Client loss early warning system based on behavior burying points

Abstract

The invention provides a customer loss early warning system based on behavior burying points, which relates to the field of behavior data processing, and comprises the following steps: dividing a behavior period into three sequential execution time periods, and sequentially running a behavior latent language module, a behavior judgment module and a ghost reverse verification module; the behavior latent language module converts the user interface operation event sequence into a behavior representation path to reflect a behavior intention; the behavior judgment module constructs a multi-dimensional behavior phase space, generates a disturbance path and establishes a mirror image loop as a comparison reference; the ghosting reverse verification module compares nodes which cannot be explained mutually in the two paths and marks the nodes as ghosting nodes, and a ghosting link is constructed to track an irregressive trend; the dislocation closing module only allows a ghost link to transmit a state; and when the ghosting link forms a head-tail closed structure in three continuous periods, interpretation continues and fails and a node growth rate exceeds a threshold value, determining that the ghosting link is a reverse detection chain path and triggering a structural customer loss early warning signal.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of behavior data processing, in particular to a customer churn early warning system based on behavior tracking. BACKGROUND

[0002] In the field of customer relationship management, accurately predicting customer churn risk is crucial for enterprises to maintain user stickiness and reduce customer acquisition costs. With the explosive growth of Internet user behavior data, building early warning models based on behavior tracking data has become the current mainstream technical trend. By capturing subtle changes in user interactions, potential churn signals can be identified. This field is developing towards more refined, real-time and intelligent analysis to cope with increasingly complex user behavior patterns and intense market competition.

[0003] Current technical solutions typically use rule-based engines or machine learning models to analyze user behavior sequences. For example, a typical solution sets threshold rules for key behavior events, triggering an early warning when user behavior reaches the pre-set threshold. Another common solution uses historical behavior data to train a classification model, inputting the user's recent behavior feature vector into the model to predict the churn probability. These solutions rely on statistical analysis or pattern recognition of explicit behavior events.

[0004] The shortcomings of existing technologies are that when dealing with gradual, atypical trends in user behavior, robustness is insufficient, making it difficult to effectively identify subtle signals indicating potential churn risk, leading to delayed or missed warnings. When user behavior is influenced by external environmental factors, existing models are weak in distinguishing between normal behavior fluctuations and real churn signs, leading to false positives or false negatives. SUMMARY

[0005] (I) Technical problems solved

[0006] To address the shortcomings of existing technologies, the present application provides a customer churn early warning system based on behavior tracking, to solve the problem of insufficient robustness when dealing with gradual, atypical trends in user behavior, making it difficult to effectively identify subtle signals indicating potential churn risk, leading to delayed or missed warnings. When user behavior is influenced by external environmental factors, existing models are weak in distinguishing between normal behavior fluctuations and real churn signs, leading to false positives or false negatives.

[0007] (II) Technical solutions

[0008] To achieve the above object, the application is implemented by the following technical solutions: A customer loss early warning system based on behavior tracking, comprising: the behavior tracking module is used to receive the user behavior data, define a set of tracking events triggered by user interface operation as behavior events generated based on user operation sequence, and convert them into behavior representation node sequence; the behavior representation node sequence is mapped into behavior representation path according to a set of configured behavior semantic templates that can be adjusted according to the scene, to reflect the behavior intention structure of the user; wherein the behavior semantic template is a rule set used to map behavior events into behavior intention units representing minimum behavior targets;

[0009] The behavior judgment module is used to receive the user behavior data, build a multi-dimensional behavior phase space containing time interval, jump depth and operation type as dimension definition, map behavior events to the coordinate points of the space according to event attributes and define them as phase labels representing disturbance context state, generate multiple disturbance paths for simulating the behavior sequence that the user may generate under different disturbance conditions, and build corresponding mirror loops as the benchmark for path similarity comparison based on the disturbance trajectory model; the disturbance trajectory model is a rule structure framework used to simulate behavior path change trend;

[0010] The residual reverse verification module is used to compare the behavior representation paths output by the behavior tracking module and the disturbance paths output by the behavior judgment module, and mark the nodes that cannot be structurally explained between the two as residual nodes, and build residual links for tracking the non-regressive trend path structure of the user behavior in multiple periods;

[0011] The misplacement sealing module divides each behavior period into three sequentially executed independent periods, and arranges the behavior tracking module, the behavior judgment module and the residual reverse verification module to execute in the first, second and third periods respectively; and establishes a data transmission isolation mechanism between the modules to block the direct transmission of the user behavior data between the modules, and only allows the residual link to be the state transmission path between the modules;

[0012] When the residual link forms a closed head-tail node, the structural explanation failure persists and the residual node growth rate is higher than the preset statistical threshold in three consecutive behavior periods, the link is determined as a reverse detection lock path, and a structural customer loss early warning signal is triggered through the misplacement sealing module.

[0013] Preferably, the behavior implicit module is configured to receive the user behavior data and process it as a behavior event; wherein the user behavior data is a set of embedded events triggered by user interface operations, the set of embedded events is defined as a behavior event generated based on a user operation sequence; after receiving the user behavior data, the behavior implicit module sequentially parses and converts the set of embedded events into a node sequence composed of a series of behavior representation nodes, each behavior representation node represents an independent behavior event triggered by user operations, the node sequence maintains the original behavior event order and completely maps the operation sequence; then the module performs mapping operation on the behavior representation node sequence according to a set of configured behavior semantic templates that can be adjusted according to the scene, and converts the node sequence into a behavior representation path; the behavior semantic template is a rule set that is used to map the behavior event to a behavior intent unit representing a minimum behavior target, each behavior intent unit defines a mapping rule between a behavior event and a user intent, which is used to identify the behavior intent structure implied in the user behavior; the generation process of the path is to sequentially find the matching relationship of each behavior representation node according to the rules in the behavior semantic template, if the node meets the rule condition, it will be included in the path structure of the behavior intent unit; the behavior representation path is an intent structure path formed by connecting a plurality of behavior representation nodes that meet the semantic mapping rules, which is used to reflect the complete behavior intent structure of the user; since the behavior semantic template can be adjusted according to the scene, the system can flexibly load different rule sets according to the business environment to ensure that the behavior representation path has corresponding semantic accuracy and structural integrity in various application backgrounds; the behavior representation path is used for subsequent disturbance path generation and structure comparison processing; the behavior semantic template is generated by a dynamically updated machine learning model, which learns the mapping relationship between behavior events and intent units based on a pre-trained behavior corpus through a Transformer architecture, the machine learning model receives a behavior event sequence and outputs a behavior intent vector with a confidence score, and triggers an artificial review mechanism when the confidence score is below a threshold.

[0014] Preferably, the behavior judging module is configured to perform a generation operation of a disturbance behavior path after receiving the user behavior data; the behavior judging module constructs a behavior characteristic index with time interval jump depth operation type as a dimension according to the user behavior data, and the behavior characteristic index constitutes a definition basis of the multi-dimensional behavior phase space; in this step, the behavior events are mapped into a behavior phase space constituted by three dimensions of time interval, jump depth and operation type as input objects, and a unique space coordinate point is generated according to the attribute characteristics of each behavior event, which is used to describe the behavior state of the behavior event under the disturbance semantics; then, the module defines each mapped behavior event as a phase label representing a disturbance context state based on the space coordinate point, and the phase label is used to depict the behavior disturbance characteristics of the user possibly generated under the influence of external conditions in operation; according to the phase label, the behavior judging module generates a plurality of disturbance paths, each path being a possible behavior sequence connected by adjacent disturbance states, and the possible behavior sequence simulates the behavior change trajectory of the user under different disturbance conditions; the disturbance path is constituted by the projection sequence of the behavior event in the behavior phase space and keeps the disturbance context characteristics unchanged; the disturbance path provides a disturbance reference sample for a subsequent comparison process; in addition, a corresponding mirror loop is constructed according to the disturbance path, the mirror loop is a reference path constructed according to the structural topological characteristics of the disturbance path, and serves as a benchmark structure for path similarity comparison; the mirror loop is formed by arranging the structural combination of all phase labels in the disturbance path and keeps the internal phase connection consistency; the disturbance trajectory model serves as a rule structure framework for simulating the behavior path change trend and is used to guide the construction process of the disturbance path and the mirror loop; the behavior judging module is the basis for generating the trajectory evolution structure in the disturbance path and provides a mode template of behavior disturbance to ensure that the disturbance path has systematic consistency and structural comparability; finally, the disturbance path and the mirror loop are transmitted to a subsequent module as output and participate in structural similarity judgment and residual node identification operation; wherein the disturbance trajectory model adopts a generative adversarial network (GAN) architecture, learns the spatiotemporal distribution characteristics of the historical user behavior sequence through a generator, and generates a disturbance path conforming to the real data distribution; a discriminator is used to calculate the JS divergence of the generated path and the real path, and the model is retrained when the difference exceeds the threshold value; the JS divergence: compares the behavior distribution reflected by the generated disturbance behavior path with the behavior distribution of the real user behavior path; the divergence value is calculated by statistics of the probability distribution difference of the two; after the behavior event sequences of the two paths are converted into probability distributions, the average logarithmic difference between them is calculated; the smaller the divergence value, the more similar the two behavior paths; the set basic divergence threshold is 0.15, and when the value exceeds the threshold, it is considered that there is a significant difference between the disturbance path and the real path.

[0015] Preferably, the residual reverse verification module is used to compare the behavior representation path output by the behavior covert module with the disturbance path output by the behavior judgment module in the same time period but in different module execution periods; the residual reverse verification module receives the behavior representation path and the disturbance path generated by the behavior covert module and the behavior judgment module respectively, and performs structural comparison on each behavior representation node in the two paths; the structural comparison does not rely on the original user behavior data between modules, but only performs comparison operation based on the structural node information in the behavior representation path and the disturbance path; in the path comparison process, the module analyzes the position attribute and behavior intention feature of each node in the path structure and whether it can be explained by the behavior semantic template or the disturbance trajectory model; if a node cannot be explained by the behavior semantic template or the disturbance trajectory model, it is determined to be a residual node; the residual node is a behavior representation unit between the two paths that cannot be explained by each other in structure, and is used to identify the abnormal state deviating from the model definition in the user behavior trajectory; the module then connects all residual nodes in chronological order to generate residual links for tracking the non-regressive trend in the user behavior path; the residual links are logically constructed as behavior structure trajectories spanning multiple behavior cycles, and the paths contain node sequences that cannot be reconstructed by the behavior semantic template or the disturbance trajectory model; the module further continuously tracks whether these nodes can be explained in the new behavior cycle or whether there is a persistent explanation failure phenomenon in the subsequent cycle; when the residual links show a growth trend and the number of nodes increases cycle by cycle in multiple cycles, it is prompted that there is a structural deviation; finally, when the subsequent module detects that the residual links have a closed feature and meet the conditions of persistent structural explanation failure and high residual node growth rate above the preset statistical threshold, an inverse detection lock chain path can be formed accordingly, and a structural customer churn early warning signal is triggered; and an LSTM time series prediction layer is added, inputting residual link features of continuous N cycles, N≥3, and outputting the probability of forming an inverse detection lock chain in the future cycle; when the probability is greater than 90% and the original closed condition is met, the early warning is triggered in advance; the probability calculation method of the future cycle forming an inverse detection lock chain is to use a recurrent neural network model, input the feature data of the residual link in continuous multiple behavior cycles, such as residual node growth rate, structural closure degree and other indicators; after the model is trained, a probability value between 0 and 1 is output, reflecting the possibility of forming an inverse detection lock chain in the future cycle.

[0016] Preferably, the misplacement closed module is used to divide each complete behavior cycle into three sequentially executed independent periods, and is respectively assigned to the behavior semantic module, the behavior judgment module and the residual inverse verification module, and the respective tasks are sequentially operated in the three periods to realize the stage separation of the processing flow; in the first period, the behavior semantic module receives and processes the user behavior data collected in the current behavior cycle, converts the set of buried point events into behavior events according to the defined rules to form a sequence of behavior representation nodes, and generates a behavior representation path according to the mapping of the behavior semantic template to completely reflect the behavior intention structure of the user in the current cycle; in the second period, the behavior judgment module receives the user behavior data in the same behavior cycle, and constructs the behavior feature index based on the three dimensions of time interval jump depth and operation type, maps the behavior events to the multi-dimensional behavior phase space to form a phase label, then generates a plurality of disturbance paths, and constructs a mirror loop formed by the disturbance trajectory model as a reference for subsequent path similarity comparison; in the third period, the residual inverse verification module compares the behavior representation path and the disturbance path output by the behavior semantic module and the behavior judgment module, judges whether there is a behavior representation node that cannot be explained by each other in the structure of the two paths in the same behavior cycle, and marks it as a residual node and constructs a residual link for multi-cycle tracking; under the division of the three periods, the misplacement closed module establishes a module-level data transmission isolation mechanism, explicitly blocks the direct transmission path of the user behavior data between the modules, and any module cannot access the behavior data representation node path results or intermediate state variables generated by other modules, and only allows the residual link to act as the only state transmission path to play the role of information bridging between modules; when the residual link forms a head-tail node closed structure in three consecutive behavior cycles and the growth rate of the residual node exceeds the preset statistical threshold, the system determines that the link is a reverse inspection lock path and triggers a structural customer loss warning signal through the module to complete the cycle processing closed loop; the intermediate state variable refers to the temporary data structure generated in the operation process of each module, including: the behavior event after the analysis of the original set of buried point events of the behavior semantic module, the sequence of behavior representation nodes generated by the conversion of the behavior event, i.e. the node set before being mapped to the path; the phase label generated by the mapping of the behavior event to the behavior phase space of the behavior judgment module, the set of disturbance state coordinate points before the construction of the disturbance path, and the intermediate result of the path topology when the mirror loop is generated; the residual node candidate set in the node comparison process of the residual inverse verification module, and the temporary cache of the semantic equivalent fragment generated when the bidirectional structure is reconstructed.

[0017] Preferably, the residual reverse verification module, in the process of performing structural comparison between the behavior representation path output by the behavior implicit language module and the disturbance path output by the behavior judgment module, when identifying that the behavior representation node cannot achieve structural interpretation in both paths within the same time period, marks the behavior representation node as a residual node and constructs a residual link based on it to track the non-regressive trend path structure existing in the user behavior cycle by cycle, the residual link records the structural position of the residual node in each behavior cycle and the comparison failure condition during the formation process, when the residual node appears in the same structural position in three consecutive behavior cycles and continues to fail to be interpreted by the behavior semantic template and the disturbance trajectory model rule, the misplacement closure module confirms that the residual link has formed a closed structure accordingly, and further judges whether it constitutes a reverse lock chain path, if the following three conditions are met at the same time, it is considered that the residual link constitutes a reverse lock chain path, the first condition is that the residual link is closed at the beginning and end nodes in three cycles, that is, the path between the starting node and the ending node of the link forms a periodic cycle structure, the second condition is that the residual node continues to be unexplained by the behavior semantic template or the disturbance trajectory model during the structure closure, the third condition is that the growth rate of the residual node is higher than the preset statistical threshold, for example, if the growth rate of the total amount of residual nodes in each cycle exceeds 20%, the lock chain determination establishment logic is triggered; when the three conditions are met, the misplacement closure module immediately triggers a structural customer loss warning signal; the structural customer loss warning signal includes the following contents: user identification; behavior representation path and residual link in the current behavior cycle; specific node sequence determined as a reverse lock chain path; risk level and recommended intervention strategy generated by the system; the warning signal can be received by the upper customer operation system, and automatically triggers the customer retention mechanism, including active contact, targeted incentive and other operation schemes.

[0018] Preferably, the structural compression format is a standardized data structure after removing redundant fields; the rule coverage is a matching degree quantification index of the behavior event sequence and the preset behavior semantic template, and the calculation formula is: coverage=(number of successfully mapped nodes by the template rule / total number of path nodes)×100%; the weighted structure comparison is that the comparison priority of the structural compression failure segment is higher than that of the ordinary node, which is reflected by the weight coefficient: comparison priority weight=base weight×position coefficient; base weight: the preset value of the ordinary node is 1; the failure segment node is promoted to 2; position coefficient: the starting segment of the path is 1.5; the terminal segment of the path is 1.3; the middle segment of the path is 1.0; the structural matching analysis is to traverse each node in the behavior representation path and compare it with the rules in the preset behavior semantic template item by item: which contains node-level matching, path segment continuity detection and bidirectional reconstruction verification; the node-level matching is to extract the key fields of the node, such as event type, page level, timestamp, and calculate the similarity with the pre-defined behavior intent unit trigger condition in the template, such as cosine similarity≥0.7 is considered to match; the path segment continuity detection is to detect whether the length of the region composed of consecutive unmatched nodes exceeds the fault tolerance threshold of 3 nodes, and verify whether the region is located at the beginning or end of the path; the bidirectional reconstruction verification is to generate an ideal behavior representation path segment based on the behavior semantic template for the residual node, and generate a disturbance path segment based on the disturbance trajectory model, if the structural topology difference degree of the residual node and any generated segment, such as edit distance> preset threshold; or the intent semantic distance exceeds the tolerance range, such as word vector cosine similarity<0.6, then it is determined as an unmergeable node.

[0019] Preferably, the misplacement closed module divides each behavior cycle into three non-overlapping execution periods in the system structure, for accurately scheduling the sequential operation of the behavior cipher module, the behavior judgment module and the residual reverse verification module; wherein the first period is the exclusive execution period of the behavior cipher module, in which the user behavior data is input into the behavior cipher module, which is responsible for identifying and defining the set of buried point events triggered by user interface operation as behavior events generated based on user operation sequence, and then converting them into behavior representation node sequence; then mapping according to the configured behavior semantic template to obtain the corresponding behavior representation path, and completing the storage of the path structure for subsequent comparison and analysis; after the end of the first period execution, the system automatically switches to the second period, and the behavior judgment module takes over the operation process, which uses the user behavior data in the same cycle as input, maps each behavior event to the behavior phase space by constructing multi-dimensional behavior feature indicators including time interval, jump depth and operation type, and labels the corresponding phase label, thereby generating a disturbance path; then the disturbance trajectory model generates a mirror loop as a behavior prediction simulation benchmark; after the second period is completely finished and the path output is completed, the system switches to the third period, and the residual reverse verification module receives the two kinds of path data generated by the first and second modules, and compares the structure of the behavior representation nodes, and then identifies the non-merging nodes and residual links; at the same time, the misplacement closed module ensures that the behavior data does not occur directly between the modules through the data transmission isolation mechanism between the modules, that is, the behavior cipher module, the behavior judgment module and the residual reverse verification module are prohibited from accessing the input data and intermediate state variables of other modules during operation, and only the residual link is allowed as the only inter-module state transmission path, ensuring that the functions of each module are independent, the data is isolated, the process is clear, and the behavior information is prevented from crossing and interfering between the modules, thereby ensuring the accuracy and reliability of the structural customer loss warning.

[0020] Preferably, the misplacement closure module further sets a module level data isolation mechanism to ensure that the behavior cryptogram module, the behavior judgment module and the residual inverse verification module strictly limit access to user behavior data during their respective operations; wherein the data isolation mechanism effectively prevents the original input user behavior data or the behavior representation node sequence, behavior representation path, disturbance path and intermediate state variable generated during operation from being called by any unauthorized module by blocking the direct transmission of behavior data among the above three modules; under this mechanism, only the residual inverse verification module is authorized to access the behavior representation path output by the behavior cryptogram module and the disturbance path generated by the behavior judgment module at the same time, and after obtaining the data of the two paths, it performs structural position comparison and semantic label comparison on the behavior representation nodes therein; the specific comparison process first extracts the structural position of each behavior representation node in the path, and analyzes the semantic label of the node in combination with the natural language processing model; then the structural position and semantic label are compared in the two paths respectively to identify those nodes that are misaligned in structural position or cannot be corresponded in semantic label; such nodes are defined as residual nodes; after identifying the residual nodes, the residual inverse verification module further performs a structure reconstruction operation, direction one is to reconstruct in the direction of the behavior representation path according to the behavior semantic template, and direction two is to reconstruct in the direction of the disturbance path according to the disturbance trajectory model; if a behavior intention unit with semantic equivalence relationship cannot be generated in any direction reconstruction process, the corresponding node is marked as a non-merging node; at the same time, the system analyzes the logical causal relationship of all non-merging nodes and their time sequence continuity in the path, and constructs the residual link as the only structure path that allows state transmission between modules, thereby realizing the closure of the state transmission process between modules, and on this basis, subsequent structural customer churn trend analysis is performed.

[0021] Preferably, when the residual link is marked as the reverse lock chain path by the misplacement closed module; the misplacement closed module immediately enters the state freezing mechanism; the freezing mechanism performs version locking on the behavior semantic template in the behavior implicit module for generating the behavior representation path and the disturbance trajectory model in the behavior judgment module for constructing the disturbance path; during the freezing period, the behavior semantic template and the disturbance trajectory model remain the current version unchanged, and any rule update is suspended; the system only reconstructs the user behavior characteristics according to the closed structure formed in the reverse lock chain path, and outputs the structural customer loss warning signal; the structural customer loss warning signal content includes user identification, behavior representation path in the current behavior period, residual link, node sequence determined as the reverse lock chain path, and risk level and suggested intervention strategy generated by the system; the signal serves as the trigger input of the upper customer operation system, and can be used for automatically enabling the customer retention mechanism, such as the active contact or the directional incentive operation scheme; if no new non-merging node appears in the three continuous behavior periods after freezing, the system automatically removes the freezing state, and restores the rule update function of the behavior semantic template and the disturbance trajectory model; through the above process, the system quickly freezes the key decision rules after identifying the structural abnormal behavior, so as to prevent model drift, and timely restores the normal operation after the abnormality disappears, thereby maintaining the stability and accuracy of the warning mechanism.

[0022] Preferably, the misplacement closure module divides each behavior cycle into three consecutive non-overlapping execution periods during system operation; the division ensures that the behavior code module, the behavior judgment module and the residual inverse verification module run independently in physical time sequence and execute in turn within their respective specified periods; the specific operation logic is: in the first period of the behavior cycle, the behavior code module receives the user behavior data in this cycle and converts it into behavior events, and then maps it into a sequence of behavior representation nodes through structural compression format; the node sequence is added to the constructed behavior representation path in turn according to the behavior event occurrence order, and does not share or transmit intermediate states with any other module; the second period is executed by the behavior judgment module, which reads the corresponding user behavior data and extracts indicators such as behavior event interval duration, jump depth and operation type, constructs the multi-dimensional behavior phase space, and maps the events into phase labels to generate the perturbation path and the corresponding mirror loop; this process also does not allow it to access the behavior representation path or node sequence generated in the first period; in the third period, the residual inverse verification module is executed, which is the only module allowed to access the behavior representation path generated in the first period and the perturbation path generated in the second period; in this stage, the residual inverse verification module performs structural position comparison and semantic matching analysis on the behavior representation nodes in the two paths, identifies the residual nodes, and on this basis performs bidirectional structural reconstruction and residual link generation; During the entire scheduling process, the system ensures that the three modules run independently in independent periods through physical isolation means and task scheduling strategies, eliminates the direct transmission of user behavior data and intermediate variables between modules, and only allows the residual link as the only cross-module state transmission path, thereby ensuring the isolation and controllability of the overall system scheduling process and providing a stable data foundation for subsequent structural customer loss analysis.

[0023] Preferably, the misplacement closed module sets a module level data isolation mechanism to strictly block the direct transmission channel of the user behavior data between the behavior cryptogram module, the behavior judgment module and the residual inverse verification module; this mechanism prohibits any module from accessing the user behavior data, intermediate state variables, behavior representation node sequence, behavior representation path or disturbance path in the processing process of other modules; specifically, the behavior cryptogram module only obtains the original user behavior data in the first period of its operation and independently completes the behavior event conversion and behavior representation path construction process; after completion, no intermediate processing result is output to other modules; the behavior judgment module starts in the second period, independently extracts event attributes from the original user behavior data and independently completes the generation of the disturbance path and the mirror loop, and its running process does not depend on any data structure or state variable generated by other modules; the residual inverse verification module, as the only module authorized to access the behavior representation path and the disturbance path at the same time, receives and loads the two path structures during the third period of operation, and only uses the structure information and the node semantic label generated by NLP model analysis to compare the structural positions of the behavior representation nodes of the two paths; in the comparison process, the residual inverse verification module relies on the behavior semantic template to reconstruct the structure in the direction of the behavior representation path, and relies on the disturbance trajectory model to reconstruct the structure in the direction of the disturbance path; if the behavior intention segment equivalent to the residual node semantics cannot be generated in any direction, the node is marked as a non-merging node and added to the residual link; the system identifies the link structure according to the connection trend of non-merging nodes in terms of logical causality and time sequence, and takes the residual link as the only state transmission path between the three modules; this mechanism ensures the logical independence between modules and the auditability of system operation through complete data isolation and structure connection transmission, prevents the risk of path pollution or abnormal spread caused by data sharing, thereby providing a solid logical foundation and technical support for the customer loss warning mechanism; the content transmitted by the only state transmission path is the structured data of the residual link, including the node ID, timestamp, structure position, semantic label analysis result of the non-merging node sequence; the time sequence continuity mark and causal dependence identifier of the logical relationship between nodes; and the periodic statistical value of the link metadata, i.e. the number of newly added nodes and the growth rate mark of the current period.

[0024] (III) Beneficial effects

[0025] The present application provides a customer loss warning system based on behavior burying. It has the following advantages

[0026] Beneficial effects:

[0027] 1、The application converts the user operation sequence into an interpretable behavior representation path through the behavior implicit language module, constructs the behavior evolution trajectory in the disturbance scene by combining the behavior judgment module to build a multi-dimensional behavior phase space, realizes the deep structural analysis of the user behavior intention, accurately identifies the residual shadow node that cannot be explained by the behavior semantic template or the disturbance trajectory model by using the residual shadow reverse verification module, constructs the residual shadow link to capture the non-returnable behavior trend, forcibly operates the three core modules independently in time periods by means of the time period isolation mechanism and the data transmission blockage of the misplacement closed module, only allows the residual shadow link to transfer the state across the modules, eliminates the data coupling interference, triggers the structural loss early warning when the residual shadow link appears a closed loop structure for three consecutive periods and the node growth exceeds the threshold, and significantly improves the reliability and early warning timeliness of the abnormal behavior detection.

[0028] 2、The application is based on the time period division and data isolation mechanism of the misplacement closed module, ensures that the behavior representation path generation, disturbance path simulation and residual shadow verification three stages are independently executed, blocks the cross-module pollution of the original data and intermediate variables, fundamentally avoids the analysis deviation, continuously captures the non-merging node through the cross-period tracking function of the residual shadow link, verifies the unexplainability of the node by combining the bidirectional structure reconstruction, and accurately locates the structural continuity fracture risk. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to make the content of the application more easily understood, the application will be further described in detail below according to the specific embodiments of the application and in combination with the drawings, in which:

[0030] Figure 1 is a flow chart of a customer loss early warning system based on behavior burying points of the application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the application will be clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0032] The embodiment of the application provides a customer loss early warning system based on behavior burying points, comprising, in the embodiment, deployment and running of the customer loss early warning system based on behavior burying points cover all behavior data collection, behavior semantic structure generation, disturbance behavior simulation, residual image comparison and customer loss early warning stages of a user from accessing a homepage to exiting a platform, and specific implementation steps are as follows, when the user logs in the platform and performs clicking, sliding, jumping and other operations on a webpage interface, user behavior data is collected in real time through a front-end burying point mechanism, a behavior implicit language module in the system first receives and analyzes the batch of user behavior data in a set behavior period first time period, the module arranges the events into user operation sequences according to preset burying point event types such as page clicking, video playing and course collecting, then converts the operation sequences into a structure compression format, standardizes after removing redundant fields, generates a group of behavior events, subsequently, the behavior events are sequentially mapped into behavior representation nodes according to actual occurrence order, and a behavior representation path is constructed, after the path is generated, a configured behavior semantic template is immediately called for structure matching analysis, the behavior semantic template is used for mapping the behavior events into corresponding behavior intention units such as 'completing a trial listening' and 'attempting payment', the system calculates the rule coverage degree of the path through three ways of node level matching, path continuity detection and bidirectional structure reconstruction, if there are three or more nodes that are not covered by the template in a path starting or ending segment, the segment is marked as a structure compression failure segment, and is given a higher comparison weight in subsequent processing;

[0033] In the second time period of the behavior period, the behavior judgment module starts running, the module extracts features such as event interval, jump depth and operation type based on the received user behavior events, and maps the features into coordinate points of a multi-dimensional behavior phase space to constitute phase labels, subsequently, a disturbance path is constructed according to a disturbance trajectory model, the path simulates behavior change trends that the user may take under disturbance conditions such as abnormal clicking, page lag and loading delay, and generates an image loop according to the same model as a path similarity comparison standard, after the above processing, the module does not output intermediate variables, and only generates final disturbance path data;

[0034] In the third period, the residual reverse check module is started; it obtains the behavior representation path and the disturbance path from the behavior hidden language module and the behavior judgment module, respectively, and performs structural comparison analysis on the behavior nodes in the same period of the two paths; the system first extracts the structural position, semantic label, and timestamp information of each node, calculates the difference nodes through structural position matching and semantic similarity evaluation, and screens out the nodes that cannot be explained by the behavior semantic template and the disturbance trajectory model as residual nodes; if the node cannot be matched to the semantic equivalent path segment in the bidirectional structure reconstruction process, it is marked as an unmerging node and added to the residual link; the system further checks whether the residual node causes continuous node loss in the behavior representation path; if so, it is defined as a structural continuity fracture and the fracture link is recorded; if the residual link has the same unmerging node in the same structural position in three consecutive behavior periods, and these nodes form a closed loop structure with a node interval of not more than one period, and cannot be explained by the semantic template or the disturbance model in all periods, the link is marked as a reverse check lock path;

[0035] At this time, the misplacement closure module immediately triggers a structural customer loss warning signal; the module will freeze the current version of the behavior semantic template and the disturbance trajectory model to prevent its update in the next behavior period, so as to ensure the consistency and traceability of the warning judgment; under the frozen state, the system only uses the reverse check lock path formed as input to reconstruct the behavior feature map of the user; and combines factors such as node growth rate, rule coverage, and residual density to comprehensively evaluate the customer loss risk level; when the risk level exceeds the medium-high risk threshold, the system will send a warning signal to the upper-layer customer operation platform; the warning signal includes user identification, behavior representation path in the current period, residual link, reverse check lock node sequence, system judgment result, and recommended intervention strategy and other detailed information; after receiving the warning signal, the operation platform can trigger customer retention mechanisms such as sending course coupons, customer service active contact, and recommending popular courses to improve user activity and reduce loss risk; if no new unmerging node is added in the next three behavior periods, the frozen state is automatically released, and the update function of the behavior semantic template and the disturbance trajectory model is restored, ensuring the continuous dynamic learning ability of the system.

[0036] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A customer churn early warning system based on behavioral tracking, characterized in that, include: The behavioral cryptic module is used to receive user behavior data, define the set of embedded events triggered by user interface operations as behavioral events generated based on user operation sequences, and convert them into behavioral representation node sequences. The behavioral representation node sequences are mapped to behavioral representation paths according to a set of configured behavioral semantic templates that can be adjusted according to the scenario, so as to reflect the user's behavioral intent structure. The behavioral semantic template is a set of rules used to map behavioral events into behavioral intention units that represent the smallest behavioral goals; The behavior judgment module receives the user behavior data, constructs a multi-dimensional behavior phase space defined by behavior feature indicators including time interval, jump depth, and operation type, maps behavior events to coordinate points in this space according to event attributes and defines them as phase labels representing the perturbation context state, generates multiple perturbation paths to simulate the behavior sequences that the user may generate under different perturbation conditions, and constructs corresponding mirror loops based on the perturbation trajectory model as a benchmark for path similarity comparison; the perturbation trajectory model is a rule structure framework used to simulate the changing trend of behavior paths. The afterimage reverse verification module is used to compare the behavior representation path output by the behavior cryptic module and the disturbance path output by the behavior judgment module, where the behavior representation nodes generated in the same time period but during the execution periods of different modules are in the same time period; nodes that cannot be structurally interpreted between the two are marked as afterimage nodes, and afterimage links are constructed to track the irreversible trend path structure of the user behavior in multiple periods; The staggered closure module divides each behavior cycle into three independent time periods that are executed sequentially, and the behavior hidden message module, the behavior judgment module, and the afterimage reverse verification module are executed in the first, second, and third time periods respectively. A data transmission isolation mechanism is established between modules to block the direct transmission of user behavior data between modules, and only the afterimage link is allowed as a state transmission path between modules. When the afterimage link forms a closed first and last node in three consecutive behavioral cycles, the structural interpretation failure persists, and the growth rate of the afterimage node is higher than the preset statistical threshold, the link is determined to be a reverse detection chain path, and a structural customer churn warning signal is triggered through the misalignment closure module.

2. The customer churn early warning system based on behavioral tracking as described in claim 1, characterized in that: In the process of constructing the behavior representation path, the behavior metaphor module converts each behavior event into a structured compression format and maps it to the behavior representation node. The behavior representation nodes are added to the behavior representation path in the order in which the behavior events occur. After the construction is completed, the behavioral semantic template is called to perform a rule-based structural matching analysis on the behavioral representation path to determine whether there are continuous non-matching node regions that the behavioral semantic template cannot cover. If the non-matching nodes are located at the beginning or end of the path and more than three appear consecutively, the non-matching node regions are marked as structural compression failure segments, and weighted structural comparisons are performed on the region in the subsequent afterimage reverse verification module.

3. A customer churn early warning system based on behavioral tracking as described in claim 1, characterized in that: After receiving user behavior data, the behavior judgment module performs dimensional identification on the interval duration and jump depth of the behavior event, and after obtaining the identification results, maps them to a multi-dimensional behavior phase space to form a phase label. The perturbation path is constructed based on the phase label to simulate the behavioral evolution trend under the perturbation scenario, and a mirror loop is constructed based on the perturbation trajectory model as a reference path. In multiple behavior cycles, the actual user behavior representation path and the structure of the mirror loop are compared. If the overlap is lower than the statistical threshold and no periodic similar structure is detected, it is marked as a track divergence state, and the relevant path structure is passed to the afterimage reverse verification module to perform structural conflict determination.

4. A customer churn early warning system based on behavioral tracking as described in claim 1, characterized in that: The afterimage reverse verification module performs structural matching analysis on the behavioral representation nodes in the two paths to identify afterimage nodes; and performs bidirectional structural reconstruction on the afterimage nodes based on behavioral semantic templates towards the behavioral representation path and based on perturbation trajectory models towards the perturbation path: direction one is reconstruction based on behavioral semantic templates towards the behavioral representation path; direction two is reconstruction based on perturbation trajectory models towards the perturbation path; if neither direction can generate a segment semantically equivalent to the behavioral intent unit of the afterimage node, it is marked as an unmergeable node and included in the afterimage link; further, it is determined whether the unmergeable node has a continuous node missing in the behavioral representation path, and if so, it is defined as a structural continuity break.

5. A customer churn early warning system based on behavioral tracking as described in claim 4, characterized in that: When the afterimage link exhibits afterimage nodes with the same structural position in three consecutive behavioral cycles, and the non-mergeable nodes form a closed-loop structure in the path network with a first-to-last node interval not exceeding a preset cycle threshold and having an event dependency relationship, and are not mutually interpreted by the behavioral semantic template and the rules of the disturbance trajectory model in subsequent cycles, the closed-loop structure is identified as a reverse detection chain path; when the proportion of non-mergeable nodes in the closed-loop structure exceeds a preset threshold, and the growth rate of the total number of link nodes exceeds a preset statistical threshold for three consecutive cycles, the misalignment closure module triggers a structural customer churn warning signal, suspends the conventional warning mechanism, and instead executes the closed-loop structure analysis process based on the reverse detection chain path, and pushes a high-priority processing signal to the management end; the closed-loop structure is defined as a first-to-last node interval of less than or equal to one behavioral cycle.

6. A customer churn early warning system based on behavioral tracking as described in claim 1, characterized in that: The misalignment closure module divides each behavior cycle into three consecutive non-overlapping execution periods. The behavior cryptic module runs in the first period and processes the user behavior data within that period to generate the behavior representation path. The behavior judgment module runs in the second period and constructs the perturbation path. The afterimage reverse verification module runs in the third period, receives the behavior representation path and the perturbation path, performs structural node comparison, and generates the afterimage link. Furthermore, the misalignment closure module prohibits any module from accessing the original input user behavior data, behavior representation node sequence, behavior representation path, perturbation path, and intermediate state variables generated during operation.

7. A customer churn early warning system based on behavioral tracking as described in claim 6, characterized in that: The misaligned closure module sets up a module-level data isolation mechanism to block the direct transmission of user behavior data between the behavior cryptic module, behavior judgment module, and afterimage reverse verification module. It prohibits any module among the behavior cryptic module, behavior judgment module, and afterimage reverse verification module from accessing the user behavior data or intermediate state variables of other modules. The afterimage reverse verification module is the only analysis node that is allowed to access both the behavior representation path and the disturbance path simultaneously. It only uses the behavior representation path and the disturbance path to compare the structural position of the behavior representation node with the semantic labels parsed by the NLP model to identify afterimage nodes. Based on the behavior semantic template, it performs structural reconstruction in the direction of the behavior representation path and based on the disturbance trajectory model in the direction of the disturbance path to mark non-mergeable nodes. It also tracks the structural connection trend between the non-mergeable nodes in terms of logical causal relationship and temporal continuity, using the afterimage link as the only state transmission path.

8. A customer churn early warning system based on behavioral tracking according to claim 7, characterized in that: When the afterimage link is marked as a reverse detection chain path, the misalignment closure module immediately triggers a state freeze mechanism, locking the behavioral semantic template used to generate the behavioral representation path in the behavioral metaphor module and the perturbation trajectory model used to construct the perturbation path in the behavioral judgment module. The current versions of the behavioral semantic template and the perturbation trajectory model remain unchanged and are suspended from updating during the freeze period. During the freeze period, the user behavior feature map is reconstructed based solely on the closed structure formed in the reverse detection chain path, and the structural customer churn warning signal is output based on the reconstruction result. If no new non-mergeable nodes are added within three consecutive behavior cycles, the freeze state will be automatically lifted and the module update function will be restored.

Images