A correlation analysis method for fusing multi-source heterogeneous enrollment data

Abstract

The present application relates to the technical field of knowledge graph, in particular to a kind of associated analysis method of fusion multi-source heterogeneous enrollment data, comprising the following steps: extracting subject, predicate and object triple form enrollment data from educational administration system and competition registration system, configure timestamp or time effective interval for each triple group.The present application integrates multi-source heterogeneous data such as educational administration system and competition registration system, and configures timestamp for each record to sequence alignment, constructs time sequence knowledge graph that can dynamically reflect the development history of students, based on time perception graph, deduces dynamic evolution function describing the continuous change of the state of each entity with time, realizes the tracking of student state, and traverses and filters the path from student to admission result in the graph, and aggregates to generate high-value evolution path pattern with universal significance, which provides objective data-driven benchmark for evaluation.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a correlation analysis method that integrates multi-source heterogeneous enrollment data. Background Technology

[0002] The field of knowledge graph technology is a technology that describes entities and their relationships in the real world in the form of structured graphs.

[0003] In practical applications, existing knowledge graph technology, when assessing students, can only demonstrate that a student has won award A and completed project B. However, it struggles to distinguish whether a student first won award A and then completed project B based on their existing knowledge, or vice versa. These distinct growth paths are treated as equivalent facts in a static graph, thus losing the underlying causal logic and crucial information about skill progression. This leads to assessments that merely aggregate and count facts, such as the number of awards or course grades, potentially resulting in biased decisions due to the neglect of critical process information. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a correlation analysis method that integrates multi-source heterogeneous enrollment data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a correlation analysis method for integrating multi-source heterogeneous enrollment data, comprising the following steps:

[0006] Enrollment data in the form of subject, predicate, and object triples are extracted from the academic affairs system and competition registration system. A timestamp or time validity interval is assigned to each triple. The timestamps of enrollment data from different sources are sequence aligned. Based on the relationship between entities, a multi-source time-series enrollment knowledge graph is established.

[0007] Based on the multi-source time-series enrollment knowledge graph, the static vector representation of each entity node in the graph is calculated iteratively to obtain the entity static vector set. Combining the timestamp information with the entity static vector set, a dynamic evolution function describing the change of the vector with time points is derived for each entity. Based on the dynamic evolution function, the state of the student or university entity at any time point is calculated to obtain the time-series dynamic state vector.

[0008] Based on the topology of the multi-source temporal enrollment knowledge graph and the temporal dynamic state vector of the nodes, path query constraints including time sequence and time span are set. All paths from the student entity to the admission result entity are traversed in the graph, and paths that do not meet the time sequence or time span requirements are removed using the query constraints to obtain event sequences that meet the time constraints. All event sequences that meet the time constraints are aggregated to generate a high-value evolution path pattern.

[0009] Based on the high-value evolution path pattern, the application material subgraph of a single student is perturbed to obtain a set of perturbation sensitivity values. Based on the set of perturbation sensitivity values, the robustness score of the application materials is calculated.

[0010] Preferably, the steps for acquiring the multi-source time-series enrollment knowledge graph are as follows:

[0011] The records of the academic affairs system and the competition registration system are analyzed, the subject, predicate and object are extracted and timestamps or time validity intervals are configured, the source is deduplicated and the literal is standardized, the subject list and object list are unified, and the sequence is aligned according to the time sequence to obtain the triplet sequence that has been aligned.

[0012] Calculate the similarity coefficient based on the triplet sequence that has been aligned.

[0013] Based on the similarity coefficient, the triplet sequence that has been aligned is traversed in pairs according to the subject candidate and the object candidate, and the minimum and maximum timestamps of each entity are extracted to form an effective interval. The ratio of the intersection length of the two intervals is calculated as the overall time span overlap, and it is jointly judged with the merging threshold and the overlap threshold. When the joint judgment is satisfied, node merging and edge redirection are performed to generate a multi-source time-series enrollment knowledge graph.

[0014] Preferably, the steps for obtaining the entity static vector set are as follows:

[0015] Based on the multi-source time-series enrollment knowledge graph, the in-degree and out-degree counts, associated predicate identifier sequences, and attribute field value coverage are summarized one by one according to the entity nodes. A relation count table, attribute coverage table, and unique neighbor count field are established according to a fixed index. The hierarchical neighbor statistics are iteratively updated until the difference ratio of the fields between two adjacent updates does not exceed the convergence threshold, and the entity static vector set is obtained.

[0016] Preferably, the step of obtaining the time-series dynamic state vector is as follows:

[0017] Based on the entity static vector set, an event index sequence is generated for each entity node in ascending order of timestamps. A fixed span sliding time window is set, and the start and end times of the window, the event count within the window, and the adjacent difference sequence of the entity static vector set within the window are extracted window by window. The correspondence between the window center time point and the adjacent difference sequence is bound and the boundary processing rules are recorded to generate a dynamic evolution function.

[0018] According to the dynamic evolution function, for any given time point, interval positioning is performed and the two nearest window center time points are retrieved. If the time point is between the two centers, the vector state is calculated by interpolation according to the corresponding relationship. If it is earlier than the minimum timestamp or later than the maximum timestamp, the vector state is extrapolated according to the boundary processing rules. The entity identifier and the vector state are merged to obtain the temporal dynamic state vector.

[0019] Preferably, the step of obtaining the event sequence that meets the time constraints is as follows:

[0020] Based on the topology of the multi-source time-series enrollment knowledge graph and the time-series dynamic state vector, the reachable links from student entities to admission result entities are analyzed, and path query constraints are set. The constraints include non-decreasing time order, continuous event interval not exceeding the upper limit of time span, the difference between the first and last events not less than the lower limit of time span, and event time falling within the effective range of the time-series dynamic state vector, thus obtaining the path query constraints.

[0021] Based on the path query constraints, the system traverses forward layer by layer from the student entity along the topology to the admission result entity. Event identifiers and event times are recorded according to the predicate order of the edges. If the event time is less than the previous event time, the branch is terminated. If the interval between consecutive events exceeds the upper limit of the time span, the branch is pruned. If the difference between the first and last events is less than the lower limit of the time span, the branch is not saved. Once the admission result entity is reached, the event is written into the time-ordered event list, generating an event sequence that meets the time constraints.

[0022] Preferably, the steps for obtaining the high-value evolutionary path pattern are as follows:

[0023] Based on the time-constrained event sequence, the sequence is grouped by predicate order sequence plus entity type sequence. The occurrence frequency and the percentage of student entities covered in each group are counted. The groups are sorted in descending order of occurrence frequency and the groups whose percentage of student entities covered reaches the minimum coverage ratio threshold are selected. Subsequences contained in longer sequences are removed and groups whose starting point is student entities and whose ending point is admission result entities are retained to form a high-value evolution path pattern.

[0024] Preferably, the step of obtaining the set of disturbance sensitivity values ​​is as follows:

[0025] Based on the high-value evolution path pattern, locate the application material subgraph of a single student, enumerate the degradable relationships, adjustable attributes, and removable secondary nodes, sort them in ascending order of event time, and record the original value, timestamp, and target element identifier in sequence to obtain the perturbation execution sequence.

[0026] Based on the perturbation execution sequence, a baseline matching degree is first generated on the unperturbed application materials subgraph, starting from the student entity and extending to the admission result entity. Then, the relationship is downgraded, the attribute is adjusted, or the minor node is removed one by one. Each time, the same path is retrieved and the predicate sequence and attribute constraints are compared in the order of events. The proportion of successfully aligned events is taken as the current matching degree. The difference between the current matching degree and the baseline matching degree is calculated, numbered, and archived to obtain the perturbation sensitivity value set.

[0027] Preferably, the step of obtaining the robustness score of the application materials is as follows: calculate the robustness score of the application materials based on the set of disturbance sensitivity values.

[0028] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0029] This invention integrates heterogeneous data from multiple sources, such as the academic affairs system and the competition registration system, and aligns each record with a timestamp to dynamically reflect the student's development process. Based on the time-aware graph, a dynamic evolution function describing the continuous change of its state over time is derived for each entity, enabling the tracking of student states. Furthermore, the invention traverses and filters the path from student to admission result in the graph, aggregating and generating high-value evolution path patterns with universal significance. This provides an objective data-driven benchmark for evaluation. Perturbation analysis, by perturbing the application material subgraph of a single student through relationship downgrading and attribute adjustment, and quantifying the degree of decrease in its matching degree with high-value patterns, ensures that the evaluation is no longer limited to the accumulation of achievements. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] Please see Figure 1 This invention provides a technical solution, a correlation analysis method for integrating multi-source heterogeneous enrollment data, comprising the following steps:

[0033] Enrollment data in the form of subject, predicate, and object triples are extracted from the academic affairs system and competition registration system. A timestamp or time validity interval is assigned to each triple. The timestamps of enrollment data from different sources are sequence aligned. Based on the relationship between entities, a multi-source time-series enrollment knowledge graph is established.

[0034] Based on the multi-source time-series enrollment knowledge graph, the static vector representation of each entity node in the graph is calculated iteratively to obtain the entity static vector set. Combining the timestamp information with the entity static vector set, a dynamic evolution function describing the change of the vector with time points is derived for each entity. Based on the dynamic evolution function, the state of the student or university entity at any time point is calculated to obtain the time-series dynamic state vector.

[0035] Based on the topological structure of the multi-source temporal enrollment knowledge graph and the temporal dynamic state vector of the nodes, path query constraints including time sequence and time span are set. All paths from the student entity to the admission result entity are traversed in the graph, and paths that do not meet the time sequence or time span requirements are removed using the query constraints to obtain event sequences that meet the time constraints. All event sequences that meet the time constraints are aggregated to generate high-value evolution path patterns.

[0036] Based on the high-value evolution path model, the application material subgraph of a single student is perturbed to obtain a set of perturbation sensitivity values. Based on the set of perturbation sensitivity values, the robustness score of the application materials is calculated.

[0037] The steps to obtain the multi-source time-series enrollment knowledge graph are as follows:

[0038] The records of the academic affairs system and the competition registration system are analyzed, the subject, predicate and object are extracted and timestamps or time validity intervals are configured, the source is deduplicated and the literal is standardized, the subject list and object list are unified, and the sequence is aligned according to the time sequence to obtain the triplet sequence that has been aligned.

[0039] The similarity coefficient is calculated based on the triplet sequence after sequence alignment. The formula is as follows:

[0040] ;

[0041] and and order ,in, For the first entity node, For the second entity node, For predicate index, The total number of predicate types. The information entropy weight of the r-th predicate, The number of different objects of the r-th predicate. For the object of the r-th predicate is The probability, For entities The set of timestamps under the r-th predicate For entities The set of timestamps under the r-th predicate The time decay kernel function value. For any entity, the set of timestamps under any predicate. For another entity, a set of timestamps under the same predicate. for The first in A timestamp, for The first in A timestamp, The time decay coefficient, To weighted normalize the denominator, The similarity coefficient;

[0042] Based on the similarity coefficient, the triplet sequence that has been aligned is traversed in pairs according to the subject candidate and the object candidate, and the minimum and maximum timestamps of each entity are extracted to form an effective interval. The ratio of the intersection length of the two intervals is calculated as the overall time span overlap, and it is jointly judged with the merging threshold and the overlap threshold. When the joint judgment is satisfied, node merging and edge redirection are performed to generate a multi-source time-series enrollment knowledge graph.

[0043] Specifically, based on records from the academic affairs system and the competition registration system, regular expression matching and keyword positioning techniques are used to identify potential subjects, predicates, and objects for the text fields of each record. For example, from the record "Student Zhang San registered for the National Undergraduate Mathematical Modeling Contest on September 5, 2022," the subject "Zhang San," the predicate "registered for," and the object "National Undergraduate Mathematical Modeling Contest" are extracted. The associated date "September 5, 2022" is then extracted from the record's metadata or text content as a timestamp. For continuous events such as "taking the 'Data Structures' course," the start and end dates are extracted from the academic affairs system and configured as "[2022-09-01, The time validity interval is in the format of "[2023-01-15]". Next, the extraction results from different systems undergo source deduplication and literal normalization. This process first performs preliminary matching based on the Jaro-Winkler similarity algorithm for entity names, setting a similarity threshold of 0.9. This threshold is based on calculating the similarity of 1000 pairs of entity names labeled "same" or "different", selecting the similarity value that reaches the highest F1 score of 0.94 on the validation set. Entities with a similarity exceeding 0.9 are considered candidates for the same entity. Then, unique identifiers such as student ID or national ID number are used for final confirmation, eliminating redundant entity records. For entities and relationships with inconsistent literal expressions, [further steps are taken]. Maintain a standardized mapping dictionary for unification. For example, map "National Mathematical Modeling Competition" and "Mathematical Modeling Contest" to "National Undergraduate Mathematical Modeling Contest", and unify predicates such as "participate" and "register" to "participate". After standardization, establish a unified list of subjects (students) and objects (courses, competitions, awards, etc.). Finally, group all cleaned and standardized triplet records according to subject entities, and sort them in each group according to the order of timestamps. For records with valid time intervals, use the start time of the interval as the sorting basis to complete sequence alignment, thereby obtaining a sequence of event triplets organized by time flow for each student, which is the triplet sequence that has been aligned.

[0044] The similarity coefficient calculation formula incorporates information entropy weights. This method quantifies the importance of different types of events (predicates) in distinguishing entities. Predicates with high discriminative power (such as "receiving a national scholarship") are given higher weights, while predicates that are common (such as "taking elective courses") are given lower weights. Furthermore, the formula incorporates a time decay kernel function. This is used to measure the temporal similarity between two entities in the same type of event. The closer the events occur, the greater the contribution to the similarity, and vice versa. This design makes the similarity calculation not only focus on "what" the entity did, but also on "when" it did it, so as to more accurately characterize the relationship between entities. Especially in scenarios with strong temporal characteristics, such as analyzing student development trajectories, it can identify groups of students with similar development paths.

[0045] and The steps to obtain the parameters are as follows: and These represent the first and second entity nodes to be compared. These nodes are selected from the subject and object candidate sets during the entity alignment stage of knowledge graph construction. For example, when deduplicating student entities, the entity "Student ID 2021001, Name Zhang San" extracted from the academic affairs system is used as the entity. The entity "ID number XXXX, name Zhang San" extracted from the competition registration system is used as the entity. These two entities were selected as candidate matching pairs for similarity calculation because they have the same name.

[0046] and The steps to obtain the parameters are as follows: For the index number of the predicate, The total number of predicate types is represented by these two parameters, which are derived from a unified predicate list constructed during the data preprocessing stage. First, all predicates extracted from the original data are collected, such as "participate," "obtain," "study," and "serve." Then, normalization is performed, for example, merging "participate" and "participate" into "participate." Finally, a unique list of predicates is formed, and each predicate is assigned a number from 1 to... Unique integer index For example, if the list of predicates is {1: "participate in a competition", 2: "take a course", 3: "receive an award"}, then... .

[0047] The steps to obtain the parameters are as follows: Representing the The number of distinct objects under each predicate is calculated statistically from the triple sequences after sequence alignment. Specifically, this involves filtering out all predicates that are... The triples are then processed, and the objects within these triples are extracted. The objects are then deduplicated and counted. For example, for the predicate "participate in the competition" (…), ), count the names of all competitions that students have participated in, such as "ACM Competition", "Mathematical Modeling Contest", "Electronic Design Contest". If there are 30 different competitions in total, then .

[0048] The steps to obtain the parameters are as follows: Indicates that when the given predicate is Under the condition that its object is The probability of this is obtained through frequency estimation, and the calculation formula is: For example, in the dataset, the predicate is "participate in the competition" ( There are a total of 1000 triples, where the object is "ACM competition" ( If there are 50 triples in a given pair, then .

[0049] The steps to obtain the parameters are as follows: As the first The information entropy weight of a predicate is calculated based on its corresponding object distribution. It reflects the degree of certainty of the information provided by the predicate. The calculation first requires obtaining... The probability of all objects Then substitute it into the information entropy formula Finally, normalization is performed to obtain the weights. This normalization ensures that the weight values ​​are within the range of [0, 1], which facilitates comparison and weighted summation between different predicates.

[0050] and The steps to obtain the parameters are as follows: It is a physical entity In the The set of timestamps of all events under a predicate is used to filter out events whose subject is a sequence of triples that have been aligned by traversing the sequence of triples. And the predicate is All triples are processed, and their timestamp fields are extracted to form a set. The method of obtaining it is the same, except that the subject is changed to the entity. For example, entities (Zhang San) has two records of "participating in a competition", with dates of "2022-10-15" and "2023-09-01" respectively. .

[0051] The steps to obtain the parameters are as follows: This is the time decay coefficient, used to control the degree of influence of time differences on similarity calculation. Its value reflects the prior judgment on the timeliness of the event. The method for setting it is to first determine a half-life. That is, the time difference between the occurrence of the events When the similarity is reduced to half, its impact is then calculated using the formula. Inverse solution The calculation formula is: For example, in the scenario of student enrollment data analysis, it is considered reasonable that the correlation between events separated by half an academic year (approximately 180 days) is halved, then the following setting is made: Day, calculated .

[0052] Calculations based on parameters:

[0053] To compute two student entities (Zhang San) and Taking the similarity coefficient of (Li Si) as an example, let the predicate set be {1: "participate in the competition", 2: "take courses"}, therefore Time decay coefficient .

[0054] Calculate predicate weights :

[0055] For the predicate "participate in the competition" ( For example, the dataset contains a total of The probabilities of occurrence for different types of competitions are as follows: .

[0056] Information entropy .

[0057] Weight .

[0058] For the predicate "courses taken" ( For example, there are a total of The courses are more evenly distributed, and their weights are calculated. .

[0059] The weighted normalized denominator is .

[0060] Collect a set of timestamps (dates converted to days from the base date):

[0061] For example, Zhang San ( The event timestamp is: (Participating in the competition) (Courses taken).

[0062] For example, Li Si ( The event timestamp is: (Participating in the competition) (Courses taken).

[0063] Calculate the time decay kernel value :

[0064] for (Participating in the competition):

[0065] .

[0066] .

[0067] .

[0068] for (Courses taken):

[0069] .

[0070] .

[0071] .

[0072] Calculate the final similarity coefficient :

[0073] The result indicates that the student entity Zhang San ( ) and Li Si ( The similarity coefficient between the two is 0.926, which is very close to 1. This indicates that the two have a high degree of similarity in the types of events examined (participating in competitions and taking courses). Not only do the types of events they participate in have a high degree of overlap, but the time points of these events are also very close.

[0074] Based on the similarity coefficients of each candidate entity pair calculated in the previous step, and the triplet sequences after sequence alignment, for each pair of subject candidates (e.g., "Wang Wu" in the academic affairs system and "Wang Wu" in the competition system) or object candidates (e.g., "School of Information" and "School of Computer and Information Engineering"), iterate through all associated events in the triplet sequence, extract the minimum and maximum values ​​of all event timestamps for each entity, and construct their respective effective time intervals. For example, if the earliest record time of entity "Wang Wu" is September 1, 2021, and the latest record time is June 30, 2023, its effective interval is [2021-09-01, 2023-06-30]. Then, calculate the ratio of the intersection length to the union length of the effective time intervals of this pair of candidate entities. This ratio is defined as the overall time span overlap, and the calculation formula is as follows: ,in and Entities and The effective time interval is then determined, followed by a joint judgment based on two preset thresholds: a similarity merging threshold and a time overlap threshold. The similarity merging threshold is set to 0.85. This value was obtained by testing different thresholds on a sample set containing 500 manually labeled pairs of entities, selecting the point that maximizes the harmonic mean (F1 score) of precision and recall. The time overlap threshold is set to 0.6. This value is based on analyzing the time overlap distribution of 1000 confirmed matching pairs, finding that 95% of the matching pairs have an overlap higher than 0.6. Only when the similarity coefficient of a pair of candidate entities is greater than 0.6 will the overlap be determined. The joint decision result is "satisfied" only when the merging threshold is met and the overall time span overlap is also greater than the time overlap threshold. Once the joint decision is met, the node merging and edge redirection operation is immediately executed. Specifically, a new, unique entity node is created as the merged canonical node, all attribute information of the two old nodes is merged (e.g., merging student ID and national ID number), and all edges pointing to these two old nodes and edges emanating from them are redirected to this new canonical node. By repeating this process for all candidate pairs, a multi-source time-series enrollment knowledge graph with redundant entities and a clear structure is finally generated.

[0075] The steps to obtain the entity static vector set are as follows:

[0076] Based on the multi-source time-series enrollment knowledge graph, the in-degree and out-degree counts, associated predicate identifier sequences, and attribute field value coverage are summarized one by one according to the entity nodes. A relation count table, attribute coverage table, and unique neighbor count field are established according to a fixed index. The hierarchical neighbor statistics are iteratively updated until the difference ratio of the fields between two adjacent updates does not exceed the convergence threshold, and the entity static vector set is obtained.

[0077] Specifically, based on the multi-source time-series enrollment knowledge graph, for each entity node in the graph, the zero-order feature initialization extraction is first performed. This involves traversing all incoming and outgoing edges directly connected to the node, counting their number to obtain the in-degree and out-degree counts. Simultaneously, the globally unique identifiers of predicates on all outgoing edges are collected, deduplicated, and sorted by identifier size to form a sequence of associated predicate identifiers. Furthermore, based on the entity type (e.g., student, university, course), the required attribute field list is obtained from a predefined attribute template. For example, a student entity should include fields such as "year of enrollment," "major," and "GPA." Then, it is checked whether the values ​​of these attribute fields in the current entity node are empty, and the ratio of the number of non-empty fields to the total number of required fields is calculated to obtain the attribute field value coverage. The in-degree, out-degree, multi-hot encoded predicate sequence, and coverage values ​​are concatenated to form the initial feature vector for each node, and then summarized into a sequence indexed by the entity's unique identifier. The relation count table and attribute coverage table are then used. Following this, the iterative update phase of hierarchical neighbor statistics begins. In the k-th iteration, for any node u, its feature vector in the (k+1)-th iteration is calculated by aggregating the feature vectors of all its first-order neighbor nodes in the k-th iteration. The aggregation operation uses mean pooling, which means averaging the feature vectors of all neighbor nodes by dimension. Then, this aggregated vector is weighted and summed with node u's own feature vector in the k-th iteration. The weight is set to 0.5, indicating that its own information is as important as its neighbor information. This iterative process continues until the change in the feature vectors of all nodes in the entire graph between two adjacent iterations becomes stable. The stability criterion is the average of the relative change rates of all feature dimensions for all nodes. The iteration terminates when this average does not exceed a preset convergence threshold of 0.001. This convergence threshold of 0.001 is determined by a grid search on the validation set, specifically within the range [0.01, ...]. Among the candidate values ​​[0.005, 0.001, 0.0005, 0.0001], the value that maximizes the accuracy of downstream tasks (such as entity classification) and has a relatively small number of iterations is selected. After the iteration converges, the final feature vector of each entity node constitutes the entity static vector set.

[0078] The steps for obtaining the time-series dynamic state vector are as follows:

[0079] Based on the entity static vector set, an event index sequence is generated for each entity node in ascending order of timestamp. A fixed span sliding time window is set, and the start and end times of the window, the event count within the window, and the adjacent difference sequence of the entity static vector set within the window are extracted window by window. The correspondence between the window center time point and the adjacent difference sequence is bound and the boundary processing rules are recorded to generate a dynamic evolution function.

[0080] Based on the dynamic evolution function, for any given time point, interval positioning is performed and the two nearest window center time points are retrieved. If the time point is between the two centers, the vector state is calculated by interpolation according to the corresponding relationship. If it is earlier than the minimum timestamp or later than the maximum timestamp, the vector state is extrapolated according to the boundary processing rules. The entity identifier and the vector state are merged to obtain the temporal dynamic state vector.

[0081] Specifically, based on the entity static vector set and the time information in the multi-source time-series enrollment knowledge graph, a dynamic evolution function is constructed for each entity node. First, for a single entity, all triples with it as the subject are extracted from the knowledge graph, and the timestamp or the start point of the effective time interval of each triple is extracted. These time points are arranged in ascending order to form the original event time series of the entity. Next, a sliding time window with a fixed span of 180 days (representing one semester) is set, and the window slides across the entity's time series with a step size of 30 days (representing one month). For each window, a feature extraction operation is performed. The extracted features include the start and end times of the window, the total number of events falling within the window, and a difference vector describing the "change" within the window. The difference vector is calculated by finding all events falling within the window. For each event, its associated object entity is found, and the entity static vector of the object entity is obtained. The static vectors of these object entities are averaged to obtain a vector representing the average characteristics of the interactive objects within the window. This vector is considered as the adjacent difference sequence within the window. Subsequently, the center time point of each window (i.e., the average of the start time and the end time) is bound to the features extracted from that window (event count, difference vector) to form a key-value pair, where the key is the center time point and the value is the set of features. The collection of key-value pairs from all windows of an entity constitutes the main part of the entity's dynamic evolution function. Finally, boundary handling rules are added to the function. The rules are defined as follows: for queries earlier than the first event time point, the state change is a zero vector; for queries later than the last event time point, the state change rate is constant at the state change rate of the last valid window. After this series of operations, a discrete function expression that can describe the evolution of its state over time, i.e., the dynamic evolution function, is generated for each entity.

[0082] Based on the dynamic evolution function, for any given entity and query time point, its temporal dynamic state vector at that moment is calculated. First, in the corresponding entity's dynamic evolution function (i.e., the list of key-value pairs between time points and change vectors), a binary search method is used to quickly locate the interval where the query time point is located, that is, to find the center time points of two consecutive windows. and , making If the query time point falls exactly on a central time point, the state at that point is used directly. If the query time point is between two central points, linear interpolation is used to calculate the current state vector. The specific calculation process is as follows: first, the relative position ratio of the query time point within that time interval is calculated, using the formula... Then, obtain the time from the initial moment to... The sum of all the difference vectors accumulated at each time step is used as... The fundamental state vector at time 1 , then obtain in The state change that occurs within the interval, that is, the change that corresponds to difference vector Then the state vector at the query time point is calculated as follows: This yields the vector state for interpolation calculation. If the query time is earlier than the minimum timestamp recorded in the dynamic evolution function (i.e., the center point of the first window), then according to the boundary handling rules, its state vector is set to zero, indicating that its dynamic state is empty before the start of the entity's historical record. If the query time is later than the maximum timestamp recorded (i.e., the center point of the last window), then extrapolation calculation is performed according to the boundary rules, first calculating the time difference from the maximum timestamp to the query time. Then, obtain the rate of change per unit time for the last window (i.e., the last difference vector divided by the window step length of 30 days), and multiply this rate of change by... The extrapolated change is obtained, and this change is added to the cumulative state vector at the maximum timestamp to obtain the extrapolated vector state. Finally, the calculated vector state is merged with the entity's unique identifier to form a complete data structure containing the entity's identity and its state at a specific point in time, namely the time-series dynamic state vector.

[0083] The steps to obtain a time-constrained event sequence are as follows:

[0084] Based on the topology and time-series dynamic state vector of the multi-source time-series enrollment knowledge graph, the reachable link from the student entity to the admission result entity is analyzed, and path query constraints are set. The constraints include non-decreasing time order, continuous event interval not exceeding the upper limit of time span, the difference between the first and last events not less than the lower limit of time span, and the event time falling within the effective range of the time-series dynamic state vector.

[0085] Based on the path query constraints, the system traverses forward layer by layer from the student entity along the topology to the admission result entity. Event identifiers and event times are recorded according to the predicate order of the edges. If the event time is less than the previous event time, the branch is terminated. If the interval between consecutive events exceeds the upper limit of the time span, the branch is pruned. If the difference between the first and last events is less than the lower limit of the time span, the branch is not saved. Once the admission result entity is reached, it is written into the time-ordered event list, generating an event sequence that meets the time constraints.

[0086] Specifically, based on the topological structure and temporal dynamic state vector of the multi-source temporal enrollment knowledge graph, a preliminary connectivity analysis is first performed on all "student" type entity nodes and "admission result" type entity nodes in the graph. A breadth-first search algorithm is used to generate an initial set containing all reachable paths for each student entity. Next, this initial path set is filtered and constrained, setting a series of path query constraints. The first constraint is that the time order is non-decreasing, meaning that for any two consecutive events in the path (represented by edges in the graph), the timestamp of the latter event must be greater than or equal to the timestamp of the former event. The second constraint is that the interval between consecutive events does not exceed the upper limit of the time span, which is set to 365 days. This upper limit is based on the statistical analysis of the event sequences of all successfully admitted students in the dataset. The analysis calculates the time interval between all adjacent events and takes the 95th percentile as the threshold. This means that the vast majority of meaningful related activities occur within one year. The third constraint is that the difference between the first and last events is not less than the lower limit of the time span, which is set at 730 days (two years). The basis for this setting is that a valuable path that can reflect the long-term development of students should cover at least two complete academic years, filtering out accidental or short-term event sequences. The fourth constraint is that the event time falls within the valid range of the time-series dynamic state vector. That is, the timestamp of each event in the path must be within the start and end time range defined by the time-series dynamic state vector of the event subject (student entity). Events outside this range are considered invalid. These four constraints are combined into a structured query object to obtain the path query constraints.

[0087] Based on the path query constraints obtained in the previous step, a path traversal and pruning process based on depth-first search is initiated. This process starts with each "student" entity and aims to reach any "admission result" entity. Starting from the student entity, it explores the neighboring nodes connected by its outgoing edges layer by layer. For each edge (representing an event), the predicate identifier, target node (object entity), and timestamp of the edge's attribute records are appended to the currently constructed path sequence. During each step of path expansion, the path query constraints are immediately applied for pruning. Specifically, when attempting to move from the current event to the next event, it first checks whether the timestamp of the next event is less than the timestamp of the current event. If so, the non-decreasing time constraint is violated, and the exploration of the current search branch is immediately terminated, backtracking to the previous section. First, calculate the difference between the timestamp of the next event and the current event. If the difference is greater than the preset time span limit (365 days), the correlation between the events is considered too weak, and this search branch is also pruned. When a path successfully reaches an "admission result" entity, the final path validity check is performed. Calculate the difference between the timestamp of the end event (admission) and the start event of the path. If the difference is less than the preset time span limit (730 days), the overall span of the path is considered insufficient and is not saved. Only when a path from a student entity to an admission result entity has no intermediate steps pruned and finally meets the first and last event difference constraint is it considered a valid event sequence and written into a time-ordered event list. After traversing all student entities, this list summarizes all event sequences that meet the time constraints.

[0088] The steps to obtain high-value evolutionary path patterns are as follows:

[0089] Based on the event sequence that meets the time constraints, the sequence is grouped by predicate order sequence plus entity type sequence. The occurrence frequency and the percentage of student entities covered in each group are counted. The groups are sorted in descending order of occurrence frequency and the groups whose percentage of student entities covered reaches the minimum coverage ratio threshold are selected. Subsequences contained in longer sequences are removed and groups whose starting point is student entities and whose ending point is admission result entities are retained to form a high-value evolution path pattern.

[0090] Specifically, based on the set of event sequences that meet time constraints, a path pattern discovery and extraction process is executed. First, each event sequence is abstracted to generate a composite key consisting of a "predicate sequence" and an "entity type sequence". For example, a sequence "student -> participate -> competition -> obtain -> award -> be admitted -> admission result" has the composite key (('participate', 'obtain', 'be admitted'), ('competition', 'award', The 'admission results' pattern is used to group all event sequences using this composite key. For each group, the total number of occurrences of the pattern (i.e., the composite key) across all sequences is calculated, along with the number of independent student entities covered by these sequences. The percentage of each covered student entity is then calculated. Next, all groups (patterns) are sorted in descending order of frequency and a minimum coverage threshold is applied for initial screening. This threshold is set to 1%. It is determined by plotting a Pareto chart of pattern coverage, selecting the smallest set of patterns that covers a cumulative 80% of the student population, and setting the coverage rate of the pattern with the lowest coverage rate in this set as the threshold. The first step is to retain common path patterns that affect at least 1% of the student population. Then, redundancy removal is performed on the filtered pattern set. Specifically, for any two patterns A and B, if the predicate sequence and entity type sequence of pattern A are both subsequences of pattern B, then pattern A is removed, retaining longer and more complete path patterns. Finally, a final verification is performed to ensure that the starting entity type of each retained pattern must be "student" and the ending entity type must be "admission result". After the complete process of grouping, statistics, sorting, threshold filtering, subsequence removal and endpoint verification, the final pattern set is the high-value evolutionary path pattern.

[0091] The steps for obtaining the set of perturbation sensitivity values ​​are as follows:

[0092] Based on the high-value evolution path pattern, locate the application material subgraph of a single student, enumerate the degradable relationships, adjustable attributes, and removable secondary nodes, sort them in ascending order of event time, and record the original value, timestamp, and target element identifier in sequence to obtain the perturbation execution sequence.

[0093] Based on the perturbation execution sequence, a baseline matching degree is first generated on the unperturbed application materials subgraph, starting from the student entity and extending to the admission result entity. Then, the relationship is downgraded, the attribute is adjusted, or the minor node is removed one by one. Each time, the same path is retrieved and the predicate sequence and attribute constraints are compared in the order of events. The proportion of successfully aligned events is taken as the current matching degree. The difference between the current matching degree and the baseline matching degree is calculated, numbered, and archived to obtain the perturbation sensitivity value set.

[0094] Specifically, based on the high-value evolutionary path pattern obtained in the previous step, the first step is to perform a local subgraph extraction operation on a single student entity within the multi-source temporal enrollment knowledge graph. This operation, centered on the student entity, uses a bidirectional breadth-first search to extract all nodes and edges within a three-hop range of the student entity, forming a subgraph of the student's application materials. Next, all possible minor negative perturbations are systematically enumerated on this subgraph. These perturbations are divided into three categories: the first category is degradable relations, determined by consulting a pre-constructed relation hierarchy mapping table. This mapping table classifies predicates into different levels based on the importance of the event; for example, the object of the "obtain" relation, "national-level award," is downgraded to "provincial-level award," or the object of the "serve" relation, "student council president," is downgraded to "student council minister." The second category is adjustable attributes, mainly targeting attributes with numerous... The first category consists of attribute fields with values ​​or ordered levels, such as lowering a course grade from 95 to 90 or a GPA from 3.9 to 3.8, with the reduction set to 5% of the original value or a minimum level unit. The second category consists of removable secondary nodes. The criteria for determining a secondary node are that its degree in the subgraph is less than or equal to 2 and it does not belong to any key entity type in any high-value evolution path pattern (such as core courses or high-level competitions). For example, a node for a one-time volunteer activity or a regular community activity. After enumerating all possible perturbations, each perturbation is recorded with its operation type, the target element of the operation (the unique identifier of the node or edge), the original value, the value after perturbation, and the timestamp of the associated event. Finally, all recorded perturbation items are sorted in ascending order from earliest to latest according to the event timestamp to form an ordered list of operations, i.e., the perturbation execution sequence.

[0095] Based on the perturbation execution sequence and the student's original application material subgraph, a high-value evolutionary path pattern that best matches the student is selected. Then, starting from the student entity node, the predicate sequence and entity type sequence in the pattern are attempted to be completely matched along the subgraph's topology. Since the original, unperturbed subgraph is used, theoretically all events should be successfully matched. Therefore, the proportion of successfully matched events to the total number of events in the path pattern is calculated to obtain a baseline matching degree with an initial value of 1.0. Subsequently, an iterative perturbation test loop is started. This loop iterates through each perturbation item in the perturbation execution sequence. At the beginning of each iteration, a temporary copy of the application material subgraph is created, and then the single perturbation operation corresponding to the current iteration is performed on this copy, for example, changing a course grade from 95 to 90. After the operation is completed, the same high-value evolutionary path pattern that was completely matched during the baseline evaluation is immediately applied to the perturbed temporary subgraph. Using the same path retrieval logic and high-value evolution path pattern, the matching degree is recalculated. Due to the presence of perturbation, a certain event node or its attribute may no longer meet the requirements of the path pattern, resulting in matching failure. For example, if the path pattern requires "course score greater than 90", and after perturbation it is exactly equal to 90, then the event matching fails. The proportion of successfully aligned events in this matching is calculated to obtain the current matching degree, which will be less than or equal to 1.0. Then, the current matching degree is subtracted from the baseline matching degree (1.0), and the result is the sensitivity value caused by this perturbation. This sensitivity value is recorded together with the number of the perturbation item (i.e., its index in the sequence). After completing one iteration, the temporary copy is discarded to ensure that the next iteration is performed on a clean copy on which only the next perturbation has been applied. After traversing the entire perturbation execution sequence, all the recorded sensitivity values ​​constitute the perturbation sensitivity value set of the student.

[0096] The steps to obtain the robustness score of application materials are as follows:

[0097] Based on the set of disturbance sensitivity values, the robustness score of the application materials is calculated using the following formula:

[0098] ;

[0099] in, To score the robustness of the application materials, The total number of disturbances. For the first The severity coefficient of each disturbance, For the first A disturbance sensitivity value, This is the sensitivity scaling factor. To sum the index from one to the last. .

[0100] Specifically, the robustness score calculation formula for application materials integrates the effects of multiple and diverse disturbances into a comprehensive score between 0 and 1 through an exponential decay function, where the sensitivity value of each disturbance is... Based on its severity coefficient Weighting is applied, which gives greater weight to significant disturbances affecting the application materials (such as downgrading core awards) in the overall score, while giving less significant disturbances (such as minor adjustments to general course grades) a smaller weight. The entire weighted average is then scaled by a sensitivity scaling factor. Adjustments are made, and then the result is passed as a negative exponent to the exponential function. This design leads to a higher average sensitivity (the more unstable the material) and a higher final robustness score. The lower the value, the higher the value, and vice versa, thus intuitively reflecting the stability of the application materials in the face of minor negative changes.

[0101] The steps to obtain the parameters are as follows: This represents the total number of perturbations performed on a single student's application material subgraph. This value is directly derived from the length of the perturbation execution sequence generated in the preceding steps. After analyzing the student's application material subgraph and enumerating all degradable relationships, adjustable attributes, and removable secondary nodes, the total number of entries contained in the resulting perturbation execution sequence is this value. For example, if an analysis of a student's materials identifies 3 awards that can be downgraded, 5 course grades that can be lowered, and 2 club activities that can be removed, then a sequence containing 10 perturbation terms will be generated. .

[0102] The steps to obtain the parameters are as follows: It is the first The perturbation sensitivity value generated by each perturbation quantifies the impact of a single perturbation on the pattern matching degree of high-value evolutionary paths. This value is calculated from the previous step, specifically, the baseline matching degree (typically 1.0) obtained on the unperturbed application material subgraph compared with the value obtained after applying the perturbation. The difference between the current matching degree obtained after each perturbation and the current matching degree is the set of elements in the perturbation sensitivity value set generated in the previous step. For example, if removing a minor node causes the path matching degree to drop from 1.0 to 0.9, then the perturbation corresponds to... Value .

[0103] The steps to obtain the parameters are as follows: It is the first A severity coefficient for each disturbance is used to distinguish the importance of different disturbances. This coefficient is set based on a multi-dimensional scoring rule that comprehensively considers the type, magnitude, and importance of the target element of the disturbance. The calculation formula is as follows: ,in The base score is for the type of disturbance. Based on experience, "node removal" is set at 5 points, "relationship downgrade" at 3 points, and "attribute adjustment" at 1 point. This is the magnitude of change factor. For numerical attributes, it is calculated as the ratio of the change to the possible range of the attribute. For example, if GPA is reduced by 0.1, and the GPA range is 4, then the factor is... For changes in level, such as an award being downgraded from national level (level 4) to provincial level (level 3), the factor is... ; This is the importance score of the target element, determined by analyzing the frequency of its appearance in all high-value evolutionary path patterns. The higher the frequency, the higher the score. For example, if a competition node appears in 80% of high-value paths, its importance score is 0.8. Finally, the three factors are multiplied together and normalized to ensure the value falls between 1 and 10. For example, a "relationship downgrade" perturbation ( If the amplitude is 0.25 and the element importance is 0.8, then its severity coefficient is... After a subsequent linear mapping to the [1,10] interval, the final coefficient is obtained, for example, 3.5.

[0104] The steps to obtain the parameters are as follows: As a sensitivity scaling factor, it is used to adjust the steepness of the curve showing the influence of the weighted average sensitivity on the final robustness score. Its value is determined by calibrating a benchmark dataset containing at least 1000 student samples. First, the weighted average sensitivity value of all students in the dataset is calculated, and its distribution is analyzed. Then, a target is set; for example, students with an average sensitivity at the median are expected to have a robustness score of 0.7, while students with an average sensitivity at the 90th percentile are expected to have a robustness score of 0.3. Based on this target, a system of equations is solved... and Determine The appropriate value is the average of the two solutions. Typically, this value is set between 2.0 and 5.0 to obtain good score discrimination. For example, if the calculated median is 0.15 and the 90th percentile is 0.4, then... and Take their average value to get .

[0105] Calculations based on parameters:

[0106] Taking a student's application materials as an example, it was analyzed... Secondary disturbance.

[0107] Disturbance 1: Relationship downgrade (national award -> provincial award), calculate the disturbance sensitivity value. Its severity coefficient .

[0108] Perturbation 2: Attribute adjustment (GPA 3.9 -> 3.8), calculate the perturbation sensitivity value. Its severity coefficient .

[0109] Perturbation 3: Removal of secondary nodes (club activities), and calculation of perturbation sensitivity value. Its severity coefficient .

[0110] Sensitivity scaling factor set to .

[0111] First, calculate the weighted average sensitivity:

[0112] ;

[0113] ;

[0114] Then, substitute this value into the robustness score formula:

[0115] ;

[0116] ;

[0117] The results indicate that the student's application materials had a robustness score of 0.601, which falls between 0 and 1, placing it in the upper-middle range. The score not being close to 1 suggests some vulnerability in the application materials, particularly in their sensitivity to changes in the level of core awards (as shown by perturbation 1). However, overall, most minor perturbations (such as perturbations 2 and 3) did not cause a significant decrease in path matching, indicating that the main structure of the materials was relatively stable. Setting a threshold, such as 0.75, materials scoring above 0.75 are considered highly robust, while those below 0.5 are considered fragile and require further examination of the authenticity and stability of key information. The score of 0.601 suggests that the student's application materials are of acceptable quality but have weaknesses.

[0118] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A correlation analysis method for integrating multi-source heterogeneous enrollment data, characterized in that, Includes the following steps: Enrollment data in the form of subject-verb-object triples are extracted from the academic affairs system and competition registration system. A timestamp or time validity interval is assigned to each triple. The timestamps of enrollment data from different sources are sequence-aligned. Based on the association between entities, a multi-source time-series enrollment knowledge graph is established; the subject is the student name. Based on the multi-source time-series enrollment knowledge graph, the static vector representation of each entity node in the graph is calculated iteratively to obtain the entity static vector set. Combining the timestamp information with the entity static vector set, a dynamic evolution function describing the change of the vector with time points is derived for each entity. Based on the dynamic evolution function, the state of the student or university entity at any time point is calculated to obtain the time-series dynamic state vector. Based on the topology of the multi-source temporal enrollment knowledge graph and the temporal dynamic state vectors of the nodes, path query constraints including temporal order and time span are set. All paths from the student entity to the admission result entity are traversed in the graph, and paths that do not meet the temporal order or time span requirements are pruned using the query constraints to obtain event sequences that meet the time constraints. All event sequences that meet the time constraints are aggregated to generate a high-value evolutionary path pattern. The admission result entity is specifically the object entity node that serves as the endpoint of the path in the multi-source temporal enrollment knowledge graph. According to the high-value evolution path pattern, the application material subgraph of a single student is perturbed to obtain a set of perturbation sensitivity values. Based on the set of perturbation sensitivity values, the robustness score of the application materials is calculated. Specifically, the perturbation sensitivity value in the set of perturbation sensitivity values ​​is the difference between the baseline matching degree obtained on the unperturbed application material subgraph and the current matching degree recalculated after applying the perturbation.

2. The correlation analysis method for integrating multi-source heterogeneous enrollment data according to claim 1, characterized in that, The steps for obtaining the multi-source time-series enrollment knowledge graph are as follows: The records of the academic affairs system and the competition registration system are analyzed, the subject, predicate and object are extracted and timestamps or time validity intervals are configured, the source is deduplicated and the literal is standardized, the subject list and object list are unified, and the sequence is aligned according to the time sequence to obtain the triplet sequence that has been aligned. Calculate the similarity coefficient based on the triplet sequence that has been aligned. Based on the similarity coefficient, the triplet sequence that has been aligned is traversed in pairs according to the subject candidate and the object candidate, and the minimum and maximum timestamps of each entity are extracted to form an effective interval. The ratio of the intersection length of the two intervals is calculated as the overall time span overlap, and it is jointly judged with the merging threshold and the overlap threshold. When the joint judgment is satisfied, node merging and edge redirection are performed to generate a multi-source time-series enrollment knowledge graph.

3. The correlation analysis method for integrating multi-source heterogeneous enrollment data according to claim 1, characterized in that, The steps for obtaining the entity static vector set are as follows: Based on the multi-source time-series enrollment knowledge graph, the in-degree and out-degree counts, associated predicate identifier sequences, and attribute field value coverage are summarized one by one according to the entity nodes. A relation count table, attribute coverage table, and unique neighbor count field are established according to a fixed index. The hierarchical neighbor statistics are iteratively updated until the difference ratio of the fields between two adjacent updates does not exceed the convergence threshold, and the entity static vector set is obtained.

4. The correlation analysis method for integrating multi-source heterogeneous enrollment data according to claim 1, characterized in that, The steps for obtaining the time-series dynamic state vector are as follows: Based on the entity static vector set, an event index sequence is generated for each entity node in ascending order of timestamps. A fixed span sliding time window is set, and the start and end times of the window, the event count within the window, and the adjacent difference sequence of the entity static vector set within the window are extracted window by window. The correspondence between the window center time point and the adjacent difference sequence is bound and the boundary processing rules are recorded to generate a dynamic evolution function. According to the dynamic evolution function, for any given time point, interval positioning is performed and the two nearest window center time points are retrieved. If the time point is between the two centers, the vector state is calculated by interpolation according to the corresponding relationship. If it is earlier than the minimum timestamp or later than the maximum timestamp, the vector state is extrapolated according to the boundary processing rules. The entity identifier and the vector state are merged to obtain the temporal dynamic state vector.

5. The correlation analysis method for integrating multi-source heterogeneous enrollment data according to claim 1, characterized in that, The steps for obtaining the event sequence that meets the time constraints are as follows: Based on the topology of the multi-source time-series enrollment knowledge graph and the time-series dynamic state vector, the reachable links from student entities to admission result entities are analyzed, and path query constraints are set. The constraints include non-decreasing time order, continuous event interval not exceeding the upper limit of time span, the difference between the first and last events not less than the lower limit of time span, and event time falling within the effective range of the time-series dynamic state vector, thus obtaining the path query constraints. Based on the path query constraints, the system traverses forward layer by layer from the student entity along the topology to the admission result entity. Event identifiers and event times are recorded according to the predicate order of the edges. If the event time is less than the previous event time, the branch is terminated. If the interval between consecutive events exceeds the upper limit of the time span, the branch is pruned. If the difference between the first and last events is less than the lower limit of the time span, the branch is not saved. Once the admission result entity is reached, the event is written into the time-ordered event list, generating an event sequence that meets the time constraints.

6. The correlation analysis method for integrating multi-source heterogeneous enrollment data according to claim 1, characterized in that, The steps for obtaining the high-value evolutionary path pattern are as follows: Based on the time-constrained event sequence, the sequence is grouped by predicate order sequence plus entity type sequence. The occurrence frequency and the percentage of student entities covered in each group are counted. The groups are sorted in descending order of occurrence frequency and the groups whose percentage of student entities covered reaches the minimum coverage ratio threshold are selected. Subsequences contained in longer sequences are removed and groups whose starting point is student entities and whose ending point is admission result entities are retained to form a high-value evolution path pattern.

7. The correlation analysis method for integrating multi-source heterogeneous enrollment data according to claim 1, characterized in that, The steps for obtaining the set of disturbance sensitivity values ​​are as follows: Based on the high-value evolution path pattern, locate the application material subgraph of a single student, enumerate the degradable relationships, adjustable attributes, and removable secondary nodes, sort them in ascending order of event time, and record the original value, timestamp, and target element identifier in sequence to obtain the perturbation execution sequence. Based on the perturbation execution sequence, a baseline matching degree is first generated on the unperturbed application materials subgraph, starting from the student entity and extending to the admission result entity. Then, the relationship is downgraded, the attribute is adjusted, or the minor node is removed one by one. Each time, the same path is retrieved and the predicate sequence and attribute constraints are compared in the order of events. The proportion of successfully aligned events is taken as the current matching degree. The difference between the current matching degree and the baseline matching degree is calculated, numbered, and archived to obtain the perturbation sensitivity value set.

8. The correlation analysis method for integrating multi-source heterogeneous enrollment data according to claim 1, characterized in that, The steps for obtaining the robustness score of the application materials are as follows: calculate the robustness score of the application materials based on the set of disturbance sensitivity values.

Images