation of this business: Calculated separately. The mean and standard deviation within the rolling window W. A larger standard deviation for paydays indicates that the company's monthly paydays are less fixed, which may suggest difficulties in cash flow management.

[0121] 2.4.2 Number of consecutive months of this business transaction: Calculate the number of months in which the transaction actually occurred within the rolling window W. If this value is less than W, it indicates that the customer has intermittently suspended transactions.

[0122] 2.5 Year-on-year and month-on-month characteristics: These characteristics are traditional, but still effective, indicators for monitoring short-term changes.

[0123] 2.5.1 Monthly Change Rate: The calculation formula is as follows .

[0124] 2.5.2 Annual Year-on-Year Growth Rate: The calculation formula is as follows .

[0125] All of the above time-series dynamic features can be calculated separately for the three windows W={3,6,12}, thereby generating hundreds of detailed dynamic profile features.

[0126] In step S53, relationship depth features are constructed. This step aims to construct comprehensive features that measure the breadth, depth, and stickiness of customer relationships by integrating multi-dimensional business data from the customer's internal business systems, in order to comprehensively assess the potential risk of customer churn. Cross-business data can be generated from perspectives such as business interactions with customers. Finally, new features are constructed using cross-business data, and the trend slope of relevant features within the rolling window W is calculated.

[0127] Step S54: Feature Fusion and Filtering. To address the curse of dimensionality, reduce the computational cost of model training, eliminate multicollinearity, and improve the model's generalization ability, the generated feature set can be effectively filtered.

[0128] Specifically, all the features generated in steps S51 to S53 can be merged according to the customer's unique identifier and month to form a large feature matrix.

[0129] Next, a strategy combining embedded and filtering methods can be used for feature selection.

[0130] During the filtering phase, an initial screening can be performed to remove invalid features. This includes removing features with zero or near-zero variance, which show almost no variation across all samples. In one embodiment, a Pearson correlation coefficient matrix can be calculated among all features. For any pair of features whose absolute correlation coefficient exceeds a preset threshold (e.g., 0.95), the feature with the higher correlation to the target variable (e.g., customer churn) can be retained, while the other can be removed.

[0131] The process then proceeds to the embedded selection stage. In one embodiment, during subsequent model training, feature selection can be performed using the feature importance scores output by the gradient boosting decision tree framework. Specifically, a dynamic risk scoring model can be constructed using the gradient boosting decision tree framework. After the initial training of the dynamic risk scoring model, the importance ranking of all features can be obtained. Then, only features whose cumulative contribution to importance scores reaches a preset threshold (e.g., 0.90) can be retained for the final formal training of the dynamic risk scoring model. This largely ensures that the selected features all have a real contribution to the prediction results, improving the model's prediction accuracy and stability.

[0132] Through detailed feature screening, a multi-dimensional, in-depth, and information-rich feature space can be successfully extracted from the original and relatively isolated features. This provides a solid data foundation and profound business insights for achieving accurate and forward-looking customer churn risk prediction.

[0133] (III) Timing Behavior Anomaly Detection Module 303

[0134] In one embodiment, the timing behavior anomaly detection module 303 can perform the operations S110 to S150 described above.

[0135] The purpose of the time-series behavioral anomaly detection module 303 is to establish a unique, dynamic, and personalized "normal behavior pattern" baseline for a customer (such as a bank's corporate client), and to calculate in real time and quantitatively the degree of deviation of the customer's current behavior from this baseline. In one embodiment, the degree of deviation can be used to determine whether the customer is at risk of churn. In another embodiment, the degree of deviation can also be used as a novel feature (such as a "behavioral anomaly score"), concatenated with the customer feature vector output by the multi-dimensional feature engineering module 302, and input into the subsequent dynamic risk scoring model. By calculating the degree of deviation of a customer's behavior, the time-series behavioral anomaly detection module 303 can capture potential risk signals that are not yet significantly changing in value but whose behavioral patterns have been broken at an earlier stage, thereby elevating the early warning capability from "event response" to a higher dimension of "pattern recognition."

[0136] In one embodiment, the temporal behavior anomaly detection module 303 can employ a deep learning architecture based on a long short-term memory network autoencoder, and independently construct a temporal behavior anomaly detection model for each eligible customer, and can automatically perform model training and anomaly scoring. The execution flow of the temporal behavior anomaly detection module 303 may include customer screening and eligibility determination, dedicated data preparation and serialization, construction and training of a personalized autoencoder model based on a long short-term memory network, and finally, reconstruction error calculation and anomaly score generation in the prediction stage.

[0137] Specifically, the first step is customer screening and eligibility assessment. The effectiveness of customer screening and eligibility assessment relies on sufficient historical data to learn customers' "normal patterns." Specifically, a configurable minimum historical data length threshold can be set. The preferred value is 24 months, meaning each customer must have at least 24 consecutive months of history. When processing each customer C, the number of valid monthly records in the wide data table generated by the data acquisition and preprocessing module 301 can be checked first. It can be done only when... Only then will the customer be identified as a "target customer" and enter the subsequent anomaly detection process. For those that do not meet the requirements... For clients with specific conditions, a default value will be assigned to the generated abnormal behavior features, indicating that the degree of abnormality is unknown or inapplicable. This ensures the integrity of the feature matrix and avoids meaningless modeling when data is insufficient, thus ensuring the accuracy and stability of subsequent model training.

[0138] Next, we will prepare and serialize customer-specific data. The purpose of this preparation and serialization is to process the historical data of each target customer C, transforming it into a format that conforms to the input requirements of the Long Short-Term Memory (LSTM) network autoencoder model. First, we will extract one or more core time-series variables from the wide table of historical data of target customer C to construct the customer's behavioral patterns.

[0139] Let's take a bank's corporate clients as an example. Monthly total amount can be selected as the input feature. Next, to eliminate the influence of the client's business scale and improve the stability of model training, the selected historical sequence (the amount sequence of the past 36 months) of the target client C can be normalized, linearly scaling its values ​​to the [0,1] interval. The normalization process is calculated independently based on the historical data of each client; the minimum value of the data for each client C can be calculated. and maximum value These values ​​are stored independently to ensure model consistency and validity. This normalization scaler will be invoked again in subsequent prediction phases to ensure a one-to-one correspondence between the model and the client.

[0140] The standardized time series can then be converted into input / output pairs suitable for training a time-series behavioral anomaly detection model using a sliding window method. Each input / output pair consists of an input sample comprising a historical sequence of data from each of W consecutive time periods (e.g., 12 months) for the customer, and an output sample also comprising this historical sequence.

[0141] Specifically, for a length of The sequence can generate There are 12 training samples. The input for each sample is data points from 12 consecutive months. The expected output is also the data for those 12 months. The input for the second sample is... The expected output is also the data from these 12 months, and so on. Finally, a three-dimensional tensor of shape (number of samples, 12, 1) is generated as the training set for the Long Short-Term Memory network autoencoder model, thus providing the required data input for the subsequent training of the temporal behavior anomaly detection model.

[0142] Next, a personalized Long Short-Term Memory (LSM) network autoencoder model is constructed and trained. The output of the trained model is a time-series behavioral anomaly detection model. A dedicated deep learning model can be built and trained for each target customer C. Specifically, a symmetric autoencoder architecture based on stacked LSM network layers is used to compress the 12-step input time series into a low-dimensional, information-dense latent representation. The encoder's structure can be designed as follows: first, a 64-unit LSM network layer, followed by a 32-unit LSM network layer. The encoder only returns the output of the last time step.

[0143] To pass the encoder's output to the decoder, a repeating vector layer can be introduced, which copies the encoder's output. (e.g., 12 times) a tensor of shape (12, 32) is generated and used as the input to the decoder. The function of the decoder is to receive the encoded information and attempt to reconstruct the original input sequence. The decoder's structure is symmetrical to the encoder and can be designed as follows: first, a Long Short-Term Memory (LSTM) network layer containing 32 units, followed by a LSM network layer containing 64 units. At the end of the decoder, a temporal distribution layer is connected, which contains a fully connected layer and has one neuron. This layer independently maps the output of the decoder at each time step to a numerical value, thus obtaining a reconstructed sequence that is completely consistent with the shape of the original input sequence.

[0144] Before training the temporal behavior anomaly detection model, the model can be compiled. A dynamic learning rate optimizer can be used, which performs well in practice and can adaptively adjust the learning rate. The loss function is the mean squared error, and its calculation formula is as follows:

[0145]

[0146] Where S′ is the original input sequence, It is the reconstructed sequence output by the model.

[0147] During the training phase of the time-series behavioral anomaly detection model, the training function can be called, simultaneously inputting the generated sequence samples as both input (x) and target (y) into the model. The number of training iterations and the batch size are configurable parameters. The time-series behavioral anomaly detection model continuously adjusts its internal weights through the backpropagation algorithm, aiming to minimize the mean squared error loss function. After training, the customized time-series behavioral anomaly detection model based on a long short-term memory network autoencoder can learn the customer's "normal" monetary fluctuation pattern, which can be serialized and stored in a model library, indexed by the customer's unique identifier.

[0148] Finally, the error calculation and behavioral anomaly score generation are reconstructed. This step can be performed in a production environment. For example, it can be done when it is necessary to calculate the risk for target customer C in a new month t. First, the latest data for target customer C before time t can be obtained. (12) Monthly Amount Sequence Then, a minimum-maximum scaler specific to target client C can be loaded, and this scaler can be used to scale the sequence. Standardization is performed to obtain the original standard sequence. Next, we can load the pre-trained temporal behavior anomaly detection model specifically for target customer C, and convert the original standard sequence... After inputting the model, the reconstructed sequence output by the model is obtained. Next, the original normalized sequence can be calculated. With reconstruction sequence The mean squared error between the two values ​​reflects the deviation between a customer's current behavior and historical behavior patterns. The calculated mean squared error value can be used as the abnormal behavior score for target customer C in month t. The larger the value, the greater the deviation of the customer's behavior pattern from historical norms, thus indicating the degree of "abnormality" in the behavior.

[0149] Through the above process, the time-series behavioral anomaly detection module 303 can independently construct a time-series behavioral anomaly detection model for each customer. This model can be similar to a highly sensitive "behavioral pattern monitor." For example, when a bank's corporate customer is a software company, a stable software company's transactions fluctuate within a small range over a long period. Its dedicated time-series behavioral anomaly detection model will learn this stability. Once a small but rare jump occurs in the transaction amount in a certain month, the model will have difficulty reconstructing, resulting in a high anomaly score. Conversely, when a bank's corporate customer is a construction company, due to the seasonal nature of construction companies' employment, its transactions experience cyclical large fluctuations every year. Its dedicated time-series behavioral anomaly detection model will also learn this seasonal pattern. Therefore, even if the transaction amount with the customer drops significantly during the off-season, as long as it conforms to its historical pattern, the model can be reconstructed well, resulting in a low anomaly score. This personalized, pattern recognition-based time-series behavioral anomaly detection model can achieve early and accurate risk warnings.

[0150] (iv) Dynamic Risk Scoring Model Module 304

[0151] In some embodiments, the dynamic risk scoring model module 304 may perform operations S250 to S270 as described above.

[0152] In one embodiment, a gradient boosting decision tree framework can be used to construct a dynamic risk scoring model. This framework offers strong interpretability, outputting a ranking of feature importance to help business personnel understand the model's decision-making rationale and improve its transparency. Furthermore, the gradient boosting decision tree framework exhibits high robustness, effectively handling outliers and noise in features, and its built-in missing value handling capabilities enable it to cope with various challenges in complex data environments. More importantly, the gradient boosting decision tree framework implements a highly optimized distributed gradient boosting algorithm, supports parallel computation, and boasts extremely fast training speeds, making it suitable for large-scale data processing and real-time prediction tasks.

[0153] The target variable for the dynamic risk scoring model is defined. Defining "churn" is crucial for model training. In one embodiment, a composite, future-oriented labeling method can be used. For example, if customer C is labeled as a "churned sample" (Label=1) at time t, then this customer is considered churned if any of the following events occur within the next K months: "Hard churn: Customer closes account or has not made any transactions for M consecutive months (e.g., M=3 months). Soft churn: Customer's average monthly spending for L consecutive months (e.g., L=3 months) decreases by more than a set threshold (e.g., 70%) compared to the average of the past 12 months, and does not recover."

[0154] If a customer does not meet the above conditions at time point t, they are labeled as a "normal sample" (Label=0). This label definition not only focuses on the customer's current behavior but also takes into account potential future churn events, thereby ensuring that the model can accurately predict future churn risks.

[0155] Gradient boosting decision tree framework: principle and derivation. It iteratively trains multiple weak learners (decision trees) and combines them to form a strong learner. For input samples... The model's predicted values Sum of all tree predictions:

[0156]

[0157] in Let be the prediction function for the k-th tree. Let this be the function space of all possible decision trees. The objective function of the gradient boosting decision tree framework consists of two parts: a loss function and a regularization term. The loss function measures the accuracy of the model's predictions, while the regularization term controls the model's complexity and prevents overfitting.

[0158]

[0159] Where L is the loss function, and logistic loss is used for binary classification problems. This is a regularization term. The definition is as follows:

[0160]

[0161] in The number of leaf nodes in the tree. Let be the weight of the j-th leaf node. and The hyperparameters are used to control the complexity of the trees. During training, the objective function is approximated through a second-order Taylor expansion and continuously optimized to obtain the optimal structure and weights for each tree.

[0162] The output of the dynamic risk scoring model is a churn risk score. The gradient boosting decision tree framework, after training, can output a dynamic risk scoring model. This dynamic risk scoring model can output a probability value for any new customer's feature vector x. This represents the probability that the customer will churn within the next K months. The feature vector x can be the feature vector output by the multi-dimensional feature engineering module 302, or a feature vector output by the multi-dimensional feature engineering module 302 with the degree of deviation from the module's output. To make the risk score more intuitive, this probability value can be converted into a percentage score.

[0163]

[0164] This score dynamically reflects the risk of customer churn within the next K months. The higher the score, the greater the likelihood of customer churn, thus providing a more accurate basis for predicting churn risk for business decisions.

[0165] (v) Adaptive Hierarchical Early Warning Module 305

[0166] This invention also provides an adaptive hierarchical early warning mechanism. First, the importance of customers is assessed. For example, a variation of the RFM concept can be used, combining multiple indicators such as customer business activity, cooperation stability, and overall contribution to automatically stratify customer importance.

[0167] In the customer segmentation process, customers can be classified based on the following three core indicators: R (Recency-like), which reflects the customer's recent business activity, such as the average amount spent in the past three months; F (Frequency-like), which measures the stability of the customer's cooperation, such as the number of business transactions in the past 12 months; and M (Monetary-like), which represents the customer's overall contribution. Based on these indicators, clustering algorithms such as K-Means or business rules can be used to divide customers into different tiers.

[0168] Next, building upon the tiered structure, an adaptive threshold mechanism can be incorporated to adjust the alert threshold based on the customer's tier. Higher importance tiers can have lower risk score thresholds. When a customer's risk score, output by the dynamic risk scoring model, exceeds the alert threshold corresponding to their tier, a churn alert is triggered. This mechanism ensures priority focus on churn risk for customers at important tiers while avoiding excessive interference with customers at other important tiers.

[0169] like Figure 6 As shown, the adaptive threshold early warning process may include: First, using a dynamic risk scoring model to calculate the customer's risk score based on their feature vector and obtaining the customer's tier. Next, comparing the customer's risk score with the early warning threshold for their tier; if the customer's risk score exceeds the tier's early warning threshold, a churn warning will be triggered.

[0170] Once a churn warning is triggered, System 300 can adopt different notification strategies (including the content and recipients of the notification) based on the customer's tier. If the customer's risk score does not reach the threshold for their tier, no warning will be triggered. This mechanism enables precise risk warning management based on customer importance, optimizes resource allocation, and improves management efficiency.

[0171] Compared with the prior art, the method of the present invention can have the following beneficial effects:

[0172] First, it can improve the accuracy of customer churn prediction and recall rate. By integrating multi-dimensional customer characteristics and utilizing machine learning technology, it is possible to more accurately identify the true risk of churn.

[0173] Secondly, it enables early warning and proactive intervention of risks. For example, by capturing subtle deviations in customer behavior patterns early on through time-series behavioral anomaly detection models built independently for each customer, rather than relying on drastic deterioration of business indicators, it can significantly advance the timing of risk warnings. This provides banks and other service institutions with ample time to take proactive customer retention measures, thereby effectively reducing the risk of customer churn.

[0174] Furthermore, it achieves quantitative risk management and optimized resource allocation. By using a dynamic risk scoring model to output continuous risk scores instead of traditional binary early warning signals, the severity of risks can be clearly quantified, further enabling refined risk management. Combined with an adaptive customer segmentation early warning mechanism, it can, to a certain extent, ensure that customer relationship management resources are accurately targeted to important customer groups, thus optimizing resource allocation.

[0175] Finally, the adaptability and intelligence of risk warning are enhanced. A machine learning model is employed, possessing continuous retraining capabilities, enabling it to learn new churn patterns and automatically adapt to changes in the market environment and customer behavior. Compared to hard-coded rules that rely on frequent manual adjustments, the method of this invention has a higher level of intelligence and lower maintenance costs, enhancing the system's adaptability and flexibility.

[0176] Based on the customer churn early warning method of this invention, this invention also provides a customer churn early warning device. In some embodiments, the customer churn early warning device may include the time-series behavior anomaly detection module 303 in the system 300 described above. In other embodiments, the customer churn early warning device may also include other modules (or parts of the functions of a certain module) in the system 300 other than the time-series behavior anomaly detection module 303.

[0177] According to embodiments of the present invention, any multiple modules among the data acquisition and preprocessing module 301, multi-dimensional feature engineering module 302, time-series behavior anomaly detection module 303, dynamic risk scoring model module 304, and adaptive hierarchical early warning module 305 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of the present invention, at least one of the data acquisition and preprocessing module 301, multi-dimensional feature engineering module 302, time-series behavior anomaly detection module 303, dynamic risk scoring model module 304, and adaptive hierarchical early warning module 305 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or any other reasonable means of integrating or packaging circuits, or implemented in hardware or firmware, or in any one of software, hardware, and firmware implementations, or in a suitable combination of any of these. Alternatively, at least one of the data acquisition and preprocessing module 301, multidimensional feature engineering module 302, time-series behavior anomaly detection module 303, dynamic risk scoring model module 304, and adaptive hierarchical early warning module 305 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0178] Figure 7 A block diagram schematically illustrates an electronic device suitable for implementing a customer churn warning method according to an embodiment of the present invention.

[0179] like Figure 7As shown, an electronic device 900 according to an embodiment of the present invention includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage portion 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include onboard memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0180] RAM 903 stores various programs and data required for the operation of electronic device 900. Processor 901, ROM 902, and RAM 903 are interconnected via bus 904. Processor 901 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 902 and / or RAM 903. It should be noted that the program may also be stored in one or more memories other than ROM 902 and RAM 903. Processor 901 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in one or more memories.

[0181] According to an embodiment of the present invention, the electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to a bus 904. The electronic device 900 may also include one or more of the following components connected to the input / output (I / O) interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 910 as needed so that computer programs read from it can be installed into the storage section 908 as needed.

[0182] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0183] According to embodiments of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, a computer-readable storage medium may include ROM 902 and / or RAM 903 and / or one or more memories other than ROM 902 and RAM 903 described above.

[0184] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of the present invention.

[0185] When the computer program is executed by the processor 901, it performs the functions defined in the system / apparatus of this invention. According to embodiments of the invention, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0186] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 909, and / or installed from a removable medium 911. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0187] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 909, and / or installed from the removable medium 911. When the computer program is executed by the processor 901, it performs the functions defined in the system of this embodiment of the invention. According to embodiments of the invention, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0188] Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

Claims

1. A customer churn early warning method, comprising: Having obtained authorization from the target customer to use their relevant data, the target customer's data is obtained for each of the W consecutive time periods up to the current time, to obtain the target sequence, where W is an integer greater than 1; The target sequence is standardized to obtain the original standard sequence; The original standard sequence is input into the target temporal behavior anomaly detection model, and the reconstructed sequence output by the target temporal behavior anomaly detection model is obtained. The target temporal behavior anomaly detection model is a temporal behavior anomaly detection model independently constructed for the target customer. Different temporal behavior anomaly detection models are constructed for different customers. By comparing the original standard sequence and the reconstructed sequence, the degree of deviation is obtained; and When the deviation meets the abnormal alarm conditions, a churn warning for the target customer is triggered.

2. The method according to claim 1, wherein, The time-series anomaly detection model, independently built for each customer, is constructed using a deep learning architecture based on a long short-term memory network autoencoder. The target training dataset for this model is derived from the historical data of the target customer. Each input sample in the target training dataset includes a historical sequence consisting of data from the target customer over W consecutive time periods; the output sample corresponding to each input sample includes the historical sequence.

3. The method according to claim 1, wherein, The comparison between the original standard sequence and the reconstructed sequence to obtain the degree of deviation includes: Calculate the mean square error between the original normalized sequence and the reconstructed sequence, wherein the mean square error characterizes the degree of deviation.

4. The method according to claim 1, wherein, The standardization process for the target sequence to obtain the original standard sequence includes: Based on the maximum and minimum values ​​of the target customer's historical data, the data in the target sequence is normalized to obtain the original standard sequence.

5. The method according to claim 1, wherein, The step of triggering the churn warning for the target customer when the deviation meets the abnormal alarm conditions includes: The degree of deviation is added to the feature vector of the target customer to obtain the target feature vector; The target feature vector is input into the trained dynamic risk scoring model, and the target risk score output by the dynamic risk scoring model is obtained. When the target risk score is higher than the preset target warning threshold, a churn warning for the target customer is triggered.

6. The method according to claim 5, wherein, W is configured to include multiple optional values, wherein multiple deviation levels are obtained for each of the multiple optional values; adding the deviation levels to the feature vector of the target customer to obtain the target feature vector includes: Multiple deviation levels are added to the feature vector of the target customer to obtain the target feature vector.

7. The method according to claim 5, wherein, The method further includes: Obtain the feature vector of the target customer, wherein the feature vector of the target customer includes n-dimensional features, which are obtained through feature fusion processing, where n is an integer greater than or equal to 1; The feature fusion process includes: Given that customers have authorized the use of their historical data, retrieve the historical data of m customers, where m is an integer greater than 1; Extract h-dimensional feature data for each customer based on their historical data, where h is an integer greater than or equal to n; Based on the h-dimensional features of the m customers, construct the Pearson correlation coefficient matrix among the h-dimensional features; Based on the Pearson correlation coefficient matrix, any pair of features whose absolute value of the Pearson correlation coefficient exceeds a preset correlation threshold is merged into one feature to finally obtain the n-dimensional feature.

8. The method according to claim 7, wherein, The step of merging any pair of features whose absolute value of the Pearson correlation coefficient exceeds a preset correlation threshold into one feature includes: One feature to be fused is obtained when the absolute value of the Pearson correlation coefficient exceeds a preset correlation threshold. Obtain the correlation between each of the features to be fused and the risk score output by the dynamic risk scoring model; The feature with higher correlation among the two features to be fused is retained, and the other feature is removed.

9. The method according to claim 5, wherein, Before triggering a churn warning for the target customer when the target risk score exceeds a preset target warning threshold, the method further includes: Obtain the target customer's tier to obtain the target tier; and Obtain the preset warning threshold corresponding to the target level to obtain the target warning threshold, wherein the warning thresholds corresponding to different levels are different.

10. A customer churn early warning device, comprising: A temporal behavior anomaly detection module is included, which deploys different temporal behavior anomaly detection models independently built for different clients. The module comprises a serialization unit, a model prediction unit, and a reconstruction error calculation and behavior anomaly score generation unit. The serialization unit is used to: obtain data of the target customer in each of W consecutive time periods up to the current time, upon obtaining authorization from the target customer to use its related data, to obtain a target sequence; and to perform standardization processing on the target sequence to obtain an original standard sequence, wherein W is an integer greater than 1; The model prediction unit is used to: input the original standard sequence into the target time-series behavior anomaly detection model and obtain the reconstructed sequence output by the target time-series behavior anomaly detection model, wherein the target time-series behavior anomaly detection model is a time-series behavior anomaly detection model independently constructed for the target customer; The reconstruction error calculation and behavior anomaly score generation unit is used to: compare the original standard sequence and the reconstructed sequence to obtain the degree of deviation, and trigger the churn warning of the target customer when the degree of deviation meets the anomaly alarm conditions.

11. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program or instructions stored thereon, wherein, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 9.

13. A computer program product comprising a computer program or instructions, wherein, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 9.

Images