User off-network prediction method and device, electronic equipment and storage medium

Abstract

The invention provides a user off-network prediction method and apparatus, an electronic device and a storage medium. The method comprises the steps of obtaining basic attribute data and time sequence behavior data of a user; determining high-order spatial domain feature information in the basic attribute data, and determining high-order time domain feature information in the time sequence behavior data; fusing the high-order spatial domain feature information and the high-order time domain feature information to obtain a fused feature set; and predicting the user off-network probability based on the fused feature set. That is to say, the time sequence behavior data and the basic attribute data of the user are comprehensively considered, and the extracted high-order time-space domain features are fused, so that multi-dimensional analysis and high-precision prediction of the user off-network behavior are realized, and the accuracy and efficiency of user off-network prediction are improved.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a method, device, electronic device, and computer-readable storage medium for predicting user churn. Background Art

[0002] The current user growth in the mobile communications market is slowing down, and the number portability policy has exacerbated the risk of user churn. Accurately predicting and retaining users who may leave the network has become a key task for operators.

[0003] In related technologies, user churn prediction methods usually rely on traditional machine learning algorithms such as Naive Bayes, Support Vector Machine (SVM), logistic regression, and decision trees. These methods mostly rely on static feature engineering of user information, such as user consumption behavior and social network behavior, and have subjective biases, resulting in the inability to fully capture user behavior and reducing the accuracy of predictions.

[0004] Therefore, how to improve the accuracy of user off-grid prediction is a problem that needs to be solved. Summary of the Invention

[0005] The present invention provides a method, device, electronic device, and computer-readable storage medium for predicting user churn, to at least address the problem in related art of low accuracy in predicting user churn due to a failure to fully capture user behavior. The technical solutions of the present invention are as follows:

[0006] According to a first aspect of an embodiment of the present invention, a method for predicting user churn is provided, comprising:

[0007] Obtain users' basic attribute data and time-series behavior data;

[0008] Determining high-order spatial domain feature information in the basic attribute data, and determining high-order temporal domain feature information in the temporal behavior data;

[0009] Fusing the high-order spatial feature information and the high-order temporal feature information to obtain a fused feature set;

[0010] Based on the fused feature set, the probability of the user leaving the network is predicted.

[0011] Optionally, obtaining the user's basic attribute data and time-series behavior data includes:

[0012] Get user data within a set time period;

[0013] Preprocessing the user data to obtain preprocessed user data;

[0014] The pre-processed user data is divided into basic attribute data and time series behavior data.

[0015] Optionally, determining the high-order spatial feature information in the basic attribute data includes:

[0016] Selecting spatial features related to the user's off-grid behavior from the user's basic attribute data using a random forest algorithm;

[0017] Determine the importance score of each selected spatial feature for predicting user churn behavior;

[0018] From the importance scores of all airspace features, select the airspace features whose importance scores are higher than the preset airspace threshold as important airspace features;

[0019] The important spatial features are input into the first deep convolutional neural network for convolution processing in the spatial dimension, and the attention weights of the important spatial features are adjusted using the attention mechanism to obtain potential high-order spatial feature information.

[0020] Optionally, determining high-order time-domain feature information in the time series behavior data includes:

[0021] Converting the time series behavior data into two-dimensional time series behavior data;

[0022] Selecting time-domain features related to the user's off-grid behavior from the two-dimensional time series behavior data using a random forest algorithm;

[0023] Determine the importance score of each selected temporal feature for predicting user behavior;

[0024] From the importance scores of all time domain features, select the time domain features whose importance scores are higher than the preset time domain threshold to obtain the important time domain features;

[0025] The important time domain features are input into the second deep convolutional neural network for convolution processing in the spatial and temporal dimensions, and the attention weights of the important time domain features are adjusted using the attention mechanism to obtain potential high-order time domain feature information.

[0026] Optionally, fusing the high-order spatial feature information and the high-order temporal feature information to obtain a fused feature set includes:

[0027] The high-order spatial domain feature information and the high-order temporal domain feature information are subjected to a cascade fusion operation along the channel axis to obtain a fusion feature set in which the high-order spatial domain features and the high-order temporal domain features are merged into a feature vector.

[0028] Optionally, predicting a probability of the user leaving the network based on the fused feature set includes:

[0029] Performing weighted processing on the fused feature set using an attention mechanism to obtain a weighted fused feature set;

[0030] The trained user churn prediction model is used to perform prediction processing on the weighted fused feature set to obtain the probability of the user churn, wherein the user churn prediction model is obtained by training the weighted fused feature set using the binary cross entropy loss function of a deep learning convolutional neural network.

[0031] Optionally, the method further includes: pre-training a user churn prediction model in the following manner:

[0032] Obtaining a training sample set, the training sample set including: a fusion feature set based on spatial domain features of basic attribute data of users and temporal domain features of temporal behavior data, and an actual user churn probability;

[0033] The fused feature set is input into a user churn prediction model based on a convolutional neural network for training. During the training process, the user churn probability output by the user churn prediction model is compared with the actual user churn probability, and the difference between the comparisons is calculated. The difference is used as the loss value of the binary cross entropy loss function. Based on the loss value, the parameters of the user churn prediction model are adjusted through a back-propagation mechanism until convergence, thereby obtaining a trained user churn prediction model.

[0034] According to a second aspect of an embodiment of the present invention, a user churn prediction device is provided, comprising:

[0035] The acquisition module is used to obtain the user's basic attribute data and time series behavior data;

[0036] A first determining module is used to determine high-order spatial feature information in the basic attribute data;

[0037] A second determining module is used to determine high-order time-domain feature information in the time series behavior data;

[0038] A fusion module, configured to fuse the high-order spatial feature information and the high-order temporal feature information to obtain a fused feature set;

[0039] The prediction module is used to predict the probability of the user leaving the network based on the fused feature set.

[0040] Optionally, the acquisition module includes:

[0041] The data acquisition module is used to obtain the user data of the user within a set time period;

[0042] A preprocessing module, configured to preprocess the user data to obtain preprocessed user data;

[0043] The division module is used to divide the pre-processed user data into basic attribute data and time series behavior data.

[0044] Optionally, the first determining module includes:

[0045] A first spatial feature selection module is configured to select spatial features related to the user's off-grid behavior from the user's basic attribute data using a random forest algorithm;

[0046] a spatial feature score determination module, configured to determine an importance score of each spatial feature selected by the first selection module for predicting user churn;

[0047] The second spatial feature selection module is used to select the spatial features with an importance score higher than a preset spatial threshold from the importance scores of all spatial features as important spatial features;

[0048] The spatial feature processing module is used to input the important spatial features into the first deep convolutional neural network for convolution processing in the spatial dimension, and use the attention mechanism to adjust the attention weights of the important spatial features to obtain potential high-order spatial feature information.

[0049] Optionally, the second determining module includes:

[0050] a conversion module, configured to convert the time series behavior data into two-dimensional time series behavior data;

[0051] A first time-domain feature selection module is configured to select time-domain features related to the user's off-grid behavior from the two-dimensional time-series behavior data using a random forest algorithm;

[0052] A time domain feature score determination module, used to determine the importance score of each selected time domain feature for predicting user behavior;

[0053] The second time domain feature selection module is used to select the time domain features whose importance scores are higher than the preset time domain threshold from the importance scores of all time domain features to obtain important time domain features;

[0054] The time domain feature processing module is used to input the important time domain features into the second deep convolutional neural network for convolution processing in the spatial dimension and the temporal dimension, and use the attention mechanism to adjust the attention weight of the important time domain features to obtain potential high-order time domain feature information.

[0055] Optionally, the fusion module is specifically used to perform a cascade fusion operation on the high-order spatial domain feature information and the high-order time domain feature information along the channel axis to obtain a fusion feature set in which the high-order spatial domain features and the high-order time domain features are merged into a feature vector.

[0056] Optionally, the prediction module includes:

[0057] A weighted processing module, configured to perform weighted processing on the fused feature set using an attention mechanism to obtain a weighted fused feature set;

[0058] The churn probability prediction module is used to use the trained user churn prediction model to predict the weighted fused feature set to obtain the probability of the user churn, wherein the user churn prediction model is obtained by training the weighted fused feature set using the binary cross entropy loss function of the deep learning convolutional neural network.

[0059] Optionally, the device further includes: a training module, configured to pre-train the user churn prediction model in the following manner:

[0060] A training sample acquisition module is used to acquire a training sample set, wherein the training sample set includes: a fusion feature set based on the spatial domain features of the user's basic attribute data and the temporal domain features of the temporal behavior data, and an actual user churn probability;

[0061] A model training module is used to input the fused feature set into a user churn prediction model based on a convolutional neural network for training. During the training process, the user churn probability output by the user churn prediction model is compared with the actual user churn probability, the difference between the comparisons is calculated, and the difference is used as the loss value of the binary cross entropy loss function. Based on the loss value, the parameters of the user churn prediction model are adjusted through a backpropagation mechanism until convergence, thereby obtaining a trained user churn prediction model.

[0062] According to a third aspect of an embodiment of the present invention, there is provided an electronic device, including:

[0063] processor;

[0064] a memory for storing instructions executable by the processor;

[0065] The processor is configured to execute the instructions to implement the user churn prediction method as described above.

[0066] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the user churn prediction method as described above.

[0067] According to a fifth aspect of an embodiment of the present invention, a computer program product is provided, including a computer program or instructions, which, when executed by a processor of an electronic device, implements the user churn prediction method as described above.

[0068] The technical solutions provided by the embodiments of the present invention bring at least the following beneficial effects:

[0069] The embodiment of the present invention obtains the basic attribute data and time-series behavior data of the user; determines the high-order spatial feature information in the basic attribute data, and determines the high-order time-domain feature information in the time-series behavior data; fuses the high-order spatial feature information and the high-order time-domain feature information to obtain a fused feature set; and predicts the probability of the user leaving the network based on the fused feature set. In other words, the embodiment of the present invention comprehensively considers the user's time-series behavior data and basic attribute data, and fuses the extracted high-order spatiotemporal features, thereby achieving multi-dimensional analysis and high-precision prediction of the user's off-network behavior, and improving the accuracy and efficiency of user off-network prediction.

[0070] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present invention, and together with the specification are used to explain the principles of the present invention, and do not constitute an undue limitation of the present invention. In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0072] Figure 1 This is a flowchart of a method for predicting user churn provided by an embodiment of the present invention.

[0073] Figure 2 This is a flowchart of determining high-order spatial feature information provided by an embodiment of the present invention.

[0074] Figure 3 This is a schematic diagram of the architecture of a first convolutional neural network provided by an embodiment of the present invention.

[0075] Figure 4 This is a flowchart of determining high-order time-domain feature information provided by an embodiment of the present invention.

[0076] Figure 5This is a schematic diagram of the architecture of a second convolutional neural network provided by an embodiment of the present invention.

[0077] Figure 6 This is a schematic diagram of feature cascade fusion provided by an embodiment of the present invention.

[0078] Figure 7 This is a schematic diagram of a user churn prediction architecture provided by an embodiment of the present invention.

[0079] Figure 8 This is a block diagram of a user churn prediction device provided by an embodiment of the present invention.

[0080] Figure 9 This is a block diagram of an electronic device provided by an embodiment of the present invention.

[0081] Figure 10 This is a block diagram of a device for predicting user churn provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0082] In order to enable ordinary persons in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0083] It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in sequences other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0084] Technical terms:

[0085] Spatial Feature Mining: This approach uses a deep learning-based approach to mine potential spatial features in the data by designing a network structure that combines a convolutional neural network with an attention mechanism. The introduction of the attention mechanism enables the model to adaptively mine high-order spatial features that have a greater impact on prediction results.

[0086] Time series feature mining: GM also adopts a deep learning-based method. By designing a network structure that combines convolutional neural networks and an attention mechanism, it mines the potential time series features in the data. By introducing the attention mechanism, the model can adaptively mine high-order spatial domain features that have a greater impact on the prediction results.

[0087] On the basis of understanding the above technical terms, please refer to the following embodiments.

[0088] See also Figure 1 , is a flow chart of a method for predicting user churn provided by an embodiment of the present invention, such as Figure 1 As shown, the method includes the following steps:

[0089] Step 101: Obtain the user's basic attribute data and time series behavior data.

[0090] Step 102: Determine high-order spatial domain feature information in the basic attribute data, and determine high-order temporal domain feature information in the temporal behavior data.

[0091] Step 103: Fusing the high-order spatial feature information and the high-order temporal feature information to obtain a fused feature set.

[0092] Step 104: Based on the fused feature set, predict the probability of the user leaving the network.

[0093] The user churn prediction method described in the present invention can be applied to terminals, servers, etc., without limitation. The terminal implementation device can be an electronic device such as a smart phone, a laptop computer, a tablet computer, a desktop computer, a personal digital assistant (PDA) and a wearable device. The server can be an independent server or a server cluster, or a server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, intermediate services, domain name services, security services, content distribution networks, or a big data and artificial intelligence platform, etc., without limitation.

[0094] The following combination Figure 1 , the specific implementation steps of a user churn prediction method provided by an embodiment of the present invention are described in detail.

[0095] In step 101, basic attribute data and time series behavior data of a user are obtained.

[0096] In this step, there are many ways to obtain the user's basic attribute data and time-series behavior data. They can be obtained locally or remotely, and this embodiment does not limit this. Specifically, they include:

[0097] 1) Get user data within a set time period.

[0098] In this embodiment, historical data within a user-set time period can be collected locally or remotely from various data sources (such as various data sources of telecommunications operators, etc.). The historical data includes but is not limited to: the user's call records, text message records, data usage, billing information, social media activities, etc., providing basic data for subsequent prediction of user off-network behavior, that is, for subsequent determination of whether the user is likely to leave the network.

[0099] The set time period may be pre-set as needed, for example, one month, three months, six months, one year, or a longer period, etc. This embodiment does not impose any limitation thereto.

[0100] 2) Preprocessing the user data to obtain preprocessed user data.

[0101] In this step, the acquired user data is preprocessed, where the preprocessing may include:

[0102] 21) Cleaning and preparation: Clean the collected raw data to remove erroneous data and outliers. For example, for data points with missing features, perform feature missing value processing; select different filling methods for different data types, including: mean filling, mode filling, zero-value filling, or other appropriate methods; for example, for numerical data, use mean or median filling; for categorical data, use mode filling. In addition, it can also include deleting invalid features: that is, deleting those features that are irrelevant to the analysis or obviously invalid, that is, removing data features that are invalid or irrelevant to user churn prediction, such as the user's temporary number.

[0103] 22) Feature encoding:

[0104] In this embodiment, feature encoding can use various encoding techniques, such as categorical encoding, one-hot encoding, or frequency encoding, to convert data into numerical features that can be processed by machine learning algorithms. In other words, categorical data is encoded to convert it into a format that can be processed by machine learning algorithms. For example, for the feature "gender," one-hot encoding can be used to convert it into a binary feature.

[0105] 3) Dividing the pre-processed user data into basic attribute data and time series behavior data.

[0106] In this step, the preprocessed user data is divided into time-series behavior data and basic attribute data. Time-series behavior data refers to a dynamic dataset that records a user's communication behavior over a period of time, including but not limited to call frequency, data usage, and SMS frequency. This data reflects the time series and changing trends of a user's communication behavior. In contrast, basic attribute data includes relatively stable user characteristics over a relatively long period of time, such as basic user information (such as age, gender, occupation, etc.), consumption habits, and account type. This data typically changes little over time and can reflect a user's basic attributes and long-term behavior patterns.

[0107] Part of the basic attribute data obtained after the division is shown in Table 1, and part of the temporal behavior data obtained is shown in Table 2. Tables 1 and 2 are only examples and are not limited to these in actual applications.

[0108] Table 1

[0109]

[0110] Table 2

[0111]

[0112]

[0113] In step 102, high-order spatial domain feature information in the basic attribute data is determined, and high-order temporal domain feature information in the temporal behavior data is determined.

[0114] In this step, the high-order spatial feature information in the basic attribute data and the high-order temporal feature information in the temporal behavior data are determined. In practical applications, they can be performed simultaneously without prioritizing one order. For example, the high-order spatial feature information can be determined first, followed by the high-order temporal feature information, or vice versa. This embodiment takes the example of first determining the high-order spatial feature information and then determining the high-order temporal feature information, specifically including:

[0115] 1) Determine high-order spatial feature information in the basic attribute data.

[0116] This step includes two parts: selecting spatial features from basic attribute data and mining high-order spatial features. The flowchart is as follows: Figure 2 As shown, Figure 2 A flowchart for determining high-order spatial feature information provided by an embodiment of the present invention includes:

[0117] Step 201: Select spatial features related to the user's off-network behavior from the user's basic attribute data.

[0118] In this step, basic user attribute data is collected and organized. A random forest algorithm (other similar algorithms are also possible) is used to filter out spatial features related to user churn from this basic attribute data. This basic user attribute data may include, but is not limited to, personal information, spending habits, and account type. This data reflects the user's long-term behavioral tendencies and basic attributes.

[0119] The random forest algorithm is used to select spatial features from user attribute data. The number of decision trees in the random forest algorithm is adjusted based on the size of the dataset and the complexity of the features. Through training, a score is obtained for each feature's importance to the prediction results, reflecting its influence in predicting user behavior.

[0120] Among them, the random forest algorithm is used to select features of user basic attribute data. The construction of each decision tree in the random forest algorithm is based on a randomly selected data subset and splits on the random feature subset.

[0121] Step 202: Determine the importance score of each selected spatial feature for predicting user churn behavior;

[0122] In this step, the spatial domain features are obtained by calculating the average impurity reduction of each feature in the random forest decision tree.

[0123] Among them, the impurity reduction Δi(s t ) is calculated using the following formula:

[0124] Δi(s t )=Impurity parent -(Impurity left child -Impurity rightchild )

[0125] The “impurity” in this embodiment is Gini impurity.

[0126] The overall importance score of each spatial feature is Importance(A_F i ) is obtained by accumulating the reduction of its impurity in all decision trees. The specific calculation formula is:

[0127]

[0128] Among them, A_F i represents the i-th spatial feature, Δi(s t ) represents the spatial feature A_F iThe amount of impurity reduction in decision tree t, trees are all decision trees in the random forest, N tress , is the number of decision trees, which can be set to 100 in this embodiment to ensure the stability and accuracy of the model, but is not limited to this in practical applications.

[0129] Step 203: From the importance scores of all spatial features, select the spatial features whose importance scores are higher than the preset spatial threshold as important spatial features.

[0130] In this step, based on the obtained feature importance scores, spatial features with importance scores above a preset spatial threshold are selected. These features are important for predicting user churn and are referred to as important spatial features. The preset spatial threshold can be set to 70% or higher for spatial features to be selected as important. Important spatial features can also be referred to as key features.

[0131] Step 204: Input the important spatial features into the first deep convolutional neural network for convolution processing in the spatial dimension, and use the attention mechanism to adjust the attention weights of the important spatial features to obtain potential high-order spatial feature information.

[0132] In this step, the architecture diagram of the first deep convolutional neural network is as follows Figure 3 As shown, Figure 3 A schematic diagram of the architecture of a first convolutional neural network provided in an embodiment of the present invention; the first convolutional neural network is used to mine high-order spatial features. Among them, the first convolutional neural network uses multiple convolutional layers to process important spatial features of the input (i.e., the spatial features in the figure). In this embodiment, the number of convolutional layers is L layers as an example. The size of the convolution kernel of each layer is 1×k, and k is the number of features that each convolution kernel can cover. In order to ensure that each unit on the feature map can fully capture and fuse all input features after multiple layers of convolution, according to the receptive field theory, the input feature data (N) is processed by the convolution layer, and the convolution kernel size K and the number of layers L of each layer are adjusted as needed. The calculation formula for the minimum number of convolution layers L required is:

[0133]

[0134] Among them, N is the total number of important spatial features of the input, and k represents the size of the convolution kernel in the feature dimension.

[0135] In this embodiment, the number of spatial features N can be 30, the convolution kernel size is 1×3, each layer of convolution expands the receptive field of k-1 units straight forward, and each layer of convolution is not padded, the step size is 1, and according to the above formula, the minimum number of convolution layers L can be 15, but in specific applications, it is not limited to this. This embodiment is only an example.

[0136] It should be noted that this embodiment introduces the attention mechanism based on the first deep convolutional neural network (CNN) architecture. Figure 3 The attention mechanism dynamically adjusts and optimizes the importance weight of the feature (i.e., the attention weight A_α i ) is used to determine the importance score of each spatial feature.

[0137] Among them, the attention mechanism calculates an attention score by evaluating the activation value of each unit on the feature map. This attention score reflects the importance of the spatial feature at a specific location to the prediction task. Specifically, the importance weight of each spatial feature i (i.e., the attention weight Λ_α i ) is calculated by the softmax function, ensuring that the sum of all spatial feature weights is 1. Specifically, the importance weight calculation formula of spatial feature i is:

[0138]

[0139] Among them, e i represents the evaluation score of spatial feature i, which is automatically learned by the first deep convolutional neural network through forward propagation; N represents the total number of spatial features; j is the index of all spatial features, which is used to calculate the denominator of the softmax function to ensure that the sum of all weights is 1, e j represents the evaluation score of any spatial feature j among spatial features 1 to N. In this way, the first deep convolutional neural network can dynamically assign weights based on the contribution of each spatial feature to the prediction target when processing a large number of features, effectively identifying and focusing on the most informative spatial features, and obtaining high-order spatial feature information of potential user behavior.

[0140] Therefore, the first deep convolutional neural network further processes and analyzes the features weighted by the attention mechanism to calculate high-order spatial feature information. The calculation formula for its high-order spatial feature is:

[0141] HAF i =CNNN(A_α i ·A_F i )

[0142] Among them, HAFi is the high-order spatial feature mined, A_Fi is the important spatial feature above the preset spatial threshold, and A_α i is the important weight (i.e., attention weight).

[0143] 2) Determining high-order time-domain feature information in the time series behavior data.

[0144] This step also includes two parts: selecting time domain features from time series behavior data and mining time domain features. The flowchart is as follows: Figure 4 As shown, Figure 4 A flowchart for determining high-order time-domain feature information provided by an embodiment of the present invention includes:

[0145] Step 401: converting the time series behavior data into two-dimensional time series behavior data;

[0146] In this step, time-domain feature selection is performed from the time-series behavior data, which is similar to the process of selecting spatial-domain features from the basic attribute data. Its purpose is to filter out key time-domain features from the time-series behavior data. However, time-series behavior data differs from basic attribute data in that basic attribute data is two-dimensional, while time-series behavior data is three-dimensional. Time-series behavior data can include: users, various communication behavior characteristics, and records of these characteristics over time. Specifically, the third dimension represents the time dimension, such as records of the number of calls per day, the number of text messages sent, and data usage. This three-dimensional data structure provides a detailed record of the specific performance of each user's communication behavior over time.

[0147] In this embodiment of the present invention, in order to adapt the random forest algorithm and effectively perform time-domain feature selection, the three-dimensional time series behavior data must first be converted into a two-dimensional form. During this conversion process, this embodiment proposes the time series change rate (TVR) as a key analysis tool to capture and quantify the changes in time-domain features over time. The calculation formula for TVR is:

[0148]

[0149] Among them, TVR i represents the time series change rate of the i-th feature, x i，t = t represents the value of the i-th feature at time point t. T represents the total length of the time dimension. In this embodiment, the time dimension can be taken as the data of the past 30 days, that is, T = 30. Of course, in actual applications, it is not limited to this.

[0150] Step 402: Selecting time-domain features related to the user's churn behavior from the two-dimensional time-series behavior data;

[0151] In this step, the random forest algorithm (of course, other similar algorithms can also be used, which is not limited in this embodiment) can be used to analyze the converted two-dimensional time series behavior data, and features that have a significant impact on predicting user behavior results can be selected by analyzing the importance scores of the features. To evaluate the importance of the features. The importance score of the feature is based on the performance of the feature in the decision tree, such as the average impurity reduction caused in the decision tree. Specifically, the time domain feature T_Fi is obtained by the same calculation method as the spatial domain feature selection. The specific process is described above and will not be repeated here. In this way, the time series features that are crucial to predicting user behavior changes can be accurately identified.

[0152] Step 403: Determine the importance score of each selected time-domain feature for predicting user behavior.

[0153] In this step, the time domain features are obtained by calculating the average impurity reduction of each feature in the random forest decision tree.

[0154] Among them, the impurity reduction Δi(s t ) is calculated using the following formula:

[0155] Δi(s t )=Impurity parent -(Impurity left child -Impurity rightchild )

[0156] The “impurity” in this embodiment is Gini impurity.

[0157] The overall importance score of each time domain feature T_Fi is Importance(A_F i ) is obtained by accumulating the reduction of its impurity in all decision trees. The specific calculation formula is:

[0158]

[0159] Among them, T_Fi represents the i-th time domain feature, Δi(s t ) represents the reduction of the impurity of the time domain feature T_Fi in the decision tree t, trees are all decision trees in the random forest, N tress is the number of decision trees, which can be set to 100 in this embodiment to ensure the stability and accuracy of the model, but is not limited to this in practical applications.

[0160] Step 404: Selecting a time domain feature whose importance score is higher than a preset time domain threshold from the importance scores of all time domain features to obtain an important time domain feature;

[0161] In this step, based on the obtained time-domain feature importance scores, time-domain features with importance scores higher than a preset time-domain threshold are selected. These features are important features for predicting user churn and are referred to as important time-domain features. The preset time-domain threshold can be set to select time-domain features with an importance score higher than 70% as important time-domain features. For example, time-domain features with an importance score higher than 75% are selected as important time-domain features, also referred to as key features.

[0162] Step 405: Input the important time domain features into the second deep convolutional neural network for convolution processing in the spatial dimension and the temporal dimension, and use the attention mechanism to adjust the attention weight of the important time domain features to obtain potential high-order time domain feature information.

[0163] In this step, the architecture diagram of the second deep convolutional neural network is as follows Figure 5 As shown, Figure 5 A schematic diagram of the architecture of a second convolutional neural network provided in an embodiment of the present invention; the second convolutional neural network is used to mine high-order time domain features.

[0164] Since the time domain features and the spatial domain features have different characteristics, the architecture of the convolutional neural network used to mine the corresponding feature parts is also different. The difference is that the second deep convolutional neural network architecture provided in this embodiment is as follows: Figure 5 As shown in the figure, the number of convolutional layers is still L, and the size of each convolution kernel is 1×k×t, which directly processes time series information, where K is the number of features that each convolution kernel can cover, and t is the number of days that can be covered in the time domain channel. This design allows the second deep convolutional neural network to process the spatial and temporal dimensions of features simultaneously in each convolution operation, effectively capturing the dynamic changes and patterns in time series data. In order to ensure that each unit on the feature map can fully capture and fuse all input features after multiple layers of convolution, the minimum number of convolutional layers L required is calculated based on the receptive field theory, and the calculation formula is:

[0165]

[0166] Among them, N is the total number of important time domain features of the input, T represents the total length of the time dimension, k represents the size of the convolution kernel in the feature dimension, and t represents the size of the convolution kernel in the time dimension.

[0167] In this implementation, each convolution layer only expands the receptive field forward by k-1 or t-1 units, and each convolution layer is not padded. The step size is 1, the total number of input time domain features is 30, the total length of the time dimension is 30 days, and the values ​​of k and t are both 3. Therefore, the minimum number of convolution layers L = 15. However, in specific applications, it is not limited to this, and this embodiment is only for illustration.

[0168] It should be noted that this embodiment also introduces the attention mechanism based on the second deep convolutional neural network (CNN) architecture. Figure 5 The attention mechanism dynamically adjusts and optimizes the importance weight of the feature (i.e., the attention weight T_α i ) is used to determine the importance score of each time-domain feature.

[0169] Among them, the attention mechanism calculates an attention score by evaluating the activation value of each unit on the feature map. This attention score reflects the importance of the feature at a specific position to the prediction task. Specifically, the importance weight of each time domain feature i (i.e., attention weight T_α i ) is calculated by the softmax function, ensuring that the sum of all time domain feature weights is 1. Specifically, the importance weight calculation formula of time domain feature i is:

[0170]

[0171] Among them, e i represents the score obtained by evaluating the time domain feature i, which is automatically learned by the second deep convolutional neural network through forward propagation; N represents the total number of time domain features; j is the index of all time domain features, which is used to calculate the denominator of the softmax function to ensure that the sum of all weights is 1, e j represents the evaluation score of any time-domain feature j among time-domain features 1 to N. In this way, the second deep convolutional neural network can dynamically assign weights based on the contribution of each time-domain feature to the prediction target when processing a large number of features, effectively identifying and focusing on the most informative time-domain features, and obtaining high-order time-domain feature information about potential user behavior.

[0172] Therefore, the second deep convolutional neural network further processes and analyzes the features weighted by the attention mechanism to calculate the high-order temporal features HTFi. The specific calculation formula is:

[0173] HTF i =CNNN(T_α i ·T_F i )

[0174] Among them, HTFi is the high-order time domain feature mined, T_Fi is the important time domain feature above the preset time domain threshold, T_α i is the important weight (i.e., attention weight).

[0175] In this example, in terms of temporal and spatial information processing, machine learning techniques, such as the random forest algorithm, are first used to perform preliminary feature screening on user data, effectively identifying key features closely related to user churn. Subsequently, deep learning techniques, combining a convolutional neural network with an attention mechanism, are used to further mine these screened features, thereby obtaining high-level characteristic information about user behavior. This process significantly improves the accuracy and sophistication of the model's predictions of user churn, enabling the model to better understand and analyze user behavior patterns.

[0176] In step 103, the high-order spatial domain feature information and the high-order temporal domain feature information are fused to obtain a fused feature set.

[0177] In this step, the high-order spatial feature information and the high-order temporal feature information are subjected to a cascade fusion operation along the channel axis to obtain a fusion feature set in which the high-order spatial feature and the high-order temporal feature are combined into a feature vector. The schematic diagram of cascade fusion is shown in FIG. Figure 6 As shown, Figure 6 A schematic diagram of feature cascade fusion according to an embodiment of the present invention.

[0178] like Figure 6 As shown, the high-order spatial features and the high-order temporal features are input into the cascade layer for cascading (ie, fusion), that is, the high-order spatial features HAFi and the high-order temporal features HTFi are cascaded to form a fused feature set. The fusion operation can be expressed by the following formula:

[0179] Fusion feature set = concat(HAF·HTF)

[0180] Among them, concat represents the cascade operation of features along the channel axis, merging the spatial domain features and the time domain features into a unified feature vector.

[0181] In step 104, the probability of the user leaving the network is predicted based on the fused feature set.

[0182] In this step, first, the fused feature set is weighted using the attention mechanism to obtain a weighted fused feature set; then, the weighted fused feature set is predicted using a trained user churn prediction model to obtain the probability of the user churn, wherein the user churn prediction model is obtained by training the weighted fused feature set using the binary cross entropy loss function of a deep learning convolutional neural network.

[0183] That is, in the embodiment of the present invention, first, the fusion feature set is input into an attention mechanism layer to automatically identify and enhance the features that have a greater impact on the prediction results, as follows: Figure 7As shown, Figure 7 Schematic diagram of a user off-network prediction architecture provided by an embodiment of the present invention. Figure 7 As shown in the figure, the fused feature set is input into the attention mechanism layer, which assigns a weight to each fused feature in the fused feature set. This weight is obtained through learning and represents the importance of the feature in off-grid prediction. The weight calculation formula still follows the aforementioned softmax function:

[0184]

[0185] Among them, M represents the total number of features in the fusion feature set, e i It represents the evaluation score of fusion feature i, which is automatically learned by the network through forward propagation, and represents the contribution score of fusion feature i to the prediction task. j is the index of traversing all fusion features, and e j Represents the evaluation score of any fusion feature j among fusion features 1 to M.

[0186] Afterwards, the weight of each feature is used to weight each fusion feature in the fusion feature set to strengthen the features that have a greater impact on the prediction results. The weighted feature set after weighting is expressed as:

[0187]

[0188] Among them, β i is the weight of feature i obtained by the softmax function. As mentioned above, it is used to strengthen the features that are more important to the prediction task. M represents the total number of features in the fusion feature set.

[0189] Secondly, the fusion feature set is weighted and sent to a trained deep neural network (i.e., a trained user behavior prediction model). The deep neural network can contain multiple hidden layers. Figure 7 Taking two convolutional layers as an example, the kernel size of each convolutional layer is 1×K, where K is the number of features each kernel can cover. The convolution results are then fed into a global max pooling layer for pooling. This pooling result is then passed through a fully connected layer, which integrates the features extracted by the convolutional and pooling layers to generate the final classification output (i.e., the predicted probability of user churn). Therefore, at this stage, the network not only processes the linear relationships between features but also explores their nonlinear connections, further mining and leveraging these features to predict user churn.

[0190] Finally, the deep neural network outputs the probability of the user leaving the network.

[0191] If the probability of the user leaving the network is greater than 0.5, it is predicted that the possibility of the user leaving the network is relatively high; otherwise, it is predicted that the possibility of the user leaving the network is relatively low.

[0192] In this embodiment, in order to adjust and optimize the deep neural network, the following objective function is used for training to obtain a trained user prediction model:

[0193] The fused feature set is used to build a user prediction model based on a deep neural network. The neural network includes two 1×k convolutional layers and a global maximum pooling layer, followed by a dense layer. The binary cross entropy loss function is used for network training to optimize the model parameters:

[0194]

[0195] Among them, L(θ) is the loss value of the binary cross entropy loss function, θ represents the parameters of the network model, M is the number of training samples, and y is the actual user off-network status. The churn prediction made by the user churn prediction model.

[0196] Among them, the user churn prediction model can be pre-trained in the following manner: obtaining a training sample set, the training sample set including: a fused feature set based on the spatial features of the user's basic attribute data and the temporal features of the temporal behavior data, and the actual user churn probability; inputting the fused feature set into the user churn prediction model based on the convolutional neural network for training; during the training process, comparing the user churn probability output by the user churn prediction model with the actual user churn probability, calculating the difference between the comparisons, and using the difference as the loss value of the binary cross entropy loss function; based on the loss value, adjusting the parameters of the user churn prediction model through the back propagation mechanism until convergence, thereby obtaining a trained user churn prediction model.

[0197] In this embodiment of the present invention, a trained deep neural network (i.e., a user churn prediction model) is used to predict user data and output the probability of user churn:

[0198]

[0199] Where σ represents the sigmoid function, which is used to compress the output to between 0 and 1. Represents the probability of disconnection. When , it is predicted that the user will leave the network.

[0200] The embodiment of the present invention obtains the basic attribute data and time-series behavior data of the user; determines the high-order spatial feature information in the basic attribute data, and determines the high-order time-domain feature information in the time-series behavior data; fuses the high-order spatial feature information and the high-order time-domain feature information to obtain a fused feature set; and predicts the probability of the user leaving the network based on the fused feature set. In other words, the embodiment of the present invention comprehensively considers the user's time-series behavior data and basic attribute data, and fuses the extracted high-order spatiotemporal features, thereby achieving multi-dimensional analysis and high-precision prediction of the user's off-network behavior, and improving the accuracy and efficiency of user off-network prediction.

[0201] Furthermore, the embodiment of the present invention combines a deep convolutional network with an attention mechanism to achieve multi-dimensional analysis and high-precision prediction of user off-line behavior. In the mining of time domain information, the embodiment of the present invention introduces an attention mechanism, aiming to accurately identify the time point that has the greatest impact on the user's off-line decision. In the mining of spatial domain information, this embodiment uses deep convolutional neural network technology to deeply explore users' personal characteristics and behavior patterns, thereby providing a more comprehensive perspective on user behavior. This method of combining time and space domain information not only effectively overcomes the limitations of traditional methods in dynamic behavior recognition and multi-dimensional feature analysis, but also significantly improves the ability to identify key time points by introducing an attention mechanism, thereby improving the accuracy and adaptability of the prediction model, and better adapting to the complexity and variability of user behavior in the 5G era and the number portability policy.

[0202] In this embodiment of the present invention, in terms of temporal and spatial information processing, machine learning techniques, such as the random forest algorithm, are first used to perform preliminary feature screening on user data, effectively identifying key features closely related to user churn. Subsequently, deep learning techniques, combining a convolutional neural network with an attention mechanism, are used to further mine these screened features, thereby obtaining high-level characteristic information about user behavior. This process significantly improves the accuracy and sophistication of the model's predictions of user churn, enabling the model to better understand and analyze user behavior patterns.

[0203] Next, by fusing high-level features from the temporal and spatial domains and jointly learning and further mining the fused features, this step not only optimizes the interaction between features but also improves the model's ability to understand complex user behavior patterns, thereby more accurately predicting the probability of user churn.

[0204] In the embodiment of the present invention, by innovatively integrating spatiotemporal features and combining deep learning with the attention mechanism, not only the limitations of traditional churn prediction methods in dynamic behavior recognition and multi-dimensional feature analysis are overcome, but also the accuracy and adaptability of the user churn prediction model are significantly improved. Therefore, the embodiment of the present invention provides an efficient and accurate user churn prediction tool, which helps to reduce user churn and improve user satisfaction and corporate competitiveness.

[0205] It should be noted that for the method embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that this disclosure is not limited by the order of the actions described, because according to the present invention, 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 this specification are all preferred embodiments, and the actions involved are not necessarily required by the present invention.

[0206] Figure 8 This is a block diagram of a user off-grid prediction device provided by an embodiment of the present invention. The device includes: an acquisition module 801, a first determination module 802, a second determination module 803, a fusion module 804 and a prediction module 805, wherein:

[0207] Acquisition module 801, used to acquire basic attribute data and time series behavior data of users;

[0208] A first determining module 802 is configured to determine high-order spatial feature information in the basic attribute data;

[0209] A second determining module 803 is configured to determine high-order time-domain feature information in the time-series behavior data;

[0210] A fusion module 804 is configured to fuse the high-order spatial feature information and the high-order temporal feature information to obtain a fused feature set;

[0211] The prediction module 805 is configured to predict the probability of a user leaving the network based on the fused feature set.

[0212] Optionally, in another embodiment, based on the above embodiment, the acquisition module includes:

[0213] The data acquisition module is used to obtain the user data of the user within a set time period;

[0214] A preprocessing module, configured to preprocess the user data to obtain preprocessed user data;

[0215] The division module is used to divide the pre-processed user data into basic attribute data and time series behavior data.

[0216] Optionally, in another embodiment, based on the above embodiment, the first determining module includes:

[0217] A first spatial feature selection module is configured to select spatial features related to the user's off-grid behavior from the user's basic attribute data using a random forest algorithm;

[0218] a spatial feature score determination module, configured to determine an importance score of each spatial feature selected by the first selection module for predicting user churn;

[0219] The second spatial feature selection module is used to select the spatial features with an importance score higher than a preset spatial threshold from the importance scores of all spatial features as important spatial features;

[0220] The spatial feature processing module is used to input the important spatial features into the first deep convolutional neural network for convolution processing in the spatial dimension, and use the attention mechanism to adjust the attention weights of the important spatial features to obtain potential high-order spatial feature information.

[0221] Optionally, in another embodiment, based on the above embodiment, the second determining module includes:

[0222] a conversion module, configured to convert the time series behavior data into two-dimensional time series behavior data;

[0223] A first time-domain feature selection module is configured to select time-domain features related to the user's off-grid behavior from the two-dimensional time-series behavior data using a random forest algorithm;

[0224] A time domain feature score determination module, used to determine the importance score of each selected time domain feature for predicting user behavior;

[0225] The second time domain feature selection module is used to select the time domain features whose importance scores are higher than the preset time domain threshold from the importance scores of all time domain features to obtain important time domain features;

[0226] The time domain feature processing module is used to input the important time domain features into the second deep convolutional neural network for convolution processing in the spatial dimension and the temporal dimension, and use the attention mechanism to adjust the attention weight of the important time domain features to obtain potential high-order time domain feature information.

[0227] Optionally, in another embodiment, based on the above embodiment, the fusion module is specifically used to perform a cascade fusion operation on the high-order spatial feature information and the high-order time domain feature information along the channel axis to obtain a fusion feature set in which the high-order spatial feature and the high-order time domain feature are merged into a feature vector.

[0228] Optionally, in another embodiment, based on the above embodiment, the prediction module includes:

[0229] A weighted processing module, configured to perform weighted processing on the fused feature set using an attention mechanism to obtain a weighted fused feature set;

[0230] The churn probability prediction module is used to use the trained user churn prediction model to predict the weighted fused feature set to obtain the probability of the user churn, wherein the user churn prediction model is obtained by training the weighted fused feature set using the binary cross entropy loss function of the deep learning convolutional neural network.

[0231] Optionally, in another embodiment, based on the above embodiment, the apparatus further comprises: a training module configured to pre-train the user churn prediction model in the following manner:

[0232] A training sample acquisition module is used to acquire a training sample set, wherein the training sample set includes: a fusion feature set based on the spatial domain features of the user's basic attribute data and the temporal domain features of the temporal behavior data, and an actual user churn probability;

[0233] A model training module is used to input the fused feature set into a user churn prediction model based on a convolutional neural network for training. During the training process, the user churn probability output by the user churn prediction model is compared with the actual user churn probability, the difference between the comparisons is calculated, and the difference is used as the loss value of the binary cross entropy loss function. Based on the loss value, the parameters of the user churn prediction model are adjusted through a backpropagation mechanism until convergence, thereby obtaining a trained user churn prediction model.

[0234] Optionally, an embodiment of the present invention further provides an electronic device, including:

[0235] processor;

[0236] a memory for storing instructions executable by the processor;

[0237] The processor is configured to execute the instructions to implement the user churn prediction method as described above.

[0238] Optionally, an embodiment of the present invention further provides a computer-readable storage medium, which, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to execute the user churn prediction method as described above.

[0239] Optionally, an embodiment of the present invention further provides a computer program product, including a computer program or instructions, which implements the above-mentioned user churn prediction method when executed by a processor of an electronic device.

[0240] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0241] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0242] Figure 9 This is a block diagram of an electronic device 900 provided in an embodiment of the present invention. For example, electronic device 900 can be a mobile terminal or a server. The present invention uses a mobile terminal as an example for description. For example, electronic device 900 can be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0243] Reference Figure 9 , the electronic device 900 may include one or more of the following components: a processing component 902 , a memory 904 , a power component 906 , a multimedia component 908 , an audio component 910 , an input / output (I / O) interface 912 , a sensor component 914 , and a communication component 916 .

[0244] The processing component 902 generally controls the overall operation of the electronic device 900, such as operations associated with display, phone calls, data communications, camera operation, and recording operations. The processing component 902 may include one or more processors 920 to execute instructions to perform all or part of the steps of the above-described method. In addition, the processing component 902 may include one or more modules to facilitate interaction between the processing component 902 and other components. For example, the processing component 902 may include a multimedia module to facilitate interaction between the multimedia component 908 and the processing component 902.

[0245] The memory 904 is configured to store various types of data to support operations on the device 900. Examples of such data include instructions for any application or method operating on the electronic device 900, contact data, phone book data, messages, pictures, videos, etc. The memory 904 can be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0246] The power supply component 906 provides power to the various components of the electronic device 900. The power supply component 906 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device 900.

[0247] The multimedia component 908 includes a screen that provides an output interface between the electronic device 900 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, slides, and gestures on the touch panel. The touch sensor can not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 908 includes a front camera and / or a rear camera. When the device 900 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera can receive external multimedia data. Each front camera and rear camera can be a fixed optical lens system or have focal length and optical zoom capabilities.

[0248] The audio component 910 is configured to output and / or input audio signals. For example, the audio component 910 includes a microphone (MIC), and when the electronic device 900 is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 904 or transmitted via the communication component 916. In some embodiments, the audio component 910 also includes a speaker for outputting audio signals.

[0249] I / O interface 912 provides an interface between processing component 902 and peripheral interface modules, such as a keyboard, click wheel, buttons, etc. These buttons may include but are not limited to: a home button, volume buttons, a start button, and a lock button.

[0250] The sensor assembly 914 includes one or more sensors for providing various aspects of status assessment for the electronic device 900. For example, the sensor assembly 914 can detect the open / closed state of the device 900, the relative positioning of components, such as the display and keypad of the electronic device 900. The sensor assembly 914 can also detect changes in the position of the electronic device 900 or a component of the electronic device 900, the presence or absence of user contact with the electronic device 900, the orientation or acceleration / deceleration of the electronic device 900, and temperature changes of the electronic device 900. The sensor assembly 914 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly 914 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 914 may also include an accelerometer, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0251] The communication component 916 is configured to facilitate wired or wireless communication between the electronic device 900 and other devices. The electronic device 900 can access a wireless network based on a communication standard, such as WiFi, an operator network (such as 2G, 3G, 4G or 5G), or a combination thereof. In an exemplary embodiment, the communication component 916 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 916 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0252] In an embodiment, the electronic device 900 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the user churn prediction method shown above.

[0253] In an embodiment, a computer-readable storage medium is further provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device 900 can perform the user churn prediction method shown above. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0254] In an embodiment, a computer program product is further provided, including a computer program or instructions. When the computer program or instructions are executed by the processor 920 of the electronic device 900, the electronic device 900 executes the above-mentioned user churn prediction method.

[0255] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0256] Figure 10 1 is a block diagram of a device 1000 for predicting user churn provided by an embodiment of the present invention. For example, the device 1000 may be provided as a server. Figure 10 The apparatus 1000 includes a processing component 1022, which further includes one or more processors, and memory resources represented by a memory 1032 for storing instructions, such as applications, that can be executed by the processing component 1022. The application stored in the memory 1032 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1022 is configured to execute the instructions to perform the above-described method.

[0257] The device 1000 may also include a power supply component 1026 configured to perform power management of the device 1000, a wired or wireless network interface 1050 configured to connect the device 1000 to a network, and an input / output (I / O) interface 1058. The device 1000 may operate based on an operating system stored in the memory 1032, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or the like.

[0258] The user information (including but not limited to the user's device information, user personal information, etc.) and related data involved in the present invention are all information authorized by the user or authorized by all parties.

[0259] Other embodiments of the present invention will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the invention being indicated by the following claims.

[0260] It should be understood that the present invention is not limited to the exact construction 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 invention is limited only by the appended claims.

Claims

1. A method for predicting user churn, characterized in that: include: Obtain users' basic attribute data and time-series behavior data; Determining high-order spatial domain feature information in the basic attribute data, and determining high-order temporal domain feature information in the temporal behavior data; Fusing the high-order spatial feature information and the high-order temporal feature information to obtain a fused feature set; Based on the fused feature set, the probability of the user leaving the network is predicted.

2. The method for predicting user churn according to claim 1, wherein: The acquisition of the user's basic attribute data and time series behavior data includes: Get user data within a set time period; Preprocessing the user data to obtain preprocessed user data; The pre-processed user data is divided into basic attribute data and time series behavior data.

3. The method for predicting user churn according to claim 1, wherein: The determining of the high-order spatial feature information in the basic attribute data includes: Selecting spatial features related to user off-grid behavior from the basic attribute data of the user; Determine the importance score of each selected spatial feature for predicting user churn behavior; From the importance scores of all airspace features, select the airspace features whose importance scores are higher than the preset airspace threshold as important airspace features; The important spatial features are input into the first deep convolutional neural network for convolution processing in the spatial dimension, and the attention weights of the important spatial features are adjusted using the attention mechanism to obtain potential high-order spatial feature information.

4. The method for predicting user churn according to claim 1, wherein: The determining of high-order time-domain feature information in the time-series behavior data includes: Converting the time series behavior data into two-dimensional time series behavior data; Selecting time domain features related to user off-grid behavior from the two-dimensional time series behavior data; Determine the importance score of each selected temporal feature for predicting user behavior; From the importance scores of all time domain features, select the time domain features whose importance scores are higher than the preset time domain threshold to obtain the important time domain features; The important time domain features are input into the second deep convolutional neural network for convolution processing in the spatial and temporal dimensions, and the attention weights of the important time domain features are adjusted using the attention mechanism to obtain potential high-order time domain feature information.

5. The method for predicting user churn according to claim 1, wherein: The fusing the high-order spatial domain feature information and the high-order temporal domain feature information to obtain a fused feature set includes: The high-order spatial domain feature information and the high-order temporal domain feature information are subjected to a cascade fusion operation along the channel axis to obtain a fusion feature set in which the high-order spatial domain features and the high-order temporal domain features are merged into a feature vector.

6. The method for predicting user churn according to any one of claims 1 to 5, characterized in that: The predicting, based on the fused feature set, the probability of the user leaving the network, includes: Performing weighted processing on the fused feature set using an attention mechanism to obtain a weighted fused feature set; The trained user churn prediction model is used to perform prediction processing on the weighted fused feature set to obtain the probability of the user churn, wherein the user churn prediction model is obtained by training the weighted fused feature set using the binary cross entropy loss function of a deep learning convolutional neural network.

7. The method for predicting user churn according to claim 6, wherein: The method further includes pre-training a user churn prediction model in the following manner: Obtaining a training sample set, the training sample set including: a fusion feature set based on spatial domain features of basic attribute data of users and temporal domain features of temporal behavior data, and an actual user churn probability; The fused feature set is input into a user churn prediction model based on a convolutional neural network for training. During the training process, the user churn probability output by the user churn prediction model is compared with the actual user churn probability, and the difference between the comparisons is calculated. The difference is used as the loss value of the binary cross entropy loss function. Based on the loss value, the parameters of the user churn prediction model are adjusted through a back-propagation mechanism until convergence, thereby obtaining a trained user churn prediction model.

8. A user churn prediction device, characterized in that: include: The acquisition module is used to obtain the user's basic attribute data and time series behavior data; A first determining module is used to determine high-order spatial feature information in the basic attribute data; A second determining module is used to determine high-order time-domain feature information in the time series behavior data; A fusion module, configured to fuse the high-order spatial feature information and the high-order temporal feature information to obtain a fused feature set; The prediction module is used to predict the probability of the user leaving the network based on the fused feature set.

9. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the user off-grid prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the user off-grid prediction method according to any one of claims 1 to 7.

Images