A funnel-based conversion analysis method for student recruitment based on student segmentation
By analyzing the behavioral characteristics of students in different groups and mapping funnel nodes, and combining data from the enrollment and student ends, a conversion obstacle matrix is generated, and targeted marketing tasks are automatically executed. This solves the problem of not being able to accurately locate conversion anomalies in existing methods, and achieves efficient automated correction and conversion rate improvement.
Patent Information
- Application Number
- CN202611136556.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-11-17
AI Technical Summary
Existing conversion analysis methods cannot accurately pinpoint the reasons for abnormal decline in specific customer groups at certain stages of business flow. This leads to excessive reliance on subjective human judgment for corrective interventions, making it impossible to achieve automated and customized corrective execution, resulting in the loss of high-potential customers and waste of resources.
By collecting front-end interaction logs of potential students, using time-weighted clustering algorithms to group students, mapping them to the enrollment funnel node sequence, calculating conversion dwell time, and combining records from the enrollment and student ends, a conversion obstacle driver matrix is generated, automatically matching correction and recovery instructions and executing targeted marketing tasks.
It enables precise capture of local abnormal stagnation points in specific groups, in-depth analysis of the pain points in matching the system with students' needs, automated execution of correction tasks, improved conversion data indicators, and reduced recovery failures caused by delayed manual intervention.
Smart Images

Figure CN123118313A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a method for hierarchical conversion analysis of the enrollment funnel based on the grouping of prospective students. Background Technology
[0002] As educational institutions accelerate their digital transformation, they accumulate massive amounts of interaction logs and behavioral details of potential students across daily operations and online promotions. In the actual student recruitment and conversion process, accurately tracking the follow-up process of potential students and implementing targeted retention actions is crucial for reducing marketing costs and increasing conversion rates. To monitor student conversion effectiveness, relevant business systems have widely deployed enrollment funnel models to present the statistical distribution of each stage of the process.
[0003] Existing conversion analysis methods typically employ a globally uniform static funnel model to process all lead data, mixing student samples with different needs to calculate the overall conversion rate. This crude statistical approach masks the true reasons for abnormal declines in specific attribute groups at local business flow stages. Furthermore, after locating conversion stagnation points, existing methods can only output surface-level churn distribution data, failing to perform penetrating calculations and comparisons based on the matching degree between the characteristics of materials distributed at the recruitment end and the actual preferences of students. This leads to excessive reliance on subjective human judgment for corrective interventions at stagnation points, failing to automatically form customized corrective execution loops targeting different obstacles. Ultimately, this results in a significant loss of high-potential customers and ineffective consumption of network communication resources. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a hierarchical conversion analysis method for student enrollment funnel based on the grouping of prospective students, thus solving the problems mentioned in the background technology.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a hierarchical conversion analysis method for enrollment funnel based on the grouping of prospective students, comprising the following steps: S1. Collecting front-end interaction behavior logs generated by prospective students at business touchpoints, extracting page browsing depth sequences and dialogue response time intervals as behavioral feature data; inputting the behavioral feature data into a time-weighted clustering algorithm for clustering calculation, outputting the intention preference type and the corresponding grouped student list; S2. Mapping the grouped student list to a preset enrollment funnel node sequence, calculating the actual conversion dwell time of the grouped student list in each single funnel node; comparing the actual conversion dwell time with a baseline time threshold associated with the intention preference type, extracting single funnel nodes whose actual conversion dwell time exceeds the baseline time threshold as abnormal break levels, and summarizing prospective students remaining in the abnormal break levels to generate a high-risk stagnation list; S3. S4. Retrieve enrollment follow-up behavior records and student feedback records within the time period corresponding to the abnormal break level of the high-risk stagnation list. Extract the feature tags of external marketing materials from the enrollment follow-up behavior records, and calculate the semantic similarity distance between the feature tags of external marketing materials and the intention preference type to obtain the text semantic misalignment degree. Extract the number of unread messages from the student feedback records, and combine the text semantic misalignment degree and the number of unread messages into a matrix to generate a conversion obstacle driver matrix. S5. Extract specific business misalignment types from the conversion obstacle driver matrix, and query the pre-built intervention action library for correction and recovery instructions that match the specific business misalignment types. Encapsulate the correction and recovery instructions into an automatic intervention task containing targeted marketing scripts and discount push conditions, call the message push interface of the front-end business system, and send targeted marketing scripts and discount push conditions to the terminals of potential students in the high-risk stagnation list according to the automatic intervention task.
[0006] Furthermore, the specific process of collecting front-end interaction behavior logs generated by prospective students at business touchpoints and extracting page browsing depth sequences and dialogue response time intervals as behavioral feature data is as follows: Parse the event tracking data of prospective students on the enrollment system interface and customer service chat window from the front-end interaction behavior logs; arrange the page access nodes contained in the event tracking data according to the timestamp order; extract the hierarchical attributes of the page access nodes to construct the page browsing depth sequence; retrieve the timestamps of customer service inquiries and prospective student responses contained in the customer service chat window; perform timestamp difference extraction; use the difference extraction result as the dialogue response time interval; and simultaneously extract the page browsing depth sequence and dialogue response time intervals to generate behavioral feature data.
[0007] Furthermore, the specific process of inputting behavioral feature data into a time-weighted clustering algorithm for clustering calculation, and outputting intention preference types and corresponding student group lists is as follows: Extract the occurrence timestamps associated with the behavioral feature data, measure the time span between the occurrence timestamps and the current system time, and assign time decay weights to the behavioral feature data according to the time span; introduce time decay weights to weight the behavioral feature data to obtain a weighted feature matrix, input the weighted feature matrix into the cluster center node of the time-weighted clustering algorithm to perform iterative clustering, output basic data clusters with similar feature dimensions, perform semantic mapping between the basic data clusters and a pre-set intention preference label library, and output intention preference types; classify and summarize the unique identifiers of the intention students that generated the behavioral feature data according to the intention preference types, and output student group lists.
[0008] Furthermore, the specific process of mapping the student group lists to the pre-set enrollment funnel node sequence and calculating the actual conversion dwell time of the student group lists in each individual funnel node is as follows: Parse the unique identifiers of the prospective students contained in the student group lists, match the business flow status records of the prospective student unique identifiers in the front-end business system, align the business flow status records with the individual funnel nodes contained in the enrollment funnel node sequence, and determine the node transfer-in timestamp and node transfer-out timestamp corresponding to the individual funnel node; extract the time span difference between the node transfer-out timestamp and the node transfer-in timestamp, and confirm the time span difference as the actual conversion dwell time. When there is a missing node transfer-out timestamp, extract the time span difference between the current system timestamp and the node transfer-in timestamp to confirm the actual conversion dwell time.
[0009] Furthermore, by comparing the actual conversion dwell time with the baseline time threshold associated with the intention preference type, the process of extracting single funnel nodes whose actual conversion dwell time exceeds the baseline time threshold as abnormal break levels and summarizing the intention students remaining in the abnormal break levels to generate a high-risk stagnation list is as follows: Read the intention preference type, query the baseline time threshold matching the intention preference type in the preset funnel timeliness configuration table, compare the actual conversion dwell time with the baseline time threshold, and record the single funnel node whose actual conversion dwell time exceeds the baseline time threshold; mark the single funnel node whose actual conversion dwell time exceeds the baseline time threshold as an abnormal break level, extract the unique identifier of the intention students remaining in the abnormal break level, and combine the unique identifiers of the intention students to generate a high-risk stagnation list.
[0010] Furthermore, the specific process of retrieving enrollment follow-up behavior records and student feedback records within the time period corresponding to the abnormal break level of the high-risk stagnation list, extracting the feature tags of outbound marketing materials from the enrollment follow-up behavior records, and calculating the semantic similarity distance between the outbound marketing material feature tags and the intention preference type to obtain the text semantic misalignment degree is as follows: Receive the unique identifiers of the intended students included in the high-risk stagnation list; delineate the time window associated with the abnormal break level in the enrollment system database; based on the unique identifiers of the intended students and the time window, retrieve and retrieve the enrollment follow-up behavior records and student feedback records; parse the outbound action details included in the enrollment follow-up behavior records; extract the material metadata bound to the outbound action details; convert the material metadata into outbound marketing material feature tags; map the outbound marketing material feature tags and intention preference types to a pre-set business semantic vector space, and output the corresponding material feature vector and preference type vector; perform the spatial distance measurement calculation between the material feature vector and the preference type vector, and confirm the spatial distance measurement result as the text semantic misalignment degree.
[0011] Furthermore, the specific process of extracting the number of unread messages from student feedback records and combining the text semantic misalignment degree with the number of unread messages to generate a conversion obstacle driving factor matrix is as follows: Scan the instant messaging interaction status bits contained in the student feedback records, summarize the frequency of the unreachable instant messaging interaction status bits to obtain the number of unread messages, introduce a pre-set misalignment level classification rule to discretize and classify the text semantic misalignment degree to be processed, generate misalignment level coordinate parameters, introduce a pre-set resistance response level classification rule to discretize and classify the number of unread messages to be processed, generate resistance level coordinate parameters; establish a two-dimensional correlation analysis architecture containing horizontal and vertical axes, assign the misalignment level coordinate parameters to the horizontal axis, assign the resistance level coordinate parameters to the vertical axis, map the misalignment level coordinate parameters and resistance level coordinate parameters to the corresponding cross nodes in the two-dimensional correlation analysis architecture, and output the conversion obstacle driving factor matrix containing the cross nodes.
[0012] Furthermore, the specific process of extracting specific business misalignment types from the conversion obstacle driver matrix and querying the pre-built intervention action library for correction and recovery instructions matching specific business misalignment types is as follows: traverse the intersection nodes contained in the conversion obstacle driver matrix, locate the extreme values of the two-dimensional coordinates pointed to by the intersection nodes, determine the specific business misalignment type based on the extreme values of the two-dimensional coordinates, and parse the business pain point field and follow-up error field contained in the specific business misalignment type; input the business pain point field and follow-up error field into the retrieval engine of the pre-built intervention action library for feature comparison, extract the set of action strategies in the pre-built intervention action library whose feature comparison consistency meets the preset threshold, and confirm the set of action strategies as correction and recovery instructions.
[0013] Furthermore, the process of encapsulating the correction and recovery instruction into an automatic intervention task containing targeted marketing scripts and discount push conditions, and calling the message push interface of the front-end business system to send the targeted marketing scripts and discount push conditions to the terminals of potential students on the high-risk stagnation list according to the automatic intervention task is as follows: Parse the trigger timing parameters, targeted marketing scripts, and discount push conditions contained in the correction and recovery instruction; according to the preset business system data exchange protocol, assemble the trigger timing parameters, targeted marketing scripts, and discount push conditions into an automatic intervention task data packet with a standard structure format; initiate a call request carrying the automatic intervention task data packet to the message push interface of the front-end business system; monitor the system clock in the front-end business system; when the system clock is aligned with the trigger timing parameters, trigger the interface release condition and transmit the targeted marketing scripts and discount push conditions to the terminals of potential students on the high-risk stagnation list.
[0014] The present invention has the following beneficial effects: (1) A funnel-based conversion analysis method for enrollment based on the grouping of prospective students. By collecting the front-end interaction behavior logs of prospective students to extract behavioral feature data, and using time-weighted clustering algorithm to output the intention preference type and the list of grouped students, it overcomes the drawback of traditional methods that perform mixed calculations on the entire sample and realizes the delineation of the dataset based on the time sensitivity of dynamic features. Furthermore, the list of grouped students is mapped to the funnel node sequence. By comparing the actual conversion dwell time with the benchmark time threshold, the abnormal break level that exceeds the time limit is accurately captured and a high-risk stagnation list is generated. From the perspective of multidimensional data flow, the local abnormal stagnation nodes of a specific group are accurately located, which effectively solves the technical problem that the calculation of the overall conversion rate covers the real loss position of the micro group.
[0015] (2) A hierarchical conversion analysis method for enrollment funnel based on the grouping of potential students. By retrieving the bidirectional interaction records of the business at the abnormal break level, the semantic similarity distance between the feature tags of the outgoing marketing materials and the intention preference type is calculated to obtain the semantic misalignment of the text. Combined with the number of unread messages, a conversion obstacle matrix is generated, which deeply analyzes the matching pain points between the actions executed by the enrollment system and the business demands of the students. Furthermore, the correction and recovery instructions are matched in the intervention action library, encapsulated as automatic intervention tasks, and the message push interface is called to execute the issuance of targeted marketing scripts and discount push conditions. This realizes the technical closed loop of the front-end business system executing correction tasks directly driven by the underlying analysis matrix, eliminates the recovery failure caused by the lag of manual intervention, and significantly improves the conversion data indicators of the target group at a specific level.
[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0017] Figure 1This is a flowchart of a funnel-based conversion analysis method for student enrollment based on the grouping of prospective students, as described in this invention.
[0018] Figure 2 Flowchart for generating the transformation obstacle factor matrix.
[0019] Figure 3 This is a flowchart for the execution of automated intervention tasks. Detailed Implementation
[0020] This application embodiment addresses the problem that existing conversion analysis models struggle to pinpoint local attrition points and automate corrective interventions by employing a funnel-based conversion analysis method based on student segmentation within prospective student groups. The overall approach of this application embodiment is as follows: First, the front-end interactive behaviors of potential students at business touchpoints are collected and their features are clustered to establish categorized group data with homogeneous preference attributes. Second, the categorized group data is imported into a pre-set enrollment funnel node sequence to perform dwell time comparison calculations, screening out abnormal levels that generate stagnant data and associated high-risk stagnant lists. Then, the interaction records between the enrollment end and the student end within the abnormal level are extracted, and a conversion obstacle matrix data model is established by calculating the semantic similarity distance between marketing materials and student preferences and the frequency of unread communication interactions. Finally, the conversion obstacle matrix data model is parsed and the corresponding intervention execution program is matched. By calling the front-end business interface, the automatic intervention data task package is dispatched, thereby realizing an automated, high-precision data tracking and system intervention response closed loop.
[0021] Please see Figure 1This invention provides a technical solution: a method for hierarchical conversion analysis of the enrollment funnel based on the grouping of potential students, comprising the following steps: S1. Collecting front-end interaction behavior logs generated by potential students at business touchpoints, extracting page browsing depth sequences and dialogue response time intervals as behavioral feature data; inputting the behavioral feature data into a time-weighted clustering algorithm for clustering calculation, outputting the intention preference type and the corresponding grouped student list; S2. Mapping the grouped student list to a preset enrollment funnel node sequence, calculating the actual conversion dwell time of the grouped student list in each single funnel node; comparing the actual conversion dwell time with the baseline time threshold associated with the intention preference type, extracting single funnel nodes whose actual conversion dwell time exceeds the baseline time threshold as abnormal break levels, and summarizing the potential students staying in the abnormal break levels to generate a high-risk stagnation list; S3. Retrieving high... The system extracts enrollment follow-up behavior records and student feedback records within the time period corresponding to the abnormal break level of the high-risk stagnation list. It extracts feature tags from external marketing materials in the enrollment follow-up behavior records and calculates the semantic similarity distance between these feature tags and the intended preference type to obtain the text semantic misalignment degree. It also extracts the number of unread messages from the student feedback records and combines the text semantic misalignment degree with the number of unread messages to generate a conversion obstacle driver matrix. S4. It extracts specific business misalignment types from the conversion obstacle driver matrix and queries the pre-built intervention action library for correction and recovery instructions matching the specific business misalignment type. The correction and recovery instructions are encapsulated into automatic intervention tasks containing targeted marketing scripts and discount push conditions. The system then calls the message push interface of the front-end business system to send targeted marketing scripts and discount push conditions to the terminals of prospective students in the high-risk stagnation list according to the automatic intervention task.
[0022] In this implementation plan, step S1 involves collecting front-end interaction behavior data of prospective students and performing dynamic feature grouping. Specifically, the system extracts two behavioral feature data from the front-end interaction behavior logs: page browsing depth sequence and dialogue response time interval. This behavioral feature data is then input into a time-weighted clustering algorithm for clustering calculation, ultimately outputting the intention preference type and the corresponding list of students in each group. Non-publicly known terms involved in this step include front-end interaction behavior logs and time-weighted clustering algorithms. Front-end interaction behavior logs refer to the data stream formed by the system's underlying records of user touch, pause, and swipe operations on the terminal interface. Time-weighted clustering algorithms are mathematical models that assign different attenuation weights based on the time elapsed between the data's generation and the current moment when performing classification and clustering calculations on data nodes. The technical advantage of this step is that it transforms massive, unordered underlying user operation logs into structured feature parameters, automatically defining classification groups with homogeneous preference attributes using algorithms. This provides a more refined data foundation with dynamic time-series characteristics for subsequent funnel analysis, overcoming the problem of ambiguous analysis objects caused by traditional analysis methods that mix all samples in the calculation.
[0023] Step S2 functions to perform dynamic and timely monitoring and anomaly location identification of funnel nodes based on categorized group data. Specifically, the student group lists obtained in the previous step are mapped to a pre-defined enrollment funnel node sequence. The actual conversion dwell time of the student group lists within each individual funnel node is calculated. By comparing the actual conversion dwell time with a baseline time threshold for associated intention preference types, individual funnel nodes exceeding the threshold are extracted as abnormal break levels. Students remaining at this level are then aggregated to generate a high-risk stagnation list. Non-publicly known terms involved in this step include the enrollment funnel node sequence and abnormal break levels. The enrollment funnel node sequence refers to a pre-configured set of business flow states with a sequential logical order in the business flow system. An abnormal break level refers to a specific system node where a specific group experiences an abnormal stagnation during state transition, as determined by a time comparison algorithm. The technical role of this step is to deeply map and couple dynamic cluster data with the standard conversion model. By comparing personalized time thresholds, it can accurately detect which data flow node is blocked for different characteristic groups, thus solving the technical problem that global unified conversion rate calculation cannot accurately locate the bottleneck of local business nodes.
[0024] Step S3's function is to delve into the root causes of business misalignment at the abnormal break level and construct a structured attribution data model. Specifically, the logic involves retrieving enrollment follow-up behavior records and student feedback records from the high-risk stagnation list corresponding to the abnormal break level. The semantic similarity distance between the feature tags of outbound marketing materials and the intention preference type is calculated to obtain the text semantic misalignment degree. Simultaneously, the number of unread messages in the student feedback records is extracted. Finally, these two parameters are combined into a matrix to generate a conversion obstacle driver matrix. Non-publicly known terms involved in this step include semantic similarity distance and the conversion obstacle driver matrix. Semantic similarity distance refers to the mathematical spatial distance calculated after converting text into vectors using natural language processing technology, used to quantify the degree of deviation between the content issued by the system and the user's actual needs. The conversion obstacle driver matrix is a multi-dimensional data structure used to comprehensively map the coordinate relationship between the degree of mismatch in communication content and the frequency of user resistance behavior. The technical role of this step is to scientifically quantify the root causes of churn at the underlying data level by calculating the distance between the supply-side data and the user demand-side data in the semantic space. This transforms the originally vague reasons for churn into a computer-readable and computable matrix model, providing precise data guidance for subsequent automated intervention.
[0025] Step S4 parses the attribution data model and automatically triggers the corresponding system-level corrective intervention command closed loop. Specifically, it extracts specific business misalignment types from the conversion obstacle driver matrix, matches corresponding corrective and recovery commands in a pre-built intervention action library, encapsulates these commands into an automatic intervention task containing targeted marketing scripts and discount push conditions, and finally calls the front-end business system's message push interface to send the corresponding conditions to terminals on the high-risk stagnation list. Non-publicly known terms involved in this step include specific business misalignment types and automatic intervention tasks. Specific business misalignment types refer to structured anomaly codes extracted from the driver matrix, such as ignored price sensitivity or unmet quality demands. Automatic intervention tasks refer to standardized data transmission packets automatically generated by the system, containing trigger conditions and execution content. The technical advantage of this step is that it directly converts the complex matrix operation results from the front end into executable system communication commands. By calling the message push interface, it achieves a fully automated closed loop of data flow from anomaly detection and cause diagnosis to strategy issuance, effectively avoiding delays caused by manual intervention and maximizing the recovery of customers in a stagnant state.
[0026] Specifically, the process of collecting front-end interaction behavior logs generated by prospective students at business touchpoints and extracting page browsing depth sequences and dialogue response time intervals as behavioral feature data is as follows: The event tracking data of prospective students on the enrollment system interface and customer service chat window are parsed from the front-end interaction behavior logs. The page access nodes contained in the event tracking data are arranged in timestamp order, and the hierarchical attributes of the page access nodes are extracted to construct the page browsing depth sequence. The timestamps of customer service inquiries and prospective student responses contained in the customer service chat window are retrieved, and timestamp difference extraction is performed. The difference extraction result is used as the dialogue response time interval, and the page browsing depth sequence and dialogue response time interval are extracted simultaneously to generate behavioral feature data.
[0027] In this implementation plan, the system first connects to the asynchronous monitoring module of the business front end to intercept event tracking data. Event tracking data refers to implicit data collection programs pre-embedded in the interactive code of web pages or mini-programs to detect specific user trigger actions. By reconstructing the timeline of the intercepted data stream, the system can reconstruct the access trajectory of potential students within the platform. During this extraction process, the hierarchical attribute of the page access node refers to the depth level of the currently accessed page in the entire website topology. The larger the hierarchical value, the more fundamental the business information contained in the page and the more vertical the user's intent. Arranging multiple consecutive access hierarchical values in chronological order yields the page browsing depth sequence, which intuitively reflects the exploration depth trend of potential students. For data processing in the customer service conversation window, the system automatically filters out the preset script response data of the system robot, accurately locating the question records issued by human customer service and the effective response records of potential students. To quantify the interaction enthusiasm of potential students, the system extracts the timestamps of multiple dialogue rounds and performs difference extraction. The calculation formula for the dialogue response time interval is as follows: The specific meanings of each parameter in the formula are as follows: : Represents the average conversation response time for a single prospective student within a business touchpoint; The total number of interaction rounds representing successful matching between customer service inquiries and valid responses from interested students; The traversal sequence number representing the interaction round; Representative at the The timestamps of the responses from interested students recorded in each round of interaction; :represents the first The system records the timestamps of customer service inquiries during each interaction round. Finally, the system spatially concatenates and combines the page browsing depth sequence and the average dialogue response time interval along the feature dimension to generate behavioral feature data for subsequent algorithm processing. This step transforms the scattered and disordered operation logs into computer-readable structured preference vectors, providing quantitative underlying data support for accurate grouping.
[0028] Specifically, the process of inputting behavioral feature data into a time-weighted clustering algorithm for clustering calculation and outputting intention preference types and corresponding student group lists is as follows: Extract the occurrence timestamps associated with the behavioral feature data, measure the time span between the occurrence timestamps and the current system time, and assign time decay weights to the behavioral feature data based on the time span; introduce time decay weights to weight the behavioral feature data to obtain a weighted feature matrix, input the weighted feature matrix into the cluster center nodes of the time-weighted clustering algorithm to perform iterative clustering, output basic data clusters with similar feature dimensions, perform semantic mapping between the basic data clusters and a pre-set intention preference label library, and output intention preference types; classify and summarize the unique identifiers of the students who generated the behavioral feature data according to the intention preference types, and output student group lists.
[0029] In this implementation plan, because the enrollment intentions and focus of prospective students will significantly decline and shift over time, the system must perform time-sensitive processing on behavioral characteristic data. First, the time span between the occurrence timestamp and the current system time is measured. Time decay weights are then assigned based on the proximity to the current moment, ensuring that more recent behaviors have a greater influence on the clustering results. The formula for calculating the time decay weights is as follows: The specific meanings of each parameter in the formula are as follows: : represents the Time decay weights are assigned to each behavioral feature data point. : Represents the traversal identifier for behavioral feature data; : The base constant of the natural logarithm; : represents the The time span difference between the occurrence timestamp of each behavioral feature data and the current system time; : Represents the time decay coefficient. The time decay coefficient is not a fixed extreme value; its determination depends on the natural follow-up and conversion cycle of the enrollment process. The formula for determining the time decay coefficient is as follows: The specific meanings of each parameter in the formula are as follows: : Represents the logarithmic value of a constant with the natural constant as its base; This represents the intended half-life time parameter preset by the business system. After obtaining the time decay weight through the above formula, the system performs scalar multiplication on the original behavioral feature data to amplify or reduce it, generating a weighted feature matrix. Then, it enters the core clustering stage, inputting the weighted feature matrix into the cluster center nodes of the time-weighted clustering algorithm for iterative calculation. During this process, the algorithm uses weighted spatial distance to measure the similarity between data points and cluster center nodes, deeply integrating the time weight into spatial classification. Its calculation formula is as follows: The specific meanings of each parameter in the formula are as follows: : represents the The behavioral characteristic data and the first Weighted spatial similarity distance between cluster center nodes; : Represents the traversal identifier of the cluster center node; : Represents the total number of feature dimensions contained in the weighted feature matrix; : Represents the traversal sequence number of the feature dimension; : represents the The behavioral characteristic data in the first Feature coefficient values in each dimension; : represents the The cluster center node at the th The system calculates coordinate coefficients across multiple dimensions. After iterating through multiple spatial distance calculations until the cluster center nodes converge and stabilize, the system outputs basic data clusters with similar feature dimensions. Since the basic data clusters are merely a set of feature vectors in a purely mathematical spatial dimension, the system performs text-level distance mapping between their multi-dimensional coordinate features and standard business templates in a pre-set intention preference label library, completing semantic association mapping. This transforms abstract mathematical classifications into intention preference types readable by admissions staff. Finally, the system scans for the unique identifiers of the students who generated these behavioral feature data, summarizes the lists according to the determined preference types, and outputs a list of students in different groups. This allows for the precise separation of student groups at different intention lifecycles and with different needs, eliminating data heterogeneity interference in subsequent funnel model analysis.
[0030] Specifically, the process of mapping the student group list to a pre-defined enrollment funnel node sequence and calculating the actual conversion dwell time of the student group list within each individual funnel node is as follows: Parse the unique identifier of the prospective student contained in the student group list, match the business flow status record of the prospective student's unique identifier in the front-end business system, align the business flow status record with the individual funnel nodes contained in the enrollment funnel node sequence, and determine the node transfer-in timestamp and node transfer-out timestamp corresponding to the individual funnel node; extract the time span difference between the node transfer-out timestamp and the node transfer-in timestamp, and confirm the time span difference as the actual conversion dwell time. When there is a missing node transfer-out timestamp, extract the time span difference between the current system timestamp and the node transfer-in timestamp to confirm the actual conversion dwell time.
[0031] In this implementation plan, to accurately track the conversion progress of different customer groups with varying preferences, the system needs to deeply couple the classification results with the data from the business front end. The system first parses the unique identifiers of the interested students in the grouped student lists. These identifiers are the system primary keys that run throughout the entire data lifecycle and can be linked to operation records across modules. Subsequently, the system matches the business flow status records of the interested student's unique identifier within the front-end business system. These records represent a series of sequential business actions recorded in the front-end business system, such as lead allocation, telephone follow-up, and trial class scheduling. The system aligns these business flow status records with a pre-defined sequence of enrollment funnel nodes to identify the absolute time points when interested students enter and leave a single funnel node. A single funnel node represents a stage in the enrollment conversion chain with independent business attributes and specific marketing objectives. To accurately measure the time students spend at each business stage, the system introduces a status indicator function to continuously and uniformly model and calculate the actual conversion dwell time. The calculation formula is as follows: The specific meanings of each parameter in the formula are as follows: : represents the The number of prospective students in the first The actual conversion dwell time within a single funnel node; The traversal index value representing the intended student; The sequence number representing a single funnel node; This represents a status indicator parameter. It takes a value of one when a node's transfer-out timestamp exists, and a value of zero when a node's transfer-out timestamp is missing. : represents the One prospective student left the first The node output timestamp corresponding to a single funnel node; : represents the The number of prospective students entered the first The node transition timestamp corresponding to a single funnel node; : Represents the current system timestamp. Through the above calculation model, the system can not only measure the historical dwell time of students who have successfully entered the next stage, but also calculate the real-time dwell time of students who are still stuck at a certain node. This allows the hidden churn risk of students who are in a state of missing node transfer timestamp, i.e., a state of stagnation in follow-up, to be dynamically quantified, thus providing a complete and continuous time span difference data foundation for subsequent anomaly interception judgment.
[0032] Specifically, the process of comparing the actual conversion dwell time with the baseline time threshold associated with the intention preference type, extracting single funnel nodes whose actual conversion dwell time exceeds the baseline time threshold as abnormal break levels, and summarizing the intention students remaining in the abnormal break levels to generate a high-risk stagnation list is as follows: Read the intention preference type, query the baseline time threshold matching the intention preference type in the preset funnel timeliness configuration table, compare the actual conversion dwell time with the baseline time threshold, and record the single funnel node whose actual conversion dwell time exceeds the baseline time threshold; mark the single funnel node whose actual conversion dwell time exceeds the baseline time threshold as an abnormal break level, extract the unique identifier of the intention students remaining in the abnormal break level, and combine the unique identifiers of the intention students to generate a high-risk stagnation list.
[0033] In this implementation plan, after obtaining the actual conversion dwell time, the system needs to further determine whether this duration constitutes abnormal congestion in the business chain. Since students with different preference types naturally have different decision-making cycles, the system abandons a globally uniform static judgment standard and instead extracts the preference type, querying a pre-set funnel timeliness configuration table for a baseline time threshold matching a specific preference type. To ensure that this threshold parameter dynamically adapts to the supply and demand changes in the actual enrollment environment, the baseline time threshold is determined by extracting the expected dwell time distribution of historically successfully converted samples and combining it with dynamic tolerance for adaptive extrapolation. The calculation formula is as follows: The specific meanings of each parameter in the formula are as follows: : represents the Types of intention preferences in the 1st The baseline time threshold associated with a single funnel node; Category number representing the type of intention / preference; : Represents the system's preset setting for the first Dynamic tolerance coefficient for each type of intention preference; Represents the historical database belonging to the first The total number of samples with each type of intention preference that have been successfully converted to the end payment node; The traversal sequence number representing the successfully transformed historical samples; : represents the The first historical successful conversion sample in the The system analyzes the historical actual dwell time generated within a single funnel node. After deriving a baseline time threshold that aligns with group characteristics, the system compares the real-time generated actual conversion dwell time with this baseline time threshold, extracting single funnel nodes where the actual conversion dwell time exceeds the baseline time threshold. These system nodes that exceed the normal decision-making lifecycle are uniformly marked as abnormal break levels, indicating that the front-end enrollment funnel has failed at this level. The system then extracts the unique identifiers of prospective students remaining within these abnormal break levels, performs multi-dimensional deduplication and merging, and finally outputs a high-risk stagnation list. This judgment process effectively filters out business flow waiting time within the normal follow-up cycle, accurately locating the real funnel shrinkage nodes and high-risk churn groups, providing targeted coordinates for subsequent triggering of highly targeted system-level automatic corrective interventions.
[0034] Please see Figure 2 Specifically, the process of retrieving enrollment follow-up behavior records and student feedback records within the time period corresponding to the abnormal break level of the high-risk stagnation list, extracting the feature tags of outbound marketing materials from the enrollment follow-up behavior records, and calculating the semantic similarity distance between the outbound marketing material feature tags and the intention preference type to obtain the text semantic misalignment is as follows: Receive the unique identifiers of the intended students included in the high-risk stagnation list; define the time window associated with the abnormal break level in the enrollment system database; based on the unique identifiers of the intended students and the time window, retrieve and retrieve the enrollment follow-up behavior records and student feedback records; parse the outbound action details included in the enrollment follow-up behavior records; extract the material metadata bound to the outbound action details; convert the material metadata into outbound marketing material feature tags; map the outbound marketing material feature tags and intention preference types to a pre-set business semantic vector space, and output the corresponding material feature vector and preference type vector; perform spatial distance measurement calculation between the material feature vector and the preference type vector; and confirm the spatial distance measurement result as the text semantic misalignment.
[0035] In this implementation plan, to delve into the root causes of potential students' stagnation at specific business nodes, the system needs to compare whether the follow-up actions of enrollment staff match the students' actual needs. The system first receives the unique identifier of the potential student and uses this as a key to define the time window corresponding to the stagnation in the enrollment system database, then batch-retrieves business two-way interaction data within that time period. The material metadata extracted after parsing the outreach details includes attributes such as course outlines, price discounts, or teacher introductions, which the system converts into external marketing material feature tags. To make the discrete text features mathematically computable, the system introduces a business semantic vector space, which is a multi-dimensional spatial coordinate system pre-trained using a natural language processing model for corpora in the education and enrollment field. The system maps the external marketing material feature tags and the student's inherent intention and preference types into this space, generating corresponding material feature vectors and preference type vectors. Subsequently, the system performs spatial distance measurement calculations to quantify the deviation in direction and magnitude between the two. The formula for calculating the text semantic misalignment is as follows: The specific meanings of each parameter in the formula are as follows: Representing the The feature tags of external marketing materials and the first Textual semantic misalignment of intention preference type; The identifier number represents the type of intention / preference; The total number of semantic dimensions in the pre-defined business semantic vector space; : Represents the material feature vector at the th Mapping projection values on each semantic dimension; : Represents the preference type vector at the th The above calculation steps scientifically quantify the degree of semantic misalignment between the content supplied by the enrollment side and the demand preferences of the students through the spatial cosine distance algorithm. The closer the misalignment value is to 1, the more the marketing materials sent deviate from the students' true intentions. This explains the business motivation for the funnel breakage from the underlying logic and provides a quantitative basis for qualitative analysis of the reasons for churn.
[0036] Specifically, the process of extracting the number of unread messages from student feedback records and combining the text semantic misalignment degree with the number of unread messages to generate a conversion obstacle driving factor matrix is as follows: Scan the instant messaging interaction status bits contained in the student feedback records, summarize the frequency of unreached instant messaging interaction status bits to obtain the number of unread messages, introduce a pre-defined misalignment level classification rule to discretize and classify the text semantic misalignment degree to be processed, and generate misalignment level coordinate parameters. Introduce a pre-defined resistance response level classification rule to discretize and classify the number of unread messages to be processed, and generate resistance level coordinate parameters. Establish a two-dimensional correlation analysis architecture containing horizontal and vertical axes, assign the misalignment level coordinate parameters to the horizontal axis, assign the resistance level coordinate parameters to the vertical axis, map the misalignment level coordinate parameters and resistance level coordinate parameters to the corresponding cross nodes within the two-dimensional correlation analysis architecture, and output a conversion obstacle driving factor matrix containing the cross nodes.
[0037] In this implementation plan, semantic misalignment only reflects the deviation on the supply side of the system. To comprehensively assess the risk of churn, the system also needs to incorporate the resistance feedback characteristics from the student demand side. The system scans the instant messaging interaction status bits in the student feedback records and summarizes the frequency of unreachable flags to obtain the number of unread messages. Since misalignment and unread messages are continuous data with different dimensions, the system introduces a pre-defined misalignment level classification rule to discretize and classify the semantic misalignment of the text to be processed. Discretization and classification refers to the data processing process of mapping continuous numerical variables to integer level constants according to a pre-defined stepped threshold range, thereby generating misalignment level coordinate parameters. Similarly, a pre-defined resistance response level classification rule is introduced to generate resistance level coordinate parameters. The system establishes a two-dimensional correlation analysis architecture with horizontal and vertical axes, assigns these two coordinate parameters to the horizontal and vertical axes respectively, and maps them to the corresponding cross nodes within the two-dimensional correlation analysis architecture. To intuitively measure the overall severity of conversion blocking caused by the cross node, the system configures a churn risk weight calculation process for the cross node, and the calculation formula is expressed as follows: The specific meanings of each parameter in the formula are as follows: : Represents the comprehensive loss risk weight borne by the corresponding cross node in the transformation obstacle driving force matrix; : Represents the identifier for traversing the set of discretized orders of the misalignment level coordinate parameters; : Represents the identifier for traversing the set of discretized orders of the resistance level coordinate parameters; : Represents the number mapped to the horizontal axis Dislocation level coordinate parameters; : Represents the number mapped to the vertical axis Level of resistance coordinate parameters; : Represents the semantic misalignment penalty correlation coefficient preset by the system; : Represents the system's preset resistance feedback penalty correlation coefficient. The method for determining the semantic misalignment penalty correlation coefficient and the resistance feedback penalty correlation coefficient is as follows: extract the variance contribution rate of the semantic feature field and the unread frequency field from the historical churn sample database to the final churn result, and assign the normalized variance contribution rate values to the two penalty coefficients mentioned above. The final system output is a conversion obstacle driver matrix containing the above-mentioned cross nodes and corresponding risk weights. This matrix transforms the ambiguous communication problems into a structured data model with clear coordinate positioning and severity rating. It not only reveals the superimposed resistance effect caused by action misalignment but also provides precise navigation coordinates for subsequent business systems to issue correction and recovery instructions of different resource intensities as needed.
[0038] Please see Figure 3 Specifically, the process of extracting specific business misalignment types from the conversion obstacle driver matrix and querying the pre-built intervention action library for correction and recovery instructions matching the specific business misalignment types is as follows: traverse the intersection nodes contained in the conversion obstacle driver matrix, locate the extreme values of the two-dimensional coordinates pointed to by the intersection nodes, determine the specific business misalignment type based on the extreme values of the two-dimensional coordinates, and parse the business pain point field and follow-up error field contained in the specific business misalignment type; input the business pain point field and follow-up error field into the retrieval engine of the pre-built intervention action library for feature comparison, extract the set of action strategies in the pre-built intervention action library whose feature comparison consistency meets the preset threshold, and confirm the set of action strategies as correction and recovery instructions.
[0039] In this implementation plan, to accurately match recovery strategies, the system first traverses the cross-nodes contained in the conversion obstacle driver matrix. The system locates the extreme values of the two-dimensional coordinates pointed to by these cross-nodes. These extreme values represent the points in the matrix where the risk of churn reaches its peak after the superposition of semantic deviation and resistance frequency. Based on these extreme values, the system determines specific business misalignment types and parses out the business pain point field and follow-up error field. The business pain point field represents the core needs of potential students, such as concerns about price or faculty, while the follow-up error field records system action defects by early recruitment staff, such as premature closing or slow response. Subsequently, the system inputs the business pain point field and follow-up error field into the retrieval engine of a pre-built intervention action library for feature comparison. To ensure that the invoked action strategies are targeted effectively, the system performs a feature comparison consistency calculation, the formula of which is as follows: The specific meanings of each parameter in the formula are as follows: : Represents the first Each business misalignment type and the first in the pre-built intervention action library Consistency of feature comparison for each action strategy; : The traversal sequence number representing the business misalignment type; : Represents the traversal sequence number of the candidate action strategy; : Represents the system's preset business pain point matching weight coefficient; : Represents the system's preset follow-up error matching weight coefficient; : Represents the first A set of business pain point field features parsed from each business misalignment type; : Represents the first A set of pain point resolution tags pre-set for each candidate action strategy; : Represents the first A set of follow-up error field features parsed from each business misalignment type; : Represents the first A set of error compensation labels pre-set for each candidate action strategy; : A function that calculates the number of characteristic elements contained within a statistical set; : Represents a mathematical operation that extracts the intersection of two sets; This represents a mathematical operation to extract the union element of two sets. The method for determining the business pain point matching weight coefficient and the follow-up error matching weight coefficient is to statistically analyze the proportion of successful transactions resulting from a single pain point resolution action and a single error correction action in the historical database, and then normalize this proportion to use as the two weight coefficients. Finally, the system extracts a set of action strategies whose feature comparison consistency meets a preset threshold. The preset threshold is determined by multiplying the median of the feature comparison consistency of all candidate action strategies by a floating constant of 1.5, thus confirming the set of action strategies with the highest score as the correction and recovery instruction. This calculation step, by evaluating the multidimensional feature consistency of the intersection-union ratio of sets, realizes the transformation from subjective human judgment to intelligent recommendation and intervention strategies by underlying algorithms, significantly improving the targeting accuracy and final conversion rate of recovery actions.
[0040] Specifically, the process of encapsulating the correction and recovery instruction into an automatic intervention task containing targeted marketing messages and discount push conditions, and calling the message push interface of the front-end business system to send the targeted marketing messages and discount push conditions to the terminals of potential students on the high-risk stagnation list according to the automatic intervention task is as follows: Parse the trigger timing parameters, targeted marketing messages, and discount push conditions contained in the correction and recovery instruction; assemble the trigger timing parameters, targeted marketing messages, and discount push conditions into an automatic intervention task data packet with a standard structure format according to the preset business system data exchange protocol; initiate a call request carrying the automatic intervention task data packet to the message push interface of the front-end business system; monitor the system clock in the front-end business system; when the system clock is aligned with the trigger timing parameters, trigger the interface release condition and transmit the targeted marketing messages and discount push conditions to the terminals of potential students on the high-risk stagnation list.
[0041] In this implementation plan, to achieve seamless system-level integration from strategy diagnosis to business execution, the system parses the trigger timing parameters, targeted marketing scripts, and discount push conditions contained in the correction and recovery instructions. The trigger timing parameter refers to the optimal message response timestamp calculated based on students' historical online patterns. Following a pre-defined business system data exchange protocol (which refers to the cross-platform encryption and format encapsulation specifications agreed upon between the front-end and back-end communication servers), the system assembles the above parameters into a standard-structured automatic intervention task data packet. Subsequently, the system initiates a call request carrying this data packet to the message push interface of the front-end business system. To ensure reasonable allocation of system computing resources and that the push action is executed within the optimal window, the system monitors the system clock within the front-end business system in real time and continuously calculates the alignment activation coefficient between the current clock and the trigger timing parameters through a trigger function. The calculation formula is as follows: The specific meanings of each parameter in the formula are as follows: : Represents the first Alignment activation coefficients of data packets for automated intervention tasks; : Represents the queue number of the automatic intervention task data package in the call queue; : Represents an exponential function with the base of the natural logarithm; : Represents the system's preset time-sensitive attenuation coefficient; : Represents the currently monitored system clock timestamp; : Represents the first Trigger timing parameters parsed from the data packet of an automatic intervention task; : Represents the first The system channel priority weights are associated with each automatic intervention task data packet. The time-sensitive attenuation coefficient is determined by calculating the average delay volatility of the historical queuing sequence of the front-end message push interface. The system channel priority weights are determined by mapping and matching different gradient integer weights based on the historical average order value prediction range of student accounts in the high-risk stagnation list. When the alignment activation coefficient reaches a preset extreme value, the interface release condition is triggered. The interface release condition refers to the control command of the business system to lift network port rate limiting and establish a low-level communication distribution link to the external network. The system then accurately sends targeted marketing messages and discount push conditions to the terminals of potential students in the high-risk stagnation list. This calculation step, by introducing a priority-based exponential attenuation alignment model, accurately controls the timing of concurrent message delivery. This not only ensures that automated intervention tasks can accurately reach students at their most active times but also effectively avoids delays in sending core high-net-worth customer recovery instructions during high-concurrency congestion periods, completely opening up the final business execution loop from funnel analysis to marketing correction at the underlying communication level.
[0042] In summary, this application has at least the following effects: A hierarchical conversion analysis method for student recruitment funnels, based on the segmentation of potential students, achieves high-precision dynamic segmentation of customer groups through time-weighted clustering of front-end interactive behavior features, overcoming the limitations of traditional static profiling segmentation. It then maps the segmented data to recruitment funnel nodes and performs continuous conversion dwell time threshold comparisons to accurately locate abnormal breakpoints and high-risk stagnation lists, solving the problem of global funnel averaging masking local group attrition. Subsequently, it deeply analyzes the two-way interaction records, combining the semantic distance between marketing materials and preference types with message unread frequency to construct a conversion obstacle driver matrix, scientifically quantifying the root causes of supply-demand mismatch between business actions and student needs. Finally, the features of this underlying matrix directly drive the instruction matching of a pre-built intervention action library, and combined with a system clock alignment model, automatically distributes targeted marketing messages and discount push conditions, completely opening up a system-level data loop from anomaly detection and root cause diagnosis to precise delivery of corrective strategies. This effectively eliminates reliance on manual experience and intervention lag, significantly improving the flow efficiency and final conversion success rate of the target student group at each funnel level.
[0043] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0044] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0046] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0047] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0048] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for admission funnel stratification conversion analysis of intention-to-student cohort, characterized in that, Includes the following steps: S1. Collect front-end interaction behavior logs generated by prospective students at business touchpoints, and extract page browsing depth sequence and dialogue response time interval as behavioral feature data; input the behavioral feature data into a time-weighted clustering algorithm for clustering calculation, and output the intention preference type and the corresponding student group list; S2. Map the student group list to the preset enrollment funnel node sequence, and calculate the actual conversion dwell time of the student group list in each single funnel node; By comparing the actual conversion dwell time with the baseline time threshold associated with the intention preference type, a single funnel node whose actual conversion dwell time exceeds the baseline time threshold is extracted as an abnormal break level, and a list of high-risk stagnation of intention students who stay in the abnormal break level is compiled. S3. Retrieve the enrollment follow-up behavior records and student feedback records of the high-risk stagnation list within the time period corresponding to the abnormal break level. Extract the feature tags of external marketing materials from the enrollment follow-up behavior records. Calculate the semantic similarity distance between the feature tags of external marketing materials and the intention preference type to obtain the text semantic misalignment. Extract the number of unread messages from the student feedback records. Combine the text semantic misalignment and the number of unread messages into a matrix to generate a conversion obstacle factor matrix. S4. Extract specific business misalignment types from the conversion obstacle driver matrix, query the pre-built intervention action library for correction and recovery instructions that match the specific business misalignment type; encapsulate the correction and recovery instructions into automatic intervention tasks containing targeted marketing scripts and discount push conditions, call the message push interface of the front-end business system, and send targeted marketing scripts and discount push conditions to the terminals of potential students in the high-risk stagnation list according to the automatic intervention task.
2. The method of claim 1, wherein the funnel layering conversion analysis of prospective student cohort is characterized by: The specific process of collecting front-end interaction behavior logs generated by prospective students at business touchpoints and extracting page browsing depth sequences and dialogue response time intervals as behavioral feature data is as follows: The event tracking data of prospective students in the enrollment system interface and customer service chat window are analyzed from the front-end interaction behavior logs. The page access nodes contained in the event tracking data are arranged in the order of timestamps, and the hierarchical attributes of the page access nodes are extracted to construct a page browsing depth sequence. Retrieve the timestamps of customer service inquiries and intended student responses from the customer service conversation window, perform timestamp difference extraction, and use the difference extraction result as the dialogue response time interval. Simultaneously extract the page browsing depth sequence and dialogue response time interval to generate behavioral feature data.
3. The enrollment funnel stratification and conversion analysis method for grouping prospective students according to claim 2, characterized in that: The specific process of inputting behavioral feature data into a time-weighted clustering algorithm for clustering calculation, and outputting the intention preference type and the corresponding list of students in the cluster is as follows: Extract the occurrence timestamps associated with behavioral feature data, measure the time span between the occurrence timestamps and the current system time, and assign time decay weights to the behavioral feature data based on the time span; A time decay weight is introduced to weight the behavioral feature data to obtain a weighted feature matrix. The weighted feature matrix is then input into the cluster center node of the time-weighted clustering algorithm to perform iterative clustering, outputting basic data clusters with similar feature dimensions. The basic data clusters are then semantically mapped to a pre-set intention preference label library to output the intention preference type. The unique identifiers of students who generate behavioral characteristic data are categorized and summarized according to their intention and preference types, and a list of students in each group is output.
4. The enrollment funnel stratification and conversion analysis method for grouping prospective students according to claim 1, characterized in that: The specific process of mapping the student group lists to a pre-defined sequence of enrollment funnel nodes and calculating the actual conversion dwell time of the student group lists within each individual funnel node is as follows: The unique identifiers of prospective students in the student group list are parsed, and the business flow status records of the prospective student unique identifiers in the front-end business system are matched. The business flow status records are aligned with the status of the single funnel nodes in the enrollment funnel node sequence, and the node transfer-in timestamp and node transfer-out timestamp corresponding to the single funnel node are determined. Extract the time span difference between the node transfer-out timestamp and the node transfer-in timestamp, and confirm the time span difference as the actual conversion dwell time. When there is a missing node transfer-out timestamp, extract the time span difference between the current system timestamp and the node transfer-in timestamp to confirm the actual conversion dwell time.
5. The enrollment funnel stratified conversion analysis method for grouping prospective students according to claim 4, characterized in that: The specific process of comparing the actual conversion dwell time with the baseline time threshold associated with the intention preference type, extracting single funnel nodes whose actual conversion dwell time exceeds the baseline time threshold as abnormal break levels, and summarizing the intention students staying in the abnormal break levels to generate a high-risk stagnation list is as follows: Read the intention preference type, query the baseline time threshold matching the intention preference type in the preset funnel time configuration table, compare the actual conversion dwell time with the value of the baseline time threshold, and record the single funnel node whose actual conversion dwell time exceeds the baseline time threshold. A single funnel node whose actual conversion dwell time exceeds the baseline time threshold is marked as an abnormal break level. The unique identifier of the intended student staying in the abnormal break level is extracted, and the combination of the unique identifiers of the intended student is merged to generate a high-risk stagnation list.
6. The enrollment funnel stratification conversion analysis method for grouping prospective students according to claim 1, characterized in that: The specific process of retrieving enrollment follow-up behavior records and student feedback records within the time period corresponding to the abnormal break level of the high-risk stagnation list, extracting feature tags of outbound marketing materials from the enrollment follow-up behavior records, and calculating the semantic similarity distance between the feature tags of outbound marketing materials and the intention preference type to obtain the semantic misalignment of the text is as follows: The system receives the unique identifiers of prospective students from the high-risk suspension list, delineates the time window associated with the abnormal break in the admissions system database, and retrieves and displays the admissions department's follow-up behavior records and the students' feedback records based on the prospective students' unique identifiers and the time window. The system analyzes the contact action details contained in the enrollment follow-up behavior record, extracts the material metadata bound to the contact action details, converts the material metadata into external marketing material feature tags, maps the external marketing material feature tags and intention preference types to the preset business semantic vector space, and outputs the corresponding material feature vector and preference type vector. Perform spatial distance measurement calculation between material feature vector and preference type vector, and confirm the spatial distance measurement calculation result as text semantic misalignment degree.
7. The enrollment funnel stratification and conversion analysis method for grouping prospective students according to claim 6, characterized in that: The specific process of extracting the number of unread messages from student feedback records and generating a transformation obstacle factor matrix by combining the text semantic misalignment degree with the number of unread messages is as follows: Scan the instant messaging interaction status bits in the student feedback records, summarize the frequency of the unreachable instant messaging interaction status bits to obtain the number of unread messages, introduce a pre-set misalignment level classification rule to discretize and classify the semantic misalignment of the text to be processed, and generate misalignment level coordinate parameters. Introduce a pre-set resistance response level classification rule to discretize and classify the number of unread messages to be processed, and generate resistance level coordinate parameters. Establish a two-dimensional correlation analysis framework with horizontal and vertical axes, assign the misalignment level coordinate parameters to the horizontal axis, assign the resistance level coordinate parameters to the vertical axis, map the misalignment level coordinate parameters and resistance level coordinate parameters to the corresponding cross nodes within the two-dimensional correlation analysis framework, and output the transformation obstacle driving force matrix containing the cross nodes.
8. The enrollment funnel stratification conversion analysis method for grouping prospective students according to claim 1, characterized in that: The specific process of extracting specific business misalignment types from the transformation obstacle driver matrix and querying the pre-built intervention action library for corrective and remedial instructions that match the specific business misalignment type is as follows: Traverse the cross nodes contained in the transformation obstacle driving force matrix, locate the two-dimensional coordinate extreme value pointed to by the cross node, determine the specific business misalignment type based on the two-dimensional coordinate extreme value, and parse the business pain point field and follow-up error field contained in the specific business misalignment type. Input the business pain point field and follow-up error field into the retrieval engine of the pre-built intervention action library for feature comparison, extract the action strategy set in the pre-built intervention action library whose feature comparison consistency meets the preset threshold, and confirm the action strategy set as the correction and recovery instruction.
9. The enrollment funnel stratification conversion analysis method for grouping prospective students according to claim 8, characterized in that: The process of encapsulating the correction and recovery instructions into an automated intervention task containing targeted marketing scripts and discount push conditions, calling the message push interface of the front-end business system, and sending targeted marketing scripts and discount push conditions to the terminals of potential students on the high-risk suspension list according to the automated intervention task is as follows: The system analyzes the trigger timing parameters, targeted marketing scripts, and discount push conditions contained in the correction and recovery instructions, and assembles these parameters into a standard structured automatic intervention task data package according to the preset business system data exchange protocol. Initiate a call request carrying an automatic intervention task data packet to the message push interface of the front-end business system, monitor the system clock in the front-end business system, and trigger the interface release condition when the system clock is aligned with the trigger timing parameter, and transmit the targeted marketing script and discount push conditions to the terminals of potential students in the high-risk stagnation list.
