User loss prediction method and device, equipment, storage medium and program product

Abstract

The embodiment of the invention provides a user loss prediction method and device, equipment, a storage medium and a program product, and relates to the field of financial science and technology. The method comprises the following steps: acquiring target data of a first user, wherein the target data comprises basic data, financial transaction data within a preset duration, account change data and user complaint data; and preprocessing the target data to obtain feature data of the first user. The preprocessing at least comprises coding processing and normalization processing. And inputting the input data of the first user into a user churn prediction model constructed by a multi-scale convolutional network based on a fusion attention mechanism to obtain a churn prediction result of the first user. The user churn prediction model is used for extracting features of the first user from different scales, fusing the extracted features based on an attention mechanism and then performing churn prediction on the first user. According to the method, the accuracy of user loss prediction is improved.

Description

Technical Field

[0001] The present application relates to the field of financial technology, and in particular to a user churn prediction method, electronic device, storage medium and program product. Background Art

[0002] With the rapid development of the economy, user needs are becoming increasingly diversified and personalized, and the competition in the financial services market is becoming increasingly fierce. The problem of user churn in financial institutions is a severe challenge for financial institutions. Therefore, reducing user churn has become an important issue that the financial institutions industry is generally concerned about and urgently needs to solve.

[0003] The user churn prediction model can predict user churn in advance, allowing financial institutions to quickly adjust strategies based on the prediction results, quickly respond to user churn problems, and achieve timely perception and refined management of user resource changes. However, traditional user churn prediction models have limitations in data preprocessing, feature extraction, algorithm selection, etc., which leads to inaccurate prediction results. Summary of the invention

[0004] The present application provides a user churn prediction method, electronic device, storage medium and program product, which are used to solve the technical problem that the prediction results of traditional user churn prediction models are not accurate enough.

[0005] In a first aspect, the present application provides a user churn prediction method, the method comprising:

[0006] Acquire target data of the first user, the target data including: basic data, financial transaction data within a preset period of time, account change data, and user complaint data;

[0007] Preprocessing the target data to obtain feature data of the first user; the preprocessing at least includes: encoding processing and normalization processing;

[0008] The input data of the first user is input into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism to obtain a churn prediction result of the first user; the user churn prediction model is used to extract features of the first user from different scales and perform churn prediction on the first user after fusing the extracted features based on the attention mechanism.

[0009] In a second aspect, the present application provides a user churn prediction device, comprising:

[0010] An acquisition module, used to acquire target data of the first user, the target data including: basic data, financial transaction data within a preset time period, account change data, and user complaint data;

[0011] A processing module, used for preprocessing the target data to obtain the characteristic data of the first user; the preprocessing at least includes: encoding processing and normalization processing;

[0012] The prediction module is used to input the input data of the first user into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism to obtain a churn prediction result of the first user; the user churn prediction model is used to extract features of the first user from different scales and perform churn prediction on the first user after fusing the extracted features based on an attention mechanism.

[0013] In a third aspect, the present application provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor;

[0014] The memory stores computer-executable instructions;

[0015] The processor executes the computer-executable instructions stored in the memory to implement the method as described in any one of the first aspects.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement the method as described in any one of the first aspects.

[0017] In a fifth aspect, the present application provides a computer program product, including a computer program, which implements the method described in any one of the first aspects when executed by a processor.

[0018] The user churn prediction method, electronic device, storage medium and program product provided in the present application predict the user's churn behavior through a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism, which can effectively capture and learn the nonlinear features in the user's complex feature data relationship, thereby further extracting useful feature data and improving the accuracy of user churn prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0020] Figure 1 A flowchart of a user churn prediction method provided in an embodiment of the present application;

[0021] Figure 2 An architectural diagram of a user churn prediction model provided in an embodiment of the present application;

[0022] Figure 3A schematic diagram of a method for extracting user information features by a multi-scale convolution module provided in an embodiment of the present application through three paths and a feature fusion layer;

[0023] Figure 4 A schematic diagram of the structure of an attention mechanism module provided in an embodiment of the present application;

[0024] Figure 5 A flowchart of a method for implementing user churn prediction provided in an embodiment of the present application;

[0025] Figure 6 A schematic diagram of the structure of a user churn prediction device provided in an embodiment of the present application;

[0026] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0027] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0028] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0029] The following uses the user churn prediction method as an example to explain the terms involved in this application:

[0030] Data preprocessing: refers to the process of cleaning, converting and normalizing raw data to improve data quality and make it more suitable for subsequent model analysis and use.

[0031] Feature extraction: The process of identifying and extracting features and attributes from data that help describe the data is critical to reducing data dimensionality and improving model performance and interpretability.

[0032] Churn prediction: It is used to predict the likelihood that a user may stop using a product or service in the future.

[0033] Attention mechanism: A mechanism that simulates human attention, allowing the model to focus on the most important part when processing information, improving the model's ability to process complex data.

[0034] Convolutional network: extract local features of images or other data through convolutional layers.

[0035] Squeeze-and-Excitation (SE) mechanism: By recalibrating the response of channel features, the network's ability to represent useful information is enhanced, thereby improving model performance.

[0036] Global average pooling: Reduces feature dimensionality by calculating the average value over the entire feature map.

[0037] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0038] In addition, this application involves conducting big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.), and using artificial intelligence technology to make automated decisions, and making technical solutions that have a significant impact on personal rights and interests based on the results of automated decisions. The application provides users with corresponding operation entrances for them to choose to agree or reject the results of automated decisions; if the user chooses to reject, the expert decision-making process will be entered.

[0039] It should be noted that the user churn prediction method, device, equipment, storage medium and program product provided in the present application can be used in the field of financial technology, and can also be used in any field other than financial technology. The application field of the user churn prediction method, device, equipment, storage medium and program product in the present application is not limited.

[0040] With the rapid development of the economy, user needs are becoming increasingly diversified and personalized, and competition in the financial market is becoming increasingly fierce. The problem of user churn in financial institutions has become more and more serious. User churn will not only cause losses to the direct economic interests of financial institutions, but also affect the brand image, market share and long-term development strategy of financial institutions. Therefore, reducing the problem of user churn has become an important issue that the financial institution industry is generally concerned about and urgently needs to solve.

[0041] The user churn prediction model can predict the user churn behavior of financial institutions in advance, allowing financial institutions to quickly adjust their strategies based on the prediction results, quickly respond to user churn problems, and achieve timely perception and refined management of user resource changes.

[0042] Traditional user churn prediction models generally include three steps: data preprocessing, feature extraction, and churn prediction.

[0043] In the data preprocessing stage, the model generally uses a group algorithm to preprocess the data. This method has the disadvantages of slow convergence, complex parameter adjustment, and low solution efficiency.

[0044] In terms of feature extraction, a convolutional network with an attention mechanism is generally used to extract feature data. The Squeeze-and-Excitation (SE) mechanism in the attention mechanism is usually used to use global average pooling in feature extraction tasks to focus on more critical information for the current task, helping the model suppress irrelevant information and improve performance.

[0045] However, although the traditional SE mechanism can improve the accuracy of the model to a certain extent, it may cause the loss of some feature information due to its direct use of global average pooling, and cannot capture all complex detail information. Therefore, although the SE mechanism is widely used, the feature extraction method is not comprehensive enough, and it is difficult to extract more effective feature data. Therefore, it is difficult for traditional user churn prediction models to achieve efficient and accurate prediction results.

[0046] In view of this, the present application proposes a user churn prediction method, which predicts the user's churn behavior by using a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism. The method can effectively capture and learn the nonlinear features in the complex feature data relationship of the user, thereby further extracting useful feature data and improving the accuracy of user churn prediction.

[0047] The above-mentioned user churn prediction method can be applied to user churn prediction of financial institutions, etc.

[0048] The execution subject of the embodiments of the present application may be an electronic device, a prediction system or prediction platform, or a system or platform integrated with the prediction function. Taking the user churn prediction scenario of a financial institution as an example, the execution subject in this scenario may be, for example, a banking system. For ease of description, the following embodiments are all described using a banking system as an example.

[0049] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0050] Figure 1 A flowchart of a user churn prediction method provided in an embodiment of the present application is shown below. Figure 1 As shown, the method comprises the following steps:

[0051] S101. Obtain target data of a first user, where the target data includes: basic data, financial transaction data within a preset period of time, account change data, and user complaint data.

[0052] Among them, basic data may include, for example, information such as the age, gender, region, and occupation of the first user; financial transaction data within a preset time period may include, for example, the number of bank card transactions, types of financial products purchased, credit score, star rating, and other information within a preset time period for the first user; account change data may include, for example, whether the first user's mobile financial institution account has been changed within a preset time period; user complaint data may include, for example, whether the first user has a complaint within a preset time period. Among them, the preset time period can be set according to the specific situation, for example, it can be within three months, within six months, etc. The embodiment of the present application does not limit the preset time period.

[0053] The first user may be any user in the banking system, or may be a suspected lost user screened from the users of the banking system. For example, the banking system may extract target data of multiple users from a database, and determine whether there is a suspected lost first user from the multiple users based on a preset financial transaction screening strategy, an account change screening strategy, and a complaint screening strategy.

[0054] Among them, the screening strategy based on the preset financial transaction, the screening strategy based on the account change, and the complaint screening strategy can be, for example, a strategy for screening financial transaction data, account change data, and user complaint data according to predefined rules. The predefined rules can be set according to the specific circumstances of the financial institution, and the embodiments of the present application are not limited to this.

[0055] For example, when a user's transaction frequency, large withdrawals, or frequent small transactions decrease, the user is determined to be the first user suspected of churn.

[0056] For another example, when a user's account balance decreases significantly, or the account activity suddenly stops or changes abnormally, the user is determined to be the first user who is suspected of being lost.

[0057] For another example, when a user has a complaint record, especially frequent or serious complaints, the user is determined to be the first user suspected of churn.

[0058] S102: Preprocess the target data to obtain feature data of the first user. The preprocessing at least includes: encoding processing and normalization processing.

[0059] Among them, encoding is the process of converting the qualitative or quantitative information in the original data of the target data into a specific format code. Normalization is the process of scaling the data to a uniform range to eliminate the impact of different dimensions and numerical ranges on model training.

[0060] Exemplarily, the target data is converted into numerical data through encoding processing, and then the data is normalized to obtain the feature data of the first user. The encoding processing can be, for example, any processing that can convert the target data into numerical data, such as one-hot encoding. The normalization processing can be implemented, for example, by the following formula (1).

[0061] (1)

[0062] in, Represents the target data after transformation. represents the target data mean, Represents the standard deviation of the target data.

[0063] Optionally, the above preprocessing may further include, for example, data cleaning, such as removing irrelevant data, filling missing values, smoothing noise data, etc. This operation may be, for example, an operation before encoding and normalization.

[0064] For example, the banking system can complete the missing data in the target data, and correct the erroneous data in the target data to obtain the corrected target data. Exemplarily, the missing data can be filled with preset values, or with the data of users with similar financial transaction profiles to the first user. Since users with similar financial transaction profiles usually have similar financial transaction behaviors, the relevant data of other users can be used to characterize the first user.

[0065] Among them, the financial transaction portrait is a detailed description constructed based on the data generated by the user in financial activities. The data may include one or more information such as the user's natural attributes, industry attributes, financial attributes, asset attributes, social attributes and interest attributes.

[0066] The above financial transaction portrait can be constructed based on the user's financial data within a preset period of time, or it can be a portrait dynamically updated based on the user's financial transaction data. For how to construct a financial transaction portrait, please refer to the construction method of user portrait in the prior art.

[0067] For example, similarity calculation can be performed between the first user's financial transaction profile and other users' financial transaction profiles to determine whether there is a second user whose similarity with the first user's financial transaction profile is greater than a preset threshold. The second user can be the user with the greatest similarity, or multiple second users whose similarities are greater than the threshold. In this implementation, the data used for data filling can be the average of the data of the multiple second users, or a randomly selected second user.

[0068] Accordingly, the determination of erroneous data in the target data of the first user can also be determined and corrected using the data of the second user. For example, when the deviation between the target data of the first user and the target data of the second user is greater than a preset threshold, it is considered that there is an error in the target data of the first user, and thus it can be corrected based on the target data of the second user.

[0069] Optionally, in this implementation, it can be determined whether there are errors in the target data of the first user based on multiple second users, so that when it is determined that there are errors in the target data of the first user based on the data of multiple second users, the erroneous data of the first user is corrected.

[0070] That is, in this implementation, the corrected target data of the first user can be one-hot encoded to obtain the encoded data of the target data. Then, the encoded data of the target data of the first user is normalized to obtain the feature data. In this way, the accuracy of the target data of the first user can be ensured.

[0071] S103: Input the input data of the first user into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism to obtain a churn prediction result for the first user. The user churn prediction model is used to extract features of the first user from different scales and perform churn prediction on the first user after fusing the extracted features based on an attention mechanism.

[0072] The above-mentioned user churn prediction model is constructed using a multi-scale convolutional network based on a fused attention mechanism, which can extract the features of the first user from different scales, so as to effectively capture and learn the nonlinear features in the user's complex feature data relationship, and fuse the extracted features based on the attention mechanism to further extract useful feature data. In this way, churn prediction based on the useful feature data can improve the accuracy of user churn prediction.

[0073] Optionally, the prediction result may be, for example, a probability value.

[0074] Optionally, if the churn prediction result of the first user indicates that the churn probability of the first user is greater than a preset probability threshold, the banking system may further analyze and output the reasons for the churn of the first user and / or corresponding improvement measures based on the target data of the first user, so as to provide accurate data support and guidance for subsequent improvements of the banking system.

[0075] The preset probability threshold may be, for example, a pre-set critical value, which is used to determine whether the user churn probability predicted by the model is high enough to consider that the user has a high churn risk.

[0076] For example, the target data of the first user may be analyzed by a pre-trained model to obtain the reason for the loss of the first user. The above model may be pre-trained using the target data of historical lost users, for example.

[0077] Optionally, the banking system may be pre-set with an expert database, which stores a mapping relationship between the churn reasons and corresponding improvement measures based on historical churn users. Therefore, the corresponding improvement measures can be obtained from the expert database based on the churn reasons of the first user obtained above. In some embodiments, the model for analyzing the churn reasons can also be pre-trained to support the output of improvement measures, so that the reasons and improvement measures can be directly derived based on the target data of the first user.

[0078] In some embodiments, the above-mentioned expert database can also be represented by a knowledge graph, etc., which will not be elaborated herein.

[0079] The method provided by the present application pre-processes the target data of the user through encoding and normalization to obtain the characteristic data of the user. The characteristic data of the user is input into the user churn prediction model constructed by the multi-scale convolutional network based on the fusion attention mechanism, the user's features are extracted from different scales and the extracted features are fused based on the improved attention mechanism, and then the user's churn behavior is predicted. The method uses the multi-scale convolutional network with the fusion attention mechanism to effectively capture and learn the nonlinear features in the complex characteristic data relationship of the user, further extract useful characteristic data, and achieve higher accuracy in user churn prediction.

[0080] The user churn prediction model can be obtained by training any multi-scale convolutional network based on a fusion attention mechanism using a sample data set. The sample data set mentioned here may include multiple sample data, each sample data includes feature data constructed based on the target data of the sample user, and the corresponding churn result.

[0081] Exemplarily, the sample data set can be divided into a training set, a test set, and a validation set. The training set can be used, for example, to train a user churn prediction model so that the user churn prediction model can learn the patterns and rules in the data. The validation set can be used, for example, to adjust and select a user churn prediction model to evaluate the performance of different user churn prediction models or different parameter settings to select the best model configuration. Prevent the user churn prediction model from overfitting on the training set. The test set can be used, for example, to finally evaluate the performance of the model. Provide a true estimate of the model's generalization ability.

[0082] To facilitate understanding, the structure of the user churn prediction model is illustrated below.

[0083] Figure 2 This is an architecture diagram of a user churn prediction model provided in an embodiment of the present application. Figure 2 As shown, the user churn prediction model may include, for example: a multi-scale convolution module, an attention mechanism module, and a prediction module.

[0084] (1) A multi-scale convolution module, used to extract first user features related to first user churn prediction from different scales.

[0085] The multi-scale convolution module captures features of different scales by setting up multiple parallel paths. Each path contains a series of convolution kernels of different sizes. The convolution layer formed by these convolution kernels is responsible for extracting features of a specific scale. Within each path, point-by-point convolution is used to adjust the relationship between channels of different convolution layers, thereby integrating features from different convolution layers and achieving efficient extraction and fusion of multi-scale and multi-dimensional features in the input data.

[0086] The scales supported by the multi-scale convolution module can be set according to actual needs. Exemplarily, the multi-scale convolution module may include: a first convolution scale path, a second convolution scale path, a third convolution scale path, and a feature fusion layer. The first convolution scale is larger than the second convolution scale, and the second convolution scale is larger than the third convolution scale.

[0087] Figure 3 A schematic diagram of a method for extracting user information features by a multi-scale convolution module provided in an embodiment of the present application through three paths and a feature fusion layer. Kernel is a convolution kernel, Stride refers to the step length of the convolution kernel when it slides on the input feature map, Input represents the input data of the first user, and X represents the output of the multi-scale convolution module.

[0088] Exemplarily, the first convolutional scale path may be used to capture the overall characteristics of the first user from the input data of the first user. The operation performed by the first convolutional scale path may be represented by the following formula (2), for example.

[0089] (2)

[0090] in, represents the first convolutional scale path The output features of the convolutional layer, represents a sample in the first convolutional scale path, It is The convolution kernel of the convolution layer is: represents the number of convolutional layers, Indicates the size of the convolution kernel in the convolution layer, for example, it can be 7. Indicates The convolution output of the convolutional layer, It is The bias term of the layer. Represents a nonlinear function, such as an activation function, applied to the result of the convolution operation to introduce nonlinearity so that the network can learn more complex features. It means that it will be performed in each convolution layer in the first convolution scale path. is the total number of convolutional layers in the network.

[0091] For example, a smaller convolution kernel and an activation function can be used to calculate the next convolution layer, where the activation function can be, for example, a ReLU function, and the convolution kernel size can be, for example, 1*1. This method can further deepen the network depth and enhance the nonlinearity and robustness of the model.

[0092] Exemplarily, the second convolutional scale path may be used to extract the input data features of the first user from the input data of the first user. The operation performed by the second convolutional scale path may be represented by the following formula (3), for example.

[0093] (3)

[0094] in, represents the first convolutional scale path in the second convolutional scale path. The output features of the convolutional layer, represents a sample in the second convolutional scale path, The effect of is as shown in the above formula (2), which will not be described in detail. For example, The value of may be, for example, 5. As shown above, for example, a smaller convolution kernel and an activation function may be used to calculate the next convolution layer, wherein the activation function may be, for example, a ReLU function, and the convolution kernel size may be, for example, 1*1.

[0095] Exemplarily, the third convolutional scale path may be used to extract detailed features of the first user from the input data of the first user. The operation performed by the third convolutional scale path may be represented by the following formula (4), for example.

[0096] (4)

[0097] in, represents the third convolutional scale path The output features of the convolutional layer, represents a sample in the second convolutional scale path, The effect of is as shown in the above formula (2), which will not be described in detail. For example, The value of can be 3, for example. In the third convolution scale path, a smaller convolution kernel can be used to obtain more detailed features. Smaller convolutions can often supplement some detailed features to enhance learning ability, avoid the model ignoring the importance of some detailed features during the learning process, and at the same time, the nonlinear activation layer can be used to improve the model's ability to discriminate input data.

[0098] As shown above, for example, a smaller convolution kernel and an activation function may be used to calculate the next convolution layer, wherein the activation function may be, for example, a linear rectification (ReLU, Rectified Linear Unit) function, and the convolution kernel size may be, for example, 1*1.

[0099] Exemplarily, the feature fusion layer is used to fuse the overall features, input data features, and detail features to obtain fused features.

[0100] For example, feature fusion can be achieved by adding the feature calculation results output in the above three convolutional scale paths element by element. The feature fusion layer can be implemented by the following formula (5), for example.

[0101] (5)

[0102] in, represents the fused feature output, and They represent the output features of the last convolutional layer in the first convolutional scale path, the second convolutional scale path, and the third convolutional scale path, respectively.

[0103] (2) An attention mechanism module, which is used to select key features from the first user features extracted by the multi-scale convolution module by assigning weights.

[0104] Figure 4 A schematic diagram of the structure of an attention mechanism module provided in an embodiment of the present application. Figure 4As shown, the attention mechanism module may include, for example, a first attention mechanism submodule, a second attention mechanism submodule, and a third attention mechanism submodule. Among them, Global avg pool represents a global average pooling operation, Conv Kernel represents the convolution kernel size, BN represents a batch normalization operation, X represents the output content of the multi-scale convolution module, Y represents the output of the attention mechanism module, B1 represents the first attention mechanism submodule, B2 represents the second attention mechanism submodule, and B3 represents the third attention mechanism submodule.

[0105] Exemplarily, the first attention mechanism submodule can be used to obtain the global features of the first user through global average pooling based on the fusion features. The global average pooling (GAP) to obtain the global features of the first user can be expressed by the following formula (6).

[0106] (6)

[0107] in, After feature fusion, The output value of each channel after GAP, Represents each channel feature dimension. Indicates the spatial position of the fused features The characteristic element value on . It means to sum the feature elements of all spatial positions after feature fusion.

[0108] For example, after performing GAP on the fused features, the global features of the first user obtained by GAP can be processed by a convolution layer with a smaller convolution kernel, so as to further capture the correlation between channels and effectively reduce the number of parameters and computational complexity of the model. The convolution kernel size can be 1*1, for example. Then, batch normalization and activation function are used to further enhance the nonlinear expression ability of the model.

[0109] For example, the global feature of the first user after this part of the processing can be expressed by the following formula (7):

[0110] (7)

[0111] In the formula, Represents the output after being processed by this method. is the loss function, Represents the processed convolution kernel function. Represents a batch normalization operation, which is used to normalize the output of the previous layer to speed up the training process and improve the stability of the model.

[0112] The first attention mechanism sub-module can better learn the overall characteristics of user features and improve the global inter-channel feature representation.

[0113] Exemplarily, the second attention mechanism submodule can be used to learn the detail features of the first user based on the fused features, and fuse the global features with the detail features to obtain intermediate features.

[0114] For example, in the second attention mechanism submodule, the fused features are subjected to batch normalization processing and activation function combination to further enhance the nonlinear mapping between the fused features and the key features, and the same data distribution as the fused features can be maintained to supplement the key detail features ignored by the first attention mechanism submodule after GAP processing. The operation of the second attention mechanism submodule can be expressed, for example, by the following formula (8):

[0115] (8)

[0116] in, is the fused feature, and The above formula (7) has the same function, so it will not be described in detail.

[0117] The second attention mechanism sub-module can enhance the abstract expression of user information.

[0118] Exemplarily, the third attention mechanism submodule can be used to obtain feature allocation weights based on intermediate features, and obtain key features based on the weights and fused features.

[0119] For example, the third attention mechanism submodule can combine the outputs of the first attention mechanism submodule and the second attention mechanism submodule and Linear addition fusion is performed, and then a certain amount of convolution layer mapping is performed to highlight more valuable features, wherein the convolution layer can be, for example, a convolution layer with a convolution kernel of 3*1.

[0120] The operation performed by the third attention mechanism submodule can be expressed, for example, by the following formula (9):

[0121] (9)

[0122] in, is an activation function, which is used to convert the output into a probability value, indicating the possibility of a certain category occurring. The activation function can be, for example, a logistic curve ( )function. represents the processed convolution kernel function, The function is the same as above and will not be described in detail.

[0123] In a specific embodiment, when the activation function is Function, The output can be, for example, The parameter between . This parameter can also be called the weight of each feature.

[0124] The third attention mechanism sub-module can further integrate the information of the previous module and learn the key information of linearly added features through convolution operation to enhance feature representation.

[0125] The third attention mechanism submodule processing may also include the relevant parameters obtained by formula (9) and Perform element-by-element multiplication, and obtain key features based on weights and fusion features. This method can be implemented, for example, by the following formula (10):

[0126] (10)

[0127] (3) A prediction module, used to obtain a churn prediction result of the first user based on the key features.

[0128] For example, the key features of the first user are input into the prediction module, and the prediction module outputs the churn probability of the first user. The prediction module may include a classification layer, and outputs the prediction result through the classification layer.

[0129] The method provided in this application inputs the user's feature data into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism, extracts the user's features from different scales, and fuses the extracted features based on the improved attention mechanism, thereby predicting the user's churn behavior. This method uses a multi-scale convolutional network with a fusion attention mechanism to effectively capture and learn the nonlinear features in the complex feature data relationship of the user, and further extracts useful key feature data, thereby improving the accuracy of the model prediction. It helps financial institutions to quickly respond to user churn problems and realize timely perception and refined management of user resource changes.

[0130] The following is a specific example to illustrate how to implement user churn prediction. Figure 5 A flowchart of a method for realizing user churn prediction provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, the method may include the following steps:

[0131] S501, extracting target data of multiple users from a database, and determining whether there is a first user suspected of being lost from the multiple users based on a preset financial transaction screening strategy, an account change screening strategy, and a complaint screening strategy. Obtaining target data of the first user.

[0132] The target data includes: basic data, financial transaction data within a preset period of time, account change data, and user complaint data;

[0133] Exemplarily, when the target data is missing, the missing data in the target data is supplemented, and the erroneous data in the target data is corrected to obtain the corrected target data.

[0134] S502: Perform one-hot encoding and normalization processing on the target data to obtain feature data of the first user.

[0135] S503: Input the feature data of the first user into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism to obtain a churn prediction result for the first user.

[0136] S504: The multi-scale convolution module extracts first user features related to the first user churn prediction from different scales by setting multiple parallel paths, and outputs fused features.

[0137] S505: Input the feature data and fusion features of the first user into the attention mechanism module, which selects key features from the first user features extracted by the multi-scale convolution module by assigning weights.

[0138] S506: Input the key features obtained by the attention mechanism module into the prediction module, and the prediction module obtains the churn prediction result of the first user based on the key features.

[0139] In an exemplary embodiment, the key features of the first user are input into the prediction module, and the prediction module outputs the churn probability of the first user. If the churn prediction result of the first user indicates that the churn probability of the first user is greater than a preset probability threshold, then based on the target data of the first user, the reasons leading to the churn of the first user are analyzed, as well as the corresponding improvement measures.

[0140] Exemplarily, the key features of the first user are input into the prediction module, and the prediction module outputs the prediction results, the reasons for the loss, and the corresponding improvement measures. The prediction module can be, for example, any model such as a neural network that can output the prediction results, the reasons for the loss, and the corresponding improvement measures.

[0141] The present application obtains corrected target data by completing missing data in the target data and correcting erroneous data in the target data. The corrected target data is uniquely encoded and normalized to obtain feature data. This method can ensure the integrity of the target data by completing and correcting the target data. Uniquely encoding and normalizing the corrected target data can ensure that different feature data are on the same scale, reduce the impact of outliers on the model, and make the model have better generalization ability and numerical stability.

[0142] By inputting the user's feature data into the user churn prediction model built on a multi-scale convolutional network based on a fusion attention mechanism, the user's features are extracted from different scales and the extracted features are fused based on the improved attention mechanism, thereby predicting the user's churn behavior. And output the prediction results, churn reasons, and corresponding improvement measures. This method uses a multi-scale convolutional network with a fusion attention mechanism to effectively capture and learn the nonlinear features in the complex feature data relationship of users, and further extract useful key feature data, thereby improving the accuracy of model prediction. It helps financial institutions to quickly respond to user churn problems and realize timely perception and refined management of user resource changes.

[0143] The above is the method of the embodiment of the present application. The device provided by the embodiment of the present application is described below.

[0144] Figure 6 A schematic diagram of the structure of a user churn prediction device provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, the user churn prediction device includes: an acquisition module 601 , a processing module 602 and a prediction module 603 .

[0145] The acquisition module 601 is used to acquire the target data of the first user, the target data including: basic data, financial transaction data within a preset time period, account change data, and user complaint data;

[0146] The processing module 602 is used to pre-process the target data to obtain the characteristic data of the first user; the pre-processing at least includes: encoding processing and normalization processing;

[0147] The prediction module 603 is used to input the input data of the first user into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism to obtain a churn prediction result of the first user; the user churn prediction model is used to extract features of the first user from different scales and perform churn prediction on the first user after fusing the extracted features based on the attention mechanism.

[0148] Optionally, the user churn prediction model includes: a multi-scale convolution module, an attention mechanism module, and a prediction module;

[0149] a multi-scale convolution module, for extracting first user features related to first user churn prediction from different scales;

[0150] an attention mechanism module, configured to select key features from the first user features extracted by the multi-scale convolution module by assigning weights;

[0151] A prediction module is used to obtain a churn prediction result of the first user based on key features.

[0152] Exemplarily, the multi-scale convolution module includes:

[0153] a first convolutional scaling path for capturing overall characteristics of the first user from input data of the first user;

[0154] a second convolutional scale path, used to extract input data features of the first user from the input data of the first user;

[0155] a third convolutional scale path, used to extract detailed features of the first user from the input data of the first user;

[0156] A feature fusion layer is used to fuse the overall feature, input data feature, and detail feature to obtain a fused feature;

[0157] The first convolution scale is larger than the second convolution scale, and the second convolution scale is larger than the third convolution scale.

[0158] Exemplarily, the attention mechanism module includes:

[0159] A first attention mechanism submodule, configured to obtain the global features of the first user through global average pooling based on the fused features;

[0160] A second attention mechanism submodule is used to learn the detail features of the first user based on the fused features, and fuse the global features with the detail features to obtain intermediate features;

[0161] The third attention mechanism submodule is used to obtain the feature allocation weight based on the intermediate feature, and obtain the key feature based on the weight and the fused feature.

[0162] Optionally, the processing module 602 is specifically used to complete missing data in the target data, and to correct erroneous data in the target data to obtain corrected target data; perform one-hot encoding on the corrected target data to obtain encoded data of the target data; and perform normalization processing on the encoded data of the target data to obtain feature data.

[0163] For example, processing module 602 is specifically used to obtain a financial transaction profile of a first user, find a second user whose similarity with the financial transaction profile of the first user is greater than a preset threshold, and correct erroneous data in the target data of the first user based on the target data of the second user.

[0164] Optionally, the acquisition module 601 is specifically used to extract target data of multiple users from the database; based on a preset financial transaction screening strategy, an account change screening strategy, and a complaint screening strategy, determine whether there is a suspected lost first user among the multiple users.

[0165] Optionally, the prediction module 603 is specifically used to analyze the reasons for the churn of the first user and the corresponding improvement measures based on the target data of the first user if the churn prediction result of the first user indicates that the churn probability of the first user is greater than a preset probability threshold; and output the prediction result, the reasons for the churn, and the corresponding improvement measures.

[0166] The user churn prediction device provided in the embodiment of the present application can be used to execute the aforementioned user churn prediction method. Its implementation principle, process and beneficial effects can be found in the aforementioned embodiment and will not be repeated here.

[0167] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 7 As shown, the electronic device 700 may include: a memory 701 and a processor 702. Optionally, the electronic device may further include a transceiver 703, wherein the memory 701 and the processor 702 communicate; illustratively, the memory 701, the processor 702 and the transceiver 703 may communicate via a communication bus 704, the memory 701 is used to store a computer program, and the processor 702 executes the computer program to implement the method of the above embodiment.

[0168] Optionally, the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps in the method embodiments disclosed in the present application may be directly implemented as being executed by a hardware processor, or may be implemented by a combination of hardware and software modules in the processor.

[0169] An embodiment of the present application further provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, a method in any of the above method embodiments is implemented.

[0170] An embodiment of the present application also provides a computer program product, including a computer program, which implements the method in any of the above method embodiments when the computer program is executed by a processor.

[0171] All or part of the steps of the above-mentioned method embodiments can be completed by hardware related to program instructions. The above-mentioned program can be stored in a readable memory. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the above-mentioned memory (storage medium) includes: read-only memory (ROM), RAM, flash memory, hard disk, solid state drive, magnetic tape, floppy disk, optical disc and any combination thereof.

[0172] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processing unit of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0173] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0174] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1The steps of the functions specified in a box or multiple boxes. It should be noted that for the aforementioned method embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited to the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0175] It should be further noted that, although the various steps in the flowchart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowchart may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0176] It should be understood that the above-mentioned device embodiments are only illustrative, and the device of the present application can also be implemented in other ways. For example, the division of units / modules in the above-mentioned embodiments is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units, modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed.

[0177] In addition, unless otherwise specified, each functional unit / module in each embodiment of the present application may be integrated into one unit / module, each unit / module may exist physically separately, or two or more units / modules may be integrated together. The above-mentioned integrated unit / module may be implemented in the form of hardware or in the form of a software program module.

[0178] If the integrated unit / module is implemented in the form of hardware, the hardware may be a digital circuit, an analog circuit, etc. The physical implementation of the hardware structure includes but is not limited to transistors, memristors, etc. Unless otherwise specified, the processor may be any appropriate hardware processor, such as CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit may be any appropriate magnetic storage medium or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0179] If the integrated unit / module is implemented in the form of a software program module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a memory, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0180] In the above embodiments, the description of each embodiment has its own emphasis. For the part not described in detail in a certain embodiment, please refer to the relevant description of other embodiments. The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0181] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0182] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for predicting user churn, characterized in that: The method comprises: Acquire target data of the first user, the target data including: basic data, financial transaction data within a preset period of time, account change data, and user complaint data; Preprocessing the target data to obtain feature data of the first user; the preprocessing at least includes: encoding processing and normalization processing; The input data of the first user is input into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism to obtain a churn prediction result of the first user; the user churn prediction model is used to extract features of the first user from different scales and perform churn prediction on the first user after fusing the extracted features based on the attention mechanism.

2. The method according to claim 1, characterized in that The user churn prediction model includes: a multi-scale convolution module, an attention mechanism module, and a prediction module; a multi-scale convolution module, for extracting first user features related to first user churn prediction from different scales; an attention mechanism module, configured to select key features from the first user features extracted by the multi-scale convolution module by assigning weights; A prediction module is used to obtain a churn prediction result of the first user based on key features.

3. The method according to claim 2, characterized in that The multi-scale convolution module comprises: a first convolutional scaling path for capturing overall characteristics of the first user from input data of the first user; a second convolutional scale path, used to extract input data features of the first user from the input data of the first user; a third convolutional scale path, used to extract detailed features of the first user from the input data of the first user; A feature fusion layer is used to fuse the overall feature, input data feature, and detail feature to obtain a fused feature; The first convolution scale is larger than the second convolution scale, and the second convolution scale is larger than the third convolution scale.

4. The method according to claim 3, characterized in that The attention mechanism module includes: A first attention mechanism submodule, configured to obtain the global features of the first user through global average pooling based on the fused features; A second attention mechanism submodule is used to learn the detail features of the first user based on the fused features, and fuse the global features with the detail features to obtain intermediate features; The third attention mechanism submodule is used to obtain the feature allocation weight based on the intermediate feature, and obtain the key feature based on the weight and the fused feature.

5. The method according to any one of claims 1 to 4, characterized in that: The preprocessing of the target data to obtain the characteristic data of the first user includes: Completing missing data in the target data, and correcting erroneous data in the target data to obtain corrected target data; Performing one-hot encoding on the corrected target data to obtain encoded data of the target data; The encoded data of the target data is normalized to obtain the feature data.

6. The method according to claim 5, characterized in that The correcting of the erroneous data in the target data includes: Obtaining a financial transaction profile of the first user; Find a second user whose financial transaction profile is more similar to the first user than a preset threshold; Based on the target data of the second user, erroneous data in the target data of the first user is corrected.

7. The method according to any one of claims 1 to 4, characterized in that: The acquiring target data of the first user includes: Extract target data of multiple users from a database; Based on a preset financial transaction screening strategy, an account change screening strategy, and a complaint screening strategy, it is determined whether there is a first user suspected of being lost from the multiple users.

8. The method according to any one of claims 1 to 4, characterized in that: After obtaining the churn prediction result of the first user, the method further includes: If the churn prediction result of the first user indicates that the churn probability of the first user is greater than a preset probability threshold, analyzing the churn reasons of the first user and corresponding improvement measures based on the target data of the first user; Output the prediction results, reasons for loss, and corresponding improvement measures.

9. A user churn prediction device, characterized in that: The device comprises: An acquisition module, used to acquire target data of the first user, the target data including: basic data, financial transaction data within a preset time period, account change data, and user complaint data; A processing module, used for preprocessing the target data to obtain the characteristic data of the first user; the preprocessing at least includes: encoding processing and normalization processing; The prediction module is used to input the input data of the first user into a user churn prediction model constructed based on a multi-scale convolutional network with a fusion attention mechanism to obtain a churn prediction result of the first user; the user churn prediction model is used to extract features of the first user from different scales and perform churn prediction on the first user after fusing the extracted features based on an attention mechanism.

10. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 8 when executed by a processor.

12. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 8 when being executed by a processor.

Images