Knowledge Graph-Based Learning Risk Assessment Method

Abstract

This invention discloses a knowledge graph-based method for assessing learning risks, belonging to the field of educational informatization and student management risk assessment. To address the problems of scattered multi-source learning data, inconsistent fields, missing or duplicate names, and time-series breakpoints leading to unstable risk identification and intervention ranking, this invention standardizes learning data, constructs a heterogeneous time-series learning knowledge graph, aggregates related features using a time-aware heterogeneous graph transformer, and combines conformal uncertainty calibration to generate intervention ranking. This achieves the technical effect of improving the relevance, interpretability, and ranking stability of risk identification.

Description

Technical Field

[0001] This invention relates to the field of educational informatization and student management risk assessment, and in particular to a learning risk assessment method based on knowledge graphs. Background Technology

[0002] In student management and educational informatization scenarios, grade systems, attendance systems, homework systems, reward and punishment systems, and consultation record systems are typically maintained by different business modules. Existing systems can record data such as student grades, class participation, homework completion, reward and punishment events, and consultation behavior, and generate academic warnings, abnormal behavior alerts, or intervention lists based on preset rules or statistical indicators, providing auxiliary decision-making basis for student management personnel.

[0003] However, existing technologies still have shortcomings in practical applications: First, there are inconsistencies in field names, units of measurement, and value dictionaries among multi-source systems, and there are cases of missing fields, duplicate student names, and incomplete identity information, making it difficult to stably correlate the same student or the same behavioral event; Second, risk identification usually relies on a single indicator threshold or simple weighted scoring, which is difficult to express the complex relationship between students, courses, behavioral events, risk factors, and intervention measures; Third, when there are time gaps or supplementary entries in attendance, assignments, rewards and punishments, and consultation records, the model has difficulty distinguishing between recent changes, historical cumulative effects, and unreliable data factors, thereby affecting the early warning results and intervention priority ranking.

[0004] Therefore, a learning risk assessment method that can address the shortcomings of existing technologies is needed. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based method for assessing learning risks. Addressing the problems of existing technologies, such as scattered multi-source learning data, inconsistent fields, missing or duplicated names, and time-series breakpoints, and the difficulty of using a single threshold to characterize the relationships between students, courses, and behavioral events, this invention proposes a technical solution combining multi-source data standardization, construction of heterogeneous time-series learning knowledge graphs, aggregation using time-aware heterogeneous graph transformers, multi-task risk assessment, and conformal uncertainty calibration. This invention effectively improves the relevance of learning risk identification, the stability of abnormal behavior identification, the interpretability of intervention ranking, and the reliability of ranking in scenarios with missing data.

[0006] This invention provides a knowledge graph-based method for assessing student learning risk, comprising: S1, acquiring student management data from grade systems, attendance systems, homework systems, reward and punishment systems, and consultation record systems; performing field semantic mapping, unit normalization, and entity disambiguation on the student management data to generate standardized student learning data; S2, constructing a heterogeneous time-series student learning knowledge graph containing student entities, course entities, behavioral event entities, risk factor entities, and intervention measure entities based on the standardized student learning data; S3, calculating the edge reliability and time decay of relation edges based on field completeness, source system credibility, entity matching confidence, event time interval, and missing window. S4. Subtract coefficients and write them into the heterogeneous time-series learning knowledge graph; input the heterogeneous time-series learning knowledge graph into the time-aware heterogeneous graph transformer, aggregate related features according to meta-paths, and correct attention weights using edge reliability and time decay coefficients to obtain student risk representations; S5. Input student risk representations into the multi-task risk assessment layer, output academic risk components, abnormal behavior risk components, intervention urgency components, and risk interpretation paths, and generate risk prediction intervals based on the non-compliance scores of calibration samples, and generate intervention ranking results based on the risk prediction intervals; wherein, the influence relationships in the heterogeneous time-series learning knowledge graph include triggering relationships and recommendation relationships.

[0007] Optionally, S1 includes:

[0008] Configure a field semantic mapping table for each source system. The field semantic mapping table records the source field name, target field name, data type, unit of measurement, and value dictionary.

[0009] For numerical fields, the conversion is performed according to the target unit of measurement; for enumerated fields, the value dictionary encoding is performed; and for missing fields, missing tags and missing start and end times are generated.

[0010] Candidate entity pairs are generated based on student name, student ID, class, year of enrollment, and contact information. The entity matching confidence is calculated based on string similarity, document field consistency, and time overlap ratio. Candidate entity pairs with an entity matching confidence greater than or equal to a preset matching threshold are retained as entity disambiguation results.

[0011] Optionally, S2 includes:

[0012] The student identifier, course identifier, behavioral event identifier, risk factor identifier, and intervention measure identifier in the standard learning data are converted into corresponding entity nodes, and the source system identifier, timestamp, missing flag, and normalized field value are written for each entity node.

[0013] Relationship edges are generated according to course selection relationship, participation relationship, occurrence relationship, triggering relationship, recommendation relationship, same class relationship, same course relationship and preceding and following event relationship. The influence relationship includes triggering relationship and recommendation relationship, and the preceding and following event relationship is determined by the timestamp sorting of the corresponding behavioral events of the same student.

[0014] The heterogeneous temporal learning knowledge graph is formed by the entity nodes and the relation edges.

[0015] Optionally, S3 includes:

[0016] The field completeness, source system credibility, and entity matching confidence are mapped to the interval between 0 and 1, and then weighted and summed according to the weights determined by historical labeled samples to obtain the basic edge reliability.

[0017] The missing correction amount is generated based on the overlap duration between the applicable window and the missing window corresponding to the time attribute of the relation edge. The applicable window is determined according to the relation type and includes event time, or effective time and failure time, or rule effective interval. The missing correction amount is used to correct the reliability of the basic edge to obtain the edge reliability.

[0018] The time decay coefficient is obtained by calculating an exponential function value that monotonically decreases as the event time interval increases, based on the event time interval and the preset half-life.

[0019] The edge reliability and the time decay coefficient are used as correction terms for relation edge attributes and attention weights;

[0020] Furthermore, the updating of the edge reliability and the time decay coefficient includes: when a new behavioral event arrives or a supplementary field is confirmed, the edge reliability of the affected relation edge is recalculated based on the source system credibility of the new behavioral event, the field completeness corresponding to the supplementary field, and the entity matching confidence.

[0021] The time decay factor is recalculated based on the time interval between the timestamp of the newly added behavioral event and the timestamp of the previous behavioral event of the same student.

[0022] The recalculated results are written to a versioned graph copy, and the edge reliability and time decay coefficient before the update are retained as rollback records.

[0023] Optionally, S4 includes:

[0024] The node attributes of student entities, course entities, and behavioral event entities are encoded into node initialization vectors, and the interval between the event timestamp and the student's current evaluation time is encoded into a time interval vector.

[0025] Generate an adjacency node set based on the meta-path from student to course to behavioral event, student to classmate to behavioral event, and student to behavioral event to risk factor;

[0026] A query vector, a key vector, and a value vector are generated for the set of adjacent nodes. The edge reliability and the time decay coefficient are multiplied into the dot product of the query vector and the key vector, scored, and then normalized to obtain the corrected attention weight.

[0027] The student risk representation is obtained by aggregating the value vector based on the modified attention weights.

[0028] Furthermore, the training process of the time-aware heterogeneous graph transformer includes: using student subgraphs with historical academic risk labels, historical abnormal behavior risk labels, and historical intervention time limit labels as training samples, and dividing the training set and validation set according to time order;

[0029] Edge shading is performed on relation edges in the training set, and the sampling probability of the edge shading is determined in reverse based on the reliability of the edge.

[0030] Construct a loss function by weighted summation of academic risk cross-entropy loss, abnormal behavior risk cross-entropy loss, intervention urgency ranking loss, and explanation path coverage loss;

[0031] The weights in the loss function are determined based on the inverse normalization of the losses on the validation set, and the model parameters are updated through backpropagation.

[0032] Optionally, S5 includes:

[0033] The multi-task risk assessment layer performs linear mapping on the student risk representation to obtain the academic risk component, the abnormal behavior risk component, and the intervention urgency component.

[0034] Calculate the non-compliance score for the calibration sample, where the non-compliance score is the absolute difference between the true label of the calibration sample and the predicted value of the corresponding risk component;

[0035] According to the preset target coverage rate, the non-compliance scores are sorted in ascending order and the quantile value is taken. The lower bound of the risk prediction interval is the corresponding risk component prediction value minus the quantile value, and the upper bound of the risk prediction interval is the corresponding risk component prediction value plus the quantile value.

[0036] The intervention ranking result is generated based on the lower bound of the risk prediction interval, the upper bound of the risk prediction interval, the interval width, and the lower bound of the synthesis obtained according to the preset synthesis rule.

[0037] Furthermore, generating the intervention ranking result includes: determining the first intervention threshold, the second review threshold, and the width threshold according to preset percentiles based on the lower bound of the risk prediction interval, the upper bound of the risk prediction interval, the interval width, and the lower bound of the synthesis obtained according to preset synthesis rules in the historical calibration samples; the lower bound of the risk prediction interval is the lower bound of the risk prediction interval of the main component or the lower bound of the synthesis obtained according to preset synthesis rules;

[0038] When the lower bound of a student's risk prediction interval or the lower bound of the synthesis obtained according to the preset synthesis rule is greater than or equal to the first intervention threshold, the student is written into the priority intervention queue and arranged in descending order according to the lower bound of the risk prediction interval or the lower bound of the synthesis.

[0039] When a student's risk prediction interval upper bound is greater than or equal to the second review threshold and the interval width of the risk component is greater than or equal to the width threshold, the student is written into the data supplementation or manual review queue, and the risk component type, missing field, source system, and entity disambiguation result that triggered the queue are recorded.

[0040] Furthermore, the generation of the risk interpretation path includes: extracting the entity and relation sequence with path contribution values ​​greater than or equal to the path contribution threshold from the meta-path that generates the corresponding risk component, wherein the path contribution threshold is determined by a preset percentile of the path contribution values ​​in the calibration sample;

[0041] The extracted entity and relation sequences are sorted by timestamp to generate a risk interpretation path that includes data source, event time, relation type, edge reliability, and risk component contribution value;

[0042] When outputting the intervention ranking results, the risk interpretation path is associated with and stored with the corresponding student identifier.

[0043] The beneficial effects of this invention are:

[0044] 1. By performing semantic mapping, unit normalization, and entity disambiguation on fields such as grades, attendance, assignments, rewards and punishments, and consultation records, data scattered across different student management systems can be converted into standardized student learning data, reducing the impact of inconsistent fields, duplicate student names, and unstable entity matching on subsequent risk assessments.

[0045] 2. By constructing a heterogeneous temporal learning knowledge graph that includes student entities, course entities, behavioral event entities, risk factor entities, and intervention measure entities, and organizing data according to relationships such as course selection, participation, occurrence, impact, same class, same course, and preceding and following events, the risk identification process can utilize individual behavior, course participation, and group association information simultaneously, avoiding reliance on a single indicator threshold for judgment.

[0046] 3. By introducing edge reliability and time decay coefficient into the time-aware heterogeneous graph transformer, and combining conformal uncertainty calibration to generate risk prediction intervals during the risk ranking stage, the impact of field completeness, source credibility, entity matching confidence, event interval, and missing window on risk results can be explicitly incorporated into the calculation, thereby improving the stability and verification direction of intervention priority ranking. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0048] Figure 1 This is a flowchart of a knowledge graph-based learning risk assessment method.

[0049] Figure 2 This is a flowchart of step S4, the time-aware heterogeneous graph converter, of the present invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0051] refer to Figures 1-2 The knowledge graph-based learning risk assessment method includes: S1. Acquiring student management data from the grading system, attendance system, homework system, reward and punishment system, and consultation record system; performing field semantic mapping, unit normalization, and entity disambiguation on the student management data to generate standardized learning data; S2. Constructing a heterogeneous time-series learning knowledge graph containing student entities, course entities, behavioral event entities, risk factor entities, and intervention measure entities based on the standardized learning data; S3. Calculating the edge reliability and time decay coefficient of relation edges based on field completeness, source system credibility, entity matching confidence, event time interval, and missing window. S4. The heterogeneous time-series learning information knowledge graph is input into a time-aware heterogeneous graph transformer. Assembled features according to meta-paths, and attention weights are corrected using edge reliability and time decay coefficients to obtain student risk representations. S5. Student risk representations are input into a multi-task risk assessment layer, outputting academic risk components, abnormal behavior risk components, intervention urgency components, and risk interpretation paths. Risk prediction intervals are generated based on the non-compliance scores of calibration samples, and intervention ranking results are generated according to the risk prediction intervals. The influence relationships in the heterogeneous time-series learning information knowledge graph include triggering relationships and recommendation relationships.

[0052] In this specific embodiment, S1 includes:

[0053] The data access module of the learning risk assessment platform uses a semester as the assessment cycle. It pulls student management data from the grade system, attendance system, homework system, reward and punishment system, and consultation record system, respectively. The student management data serves as the input object for this step. The input records include student name, student ID, class, year of enrollment, contact information, course identifier, grade value, attendance status, homework submission status, reward and punishment type, consultation record summary, event occurrence time, and source system identifier. The data access module establishes a field semantic mapping table for each source system. The field semantic mapping table is part of the project data structure and records the source field name, target field name, data type, source unit of measurement, target unit of measurement, value dictionary, whether the field is required, update time, and table version number. The table version number increments after the field definition changes. The mapping table is stored in the metadata repository and is called by this step.

[0054] The field semantic mapping table is generated by the metadata maintenance module before deployment based on the interface specifications of the five types of source systems. The generation rules are: synonymous fields are merged into the same target field, only one target unit of measurement is retained for the same target field, and enumerated values ​​are mapped to stable codes according to business meaning. When the source field name does not match the mapping table, the data access module writes the source field into the unmapped field queue and retains the original value, without writing it into the standard chemical intelligence data table. The incremental normalization task is triggered after the metadata maintenance module supplements the mapping relationship, and the unmapped field queue number is written into the batch summary for subsequent review and location.

[0055] The conversion of units of measurement for numeric fields is controlled by a unit conversion table, which includes the source unit of measurement, target unit of measurement, conversion factor, offset, applicable fields, and version number; the data access module follows... Obtain the converted value in the target unit of measurement, where For the first The converted value of each numeric field. For the source field value, To convert the unit to the ratio, Offset; converted Continue to prune to the target range according to the field value range. Records that exceed the range are written to the exception log and the values ​​before pruning are retained.

[0056] After reading the field semantic mapping table, the data access module generates a standard field set according to the target field name. Numerical fields are converted according to the target unit of measurement, enumerated fields are encoded according to the value dictionary, and text fields are written into the standard text field after removing spaces, normalizing full-width and half-width characters, and replacing synonyms. The percentage system for grades is converted to zero to one with 100 as the denominator. The five-level system and grade system are converted to zero to one by dividing the grade number in the value dictionary by the highest grade number. When the conversion result is less than zero, it is set to zero, and when it is greater than one, it is set to one, to obtain the normalized grade value, which is in the range of 0 to 1 and is based on the field configuration range. Attendance status is encoded into five categories: present, late, early departure, leave, and absent. Assignment status is encoded into four categories: submitted on time, submitted late, not submitted, and submitted retroactively. Missing fields are written with a missing tag and missing start and end time. The missing start and end time is limited by the collection time when the missing is first discovered and the field supplement confirmation time. The target unit of measurement, value dictionary, and required field configuration in the field semantic mapping table constitute the parameter source for this step.

[0057] For duplicate records within the same source system, after field mapping and unit normalization but before entity disambiguation, the data access module establishes a collection batch cache using the source system student identifier, source system course identifier, event type, and event occurrence time as pre-deduplication keys. When the same pre-deduplication key corresponds to multiple records, the record with the latest source system timestamp and the highest field completeness is retained, and the source system identifier and collection batch number of the removed record are written to the duplicate record log. Field completeness is determined according to... Calculation, where For the first Record-level field completeness of each record. This represents the number of non-missing required fields in the record. This represents the total number of required fields configured in the field semantic mapping table for this source system. The value is truncated to between zero and one and serves as the sole basis input for the completeness of the S3 summary relation edge field;

[0058] The entity disambiguation module generates candidate entity pairs based on student name, student ID, class, year of enrollment, and contact information. During candidate generation, an exact index is first created using the student ID and year of enrollment, followed by a ternary inverted index for the name and contact information. Two student records within the same index bucket are added to the candidate entity pair set. For each candidate entity pair, the module calculates the entity matching confidence score. ,in For the first Entity matching confidence of candidate entity pairs For name string similarity, To ensure consistency of the document fields, The percentage of overlap between classes or time periods of study. , and The weights are determined by historical manually merged samples, and the sum of the three is one; in this specific implementation, the preset matching threshold is 0.82, when... When the value is greater than or equal to 0.82, the candidate entity pair is retained as the entity disambiguation result. When the value is less than 0.70, the candidate entity pair is removed. When the value is between 0.70 and 0.82, write the candidate entity pairs to the cache to be reviewed and do not merge them for the time being;

[0059] After the candidate entity pair set is generated, the entity disambiguation module first deletes candidate entity pairs with completely different student IDs and valid ID fields. Then, it sorts the remaining candidate entity pairs according to their entity matching confidence from highest to lowest. When two candidate entity pairs... When identical and shared student records are used, the module prioritizes maintaining consistency in the identification document fields. Higher candidate entity pairs, The retention time overlap ratio remains the same. If higher-ranking candidate entity pairs are still tied, they are written to the manual review queue and the automatic merging of the student record is paused.

[0060] After entity disambiguation is completed, the data access module performs a second deduplication using unified student identifiers, unified course identifiers, event types, and event occurrence times. During the second deduplication, if two records come from different source systems, the module retains the primary record based on the source system's credibility from highest to lowest, and adds the other record as a supplementary source to the source merging field. If the source system credibility is the same, the module prioritizes the record based on field completeness. If the numbers are still the same, retain them from largest to smallest; if they are still the same, retain them from smallest to largest according to the collection batch number.

[0061] After entity disambiguation and the generation of unified student identifiers, unified course identifiers, and behavioral event identifiers, the risk factor generation module reads the risk factor generation rule table. The risk factor generation rule table includes rule identifier, applicable event type, target field name, comparison operator, threshold, statistical window, risk factor type, risk factor identifier generation rule, rule source, statistical triggering method, rule priority, rule priority coefficient, default value of priority coefficient, and rule version number. The rule source includes manual configuration by academic affairs staff, statistical analysis of historical labeled samples, and model output that has been manually verified. In this specific implementation, when the normalized value of the grade is less than 0.60 and the statistical window is for one semester, an academic decline risk factor is generated; when the number of absences reaches three within thirty days, an attendance abnormality risk factor is generated; when three consecutive assignments are not submitted, an assignment procrastination risk factor is generated; and when the consultation record summary matches the psychological support dictionary and passes manual review, a consultation concern risk factor is generated.

[0062] The risk factor generation module aggregates the student's behavioral event sequence by unified student ID and performs sliding window statistics in ascending order of event occurrence time. When a new behavioral event arrives, the module performs incremental recalculation of the affected statistical window corresponding to the unified student ID. During batch processing on the assessment day, it performs batch recalculation of all unconfirmed windows in the current semester. When the window statistical results first meet the risk factor generation rules, the trigger time is the occurrence time of the last behavioral event that meets the conditions. If the rules are generated by batch processing on the assessment day and there is no single end event, the trigger time is the current assessment time.

[0063] The risk factor generation module generates risk factor identifiers using unified student identifiers, behavioral event identifiers, risk factor types, trigger rule identifiers, and trigger times. It then writes the risk factor identifier, risk factor type, trigger rule identifier, trigger field name, trigger field value, comparison operator, threshold, statistics window, statistics trigger method, trigger time, and rule version number into a standard student behavior data table. When the same behavioral event simultaneously matches multiple risk factor generation rules, the module writes multiple risk factor identifiers according to rule priority from highest to lowest. If priorities are the same, they are written in lexicographical order. The rule priority coefficient is obtained by normalizing the rule priority and is clipped to between zero and one. If a rule does not have a configured rule priority coefficient, it is written with the default value of 1, ensuring that S2 can convert each risk factor identifier into an independent risk factor entity and that the static trigger scoring of S4 can reuse the same rule parameters.

[0064] After the risk factor identifier is written into the standard student performance data table, the intervention measure generation module reads the intervention measure generation rule table. The intervention measure generation rule table includes rule identifier, risk factor type, student grade, responsible role, resource type, intervention measure type, intervention measure identifier generation rule, expected completion time limit, rule source, resource status field, and rule version number. The rule source is manually configured by student management personnel and verified through historical intervention records. In this specific implementation, the risk factor of academic decline corresponds to the course teacher tutoring measures, the risk factor of attendance abnormality corresponds to the counselor contact measures, the risk factor of homework procrastination corresponds to the learning plan follow-up measures, and the risk factor of consultation and attention corresponds to the psychological counseling appointment measures.

[0065] Before the standardized data table is written, the data quality verifier checks the integrity of the primary key with unified student identifier, unified course identifier and behavior event identifier, and checks the integrity of the source link with source system identifier and collection batch number. When the primary key field is missing, the record is written to the data supplement queue and does not enter the S2 graph. When a non-primary key field is missing, the record carries a missing mark and enters the S2 graph and reduces the reliability of the corresponding relationship edge in S3.

[0066] The data access module generates a batch summary for each collection batch. The batch summary includes the number of source systems, the number of input records, the number of successful field mappings, the number of successful unit conversions, the number of entity disambiguation merges, the number of candidate entity pairs to be reviewed, and the number of output records. This batch summary is stored together with the standardized data table, so that subsequent steps can trace back to the source system that triggered the width, the missing fields, and the entity disambiguation results when the risk prediction range is found to be too wide.

[0067] After student management data undergoes field semantic mapping, unit normalization, pre-duplicate removal within the source system, entity disambiguation and unified identifier generation, secondary deduplication, risk factor generation, and intervention measure generation, standardized student learning data is written into a standardized data table. Each record in the standardized data table includes a unified student identifier, a unified course identifier, a behavioral event identifier, a risk factor identifier, a risk factor type, a trigger rule identifier, a trigger field value, a statistical trigger method, a trigger time, an intervention measure identifier, an intervention measure type, a responsible role, a source system identifier, an event occurrence time, a normalized field value, record-level field completeness, a missing marker, a missing start and end time, an entity matching confidence level, and a collection batch number. This standardized data table serves as the output object of this step and as the input for S2 to construct a heterogeneous time-series student learning knowledge graph. Simultaneously, the record-level field completeness, source system confidence level, and entity matching confidence level are passed to S3 to calculate the edge reliability of relation edges.

[0068] In this specific embodiment, S2 includes:

[0069] The graph construction module takes the standard chemical intelligence data table output by S1 as input and converts the unified student identifier, unified course identifier, behavioral event identifier, risk factor identifier, and intervention measure identifier into student entity, course entity, behavioral event entity, risk factor entity, and intervention measure entity respectively according to the entity type dictionary. Each entity node is written into the node table, which serves as the engineering data structure for this step. Its fields include node identifier, node type, source system identifier, timestamp, missing flag, normalized field value, field completeness, entity matching confidence, and node version number. The node identifier is obtained by hashing the node type and the corresponding business identifier.

[0070] For student entity nodes, the graph construction module writes a unified student identifier, class, year of enrollment, current enrollment status, and contact information hash value; for course entity nodes, it writes the course identifier, course semester, and course category; for behavioral event entity nodes, it writes the event type, event occurrence time, normalized grade value, attendance status code, assignment status code, or consultation record summary vector; for risk factor entity nodes, it writes the risk factor type and trigger source; for intervention measure entity nodes, it writes the intervention measure type, responsible role, and expected completion deadline; when the same node identifier appears repeatedly in the same version of the graph, the module retains the latest attribute according to the timestamp, while appending the old attribute to the node's historical attribute table;

[0071] The relationship edge generator generates relationship edges according to course selection, participation, occurrence, triggering, recommendation, same-class, same-course, and preceding-following event relationships. Course selection connects student entities and course entities; participation connects student entities and behavioral event entities; occurrence connects course entities and behavioral event entities; triggering connects behavioral event entities and risk factor entities, indicating that the behavioral event satisfies the risk factor generation rule; recommendation connects risk factor entities and intervention measure entities, indicating that the risk factor matches the intervention measure generation rule; same-class connects student entities within the same class; same-course connects student entities under the same course; and preceding-following event connects adjacent behavioral event entities of the same student. Event-type relationship edges are written with the event time, while persistent relationship edges are written with the effective time, expiration time, and semester interval. All relationship edges are written with the starting node identifier, ending node identifier, relationship type, source system identifier, field completeness, entity matching confidence, and edge version number. Recommendation relationships are classified as persistent relationship edges, with their effective time and expiration time taken from the effective start and end times of the intervention rule. The risk factor generation time is only written as a secondary attribute of the recommendation relationship in the recommendation generation time field.

[0072] The sequential event relationship is generated by the graph construction module by sorting the behavioral event entities corresponding to the same unified student identifier in ascending order of event occurrence time. If the time interval between two adjacent behavioral events exceeds 180 days as set in this specific implementation, the sequential event relationship is not generated and the breakpoint reason is recorded in the student time breakpoint table. If two event timestamps are the same, they are stably sorted according to the source system credibility, field completeness, and collection batch number to ensure that the same sequential event relationship is obtained when the same batch of data is repeatedly constructed. The entity type dictionary, relationship type dictionary, and 180-day breakpoint threshold are stored in the graph configuration table as the parameter source for this step.

[0073] Before writing the relationship edge table, the graph construction module performs edge validity checks on each relationship type. The validation rules are provided by the relationship type dictionary, which records the start node type, end node type, directional semantics, generation conditions, whether multiple edges are allowed, time attribute type, and default validity period. The starting point of a triggering relationship is a behavioral event entity, and the end point is a risk factor entity. The generation condition is that a corresponding risk factor identifier exists in the standard learning and evaluating data, and the directional semantics is that the event triggers the risk factor. The starting point of a recommendation relationship is a risk factor entity, and the end point is an intervention measure entity. The generation condition is that a corresponding intervention measure identifier exists in the standard learning and evaluating data, and the directional semantics is that the risk factor recommends intervention measures. Both types of relationships are allowed to exist simultaneously in the same student subgraph, but with different relationship types.

[0074] Persistent relationship edges include course selection relationships, same class relationships, same course relationships, and recommendation relationships. The effective time of course selection relationships, same class relationships, and same course relationships is the start time of the course semester or the effective time of the class, and the expiration time is the end time of the course semester or the end time of the class relationship. The effective time and expiration time of recommendation relationships are the start and end times of the intervention rule. The risk factor generation time is only used as the recommendation generation time field of the recommendation relationship and not as the decay window of the edge. When a persistent relationship lacks an effective time, the graph construction module uses the start time of the current semester as the effective time and writes a missing time mark. When an expiration time is missing, the end time of the current semester is used as the expiration time. Event-type relationship edges include participation relationships, occurrence relationships, triggering relationships, and before-and-after event relationships. The event time of event-type relationship edges is the event occurrence time of the behavioral event entity.

[0075] The graph construction module generates node incremental files and relation edge incremental files for each graph version. The incremental files record the node identifiers and edge identifiers added, updated and deleted in this batch. When the same entity node or relation edge is updated in consecutive batches, the module retains the old version attributes and writes the new version attributes into the current graph version, so that S3 can locate the set of affected relation edges when performing incremental edge reliability updates.

[0076] The graph construction module encapsulates the node table and relation edge table into a heterogeneous time-series learning knowledge graph. The graph includes a node index, a relation type index, a student time series index, and a meta-path query index. The node index uses the node identifier as the key, the relation type index uses the relation type and the starting node type as a composite key, the student time series index uses the unified student identifier and the event occurrence time as the sorting key, and the meta-path query index uses three types of meta-paths as query keys: student to course to behavior event, student to classmate to behavior event, and student to behavior event to risk factor. The meta-path index records the neighborhood size and number of triggering relationships of the risk factor corresponding to the third type of meta-path. When there are missing markers in the standard learning data, the corresponding nodes and edges are still written into the graph, but the missing markers are retained as node attributes and relation edge attributes for S3 to calculate edge reliability.

[0077] After construction, the graph construction module writes the heterogeneous time-series learning knowledge graph into the current graph version of the graph database, and writes the number of nodes, the number of relation edges, the number of each relation type, the number of missing nodes, and the graph construction batch number into the graph version table. The current graph version, node index, relation type index, student time series index, and meta-path query index are used as the output objects of this step and are passed to S3 to calculate the edge reliability and time decay coefficient, and at the same time as the graph structure input of the S4 time-aware heterogeneous graph transformer. The graph version table also stores the entity type dictionary version and the relation type dictionary version for subsequent steps to reproduce the same graph construction rule.

[0078] In this specific embodiment, S3 includes:

[0079] The edge attribute calculation module uses the heterogeneous time-series learning knowledge graph and its node table and relation edge table generated by S2 as input objects, and reads the field completeness, source system identifier, entity matching confidence, event occurrence time and missing window from each relation edge and its two end nodes. The source system confidence is provided by the source confidence table, which serves as the engineering data structure for this step. Each record in the source confidence table includes the source system identifier, system type, confidence value, number of calibration samples, calibration date and table version number. The confidence value is obtained by the ratio of the number of correct records in the historical manual verification records to the total number of verification records and is clipped to between zero and one.

[0080] Field completeness mapping is based on record-level field completeness output by S1. As the sole basic input, the field quality rule table is only used to determine the aggregate weights of the nodes at both ends of a relation edge and the records of the relation edge itself. These aggregate weights are dimensionless proportions and sum to one. The field quality rule table includes field names, required status, missing value penalty coefficients, outlier penalty coefficients, and rule version numbers. The edge attribute calculation module reads the records corresponding to the starting node of the relation edge. The record corresponding to the endpoint node and the source record of the relationship edge ,according to The completeness of the relation edge fields is obtained by summarizing. ,in , and The values ​​are given by the field quality rule table and the sum of the three values ​​is one.

[0081] The source credibility table is constructed using historical manually verified samples. The verification samples include the source system identifier, the number of sampled records, the number of correct records, the error type, and the verification time. The source credibility value is obtained by dividing the number of correct records by the number of sampled records. When the number of sampled records is less than the thirty records set in this specific implementation, the edge attribute calculation module marks the source system credibility as low sample credibility and uses the median credibility of the same type of source system as a temporary credibility. Once the number of samples reaches thirty, the source system's own credibility is used.

[0082] The edge attribute calculation module first maps field completeness, source system credibility, and entity matching confidence to a range of zero to one, and then calculates the basic edge reliability according to the weights determined by historical labeled samples. ,in For relation edges Basic edge reliability, For relation edges Corresponding field completeness For relation edges The credibility of the source system For relation edges Confidence of matching related entities , and Determined by a grid search on historical labeled samples, and the sum of the three is one; in this specific implementation method Take 0.35, Take 0.35, Setting it to 0.30 reduces the reliability of the basic edge due to missing fields, untrusted sources, and unstable entity matching.

[0083] For fields with missing start and end times, the edge attribute calculation module first determines the applicable window for missing correction based on the relationship type; participation relationships, occurrence relationships, triggering relationships, and preceding and following event relationships use an event coverage window, with the starting point of the event coverage window being the event time corresponding to the relationship edge, and the ending point being the event time of the next row of the same student or the current evaluation time; course selection relationships, same class relationships, and same course relationships use a persistent relationship window, which is determined by the effective time, expiration time, and semester interval; recommendation relationships use the effective interval of the intervention rule as the window;

[0084] The overlap duration between the missing window and the applicable window is calculated from the window intersection. The edge attribute calculation module follows... Calculate the missing overlap ratio, where For relation edges The proportion of missing overlap, The duration of overlap between the missing window and the applicable window. To determine the applicable window duration; when the persistent relationship window of a persistent relationship edge is complete and the missing field is not a required field for that relationship type, Set to zero; when the applicable window duration is zero, the module sets... Set one day as the baseline for zero-case prevention and put The cutting distance is between zero and one; the edge reliability is based on... Calculate the result and write it to the relation edge attribute.

[0085] Before writing the edge reliability and time decay coefficient into the relation edge attributes, the edge attribute calculation module performs an edge attribute range check; when Set to zero when less than zero, and when When the value is greater than one, it is set to one; when... Set to zero when less than zero, and when If the value is greater than one, it is set to one; the value after range verification is written to the relation edge attribute table, and the original calculated value is written to the edge attribute verification log. The edge attribute verification log includes the edge identifier, the original value, the pruning value, and the pruning reason.

[0086] The time decay coefficient is determined by the event time interval and the preset half-life. The edge attribute calculation module defines the event time interval as the interval between the current evaluation time and the timestamp of the event corresponding to the relation edge. For consecutive event relationships, the module simultaneously records the adjacent event interval between two adjacent behavioral events as an auxiliary field but does not replace the current evaluation time interval. Event-type relation edges are calculated according to... Calculate the time decay coefficient, where For relation edges Time decay coefficient, In days The half-life corresponds to the relation type; for persistent relation edges, when the current evaluation time falls between the effective time and the expiration time... Set to 1, and after the expiration time, perform exponential decay based on the semester distance; the half-life table is stored according to relational type. In this specific implementation, the relationship between grades and assignments is taken as 90 days, the relationship between attendance and abnormal behavior is taken as 45 days, the relationship between consultation records is taken as 120 days, the semester half-life of continuous relationships is taken as 1 semester, and the unmatched relationship type is taken as 60 days and written to the half-life default log.

[0087] The versioned graph copy is determined by the graph version number, update batch number, and rollback record table. After the incremental update task is completed, the edge attribute calculation module writes the number of updated objects in this batch, the change in average edge reliability, the change in average time decay coefficient, and the number of failed edges into the edge attribute update summary. When the number of failed edges is not zero, the module retains the old edge reliability and old time decay coefficient of the failed edges and writes the reason for failure into the update failure queue to prevent uncomputed relation edges from entering the S4 attention weight correction.

[0088] For new behavioral events of the same student arriving consecutively within a short period of time, the edge attribute calculation module processes incremental update tasks according to the event time from smallest to largest, and processes them according to the collection batch number from smallest to largest when the event time is the same; this order is written into the update task table as a parallel rule, so that when the same batch of supplementary data is repeatedly executed, the same versioned map copy and the same rollback record are obtained.

[0089] When a new behavioral event arrives or a supplementary field is confirmed, the edge attribute calculation module selects only the affected unified student identifier, affected course identifier, and relation edges within the time range of the affected behavioral event as the incremental update candidate set. It then rereads the source system credibility of the new event, the field completeness corresponding to the supplementary field, and the entity matching confidence. The relation edges in the candidate set are first sorted by the interval between the event time and the new behavioral event time from smallest to largest. If the time intervals are the same, they are sorted by the source system credibility from largest to smallest. If they are still the same, the order is determined by the lexicographical order of the edge identifier. The module then filters all affected relation edges after sorting as the update objects for this update. For each relation edge in the update objects, the module performs the following steps: , , and The calculation process regenerates the edge reliability and time decay coefficient, and writes the values ​​before the update into the rollback record table. The rollback record table fields include the graph version number, edge identifier, old edge reliability, old time decay coefficient, new edge reliability, new time decay coefficient, update reason, and update time.

[0090] The edge attribute calculation module will and Write the relation edge attribute table of the current graph version and form the attention weight correction term. ,in For relation edges The correction factor involved in attention scoring in S4 The value is between zero and one, and the larger the value, the more reliable the relationship edge is and the closer the time is to the current evaluation time; the relationship edge attribute table, rollback record table, and attention weight correction term are the output objects of this step, and are passed to S4 along with the heterogeneous time-series learning knowledge graph; when When the value is zero, the relation edge is still retained in the graph but does not contribute adjacency information in the S4 attention scoring; the relation edge attribute table also records the calculation batch number and parameter table version number, which are used to reproduce the reliability calculation process of the same edge; the table is indexed by edge identifier, so that S4 can directly read the corresponding correction factor by edge identifier.

[0091] In this specific embodiment, S4 includes:

[0092] The risk characterization module takes the heterogeneous time-series learning knowledge graph after writing the edge reliability and time decay coefficient in S3 as the input object, and inputs it into the time-aware heterogeneous graph transformer. The time-aware heterogeneous graph transformer includes a node attribute encoding layer, a meta-path neighborhood sampling layer, a relation edge correction attention layer, and a multi-task shared characterization layer. The model configuration table serves as the engineering data structure for this step and records the model version number, hidden vector dimension, meta-path set, training batch number, and input field list. In this specific implementation, the hidden vector dimension is 128.

[0093] The node attribute encoding layer encodes the node attributes of student entities, course entities, and behavioral event entities into node initial vectors and generates time interval vectors according to node type. The time interval vector for behavioral event entities is generated from the interval between the event occurrence time and the student's current assessment time; the time interval vector for student entities is generated from the interval between the enrollment time or the start time of the current semester and the assessment time; and the time interval vector for course entities is generated from the interval between the course start time or the start time of the course semester and the assessment time. For numerical attributes, the encoding layer uses the normalized field values ​​obtained from S1; for enumerated attributes, the encoding layer reads the enumerated encoding vector table; for consultation record summaries, the encoding layer uses a text encoder fixed by training samples to output a summary vector. The encoded node initial vector and the corresponding node type time interval vector are added together to form the node input vector. ,in For nodes The input vectors of the nodes are written into the transformer batch cache;

[0094] The meta-path neighborhood sampling layer generates a candidate set of adjacent nodes based on three types of meta-paths: student to course to behavioral event, student to classmate to behavioral event, and student to behavioral event to risk factor. For each student entity, candidate adjacent nodes are read from the meta-path query index within the layer, and are filtered according to meta-path type, event time, and relational edge reliability. First, they are sorted by event time from nearest to farthest. Then, if the event times are the same, they are sorted by edge reliability from largest to smallest. If they are still the same, the lexicographical order of the node identifiers determines the parallel order. In this specific implementation, the top 64 most recent behavioral event nodes and the top 32 most recent classmate related event nodes are retained, and the student to course to behavioral event... The path from the event to the risk factor metapath retains the risk factor nodes triggered by the 64 most recent behavioral events. For each behavioral event, a maximum of 8 risk factor nodes with the highest static trigger scores are retained. The static trigger scores are obtained by multiplying the reliability of the trigger relationship edge, the time decay coefficient, and the rule priority coefficient in the S1 risk factor generation rule table. All three are dimensionless quantities between zero and one. When the rule priority coefficient is missing, the sampling layer takes the default value of 1 for calculation. When the candidate set of adjacent nodes under a certain metapath is empty, the sampling layer writes an empty neighborhood marker and retains the student entity's own node as the self-loop adjacent node of the metapath, so that the subsequent attention layer can still output the student risk representation.

[0095] The metapath neighborhood sampling layer also reads the metapath threshold table, which records the metapath type, minimum edge reliability, maximum time interval, maximum number of adjacent nodes, and table version number. In this specific implementation, the minimum edge reliability of the metapath from student to course to behavioral event is 0.40, the minimum edge reliability of the metapath from student to classmate to behavioral event is 0.55, and the minimum edge reliability of the metapath from student to behavioral event to risk factor is 0.50. Candidate adjacent nodes with a reliability lower than the corresponding minimum edge reliability are filtered out and written to the neighborhood filtering log. The neighborhood filtering log also records whether the third type of metapath is enabled, the number of risk factor neighborhoods, static trigger score, and the number of trigger relationships that are filtered out.

[0096] When the number of candidate adjacent nodes in the metapath exceeds the maximum number of adjacent nodes, the sampling layer truncates the candidate set from largest to smallest according to the product of the edge reliability before correction and the time decay coefficient, and retains the node with the smallest lexicographical order of the node identifier among the parallel nodes at the truncation boundary; when the number of candidate adjacent nodes is lower than the minimum number of neighborhoods in the model configuration table, the sampling layer enables the self-loop completion rule and writes the completion status into the student risk representation table, so that S5 can identify the widening of the risk prediction interval caused by insufficient neighborhoods;

[0097] The relation edge correction attention layer generates query vectors, key vectors, and value vectors for the set of adjacent nodes, and scores the edge reliability by multiplying the time decay coefficient by the dot product of the query vector and key vector; for student nodes... and adjacent nodes In-layer calculation ,in For nodes Point to adjacent nodes Correction of attention score For student nodes The query vector, Adjacent nodes The key vector, To determine the reliability of the relation edge connecting two nodes, This is the time decay coefficient for the relation edge; subsequently, for the same metapath... Perform normalization to obtain the corrected attention weights. or When the value is reduced, the contribution of the adjacent nodes to the risk representation decreases.

[0098] After the attention weights are normalized, the attention layer of the relation edge will reset the weights that are less than the effective attention threshold to zero. The effective attention threshold is determined by the preset percentile of the explained path coverage on the validation set. In this specific embodiment, the effective attention threshold is 0.01. After the weights are reset to zero, they are normalized again and the value vector is aggregated. If all weights under the same metapath are lower than the effective attention threshold, the neighboring node with the largest weight is retained and a low attention branch mark is written.

[0099] During model inference, if the graph batch cache is missing node input vectors, the set of adjacent nodes is empty and self-loop completion fails, or the input field list in the model configuration table is inconsistent with the field set of the current graph version and cannot be completed, the risk characterization module marks the student as in an infeasible reasoning state. In the infeasible reasoning state, the student risk characterization for the current evaluation cycle is not output, and the unified student identifier, graph version number, evaluation time and infeasibility reason are written into the model inference anomaly table and manual review queue to avoid mixing the characterization of the previous evaluation cycle with the current graph version.

[0100] The multi-task shared representation layer obtains student risk representations based on the aggregated value vector of the modified attention weights. ,in For student nodes Student risk characterization under a single attentional head Adjacent nodes The value vector; the multi-head representations are concatenated and written into the student risk representation table. The fields of the student risk representation table include unified student identifier, graph version number, model version number, assessment time, academic related representations, abnormal behavior related representations, intervention urgency related representations, empty neighborhood markers, and feature update time.

[0101] The training process of the time-aware heterogeneous graph transformer uses student subgraphs labeled with historical academic risk, historical abnormal behavior risk, and historical intervention time limits as training samples, and divides the training and validation sets according to the chronological order of events; during training, relations are executed and occluded simultaneously, with the occlusion probability following the order of events. Confirmed, among which Let be the probability of edge occlusion. The reliability of a relation edge is determined by the reliability of the edge. Relation edges with lower reliability are more easily hidden, thereby reducing the model's dependence on low-reliability relations during the training phase.

[0102] After each training batch, the model training module calculates the validation set academic risk recall rate, abnormal behavior risk recall rate, intervention urgency ranking consistency rate, and explanatory path coverage rate. When any of these indicators falls below the acceptance threshold in the model configuration table, the best model version is not written for that training batch, and the training module continues to use the previous best model version as the inference version. When three consecutive batches are below the acceptance threshold, the training module freezes model parameter updates and outputs a training infeasibility status. Training resumes after the training samples or label quality is reviewed.

[0103] The candidate path explanation cache is generated synchronously by the risk characterization module during the inference phase. The cache records the unified student identifier, meta-path type, entity node sequence in the path, relation edge sequence, corrected attention weight, edge reliability, time decay coefficient, and model version number. When S5 requests a risk explanation path in the future, S5 directly reads the cache and filters it according to the risk component contribution value to ensure that the risk explanation path is consistent with the graph version and model version used when generating the student risk characterization in this step.

[0104] The training loss function is obtained by weighted summation of academic risk cross-entropy loss, abnormal behavior risk cross-entropy loss, intervention urgency ranking loss, and explanatory path coverage loss. The model training module calculates this. ,in For the total loss, , , and The losses are the four items mentioned above. to The loss weights are determined based on the inverse normalized results of the losses on the validation set. Tasks with larger validation set losses receive higher training weights. The training module updates the model parameters through backpropagation and writes the model version corresponding to the best validation batch into the model configuration table.

[0105] Before outputting student risk representations, the risk characterization module performs a characterization integrity check. The check fields include unified student identifier, graph version number, model version number, evaluation time, empty neighborhood marker, third-type meta-path enabled marker, risk factor neighborhood size, and completion status. When any primary key field is missing, the student record is written to the characterization anomaly queue and does not enter S5. When only the completion status exists or the risk factor neighborhood size is zero, the student record continues to enter S5 and carries the completion status and third-type meta-path marker for the risk prediction interval generation module to expand the scope of the review.

[0106] During inference, the risk representation module reads the best model version from the model configuration table and generates a student risk representation for each student entity in the current graph version. When the input graph version is inconsistent with the training graph field set recorded in the model configuration table, the module uses missing markers and empty neighborhood markers from the field semantic mapping table to fill in the gaps and writes the completion status into the student risk representation table. The student risk representation table serves as the output object of this step and as the input of the S5 multi-task risk assessment layer. The meta-paths, corrected attention weights, and adjacent node sequences required for risk interpretation are synchronously written into the interpretation candidate path cache.

[0107] In this specific embodiment, S5 includes:

[0108] The multi-task risk assessment layer uses the student risk representation table generated by S4 as input and outputs academic risk components, abnormal behavior risk components, and intervention urgency components through three linear mapping heads; for each task... Linear mapping head calculation ,in For the task The predicted risk component value, For the task linear mapping weights, For student risk characterization, For the task The bias term; the academic risk component and the abnormal behavior risk component, after normalization, take values ​​between zero and one. The larger the value, the higher the risk. The larger the value of the intervention urgency component, the more priority intervention is needed.

[0109] The three linear mapping heads share the student risk representation of S4 but use independent task weights. The task weights are obtained by minimizing the corresponding task loss from labeled training samples and are written into the risk assessment layer parameter table along with the model version number. When there are empty neighborhood labels, low attention branch labels, third-class meta-path not enabled labels, or risk factor neighborhood size of zero in the student risk representation table, the multi-task risk assessment layer writes an input quality label into the corresponding student record and passes the label to the calibration module to interpret the risk prediction interval width. Infeasible reasoning records in the model inference anomaly table are not input into the multi-task risk assessment layer but are entered into the manual review queue.

[0110] The calibration module reads historical calibration samples, which serve as the engineering data structure for this step and include student risk representations, corresponding ground truth labels, predicted risk components, event times, map version numbers, and manual review results. Academic risk tasks and abnormal behavior risk tasks use binary or multi-level risk labels, encoding the ground truth labels of the target risk categories as values ​​between zero and one, and the predicted risk components are the predicted probabilities of the target risk categories. Intervention urgency tasks use normalized rank or normalized time-limited regression labels. For each task, the module follows... Calculate the non-compliance score, where For the first Non-compliance score of each calibration sample Encode the actual label between zero and one. The corresponding risk component prediction value; the larger the non-compliance score, the further the prediction deviates from the true label. The non-compliance scores of different tasks are written into the corresponding calibration score table.

[0111] The risk prediction interval generation module reads quantile values ​​from the calibration score table according to a preset target coverage rate. In this specific embodiment, the target coverage rate is 0.90, and... Determine the quantile number, where For quantile numbers, To calibrate the sample size, The target coverage rate; the module is taken as the first in ascending order. The non-compliance score was used as the quantile. ,when When the number of calibration samples is greater than the number of non-compliance samples, the maximum non-compliance score is taken as... The target coverage, the number of calibration samples, and the quantile values ​​are written into the calibration version table. The target coverage, the lower limit of the number of calibration samples, and the quantile value selection rules serve as the source of parameters for this step.

[0112] For each risk component of the current student, the risk prediction interval generation module calculates the lower bound of the risk prediction interval. and the upper limit of the risk prediction range ,in This represents the lower bound of the risk prediction interval corresponding to the student and the risk component. This is the upper bound of the risk prediction range. This is the current predicted value of the risk component. The quantile value; if If less than zero, set to zero; if If the value is greater than 1, set it to 1; the interval width is determined by... Calculate and record the risk interval table; the risk interval table records the academic risk interval, abnormal behavior risk interval, and intervention urgency interval according to the risk component type;

[0113] The risk prediction interval table is established with a primary key based on a unified student identifier, risk component type, and assessment time. Table fields include the predicted risk component value, non-compliance quantile, lower bound of the risk prediction interval, upper bound of the risk prediction interval, interval width, calibration version number, input quality flag, component type used, and synthesis rule version number. When the number of calibration samples is less than one hundred as defined in this specific implementation, the module does not update the calibration version table but instead uses the previous calibration version. And write a flag indicating insufficient calibration samples in the risk interval table;

[0114] For the academic risk component, the abnormal behavior risk component, and the intervention urgency component, the risk prediction interval generation module maintains a calibration score table and a calibration version table respectively, so that the non-compliance scores of different tasks are not mixed. When the calibration version of a certain task becomes invalid, only the previous valid calibration version is used for that task, while the other tasks continue to use the current calibration version to generate risk prediction intervals.

[0115] The risk interpretation module extracts entities and relation sequences whose path contribution values ​​are greater than or equal to the path contribution threshold from the meta-paths that generate the corresponding risk components. The path contribution threshold is determined by the preset percentile of the path contribution values ​​in the calibration samples. For relation edges in the path, the module first calculates the task contribution scalar of the adjacent node value vector based on the task linear mapping head. ,in Adjacent nodes For the task The value vector contributes a scalar. For the task linear mapping weights, Adjacent nodes The value vector; then according to Calculate the contribution value of the relation edge, where For relation edges For the task Risk explanation contribution value, The corrected attention weights are generated for S4; the extracted entity and relation sequences are sorted by timestamp to generate risk interpretation paths that include data source, event time, relation type, edge reliability, time decay coefficient, and risk component contribution value;

[0116] The risk interpretation module uses a meta-path contribution backtracking algorithm to generate risk interpretation paths. This algorithm uses the S4 interpretation candidate path cache as the candidate path set and retrieves candidate paths using risk component type, unified student identifier, and model version number as search keys. The algorithm first calculates based on the attention heads of the relation edges. Then, for multiple attention heads within the same metapath The average value is taken as the contribution value of the relation edge under the meta-path, and the cumulative sum of the contribution values ​​of the relation edges in each candidate path is calculated; the candidate paths are sorted from largest to smallest according to the cumulative contribution value, and candidate paths with a cumulative contribution value greater than or equal to the path contribution threshold are selected as the candidate set of explanatory paths;

[0117] In the candidate set of explanatory paths, the risk explanation module prioritizes retaining paths that contain both behavioral event entities and risk factor entities. When the cumulative contribution values ​​of two candidate paths are the same, the path with the higher average edge reliability is prioritized. If they are still the same, the path with the event time closer to the current assessment time is retained. The filtered entity and relation sequences are converted into risk explanation path records. The fields of the risk explanation path records include path identifier, path contribution value, data source, event time, relation type, edge reliability, time decay coefficient, risk component type, and student identifier.

[0118] The path contribution threshold is determined by the preset percentile of the path contribution value in the calibration sample. The calibration module calculates the threshold according to the risk component type and writes it into the interpretation threshold table. When a student has no candidate path that reaches the path contribution threshold, the risk interpretation module outputs the candidate path with the highest contribution value and writes it into the low contribution interpretation mark to ensure that the intervention ranking result can still be associated with at least one traceable data source and event time link.

[0119] The intervention ranking module determines the first intervention threshold, the second review threshold, and the width threshold based on the lower bound, upper bound, and width of the risk prediction interval in the historical calibration samples, respectively, according to preset percentiles. In this specific implementation, the intervention urgency interval is the main component of the priority intervention queue, and the academic risk interval and abnormal behavior risk interval are auxiliary components. The synthesis rule table records the main component type, auxiliary component weight, threshold version number, and effective time. The module calculates the synthesis lower bound, where is the lower bound of the student's synthesis risk prediction interval, is the lower bound of the intervention urgency interval, is the lower bound of the academic risk interval, and is the lower bound of the abnormal behavior risk interval. When the lower bound is greater than or equal to the first intervention threshold, the student is added to the priority intervention queue. When the upper bound of the risk prediction interval of any risk component is greater than or equal to the second review threshold and the width of the interval of the risk component is greater than or equal to the width threshold, the student is added to the data entry or manual review queue, and the risk component type, missing field, source system, and entity disambiguation result that triggered the queue are recorded.

[0120] After the intervention ranking module generates a priority intervention queue, students in the queue are ranked according to the lower bound of the synthetic risk prediction interval. Sort from largest to smallest If the values ​​are the same, sort them from largest to smallest according to the lower bound of the intervention urgency component. If they are still the same, sort them from smallest to largest according to the largest interval width among the three components. The data entry or manual review queue is sorted from largest to smallest according to the interval width of the triggering component. If the interval widths are the same, sort them from largest to smallest according to the upper bound of the risk prediction interval of the triggering component. The sorting results are written to the queue number field.

[0121] When the same student meets the conditions of both the priority intervention queue and the data entry or manual review queue, the intervention sorting module retains the priority intervention queue identifier and writes a parallel review mark into the review queue identifier. After reading the parallel review mark, the counselor's workbench displays the relationship between the intervention task and the data entry task, so that students whose risk prediction interval lower bound reaches the first intervention threshold and whose interval width reaches the width threshold will not have their intervention delayed due to entering the review queue.

[0122] When the multi-task risk assessment layer outputs the intervention ranking results, it associates and stores the academic risk component, abnormal behavior risk component, intervention urgency component, lower bound of the risk prediction interval for each component, upper bound of the risk prediction interval for each component, interval width for each component, lower bound of the composite risk prediction interval, component type used, composite rule version number, priority intervention queue identifier, review queue identifier, and risk interpretation path with the unified student identifier. The intervention ranking result table and risk interpretation path are the output objects of this step and the inputs of the counselor workbench and data entry process. The risk interpretation path is also used to display the source system, event time, relationship type, and edge reliability that caused the corresponding risk component to increase.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0124] This invention standardizes multi-source student management data into learning entities and behavioral events, and further organizes them into a heterogeneous time-series learning knowledge graph. This enables data that was originally scattered in the systems of grades, attendance, homework, rewards and punishments, and consultation records to be associated and processed under a unified graph structure. The time-aware heterogeneous graph transformer aggregates the features between students, courses, and behavioral events according to the meta-path, and can form risk representations corresponding to business relationships for academic risks, abnormal behavior risks, and intervention urgency.

[0125] This invention further calculates edge reliability and time decay coefficient based on field completeness, source system credibility, entity matching confidence, event time interval, and missing window, and uses them for attention weight correction; at the same time, it generates risk prediction intervals based on calibration samples, and includes students with higher lower bounds of principal components or composite components in the priority intervention queue, and students with higher upper bounds but wider intervals in the data supplementation or manual review queue, so that the impact of missing, duplicate names, and time-series breakpoints on risk identification and ranking results can be identified, calibrated, and diverted.

Claims

1. A knowledge graph-based learning risk assessment method, characterized in that, include: S1. Obtain student management data from the grade system, attendance system, homework system, reward and punishment system, and consultation record system; perform field semantic mapping, unit normalization, and entity disambiguation on the student management data to generate standardized student behavior data. S2. Construct a heterogeneous temporal learning knowledge graph containing student entities, course entities, behavioral event entities, risk factor entities, and intervention measure entities based on standard learning data; S3. Calculate the edge reliability and time decay coefficient of the relation edge based on the field completeness, source system credibility, entity matching confidence, event time interval and missing window, and write it into the heterogeneous time series learning knowledge graph. S4. Input the heterogeneous time-series learning knowledge graph into the time-aware heterogeneous graph transformer, aggregate the associated features according to the meta-path, and use edge reliability and time decay coefficient to correct the attention weight to obtain the student risk representation. S5. Input the student risk representation into the multi-task risk assessment layer, output academic risk component, abnormal behavior risk component, intervention urgency component and risk interpretation path, and generate risk prediction interval based on the non-compliance score of the calibration sample, and generate intervention ranking results according to the risk prediction interval; wherein, the influence relationship in the heterogeneous time-series learning knowledge graph includes triggering relationship and recommendation relationship.

2. The knowledge graph-based learning risk assessment method according to claim 1, characterized in that, S1 includes: Configure a field semantic mapping table for each source system. The field semantic mapping table records the source field name, target field name, data type, unit of measurement, and value dictionary. For numerical fields, the conversion is performed according to the target unit of measurement; for enumerated fields, the value dictionary encoding is performed; and for missing fields, missing tags and missing start and end times are generated. Candidate entity pairs are generated based on student name, student ID, class, year of enrollment, and contact information. The entity matching confidence is calculated based on string similarity, document field consistency, and time overlap ratio. Candidate entity pairs with an entity matching confidence greater than or equal to a preset matching threshold are retained as entity disambiguation results.

3. The knowledge graph-based learning risk assessment method according to claim 1, characterized in that, S2 includes: The student identifier, course identifier, behavioral event identifier, risk factor identifier, and intervention measure identifier in the standardized learning information data are converted into corresponding entity nodes, and a source system identifier, timestamp, missing value flag, and normalized field value are written for each entity node. Relationship edges are generated according to course selection relationship, participation relationship, occurrence relationship, trigger relationship, recommendation relationship, same class relationship, same course relationship, and before-and-after event relationship. The influence relationship includes trigger relationship and recommendation relationship, and the before-and-after event relationship is determined by the timestamp sorting of the corresponding behavioral events of the same student. The heterogeneous time-series learning information knowledge graph is formed by the entity nodes and the relationship edges.

4. The knowledge graph-based learning risk assessment method according to claim 1, characterized in that, S3 includes: mapping field completeness, source system credibility, and entity matching confidence to the interval between 0 and 1, and summing them according to the weights determined by historical labeled samples to obtain the basic edge reliability; generating a missing correction amount based on the overlap duration between the applicable window and the missing window corresponding to the time attribute of the relation edge, wherein the applicable window is determined according to the relation type and includes event time, or effective time and expiration time, or rule effective interval, and using the missing correction amount to correct the basic edge reliability to obtain the edge reliability; calculating the time decay coefficient based on the event time interval and the preset half-life, which monotonically decreases as the event time interval increases; and using the edge reliability and the time decay coefficient as the relation edge attribute and attention weight correction terms.

5. The knowledge graph-based learning risk assessment method according to claim 1, characterized in that, S4 includes: encoding the node attributes of student entities, course entities, and behavioral event entities into node initial vectors, and encoding the interval between the event timestamp and the student's current evaluation time into a time interval vector; generating a set of adjacent nodes according to the meta-paths from student to course to behavioral event, student to classmate to behavioral event, and student to behavioral event to risk factor; generating a query vector, key vector, and value vector for the set of adjacent nodes; multiplying the edge reliability and the time decay coefficient into the dot product of the query vector and key vector, scoring, and normalizing to obtain a corrected attention weight; and aggregating the value vector based on the corrected attention weight to obtain the student risk representation.

6. The knowledge graph-based learning risk assessment method according to claim 1, characterized in that, S5 includes: a multi-task risk assessment layer linearly mapping the student risk representations to obtain academic risk components, abnormal behavior risk components, and intervention urgency components; calculating a non-compliance score for calibration samples, where the non-compliance score is the absolute difference between the true label of the calibration sample and the predicted value of the corresponding risk component; taking the quantile value of the non-compliance score after ascending sorting according to a preset target coverage rate, and using the predicted value of the corresponding risk component minus the quantile value as the lower bound of the risk prediction interval, and using the predicted value of the corresponding risk component plus the quantile value as the upper bound of the risk prediction interval; generating the intervention ranking result based on the lower bound of the risk prediction interval, the upper bound of the risk prediction interval, the interval width, and the lower bound of the synthesis obtained according to a preset synthesis rule.

7. The knowledge graph-based learning risk assessment method according to claim 5, characterized in that, The training process of the time-aware heterogeneous graph transformer includes: using student subgraphs labeled with historical academic risk, historical abnormal behavior risk, and historical intervention time limit as training samples, and dividing the training set and validation set according to time order; performing edge masking on relational edges in the training set, with the sampling probability of the edge masking determined inversely based on the edge reliability; constructing a loss function obtained by weighted summation of academic risk cross-entropy loss, abnormal behavior risk cross-entropy loss, intervention urgency ranking loss, and explanation path coverage loss; determining the weights in the loss function based on the inverse normalization results of each loss on the validation set, and updating the model parameters through backpropagation.

8. The knowledge graph-based learning risk assessment method according to claim 6, characterized in that, The generation of the intervention ranking results includes: determining a first intervention threshold, a second review threshold, and a width threshold based on the lower bound, upper bound, and width of the risk prediction interval in the historical calibration samples, as well as the lower bound of the synthesis obtained according to a preset synthesis rule, according to preset percentiles; the lower bound of the risk prediction interval is the lower bound of the risk prediction interval of the main component or the lower bound of the synthesis obtained according to the preset synthesis rule; when the lower bound of a student's risk prediction interval or the lower bound of the synthesis obtained according to the preset synthesis rule is greater than or equal to the first intervention threshold, the student is written into a priority intervention queue and arranged in descending order according to the lower bound of the risk prediction interval or the lower bound of the synthesis; when the upper bound of a student's risk prediction interval on any risk component is greater than or equal to the second review threshold and the width of the interval of the risk component is greater than or equal to the width threshold, the student is written into a data supplementation or manual review queue, and the risk component type, missing field, source system, and entity disambiguation result that triggered the queue are recorded.

9. The knowledge graph-based learning risk assessment method according to claim 4, characterized in that, The update of the edge reliability and the time decay coefficient includes: when a new behavior event arrives or a supplementary field is confirmed, recalculating the edge reliability of the affected relation edge based on the source system credibility of the new behavior event, the field completeness corresponding to the supplementary field, and the entity matching confidence; recalculating the time decay coefficient based on the time interval between the timestamp of the new behavior event and the timestamp of the previous behavior event of the same student; writing the recalculation result into the versioned graph copy, and retaining the edge reliability and time decay coefficient before the update as a rollback record.

10. The knowledge graph-based learning risk assessment method according to claim 8, characterized in that, The generation of the risk interpretation path includes: extracting entity and relation sequences with path contribution values ​​greater than or equal to a path contribution threshold from the meta-paths that generate the corresponding risk components, wherein the path contribution threshold is determined by a preset percentile of the path contribution values ​​in the calibration sample; sorting the extracted entity and relation sequences according to timestamps to generate a risk interpretation path that includes data source, event time, relation type, edge reliability, and risk component contribution value; and storing the risk interpretation path in association with the corresponding student identifier when outputting the intervention ranking result.

Images