Method for analyzing root cause of user portability network and electronic equipment

Abstract

The embodiment of the invention discloses a method for analyzing a root cause of a user carrying network and electronic equipment, and the method comprises the steps: determining a first weight of each event in a personalized service process based on the personalized service process of a to-be-analyzed user and basic weights of a plurality of events, for example, the basic weight, and further, determining a second weight of each event in the personalized service process; according to the method, the basic weight of each event in the personalized service process is corrected, the corrected weight of the event reflects the influence degree of the event on the to-be-analyzed user carrying network, and further, the root cause (or root cause event) of the to-be-analyzed user carrying network is determined based on the corrected weight of each event in the personalized service process. According to the method, the root cause of the carrying network of the user can be analyzed differentially, and the accuracy of root cause identification is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and electronic equipment for analyzing the root cause of user network portability. Background Art

[0002] After experiencing the demographic dividend of the telecommunications industry, operators have completely entered the era of stock competition. As a result, the competition focus of operators has shifted from acquiring new users to how to maintain existing users and improve user satisfaction and loyalty. Currently, operators need to pay more attention to user experience and provide better services to retain users. At the same time, operators also need to pay more attention to data analysis, and understand user needs and behaviors through data analysis, so as to better meet user needs.

[0003] The mobile phone switching is also born in the context of the stock competition. Users can choose operators with better services at any time. Therefore, in order to better retain users, operators have carried out a series of predictions on potential mobile phone switching users. After identifying potential mobile phone switching users, operators need to formulate retention strategies, so the root cause analysis of potential users' mobile phone switching is particularly important.

[0004] However, the current main method for analyzing the root causes of users is to count the characteristics of potential users who switch networks, such as the number of times they exceed their coverage, the number of complaints, etc., and then artificially set some thresholds to roughly determine the root causes of users' network switching. For example, if a user exceeds their coverage more than 4 times a year, it is determined that the root cause of the user's network switching includes exceeding the coverage, and if a user complains more than 5 times a year, it is determined that the root cause of the user's network switching includes exceeding the coverage, etc. This method is simple and crude, and the accuracy of identifying the root cause is poor. Summary of the invention

[0005] The embodiments of the present invention provide a method and electronic device for analyzing the root cause of user network portability, which solve the problem of poor accuracy in analyzing the root cause of user network portability in the prior art.

[0006] In a first aspect, an embodiment of the present invention provides a method for analyzing the root cause of a user's network portability, comprising: determining a first weight of each event in the personalized service history based on a personalized service history of a user to be analyzed and basic weights of multiple events; the personalized service history includes multiple events arranged in a chronological order of occurrence; the basic weight is used to indicate the degree of influence of the event on the user's network portability; the first weight is used to indicate the degree of influence of the event on the user's network portability;

[0007] Modifying the first weight of each event in the personalized service history to obtain a modified weight of each event in the personalized service history;

[0008] Based on the modified weight of each event in the personalized service history, the root cause of the user to be analyzed switching networks is determined.

[0009] Optionally, the method may be executed by an electronic device, which may be a server, a mobile phone, a personal computer, a tablet computer, etc.

[0010] In the present application, based on the personalized service history of the user to be analyzed and the basic weights of multiple events, the first weight of each event in the personalized service history is determined, for example, as its basic weight. Further, the basic weight of each event in the personalized service history is corrected, and the corrected weight of the event reflects the degree of influence of the event on the network portability of the user to be analyzed. Further, based on the corrected weight of each event in the personalized service history, the root cause (or root cause event) of the network portability of the user to be analyzed is determined. This method can differentially analyze the root cause of the user's network portability and improve the accuracy of root cause identification.

[0011] Optionally, the first weight of each event in the personalized service history is corrected to obtain a corrected weight of each event in the personalized service history, including: based on the time interval of each event in the personalized service history, the first weight of each event is corrected to obtain a second weight of each event; the time interval of each event is the interval between the occurrence time of each event and the occurrence time of the most recent event in the personalized service history; the larger the time interval between each event, the smaller the second weight of each event.

[0012] In one implementation, the second weight of the event is the modified weight of the event.

[0013] In an embodiment of the present application, based on the time interval of each event in the personalized service history, the first weight of each event is corrected, taking into account the impact of the time interval between the events on the weight of the events, so that the larger the time interval between the events, the smaller the second weight of the events, which conforms to the changing law of the impact of events on user network portability, so as to improve the accuracy of identifying the root cause of user network portability.

[0014] Optionally, based on the time interval of each event in the personalized service history, the first weight of each event is modified to obtain the second weight of each event, including:

[0015] Establish the decay function of event heat and time interval;

[0016] Determine the event heat of each event based on the decay function and the time interval of each event;

[0017] The basic weight of each event is modified based on the event heat of each event to obtain a second weight of each event.

[0018] Optionally, different events correspond to different attenuation functions to take into account the different changing patterns of event heat of different events, so that the correction weight more accurately reflects the impact of the event on user network portability.

[0019] Optionally, the first weight of each event in the personalized service history is corrected to obtain the corrected weight of each event in the personalized service history, and also includes: based on the interaction between two events in the personalized service history, the second weight of each event is corrected to obtain the corrected weight of each event.

[0020] The embodiment of the present application also takes into account the impact of the interaction between various events on the weight of the events, so as to further improve the accuracy of identifying the root cause of the user's network portability.

[0021] Optionally, the personalized service history includes a first event and a second event that are adjacent to each other, the first event occurs before the second event; the first event belongs to a first event type, and the second event belongs to a second event type; the modified weight W of the second event is * for:

[0022] W * =W+(V a +V b ) / 2*S ab

[0023] Among them, the V a is the weight of the first event type, the V b is the weight of the second event type, and the weight of the event type is used to indicate the impact of the event type on the user's network portability; W is the second weight of the second event, and S ab A coefficient indicating the interaction between the first event type occurring first and the second event type occurring later.

[0024] The method provides a method for calculating the correction weight of an event, which can simply and effectively reflect the influence of the correlation between events on the user's network portability.

[0025] Optionally, the event type includes multiple event types, the multiple event types include the first event type and the second event type, and the weights of the multiple event types are obtained based on a hierarchical analysis model; the hierarchical analysis model includes a target layer, a criterion layer and a solution layer; the target layer is the user's tendency to switch networks, the criterion layer includes multiple first features for judging whether the user has the tendency to switch networks; the solution layer includes the multiple event types.

[0026] In the embodiment of the present application, based on the hierarchical analysis model, the weights of multiple event types are obtained, which can quantitatively analyze the impact of multiple event types on the target layer and accurately reflect the impact of multiple event types on the network portability of the user to be analyzed.

[0027] Optionally, the basic weight of the third event is obtained based on the average value of the weight of at least one second feature corresponding to the third event, and the third event is any event in the personalized service history of the user to be analyzed.

[0028] This embodiment, based on the correspondence between events and features, calculates the basic weight of an event by the weight of at least one feature corresponding to the event, thereby providing a simple and effective method for calculating the basic weight of an event.

[0029] Optionally, the weight of the second feature is obtained based on the Gini coefficient when the second feature is used as a split feature and the number of times the second feature is used as a split feature, which are calculated during training of the ensemble learning classification model; the ensemble learning classification model is trained by a training sample set, and the training sample set includes multiple first-category samples and multiple second-category samples, the first-category samples include multi-domain static event data of users who have switched networks, and the second-category samples include multi-domain static event data of users who have not switched networks, and the multi-domain static event data include multiple second features.

[0030] This embodiment provides a method for calculating feature weights.

[0031] Optionally, the first weight of each event in the personalized service history is corrected to obtain the corrected weight of each event in the personalized service history, including: based on the interaction between two events in the personalized service history, the first weight of each event is corrected to obtain the corrected weight of each event.

[0032] This embodiment takes into account the influence of the interaction between various events on the weight of the events, so as to further improve the accuracy of identifying the root cause of the user's network portability.

[0033] Optionally, the method further includes:

[0034] Obtain multi-domain static event data of each user to be predicted in the set of users to be predicted;

[0035] Based on the multi-domain static event data of each to-be-predicted user in the to-be-predicted user set, predict the network portability probability of each to-be-predicted user;

[0036] Based on the network portability probability, multiple users to be predicted are selected from the set of users to be predicted to form a set of users to be analyzed; the set of users to be analyzed includes the users to be analyzed.

[0037] In the embodiment of the present application, before analyzing the root cause of the user's network porting, the user is screened based on the user's network porting probability, which reduces the number of users to analyze the root cause and improves the calculation efficiency. On the other hand, the user's network porting probability is predicted based on the user's multi-domain static event data, which is more accurate.

[0038] Optionally, the multi-domain static event data includes at least one of network domain data, consumption domain data and service domain data, the consumption domain data is used to characterize the user's consumption habits, the network domain data is used to characterize the network quality corresponding to the user, and the service domain data is used to characterize the user's user perception.

[0039] This embodiment takes into account multi-domain data and can more accurately predict the probability of a user switching networks.

[0040] Optionally, the set of users to be predicted includes a first user to be predicted, and predicting the network portability probability of each user to be predicted based on the multi-domain static event data of each user to be predicted in the set of users to be predicted includes:

[0041] Based on the multi-domain static event data of the first to-be-predicted user, a first probability of the first to-be-predicted user switching networks is predicted.

[0042] Optionally, predicting a first probability of the first to-be-predicted user switching networks based on the multi-domain static event data of the first to-be-predicted user includes:

[0043] Inputting the multi-domain static event data of the first to-be-predicted user into an integrated learning classification model, wherein the integrated learning classification model includes a plurality of decision trees;

[0044] Predicting the first user to be predicted based on the multi-domain static event data of the first user to be predicted by each decision tree in the multiple decision trees, respectively, obtains a prediction result of each decision tree; the prediction result includes users who carry over and users who do not carry over;

[0045] The ratio of the number of users whose prediction results are network porting to the total number of prediction results is used as the first probability of the first to-be-predicted user porting.

[0046] Optionally, the integrated learning classification model is trained by a training sample set, and the training sample set includes multiple first-category samples and multiple second-category samples, the first-category samples include multi-domain static event data of users who have switched networks, and the second-category samples include multi-domain static event data of users who have not switched networks.

[0047] In an embodiment of the present application, the electronic device trains the ensemble learning classification model based on the training sample set, and the trained ensemble learning classification model can more accurately predict the probability of user network portability.

[0048] Optionally, the first probability of the first to-be-predicted user switching networks is a probability of the first to-be-predicted user switching networks.

[0049] Optionally, predicting the network portability probability of each to-be-predicted user in the to-be-predicted user set based on the multi-domain static event data of the to-be-predicted user includes:

[0050] A second probability of the first to-be-predicted user switching networks is determined based on the matching degree between the individual service history of the first to-be-predicted user and the best matching common service history.

[0051] Optionally, the probability of the first to-be-predicted user switching networks is a weighted sum of a first probability of the first to-be-predicted user switching networks and a second probability of the first to-be-predicted user switching networks.

[0052] This method not only considers the user's multi-domain static event data, but also considers the similarity between the user's individual service history and the common service history of the customer group that has switched networks, and is therefore more accurate.

[0053] Optionally, the method further includes:

[0054] Obtaining the individual service history of the first to-be-predicted user and the common service history of multiple customer groups;

[0055] Matching the individual service history of the first to-be-predicted user with the common service history of the multiple customer groups to obtain a common service history that matches the individual service history of the first to-be-predicted user;

[0056] Based on the occurrence time of the matching event in the individual service history of the first user to be predicted, determining the common service history that best matches the first user to be predicted from the matched common service history;

[0057] Among them, the matched common service history includes the first common service history, and the matching events between the personalized service history of the first user to be predicted and the first common service history are events included in both the personalized service history of the first user to be predicted and the first common service history.

[0058] Optionally, the determining, based on the occurrence time of the matching event in the individual service history of the first user to be predicted, from the matched common service history the common service history that best matches the first user to be predicted includes:

[0059] The target value S between the individual service history of the first to-be-predicted user and the first common service history is determined as:

[0060]

[0061] The matching events between the first to-be-predicted user's individual service history and the first common service history include n matching events, and the R n is the last matching event among the n matching events, R i is the i-th matching event among the n matching events, where i and n are positive integers, and i is not greater than n; is the R in the personalized service history of the first user to be predicted n The time of occurrence, is the R in the personalized service history of the first user to be predicted i The time of occurrence;

[0062] The common service history that best matches the first to-be-predicted user is the common service history with the smallest S among the matched common service history.

[0063] Optionally, it is characterized in that the matching degree is the ratio of the number of all events in the common service history of the best match to the number of all events in the individual service history of the first user to be predicted; the matching degree is the second probability.

[0064] In a second aspect, a device for analyzing the root cause of a user's network portability is provided. The device may include: a functional unit for implementing the method in the first aspect or any possible implementation of the first aspect.

[0065] In a third aspect, an electronic device is provided, which may include: a memory; a communication interface; one or more processors;

[0066] Among them, one or more computer programs are stored in the memory, and the one or more computer programs include instructions. When the instructions are executed by the electronic device, the electronic device implements any one of the methods for optionally analyzing the root causes of user network switching in the above-mentioned first aspect.

[0067] In a fourth aspect, a computer-readable storage medium is provided, on which instructions are stored. When the instructions in the computer-readable storage medium are executed by an electronic device, the electronic device is enabled to execute any one of the methods for optionally analyzing the root causes of user network switching in the above-mentioned first aspect.

[0068] In a fifth aspect, a computer program product is provided, which includes computer instructions. When the computer instructions are executed on an electronic device, the electronic device executes any method of optionally analyzing the root cause of a user's network portability as described in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1A This is a system architecture diagram provided by this application.

[0070] Figure 1B It is a schematic diagram of a scenario provided by this application.

[0071] Figure 2A It is a flowchart of a training method for an integrated learning classification model provided in this application.

[0072] Figure 2B The present application provides a schematic diagram of the principles of a training method for an integrated learning classification model.

[0073] Figure 3 It is a flowchart of a method for calculating the basic weight of an event provided in this application.

[0074] Figure 4 This is a flow chart of a method provided by the present application for predicting network porting users and analyzing the root causes of user network porting.

[0075] Figure 5 It is a flowchart of a method for calculating the network portability probability of a user to be predicted provided by the present application.

[0076] Figure 6 It is a flowchart of a method for calculating a second probability provided by the present application.

[0077] Figure 7 It is a schematic diagram of a method for matching the personalized service history of a user to be predicted with the common service history of multiple customer groups provided by the present application.

[0078] Figure 8 This is an example diagram provided by the present application for modifying the weights of events in the personalized service process.

[0079] Fig. 9 It is a flowchart of a method for obtaining the modified weights of events in a user's personalized service history provided by the present application.

[0080] Fig.10 This is a schematic diagram of the relationship between event heat and time interval provided by this application.

[0081] Fig.11It is a flowchart of a method provided by the present application for correcting the weight of events in the personalized service history of a user to be analyzed based on the time interval of the events.

[0082] Fig.12 It is a structural diagram of a hierarchical analysis model provided by this application.

[0083] Fig.13A This is a schematic diagram provided by the present application for visually displaying a user's personalized service history, event weights, network portability probability, and root causes of network portability.

[0084] Fig. 13B This is a schematic diagram of another method provided by the present application for visually displaying a user's personalized service history, the probability of network porting, and the root cause of network porting.

[0085] Fig.14 It is a flowchart of a method for calculating the weights of network type, home broadband type and market type provided by the present application.

[0086] Fig.15 A schematic diagram showing a principle of obtaining a user's behavior sequence provided by an embodiment of the present application is shown.

[0087] Fig.16 A user behavior history diagram provided in an embodiment of the present application is shown.

[0088] Fig.17 A schematic diagram showing a principle of obtaining a user's behavior sequence based on the user's original behavior sequence provided by an embodiment of the present application.

[0089] Fig.18 A schematic diagram showing a principle of obtaining a convergence behavior sequence provided by an embodiment of the present application is shown.

[0090] Fig.19 A schematic diagram showing a principle of determining a user type provided in an embodiment of the present application is shown.

[0091] Fig. 20 It is a structural schematic diagram of an electronic device provided by this application. DETAILED DESCRIPTION

[0092] First, the terms involved in the embodiments of the present application are introduced.

[0093] (1) Multi-domain static event data

[0094] The multi-domain static event data includes one or more of user network domain data, consumption domain data, and service domain data.

[0095] The user's consumption domain data is used to characterize the consumption habits of the target user, the user's network domain data is used to characterize the network quality corresponding to the target user, and the user's service domain data is used to characterize the user's user perception.

[0096] Network domain data (or O domain data (operation support system)) Network domain data may include network perception data and network performance data related to the user, mainly including user network data, such as the resident cell of voice service, the resident cell of data service, network quality data, etc. One or more. Among them, network quality data includes the number of wireless disconnection, slow speed, weak signal strength, and the number of home broadband disconnection, slow speed, weak signal strength, etc. within X months. Among them, slow speed means that the uplink rate or downlink rate is less than the first threshold, such as 2Mbps, and weak signal strength means that the signal strength is less than the second threshold, such as -50dBm.

[0097] Consumption domain data (also called B domain data (business support system)) mainly includes user consumption data, such as obtaining one or more of the user's basic information, online information, package information, package change information, and over-package information. Among them, the user's basic information may include one or more of the user's age, region, gender, etc. The online information includes one or more of the network access time and network access duration of wireless and home broadband (hereinafter referred to as home broadband). Package information includes one or more of wireless package information and home broadband package information. Package change information includes change time, changed wireless package information, and changed home broadband package information. Over-package information includes one or more of the number of over-packages, the number of low-saturation times, and the number of speed limits within X months, where X is a positive integer, which can be 3, 6, 12, etc.

[0098] Service domain data (also called S domain data) mainly includes customer service data, such as one or more of customer service work order recordings, complaint information, complaint information, service evaluation information of data services, service evaluation information of voice services, and home broadband service evaluation. Among them, customer service work order recordings can be used to analyze and obtain at least one of the user's demands, intentions, and content of concern. Complaint information may include at least one of the number of complaints about the market, wireless, home broadband, etc. within X months, and complaint information may include at least one of the number of complaints about the market, wireless, home broadband, etc. within X months. Service evaluation information of data services includes at least one of the star rating, evaluation times, and evaluation time of user evaluation of customer service services related to data services. Service evaluation information of voice services includes at least one of the star rating, evaluation times, and evaluation time of user evaluation of customer service services related to voice services. Home broadband service evaluation includes at least one of the star rating, evaluation times, and evaluation time of user evaluation of customer service services related to home broadband.

[0099] It should be understood that "over-package" means that the data traffic or voice and other services consumed by the user terminal in a certain month exceeds the data traffic or voice in the package subscribed by the user.

[0100] One or more features in the above multi-domain static event data may correspond to an event, for example, the feature "number of times of exceeding the quota" corresponds to the event of exceeding the quota. The feature "number of complaints" corresponds to the event of complaints. The feature "complaints about home broadband" corresponds to the event of complaints.

[0101] (2) Personalized service experience

[0102] The user's personalized service history includes events with time-series characteristics, which are events experienced by the user in the process of carrying out telecommunications services, also known as "user events". Events can represent the state of the user in the telecommunications service at the time corresponding to the event.

[0103] Specifically, the user's personalized service history includes multiple events arranged in the order of occurrence of the events, which are used to record the user's service history. The personalized service history also includes the occurrence time of each event.

[0104] In this application, the "personal service process" is also referred to as a "behavior sequence", and the time when an event occurs is also referred to as the timestamp of the event.

[0105] Exemplarily, events may include network access, complaints, traffic exceeding the limit, mobile secondary card->mobile primary card, average revenue per user (ARPU) increase, ARPU decrease, query portability transcoding, short board label, voice exceeding the limit, package transfer, cancellation of home broadband service, complaints, etc. Complaints may include complaints about customer service, voice, charges, traffic, network quality, etc. Short board labels may include wireless disconnection, wireless slowness, etc. ARPU increase may include high ARPU increase, low ARPU increase, etc. Traffic exceeding the limit may include high traffic exceeding the limit and low traffic exceeding the limit, voice exceeding the limit may include high voice exceeding the limit and low voice exceeding the limit, and complaints may include at least one of complaints about wireless, complaints about voice, complaints about bandwidth, and complaints about tariffs, etc. Not limited to the above events, other user events generated in the process of users using telecommunications services may also be included.

[0106] The user's personalized service history can be obtained based on the user's user data, and the user data can include multi-domain static event data. The method for generating the user's personalized service history (behavior sequence) can refer to the embodiment of the method for generating the personalized service history (behavior sequence) below, which will not be described here.

[0107] (3) Common service experience of customer groups

[0108] Customer groups are the classification of users based on their characteristics. A customer group can also be understood as a user group. Prior knowledge shows that location information, perception type, risk level, age, and gender are important factors that may affect the user's switching / leaving the network. Therefore, based on information entropy gain, the grouping information (or important tags) that affects the user's switching / leaving the network can be identified, and the user groups (or groups) can be distinguished based on this.

[0109] For example, customer groups based on regional classification include users in Guangdong, users in Guangxi, etc. Another example is customer groups based on service type classification, including home broadband users, wireless users, etc. Customer groups based on service content classification include voice users, data users, etc.

[0110] The common service history includes events with time series characteristics, which can specifically include events arranged in chronological order of occurrence. It represents the characteristics contained in the individual service history of users who have switched networks in a customer group and can be used to predict users who have switched networks.

[0111] In the present application, the common service history of a customer group may include events common to the individual service history of users in the customer group who have switched networks and the order in which these events occurred.

[0112] The common service history of a customer group can be generated based on the individual service history of users who have switched networks in the customer group, or based on multi-domain static data of users who have switched networks. For specific implementation, please refer to the following embodiment of the method for generating the common service history of a customer group (common behavior sequence of a user group), which will not be described here.

[0113] (4) Carry-over

[0114] In this application, "carrying and switching networks" refers to the behavior of users changing their contracted telecommunications service operators or leaving the network, which can also be expressed as carrying and switching / leaving the network.

[0115] (5) Users to be predicted

[0116] In this application, "users to be predicted" refer to users who have not yet switched networks.

[0117] (6) Users to be analyzed

[0118] In the present application, "users to be analyzed" are some users selected from multiple users to be predicted, who are users with a relatively high probability of network portability, that is, high-risk users, potential network portability users. The specific implementation of selecting users to be analyzed from multiple users to be predicted can be described in the following embodiments.

[0119] (7) Basic weight of events

[0120] In this application, the "basic weight of an event" is used to characterize the impact / contribution of an event on the user's network portability. Each event has a basic weight. The larger the basic weight of an event, the higher the impact of the event on the user's network portability; the smaller the basic weight of the event, the lower the impact of the event on the user's network portability to be analyzed. The specific implementation of obtaining the basic weight of an event can be found in the relevant description of calculating the basic weight of an event below.

[0121] (8) Modification weights of events in the personalized service process

[0122] In this application, the "corrected weight of events in the personalized service history" is obtained by correcting the basic weight of the event through the application server. It is used to characterize the impact / contribution of events in the personalized service history on the user's network portability. Each event in the personalized service history has a correction weight. The larger the correction weight of the event, the higher the impact of the event on the user's network portability; the smaller the correction weight of the event, the lower the impact of the event on the user's network portability.

[0123] At present, the identification of network-switching users is mainly based on the threshold calculation of a single feature, such as identifying a user as a network-switching user if the user has complained more than 5 times or exceeded the limit more than 10 times. However, a large number of potential network-switching users will be identified based on a single feature, which not only has a low accuracy rate, but also has a high maintenance cost when each identified network-switching user needs to be maintained.

[0124] In order to improve the accuracy of identifying users who are switching networks, machine learning can be used to identify users who are switching networks based on multiple features. For example, an ensemble learning classification model is used, such as random forest (RF) of parallel ensemble learning (bagging) and gradient boosted tree (GBDT) of serial ensemble learning (boosting). Since the ensemble learning method comprehensively considers the impact of multiple features on the user's tendency to switch networks and considers the mutual influence between features, it is more accurate than the single feature threshold method.

[0125] In order to further improve the accuracy of identifying users who have switched networks, the present application also provides a method for screening users who have switched networks. The method can quantitatively characterize the user's tendency to switch networks. It predicts the first probability of the user's switching networks through an integrated learning classification model based on the user's multi-domain static event data, and determines the second probability of the user's switching networks based on the matching degree between the user's individual service history and the common service history of the customer group that has switched networks. Then, the probability of the user's switching networks is comprehensively determined based on the first probability and the second probability. This method not only considers the user's multi-domain static event data, but also considers the similarity between the user's individual service history and the common service history of the customer group that has switched networks, so it is more accurate.

[0126] After identifying potential users who switch networks, it is very important to analyze the root causes of users' network switching, because the accuracy of root cause identification seriously affects the accuracy of the retention strategy for the user. However, users generate a large amount of O-domain, B-domain and S-domain data every day, and the root causes of network switching for each user are different, so root cause mining is very difficult and has a low accuracy rate. The current main way to analyze the root causes of users is to count the characteristics of potential users who switch networks, such as the number of times they exceed the plan limit and the number of complaints, and then manually set some thresholds to roughly determine the root causes of users' network switching. For example, if a user exceeds the plan limit more than 4 times a year, it is determined that the root cause of the user's network switching includes the plan limit, and if the user complains more than 5 times a year, it is determined that the root cause of the user's network switching includes the plan limit, etc. This method is simple and crude, and the accuracy of identifying the root cause is poor.

[0127] In order to further improve the accuracy of identifying the root cause of a user's network portability, the present application also provides a method for analyzing the root cause of a user's network portability, which method determines the first weight of each event in the personalized service history, such as its basic weight, based on the personalized service history of the user to be analyzed and the basic weights of multiple events. Further, the basic weight of each event in the personalized service history is corrected, and the corrected weight of the event reflects the degree of influence of the event on the network portability of the user to be analyzed. Further, based on the corrected weight of each event in the personalized service history, the root cause (or root cause event) of the network portability of the user to be analyzed is determined. This method can differentially analyze the root cause of the user's network portability and improve the accuracy of root cause identification.

[0128] Moreover, the method not only considers the influence of the time of event occurrence on the weight of the event, but also considers the influence of the interaction between various events on the weight of the event, and uses the modified weight to characterize the influence of the event on the user's network portability, thereby improving the accuracy of identifying the root cause of the user's network portability.

[0129] First, the system architecture involved in the embodiments of the present application is described.

[0130] like Figure 1A The following is a schematic diagram of the architecture of a system provided by an embodiment of the present application. Figure 1B The diagram is a schematic diagram of an application scenario provided by an embodiment of the present application. The system includes a user device 11, a big data platform 12, a training device 13, an application server 14 and a client device 15.

[0131] The big data platform 12 can regularly collect multi-domain static event data of users, that is, collect data from its O domain, B domain and S domain, such as network quality data, consumption data, customer service work order recordings, complaint information, complaint information, etc.

[0132] When the user recharges phone bills, uses call duration and traffic, and provides customer service feedback, the user device 11 will send the above-mentioned multi-domain static event data to the big data platform 12. At this time, the big data platform 12 can store this data.

[0133] First-line data analysts can regularly (e.g., every 3 months) access the big data platform 12 and capture the incremental historical data of users and place it in the database of the application server 14, including the following data: multi-domain static event data, the user's personalized service history, and the common service history of multiple customer groups. Among them, the multi-domain static event data mainly includes some inherent attributes of the user and the statistical number of events that occurred within a period of time; the user's personalized service history includes events with time series characteristics; the common service history of customer groups also includes events with time series characteristics, but they are time series characteristic events of users who have switched networks.

[0134] First-line data analysts can also generate samples based on the redundant static event data of users in the big data platform 12. Among them, the samples include the first type of samples (also called negative samples) and the second type of samples (also called positive samples), wherein the first type of samples include the multi-domain static event data of users who have switched from T-3 to T month from T-3-X month to T-3 month, and the second type of samples include the multi-domain static event data of users who have not switched from T-3 to T month from T-3-X month to T-3 month. It should be understood that the above T-3 can also be replaced by T-2 or T-1, etc. T is the time before the current moment.

[0135] Furthermore, the big data platform 12 can send these samples to the training device 13, and the training device 13 can use these samples to train an integrated learning classification model, which is used to predict the probability of a user porting network based on the user's multi-domain static event data. The training device 13 can send the trained integrated learning classification model to the application server 14. The specific implementation of training the integrated learning classification model can refer to the relevant description in the following part of training the integrated learning classification model.

[0136] Application server 14 may include an analytics system and a customer experience system.

[0137] The analysis system can predict the first probability of the user's network portability through an integrated learning classification model based on the user's multi-domain static event data, and determine the second probability of the user's network portability based on the matching degree between the user's individual service history and the common service history of the customer group that has already ported the network, and then comprehensively determine the user's final network portability probability based on the first probability and the second probability. After identifying potential network portability users, the analysis system can also modify the weight of the user's event based on the events in the user's individual service history and the time when the event occurred, and further modify the weight of the user's event based on the interaction between the user's events, obtain the modified weight of the user's event, and select the top K events with the largest weight as the root cause of the user's network portability. For specific implementation, please refer to the relevant description in the following part of analyzing the root cause of the user's network portability.

[0138] Furthermore, the analysis system can send the list of high-risk users, the probability of network switching, and the root cause of network switching to the customer experience system. Then the customer experience system can distribute it to the customer device 15. The customer device 15 can visually display the list of high-risk users, the probability of network switching, and the root cause of network switching, and exemplarily highlight the root cause events in the user's personalized service history. The customer experience system can also generate corresponding retention strategies for the root causes of network switching of high-risk users for reference by customer service personnel. Among them, the list of high-risk users includes users whose probability of network switching is greater than a preset value or includes the top 1000 users with the highest probability of network switching.

[0139] The customer experience system is also used to present the user's consumption tendency and satisfaction with the operator's services to the equipment of the operator's customer experience department or marketing department (ie, the customer service equipment 15).

[0140] Optionally, the application server 14 or the analysis system device can also generate a user's behavior sequence (also called a personalized service history) based on the user data, and generate a common behavior sequence (also called a common service history) for the customer group based on the personalized service history of each user in the customer group, and match corresponding maintenance strategies for the users.

[0141] Optionally, the training device 13 may be the same device as the application server 14. The analysis system and the customer experience system may also be located in different devices.

[0142] Not limited to the above Figure 1A The system architecture shown may include more or fewer devices in other embodiments.

[0143] Training an ensemble learning classification model

[0144] Before introducing the method for predicting network portability users in the present application, the integrated learning classification model used by the method is first introduced.

[0145] Figure 2A The flowchart of the training method of the integrated learning classification model provided by the present invention is exemplarily shown. Figure 2B The schematic diagram of the principle of a training method for an integrated learning classification model provided by the present invention is exemplarily shown. The method can be Figure 1A The method includes but is not limited to some or all of the following steps:

[0146] S11: Obtain a training sample set, the training sample set includes multiple first-category samples and multiple second-category samples, the first-category samples include multi-domain static event data of users who have switched networks, and the second-category samples include multi-domain static event data of users who have not switched networks, wherein the multi-domain static event data includes multiple features.

[0147] S12: Select e samples randomly and with replacement from the training sample set, where e is a positive integer not less than 1.

[0148] S13: For the e samples, randomly select Q features.

[0149] S14: Calculate the Gini coefficients of Q features based on the samples on the current node, where the samples on the root node are the e samples mentioned above.

[0150] For example, Q features can be represented as f1, f2, ..., f i …、f Q .

[0151] At the i-th node, feature f i Taking the calculation of the Gini coefficient as an example, f i The Gini coefficient calculation formula is as follows:

[0152]

[0153] Among them, E i represents the set of samples at the ith node, Gini(E i , f i ) means that at the i-th node, the feature f i The size of the Gini coefficient.

[0154] E il and E ir For E i Based on the feature f i The two sample sets obtained by splitting. l Indicates E il The number of samples in E i The ratio of the number of samples in . r Indicates E ir The number of samples in E iThe ratio of the number of samples in .

[0155] S15: Select the split feature with the largest Gini coefficient as the split feature of the current node.

[0156] S16: Determine whether the conditions for stopping splitting of the decision tree are met. If the conditions for stopping splitting are not met, execute the next step. If the conditions for stopping splitting are met, execute S18. Exemplarily, when the number of samples on the nodes in the current decision tree is less than the preset value, the current node stops splitting. The preset value can be 1, 2 or other values. Otherwise, the conditions for stopping splitting are not met, and execute step S17.

[0157] S17: Split the sample on the current node to the next layer of nodes based on the split feature.

[0158] At the next layer of nodes, the training device continues to execute steps S14-S16.

[0159] S18: Generate a decision tree.

[0160] S19: Repeat S12-S18m times to generate an integrated learning classification model including m decision trees.

[0161] The above-mentioned ensemble learning classification model is explained by taking the random forest model as an example. In other embodiments, other ensemble learning classification models can also be used. Among them, the random forest algorithm is an ensemble learning method based on bagging. The random forest model consists of a series of decision trees and can be used to deal with one or more problems such as prediction problems. Among them, the decision tree can be a classification / regression decision tree (categorization and regression tree, CART). In other embodiments, the decision tree can also be C4.5, iterative binary tree three generations (iteractive dichotomiser 3, ID3), etc.

[0162] In other implementations, the decision tree may also be split based on one or more of information gain, information entropy, etc. of the feature.

[0163] Calculate the base weight of the event

[0164] Before introducing the method for analyzing the root cause of user network portability in the present application, a method for calculating the basic weight of an event is described first.

[0165] Because there is a certain correspondence between the features and events in the multi-domain static event data. For example, the feature "number of times of exceeding the package" in the multi-domain static event data corresponds to the event "exceeding the package"; the feature "number of times of complaining about home broadband" corresponds to the event "complaint". The Gini coefficient of a feature can indicate the impact of the feature on the user's network portability. Therefore, the Gini coefficient of the feature obtained during the training of the integrated classification model can be used to calculate the basic weight of the event. Among them, the basic weight of the event is used to indicate the impact of the event on the general user's network portability.

[0166] Figure 3 The flowchart of the method for calculating the basic weight of an event is shown. The method can be Figure 1A The training device or application server in the embodiment is executed, and the method may include but is not limited to some or all of the following steps:

[0167] S21: Get the Gini coefficient of each feature in the ensemble learning classification model when it is used as a split feature and the number of times each feature is used as a split feature.

[0168] For example, Figure 2B As shown in , the weight of the feature can be obtained by the Gini coefficient of the feature calculated during the training process. Figure 2B The Gini coefficient is shown in . At each split node, the Gini coefficient of each of the Q selected features is calculated. However, the weight of the feature here is calculated using the Gini coefficient of the feature calculated when the feature is used as a split feature. It should be understood that for a decision tree, there are as many Gini coefficients of split features as there are non-leaf nodes. A feature can be used as a split feature once or multiple times.

[0169] S22: For each feature, determine the weight of the feature based on the average value of the Gini coefficient when the feature is used as a split feature.

[0170] For example, when the feature "number of times of over-set" is used as a split feature, the Gini coefficients obtained are 0.2, 0.4 and 0.5 respectively, and the weight of the feature is (0.2+0.4+0.5) / 3.

[0171] Through the above S301 and S302, the weight of each feature can be obtained.

[0172] S23: For each event, determine a basic weight of the event based on the weight of the feature corresponding to the event.

[0173] For example, the feature "number of times of exceeding the package" in the multi-domain static event data corresponds to the event "exceeding the package", the feature "number of times of changing the package" corresponds to the event "package change", the feature "complaint market" corresponds to the event "complaint", and the feature "complaint home broadband" corresponds to the event "complaint".

[0174] In one implementation, one event corresponds to one feature, and in this case, the weight of the feature corresponding to the event is the basic weight of the event.

[0175] In another implementation, an event corresponds to multiple features. In this case, the basic weight of the event may be the weighted sum of the weights of the features corresponding to the event or their average value.

[0176] After obtaining the basic weight of each event, it can be normalized and the normalized basic weight is used as the final basic weight.

[0177] For example, each event can be ranked based on the size of the basic weight. The size of the basic weight indicates the impact of the event on the user's network portability and reflects the importance of the event to the user.

[0178] Methods for analyzing the root causes of user network portability.

[0179] Figure 4 A flow chart of a method for analyzing the root cause of a user's network portability is exemplarily shown.

[0180] In some embodiments, the network portability probability of each user in the set of users to be predicted is first predicted, and then the set of users to be analyzed is screened out from the set of users to be predicted based on the network portability probability of the user, and then the root cause of the network portability of each user in the set of users to be analyzed is analyzed. Figure 1A or Figure 1B The method may include but is not limited to some or all of the following steps:

[0181] S10: The big data platform sends the multi-domain static event data of each user in the set of users to be predicted to the application server.

[0182] The to-be-predicted user set includes multiple to-be-predicted users.

[0183] The set of users to be predicted may be all or part of the users in the big data platform. The multi-domain static event data may be the user data of the users to be predicted generated by the big data platform in the previous X months (eg, 3 months).

[0184] The big data platform can periodically send the multi-domain static event data of each user in the set of users to be predicted to the application server, for example, once or more every three months.

[0185] S20: The application server calculates the network portability probability of each user in the set of users to be predicted based on the multi-domain static event data of the user.

[0186] For example, Figure 5As shown, taking the calculation method of the network portability probability of a user to be predicted in the set of users to be predicted as an example, a flow chart of the calculation method of the network portability probability of the user is exemplarily shown. The method may include but is not limited to some or all of the following steps:

[0187] S201: The application server predicts a first probability of the user to be predicted switching networks based on the multi-domain static event data of the user to be predicted.

[0188] Among them, the first probability can be obtained based on the integrated learning classification model obtained by the above training. The method is to use the integrated learning classification model obtained by the above training to calculate the first probability of the user to be predicted to switch networks. Specifically, the multi-domain static event data of the user to be predicted is input into the integrated learning classification model, and each decision tree in the integrated learning classification model predicts the user to be predicted based on the multi-domain static event data of the user to be predicted, and obtains the prediction result of each decision tree. The prediction result includes users who switch networks and users who do not switch networks. Then, the ratio of the number of users whose prediction results are users who switch networks to the total number of prediction results is used as the first probability of the user to be predicted to switch networks.

[0189] S202: The application server calculates a second probability of the user to be predicted switching networks based on the matching degree between the individual service history of the user to be predicted and the best matching common service history.

[0190] Figure 6 A method for calculating the second probability is exemplarily shown, and the method includes some or all of the following steps.

[0191] S2021: The application server obtains the personalized service history of the user to be predicted and the common service history of multiple customer groups.

[0192] After identifying the long-term negative perception events and short-term severe mutation events in the original service history of the user to be predicted from month TX to month T, these identified events are connected in series according to the time of occurrence to obtain the personalized service history of the user to be predicted. Wherein, T is the time, which can be the current time, the current month or the month before the current month, and X is a positive integer, which can be 6 months, 12 months, 24 months, 36 months or more months. The specific method of generating the personalized service history of the user to be predicted can refer to the relevant description in the above-mentioned method for generating the personalized service history.

[0193] After the individual service process is divided into groups, frequent pattern mining and process aggregation can be used to obtain the common service process of the customer group. For specific methods, please refer to the relevant description of the method for generating the common service process of the customer group mentioned above.

[0194] S2022: The application server matches the individual service history of the user to be predicted with the common service history of multiple customer groups to obtain the common service history of one or more customer groups that matches the individual service history of the user to be predicted.

[0195] The matching formula is as follows:

[0196]

[0197] Among them, G represents the set of events in the common service history of the customer group, and S represents the set of events in the individual service history of the user to be predicted.

[0198] For example, Figure 7 As shown, a schematic diagram of the individual service history of the user to be predicted and the common service history of multiple customer groups is shown.

[0199] Assume that the user's personalized service history includes event A occurring at time t1, event B occurring at time t2, event C occurring at time t3, event D occurring at time t4, and event B occurring at time t5.

[0200] Figure 7 The common service history of multiple customer groups is also shown as an example.

[0201] The common service history of customer group 1 includes event A, event B, and event C arranged in chronological order.

[0202] The common service history of customer group 2 includes event A, event B, and event F arranged in chronological order.

[0203] The common service history of customer group 3 includes event A, event C, and event B arranged in chronological order.

[0204] The common service history of customer group 4 includes event C, event D, and event B arranged in chronological order.

[0205] It should be understood that an actual user's individual service history or a customer group's common service history may include more or fewer events.

[0206] In one implementation, when all events in the common service history of a certain customer group are included in the individual service history of the user to be predicted, it is considered that the common service history of the customer group matches the individual service history of the user to be predicted. At this time, the events included in the common service history of customer groups 1, 3, and 4 are all included in the events included in the individual service history to be predicted, so the common service history of customer groups 1, 3, and 4 all match the individual service history of the user to be predicted.

[0207] At this time, event A occurring at time t1, event B occurring at time t2, and event C occurring at time t3 in the personalized service history of the user to be predicted are matching events with the common service history of customer group 1.

[0208] In another implementation, when all events in the common service history of a certain customer group are included in the individual service history of the user to be predicted, and the order of the events in the individual service history of the user to be predicted that are the same as those in the common service history of the customer group is the same as the order of the events in the common service history of the customer group, then it is considered that the common service history of the customer group matches the individual service history of the user to be predicted. At this time, the events included in the individual service history of the user to be predicted and the common service history of the customer groups are matching events. At this time, the events included in the common service history of customer groups 1, 3, and 4 are all included in the events included in the individual service history to be predicted, but the order of events in customer group 3 is different from the order of the individual service history of the user to be predicted, therefore, only the common service history of customer groups 1 and 4 matches the individual service history of the user to be predicted.

[0209] At this time, event A occurring at time t1, event B occurring at time t2, and event C occurring at time t3 in the personalized service history of the user to be predicted are matching events with the common service history of customer group 1.

[0210] At this time, event C occurring at time t3, event D occurring at time t4, and event B occurring at time t5 in the personalized service history of the user to be predicted are matching events with the common service history of customer group 4.

[0211] S2023: The application server determines the common service history of the best matching customer group for the user to be predicted from the matched common service history based on the occurrence time of the matching event in the individual service history of the user to be predicted.

[0212] For any common service history R of a customer group that matches the individual service history of the user to be predicted, the target value S between the individual service history of the user to be predicted and the common service history of the matching customer group is:

[0213]

[0214] The number of events (also called matching events) in the personalized service history of the user to be predicted that match a common service history R is n, and R n The last event among the n matching events, that is, the event whose occurrence time is closest to the current time. i is the i-th matching event among n matching events, where i and n are positive integers and i is not greater than n. The event R in the personalized service history of the user to be predictedn The time of occurrence, The event R in the personalized service history of the user to be predicted i The time of occurrence.

[0215] The target value S between the individual service history of the user to be predicted and the common service history of each matching customer group is calculated, and the common service history of the matching customer group that minimizes S is determined as the best matching common service history of the user to be predicted.

[0216] Exemplarily, the target value S between the individual service history of the user to be predicted and the common service history of the matched customer group 1 is: S=|t3-t1|+|t3-t2|+|t3-t3|.

[0217] As another example, the target value T between the individual service history of the user to be predicted and the common service history of the matched customer group 4 is: S=|t5-t3|+|t5-t4|+|t5-t5|.

[0218] S2024: The application server obtains a second probability of the user to be predicted switching networks based on the matching degree between the best matching common service history of the user to be predicted and the individual service history of the user to be predicted.

[0219] In one implementation, the matching degree of the user to be predicted is the ratio of the number of all events in the best matching common service history to the number of all events in the personalized service history of the user to be predicted. The matching degree of the user to be predicted is the second probability of the user to be predicted switching networks.

[0220] In another implementation, the matching degree of the to-be-predicted user is proportional to the second probability of the to-be-predicted user switching networks.

[0221] The above method can be used to obtain the second probability of network portability of all users to be predicted in the sample set to be predicted.

[0222] S203: The application server determines the network portability probability of the user based on the first probability and the second probability of the user to be predicted.

[0223] In a possible implementation, it is determined that the weight of the first probability p1 of network portability of a certain user to be predicted in the final probability is w1, and the weight of the second probability p2 of the network portability of the user to be predicted in the final probability is w2, then the network portability probability p of the user to be predicted is: p=p1*w1+p2*w2.

[0224] It should be understood that there is no order of precedence for the steps of calculating the first probability and calculating the second probability, and they can be performed in any order.

[0225] In other embodiments, only the first probability or the second probability of the user to be predicted may be used as the network portability probability of the user to be predicted.

[0226] S30: The application server selects a plurality of users from the set of users to be predicted to form a set of users to be analyzed based on the network portability probability of each user in the set of users to be predicted.

[0227] In one implementation, all users to be predicted are sorted according to their probability of network porting, and the first P users with the largest probability of network porting are selected as users to be analyzed, where P is a positive integer not less than 1. Exemplarily, P is 1000, or 10000, etc. Alternatively, P is determined based on the total number of users in the set of users to be predicted, such as 70% or 80% of the total number.

[0228] In another implementation, users whose network porting probability is greater than a preset value (such as 0.8, 0.85, 0.91, etc.) in the set of users to be predicted may be selected to form the set of users to be analyzed.

[0229] In another implementation, users whose first probability is greater than a preset value (such as 0.8, 0.85, 0.91, etc.) in the user set to be predicted can be selected to form the user set to be analyzed, or the first P users with the largest first probability can be selected to form the user set to be analyzed.

[0230] In another implementation, all the users to be predicted are selected as the users to be analyzed.

[0231] S40: The application server obtains the basic weight of each event in the personalized service history of the user to be analyzed based on the personalized service history of each user in the set of users to be analyzed and the basic weight of each event.

[0232] In one implementation, the basic weight of the event is assigned to the event of the personalized service history of the user to be analyzed, so as to obtain the personalized service history of the user to be analyzed with the basic weight.

[0233] For example, the basic weights of event A, event B, event C, and event D are W and W, respectively. A , W B , W C , W D .

[0234] like Figure 8 Figure (a) shows the personalized service history of the user to be analyzed, including event A occurring at time t1, event B occurring at time t2, event C occurring at time t3, event D occurring at time t4, and event B occurring at time t5.

[0235] If Figure 8 As shown in Figure (b), it is the personalized service history of the user to be analyzed with basic weights.

[0236] In another implementation, the weight of an event in the personalized service process can also be different from the basic weight of the event. For example, the weight of any event in the personalized service process is a multiple of the basic weight of the event or the weight after normalizing the basic weight of the event.

[0237] S50: For each user in the set of users to be analyzed, the application server corrects the basic weight of each event in their personalized service process to obtain the corrected weight of each event in their personalized service process.

[0238] A method for obtaining the corrected weight of an event in the personalized service process of a user to be analyzed provided by the present invention is based on one or more ways such as the time interval between events or the interaction between events in the personalized service process of the user to be analyzed, and obtains the corrected weight of the events in the personalized service process of the user to be analyzed.

[0239] Fig. 9 Exemplarily shown is a method for obtaining the corrected weight of an event in the personalized service process of a user, which includes S501 - S502, or includes S501, or includes S502. Specifically as follows:

[0240] S501: Based on the time interval between events in the personalized service process of the user to be analyzed, obtain the corrected weight of the events in the personalized service process of the user to be analyzed.

[0241] In some embodiments, the basic weight of each event in the personalized service process can be corrected based on the time interval between events. Among them, the time interval of an event in the personalized service process refers to the difference between the occurrence time of the most recently occurred event in the personalized service process and the occurrence time of this event. Taking the time interval of an event in the personalized service process of a user to be analyzed as an example, among the events in the personalized service process of this user, the occurrence time of the most recently occurred event is t n , the occurrence time of a certain event is t i , then the time interval of this event is t n - t i . Among them, i < n, and i and n are positive integers not less than 1.

[0242] This application takes the time series cooling model as an example for illustration. In other embodiments, it can also be other models.

[0243] Based on the idea of the time series cooling algorithm, each event that occurs at the current moment of the user has a default event heat. The closer it is to the current moment, the higher the event heat of this event. As time goes by, the event heat of all events gradually decreases slowly. The time series cooling model imitates Newton's law of cooling in physics and correlates the event heat of an event with the time interval of the event.

[0244] Fig.10 The relationship between event popularity and time interval is shown as an example. For a user to be analyzed, t n Indicates the time of occurrence of the last event in the user's personalized service history. t1 indicates the time of occurrence of the first event in the user's personalized service history. t2 indicates the time of occurrence of the second event in the user's personalized service history. t3 indicates the time of occurrence of the third event in the user's personalized service history. For an event in the personalized service history, the time of occurrence of the event is t n The closer it is, the greater the event popularity of the event.

[0245] Based on the relationship between event heat and time interval, the event heat of each event in the personalized service history of the user to be analyzed can be calculated.

[0246] For example, the event heats of event A occurring at time t1, event B occurring at time t2, event C occurring at time t3, event D occurring at time t4, and event B occurring at time t5 are η1, η2, η3, η4, and η5 respectively. Figure 8 As shown in Figure (c), the weight of the above events is adjusted to η1*W A ,η2*W B ,η3*W C ,η4*W D ,η5*W B In a possible implementation, by presetting the event heat of the events in the personalized service history of the user to be analyzed at the current moment, the application server calculates the modified weight of each event in the personalized service history of the user to be analyzed using the decay function relationship between each event heat and the time interval.

[0247] Fig.11 The flowchart of the specific implementation of step S501 is exemplarily shown.

[0248] S5011: The application server establishes a decay function of the event heat and time interval of each event.

[0249] The attenuation function of an event's event heat and time interval describes the corresponding relationship between the event heat and time interval of the event. As the time interval increases, the event heat decreases. The attenuation functions of different events can be different or the same.

[0250] In a possible implementation, the decay function may be an exponential decay function, and the function expression is as follows:

[0251] S5012: The application server obtains the event heat corresponding to the time interval of each event in the personalized service history of the user to be analyzed based on the established attenuation function of the event heat and time interval of each event.

[0252] S5013: The application server modifies the basic weight of each event in the personalized service history of the user to be analyzed based on the event heat of each event, and obtains the modified weight of each event in the personalized service history of the user to be analyzed.

[0253] Exemplarily, for any event in the personalized service history of the user to be analyzed, the modified weight of the event is the product of the event heat of the event and the basic weight of the event.

[0254] Based on the relationship between event heat and time interval, the event heat of each event in the personalized service history of the user to be analyzed can be calculated.

[0255] For example, the event heats of event A occurring at time t1, event B occurring at time t2, event C occurring at time t3, event D occurring at time t4, and event B occurring at time t5 are η1, η2, η3, η4, and η5 respectively. Figure 8 As shown in Figure (c), the weight of the above events is adjusted to η1*W A ,η2*W B ,η3*W C ,η4*W D ,η5*W B .

[0256] As another example, the personalized service history of a user based on the weight modified by the time series cooling model is recorded as follows. The recording format of each event is "event name: time; event ID; current weight; cumulative weight", "=>" identifies the next event, where the cumulative weight is the sum of the weights before the current event (including the current event). The personalized service history of the user is as follows:

[0257] "Access: 2009-12-20 11:23:01; 0; 0.0; 0.0

[0258] = >ARPU increased: 2022-12-31 23:59:59; 11; 0.015730754840946968

[0259] =>Traffic exceeded: 2022-12-31 23:59:59; 18; 0.0023371407192264064; 0.018067895560173373

[0260] =>Mobile secondary card->Mobile primary card: 2022-12-31 23:59:59; 19; 0.0; 0.018067895560173373

[0261] = >ARPU increased: 2023-01-31 23:59:59; 11; 0.02278639092734213

[0262] =>Traffic exceeded: 2023-01-31 23:59:59; 18; 0.01752583993519823; 0.04031223086254036

[0263] =ARPU increased: 2023-03-01 23:59:59; 11; 0.0724184073055541

[0264] =>Traffic exceeds the limit: 2023-03-01 23:59:59; 18; 0.5155815926944459; 0.58800000000000001”

[0265] It can be seen that after the correction, the same event has different weights after correction due to the different time intervals between its occurrence and the most recent event. The changes in the weights of events in the user's personalized service history before and after the correction based on the time interval of the event are shown exemplarily. For the same event "ARPU increase" in the personalized service history of the same user, since the time interval between the subsequent event "ARPU increase" is shorter, its weight "0.03210617644301374" is significantly higher than the weight "0.015730754840946968" of the previous "ARPU increase". It should be understood that when a positive event occurs, the weights of one or more subsequent negative events will decrease. The weight of "ARPU increase" after the positive event "mobile secondary card->mobile primary card" is "0.004718495367168755", which is lower than the previous "0.015730754840946968".

[0266] As another example, the personalized service history of another user based on the weight modified by the time series cooling model is recorded as follows. The recording format of each event is "event name: time; event ID; current weight; cumulative weight", "=>" identifies the next event, where the cumulative weight is the sum of the weights before the current event (including the current event). The personalized service history of the user is as follows:

[0267] "Network access: 2010-10-24 14:03:46:0:0.0:0.0

[0268] => Voice superset: 2022-09-3023:59:59; 18; 0.004814041822069206; 0.004814041822069206

[0269] => Short board label, wireless disconnection: 2022-09-30 23:59:59; 31; 0.0010780509308732618; 0.005892092752942468

[0270] = >ARPU increased: 2022-10-31 23:59:59; 11; 0.32152787482072637; 0.3274199675736688

[0271] => Traffic exceeds the limit: 2022-10-31 23:59:59; 18; 0.33697393871119896

[0272] = > ARPU increased: 2022-11-3023: 59: 59; 11; 0.09548379141574946; 0.43245773012694844

[0273] => Traffic exceeds the limit: 2022-11-30 23:59:59; 18; 0.5033885466072194

[0274] = > Traffic exceeds the limit: 2022-12-0415:01:50; 18; 0.2686114533927806: 0.772

[0275] =>Query the conversion code: 2022-12-04 15:01:50; 1; 0.0; 0.772” It can be seen that the same event “traffic overrun” in the user’s personalized service history has a weight of 0.07093081648027104, which is significantly higher than the weight of 0.009553971137530155 of the previous “traffic overrun”, because the time interval between the subsequent events is shorter.

[0276] S502: Based on the interaction between events, obtain the modified weights of the events in the personalized service history of the user to be analyzed.

[0277] In some embodiments, the basic weight of each event in the personalized service history of the user to be analyzed can be modified based on the interaction between events. The interaction between events means that the events will affect each other, and the mutual influence between events has a transmission relationship. For example, the event "exceeding the package" will affect the event "carrying and switching the network". The event "wireless type user" will affect the event "exceeding the package", which in turn will affect the event "carrying and switching the network".

[0278] like Fig.12 The hierarchical analysis model shown includes a target layer, a criterion layer and a solution layer. The target layer is "the tendency of users to switch networks". The criterion layer includes features for judging whether users have a tendency to switch networks, including one or more features such as the time interval from the occurrence of the event to the final recording time (referred to as the time interval), the user's age, the network access time, the user type, the traffic preference, and the user's rating of customer service (referred to as the user rating). For the events of the solution layer, the solution layer includes multiple event types, exemplarily including three event types, including "network type", "home broadband type" and "market type". Exemplarily, the events of "network type" include events such as "wireless disconnection" and "slow wireless speed". The events of "home broadband type" include events such as "home broadband disconnection" and "slow home broadband speed". The events of "market type" include events such as "package change" and "traffic exceeding the package".

[0279] Further, the weight of each event type at the solution layer is determined. For example, the weight of the network type is V1, the weight of the home broadband type is V2, and the weight of the market type is V3. The calculation method of the weight can refer to the calculation method of the following network type weight V1, home broadband type weight V2, and market type weight V3, which will not be repeated here.

[0280] For two adjacent events in the user's personalized service history, they are exemplarily arranged in order of occurrence time as event A and event B, that is, event A occurs before event B. If event A belongs to the first event type and event B belongs to the second event type, and there is an interaction between the first event type and the second event type, then the modified weight of event B is:

[0281] W * =W+ΔH=W+(V a +V b ) / 2*S ab (1)

[0282] W is the current weight of event B.

[0283] In some embodiments, the current weight is the weight modified by the above step S501. In other embodiments, the current weight is the basic weight. 12is a preset value, used to indicate the coefficient of interaction between the first event type that occurs first and the second event type that occurs later, S 12 Not greater than 1.

[0284] In some embodiments, when a positive event occurs, the weights of one or more negative events immediately following the positive event are reduced, for example, set to 0, or reduced to the original 0.8 or other values.

[0285] For example, Figure 8 As shown in Figure (d), events A and D belong to the network type, event B belongs to the home broadband type, and event C belongs to the market type. The weights of event A occurring at time t1, event B occurring at time t2, event C occurring at time t3, event D occurring at time t4, and event B occurring at time t5 are adjusted to η1*W A ,η2*W B +ΔH1、η3*W C +ΔH2、η4*W D +ΔH3、η5*W B +ΔH4.

[0286] Among them, ΔH1, ΔH2, ΔH3, and ΔH4 can be obtained based on the above formula (1).

[0287] S60: Determine the root cause of the user's network portability based on the modified weight of each event in the personalized service history of each user in the set of users to be analyzed.

[0288] In a possible implementation, for a certain user to be analyzed, the events of the personalized service history of the user to be analyzed are sorted according to the size of the correction weight, and the first K events with the largest correction weight are selected as the root cause of the user's portability, that is, the root cause event. The root cause event, also known as the key event, is a major event that affects the user's portability. K is a positive integer greater than 1.

[0289] S70: The application server sends at least one of the personalized service history, event weight, network portability probability and root cause of network portability of each user in the user set to be analyzed to the customer service device.

[0290] S80: The customer service device visually displays at least one of the personalized service history, event weight, network portability probability, and root cause of network portability of each user in the set of users to be analyzed.

[0291] like Fig.13AAs shown, a method for visually displaying a user's personalized service history, event weight, network portability probability, and root cause of network portability provided by an embodiment of the present application. This method will represent the root cause event in different colors, or in different font sizes, thicknesses, etc. based on the size of the event weight. Fig.13A The root cause events are indicated by bold and larger fonts. The recording format of each event is "event name: time; weight", and "=>" marks the next event.

[0292] like Fig. 13B As shown, another method for visually displaying a user's personalized service history, the probability of network switching, and the root cause of network switching provided by the embodiment of the present application. This method uses different graphic elements to represent different events, displays the weight of each event, and based on the size of the event weight, highlights the root cause event. The highlighting method may include enlarging the graphic element, setting it to a preset color, thickening the border, etc. Fig. 13B Let’s take the example of making the border thicker.

[0293] In another implementation, the application server may send text, pictures, charts, etc. generated by visualizing the personalized service history, network portability probability, and root causes of network portability of each user in the set of users to be analyzed to the customer service device. Then, the customer service device may display the text, pictures, or charts.

[0294] Calculation method of weights for network type, home broadband type and market type

[0295] Combination Fig.12 The analytic hierarchy model shown and Fig.14 The flowchart of the method for calculating the weights of network type, home broadband type and market type shown in the figure illustrates the method for calculating the weights of network type, home broadband type and market type involved in the present application. The method can be executed by an application server or other device, and may include but is not limited to some or all of the following steps:

[0296] S5021: Constructing a hierarchical analysis model

[0297] Fig.12 The example shows how to calculate the modified weights of events in a user's personalized service history using a hierarchical analysis model.

[0298] Assume that there are c features in the criterion layer, where c is a positive integer greater than 1. Construct a solution layer, where there are d event categories in the solution layer, where d is a positive integer not less than 1. Exemplarily, the event categories in the solution layer include network type, home broadband type, and market type.

[0299] S5022: The application server constructs a judgment matrix A of the criterion layer.

[0300] Assume that there are c features in the criterion layer, and c is a positive integer greater than 1. Based on prior experience, the application server determines the relative importance between any two features of the criterion layer and constructs a judgment matrix A of the criterion layer.

[0301] Specifically, the relative importance of any two features in the criterion layer relative to the target layer is determined. The relative importance can be represented by discrete values ​​1-9 and 1 / 9-1. The discrete values ​​of each two features in the criterion layer relative to the target layer are used as matrix elements of the criterion layer judgment matrix A. The criterion layer judgment matrix A is a c-order square matrix, and the matrix element a in the i-th row and j-th column of the criterion layer judgment matrix A is ij It is the discrete value between the i-th feature and the j-th feature among the c features relative to the target layer, where i and j are positive integers not greater than c.

[0302] S5023: The application server performs a consistency check on the judgment matrix A of the criterion layer.

[0303] Specifically, based on the eigenvalues ​​and eigenvectors of the judgment matrix A, the consistency test of the judgment matrix A is performed using the test index CI and the Satty consistency index RI, where the RI value is given by prior experience and the CI value is given by the following formula:

[0304]

[0305] where λ A max is the maximum eigenvalue of the judgment matrix A, c is the order of the judgment matrix A, and the CR value is calculated: CR = CI / RI. If CR < 0.1, the judgment matrix A satisfies the consistency test, otherwise the judgment matrix A needs to be corrected.

[0306] S5024: Calculate the weight of the criterion layer to the target layer (also called the first weight).

[0307] Calculate the first weight of each of the c features and take the maximum eigenvalue λ of the judgment matrix A that passes the consistency test A max The corresponding eigenvector is normalized, and the maximum eigenvalue λ of the normalized judgment matrix A is recorded A max The corresponding eigenvector is w A , w A The i-th element in is the first weight of the i-th feature among the c features.

[0308] S5025: The application server constructs a judgment matrix of the solution layer for each feature of the criterion layer based on the relative importance of the feature between any two event types of the solution layer.

[0309] The criterion layer has c features, so the judgment matrix of c solution layers is constructed, which can be expressed as {B i}, where i = 1, 2, ..., c. For the i-th feature in the criterion layer, its corresponding judgment matrix B i is a d-order matrix. Judgment matrix B i The principle of determining the elements within is the same as the judgment matrix A mentioned above.

[0310] S5026: The application server performs a calculation on all judgment matrices B1, B2, ..., B c Perform consistency check. Refer to S5023 for the test method.

[0311] S5027: The application server determines the weight (also called the second weight) of each event type to the target layer based on the judgment matrix of each solution layer.

[0312] Specifically, the judgment matrices B1, B2, ..., B c The eigenvector corresponding to the maximum eigenvalue of . For the judgment matrix B i , let the judgment matrix B i The maximum eigenvalue λ i B The corresponding normalized eigenvector is denoted as W B i , W B i It includes d elements, which are expressed as w i1 B , w i2 B ,…,w id B , calculate the second weight of each event type in the solution layer. Taking the kth event type in the solution layer as an example, k = 1, 2, ..., d, its second weight V k The calculation formula is as follows:

[0313]

[0314] The above classification method of events is not limited. In some embodiments, events may not be classified or may be classified in other ways. The weights of the three types of network type, home broadband type and market type may also be calculated in other ways, such as pre-setting.

[0315] Method for generating personalized service history (i.e. user behavior sequence)

[0316] The following takes the generation of the target user's behavior sequence as an example to illustrate.

[0317] The target user's behavior sequence includes multiple user events and respective timestamps of the multiple user events. The target user's behavior sequence is obtained by sorting the multiple user events according to respective timestamps of the multiple user events.

[0318] In combination with the description of the above embodiment, the electronic device can determine the user events (ie, the multiple events) included in the behavior sequence of the target user based on the user data of the target user and can also determine the timestamps of the multiple events based on the user data.

[0319] In an embodiment of the present application, after determining (or obtaining) multiple user events and respective timestamps of the multiple user events, the electronic device sorts the multiple user events according to respective timestamps of the multiple user events, and can accurately and effectively reflect the user events that occur (or are executed) in chronological order when the target user is performing telecommunications services, that is, a correct and relevant behavior sequence that can effectively reflect the user behavior of each user can be obtained.

[0320] For example, Fig.15 As shown, the user data of a user may include the user's O-domain data, B-domain data, and service domain data. The user's O-domain data includes product tariff data and consumption attribute data related to the user, including base station engineering parameter table, home broadband resource service topology, resident base station level value, and OLT / BRAS faults.

[0321] The user's B-domain data includes the user's network perception data and network performance data, including voice / traffic low saturation, voice / traffic over-package, speed limit, and main tariff upgrade / downgrade.

[0322] The user's service domain data includes the user's contact service data, including business consultation, business processing, complaints and post-use reviews, etc.

[0323] After the electronic device obtains the user data of the user (that is, including the above-mentioned O domain data, B domain data and service domain data), it can obtain the behavior sequence of the user based on the user data of the user.

[0324] The user's behavior sequence includes 7 user events, namely, network access, voice overage, low measurement report (MR) coverage, complaint about wireless Internet access, speed limit, processing of phone bill contract, and number portability. The time points corresponding to the 7 user events are 2021-02-01 12:59:01, 2021-03-31 08:00:00, 2021-04-30 12:00:01, 2021-06-01 14:08:11, 2021-08-20 09:20:11, 2021-10-02 16:24:02, and 2022-01-01 19:07:09. That is, joining the network → voice coverage exceeding the limit → low MR coverage → complaining about wireless Internet access issues → speed limit → applying for a phone bill contract → number portability can be understood as a user's behavior sequence, and each user event included in the behavior sequence has a corresponding timestamp.

[0325] In one implementation of an embodiment of the present application, the behavior sequence of the above-mentioned target user is determined based on a user behavior history graph, which includes a user node, an event node, and an edge connecting the user node and the event node. The user node is used to represent the user, the event node is used to represent the user event, and the edge is used to represent the timestamp of the user executing the user event.

[0326] Specifically, for a certain user (for example, a target user), the electronic device can determine the user node corresponding to the target user from the user behavior history graph, and determine the event node that has an edge (or connection relationship) with the user node as the event node corresponding to the above-mentioned multiple user events, and determine the timestamp represented by the edge between the user node and the event nodes corresponding to the multiple user events as the timestamp of each of the multiple user events, so that the behavior sequence of the target user can be obtained.

[0327] In the embodiment of the present application, since the user behavior history graph includes user nodes for representing users, event nodes for representing user events, and edges for representing the timestamps of users executing user events, the electronic device can easily and quickly determine the behavior sequence of each user based on the user behavior history graph, thereby improving the rate of generating corresponding strategies for each user.

[0328] The electronic device obtains the behavior sequence of the target user, which may specifically include S1011-S1015.

[0329] S1011. The electronic device obtains a user ID of a target user.

[0330] S1012: The electronic device queries a user behavior history graph based on the user identification of the target user, and determines a target user node from the user behavior history graph.

[0331] Among them, the node identifier of the target user node is the same as the user identifier of the target user, and the user behavior history graph includes a user node, an event node, and an edge connecting the user node and the event node. The user node is used to represent the user, the event node is used to represent the user event, and the edge is used to represent the timestamp of the user executing the user event.

[0332] Specifically, the user behavior history graph includes multiple user nodes, and the electronic device (or the user behavior history graph) stores the respective identifiers of the multiple user nodes. The electronic device can determine the target user node from the multiple user nodes based on the user identifier of the target user, and the target user node is used to characterize the target user.

[0333] S1013: The electronic device determines the user events represented by the multiple event nodes as the multiple user events included in the behavior sequence of the target user.

[0334] The multiple event nodes are event nodes that have edges with the target user node.

[0335] It should be understood that there are edges between the multiple event nodes and the target user node, indicating that the user represented by the target user node has executed the user events represented by the multiple event nodes. Since the user represented by the target user node is the target user, the user events represented by the multiple event nodes are user events executed by the target user, so it can be concluded that the user events represented by the multiple event nodes are multiple user events included in the behavior sequence of the target user.

[0336] S1014: The electronic device determines the timestamps represented by the multiple edges as the timestamps of the multiple user events respectively.

[0337] Among them, one edge is used to connect the above-mentioned target user node and one event node among multiple event nodes.

[0338] In combination with the description of the above embodiment, the edge connecting a user node and an event node is used to represent the timestamp of the user represented by the user node executing the user event represented by the event node. In this way, after the electronic device determines the multiple edges between the multiple event nodes and the target user node, the timestamps represented by each of the multiple edges can be determined as the timestamps of each of the multiple user events included in the behavior sequence of the target user. At this point, the electronic device can obtain the multiple user events and the timestamps of each of the multiple user events.

[0339] S1015. The electronic device sorts the multiple user events according to their respective timestamps to obtain a behavior sequence of the target user.

[0340] In an embodiment of the present application, after obtaining the user identification of the target user, the electronic device can determine the target user node from the user behavior history graph, the node identification of the target user node is the same as the user identification of the target user, and the target user node is used to characterize the target user; then the electronic device can determine the user events represented by multiple event nodes in the user behavior history graph that have edges (or connection relationships) with the target user node as multiple user events included in the behavior sequence of the target user, and determine the timestamps represented by the multiple edges between the target user node and the multiple event nodes in the user behavior history graph as the timestamps of the multiple user events; finally, the electronic device can sort the multiple user events according to the timestamps of the multiple user events to obtain the behavior sequence of the target user. That is, the electronic device can conveniently and quickly determine the user events performed by each user and the timestamps of the user events performed by each user based on the user identification of each user and the user behavior history graph, which can improve the efficiency of obtaining the behavior sequence of each user.

[0341] Optionally, the electronic device may store the behavior sequence of the target user in the form of a user behavior history map.

[0342] For example, Fig.16 As shown, for a single user (for example, user A), user node A can represent user A, and there are 6 event nodes with edges (or connection relationships) between user node A, including event node 201, event node 202, event node 203, event node 204, event node 205 and event node 206. The timestamps of the 6 user events represented by the 6 event nodes are timestamp 1, timestamp 2, timestamp 3, timestamp 4, timestamp 5 and timestamp 6 respectively. The 6 user events are speed limit 2-4 times, voice package exceeding, traffic package exceeding, package cancellation, complaint about network migration and query of transcoding in the order of timestamps.

[0343] Continue as Fig.16 As shown, for multiple users (including user A and user B), user node B can represent user B, and there are four event nodes with edges (or connection relationships) between user node B, including event node 204, event node 205, event node 207, and event node 208. The timestamps corresponding to the four user events represented by the four event nodes are timestamp 1, timestamp 2, timestamp 3, and timestamp 4, respectively. The four user events are package cancellation, low traffic saturation, complaint about home broadband, and complaint about network migration in the order of timestamps. Since the user events executed by user A and user B both include package cancellation and complaint about network migration, that is, there are edges between user node A and user node B and event node 204 and event node 205.

[0344] If the number of user nodes is m and the number of event nodes is n, then the user behavior history graph contains a total of m+n nodes, and the number of edges (or association relationships) is less than m*n, where m is an integer greater than or equal to 1, and n is an integer greater than or equal to 1.

[0345] For user A and user B, the behavior sequence of user A determined by the electronic device based on user A's user identification and user behavior history graph is: speed limit 2-4 times → voice exceeding package → data exceeding package → cancel package → complain about mobile network → query portability transcoding. The behavior sequence of user B determined by the electronic device based on user B's user identification and user behavior history graph is: cancel package → low data saturation → complain about home broadband → complain about mobile network.

[0346] In one implementation of an embodiment of the present application, the user events contained in the behavior sequence of the above-mentioned target user include negative perception events and / or subjective feedback events, wherein the negative perception event is used to characterize the passive consumption behavior corresponding to the target user, and the subjective feedback event is used to characterize the active complaint behavior corresponding to the target user.

[0347] Specifically, passive consumption behaviors may include voice overage, traffic overage, voice undersaturation, traffic undersaturation, and speed limit, etc. Active complaint behaviors may include complaints and complaints, etc.

[0348] In the embodiment of the present application, since the user events included in the target user's behavior sequence include negative perception events and / or subjective feedback events, that is, the electronic device can characterize the user behavior of each user in the process of performing telecommunications services in the form of passive consumption behavior corresponding to each user and active complaint behavior corresponding to each user, and can comprehensively and completely obtain the behavior sequence of each user.

[0349] Optionally, the electronic device can subdivide negative perception events into long-term negative perception events and short-term severe perception events. Long-term negative perception events are used to characterize that a negative perception event occurs a large number of times (or frequency) within a preset time period, while short-term severe perception events are used to characterize that a negative perception event has a large range of changes during a single occurrence.

[0350] For example, long-term negative perception events may be speed limit 2-4 times, speed limit greater than 4 times, and voice overage greater than 4 times, etc. Short-term serious perception events may be voice overage (for example, voice exceeds the package by 1 times) and traffic overage.

[0351] In an optional implementation, the electronic device may classify the above-mentioned long-term negative perception events as long-term negative perception modes, and classify the above-mentioned subjective feedback events and short-term severe perception events as short-term severe mutation modes, and use the user events included in the two modes to form a user's behavior sequence.

[0352] In another optional implementation, porting and leaving the network can also be understood as a landmark event that occurs (or is performed) by the user in the process of carrying out telecommunications services. In the process of carrying out telecommunications services, the user can also perform a bundled event that is closely related to (or has a collateral effect on) the landmark event, and the electronic device can delete the bundled event to avoid the negative impact of the bundled event on the user's behavior sequence.

[0353] Exemplarily, bundling events may include user complaints after querying the code portability, broadband machine disassembly before portability, network withdrawal before portability, and package withdrawal before portability, etc.

[0354] Optionally, the electronic device can determine the original behavior sequence of the target user based on the user data of the above-mentioned target user, and then identify the user events included in the above-mentioned long-term negative perception pattern and the user events included in the short-term severe mutation pattern from all user events included in the original behavior sequence, and delete the bundled events from the original behavior sequence to obtain the behavior sequence of the target user.

[0355] It is worth noting that a user's original behavior sequence can be understood as the user's original service history, and the user's behavior sequence can be understood as the user's personalized service history.

[0356] For example, Fig.17 As shown, the user's original behavior sequence includes 3 speed limits, 5 voice over-packages, 1 serious voice over-package, 1 package cancellation, 1 complaint about network migration, 1 package withdrawal before porting, and 1 porting.

[0357] The three speed limits and the five voice over-sets correspond to a long-term negative perception event respectively. The corresponding user events are speed limits 2-4 times and voice over-sets more than 4 times. The two long-term negative perception events belong to the long-term negative perception mode.

[0358] This severe voice overshoot corresponds to a short-term severe perception event, and the specific corresponding user event is severe voice overshoot.

[0359] The cancellation of the package and the complaint about the network migration each correspond to a subjective feedback event, and the specific corresponding user events are cancellation of the package and complaint about the network migration. The short-term severe perception event and the two subjective feedback events belong to the short-term severe mutation mode.

[0360] This one package cancellation before transfer corresponds to a bundling event, and this one package transfer corresponds to a landmark event.

[0361] That is, the user events included in the behavior sequence of the user obtained by the electronic device are: speed limit 2-4 times, voice over limit more than 4 times, voice over limit seriously, package cancellation, complaint about network migration, and portability.

[0362] A method for generating the common service journey of customer groups (i.e. the common behavior sequence of user groups)

[0363] The common behavior sequence corresponding to the target user group is determined based on the behavior sequence of each user included in the target user group. Among them, the users in the target user group are all users who have switched networks. At this time, the common behavior sequence (that is, the common service history) of the target user group can reflect the characteristics contained in the individual service history of the users who have switched networks in this target user group.

[0364] In combination with the description of the above embodiment, the target user group is the user group to which the above target user belongs, and each user included in the target user group has the same user type.

[0365] It is worth noting that the common behavior sequence corresponding to the target user group can characterize the most common behavior pattern with time series characteristics for each user included in the target user group (or the user type corresponding to the target user group). After determining the target user group, the electronic device can determine the common behavior sequence corresponding to the target user group based on the behavior sequence of each user included in the target user group, and accurately and effectively determine the target strategy based on the common behavior sequence corresponding to the target user group. That is, the electronic device can accurately and effectively determine the maintenance strategy corresponding to each user based on the most common behavior pattern with time series characteristics corresponding to the user group to which each user belongs, thereby reducing the portability / off-network situation of each user.

[0366] In an optional implementation, the common behavior sequence corresponding to the target user group includes at least two behavior sequences, the target user event is a user event included in the converged behavior sequence corresponding to the target user group, and the converged behavior sequence corresponding to the target user group is determined based on the at least two behavior sequences.

[0367] Specifically, the electronic device can converge the at least two behavior sequences to obtain a converged behavior sequence corresponding to the target user group, and the converged behavior sequence includes user events included in each of the at least two behavior sequences. In an embodiment of the present application, since the converged behavior sequence corresponding to a user group converges all common behavior sequences corresponding to the user group, the converged behavior sequence can reflect the behavioral habits corresponding to the user group as a whole. And since the target user event is the user event included in the converged behavior sequence corresponding to the target user group, the electronic device can accurately screen out key events associated with the user type corresponding to the target user group, thereby improving the retention success rate of each type of user.

[0368] Optionally, the electronic device may further aggregate the same user events included in at least two behavior sequences, and sort the different user events included in at least two behavior sequences in chronological order of timestamps to obtain an aggregated behavior sequence corresponding to the target user group.

[0369] It is worth noting that for all user events included in at least two behavior sequences included in the common service sequence corresponding to the target user group, the electronic device aggregates the same user events, that is, the same user events can be merged and deduplicated, reducing the size of the behavior sequence (specifically the converged behavior sequence) and saving storage resources. The electronic device sorts different user events in the order of timestamps, which can fully cover every user event included in the common service sequence corresponding to the target user group, ensuring the integrity of the common service sequence.

[0370] In one implementation of an embodiment of the present application, the number of users corresponding to the above-mentioned target user event is determined based on a common behavior history graph, which includes at least two event nodes and an edge connecting the at least two event nodes. An event node is used to represent a user event, and the edge is used to represent the number of users corresponding to the downstream event node connected to the edge.

[0371] Specifically, the electronic device can determine the event node used to characterize the target user event from the common behavior history graph, and determine the direct upstream event node of the event node (specifically, the upstream node that has a connection relationship with the event node). Finally, the number of users represented by the edge between the direct upstream event node and the event node is determined as the number of users corresponding to the target user event. In the embodiment of the present application, since the common behavior history graph can characterize multiple user events and the number of users corresponding to each of the multiple user events, the electronic device can accurately and effectively determine the number of users corresponding to the target user event from the common behavior history graph, thereby improving the efficiency of determining the number of users corresponding to each user event.

[0372] In an optional implementation, the common behavior history diagram may be pre-stored in an electronic device.

[0373] In an optional implementation, the common behavior history diagram can be pre-stored in an electronic device. Based on this, the method can also include step M-step O.

[0374] Step M: The electronic device obtains the event identifier of the target user event.

[0375] Step N: the electronic device queries the common behavior history graph based on the event identifier of the target user event, and determines the target event node from the common behavior history graph.

[0376] Among them, the node identifier of the target event node is the same as the event identifier of the target user event, the common behavior history diagram includes at least two event nodes and an edge connecting the at least two event nodes, one event node is used to represent a user event, and the edge is used to represent the number of users corresponding to the downstream event node connected to the edge.

[0377] Step O: The electronic device determines the number of users represented by the target edge as the number of users corresponding to the target user event.

[0378] Among them, the downstream event node connected by the target edge is the target event node.

[0379] In an embodiment of the present application, after obtaining the event identifier of the target user event, the electronic device can determine the target event node from the common behavior history graph, the node identifier of the target event node is the same as the event identifier of the target user event, and the target event node can represent the target user event; then the electronic device can determine the target edge from the common behavior history graph, and the downstream event node connected to the target edge is the target event node; finally, the electronic device can determine the number of users represented by the target edge as the number of users corresponding to the target user event. That is, the electronic device can conveniently and quickly determine the number of users corresponding to each user event from the common behavior history graph based on the event identifier of each user event, thereby improving the rate at which the electronic device identifies key events and further improving the success rate of maintaining (off-network) users.

[0380] Optionally, the electronic device may use a graph multi-hop query technology to query at least two behavior sequences included in the common behavior history graph (specifically, the common behavior sequences corresponding to the target user group), thereby obtaining a converged behavior sequence corresponding to the target user group.

[0381] For example, Fig.18As shown, the common behavior sequences corresponding to the target user group include two behavior sequences, namely, behavior sequence 1 and behavior sequence 2. Specifically, behavior sequence 1 is: network access → speed limit 2-4 times → cancel package → complain about network migration → query portability and transcoding, and behavior sequence 2 is: network access → speed limit 2-4 times → cancel package → low traffic saturation → complain about wireless.

[0382] Based on the graph multi-hop query technology, the electronic device determines from the common behavior history graph that the same user events in the two behavior sequences include network access, speed limit 2-4 times and package cancellation, and different user events include complaints about network migration, querying portability and transcoding, low traffic saturation and complaints about wireless.

[0383] Therefore, the converged behavior sequence corresponding to the target user group obtained by the electronic device is network access → speed limit 2-4 times → cancel package → complain about network migration / low traffic saturation → query portability / complain about wireless. Among them, the two sides of / represent two branches.

[0384] It is worth noting that for the same user event in the two behavior sequences (eg, speed limit 2-4 times), the number of users corresponding to the same user event is the sum of the number of users corresponding to the user event in the two behavior sequences.

[0385] In one implementation of an embodiment of the present application, the user type of the above-mentioned target user is used to characterize the grouping information of the target user, and the grouping information of the target user includes one or more of the location information of the target user, the perception type corresponding to the target user, the risk level of the target user, the age of the target user, and the gender of the target user.

[0386] In the embodiment of the present application, the user type of the target user can also be understood as the user type corresponding to the user group to which the target user belongs (ie, the target user group), and a user group can also be understood as a customer group.

[0387] Prior knowledge shows that location information, perception type, risk level, age, and gender are important factors that may determine whether a user is likely to switch or leave the network. Therefore, based on information entropy gain, we can identify the grouping information (or important tags) that affect user switching or leaving the network, and use this to distinguish user groups (or groups).

[0388] Specifically, the electronic device can group users according to the importance of each piece of information included in the grouping information, first according to the most important information (or label), and then continue to group based on the second most important information, until the user type of each user can be obtained by grouping based on each piece of information. Through a hierarchical grouping strategy, similar users can be grouped into specific user groups (or specific customer groups), which facilitates the subsequent determination of the common behavior sequence corresponding to the user group.

[0389] For example, Fig.19As shown, assuming that the target user's grouping information includes the target user's location information, the target user's corresponding perception type, and the target user's risk level, the importance of the location information is greater than the importance of the perception type, and the importance of the perception type is greater than the importance of the risk level. Then the user type of the target user (or the specific customer group to which the target user belongs) can be Dazhou City-Market Category-Other Number Primary Card.

[0390] In the embodiment of the present application, since the user type of the target user is used to characterize the grouping information of the target user, the grouping information of the target user includes one or more of the location information of the target user, the perception type corresponding to the target user, the risk level of the target user, the age of the target user, and the gender of the target user. That is, the electronic device can group each user or determine the user type of each user from multiple perspectives such as location information, perception type, risk level, age, and gender, and can comprehensively and accurately determine the user group (or customer group) to which each user belongs based on the tags that may cause porting / offline, and then determine the maintenance strategies with specific service values ​​for different user groups.

[0391] Fig. 20 is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. The electronic device may be the above Figure 1A or Figure 1B application servers, analytics systems, customer experience systems, or customer devices in the environment.

[0392] like Fig. 20 As shown, the electronic device 1500 may include a memory 1501, a processor 1502, a communication interface 1503 and a bus 1504. The memory 1501, the processor 1502 and the communication interface 1503 are connected to each other through the bus 1504.

[0393] The memory 1501 may be a read-only memory (ROM), a static storage device, a dynamic storage device or a random access memory (RAM). The memory 1501 may store programs. When the programs stored in the memory 1501 are executed by the processor 1502, the processor 1502 is used to execute the above-mentioned embodiments of the present application. Figure 2A , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Fig. 9 , Fig.11 , Fig.14 The processor 1502 is also used to execute the above-mentioned method for generating a user's personalized service history and the method for generating a customer group's common service history.

[0394] Processor 1502 can adopt a general central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), a graphics processing unit (GPU) or one or more integrated circuits to execute relevant programs to implement the functions required to be performed by the units in the electronic device of the embodiment of the present application, or to execute all or part of the steps in the method of analyzing the root cause of user network portability in the method of the present application.

[0395] Processor 1502 can also be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the training method of the neural network of the present application can be completed by the hardware integrated logic circuit or software instructions in the processor 1502. The above-mentioned processor 1502 can also be a general-purpose processor, a digital signal processor (Digital Signal Processing, DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application can be directly embodied as a hardware decoding processor to execute, or the hardware and software modules in the decoding processor are combined and executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 1501, and the processor 1502 reads the information in the memory 1501, and combines its hardware to complete the functions required to be performed by the units included in the search device of the machine learning model of the embodiment of the present application, or executes all or part of the steps in the search method of the machine learning model of the method embodiment of the present application.

[0396] The communication interface 1503 uses a transceiver device such as but not limited to a transceiver to implement communication between the electronic device 1500 and other devices or a communication network.

[0397] It should be understood that each step in the above method embodiment provided by the present application can be completed by an integrated logic circuit of hardware in a processor or by instructions in the form of software. The method steps disclosed in the embodiments of the present application can be directly embodied as being executed by a hardware processor, or by a combination of hardware and software modules in a processor.

[0398] The present application also provides an electronic device, which may include: a memory and a processor, wherein the memory may be used to store a computer program; and the processor may be used to call the computer program in the memory so that the electronic device executes the method in any one of the above embodiments.

[0399] The present application also provides a computer program product, which includes: a computer program (also referred to as code, or instruction), which enables a computer to execute the method executed by the application server in any of the above embodiments when the computer program is executed.

[0400] The present application also provides a computer-readable storage medium, which stores a computer program (also referred to as code or instruction). When the computer program is executed, the computer executes the method executed by the application server in any of the above embodiments.

[0401] The various implementation modes of the present application can be combined arbitrarily to achieve different technical effects.

[0402] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part 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 the computer, the process or function according to the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. 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 a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integration. 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 hard disk Solid State Disk), etc.

[0403] Those skilled in the art can understand that to implement all or part of the processes in the above-mentioned embodiments, the processes can be completed by computer programs to instruct related hardware, and the programs can be stored in computer-readable storage media. When the programs are executed, they can include the processes of the above-mentioned method embodiments. The aforementioned storage media include: ROM or random access memory RAM, magnetic disk or optical disk and other media that can store program codes.

[0404] In short, the above are only embodiments of the technical solution of the present invention, and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made according to the disclosure of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for analyzing the root cause of user network portability, characterized in that: include: Based on the personalized service history of the user to be analyzed and the basic weights of multiple events, determining a first weight of each event in the personalized service history; The personalized service history includes multiple events arranged in the order in which the events occurred; The basic weight is used to indicate the impact of the event on the user's network portability; The first weight is used to indicate the impact of the event on the network portability of the user to be analyzed; Modifying the first weight of each event in the personalized service history to obtain a modified weight of each event in the personalized service history; Based on the modified weight of each event in the personalized service history, the root cause of the user to be analyzed switching networks is determined.

2. The method according to claim 1, characterized in that The step of modifying the first weight of each event in the personalized service history to obtain a modified weight of each event in the personalized service history includes: Based on the time interval between each event in the personalized service history, the first weight of each event is modified to obtain the second weight of each event; the time interval between each event is the interval between the occurrence time of each event and the occurrence time of the most recent event in the personalized service history; the larger the time interval between each event, the smaller the second weight of each event.

3. The method according to claim 2, characterized in that Based on the time interval of each event in the personalized service history, the first weight of each event is modified to obtain the second weight of each event, including: Establish the decay function of event heat and time interval; Determine the event heat of each event based on the decay function and the time interval of each event; The basic weight of each event is modified based on the event heat of each event to obtain a second weight of each event.

4. The method according to claim 3, characterized in that Different events correspond to different attenuation functions.

5. The method according to any one of claims 2 to 4, characterized in that: The step of modifying the first weight of each event in the personalized service history to obtain a modified weight of each event in the personalized service history further includes: Based on the interaction between two events in the personalized service history, the second weight of each event is modified to obtain a modified weight of each event.

6. The method according to claim 5, characterized in that The personalized service history includes a first event and a second event that are adjacent to each other, the first event occurs before the second event; the first event belongs to a first event type, and the second event belongs to a second event type; the modified weight W of the second event * for: W * =W+(V a +V b ) / 2*S ab Among them, the V a is the weight of the first event type, the V b is the weight of the second event type, and the weight of the event type is used to indicate the impact of the event type on the user's network portability; W is the second weight of the second event, and S ab A coefficient indicating the interaction between the first event type occurring first and the second event type occurring later.

7. The method according to claim 6, characterized in that The event type includes a plurality of event types, the plurality of event types include the first event type and the second event type, and the weights of the plurality of event types are obtained based on a hierarchical analysis model; The hierarchical analysis model includes a target layer, a criterion layer and a solution layer; the target layer is the user's tendency to switch networks, and the criterion layer includes a plurality of first features for judging whether the user has the tendency to switch networks; The scenario layer includes the plurality of event types.

8. The method according to any one of claims 1 to 7, characterized in that: The basic weight of the third event is obtained based on the average value of the weight of at least one second feature corresponding to the third event, and the third event is any event in the personalized service history of the user to be analyzed.

9. The method according to claim 8, characterized in that The weight of the second feature is obtained based on the Gini coefficient when the second feature is used as a split feature and the number of times the second feature is used as a split feature, which are calculated during training of the ensemble learning classification model; the ensemble learning classification model is trained by a training sample set, and the training sample set includes multiple first-category samples and multiple second-category samples, the first-category samples include multi-domain static event data of users who have switched networks, and the second-category samples include multi-domain static event data of users who have not switched networks, and the multi-domain static event data include multiple second features.

10. The method according to any one of claims 1 to 9, characterized in that: The method further comprises: Obtain multi-domain static event data of each user to be predicted in the set of users to be predicted; Based on the multi-domain static event data of each to-be-predicted user in the to-be-predicted user set, predict the network portability probability of each to-be-predicted user; Based on the network portability probability, multiple users to be predicted are selected from the set of users to be predicted to form a set of users to be analyzed; the set of users to be analyzed includes the users to be analyzed.

11. The method according to claim 10, characterized in that The multi-domain static event data includes at least one of network domain data, consumption domain data and service domain data, wherein the consumption domain data is used to characterize the consumption habits of the user, the network domain data is used to characterize the network quality corresponding to the user, and the service domain data is used to characterize the user perception of the user.

12. The method according to claim 11, characterized in that The set of users to be predicted includes a first user to be predicted, and predicting the network portability probability of each user to be predicted based on the multi-domain static event data of each user to be predicted in the set of users to be predicted includes: Based on the multi-domain static event data of the first to-be-predicted user, a first probability of the first to-be-predicted user switching networks is predicted.

13. The method according to claim 12, characterized in that The predicting, based on the multi-domain static event data of the first to-be-predicted user, a first probability of the first to-be-predicted user switching networks includes: Inputting the multi-domain static event data of the first to-be-predicted user into an integrated learning classification model, wherein the integrated learning classification model includes a plurality of decision trees; Predicting the first user to be predicted based on the multi-domain static event data of the first user to be predicted by each decision tree in the multiple decision trees, respectively, obtains a prediction result of each decision tree; the prediction result includes users who carry over and users who do not carry over; The ratio of the number of users whose prediction results are network porting to the total number of prediction results is used as the first probability of the first to-be-predicted user porting.

14. The method according to claim 13, characterized in that The integrated learning classification model is obtained by training a training sample set, and the training sample set includes multiple first-category samples and multiple second-category samples. The first-category samples include multi-domain static event data of users who have switched networks, and the second-category samples include multi-domain static event data of users who have not switched networks.

15. The method according to claim 12 or 13, characterized in that The first probability of the first to-be-predicted user switching networks is a probability of the first to-be-predicted user switching networks.

16. The method according to claim 12 or 13, characterized in that: The predicting the network portability probability of each to-be-predicted user in the to-be-predicted user set based on the multi-domain static event data of each to-be-predicted user includes: A second probability of the first to-be-predicted user switching networks is determined based on the matching degree between the individual service history of the first to-be-predicted user and the best matching common service history.

17. The method according to claim 16, characterized in that The network portability probability of the first to-be-predicted user is a weighted sum of a first probability of the first to-be-predicted user porting and a second probability of the first to-be-predicted user porting.

18. The method according to claim 16 or 17, characterized in that The method further comprises: Obtaining the individual service history of the first to-be-predicted user and the common service history of multiple customer groups; Matching the individual service history of the first to-be-predicted user with the common service history of the multiple customer groups to obtain a common service history that matches the individual service history of the first to-be-predicted user; Based on the occurrence time of the matching event in the individual service history of the first user to be predicted, determining the common service history that best matches the first user to be predicted from the matched common service history; Among them, the matched common service history includes the first common service history, and the matching events between the personalized service history of the first user to be predicted and the first common service history are events included in both the personalized service history of the first user to be predicted and the first common service history.

19. The method according to claim 18, characterized in that The determining, based on the occurrence time of the matching event in the individual service history of the first user to be predicted, from the matched common service history the common service history that best matches the first user to be predicted comprises: The target value S between the individual service history of the first to-be-predicted user and the first common service history is determined as: The matching events between the first to-be-predicted user's individual service history and the first common service history include n matching events, and the R n is the last matching event among the n matching events, R i is the i-th matching event among the n matching events, where i and n are positive integers, and i is not greater than n; is the R in the personalized service history of the first user to be predicted n The time of occurrence, is the R in the personalized service history of the first user to be predicted i The time of occurrence; The common service history that best matches the first to-be-predicted user is the common service history with the smallest S among the matched common service history.

20. The method of claim 19, wherein: The matching degree is the ratio of the number of all events in the common service history of the best match to the number of all events in the individual service history of the first user to be predicted; the matching degree is the second probability.

21. An electronic device, characterized in that: The electronic device comprises: Memory; Communication interface; one or more processors; Wherein, one or more computer programs are stored in the memory, and the one or more computer programs include instructions. When the instructions are executed by the electronic device, the electronic device executes the method for analyzing the root cause of user network portability as described in any one of claims 1-20.

22. A computer-readable storage medium, characterized in that: The method comprises computer instructions, which, when executed on an electronic device, enable the electronic device to execute the method for analyzing the root cause of user network portability as described in any one of claims 1 to 20.

23. A computer program product comprising instructions, characterized in that When the instruction is executed on an electronic device, the electronic device executes the method for analyzing the root cause of user network portability as described in any one of claims 1 to 20.

Images