User churn prediction method, device, electronic device and readable storage medium

Abstract

The embodiment of the present application discloses a user churn prediction method, device, electronic device and readable storage medium, the method comprising: obtaining user data, the user data comprising first time series feature matrix data and first non-time series feature matrix data of user behavior; inputting the first time series feature matrix data and the first non-time series feature matrix data into a preset hybrid multimodal churn prediction model of latent features, and calculating a first probability of user churn through the hybrid multimodal churn prediction model of latent features; wherein the hybrid multimodal churn prediction model of latent features is obtained by training based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; and determining the churn user data within the target time period based on the first probability of user churn. According to the embodiment of the present application, the efficiency of user churn prediction can be improved.

Description

Technical Field

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

[0002] At present, with the rapid development of computer technology, there are more and more application platforms, and people are paying more and more attention to the retention of platform users. For example, film and television application platforms have launched membership services, and members enjoy more extensive rights. Due to various reasons, some members will churn every month. In order to achieve user retention, it is necessary to analyze the user churn situation, and accurately predicting the churn behavior has important commercial value.

[0003] However, the accuracy of current user churn prediction methods is not high and cannot meet the growing prediction needs. Summary of the invention

[0004] The embodiments of the present application provide a user churn prediction method, apparatus, device, and readable storage medium, which can solve the problem of low accuracy of current user churn prediction methods.

[0005] In a first aspect, an embodiment of the present application provides a method for predicting user churn, the method comprising:

[0006] Acquire user data, where the user data includes first time series feature matrix data and first non-time series feature matrix data of user behavior;

[0007] Input the first time series feature matrix data and the first non-time series feature matrix data into a preset hybrid multimodal churn prediction model of latent features, and calculate the first probability of user churn through the hybrid multimodal churn prediction model of latent features; wherein the hybrid multimodal churn prediction model of latent features is obtained by training based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights, and preset latent feature weights; the latent feature matrix data is latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data;

[0008] The churned user data within the target time period is determined based on the first probability of user churn.

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

[0010] An acquisition module, used to acquire user data, the user data including first time series feature matrix data and first non-time series feature matrix data of user behavior;

[0011] A processing module, used for inputting the first time series feature matrix data and the first non-time series feature matrix data into a preset hybrid multimodal churn prediction model of latent features, and calculating a first probability of user churn through the hybrid multimodal churn prediction model of latent features; wherein the hybrid multimodal churn prediction model of latent features is obtained by training based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; and the latent feature matrix data is latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data;

[0012] The determination module is used to determine the lost user data within the target time period based on the first probability of user loss.

[0013] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the method in the first aspect or any possible implementation of the first aspect.

[0014] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, the method in the first aspect or any possible implementation of the first aspect is implemented.

[0015] In the embodiment of the present application, by acquiring user data, the user data includes the first time series feature matrix data and the first non-time series feature matrix data of user behavior; the first time series feature matrix data and the first non-time series feature matrix data are input into the preset hybrid multimodal churn prediction model of latent features, and the first probability of user churn is calculated by the hybrid multimodal churn prediction model of latent features; wherein, since the hybrid multimodal churn prediction model of latent features is obtained by training based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; the latent feature matrix data is the latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data; here, accurate discovery of latent features is achieved, and finally the lost user data within the target time period is determined based on the first probability of user churn. Thus, the lost user data can be accurately identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solution of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0017] Figure 1is a flow chart of a user churn prediction method provided by an embodiment of the present application;

[0018] Figure 2 It is a flow chart of a training method of a hybrid multimodal churn prediction model provided in an embodiment of the present application;

[0019] Figure 3 It is a formula transformation schematic diagram provided in an embodiment of the present application;

[0020] Figure 4 It is a schematic diagram of the structure of a user churn prediction device provided in an embodiment of the present application;

[0021] Figure 5 It is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only configured to explain the present application and are not configured to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by illustrating the examples of the present application.

[0023] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0024] First, the technical terms involved in the embodiments of the present application are introduced.

[0025] Domain B is the data domain of the business support system. Domain B contains user data and business data, such as user consumption habits, terminal information, ARPU grouping, business content, and business target groups.

[0026] The O domain is the data domain of the operation support system. The O domain contains network data, such as signaling, alarms, faults, and network resources.

[0027] Data analysis and mining from the above data domains are usually used to implement big data applications in the telecommunications industry, such as geographic precision marketing and cost actuarial calculations.

[0028] The xgboost algorithm (eXtreme Gradient Boosting) is based on the gradient boosting iterative decision tree (Gradient Boosting Decision Tree, GBDT). Xgboost is an efficient implementation of the GB algorithm. In addition to being a linear classifier (gblinear), the base learner in xgboost belongs to a tree algorithm and is suitable for classification and regression.

[0029] The user churn prediction method provided in the embodiment of the present application can be applied to at least the following application scenarios, which are described below.

[0030] At present, massive user behavior data is usually extracted regularly, and then imported into a pre-trained prediction model. Based on the prediction model, the regularly extracted user behavior data is predicted and analyzed to obtain user churn prediction results.

[0031] Specifically, we can manually analyze the user's labeled information, construct user features based on the labeled information, and then put the feature engineering data set into the xgboost model training to obtain a prediction model based on the labeled information features and prediction results. However, the training and prediction of this method rely on manual feature analysis, and it is not easy to find the hidden connections between feature values.

[0032] Therefore, on the one hand, it will lead to inaccurate feature analysis, that is, traditional churn prediction is based on manual analysis of user features, and feature analysis is based on individual information. It is impossible to accurately analyze the characteristics of a group of multiple individuals. For example, the member churn analysis of a certain business cannot obtain the overall characteristics of the family, and cannot analyze the important implicit characteristics of the original data. On the other hand, it is impossible to predict the situation of missing data. That is, traditional churn prediction is based on the analysis of user attribute information, and it is impossible to accurately predict users with missing attribute information, such as missing data on whether they have subscribed to membership and missing data on historical viewing behavior.

[0033] Based on the above application scenarios, the user churn prediction method provided in the embodiments of the present application is described in detail below.

[0034] Figure 1 A flowchart of a user churn prediction method provided in an embodiment of the present application.

[0035] like Figure 1 As shown, the user churn prediction method may include steps 110 to 130, and the method is applied to a user churn prediction device, as shown below:

[0036] Step 110: Acquire user data, where the user data includes first time series feature matrix data and first non-time series feature matrix data of user behavior.

[0037] Step 120, input the first time series feature matrix data and the first non-time series feature matrix data into a preset latent feature hybrid multimodal churn prediction model, and calculate the first probability of user churn through the latent feature hybrid multimodal churn prediction model; wherein the latent feature hybrid multimodal churn prediction model is trained based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; the latent feature matrix data is latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data.

[0038] Step 130: Determine the churned user data within the target time period based on the first probability of user churn.

[0039] In the user churn prediction method provided by the present application, by acquiring user data, the user data includes the first time series feature matrix data and the first non-time series feature matrix data of user behavior; the first time series feature matrix data and the first non-time series feature matrix data are input into the preset hybrid multimodal churn prediction model of latent features, and the first probability of user churn is calculated by the hybrid multimodal churn prediction model of latent features; wherein, since the hybrid multimodal churn prediction model of latent features is obtained by training based on the latent feature matrix data of historical user data, the preset feature weights, the preset time factor weights and the preset latent feature weights; the latent feature matrix data is the latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data; here, accurate discovery of latent features is achieved, and finally the churn user data within the target time period is determined based on the first probability of user churn. Thus, the churn user data can be accurately identified.

[0040] The contents of step 110 to step 130 are described below respectively:

[0041] First, step 110 is involved.

[0042] Step 110, obtaining user data, the user data including first time series feature matrix data and first non-time series feature matrix data of user behavior, including:

[0043] Get the original user data;

[0044] The original user data is extracted by word vectors, data that meets preset conditions is eliminated, and churn labels are extracted to obtain user data.

[0045] The raw data is preprocessed, including the processing of time series data and non-time series data. This may include user playback behavior data; non-time series data may include BO domain data and business platform data. Among them, user data includes user playback behavior data, BO domain data and business platform data; the above-mentioned steps of extracting word vectors from the raw user data, eliminating data that meets the preset conditions, and extracting churn labels to obtain user data include:

[0046] Extract the word vector of the user's playback behavior data to obtain a first time series feature matrix;

[0047] Eliminate the BO domain data that meets the preset conditions to obtain the first non-time series feature matrix of the user data;

[0048] Extracting the user's churn label from the business platform data to obtain the first label feature matrix data;

[0049] The first label feature matrix data is respectively matrix-connected with the first time series feature matrix and the first non-time series feature matrix to obtain user data.

[0050] Among them, BO domain data includes B domain data and O domain data.

[0051] The word vector is extracted from the original user data, and specifically, the word vector can be extracted from the user playback behavior log data.

[0052] Specifically, the processing of time series data can be specifically carried out by extracting the names of programs watched in the past 1 week to the past n weeks in units of weeks, and constructing an N*M order time series data matrix of (user, program name matrix), wherein the program name matrix is ​​an n-tuple containing the names of programs watched in the past i weeks, and obtaining the first time series feature matrix.

[0053] The processing of non-time series data can specifically include removing features with missing values ​​exceeding 10% of the business platform data and BO domain data, and filling in "0" or "none" with empty meaning characters if the missing values ​​are less than 10%. Abnormal features with feature values ​​that are 70% the same are removed; thus, the processing of missing values ​​and abnormal values ​​of the business platform data and BO domain data is completed, and the first non-time series feature matrix is ​​obtained.

[0054] The extraction of churn labels can be specifically based on the business platform data, marking users who have unsubscribed within a week, marking churned users as 1, and non-churned users as 0, and constructing an N*2-order matrix of (user, churn or not). In this way, the extraction of churn label information is completed, and the first label feature matrix data is obtained.

[0055] Finally, the first label feature matrix data is respectively matrix-connected with the first time series feature matrix and the first non-time series feature matrix to obtain user data, that is, synthesized data to be predicted.

[0056] In a possible embodiment, before step 120, the following steps may also be included:

[0057] Step 210, obtaining historical user data, the historical user data including historical user playback behavior data, historical BO domain data and historical business platform data;

[0058] Step 220, extracting the word vector of historical user playback behavior data to obtain a second time series feature matrix;

[0059] Step 230, eliminating historical BO domain data that meets preset conditions to obtain a second non-time series feature matrix of user data;

[0060] Step 240, extracting the user churn labels from the historical business platform data to obtain second label feature matrix data;

[0061] Step 250, respectively performing matrix connection on the second label feature matrix data, the second time series feature matrix, and the second non-time series feature matrix to obtain the second time series feature matrix data and the second non-time series feature matrix data of the historical user data;

[0062] Step 260, using a natural language processing algorithm to process the second time series feature matrix data and the second non-time series feature matrix data respectively, to obtain latent feature matrix data having connections between the features;

[0063] Step 270, training a hybrid multimodal churn prediction model of latent features based on latent feature matrix data, preset feature weights, preset time factor weights, and preset latent feature weights.

[0064] Before step 120, a hybrid multimodal churn prediction model with latent features needs to be trained.

[0065] The data processing process of steps 210 - 250 is similar to the preprocessing process of the original user data involved in step 110 , and will not be described in detail here.

[0066] Applying an algorithm based on natural language processing to process the second time series feature matrix data and the second non-time series feature matrix data respectively, to obtain latent feature matrix data with connections between the features;

[0067] Among them, hidden features refer to the hidden common features between each user in addition to these labeled features. Generally speaking, the data records of each user are shown in Table 1, and each user data will have labeled explicit features such as "watching a movie".

[0068] Table 1

[0069] User ID Watch the video 13452 "Lost and Love" and "Mission: Impossible 3" 59863 "The Final Battle" "A Simple Life" 35845 "Awakening Age" "The Hidden Corner" 43859 White Snake, Avatar 74583 Pleasant Goat and Big Big Wolf, Tom and Jerry

[0070] But in real life, Figure 2 As shown in the figure, for example, users who watch "The Decisive Battle" and "Awakening Age" may have hidden internal connections such as "history enthusiasts, college students, and patriotic enthusiasm", and users who watch "White Snake" and "Pleasant Goat and Big Big Wolf" may have hidden internal connections such as "having a childlike heart and watching on demand for children". The true value of each label feature of the user can be described by these hidden internal connections. These hidden internal connections are represented by feature vectors, and these feature vectors describing the internal connections are called hidden features. For example, "The Decisive Battle" can be described as Table 2 using hidden features.

[0071] Table 2

[0072]

[0073] Here, by introducing a natural language processing algorithm, the labeled user information is converted into user information described in natural language, the analysis data is converted into analysis semantics, and the natural language processing algorithm is used to actively discover the group characteristics and implicit characteristics between the original feature values, solving the problem of incomplete and inaccurate feature selection caused by missing group characteristics or ignoring implicit characteristics in the labeled user feature analysis. Here, the group feature values ​​can be discovered through natural language processing methods, and the model can be trained based on group characteristics, which shields the model training from the dependence on individual user historical data and can quickly adapt to new users. In addition, by describing individual information in natural language, the hidden characteristics of users can be fully discovered, and the intention of each user to leave the network can be accurately identified.

[0074] The mixed multi-modal prediction model based on hidden features (MMPM-HF) performs fusion training on the second time series feature matrix data and the second non-time series feature matrix data, recursively optimizes the feature weights, and outputs the probability of user churn in a future time period, solving the problem of missing non-time information of users and being unable to make predictions, and can quickly adapt to the data added over time.

[0075] The latent feature matrix data includes first K-dimensional latent feature matrix data and second K-dimensional latent feature matrix data; step 260 includes:

[0076] The second time series feature matrix data are processed respectively by using a natural language processing algorithm to obtain a first K-dimensional latent feature matrix reflecting the internal connection of each feature in the second time series feature matrix data; wherein the first K-dimensional latent feature matrix includes the second time series feature matrix data and a first d-dimensional latent feature value, and the first d-dimensional latent feature value is the d-dimensional latent feature value in the second time series feature matrix data;

[0077] The second non-time series feature matrix data is processed respectively by using a natural language processing algorithm to obtain a second K-dimensional latent feature matrix reflecting the internal relationship between each feature in the second non-time series feature matrix data; wherein the second K-dimensional latent feature matrix includes the second non-time series feature matrix data and the second d-dimensional latent feature value, and the second d-dimensional latent feature value is the d-dimensional latent feature value in the second non-time series feature matrix data;

[0078] The first d-dimensional latent feature value in the first K-dimensional latent feature matrix replaces the second time series feature matrix data in the first K-dimensional latent feature matrix to obtain the first K-dimensional latent feature matrix data;

[0079] The second non-time-sequential feature matrix data in the second K-dimensional latent feature matrix is ​​replaced by the second d-dimensional latent feature value in the second K-dimensional latent feature matrix to obtain the second K-dimensional latent feature matrix data.

[0080] First, the second time series feature matrix data is processed respectively by using a natural language processing algorithm to obtain a first K-dimensional latent feature matrix reflecting the internal connection of each feature in the second time series feature matrix data; wherein the first K-dimensional latent feature matrix includes the second time series feature matrix data and the first d-dimensional latent feature value, and the first d-dimensional latent feature value is the d-dimensional latent feature value in the second time series feature matrix data, which may specifically include:

[0081] Generate the hidden eigenvalues ​​of each eigenvalue in the original feature matrix based on the natural language processing algorithm: The word2vec algorithm can be introduced to model the time series feature matrix and the non-time series feature matrix to generate the first K-dimensional hidden feature matrix that reflects the internal connection of each eigenvalue in the original feature matrix. The expression is Y = (A, A'), where A represents the eigenvalue A in the original feature matrix, and A' represents the d-dimensional hidden eigenvalue of the eigenvalue A, that is, A is represented by the d-dimensional vector A'.

[0082] The second non-time series feature matrix data is processed respectively by using a natural language processing algorithm to obtain a second K-dimensional latent feature matrix reflecting the internal relationship between the features in the second non-time series feature matrix data; wherein the second K-dimensional latent feature matrix includes the second non-time series feature matrix data and the second d-dimensional latent feature value, and the second d-dimensional latent feature value is the d-dimensional latent feature value in the second non-time series feature matrix data, which may specifically include:

[0083] Generate the hidden eigenvalues ​​of each eigenvalue in the original feature matrix based on the natural language processing algorithm: The word2vec algorithm can be introduced to model the time series feature matrix and the non-time series feature matrix to generate a second K-dimensional hidden feature matrix that reflects the internal connection between the eigenvalues ​​in the original feature matrix. The expression is Y = (A, A'), where A represents the eigenvalue A in the original feature matrix, and A' represents the d-dimensional hidden eigenvalue of the eigenvalue A, that is, A is represented by the d-dimensional vector A'.

[0084] Then, replacing the second time series feature matrix data in the first K-dimensional latent feature matrix with the first d-dimensional latent feature value in the first K-dimensional latent feature matrix to obtain the first K-dimensional latent feature matrix data may specifically include:

[0085] The latent eigenvalue replaces the original matrix eigenvalue to form a latent feature matrix; the original eigenvalue (second time series feature matrix data) in the first K-dimensional latent feature matrix is ​​replaced with the first d-dimensional latent eigenvalue, and the feature matrix is ​​transformed from E to E', V ij represents the original eigenvalue, V ij ' is represented by V ij The d-dimensional latent feature vector is as follows: Figure 3 shown.

[0086] Replacing the second non-time-sequential feature matrix data in the second K-dimensional latent feature matrix with the second d-dimensional latent feature value in the second K-dimensional latent feature matrix to obtain the second K-dimensional latent feature matrix data may specifically include:

[0087] Replace the original eigenvalues ​​(second non-time-sequential feature matrix data) in the second K-dimensional latent feature matrix with the corresponding second d-dimensional latent eigenvalues, and transform the feature matrix from E to E', V ij represents the original eigenvalue, V ij ' is represented by V ij The d-dimensional latent feature vector is, as follows Figure 3 shown.

[0088] Therefore, the natural language processing method is used to solve the recognition problem, and the internal connections between the telecommunications BO domain data, business platform data, and user viewing behavior log data are analyzed. Important features are actively discovered in the depth of the group and the breadth of time, which simplifies the feature analysis process and solves the problem that manual feature analysis cannot discover group characteristics and hidden important features.

[0089] Wherein, step 270 includes:

[0090] Performing matrix transformation on latent feature matrix data based on preset feature weights, preset time factor weights, and preset latent feature weights to obtain transformed matrix data;

[0091] The initial prediction model is trained based on the changed matrix data to obtain a hybrid multimodal churn prediction model with latent features.

[0092] Here, through the preset time factor weight, the positive feedback of the time factor on the prediction result can be increased and the preset feature weight can be assigned according to the influence. The latent feature matrix data is transformed based on the preset feature weight, the preset time factor weight and the preset latent feature weight to obtain the transformed matrix data; and the initial prediction model is trained based on the changed matrix data to construct a hybrid multimodal churn prediction model based on latent features, which can accurately predict the user churn data in the target time period.

[0093] Among them, the above-mentioned step of performing matrix transformation on latent feature matrix data based on preset feature weights, preset time factor weights and preset latent feature weights to obtain the transformed matrix data may specifically include the following steps:

[0094] Performing eigenvalue vector factor transformation on the time series feature vector of each user in the second time series feature matrix data based on a preset feature weight to obtain a first time series feature vector matrix;

[0095] Performing a time factor weight classification matrix transformation on the first time series feature vector matrix based on a preset time factor weight to obtain a second time series feature vector matrix, wherein in the matrix of the preset time factor weight, the feature of each row includes only one non-zero value;

[0096] Combining the second time series eigenvector matrix to obtain a third time series eigenvector matrix;

[0097] Superimposing the third time series feature vector and the second non-time series feature matrix data to obtain a feature matrix reflecting the user's historical behavior and attribute information;

[0098] Based on the preset latent feature weights, the feature matrix is ​​weighted to obtain the transformed matrix data.

[0099] First, the above-mentioned time series feature vector of each user in the second time series feature matrix data is transformed into an eigenvalue vector factor based on a preset feature weight to obtain a first time series feature vector matrix, which specifically includes the following steps:

[0100] Initialize the feature weights α, β, and γ, and update the weight matrix; α is the weight matrix constructed according to the number of time series feature vectors, β is the matrix constructed according to the time factor of the business model, and γ is the (d+e)*1 order weight matrix after the time series and non-time series feature vectors are superimposed. Then, update the weight matrix; then, transform the time series feature matrix according to the eigenvalue weight vector α and the time factor weight classification matrix; finally, the eigenvalue vector factor is transformed to obtain the first time series feature vector matrix U′=U*α.

[0101] Next, the first time series feature vector matrix is ​​transformed into a time factor weight classification matrix based on the preset time factor weight to obtain a second time series feature vector matrix, wherein in the matrix of the preset time factor weight, the feature of each row includes only one non-zero value, specifically including the following steps:

[0102] Among them, in the matrix of preset time factor weights, the features of each row include only one non-zero value, wherein row, represents the number of users.

[0103] Based on the preset time factor weight, the first time series feature vector matrix is ​​transformed into a time factor weight classification matrix to obtain a second time series feature vector matrix, which can be specifically: U″=U′*β / d

[0104] Specifically, construct the time factor weight classification matrix The time factor weight classification matrix is ​​a weight classification matrix. Each row can only have one non-zero value, and each column has the same value. i represents the number of eigenvectors of the user vector U, j represents the number of classifications, and β ij Indicates the weight of the impact of this classification on the result.

[0105] Assume that each user data of the time series feature matrix is ​​a vector U i =(V i1 ',…,V ij '), the eigenvector V ij ' is a d-dimensional vector containing d latent eigenvalues, and the initialized eigenvalue factor weight vector is α i =(α i1 ,…,α ij ).

[0106] The time factor weight classification matrix is: Then the time series feature matrix transformation needs to be i Perform eigenvalue factor transformation, U i ′=U i *α i =(α i1 V i1 ',…, α ij V ij '), then i ′ is transformed by time factor to obtain the second time series eigenvector matrix:

[0107]

[0108] Then, the above-mentioned merging of the second time series eigenvector matrix to obtain the third time series eigenvector matrix specifically includes the following steps:

[0109] The second time series eigenvector matrix U i ″ is merged into the final time series feature matrix, that is, the third time series feature vector matrix:

[0110]

[0111] Then, the third time series feature vector and the second non-time series feature matrix data are superimposed to obtain a feature matrix reflecting the user's historical behavior and attribute information, which specifically includes the following steps:

[0112] The non-time series feature matrix superposition transformation superimposes the third time series feature vector matrix E″ and the second non-time series feature matrix data F to form a feature matrix T=(E″F) reflecting the user's historical behavior and attribute information.

[0113] The third time series feature vector matrix E″ is an N*d-order matrix, d is the latent feature dimension of the time series feature matrix, and F is an N*e-order matrix, e is the latent feature dimension of the non-time series feature matrix.

[0114] Finally, the above-mentioned weighting of the feature matrix based on the preset latent feature weights to obtain the transformed matrix data specifically includes the following steps:

[0115] Construct a latent feature weight selection matrix that reflects the latent eigenvalue weights

[0116] Use the latent eigenvalue weight classification selection matrix γ to superimpose and calculate the weighted average value of each eigenvector in the time series and non-time series latent feature selection matrices; superimpose the non-time series feature matrix on the transformed time series feature matrix, construct the superimposed matrix and perform weighted transformation. The formula is: T′=T*γ / (d+e).

[0117] Based on the preset latent feature weights, the feature matrix is ​​weighted to obtain the transformed matrix data. The formula is:

[0118]

[0119] Among them, t ij Represents each latent eigenvalue in the matrix T.

[0120] Among them, the above-mentioned step of training the initial prediction model based on the changed matrix data to obtain the hybrid multimodal churn prediction model of latent features may specifically include the following steps:

[0121] The transformed matrix data is input into the initial prediction model, and the second probability of user diversion is calculated by the initial prediction model;

[0122] Obtain a loss function value based on the comparison between the second probability and a preset true result;

[0123] When the loss function value meets the preset conditions, it is determined that the model corresponding to the second probability value is a hybrid multimodal churn prediction model of latent features.

[0124] According to the superimposed feature matrix, the classification probability, that is, the second probability of user diversion, is calculated.

[0125] The probability formula is The numerator represents the exponent of the prediction value of the classification node, and the denominator represents the sum of the exponents of the prediction values ​​of all nodes. Compare the two columns of prediction values ​​of T', and the one with a higher probability is the output result. The output result is compared with the actual result, and the loss function value is obtained based on the comparison between the second probability and the preset actual result; when the loss function value meets the preset conditions, it is determined that the model corresponding to the second probability value is a hybrid multimodal churn prediction model with hidden features.

[0126] Specifically, the training parameters of the model can be iteratively tuned according to the gradient descent algorithm, and the optimal model is output when the loss function is minimized. Otherwise, the weight matrix is ​​iteratively updated.

[0127] Therefore, a hybrid multimodal churn prediction model based on latent features is proposed: the model constructs eigenvalue weight vectors, time factor weight classification matrices, and latent eigenvalue weight classification selection matrices, performs multimodal hybrid training on time-series-based user behavior feature matrices and non-time-series user information feature matrices, recursively optimizes feature weights, and outputs the probability of user churn within the target time period. This solves the problem that the model cannot make predictions due to the lack of non-time-series information, can quickly adapt to new data added over time, and improves prediction accuracy.

[0128] This involves step 120 .

[0129] The first time series feature matrix data and the first non-time series feature matrix data are input into a preset latent feature hybrid multimodal churn prediction model, and a first probability of user churn is calculated by the latent feature hybrid multimodal churn prediction model.

[0130] The hybrid multimodal churn prediction model of latent features is obtained by training based on latent feature matrix data of historical user data, preset feature weights, preset time factor weights, and preset latent feature weights; the latent feature matrix data is latent feature matrix data generated based on the second temporal feature matrix data and the second non-temporal feature matrix data of historical user data. The training process of the hybrid multimodal churn prediction model of latent features is specifically shown in the above steps 210 to 270.

[0131] This involves step 130 .

[0132] The churned user data within the target time period is determined based on the first probability of user churn.

[0133] The data of users who have lost during the target time period, that is, the list of users who may have lost during the target time period.

[0134] In summary, in the embodiment of the present application, by acquiring user data, the user data includes the first time series feature matrix data and the first non-time series feature matrix data of user behavior; the first time series feature matrix data and the first non-time series feature matrix data are input into the preset hybrid multimodal churn prediction model of latent features, and the first probability of user churn is calculated by the hybrid multimodal churn prediction model of latent features; wherein, since the hybrid multimodal churn prediction model of latent features is obtained by training based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; the latent feature matrix data is the latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data; here, accurate discovery of latent features is achieved, and finally the lost user data within the target time period is determined based on the first probability of user churn. Thus, the lost user data can be accurately identified.

[0135] Based on the above Figure 1 The user churn prediction method shown in the embodiment of the present application also provides a data processing device, such as Figure 4 As shown, the apparatus 400 may include:

[0136] The acquisition module 410 is used to acquire user data, where the user data includes first time series feature matrix data and first non-time series feature matrix data of user behavior.

[0137] Processing module 420 is used to input the first time series feature matrix data and the first non-time series feature matrix data into a preset latent feature hybrid multimodal churn prediction model, and calculate the first probability of user churn through the latent feature hybrid multimodal churn prediction model; wherein the latent feature hybrid multimodal churn prediction model is trained based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; the latent feature matrix data is latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data.

[0138] The determination module 430 is configured to determine lost user data within a target time period based on the first probability of user loss.

[0139] In a possible implementation, the acquisition module 410 is specifically configured to:

[0140] Get the original user data;

[0141] The processing module 420 is also used to extract word vectors from the original user data, remove data that meets preset conditions, and extract churn labels to obtain user data.

[0142] In a possible implementation, the processing module 420 is specifically configured to:

[0143] Extract the word vector of the user's playback behavior data to obtain a first time series feature matrix;

[0144] Eliminate the BO domain data that meets the preset conditions to obtain the first non-time series feature matrix of the user data;

[0145] Extracting the user's churn label from the business platform data to obtain the first label feature matrix data;

[0146] The first label feature matrix data is respectively matrix-connected with the first time series feature matrix and the first non-time series feature matrix to obtain user data.

[0147] In a possible implementation, the acquisition module 410 is further used to acquire historical user data, where the historical user data includes historical user playback behavior data, historical BO domain data, and historical business platform data.

[0148] The processing module 420 is further used to extract the word vector of the historical user playback behavior data to obtain a second time series feature matrix.

[0149] The processing module 420 is further used to remove the historical BO domain data that meets the preset conditions to obtain a second non-time series feature matrix of the user data.

[0150] The processing module 420 is further used to extract the user churn labels in the historical business platform data to obtain second label feature matrix data.

[0151] The processing module 420 is further used to perform matrix connection on the second label feature matrix data, the second time series feature matrix and the second non-time series feature matrix to obtain the second time series feature matrix data and the second non-time series feature matrix data of the historical user data.

[0152] The processing module 420 is further used to process the second time series feature matrix data and the second non-time series feature matrix data respectively using a natural language processing algorithm to obtain latent feature matrix data with connections between the features.

[0153] The processing module 420 is also used to train a hybrid multimodal churn prediction model of latent features based on latent feature matrix data, preset feature weights, preset time factor weights and preset latent feature weights.

[0154] In a possible implementation, the latent feature matrix data includes first K-dimensional latent feature matrix data and second K-dimensional latent feature matrix data; the processing module 420 is specifically used to:

[0155] The second time series feature matrix data are processed respectively by using a natural language processing algorithm to obtain a first K-dimensional latent feature matrix reflecting the internal connection of each feature in the second time series feature matrix data; wherein the first K-dimensional latent feature matrix includes the second time series feature matrix data and a first d-dimensional latent feature value, and the first d-dimensional latent feature value is the d-dimensional latent feature value in the second time series feature matrix data;

[0156] The second non-time series feature matrix data is processed respectively by using a natural language processing algorithm to obtain a second K-dimensional latent feature matrix reflecting the internal relationship between each feature in the second non-time series feature matrix data; wherein the second K-dimensional latent feature matrix includes the second non-time series feature matrix data and the second d-dimensional latent feature value, and the second d-dimensional latent feature value is the d-dimensional latent feature value in the second non-time series feature matrix data;

[0157] The first d-dimensional latent feature value in the first K-dimensional latent feature matrix replaces the second time series feature matrix data in the first K-dimensional latent feature matrix to obtain the first K-dimensional latent feature matrix data;

[0158] The second non-time-sequential feature matrix data in the second K-dimensional latent feature matrix is ​​replaced by the second d-dimensional latent feature value in the second K-dimensional latent feature matrix to obtain the second K-dimensional latent feature matrix data.

[0159] In a possible implementation, the processing module 420 is specifically configured to:

[0160] Performing matrix transformation on latent feature matrix data based on preset feature weights, preset time factor weights, and preset latent feature weights to obtain transformed matrix data;

[0161] The initial prediction model is trained based on the changed matrix data to obtain a hybrid multimodal churn prediction model with latent features.

[0162] In a possible implementation, the processing module 420 is specifically configured to:

[0163] Performing eigenvalue vector factor transformation on the time series feature vector of each user in the second time series feature matrix data based on a preset feature weight to obtain a first time series feature vector matrix;

[0164] Performing a time factor weight classification matrix transformation on the first time series feature vector matrix based on a preset time factor weight to obtain a second time series feature vector matrix, wherein in the matrix of the preset time factor weight, the feature of each row includes only one non-zero value;

[0165] Combining the second time series eigenvector matrix to obtain a third time series eigenvector matrix;

[0166] Superimposing the third time series feature vector and the second non-time series feature matrix data to obtain a feature matrix reflecting the user's historical behavior and attribute information;

[0167] Based on the preset latent feature weights, the feature matrix is ​​weighted to obtain the transformed matrix data.

[0168] In a possible implementation, the processing module 420 is specifically configured to:

[0169] The transformed matrix data is input into the initial prediction model, and the second probability of user diversion is calculated by the initial prediction model;

[0170] Obtain a loss function value based on the comparison between the second probability and a preset true result;

[0171] When the loss function value meets the preset conditions, it is determined that the model corresponding to the second probability value is a hybrid multimodal churn prediction model of latent features.

[0172] In summary, in the embodiment of the present application, by acquiring user data, the user data includes the first time series feature matrix data and the first non-time series feature matrix data of user behavior; the first time series feature matrix data and the first non-time series feature matrix data are input into the preset hybrid multimodal churn prediction model of latent features, and the first probability of user churn is calculated by the hybrid multimodal churn prediction model of latent features; wherein, since the hybrid multimodal churn prediction model of latent features is obtained by training based on the latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; the latent feature matrix data is the latent feature matrix data generated based on the second time series feature matrix data and the second non-time series feature matrix data of historical user data; here, accurate discovery of latent features is achieved, and finally the lost user data within the target time period is determined based on the first probability of user churn. Thus, the lost user data can be accurately identified.

[0173] Figure 5 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application is shown.

[0174] The electronic device may include a processor 501 and a memory 502 storing computer program instructions.

[0175] Specifically, the processor 501 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0176] The memory 502 may include a large capacity memory for data or instructions. By way of example and not limitation, the memory 502 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. In appropriate cases, the memory 502 may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory 502 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, the memory 502 is a non-volatile solid-state memory. In a specific embodiment, the memory 502 includes a read-only memory (ROM). In appropriate cases, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM) or a flash memory or a combination of two or more of these.

[0177] The processor 501 reads and executes the computer program instructions stored in the memory 502 to implement any one of the user churn prediction methods in the embodiments shown in the figure.

[0178] In one example, the electronic device may further include a communication interface 503 and a bus 510. Figure 5 As shown, the processor 501, the memory 502, and the communication interface 503 are connected via a bus 510 and communicate with each other.

[0179] The communication interface 503 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0180] Bus 510 includes hardware, software or both, and the parts of electronic equipment are coupled to each other. For example, but not limitation, bus may include accelerated graphics port (AGP) or other graphics bus, enhanced industrial standard architecture (EISA) bus, front side bus (FSB), hypertransport (HT) interconnection, industrial standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI- Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 510 may include one or more buses. Although the present application embodiment describes and shows a specific bus, the present application considers any suitable bus or interconnection.

[0181] The electronic device can execute the user churn prediction method in the embodiment of the present application, thereby realizing the combination Figure 1 to Figure 2 Describe the user churn prediction method.

[0182] In addition, in combination with the user churn prediction method in the above embodiment, the present application embodiment can provide a computer-readable storage medium for implementation. The computer-readable storage medium stores computer program instructions; when the computer program instructions are executed by the processor, the Figure 1 to Figure 2 User churn prediction method in .

[0183] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0184] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present application are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0185] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be performed simultaneously.

[0186] The above is only a specific implementation of the present application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the protection scope of the present application is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the protection scope of this application.

Claims

1. A method for predicting user churn, characterized in that: The method comprises: Acquire user data, wherein the user data includes first time series feature matrix data and first non-time series feature matrix data of user behavior; The first time series feature matrix data and the first non-time series feature matrix data are input into a preset hybrid multimodal churn prediction model of latent features, and the first probability of user churn is calculated by the hybrid multimodal churn prediction model of latent features; wherein the hybrid multimodal churn prediction model of latent features is obtained by training based on latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; the latent feature matrix data is latent feature matrix data generated based on second time series feature matrix data and second non-time series feature matrix data of the historical user data; Determine the lost user data within the target time period based on the first probability of user loss; Before inputting the first time series feature matrix data and the first non-time series feature matrix data into a preset hybrid multimodal churn prediction model of latent features and calculating the first probability of user churn by using the hybrid multimodal churn prediction model of latent features, the method further includes: Performing a matrix transformation on the latent feature matrix data based on the preset feature weight, the preset time factor weight and the preset latent feature weight to obtain transformed matrix data; Training an initial prediction model based on the transformed matrix data to obtain a hybrid multimodal churn prediction model of the latent features; The obtaining of user data comprises: Get the original user data; Extract word vectors from the original user data, remove data that meets preset conditions, and extract churn labels to obtain the user data; The user data includes user playback behavior data, BO domain data and business platform data; the word vector extraction, elimination of data that meets preset conditions, and extraction of churn labels are performed on the original user data to obtain the user data, including: Extract the word vector of the user's playback behavior data to obtain a first time series feature matrix; Eliminating the BO domain data that meets a preset condition to obtain a first non-time series feature matrix of user data; Extracting the churn label of the user from the business platform data to obtain first label feature matrix data; The first label feature matrix data is respectively matrix-connected with the first time series feature matrix and the first non-time series feature matrix to obtain the user data.

2. The method according to claim 1, characterized in that: Before inputting the first time series feature matrix data and the first non-time series feature matrix data into a preset hybrid multimodal churn prediction model of latent features and calculating the first probability of user churn by using the hybrid multimodal churn prediction model of latent features, the method further includes: Acquire the historical user data, wherein the historical user data includes historical user playback behavior data, historical BO domain data, and historical service platform data; Extracting the word vector of the historical user playback behavior data to obtain a second time series feature matrix; Eliminating the historical BO domain data that meets the preset conditions to obtain a second non-time series feature matrix of user data; Extracting the churn label of the user from the historical business platform data to obtain second label feature matrix data; Respectively performing matrix connection on the second label feature matrix data, the second time series feature matrix, and the second non-time series feature matrix to obtain the second time series feature matrix data and the second non-time series feature matrix data of the historical user data; The second time series feature matrix data and the second non-time series feature matrix data are processed respectively using a natural language processing algorithm to obtain latent feature matrix data with connections between the features.

3. The method according to claim 2, characterized in that The latent feature matrix data includes first K-dimensional latent feature matrix data and second K-dimensional latent feature matrix data; the natural language processing algorithm is used to process the second time series feature matrix data and the second non-time series feature matrix data respectively to obtain latent feature matrix data with connections between the features, including: The second time series feature matrix data are processed respectively by a natural language processing algorithm to obtain a first K-dimensional latent feature matrix reflecting the internal relationship between each feature in the second time series feature matrix data; wherein the first K-dimensional latent feature matrix includes the second time series feature matrix data and a first d-dimensional latent feature value, and the first d-dimensional latent feature value is the d-dimensional latent feature value in the second time series feature matrix data; The second non-time series feature matrix data is processed respectively by a natural language processing algorithm to obtain a second K-dimensional latent feature matrix reflecting the internal relationship between the features in the second non-time series feature matrix data; wherein the second K-dimensional latent feature matrix includes the second non-time series feature matrix data and the second d-dimensional latent feature value, and the second d-dimensional latent feature value is the d-dimensional latent feature value in the second non-time series feature matrix data; Replacing the second time series feature matrix data in the first K-dimensional latent feature matrix with the first d-dimensional latent feature value in the first K-dimensional latent feature matrix to obtain the first K-dimensional latent feature matrix data; The second non-time-series feature matrix data in the second K-dimensional latent feature matrix is ​​replaced by the second d-dimensional latent feature values ​​in the second K-dimensional latent feature matrix to obtain second K-dimensional latent feature matrix data.

4. The method according to claim 1, characterized in that: The performing matrix transformation on the latent feature matrix data based on the preset feature weight, the preset time factor weight and the preset latent feature weight to obtain the transformed matrix data includes: Performing eigenvalue vector factor transformation on the time series feature vector of each user in the second time series feature matrix data based on the preset feature weight to obtain a first time series feature vector matrix; Based on the preset time factor weight, the first time series feature vector matrix is ​​transformed into a time factor weight classification matrix to obtain a second time series feature vector matrix, wherein in the matrix of the preset time factor weight, the feature of each row includes only one non-zero value; Combining the second time series eigenvector matrix to obtain a third time series eigenvector matrix; Superimposing the third time series feature vector and the second non-time series feature matrix data to obtain a feature matrix reflecting the user's historical behavior and attribute information; Based on the preset latent feature weights, the feature matrix is ​​weighted to obtain the transformed matrix data.

5. The method according to claim 4, characterized in that The initial prediction model is trained based on the transformed matrix data to obtain the hybrid multimodal churn prediction model of the latent features, including: Inputting the transformed matrix data into an initial prediction model, and calculating a second probability of user diversion through the initial prediction model; Obtain a loss function value based on comparison between the second probability and a preset true result; When the loss function value satisfies a preset condition, it is determined that the model corresponding to the second probability value is a hybrid multimodal churn prediction model of the latent features.

6. A user churn prediction device, characterized in that: The device comprises: An acquisition module, used to acquire user data, wherein the user data includes first time series feature matrix data and first non-time series feature matrix data of user behavior; A processing module, used for inputting the first time series feature matrix data and the first non-time series feature matrix data into a preset hybrid multimodal churn prediction model of latent features, and calculating the first probability of user churn through the hybrid multimodal churn prediction model of latent features; wherein the hybrid multimodal churn prediction model of latent features is obtained by training based on latent feature matrix data of historical user data, preset feature weights, preset time factor weights and preset latent feature weights; the latent feature matrix data is latent feature matrix data generated based on second time series feature matrix data and second non-time series feature matrix data of the historical user data; A determination module, configured to determine lost user data within a target time period based on the first probability of user loss; The processing module is further used for: Performing a matrix transformation on the latent feature matrix data based on the preset feature weight, the preset time factor weight and the preset latent feature weight to obtain transformed matrix data; Training an initial prediction model based on the transformed matrix data to obtain a hybrid multimodal churn prediction model of the latent features; The acquisition module is specifically used for: Get the original user data; The processing module is further used to extract word vectors from the original user data, remove data that meets preset conditions, and extract churn labels to obtain the user data; The user data includes user playback behavior data, BO domain data and business platform data; the processing module is specifically used for: Extract the word vector of the user's playback behavior data to obtain a first time series feature matrix; Eliminating the BO domain data that meets a preset condition to obtain a first non-time series feature matrix of user data; Extracting the churn label of the user from the business platform data to obtain first label feature matrix data; The first label feature matrix data is respectively matrix-connected with the first time series feature matrix and the first non-time series feature matrix to obtain the user data.

7. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the user churn prediction method according to any one of claims 1 to 5 is implemented.

8. A readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the user churn prediction method according to any one of claims 1 to 5 is implemented.

Images