Comprehensive risk early warning decision support system and method based on big data

Abstract

The invention provides a comprehensive risk early warning decision support system and method based on big data, relates to the field of intelligent data analysis, and adopts a semantic analysis technology based on a natural language. Whether a user is lost or not is judged by performing semantic analysis on user behavior data including user login frequency, use duration, use frequency and last login time. In this way, the user loss risk can be predicted more accurately by understanding the user behavior data in a deeper level.

Description

Technical Field

[0001] The present application relates to the field of intelligent data analysis, and more specifically, to a comprehensive risk warning decision support system and method based on big data. Background Art

[0002] User churn behavior refers to a series of behaviors in which users originally continued to use a product, service, or were active on a platform, but then gradually reduced the frequency of use, stopped using it, or even completely stopped interacting with the product, service, or platform. In a highly competitive market environment, user churn means a reduction in market share. In order to more effectively manage user relationships and reduce user churn, it is necessary to conduct user churn behavior risk warnings. The traditional method may simply count the values ​​of user behavior data, such as simply looking at basic indicators such as login frequency and usage time to roughly judge the risk of user churn, without deeply digging into the semantic information hidden behind these data. In addition, in traditional predictions, each behavioral data indicator is often viewed in isolation, such as analyzing the relationship between login frequency and churn, the relationship between usage time and churn, etc., while ignoring the inherent relationship between these behavioral data.

[0003] Therefore, an optimized comprehensive risk early warning decision support solution is needed. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present application provides a comprehensive risk warning decision support system and method based on big data.

[0005] A comprehensive risk early warning decision support system based on big data, comprising:

[0006] A user behavior data collection module is used to obtain the behavior data of the target evaluation user, and the behavior data includes login frequency, usage duration, usage frequency, and the most recent login data;

[0007] A user behavior preprocessing module, used for preprocessing the behavior data of the target evaluation user to obtain preprocessed user behavior data;

[0008] A user behavior analysis module, used to perform user behavior analysis on the preprocessed user behavior data to obtain a user behavior fusion semantic feature vector;

[0009] The churn risk identification module is used to obtain the churn risk identification result of the target evaluation user by fusing the semantic feature vector based on the user behavior.

[0010] A comprehensive risk early warning decision support method based on big data, comprising:

[0011] Obtaining the behavior data of the target evaluation user, the behavior data including login frequency, usage duration, usage frequency, and the most recent login data;

[0012] Preprocessing the behavior data of the target evaluation user to obtain preprocessed user behavior data;

[0013] Performing user behavior analysis on the preprocessed user behavior data to obtain a user behavior fusion semantic feature vector;

[0014] Based on the user behavior fusion semantic feature vector, a churn risk identification result of the target assessment user is obtained.

[0015] This application has significant technical effects due to the adoption of the above technical solutions:

[0016] The comprehensive risk warning decision support system and method based on big data provided by this application adopts semantic analysis technology based on natural language, and determines whether the user will churn by performing semantic analysis on user behavior data including user login frequency, usage duration, usage frequency, and the last login time. In this way, by understanding user behavior data at a deeper level, the risk of user churn can be predicted more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0018] Figure 1 This is a system block diagram of a comprehensive risk warning decision support system based on big data according to an embodiment of the present application.

[0019] Figure 2 It is a block diagram of a user behavior analysis module in a comprehensive risk warning decision support system based on big data according to an embodiment of the present application.

[0020] Figure 3 The present invention is a flowchart of a comprehensive risk warning decision support method based on big data according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0022] As mentioned in the background technology above, user churn behavior refers to a series of behaviors in which users go from continuously using a product, service, or being active on a platform to gradually reducing the frequency of use, eventually stopping using it, or even completely no longer interacting with the product, service, or platform. In a highly competitive market environment, user churn directly leads to a reduction in market share and may have a negative impact on brand reputation and long-term profitability. Therefore, in order to more effectively manage user relationships and improve user retention rates, companies need to conduct risk warnings for user churn behavior.

[0023] Traditional methods usually rely on simple statistical user behavior data, such as basic indicators such as login frequency and usage time, to roughly judge the risk of user churn. However, this method often fails to deeply explore the semantic information hidden behind the data, nor does it identify the complex patterns of user behavior. In addition, traditional prediction methods often analyze each behavioral data indicator in isolation, such as studying the relationship between login frequency and churn or the relationship between usage time and churn separately, ignoring the possible inherent correlation and mutual influence between these behavioral data. Therefore, a comprehensive risk warning decision support solution based on big data is expected.

[0024] In recent years, deep learning and neural networks have been widely used in computer vision, natural language processing, text signal processing and other fields. In addition, deep learning and neural networks have also shown a level close to or even beyond that of humans in image classification, object detection, semantic segmentation, text translation and other fields. The development of deep learning and neural networks provides new solutions and solutions for comprehensive risk warning decision support.

[0025] Figure 1 : is a system block diagram of a comprehensive risk warning decision support system based on big data according to an embodiment of the present application. Figure 1 As shown, in the comprehensive risk warning decision support system 100 based on big data, it includes: a user behavior data acquisition module 110, which is used to obtain the behavior data of the target evaluation user, and the behavior data includes login frequency, usage time, usage frequency, and the most recent login data; a user behavior preprocessing module 120, which is used to preprocess the behavior data of the target evaluation user to obtain preprocessed user behavior data; a user behavior analysis module 130, which is used to perform user behavior analysis on the preprocessed user behavior data to obtain a user behavior fusion semantic feature vector; a churn risk identification module 140, which is used to obtain the churn risk identification result of the target evaluation user based on the user behavior fusion semantic feature vector.

[0026] In an embodiment of the present application, the user behavior data acquisition module 110 is used to obtain the behavior data of the target evaluation user, and the behavior data includes login frequency, usage time, usage frequency, and the most recent login data. It should be understood that the login frequency refers to the number of times a user logs in to a related product or platform within a certain period of time. For example, how many times a user has logged in in the past month, or the average number of logins per week, etc. This data can intuitively reflect the frequency of user contact with the product. A large number of logins means that the user actively uses the product more frequently, while conversely, it may indicate that the user is not very motivated to use the product or the attention to the product is decreasing. The usage time refers to the cumulative length of time that the user continues to use the product or platform each time he logs in. It reflects the time resources that the user actually invests in the use of the product. A longer usage time often implies that the user has a certain interest and demand in product functions, content, etc., and the continuous shortening of the usage time may indicate that the user's dependence on the product is decreasing. The usage frequency mainly focuses on describing the frequency of users using specific functional modules of the product and performing specific operations within a specific time range. It can show the user's active participation in different functions of the product. A high frequency of use means that the user is actively engaging in various interactive behaviors in the product, while a decrease in frequency may mean that the user's willingness to use the product is weakening. The data of the last login time records the specific time point when the user last logged into the product or platform from the current time. It can know the latest interaction moment between the user and the product. Not logging in for a long time is often an important signal that the user may be about to churn or has begun to churn. The interaction between users and products is complex and diverse, and a single data indicator is difficult to fully reflect the user's actual usage status and attitude towards the product. The indicators of login frequency, usage duration, usage frequency, and the last login time quantify user behavior from different angles. Taken together, they can form a more three-dimensional and comprehensive portrayal of the user's use of the product, avoiding misjudgments caused by evaluating users based solely on one-sided information.

[0027] In an embodiment of the present application, the user behavior preprocessing module 120 is used to preprocess the behavior data of the target evaluation user to obtain preprocessed user behavior data. It should be understood that in the process of collecting user behavior data, some abnormal values ​​that do not conform to the actual situation or are caused by collection errors, that is, noise data, will inevitably be mixed in. At the same time, sometimes due to technical failures, some user operations are not fully recorded, and other factors, some behavior data will be missing. In addition, in the process of collecting user behavior data for a long time, there will be some duplicate records or redundant information that is of no substantial help to subsequent analysis. In order to improve the data quality of the behavior data of the target evaluation user, in the technical solution of the present application, it is necessary to preprocess the behavior data of the target evaluation user. By removing noise data, filling missing values, unifying formats and other preprocessing operations, the accuracy and reliability of the data can be improved, thereby laying a good data foundation for the analysis of the system.

[0028] The process of preprocessing the behavior data of the target evaluation users to obtain high-quality preprocessed user behavior data mainly involves key links such as outlier processing, missing value processing and data format unification. Each link needs to be carefully designed and implemented to ensure the accuracy, completeness and consistency of the data, thereby providing a reliable data foundation for subsequent data analysis. The following is a detailed description of its implementation process:

[0029] Outlier processing aims to identify and process data points that are obviously deviated from the normal range. These outliers may be caused by system errors, data entry errors, or abnormal user behavior patterns (such as malicious volume brushing, etc.). For login frequency, a reasonable threshold range needs to be set based on the characteristics of the product or platform and the regular usage habits of the user group. Taking social platforms as an example, if the number of active users logging in per week is usually between 5-20 times, then too low a login number (such as less than 1) may be an abnormal rare login situation, such as a system error record or a test account login; while too high a login number (such as more than 50 times) may be a machine brushing or data error. When processing these outliers, obviously wrong data (such as a negative login frequency) should be deleted directly because it is meaningless to the analysis and will seriously interfere with the results. For uncertain outliers, such as login frequencies slightly higher than the normal range, it is necessary to combine other data or manual review and judgment. If it is found that it is caused by reasonable special circumstances such as promotional activities after investigation, it can be marked and retained; if the rationality cannot be determined, it can be corrected to a value within a reasonable range (such as taking the maximum value or average value of the normal range) or deleted according to the data distribution.

[0030] In terms of usage time, the product functions and normal usage scenarios should also be considered to set thresholds. For example, for video playback applications, the normal user's single usage time is generally between 10 minutes and 3 hours. Single usage time less than 1 minute (may be opened by mistake and then closed immediately) or more than 10 hours (may be data recording errors or abnormal situations) can be regarded as abnormal values. The handling method is similar to the login frequency, and the abnormal value will be deleted or corrected according to the obviousness and rationality of the abnormal value.

[0031] The judgment of abnormal values ​​of usage frequency is based on the probability distribution of usage of different functional modules of the product. Taking the e-commerce platform as an example, the normal range of users browsing the product details page every day may be 10-100 times. It is considered abnormal if it is higher than 500 times (may be program abnormality or malicious brushing) or lower than 1 time (may be data missing). The same principles as above are followed when processing to ensure the rationality of the data.

[0032] The outlier judgment of the most recent login time is related to the activity requirements of the product. If the product expects users to log in at least once a week to stay active, then a user who has not logged in for more than 3 months (for products with higher activity requirements) can be considered an outlier. The handling of outliers also needs to comprehensively consider product characteristics and user behavior patterns and adopt appropriate handling methods.

[0033] In the missing value processing phase, the missing value situation must be determined first. Through data statistical analysis, the number and proportion of missing values ​​in each user behavior data field, as well as their distribution pattern (random missing or regular missing) can be clarified. When filling missing values, different methods can be selected according to specific circumstances. The mean / median filling method is suitable for situations where missing values ​​are randomly distributed and the data distribution is relatively stable. For numerical data such as login frequency and usage time, the mean or median of all non-missing values ​​is calculated to fill missing values. However, this method may smooth the true characteristics of the data and should be used with caution when the data distribution is uneven. The model-based filling method uses a machine learning model to predict missing values, such as using other complete user behavior data as independent variables and missing login frequency or usage time as dependent variables to establish a regression model for training, and then use the model to predict missing values. This method is more flexible, but it relies on sufficient complete data to train the model and is affected by the accuracy of the model. The special value filling method is used for data with special meanings. For example, when the last login time is missing, a reasonable default value can be filled according to business logic. For example, for platforms with a membership level system and related to login activity, a relatively reasonable time can be filled according to the user's membership level and previous login patterns.

[0034] Data format unification is also an important part of preprocessing. For date formats, such as the last login time, it is necessary to ensure that all data is in a unified format. In this application, all data is converted into text format for subsequent semantic analysis.

[0035] Through the above comprehensive and detailed preprocessing operations such as outlier processing, missing value processing and data format unification, the quality of the target evaluation user behavior data can be significantly improved.

[0036] In the embodiment of the present application, the user behavior analysis module 130 is used to perform user behavior analysis on the pre-processed user behavior data to obtain a user behavior fusion semantic feature vector. Specifically, Figure 2 FIG. 1 is a block diagram of a user behavior analysis module in a comprehensive risk warning decision support system based on big data according to an embodiment of the present application. Figure 2 As shown, the user behavior analysis module 130 includes: a user behavior semantic feature extraction unit 131, used to perform preliminary semantic feature extraction on the preprocessed user behavior data to obtain a user behavior semantic feature vector; a user behavior association semantic feature extraction unit 132, used to perform semantic association feature extraction on the user behavior semantic feature vector to obtain a user behavior association semantic feature vector; a user behavior semantic feature fusion unit 133, used to perform feature internal consistency alignment based on scale difference elimination calibration on the user behavior semantic feature vector and the user behavior association semantic feature vector to obtain the user behavior fusion semantic feature vector.

[0037] In an embodiment of the present application, the user behavior semantic feature extraction unit 131 is used to perform preliminary semantic feature extraction on the preprocessed user behavior data to obtain a user behavior semantic feature vector. Specifically, in an embodiment of the present application, the user behavior semantic feature extraction unit is used to: segment the preprocessed user behavior data and obtain the user behavior semantic feature vector through a user behavior semantic capturer. It should be understood that although the preprocessed user behavior data has been optimized in terms of quality and format, it is difficult for a computer to directly mine deep semantic information from it in the form of continuous text. In order to split these relatively continuous text data into meaningful word units, so that they are more in line with the basic input requirements for subsequent semantic analysis, and to facilitate further mining of the semantic features hidden therein, it is first necessary to perform word segmentation on the preprocessed user behavior data. After word segmentation, the originally complex behavior data expression is broken down into words. In order to more clearly explore the meaning carried by each word and the semantic relationship reflected after their combination, and then more accurately capture the hidden meaning behind the user behavior, the user behavior data after word segmentation needs to be input into the user behavior semantic capturer. Here, the user behavior semantics capturer is a bidirectional long short-term memory neural network model including a word embedding layer. Those skilled in the art should understand that the word embedding layer is a layer structure that converts the input segmented words into a low-dimensional vector representation that the computer can better understand and process. Its basic principle is to learn the semantics and grammar of words based on a large-scale corpus, and assign a distributed vector representation to each word. The bidirectional long short-term memory neural network is an improved structure developed on the basis of the traditional long short-term memory neural network (LSTM). It consists of two parts: the forward LSTM and the backward LSTM. The forward LSTM reads and processes the word vectors in the order of word input in sequence to capture the semantic information from front to back; the backward LSTM reads the word vectors in reverse order to capture the semantic information from back to front. This bidirectional processing method can make full use of the context information of the entire word sequence, avoid the problem that one-way processing may lose some semantic information, and finally output a hidden state vector containing more comprehensive semantic features by combining the forward and backward results for subsequent analysis and feature extraction. That is, through the processing of the user behavior semantics capturer, the semantic features of user behavior can be effectively extracted to realize user churn risk analysis.

[0038] In the embodiment of the present application, the user behavior association semantic feature extraction unit 132 is used to extract semantic association features from the user behavior semantic feature vector to obtain the user behavior association semantic feature vector. Specifically, in the embodiment of the present application, the user behavior association semantic feature extraction unit is used to: pass the user behavior semantic feature vector through the behavior semantic association capturer to obtain the user behavior association semantic feature vector. It should be understood that although the user behavior semantic feature vector has been obtained through the user behavior semantic capturer, the vector can reflect the semantic information contained in each behavior data, but in actual situations, there are complex internal associations between different behaviors of users, and these associations cannot be fully presented by a single behavior semantic feature. For example, the change in login frequency may have a certain linkage relationship with the change in usage time. It is possible that when the user login frequency decreases, the usage time is also shortened accordingly, and this cross-behavior dimension association information needs to be further mined to fully grasp the overall picture of user behavior and more accurately judge the risk of churn. The semantic feature vector obtained by relying solely on the previous steps is not enough to fully reveal these deep-level associations hidden between behaviors, so it is necessary to process again to mine these potential semantic associations. Specifically, the user behavior semantic feature vector is input into the behavior semantic association capturer. Here, the behavior semantic association capturer is a convolutional neural network model including a first convolution layer and a second convolution layer, wherein the first convolution layer uses a one-dimensional convolution kernel of a first scale, and the second convolution layer uses a one-dimensional convolution kernel of a second scale, wherein the first scale is larger than the second scale. When processing the user behavior semantic feature vector, the one-dimensional convolution kernel of a larger scale (the convolution kernel of the first scale) can cover a wider feature range and capture the correlation of the behavior semantic features within a longer sequence range. The one-dimensional convolution kernel of a smaller scale (the convolution kernel of the second scale) is more focused on capturing local and detailed semantic associations. It can focus on a shorter feature vector interval and discover the connection between those subtle but potentially critical behavior semantics. The combination of these two convolution kernels can give full play to the advantages of both and realize the capture of comprehensive correlation information of the user behavior semantic feature vector from macro to micro, from the whole to the local. In this way, the complex correlation between the user behavior semantics can be more completely and accurately reflected, thereby providing more comprehensive information support for the subsequent judgment of the risk of user churn.

[0039] In an embodiment of the present application, the user behavior semantic feature fusion unit 133 is used to perform characteristic internal consistency alignment based on scale difference elimination calibration on the user behavior semantic feature vector and the user behavior associated semantic feature vector to obtain the user behavior fusion semantic feature vector. It should be understood that the user behavior semantic feature vector focuses on the semantic information mined from each behavior data itself, while the user behavior associated semantic feature vector focuses on revealing the intrinsic association semantics between different behaviors. Considering that the judgment of user churn risk is a complex process, it is affected by multiple factors and the interaction between these factors. Only based on a single semantic feature vector or an associated semantic feature vector to make a judgment, the feature information based on is relatively limited, which may cause omissions or misjudgments. In order to comprehensively utilize the information advantages of these two vectors, realize information complementarity, and thus more comprehensively reflect the user's behavior characteristics, it is necessary to fuse the user behavior semantic feature vector and the user behavior associated semantic feature vector to obtain the user behavior fusion semantic feature vector.

[0040] In particular, it is considered that the user behavior semantic feature vector and the user behavior association semantic feature vector respectively capture user behavior information of different dimensions and granularity. Specifically, the user behavior semantic feature vector mainly focuses on the behavior patterns of individual users, such as login frequency, usage time, etc., while the user behavior association semantic feature vector may involve the correlation between user behaviors, such as time correlation and event correlation between different behaviors. This difference will cause the two feature vectors to have essential differences in numerical range and distribution characteristics, that is, there are different scale ranges. Different scale ranges will destroy the balance of fusion, so that the fused user behavior fusion semantic feature vector cannot well integrate the effective information originally carried by the two, resulting in information loss or imbalance of information weight. Inputting this inaccurate fusion feature into the classifier will affect the accuracy of the final churn risk identification result. Therefore, in the technical solution of the present application, the user behavior semantic feature vector and the user behavior association semantic feature vector are aligned based on the internal consistency of the characteristics of scale difference elimination calibration to obtain the user behavior fusion semantic feature vector.

[0041] Specifically, in an embodiment of the present application, the user behavior semantic feature vector and the user behavior associated semantic feature vector are aligned for internal consistency based on scale difference elimination calibration to obtain a user behavior fused semantic feature vector, including: performing normalization modulation based on vector norming on the user behavior semantic feature vector and the user behavior associated semantic feature vector to obtain a pre-aligned user behavior semantic feature vector and a pre-aligned user behavior associated semantic feature vector; performing intrinsic feature extraction on the pre-aligned user behavior semantic feature vector and the pre-aligned user behavior associated semantic feature vector to obtain a first user behavior semantic principal component feature vector and a second user behavior semantic principal component feature vector; performing intrinsic fine-grained alignment and fusion on the first user behavior semantic principal component feature vector and the second user behavior semantic principal component feature vector to obtain the user behavior fused semantic feature vector.

[0042] In the embodiment of the present application, specifically, another implementation expression of performing characteristic internal consistency alignment based on scale difference elimination calibration on the user behavior semantic feature vector and the user behavior associated semantic feature vector to obtain the user behavior fusion semantic feature vector can be: processing the user behavior semantic feature vector and the user behavior associated semantic feature vector according to the following formula to obtain the user behavior fusion semantic feature vector; wherein the formula is:

[0043]

[0044]

[0045] Among them, V1 represents the user behavior semantic feature vector, V2 represents the user behavior association semantic feature vector, and v 1i represents the feature value of the i-th position of the user behavior semantic feature vector, v 2i represents the eigenvalue of the i-th position of the semantic feature vector associated with user behavior, exp represents the natural exponential function, represents the square of the binary norm of the vector, V1′ represents the pre-aligned user behavior semantic feature vector, V2′ represents the pre-aligned user behavior association semantic feature vector, PCA(·) represents the intrinsic feature extraction, U1 represents the sequence of the first intrinsic decomposition vector, Λ1 represents the first diagonal matrix, λ 11 and λ 1m denote the eigenvalues ​​of the first and mth positions of the first diagonal matrix, respectively. T denotes the transpose of the vector. concat{·;·} denotes vector concatenation. v 11 、v 12 、v 1m Respectively represent the first, second and mth eigenvectors of the sequence of the first user behavior semantic principal component eigenvector, V krepresents the first user behavior semantic principal component eigenvector, U2 represents the sequence of the second eigendecomposition vector, Λ2 represents the second diagonal matrix, λ 21 and λ 2m denote the eigenvalues ​​of the first and mth positions of the second diagonal matrix, respectively, v 21 、v 22 、v 2m Respectively represent the first, second and mth eigenvectors of the sequence of the second user behavior semantic principal component eigenvector, V s represents the semantic principal component feature vector of the second user behavior, represents positional subtraction, ⊙ represents positional multiplication, Indicates positional addition, Conv 1×1 represents point convolution, f vp (V k ,V s ) represents the user behavior fusion semantic feature vector.

[0046] That is, in the technical solution of the present application, the user behavior semantic feature vector and the user behavior associated semantic feature vector are aligned for internal consistency based on scale difference elimination calibration to obtain a user behavior fusion semantic feature vector, which first performs normalization modulation based on vector norming on the user behavior semantic feature vector and the user behavior associated semantic feature vector to obtain a pre-aligned user behavior semantic feature vector and a pre-aligned user behavior associated semantic feature vector. This process is not just for simply adjusting the scale, but an important technology in feature engineering, which aims to solve the problem of inconsistent scales between features in a data set. By calculating the modulus of the feature vector and dividing each vector by its own modulus, it can be ensured that all feature vectors have the same length, thereby eliminating the impact of scale differences between different feature vectors. Normalization processing not only contributes to numerical stability in subsequent processing, but also makes the comparison between feature vectors more fair and reasonable, laying the foundation for subsequent feature alignment. In machine learning models, especially algorithms such as neural networks, the scale of the input features directly affects the speed and direction of weight updates, so normalization can be regarded as a regularization method that helps prevent overfitting and also helps avoid the problem of gradient disappearance or explosion.

[0047] Next, the pre-aligned user behavior semantic feature vector and the pre-aligned user behavior association semantic feature vector are subjected to intrinsic feature extraction to obtain a first user behavior semantic principal component feature vector and a second user behavior semantic principal component feature vector. That is, after the normalization modulation is completed, the pre-aligned user behavior semantic feature vector and the user behavior association semantic feature vector are subjected to intrinsic feature extraction. As an unsupervised linear dimensionality reduction method, intrinsic feature extraction is widely used in data compression and feature extraction. This process determines the principal components by finding the direction of the maximum variance of the data. These principal components are new coordinate axes in the original feature space and can capture the structure of the data with minimal information loss. It is worth mentioning that by selecting the eigenvectors corresponding to the first few largest eigenvalues, the dimension of the feature space can be effectively reduced while maintaining the key information of the data. This dimensionality reduction can not only speed up the model training process, but also help the model generalize better and reduce the impact of noise, thereby improving the performance of the model.

[0048] Finally, the first user behavior semantic principal component feature vector and the second user behavior semantic principal component feature vector are intrinsically aligned and fused to obtain the user behavior fused semantic feature vector. The goal of this step is to find the best matching point or area between the principal component feature vectors of the two target analysis original data to achieve alignment between the two. Fine-grained alignment goes beyond the traditional rough fusion methods such as simple weighted average or splicing, and needs to consider the time series characteristics or spatial distribution characteristics of the feature vector. The fusion after alignment is not a simple numerical operation, but requires the design of a specific mechanism to determine how to combine the information of the two feature vectors. This content-based intelligent fusion method can more effectively utilize multi-source information and improve the representation ability and decision-making quality of the model.

[0049] In an embodiment of the present application, the churn risk identification module 140 is used to obtain the churn risk identification result of the target evaluation user based on the user behavior fusion semantic feature vector. Specifically, in an embodiment of the present application, the churn risk identification module is used to: pass the user behavior fusion semantic feature vector through the user churn identification classifier to obtain the churn risk identification result, and the churn risk identification result is used to indicate whether the target evaluation user has a churn risk. It should be understood that although the user behavior fusion semantic feature vector has integrated rich information such as user behavior semantic features and behavior association semantic features, it is still only a feature representation form at the data level. In order to judge whether the user has a churn risk based on the key information in this vector, it is necessary to input the user behavior fusion semantic feature vector into the user churn identification classifier. After being trained based on a large amount of historical user behavior data and the corresponding churn annotation situation, the user churn identification classifier can automatically learn and discover the intrinsic correlation pattern between these complex features and churn risks, so as to accurately judge the churn risk of the target evaluation user based on the newly input fusion semantic feature vector. The churn risk identification results clearly inform the target assessment user whether there is a churn risk. For users judged to be at risk of churn, the company can formulate targeted user retention strategies based on this, such as providing promotional activities, improving products, etc.

[0050] In summary, the comprehensive risk warning decision support system 100 based on big data according to the embodiment of the present application is explained, which adopts the semantic analysis technology based on natural language, and determines whether the user will churn by performing semantic analysis on the user behavior data including the user login frequency, usage duration, usage frequency, and the last login time. In this way, by understanding the user behavior data at a deeper level, the risk of user churn can be predicted more accurately.

[0051] As described above, the comprehensive risk warning decision support system 100 based on big data according to the embodiment of the present application can be implemented in various terminal devices, such as a server for comprehensive risk warning decision support based on big data. In one example, the comprehensive risk warning decision support system 100 based on big data according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the comprehensive risk warning decision support system 100 based on big data can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the comprehensive risk warning decision support system 100 based on big data can also be one of the many hardware modules of the terminal device.

[0052] Alternatively, in another example, the comprehensive risk warning decision support system 100 based on big data and the terminal device may also be separate devices, and the comprehensive risk warning decision support system 100 based on big data may be connected to the terminal device via a wired and / or wireless network, and transmit interactive information in accordance with an agreed data format.

[0053] Figure 3 FIG. 1 is a flowchart of a comprehensive risk warning decision support method based on big data according to an embodiment of the present application. Figure 3 As shown, in the comprehensive risk warning decision support method based on big data, it includes: S110, obtaining the behavior data of the target evaluation user, the behavior data including login frequency, usage time, usage frequency, and the most recent login data; S120, preprocessing the behavior data of the target evaluation user to obtain preprocessed user behavior data; S130, performing user behavior analysis on the preprocessed user behavior data to obtain a user behavior fusion semantic feature vector; S140, based on the user behavior fusion semantic feature vector, obtaining the churn risk identification result of the target evaluation user.

[0054] Here, those skilled in the art can understand that the specific operations of each step in the above-mentioned comprehensive risk warning decision support method based on big data have been referred to above. Figure 1 to Figure 2 It has been introduced in detail in the description of the comprehensive risk early warning decision support system based on big data, and therefore, its repeated description will be omitted.

[0055] In summary, the comprehensive risk warning decision support method based on big data according to the embodiment of the present application is explained, which uses semantic analysis technology based on natural language to determine whether the user will churn by performing semantic analysis on user behavior data including user login frequency, usage duration, usage frequency, and the last login time. In this way, by understanding user behavior data at a deeper level, the risk of user churn can be predicted more accurately.

Claims

1. A comprehensive risk early warning decision support system based on big data, characterized in that: include: A user behavior data collection module is used to obtain the behavior data of the target evaluation user, and the behavior data includes login frequency, usage duration, usage frequency, and the most recent login data; A user behavior preprocessing module, used to preprocess the behavior data of the target evaluation user to obtain preprocessed user behavior data; A user behavior analysis module, used to perform user behavior analysis on the preprocessed user behavior data to obtain a user behavior fusion semantic feature vector; The churn risk identification module is used to obtain the churn risk identification result of the target evaluation user by fusing the semantic feature vector based on the user behavior.

2. The comprehensive risk early warning decision support system based on big data according to claim 1 is characterized in that: The user behavior analysis module includes: A user behavior semantic feature extraction unit, used for performing preliminary semantic feature extraction on the preprocessed user behavior data to obtain a user behavior semantic feature vector; A user behavior associated semantic feature extraction unit, used for performing semantic associated feature extraction on the user behavior semantic feature vector to obtain a user behavior associated semantic feature vector; The user behavior semantic feature fusion unit is used to perform characteristic internal consistency alignment based on scale difference elimination calibration on the user behavior semantic feature vector and the user behavior associated semantic feature vector to obtain the user behavior fused semantic feature vector.

3. The comprehensive risk early warning decision support system based on big data according to claim 2 is characterized in that: The user behavior semantic feature extraction unit is used to: segment the preprocessed user behavior data and pass it through a user behavior semantic capturer to obtain the user behavior semantic feature vector.

4. The comprehensive risk early warning decision support system based on big data according to claim 3 is characterized in that: The user behavior associated semantic feature extraction unit is used to: pass the user behavior semantic feature vector through a behavior semantic association capturer to obtain the user behavior associated semantic feature vector.

5. The comprehensive risk early warning decision support system based on big data according to claim 4 is characterized in that: The user behavior semantic capturer is a bidirectional long short-term memory neural network model including a word embedding layer, and the behavior semantic association capturer is a convolutional neural network model including a first convolutional layer and a second convolutional layer, wherein the first convolutional layer uses a one-dimensional convolutional kernel of a first scale, and the second convolutional layer uses a one-dimensional convolutional kernel of a second scale, wherein the first scale is larger than the second scale.

6. The comprehensive risk early warning decision support system based on big data according to claim 5 is characterized in that: The user behavior semantic feature fusion unit is used to: Performing normalization modulation based on vector norming on the user behavior semantic feature vector and the user behavior association semantic feature vector to obtain a pre-aligned user behavior semantic feature vector and a pre-aligned user behavior association semantic feature vector; Performing intrinsic feature extraction on the pre-aligned user behavior semantic feature vector and the pre-aligned user behavior associated semantic feature vector to obtain a first user behavior semantic principal component feature vector and a second user behavior semantic principal component feature vector; The first user behavior semantic principal component feature vector and the second user behavior semantic principal component feature vector are intrinsically aligned and fused to obtain the user behavior fused semantic feature vector.

7. The comprehensive risk early warning decision support system based on big data according to claim 6 is characterized in that: The churn risk identification module is used to: pass the user behavior fusion semantic feature vector through the user churn identification classifier to obtain the churn risk identification result, and the churn risk identification result is used to indicate whether the target assessment user has a churn risk.

8. A comprehensive risk warning decision support method based on big data, characterized in that: include: Obtaining the behavior data of the target evaluation user, the behavior data including login frequency, usage duration, usage frequency, and the most recent login data; Preprocessing the behavior data of the target evaluation user to obtain preprocessed user behavior data; Performing user behavior analysis on the preprocessed user behavior data to obtain a user behavior fusion semantic feature vector; Based on the user behavior fusion semantic feature vector, a churn risk identification result of the target assessment user is obtained.

9. The comprehensive risk early warning decision support method based on big data according to claim 8 is characterized in that: Performing user behavior analysis on the preprocessed user behavior data to obtain a user behavior fusion semantic feature vector includes: Performing preliminary semantic feature extraction on the preprocessed user behavior data to obtain a user behavior semantic feature vector; Performing semantic association feature extraction on the user behavior semantic feature vector to obtain a user behavior association semantic feature vector; The user behavior semantic feature vector and the user behavior associated semantic feature vector are aligned based on the scale difference elimination calibration to obtain the user behavior fusion semantic feature vector.

10. The comprehensive risk early warning decision support method based on big data according to claim 9 is characterized in that: Based on the user behavior fusion semantic feature vector, a churn risk identification result of the target evaluation user is obtained, including: passing the user behavior fusion semantic feature vector through a user churn identification classifier to obtain the churn risk identification result, and the churn risk identification result is used to indicate whether the target evaluation user has a churn risk.

Images