Customer label determination method and related equipment

Abstract

The invention provides a customer label determination method, and the label of a customer is used for indicating whether the customer is a potential lost customer. According to the method, features of a plurality of time sequences in customer historical activity data are extracted, a feature graph structure is constructed, and a customer loss prediction result is obtained based on a customer loss prediction model capable of aggregating node feature information, edge feature information and graph structure information in the feature graph structure. According to the method, the correlation among multiple dimensions in the customer historical activity data is effectively utilized for analysis, the customer loss probability is predicted, and the prediction accuracy is improved, so that a bank can pertinently execute a retention strategy on the customer which may be lost, and the loss is reduced.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a customer label determination method, a customer label determination device, and a computer device. Background Art

[0002] A bank's customers are one of its most important assets. Customer churn may cause the bank to lose its stable customer base, affecting its profitability and long-term development. Therefore, timely and accurate prediction of customer churn is crucial for banks to maintain a stable customer base, reduce operating costs, optimize products and services, and improve market competitiveness.

[0003] The customer's historical activity data reflects a variety of information such as the customer's transaction behavior, preference changes, active cycle, and interaction pattern. By analyzing these historical activity data, banks can identify customers' transaction habits, value contribution, loyalty, satisfaction, and potential needs, and then predict the probability of customer churn. For example, signs such as a decrease in customer transaction frequency, a decrease in the number of interactions, and a cold response to promotional activities in historical activity data often indicate an increase in the risk of customer churn. Therefore, banks should choose appropriate methods to predict customer churn based on customers' historical activity data.

[0004] However, the prediction effect of customer churn in related technologies is poor, which makes it impossible for banks to formulate targeted retention strategies, resulting in losses. Summary of the invention

[0005] The present application provides a method for determining a customer label, wherein the customer label is used to indicate whether the customer is a potential churn customer. The method can solve the problem in related technologies that the prediction effect of customer churn is poor, making it impossible for banks to formulate targeted retention strategies.

[0006] In a first aspect, the present application provides a method for determining a customer tag. The method comprises:

[0007] Acquire historical activity data of a customer, wherein the historical activity data includes a plurality of time series, each of which includes time series data of a financial activity of the customer within a target time period;

[0008] Decomposing the multiple time series respectively, extracting the features of each of the multiple time series, wherein the features include a trend item and at least one feature sequence;

[0009] Constructing a feature graph structure according to the features, the feature graph structure comprising a node matrix, an edge feature and an adjacency matrix, wherein the node in the t-th row and the n-th column of the node matrix comprises t historical values ​​in the n-th time series, the edge feature is determined according to the correlation between the feature sequences, and the adjacency matrix is ​​determined according to the correlation between the trend items of each of the time series;

[0010] The feature graph structure is input into a customer churn prediction model to obtain a first customer churn prediction result, wherein the first customer churn prediction result is used to determine whether the customer's label is a potential churn customer, wherein the customer churn prediction model includes an encoder, wherein the encoder includes a self-attention layer, and the self-attention layer is used to respectively aggregate feature information of a node matrix, edge features, and feature information of an adjacency matrix in the feature graph structure to obtain fused feature information.

[0011] In some possible implementations, the method further includes:

[0012] A recurrent neural network is used as a position encoding layer of the node matrix, and the position encoding layer is used to add a position vector to each node in the node matrix, and the position vector is used to indicate the order of the historical values ​​in each of the time series.

[0013] In some possible implementations, the method further includes:

[0014] The output of the self-attention layer is processed based on a causal convolutional neural network.

[0015] In some possible implementations, the method further includes:

[0016] Inputting the historical activity data into an autoregressive linear model to obtain a second customer churn prediction result;

[0017] According to the first customer churn prediction result and the second customer churn prediction result, it is determined whether the customer is labeled as a potential churn customer.

[0018] In some possible implementations, the method further includes:

[0019] The historical activity data is preprocessed, wherein the preprocessing includes at least one of missing value processing, outlier processing, format conversion or normalization processing.

[0020] In some possible implementations, the method further includes:

[0021] Based on the explanatory machine learning method, the key factors of customer churn are obtained, and an analysis report is generated according to the prediction results and the key factors of customer churn to implement a customer retention strategy.

[0022] In a second aspect, the present application provides a customer tag determination device. The device includes:

[0023] An acquisition module, used to acquire historical activity data of a customer, wherein the historical activity data includes a plurality of time series, each of which includes time series data of a financial activity of the customer within a target time period;

[0024] A feature extraction module, used to decompose the multiple time series respectively, and extract the feature of each of the multiple time series, wherein the feature includes a trend item and at least one feature sequence;

[0025] The feature extraction module is further used to construct a feature graph structure according to the features, wherein the feature graph structure includes a node matrix, an edge feature and an adjacency matrix, wherein the node in the t-th row and n-th column of the node matrix includes t historical values ​​in the n-th time series, the edge feature is determined according to the correlation between the feature sequences, and the adjacency matrix is ​​determined according to the correlation between the trend items of each of the time series;

[0026] A prediction module is used to obtain a first customer churn prediction result according to a customer churn prediction model, wherein the first customer churn prediction result is used to determine whether the customer's label is a potential churn customer, wherein the customer churn prediction model includes an encoder, wherein the encoder includes a self-attention layer, and the self-attention layer is used to respectively aggregate feature information of a node matrix, edge features, and feature information of an adjacency matrix in the feature graph structure to obtain fused feature information.

[0027] In some possible implementations, the prediction module is further used to:

[0028] In the customer churn prediction model, a recurrent neural network is used as a position encoding layer of the node matrix, and the position encoding layer is used to add a position vector to each node in the node matrix, and the position vector is used to indicate the order of the historical values ​​in each time series.

[0029] In some possible implementations, the prediction module is further used to:

[0030] In the customer churn prediction model, the output of the self-attention layer is processed based on a causal convolutional neural network.

[0031] In some possible implementations, the prediction module is further used to:

[0032] According to the autoregressive linear model and historical activity data, the second customer churn prediction result is obtained;

[0033] According to the first customer churn prediction result and the second customer churn prediction result, it is determined whether the customer is labeled as a potential churn customer.

[0034] In some possible implementations, the device further includes a preprocessing module, which is specifically configured to:

[0035] The historical activity data is preprocessed, wherein the preprocessing includes at least one of missing value processing, outlier processing, format conversion or normalization processing.

[0036] In some possible implementations, the device further includes an analysis module, which is specifically configured to:

[0037] Based on the explanatory machine learning method, the key factors of customer churn are obtained, and an analysis report is generated according to the prediction results and the key factors of customer churn to implement a customer retention strategy.

[0038] In a third aspect, the present application provides a device, wherein the device includes a processor and a memory.

[0039] The processor is used to execute instructions stored in the memory, so that the device performs the customer tag determination method as described in the first aspect or any implementation manner of the first aspect.

[0040] In a fourth aspect, the present application provides a computer-readable storage medium. The computer-readable storage medium includes instructions, which, when executed on a device, enable the device to execute the customer tag determination method as described in the first aspect or any implementation of the first aspect.

[0041] In a fifth aspect, the present application provides a computer program product comprising instructions, which, when executed on a device, enables the device to execute the customer tag determination method described in the first aspect or any one of the implementations of the first aspect.

[0042] Based on the implementations provided in the above aspects, this application can also be further combined to provide more implementations.

[0043] It can be seen from the above technical solutions that the embodiments of the present application have at least the following advantages:

[0044] This method extracts the features of multiple time series in the customer's historical activity data, constructs a feature graph structure, and obtains customer churn prediction results based on a customer churn prediction model that can aggregate the feature information in the feature graph structure. It effectively utilizes the correlation analysis between multiple dimensions and predicts the probability of customer churn, thereby improving the accuracy of the prediction, so that banks can implement retention strategies for customers who may churn in a targeted manner and reduce losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0046] Figure 1 A flowchart of a method for determining a customer tag disclosed in an embodiment of the present application;

[0047] Figure 2 A schematic diagram of the structure of a customer label determination device disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0048] The scheme in the embodiments provided in this application will be described below in conjunction with the drawings in this application.

[0049] The terms used in the following embodiments are only for the purpose of describing specific embodiments and are not intended to be limiting of the present application. The terms "first" and "second" in the embodiments of the present application are only used for descriptive purposes and cannot be understood as indicating or implying relative importance, operation time sequence, or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0050] In order to facilitate understanding of the technical solution of the present application, the background technology involved in the present application is first introduced in detail.

[0051] A bank's customers are one of its most important assets. Customer churn may cause the bank to lose its stable customer base, affecting the bank's profitability and long-term development. Predicting customer churn is crucial for banks to maintain a stable customer base, reduce operating costs, optimize products and services, and improve market competitiveness.

[0052] Traditional statistical methods often use the Auto Regressive Integrated Moving Average (ARIMA) model or exponential smoothing method to analyze the historical activity data of customers and predict the probability of customer churn. Among them, the ARIMA model combines the characteristics of the Auto Regressive (AR) and Moving Average (MA) models, and uses historical data to predict future values ​​by identifying the autocorrelation structure in the time series. The ARIMA model assumes that the relationship in the time series is linear. However, the historical activity data of customers is usually time series data, which is affected by many factors, such as the customer's personal economic situation, customer consumption or investment preferences, market environment, policy changes, economic cycles, etc., and has obvious nonlinear characteristics. Faced with complex and changeable time series data, the ARIMA model may not be able to adapt to changes in the data in a timely manner, resulting in inaccurate prediction results. The exponential smoothing method reduces the random fluctuations of the data by exponentially smoothing the level, trend and seasonal components in the time series. However, similar to the ARIMA model, the exponential smoothing method also assumes that the relationship in the time series is linear, which limits its ability to capture nonlinear relationships.

[0053] Deep learning methods can extract data features and capture nonlinear patterns between time series, and are widely used in time series tasks. For example, recurrent neural networks (RNNs) are used. By introducing recurrent connections, recurrent neural networks enable neural networks to store historical information, thereby having the ability to process time series data. Among them, recurrent neural networks include long short-term memory networks (LSTMs), which can capture short-term and long-term patterns in sequence inputs and alleviate the gradient vanishing and explosion problems in recurrent neural network models. However, customers' historical activity data usually contains data in multiple attribute dimensions, indicating a variety of financial activities of customers. For example, customers' historical activity data can include a sequence of customer financial data changing over time, a sequence of customer savings data changing over time, a sequence of customer loan data changing over time, and so on. There are complex correlations between these financial activities. However, the relevant deep learning network only considers the time correlation of the data sequence itself, and cannot analyze and utilize the correlation between the various dimensions of the data, so the prediction results of customer churn are not accurate enough.

[0054] In view of this, the present application provides a method for determining a customer label, wherein the customer label is used to indicate whether the customer is a potential churn customer. The method can solve the problem that the prediction effect of customer churn in the related art is poor, making it impossible for banks to make targeted retention strategies. The method includes:

[0055] The customer label determination device obtains the customer's historical activity data, the historical activity data includes multiple time series, each time series includes the time series data of a financial activity of the customer in the target time period, and decomposes the multiple time series respectively, extracts the features of each time series in the multiple time series, wherein the features include a trend item and at least one feature sequence. Then, a feature graph structure is constructed according to the features, the feature graph structure includes a node matrix, an edge feature and an adjacency matrix, wherein the node in the tth row and the nth column in the node matrix includes t historical values ​​in the nth time series, the edge feature is determined according to the correlation between the feature sequences, and the adjacency matrix is ​​determined according to the correlation between the trend items of each time series. Then, the feature graph structure is input into the customer churn prediction model to obtain the first customer churn prediction result, and the first customer churn prediction result is used to determine whether the customer's label is a potential churn customer. Among them, the customer churn prediction model includes an encoder, the encoder includes a self-attention layer, and the self-attention layer is used to respectively aggregate the feature information of the node matrix, the edge feature and the feature information of the adjacency matrix in the feature graph structure to obtain fused feature information.

[0056] This method extracts the features of multiple time series in the customer's historical activity data, constructs a feature graph structure, and obtains customer churn prediction results based on a customer churn prediction model that can aggregate the feature information in the feature graph structure. It effectively utilizes the correlation analysis between multiple dimensions and predicts the probability of customer churn, thereby improving the accuracy of the prediction, so that banks can implement retention strategies for customers who may churn in a targeted manner and reduce losses.

[0057] Next, the customer label determination method provided in the embodiment of the present application will be introduced with reference to the accompanying drawings.

[0058] See also Figure 1 This is a flow chart of a method for determining a customer tag disclosed in an embodiment of the present application, the method comprising:

[0059] S102: Acquire historical activity data of the customer, where the historical activity data includes multiple time series.

[0060] The customer's historical activity data includes multiple time series, each of which includes the time series data of a financial activity of the customer within the target time period. For example, the customer's historical activity data may include the sequence of the customer's financial data changing over time, the sequence of the customer's savings data changing over time, the sequence of the customer's loan data changing over time, etc. There are complex correlations between these financial activities, indicating information on various dimensions of the customer's historical activities.

[0061] In some possible implementations, the input data needs to be preprocessed first to ensure data quality. The preprocessing operation includes at least one of missing value processing, outlier processing, format conversion or normalization processing.

[0062] Missing value processing refers to the processing of missing values ​​in data by methods such as deletion, filling or interpolation.

[0063] Outlier processing refers to the use of methods such as deletion, replacement or correction to deal with outliers in data. Outlier detection can be based on statistical methods (such as the 3σ principle, box plots, etc.) or machine learning methods (such as clustering, classification, etc.).

[0064] In deep learning algorithms, data needs to be recognizable and needs to be converted from other formats into character data. For example, one-hot encoding can be used for format conversion. One-hot encoding can assign a unique binary vector to each data in other formats, where one bit of the binary vector is 1 and the rest are 0.

[0065] Normalization refers to scaling the data features to eliminate the impact of dimension, which can increase the robustness of the model. For example, normalization methods such as minimum-maximum normalization and zero-mean normalization can be used.

[0066] It should be noted that the specific selection of the above preprocessing method needs to be determined according to the actual situation such as the characteristics of the input data, and the embodiments of the present application do not impose any restrictions on this.

[0067] S104: Decompose the multiple time series respectively, extract the features of each time series in the multiple time series, and construct a feature graph structure according to the features.

[0068] The customer tag determination device can decompose the multiple time series respectively, and extract the features of each time series in the multiple time series, wherein the features include a trend item and at least one feature sequence.

[0069] In some possible implementations, the empirical mode decomposition method can be used to decompose and extract features from the input data. The specific formula of this method is:

[0070]

[0071] f i (t) = A i (t)cos(2πφ i (t))

[0072] Among them, x(t) is a time series, φi( t ) refers to the non-stationary phase function, whose derivative is the instantaneous frequency, Ai( t ) is the instantaneous amplitude, and R(t) is the trend term. Then the image is constructed based on the features extracted by empirical mode decomposition.

[0073] The feature graph structure includes a node matrix, edge features, and an adjacency matrix. The node in the tth row and nth column of the node matrix includes t historical values ​​in the nth time series. In this way, it contains some repeated structures and can provide sufficient data to accurately extract local features.

[0074] The edge feature is determined according to the correlation between feature sequences. Specifically, it is determined according to the following formula:

[0075]

[0076] The adjacency matrix A is determined based on the correlation between the trend items of each time series, which can be specifically determined according to the formula:

[0077]

[0078] Among them, a ij Defines the correlation between two adjacent items in the adjacency matrix, c is the threshold, p ij is the correlation between the trend items of two time series. The adjacency matrix can indicate the characteristic information in the graph structure.

[0079] S106: Input the feature graph structure into the customer churn prediction model to obtain a first customer churn prediction result, where the first customer churn prediction result is used to determine whether the customer's label is a potential churn customer.

[0080] The customer churn prediction model can be a deep learning model based on the attention mechanism, such as a Transformer model. The Transformer model includes an encoder and a decoder, wherein the encoder includes a self-attention layer and a feed forward neural network blayer (FFN).

[0081] The self-attention layer can capture the dependencies in the input sequence by calculating the correlation between the node in the input sequence and all other nodes. In the self-attention mechanism, the input sequence is first converted into three matrices: query, key, and value. Then, the dot product of the query matrix and the key matrix is ​​calculated, and the softmax function is applied to obtain the attention weights. Finally, these weights are used to perform weighted summation on the value matrix to obtain the output representation. In order to improve the efficiency of the model, the Transformer model can use a multi-head self-attention mechanism to process the input data in parallel. It divides the input sequence into multiple heads (usually 8 or 16), each head performs self-attention calculations independently, and then concatenates the results and obtains the final output representation through linear transformation.

[0082] In order to enable the Transformer model to utilize the feature graph structure constructed above, the multi-head attention algorithm of the model needs to be modified as follows:

[0083] The correlation matrix of the kth feature sequence obtained by decomposition integrates the edge features; A is the adjacency matrix, which is used to integrate graph structure information.

[0084] In this way, the modified Transformer model can utilize the correlation characteristics between various dimensions in the customer's historical activity data to make the prediction results more accurate, helping banks to predict potential lost customers in advance and then adopt corresponding retention strategies.

[0085] The attention mechanism ignores the order of elements in the sequence and treats each input element in the same way. In some possible implementations, in order to utilize the order information, a position vector can be added to each element in the input sequence. Specifically, the recurrent neural network can be used as a position encoding layer of the node matrix, where the position encoding layer is used to add a position vector to each node in the node matrix, and the position vector is used to indicate the order of historical values ​​in each time series. In this way, the timing information between data in the same time series can be used to refine the feature relationship within each dimension of the customer's historical activity data.

[0086] In some possible implementations, the output of the self-attention layer can also be processed based on a causal convolutional neural network in a transformer model. The value of the causal convolution output at time point t depends only on the input value before time point t, ensuring that the output of the model will not be affected by future information and can also be easily parallelized.

[0087] The deep learning model has a nonlinear property, which causes the scale of the model output to be insensitive to the scale of the input. However, the scale of the input signal is constantly changing in a non-periodic manner, which will affect the prediction accuracy of the deep learning model. In some possible implementations, a linear model can be introduced into the prediction process, thereby utilizing linear features on the basis of deep learning, making the method more robust. Specifically, historical activity data can be input into an autoregressive linear model to obtain a second customer churn prediction result, and then based on the first customer churn prediction result and the second customer churn prediction result, it is determined whether the customer's label is a potential churn customer. For example, the first customer churn prediction result and the second customer churn prediction result can be weighted and summed, and the embodiments of the present application do not impose any restrictions on this.

[0088] In some possible implementations, the customer label determination device can also obtain key factors of customer churn based on an explanatory machine learning method, and generate an analysis report based on the prediction results and the key factors of customer churn to implement a customer retention strategy. For example, the SHAP method can be used to calculate the contribution of different features to the prediction results, and to provide an explanation of the model prediction results.

[0089] Based on the above description, the embodiment of the present application discloses a method for determining a customer label, wherein the customer label is used to indicate whether the customer is a potential churn customer. The method extracts the features of multiple time series in the customer's historical activity data, constructs a feature graph structure, and obtains a customer churn prediction result based on a customer churn prediction model that can aggregate feature information in the feature graph structure. It effectively utilizes the correlation analysis between multiple dimensions, predicts the probability of customer churn, and improves the accuracy of the prediction, so that the bank can implement a retention strategy for customers who may churn in a targeted manner and reduce losses.

[0090] The above is a detailed introduction to the customer tag determination method provided in the embodiment of the present application. Next, the customer tag determination device and equipment provided in the embodiment of the present application will be introduced in conjunction with the accompanying drawings. The customer tag determination device can be any terminal having the aforementioned feature extraction function and model running function. Of course, in some possible implementations, the customer tag determination device can also be a server.

[0091] See also Figure 2 A customer tag determination device 200 is shown. The device comprises:

[0092] An acquisition module 202 is used to acquire historical activity data of a customer, where the historical activity data includes multiple time series, each of which includes time series data of a financial activity of the customer within a target time period;

[0093] A feature extraction module 204 is used to decompose the multiple time series respectively and extract the features of each time series in the multiple time series, wherein the features include a trend item and at least one feature sequence;

[0094] The feature extraction module 204 is further used to construct a feature graph structure according to the features, the feature graph structure including a node matrix, an edge feature and an adjacency matrix, the node in the t-th row and the n-th column in the node matrix includes t historical values ​​in the n-th time series, the edge feature is determined according to the correlation between the feature sequences, and the adjacency matrix is ​​determined according to the correlation between the trend items of each time series;

[0095] The prediction module 206 is used to obtain a first customer churn prediction result according to the customer churn prediction model, and the first customer churn prediction result is used to determine whether the customer's label is a potential churn customer. The customer churn prediction model includes an encoder, and the encoder includes a self-attention layer. The self-attention layer is used to respectively aggregate the feature information of the node matrix, the edge feature and the feature information of the adjacency matrix in the feature graph structure to obtain fused feature information.

[0096] In some possible implementations, the prediction module 206 is further configured to:

[0097] In the customer churn prediction model, the recurrent neural network is used as the position encoding layer of the node matrix. The position encoding layer is used to add a position vector to each node in the node matrix. The position vector is used to indicate the order of historical values ​​in each time series.

[0098] In some possible implementations, the prediction module 206 is further configured to:

[0099] In the customer churn prediction model, the output of the self-attention layer is processed based on a causal convolutional neural network.

[0100] In some possible implementations, the prediction module 206 is further configured to:

[0101] According to the autoregressive linear model and historical activity data, the second customer churn prediction result is obtained;

[0102] According to the first customer churn prediction result and the second customer churn prediction result, it is determined whether the customer is labeled as a potential churn customer.

[0103] In some possible implementations, the device further includes a preprocessing module, which is specifically configured to:

[0104] The historical activity data is preprocessed, where the preprocessing includes at least one of missing value processing, outlier processing, format conversion or normalization processing.

[0105] In some possible implementations, the device further includes an analysis module, which is specifically configured to:

[0106] Based on the explanatory machine learning method, the key factors of customer churn are obtained, and an analysis report is generated based on the prediction results and the key factors of customer churn to implement customer retention strategies.

[0107] The present application provides a device for implementing a customer tag determination method. The device includes a processor and a memory. The processor and the memory communicate with each other. The processor is used to execute instructions stored in the memory so that the device executes the customer tag determination method.

[0108] The present application provides a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is run on a device, the device executes the above-mentioned customer tag determination method.

[0109] The present application provides a computer program product comprising instructions, which, when executed on a device, enables the device to execute the above-mentioned customer tag determination method.

[0110] It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

[0111] Through the description of the above implementation mode, the technicians in the field can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present application, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0112] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0113] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

[0114] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for determining a customer label, characterized in that: The method comprises: Acquire historical activity data of a customer, wherein the historical activity data includes a plurality of time series, each of which includes time series data of a financial activity of the customer within a target time period; Decomposing the multiple time series respectively, extracting the features of each of the multiple time series, wherein the features include a trend item and at least one feature sequence; Constructing a feature graph structure according to the features, the feature graph structure comprising a node matrix, an edge feature and an adjacency matrix, wherein the node in the t-th row and the n-th column of the node matrix comprises t historical values ​​in the n-th time series, the edge feature is determined according to the correlation between the feature sequences, and the adjacency matrix is ​​determined according to the correlation between the trend items of each of the time series; The feature graph structure is input into a customer churn prediction model to obtain a first customer churn prediction result, wherein the first customer churn prediction result is used to determine whether the customer's label is a potential churn customer, wherein the customer churn prediction model includes an encoder, wherein the encoder includes a self-attention layer, and the self-attention layer is used to respectively aggregate feature information of a node matrix, edge features, and feature information of an adjacency matrix in the feature graph structure to obtain fused feature information.

2. The method according to claim 1, characterized in that: The method further comprises: A recurrent neural network is used as a position encoding layer of the node matrix, and the position encoding layer is used to add a position vector to each node in the node matrix, and the position vector is used to indicate the order of the historical values ​​in each of the time series.

3. The method according to claim 1, characterized in that The method further comprises: The output of the self-attention layer is processed based on a causal convolutional neural network.

4. The method according to claim 1, characterized in that: The method further comprises: Inputting the historical activity data into an autoregressive linear model to obtain a second customer churn prediction result; According to the first customer churn prediction result and the second customer churn prediction result, it is determined whether the customer is labeled as a potential churn customer.

5. The method according to any one of claims 1 to 4, characterized in that The method further comprises: The historical activity data is preprocessed, wherein the preprocessing includes at least one of missing value processing, outlier processing, format conversion or normalization processing.

6. The method according to any one of claims 1 to 4, characterized in that The method further comprises: Based on the explanatory machine learning method, the key factors of customer churn are obtained, and an analysis report is generated according to the prediction results and the key factors of customer churn to implement a customer retention strategy.

7. A customer label determination device, characterized in that: The device comprises: An acquisition module, used to acquire historical activity data of a customer, wherein the historical activity data includes a plurality of time series, each of which includes time series data of a financial activity of the customer within a target time period; A feature extraction module, used to decompose the multiple time series respectively, and extract the feature of each of the multiple time series, wherein the feature includes a trend item and at least one feature sequence; The feature extraction module is further used to construct a feature graph structure according to the features, wherein the feature graph structure includes a node matrix, an edge feature and an adjacency matrix, wherein the node in the t-th row and n-th column of the node matrix includes t historical values ​​in the n-th time series, the edge feature is determined according to the correlation between the feature sequences, and the adjacency matrix is ​​determined according to the correlation between the trend items of each of the time series; A prediction module is used to obtain a first customer churn prediction result according to a customer churn prediction model, wherein the first customer churn prediction result is used to determine whether the customer's label is a potential churn customer, wherein the customer churn prediction model includes an encoder, wherein the encoder includes a self-attention layer, and the self-attention layer is used to respectively aggregate feature information of a node matrix, edge features, and feature information of an adjacency matrix in the feature graph structure to obtain fused feature information.

8. A computer device, characterized in that: The computer equipment comprises: Memory for storing computer programs or computer instructions; A processor, configured to execute a computer program or computer instruction stored in the memory, so that the computer device executes the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and when the computer program is executed, it is used to implement the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that When the computer program product is run on a computer, the computer is caused to execute the method according to any one of claims 1 to 6.

Images