Education management platform fusing multi-level data

Abstract

The invention discloses an education management platform fusing multi-level data, and particularly relates to the field of education informatization, and the platform comprises a terminal collection module which is used for collecting an evaluation record, an IEP record, a teaching record, a teaching and research record, a training record and a supervision record, and outputting a process record with an object identifier, an item identifier, a record content and a time identifier; and the edge decoupling module is used for receiving a process record at an edge computing node, constructing a process tensor according to an object identifier, an item identifier and a time identifier, performing tensor decomposition on the process tensor to obtain decomposition factors representing object continuation, item continuation and time sequence continuation respectively, and outputting a corresponding process factor sequence. According to the method, field change content is extracted from front and back process records, a process factor sequence is generated, an action chain is constructed, actual effect verification is carried out in combination with inheritance probability, deviation and a convergence relation, and the process record which truly causes subsequent changes and the form record which only completes trace leaving can be relatively distinguished.

Description

Technical Field

[0001] This invention relates to the field of educational informatization technology, and more specifically, to an educational management platform that integrates multi-level data. Background Technology

[0002] In the construction of the integrated education management platform, the existing processing methods mainly revolve around the standardized record-keeping and completion statistics of various links such as assessment, smart IEP, lesson preparation, teaching, teaching research, lesson observation and evaluation, training, supervision and data reporting. Usually, the operation results of teachers, students and administrators are recorded separately in each business module. Then, combined with the cloud aggregation and the local collection, preprocessing and uploading mechanism of edge computing nodes, the statistics, display and assessment of whether the assessment is completed, whether the lesson preparation is submitted, whether teaching research is participated in, whether training is completed, and whether supervision records are formed are carried out.

[0003] For example, in a scenario where a county-level resource center coordinates multiple inclusive schools to provide individualized education support for students with special needs, the platform needs to continuously receive assessment records, IEP adjustment records, teaching implementation records, lesson observation and evaluation notes, training completion records, and supervision feedback records generated by school-side edge computing devices or local business terminals without changing the existing operating habits of front-line teachers. At the same time, it needs to meet hard constraints such as asynchronous data uploads across institutions, inconsistent business rhythms, network condition fluctuations, and the need for the management side to quickly form decision-making basis on a periodic basis.

[0004] Under these conditions, while existing mainstream practices can create a seemingly complete process ledger, phenomena such as assessment results being submitted but not triggering subsequent IEP adjustments, teaching and research activities being completed but not resulting in changes to lesson preparation content, training certificates being issued but not showing any application traces in subsequent teaching, and supervisory opinions being recorded but not forming a rectification closure may occur. The reason for this is that most existing platforms record and statistically analyze the completion status of single links, lacking a unified correlation and judgment mechanism for the inheritance relationship, feedback relationship, and subsequent trajectory of cross-process results. Therefore, it is difficult to distinguish which records correspond to real educational improvements and which records only correspond to formal compliance.

[0005] The technical problem this application aims to solve is: how to, within an integrated education platform, without altering existing business operations, use multi-stage business data collection results supported by edge computing to uniformly correlate and determine whether the execution results of processes such as assessment, IEP, teaching, teaching research, training, and supervision have formed a subsequent substantial effect, thereby identifying the authenticity and validity of process records. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an education management platform that integrates multi-level data. By collecting and merging process records from assessment, IEP, teaching, teaching research, training, and supervision at the edge, decoupling factors, generating links, and verifying effectiveness, the platform identifies whether changes in the previous process are continuously inherited and implemented in the subsequent process, thereby solving the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an education management platform integrating multi-level data, comprising:

[0008] The terminal data acquisition module is used to collect assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records, and output process records with object identifiers, item identifiers, record content, and time identifiers;

[0009] The edge decoupling module is used to receive process records at the edge computing node, construct a process tensor according to object identifier, event identifier and time identifier, perform tensor decomposition on the process tensor to obtain decomposition factors that respectively represent object continuation, event continuation and time sequence continuation, and output the corresponding process factor sequence.

[0010] The link generation module is used to perform change point detection and graph matching on the preceding and following process factor sequences. It connects the changed item factors in the preceding process factor sequence with the item factors that are retained, expanded, or implemented in the following process factor sequence using the shortest path to generate the corresponding action chain.

[0011] The effectiveness verification module is used to independently decompose the execution conditions of the action chain, perform posterior inference and KL divergence calculation, obtain the inheritance probability and deviation of the previous process change on the subsequent process change, and mark the action chain with the inheritance probability higher than the deviation probability and the deviation continuously converging as the effective chain.

[0012] The results output module is used to determine whether the corresponding process record is an effective process based on the effective chain, and to determine whether the process record that has not formed an effective chain is a formal process, and output the corresponding process effectiveness results.

[0013] In a preferred embodiment, the execution of the terminal acquisition module includes:

[0014] The collected assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records are categorized into corresponding record sets according to object identifiers and item identifiers. When collecting the content of the current record, the previous record with the same object identifier and the same item identifier whose time identifier is immediately preceding it is read, and the same-track record pair consisting of the content of the current record and the content of the previous record is output.

[0015] Based on the current record content and the previous record content in the same track record pair, the field values ​​are compared item by item according to the same field name. The field names with different field values, the previous field value and the current field value are written into the process record, and the process record with the field change content is output.

[0016] Based on the process record with field changes, the record content, field changes, object identifier, item identifier, and time identifier are concatenated in a predetermined field order to generate the current record fingerprint. The current record fingerprint and the fingerprint of the previous record corresponding to the previous record content are written together into the process record, and the process record with the previous and next fingerprint relationships is output.

[0017] In a preferred embodiment, the execution of the edge decoupling module includes:

[0018] The received process records are sorted by object identifier, item identifier, and time identifier. Each process record is assigned an object sequence number, item sequence number, and time sequence number. The record content of the process record is converted into a fixed-length numerical vector and written to the corresponding three-dimensional coordinate position using the object sequence number, item sequence number, and time sequence number as three-dimensional coordinates to generate a process tensor.

[0019] Expand the process tensor into an object expansion matrix along the object index direction, an item expansion matrix along the item index direction, and a time expansion matrix along the time index direction. Calculate the object product matrix obtained by multiplying the object expansion matrix by its transpose, the item product matrix obtained by multiplying the item expansion matrix by its transpose, and the time product matrix obtained by multiplying the time expansion matrix by its transpose. Solve for the eigenvalues ​​and eigenvectors of the object product matrix, item product matrix, and time product matrix respectively. Then, write the first few eigenvectors in each product matrix, arranged in descending order of eigenvalues, as object factor matrices, item factor matrices, and time factor matrices.

[0020] In a preferred embodiment, the execution of the edge decoupling module further includes:

[0021] Based on the object number, item number, and time number corresponding to each process record, retrieve the object factor row vector corresponding to the object number from the object factor matrix, retrieve the item factor row vector corresponding to the item number from the item factor matrix, and retrieve the time factor row vector corresponding to the time number from the time factor matrix.

[0022] The object factor row vector, the event factor row vector, and the time factor row vector are concatenated in a fixed order to generate the process factor vector corresponding to the process record. Then, the process factor vectors of each process record under the same object identifier are arranged in ascending order by time identifier to output the process factor sequence.

[0023] In a preferred embodiment, the execution of the link generation module includes:

[0024] Arrange the preceding and following process factor sequences in ascending order according to the time identifier under the same object identifier. Calculate the difference vector between the process factor vector at each moment and the process factor vector at the previous moment. Sum the absolute values ​​of each component of the difference vector to obtain the variation value at the corresponding moment. Subtract the variation value of the previous moment from the variation value at the current moment. Take the moment when the variation value changes from decreasing to increasing as the change point position. Output the change point position and the item factor corresponding to the change point position.

[0025] Based on the item factors corresponding to each change point position in the previous process factor sequence and each item factor in the next process factor sequence, calculate the sum of the absolute values ​​of the factor component differences, the difference of the time identifier, and the difference of the sequence position for each pair. Then, add the sum of the absolute values ​​of the factor component differences, the difference of the time identifier, and the difference of the sequence position to obtain the matching cost value. Delete duplicate matches by retaining only the next item factor with the smallest matching cost value for each previous item factor and only the previous item factor with the smallest matching cost value for each next item factor. Output the one-to-one matching edge between the previous and next item factors.

[0026] In a preferred embodiment, the execution of the link generation module further includes:

[0027] Construct a directed link graph based on the one-to-one matching edges. Treat each preceding event factor and each following event factor as nodes, and each one-to-one matching edge as a directed edge. Add the matching cost of the directed edge to the difference between the time identifiers at both ends of the directed edge to obtain the edge weight. Starting from the preceding event factor with the smallest time identifier, accumulate the edge weights edge by edge in the direction of increasing time identifiers. Select the node sequence with the smallest accumulated edge weight as the candidate action chain and output the candidate action chain.

[0028] For two adjacent event factors in a candidate action chain, compare whether each component of the latter event factor is equal to the corresponding component of the former event factor, whether a new non-zero component is added in addition to the equal components, and whether the latter event factor reappears at at least one time position afterward. If the corresponding components are equal, it is recorded as "continue to retain". If a new non-zero component is added, it is recorded as "continue to expand". If it reappears afterward, it is recorded as "continue to implement". The candidate action chain containing "continue to retain", "continue to expand", or "continue to implement" is output as the action chain.

[0029] In a preferred embodiment, the execution of the practical verification module includes:

[0030] The action chain is divided into multiple segments according to time identifiers. The number of common components, the number of unique components in the next process change, and the number of unique components in the previous process change are counted in each segment. The inheritance probability of the segment is obtained by dividing the number of common components by the sum of the number of common components, the number of unique components in the next process change, and the number of unique components in the previous process change. The deviation of the segment is obtained by dividing the sum of the number of unique components in the next process change and the number of unique components in the previous process change by the sum of the number of common components, the number of unique components in the next process change, and the number of unique components in the previous process change.

[0031] The inheritance probability and deviation are recursively calculated according to the time sequence of each chain segment. The inheritance probability of the first chain segment is the inheritance probability of that chain segment divided by the sum of the inheritance probability and the deviation of that chain segment. The deviation of the first chain segment is the deviation of that chain segment divided by the sum of the inheritance probability and the deviation of that chain segment. The inheritance probability of the next chain segment is the product of the inheritance probability of that chain segment and the inheritance probability of the previous chain segment divided by the sum of the product of the product and the deviation of that chain segment and the deviation of the previous chain segment. The deviation of the next chain segment is calculated in reverse order.

[0032] In a preferred embodiment, the execution of the practical verification module further includes:

[0033] The proportion of each component of the subsequent process change in each chain segment is taken as the actual distribution. The KL divergence value between the inheritance distribution composed of inheritance probability and the actual distribution and the deviation distribution composed of deviation amount are calculated respectively. When the KL divergence value of the previous chain segment is greater than the KL divergence value of the next chain segment, the previous chain segment and the next chain segment are merged and recalculated until the KL divergence value of each chain segment decreases or becomes equal in time order, which is used as the deviation amount to continuously converge.

[0034] The action chain whose inheritance probability of the last chain segment is greater than the deviation and whose deviation continues to converge is marked as the effective chain.

[0035] In a preferred embodiment, the execution of the result output module includes:

[0036] The starting process record, ending process record, and intermediate process record of the effective chain are grouped into the same process group according to the object identifier and the item identifier. The process group is marked as an effective process, and the process record that is not included in any effective chain is marked as a formal process.

[0037] Based on the effective process and the formal process, count the number of effective processes and the number of formal processes under the same object identifier, calculate the ratio of the number of effective processes to the sum of the number of effective processes and the number of formal processes, generate the process effectiveness results for the corresponding object identifier and item identifier, and output them.

[0038] The technical effects and advantages of this invention are as follows:

[0039] 1. By extracting the changes in fields from the records before and after the process, generating a process factor sequence and constructing an action chain, and then combining inheritance probability, deviation amount and convergence relationship for effective verification, it is possible to relatively distinguish between process records that actually cause subsequent changes and records that only complete the traceability.

[0040] 2. By performing tensor decoupling, factor extraction, and link generation on multi-stage process records at edge computing nodes, assessment, IEP, teaching, teaching research, training, and supervision records can be incorporated into a unified comparison caliber, thereby relatively improving the distortion of correlation judgment caused by data fragmentation across modules;

[0041] 3. By writing the effective chain back to the effective process and statistically analyzing the effective percentage corresponding to the object identifier and the item identifier, the management side can obtain results output oriented towards the actual effect of the process, thereby relatively improving the pertinence of periodic supervision, assessment and decision-making. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Refer to the instruction manual appendix Figure 1 The present invention provides an education management platform that integrates multi-level data, comprising:

[0045] The terminal data acquisition module is used to collect assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records, and output process records with object identifiers, item identifiers, record content, and time identifiers;

[0046] In this embodiment, the terminal acquisition module is used to organize consecutive business records of the same object under the same event into directly comparable continuous records before the process records enter the edge computing node. It further extracts the changes in fields and the fingerprint relationships between the preceding and following records, thus providing a unified and traceable data foundation for subsequent process tensor construction, process factor extraction, and action chain verification. The record types processed by the terminal acquisition module include assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records. Each type of record carries an object identifier, event identifier, record content, time identifier, and record writing sequence number when entering the terminal acquisition module. The object identifier is used to identify students, teachers, etc. For specific business entities within a teacher or organization, the event identifier is used to identify a single, continuously traceable event under the same business entity. The time identifier uniformly adopts the write time of the record to the platform database. The record write sequence number is used to determine the order of events when the time identifiers are the same. The record content is organized using a unified field slot table. The unified field slot table is pre-configured according to the record type, including the field name order, field type, and null placeholder code. Numeric fields directly store the numerical value, enumerated fields store the corresponding coded value, and text fields are converted into a fixed-length term counting sequence according to the fixed term table pre-written in the field slot table to ensure that different record types can be read in a fixed order when entering subsequent processing. This implementation process includes the following steps:

[0047] First, the collected assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records are categorized into corresponding record sets according to object identifiers and item identifiers. This is used to establish a continuous record trajectory for the same item identifier under the same object identifier. The terminal acquisition module receives the record type, object identifier, item identifier, record content, time identifier, and record writing sequence number submitted by the business terminal. It then sorts the records first by object identifier in ascending order, then by item identifier in ascending order, then by time identifier in ascending order, and finally by record writing sequence number in ascending order when time identifiers are the same. After sorting, records with the same object identifier and the same item identifier are grouped together. Records are written to the same record set. When writing the current record content, the content of the previous record immediately preceding the current record in the sorting result is read from the record set. If the current record content is the first record under the object identifier and the item identifier, an empty record of the same type as the current record is generated as the previous record content. Each field slot in the empty record is written with the corresponding null value placeholder code in the field slot table. At the same time, the record entry number corresponding to the previous record content is written to zero. After processing, a pair of records on the same track consisting of the current record content and the previous record content is output and written to the terminal acquisition buffer for direct reading in the next processing step.

[0048] Secondly, based on the current and previous record content in the same track record pair, field values ​​are compared item by item according to the same field name. This is used to compress business changes between records into field change content that can be directly written into the process record. The terminal acquisition module reads the field slot table corresponding to the current record type and extracts the field values ​​with the same name from the current and previous record content in the order of the field names determined in the field slot table. For numeric fields, the numerical values ​​are directly compared for equality. For enumerated fields, the corresponding encoded values ​​are compared for equality. For text fields, the text is first split into a fixed-length term count sequence according to the fixed term table pre-written in the field slot table, and then the term count values ​​are compared for equality. If the field value of a certain field in the current record content is different from the field value in the previous record content, the module will take appropriate action. If the current field value, the field name, the previous field value, and the current field value are written to the field change content in a fixed order; if the current field value is not null and the previous field value is a null placeholder, the field change content indicates addition; if the current field value is a null placeholder and the previous field value is not null, the field change content indicates deletion; if both the current and previous field values ​​are not null but have different values, the field change content indicates replacement; multiple field change contents are written sequentially to the change area of ​​the process record after being arranged in ascending order of field name; if all field values ​​of the current record content are the same as those of the previous record content, a null change marker is written to the change area of ​​the process record; after processing, the process record with the field change content is output and written to the local cache for the current record fingerprint generation process to read;

[0049] Finally, based on the process record with field changes, the record content, field change content, object identifier, event identifier, and time identifier are concatenated end-to-end according to a predetermined field order to generate the current record fingerprint. This fingerprint is used to provide a stable identifier for subsequent record consistency verification and traceability of previous and subsequent records on the edge computing node side. The terminal acquisition module first reads the predetermined field order table, which is written into the configuration area by the platform during deployment. The field order is fixed as object identifier, event identifier, time identifier, values ​​of each field in the record content, and triplet values ​​of each field change content. The terminal acquisition module extracts the values ​​of each field sequentially according to this field order and... The first and last characters are concatenated to form the original fingerprint string. Then, each character in the original fingerprint string is converted into its corresponding character encoding value. The character encoding values ​​are accumulated sequentially according to the order of character appearance, and the remainder is taken with respect to a fixed modulus of the system to obtain the fingerprint of the current record. Then, the fingerprint of the previous record corresponding to the content of the previous record is read. If the content of the previous record is empty, the fingerprint of the previous record is written as a fixed empty fingerprint value. The terminal acquisition module writes the fingerprint of the current record and the fingerprint of the previous record together into the fingerprint area of ​​the process record to form a process record with the relationship between the previous and next fingerprints. The process record is then sent to the receiving queue of the edge computing node for the edge decoupling module to read directly.

[0050] Through the above processing, the terminal acquisition module completes the same-track aggregation, field change extraction, and front-to-back fingerprint binding at the record generation end. This allows the subsequent edge decoupling module to directly use the unified process records to perform process tensor construction and factor decomposition without having to backtrack to the original business forms. This reduces the repetitive data processing by edge computing nodes and ensures that the value acquisition and writing methods of object identifiers, item identifiers, time identifiers, field change content, and front-to-back fingerprint relationships remain consistent throughout the entire process. In practical applications: when the same student submits two assessment records under the same enrollment adaptability assessment item, the terminal acquisition module first groups the two assessment records into the same record set, and then... When writing the second assessment record, the first assessment record is read as the previous record content. Then, the field values ​​corresponding to the self-care ability, motor ability, and perception ability fields in the two assessment records are compared item by item according to the field slot table. The changed field name, previous field value, and current field value are written into the field change content. Then, the object identifier, event identifier, time identifier, the record content of the second assessment record, and the field change content are concatenated in the predetermined field order to generate the current record fingerprint. The previous record fingerprint corresponding to the first assessment record is also written into the process record. Finally, a process record with the previous and next fingerprint relationship is formed and sent to the edge computing node for subsequent process tensor writing and process factor extraction.

[0051] The edge decoupling module is used to receive process records at the edge computing node, construct a process tensor according to object identifier, event identifier and time identifier, perform tensor decomposition on the process tensor to obtain decomposition factors that respectively represent object continuation, event continuation and time sequence continuation, and output the corresponding process factor sequence.

[0052] In this embodiment, the edge decoupling module is used to uniformly sort, encode, and decompose the process records output by the terminal acquisition module, transforming the process records, originally existing in the form of business forms, into a sequence of process factors that can be directly compared by the subsequent link generation module. Its processing principle is to first establish a stable three-dimensional index relationship using object identifiers, item identifiers, and time identifiers; then map the record content and field changes in the process records into fixed-length numeric vectors and write them into the process tensor; subsequently, expand the process tensor along the three dimensions and solve for the eigenvalues ​​and eigenvectors of the corresponding product matrices to obtain the object factor matrix, item factor matrix, and... The time factor matrix is ​​then used to extract corresponding factors for each process record based on the object number, item number, and time number, and these factors are concatenated into a process factor vector. This vector is then organized into a process factor sequence based on the object identifier and time identifier. To ensure a closed calculation chain, in this implementation, the object identifier uniformly corresponds to a single business entity, the item identifier uniformly corresponds to a single, continuously traceable item under that business entity, and the time identifier uniformly uses the write time of the process record to the platform database. When a process record involves multiple business entities, the terminal acquisition module has already split it into multiple process records, each corresponding to a single object identifier. This implementation process includes the following steps:

[0053] First, the received process records are sorted by object identifier, item identifier, and time identifier and written into the process tensor. The purpose of this is to establish stable three-dimensional coordinates required for subsequent decomposition calculations. After receiving the process records with object identifier, item identifier, record content, field change content, and time identifier, the edge decoupling module sorts them first by object identifier in ascending order, then by item identifier in ascending order, then by time identifier in ascending order, and in the case of the same time identifier, sorts them by the record writing sequence number in ascending order.

[0054] Subsequently, based on the sorting results, consecutive object serial numbers are assigned to each different object identifier, consecutive event serial numbers are assigned to each different event identifier, and consecutive time serial numbers are assigned to each different time identifier. After the serial number assignment is completed, the edge decoupling module reads the field slot table and vector mapping table of the corresponding record type, and converts the record content and field change content in the process record into a fixed-length numeric vector. Numeric fields are directly written with numerical values, enumeration fields are written with pre-numbered values, and text fields are converted into a fixed-length term counting vector according to the fixed term table pre-written in the field slot table. Each change item in the field change content is assigned according to the field name serial number, the previous field value, and the current field value. The data is written to the change vector area. Missing fields are uniformly written with zero values, and placeholder fields are uniformly written with the number corresponding to the null placeholder code. Then, using the object number, item number, and time number as three-dimensional coordinates, the fixed-length digital vector is written to the corresponding three-dimensional coordinate position. If there are multiple process records for the same three-dimensional coordinate, the sequence number is written according to the record, the fixed-length digital vector corresponding to the last process record is retained, and the remaining records are written to the supplementary table for traceability. After processing, a process tensor is generated, and the mapping tables of object identifier and object number, item identifier and item number, and time identifier and time number are written to the local cache area of ​​the edge computing node for subsequent expansion calculation and reading.

[0055] Secondly, the process tensor is expanded along the object number direction, the item number direction, and the time number direction, respectively, and three factor matrices are obtained. The purpose is to decompose the three-dimensional relationship in the process tensor into a computable representation of object continuation, item continuation, and time sequence continuation. After reading the process tensor, the edge decoupling module expands the process tensor into an object expansion matrix along the object number direction. Specifically, the order of item number and time number is fixed, and the fixed-length numerical vectors of each three-dimensional coordinate position under the same object number are concatenated end to end to form a row in the object expansion matrix. When expanding into an item expansion matrix along the item number direction, the object number and time are fixed. The order of the serial numbers is determined by concatenating the fixed-length numerical vectors of each three-dimensional coordinate position under the same item serial number to form a row in the item expansion matrix. When expanding into a time expansion matrix along the time serial number direction, the order of the object serial number and the item serial number is fixed, and the fixed-length numerical vectors of each three-dimensional coordinate position under the same time serial number are concatenated to form a row in the time expansion matrix. After expansion, the product of the object expansion matrix and its transpose matrix is ​​calculated to obtain the object product matrix, the product of the item expansion matrix and its transpose matrix is ​​calculated to obtain the item product matrix, and the product of the time expansion matrix and its transpose matrix is ​​calculated to obtain the time product matrix.

[0056] Subsequently, eigenvalue decomposition was performed on the object product matrix, event product matrix, and time product matrix respectively to obtain their respective eigenvalue sets and corresponding eigenvector sets, which were then sorted in descending order of eigenvalues.

[0057] In addition, to eliminate the problem of unclear first few feature vectors mentioned above, this embodiment reads the factor dimension values ​​written into the configuration table during system deployment. The factor dimension values ​​are positive integers and remain unchanged within one platform operation cycle. The edge decoupling module sequentially takes the first few factor dimension value feature vectors from the feature vector set after sorting each product matrix, and writes them into object factor matrix, event factor matrix and time factor matrix by column. If the number of feature vectors corresponding to a certain product matrix is ​​less than the factor dimension value, the missing column number is supplemented with zero vectors. After processing, the object factor matrix, event factor matrix and time factor matrix are written into the factor cache area of ​​the edge computing node for reading in the process factor vector generation process.

[0058] Finally, based on the object number, item number, and time number corresponding to each process record, corresponding factors are extracted from the three factor matrices to generate a process factor sequence. The purpose is to uniformly map the same process record to a process factor vector that can be directly processed by the subsequent link generation module. After reading the object factor matrix, item factor matrix, time factor matrix, and the aforementioned mapping table, the edge decoupling module extracts the corresponding row from the object factor matrix according to its object number as the object factor row vector, the corresponding row from the item factor matrix according to its item number as the item factor row vector, and the corresponding row from the time factor matrix according to its time number as the time factor row vector. Then, the object factor row vector, item factor row vector, and time factor row vector are concatenated in a fixed order, with the fixed order being object factor row vector first, item factor row vector in the middle, and time factor row vector last, to form the process factor vector corresponding to the process record. If a corresponding row of a process record is missing in the object factor matrix, item factor matrix, or time factor matrix, a zero vector of the same length is used to replace the missing row vector, and a missing mark is written in the anomaly marking area.

[0059] After all process factor vectors are generated, the edge decoupling module uses the object identifier as the merging key to sort all process factor vectors under the same object identifier in ascending order by time identifier. In the case of the same time identifier, they are sorted in ascending order by the record writing sequence number to form a process factor sequence corresponding to the object identifier. The process factor sequence, together with the correspondence between the object identifier, the item identifier, and the time identifier, is written into the reading area of ​​the link generation module for direct use by subsequent change point detection, graph matching, and shortest path connection.

[0060] Through the above processing, the edge decoupling module transforms the process records output by the terminal acquisition module into a process factor sequence with fixed dimensions, fixed order and stable index relationship. This ensures that the subsequent link generation module no longer relies on the original form content when performing change point detection and factor matching. Instead, it directly performs continuous calculations based on the decoupled object factors, event factors and time factors, reducing the field misalignment and dimension inconsistency problems caused by direct comparison of cross-type records.

[0061] In practical applications: When the same student generates assessment records, IEP records, and teaching records consecutively under the same event identifier, the edge decoupling module first sorts the corresponding process records by object identifier, event identifier, and time identifier, and assigns object sequence number, event sequence number, and time sequence number. Then, it maps the assessment result field, IEP adjustment field, teaching execution field, and corresponding field change content in each process record into a fixed-length numeric vector of the same length and writes it into the process tensor. Subsequently, it expands the process tensor along the object sequence number direction, event sequence number direction, and time sequence number direction respectively to obtain the object factor matrix, event factor matrix, and time factor matrix. Then, it extracts the object factor row vector, event factor row vector, and time factor row vector according to the object sequence number, event sequence number, and time sequence number corresponding to each process record of the student, and concatenates them to form the process factor vector. Finally, it outputs the process factor sequence corresponding to the student in ascending order of time identifier, so that the link generation module can determine whether the changes in the previous process are to be retained, expanded, or implemented in the next process.

[0062] The link generation module is used to perform change point detection and graph matching on the preceding and following process factor sequences. It connects the changed item factors in the preceding process factor sequence with the item factors that are retained, expanded, or implemented in the following process factor sequence using the shortest path to generate the corresponding action chain.

[0063] In this embodiment, the link generation module is used to identify the location where a process change actually occurs from a process factor sequence arranged by time identifier under the same object identifier, and to establish a verifiable continuous connection between the preceding and following processes. This allows the subsequent effectiveness verification module to determine whether a change in the preceding process has entered into a change in the following process based on a clear chain of action. Its processing principle is to first identify the location where abrupt changes occur in the event factors using the variation of the process factor vector in the time direction, and then construct a matching cost value based on the component differences, time differences, and sequence position differences between the event factors, thereby establishing a one-to-one matching edge between the preceding and following process factor sequences. Subsequently, a directed link graph is constructed using the one-to-one matching edges, and... Under the constraint of monotonically increasing time identifiers, the node sequence with the smallest cumulative edge weight is selected as the candidate action chain. Finally, the continuation method of adjacent factors in the candidate action chain is verified item by item, and the candidate action chains that meet the requirements of continuing to retain, continue to expand, or continue to implement are output as action chains. To ensure consistency in the chain generation process, in this embodiment, the preceding process factor sequence and the following process factor sequence are both taken from two adjacent process factor sequences arranged in ascending order of time identifier under the same object identifier. The preceding process factor sequence corresponds to the set of continuous process factor vectors that are earlier in time, and the following process factor sequence corresponds to the set of continuous process factor vectors that are immediately following it in time. This implementation process includes the following steps:

[0064] First, the previous and next process factor sequences are arranged in ascending order according to the time identifier under the same object identifier, and the change point is identified. The purpose is to extract the moment when the change of the event factor is most significant from the continuous process factor vector, which is used as the starting object for subsequent matching. After receiving the process factor sequence output by the edge decoupling module, the link generation module first arranges the process factor vectors in ascending order according to the time identifier. If the time identifiers are the same, they are arranged in ascending order according to the record writing sequence number. Then, the difference vector between the process factor vector at each moment and the process factor vector at the previous moment is calculated in turn. The absolute values ​​of each component of the difference vector are then added to obtain the variation value at that moment. For the first moment, since there is no process factor vector at the previous moment, the variation value at the first moment is written as zero.

[0065] After calculating the variance values ​​for all time points, the variance value at the current time point is subtracted from the variance value at the previous time point to obtain the variance increment. If the variance increment at a certain time point is less than zero and the variance increment at the current time point is greater than zero, then the current time point is determined as the change point location, and the item factor part in the process factor vector corresponding to the change point location is extracted as the change point item factor. Here, the truncation range of the item factor part directly adopts the fixed splicing order and fixed factor dimension value when the edge decoupling module generates the process factor vector, that is, the continuous factor component interval after the object factor row vector and before the time factor row vector. If there is no time point in the previous process factor sequence where the variance increment changes from decreasing to increasing, then the time point with the largest variance value is taken as the change point location. If there are multiple times where the variance values ​​are all at the maximum value, then the time point with the earliest time identifier is taken as the change point location. After processing, the change point location and the item factor corresponding to the change point location are output and written to the matching calculation area for the next processing step to read.

[0066] Secondly, a one-to-one matching edge is generated based on the item factors corresponding to the positions of each change point in the previous process factor sequence and the item factors in each item factor sequence of the next process. The purpose is to screen out the item connection relationship with the lowest cost and no repetition between the two processes. The link generation module reads all change point item factors in the previous process factor sequence and all item factors in the next process factor sequence, and calculates three types of differences for each pair. The first type is the sum of the absolute values ​​of the factor component differences, that is, subtracting the factor components at corresponding positions in the two item factors one by one and taking the absolute value, and then summing all the absolute values ​​to obtain the factor component difference. The second type is the difference of time identifiers, that is, the time identifier of the next item factor minus the time identifier of the previous item factor. If the difference is less than zero, the item factor pair will not participate in the subsequent matching. The third type is the difference of sequence position, that is, the position number of the next item factor in the next process factor sequence minus the position number of the previous item factor in the previous process factor sequence. If the difference is less than zero, the calculation result is still retained, but it participates in the accumulation of the cost in the subsequent matching according to its absolute value.

[0067] After completing the above calculations, the sum of the absolute values ​​of the factor component differences, the difference of time identifiers, and the difference of sequence positions are directly added together to obtain the matching cost of the corresponding event factor pair. Then, taking the previous event factor as the benchmark, only the matching relationship with the smallest matching cost is retained among all its candidate subsequent event factors. If there are multiple subsequent event factors with the same matching cost, the one with the smallest time identifier difference is retained. If the time identifier differences are still the same, the one with the smallest absolute value of sequence position difference is retained. Then, taking the subsequent event factor as the benchmark, the retained results are reverse-filtered, and only the matching relationship with the smallest matching cost is retained among all its candidate previous event factors. If the matching cost is the same, the unique match is determined in the order of smallest time identifier difference, smallest absolute value of sequence position difference, and earliest time identifier of the previous event factor. After bidirectional filtering, the one-to-one matching edge between the previous and subsequent event factors is output, and the matching cost, the time identifier of the previous event factor, and the time identifier of the subsequent event factor are written into the link graph construction area for the next processing step to read.

[0068] Next, a directed link graph is constructed based on the one-to-one matching edges, and candidate action chains are generated. The purpose is to extract the continuous connection path with the minimum global cumulative cost from local matching relationships. After reading the one-to-one matching edges, the link generation module writes each preceding and following event factor as a node in the directed link graph, and writes each one-to-one matching edge as a directed edge connecting the corresponding preceding and following event factor nodes. Then, the matching cost corresponding to the directed edge is added to the difference between the time identifiers at both ends of the directed edge to obtain the edge weight. After the graph is constructed, the node with the smallest time identifier among all preceding event factor nodes is selected as the starting node, and... Initialize the cumulative edge weight of the starting node to zero. Then, traverse all directed edges in ascending order of time identifier. For each directed edge, add the cumulative edge weight of its starting node to the weight of the directed edge to obtain the candidate cumulative edge weight of the ending node. If the ending node has not yet been written with a cumulative edge weight, write the candidate cumulative edge weight directly and record the starting node of the directed edge as the predecessor node. If the ending node has already been written with a cumulative edge weight and the candidate cumulative edge weight is less than the already written cumulative edge weight, replace the original cumulative edge weight with the candidate cumulative edge weight and simultaneously replace the predecessor node record. If the candidate cumulative edge weight is equal to the already written cumulative edge weight, retain the path with the earlier time identifier of the predecessor node.

[0069] After traversing all directed edges, select the node with the smallest cumulative edge weight among all subsequent event factor nodes as the endpoint node, and backtrack from the endpoint node to the starting node according to the predecessor node record to obtain a node sequence with monotonically increasing time identifiers. This node sequence is determined as a candidate action chain. If there is no reachable endpoint node, write the candidate action chain corresponding to the current object identifier as an empty chain and write the empty chain mark into the subsequent verification area. After processing, output the candidate action chain for the next processing step to read.

[0070] Finally, the continuation method is verified for two adjacent event factors in the candidate action chain. The purpose is to filter out the event succession relationship that truly reflects the transition from the previous process change to the next process change from the candidate path with the lowest cumulative cost. The link generation module reads two adjacent event factors one by one according to the node order in the candidate action chain and compares the factor components at the corresponding positions item by item. If the corresponding component of the subsequent event factor is equal to the corresponding component of the previous event factor, and this corresponding component is not zero, then this corresponding component is recorded as to be retained. If the subsequent event factor has a new non-zero component in addition to the corresponding component equal to the previous event factor, then the newly added non-zero component is recorded as to be expanded. If the subsequent event factor reappears in the process factor vector corresponding to at least one subsequent time position under the object identifier, and the event identifier at the time of reappearance belongs to teaching records, execution... If either the line record or the feedback record is selected, the subsequent event factor is marked as to be continued. The judgment criterion here is that in the process factor sequence following the subsequent event factor, there exists at least one process factor vector whose event factor portion is entirely equal to the corresponding component of the subsequent event factor. If there is no option to continue retaining, expanding, or implementing between two adjacent event factors, the connection between the adjacent nodes is deleted. If at least one continuous connection path remains after deletion of the candidate action chain, the remaining continuous node sequence is output as the action chain. If no continuous connection path remains after deletion of the candidate action chain, the candidate action chain is written as an empty chain. After outputting the action chain, the object identifier, event identifier, time identifier, and node order within the chain are written into the effective verification module's reading area for direct use in subsequent inheritance probability and deviation calculations.

[0071] Through the above processing, the link generation module first accurately identifies the change points in the process factor sequence, then connects the preceding and following factors into a continuous path with the minimum cumulative cost by matching edges one by one and a directed link graph. Finally, it filters out matching relationships that are only similar in form but do not form a real subsequent effect by combining three interface paths: continue to retain, continue to expand, and continue to implement. This provides the effectiveness verification module with a clear function chain with a clear structure, clear sequence, and clear connection method.

[0072] In practical applications: When the same student successively forms an assessment process factor sequence, an IEP process factor sequence, and a teaching process factor sequence under the same item identifier, the link generation module first calculates the variation value of the process factor vector at adjacent times in the assessment process factor sequence, and determines the time when the variation value changes from decreasing to increasing as the change point position, and then extracts the item factor corresponding to this position as the change point item factor; then, it calculates the sum of the absolute values ​​of the factor component differences, the difference of time identifiers, and the difference of sequence positions for each item factor in the IEP process factor sequence, and retains the one-to-one matching edge with the smallest matching cost; then, it constructs a directed link graph based on the one-to-one matching edge, and searches for the node sequence with the smallest cumulative edge weight starting from the item factor with the smallest time identifier as a candidate action chain; finally, it verifies whether the IEP item factor retains the corresponding component of the assessment item factor, whether a new non-zero component has been added, and whether the IEP item factor reappears in the subsequent teaching process factor sequence and enters the teaching record, so as to output the candidate action chain that meets the above conditions as the action chain, for the effectiveness verification module to continue to judge whether the assessment change has truly entered the subsequent IEP adjustment and teaching execution.

[0073] The effectiveness verification module is used to independently decompose the execution conditions of the action chain, perform posterior inference and KL divergence calculation, obtain the inheritance probability and deviation of the previous process change on the subsequent process change, and mark the action chain with the inheritance probability higher than the deviation probability and the deviation continuously converging as the effective chain.

[0074] In this embodiment, the effectiveness verification module is used to verify the action chain output by the link generation module segment by segment to determine whether the change in the previous process forms a true inheritance in the change in the next process, rather than just forming a superficially similar continuous record. Its basic principle is to first split the action chain into segments connected end to end according to time markers, then calculate the inheritance probability and deviation based on the component correspondence between the change in the previous process and the change in the next process in each segment. Subsequently, the inheritance probability and deviation are recursively extrapolated along the time direction, and the KL divergence values ​​of the inheritance distribution and the deviation distribution are calculated with reference to the actual distribution formed by the change in the next process. Then, the KL divergence values ​​corresponding to the deviation distribution are reduced or equalized segment by segment by segment through segment merging and recalculation. Finally, the action chain with an inheritance probability greater than the deviation and a continuously converging deviation is marked as an effective chain. To ensure consistent calculation methods, in this embodiment, both the change in the previous process and the change in the next process are taken from two adjacent event factors in the action chain. The components refer to the factor components of these two event factors at the same position. A component value of zero is considered as no effective change at that position. This implementation process includes the following steps:

[0075] First, the action chain is divided into multiple segments based on time identifiers, decomposing the entire action chain into verifiable sequential units. The effectiveness verification module reads the node sequence, corresponding time identifiers, and event factors in the action chain, arranges all nodes in ascending order of time identifiers, and defines the connection between two adjacent nodes as a segment. The event factor corresponding to the preceding node is considered the change in the preceding process, and the event factor corresponding to the following node is considered the change in the following process. Then, the changes in the preceding and following processes are compared item by item according to the component positions of the event factors. If both components at a corresponding position are non-zero and have equal values, that position is recorded as a common component. If only the component of the following process change is present at the corresponding position... If the value is non-zero, the position is recorded as the next unique component. If only the component that changed in the previous process is non-zero at the corresponding position, the position is recorded as the previous unique component. After completing the statistics of all component positions, the number of common components, the number of next unique components, and the number of previous unique components are obtained. The inheritance probability of the chain segment is obtained by dividing the number of common components by the sum of the three. The deviation of the chain segment is obtained by dividing the sum of the number of next unique components and the number of previous unique components by the sum of the three. If the sum of the three is zero, the inheritance probability of the chain segment is written as zero and the deviation is written as one. After processing, the number of common components, the number of next unique components, the number of previous unique components, the inheritance probability, and the deviation of each chain segment are written into the chain segment calculation table for subsequent recursive calculation.

[0076] Secondly, the inheritance probability and deviation are recursively calculated according to the time sequence of each chain segment, which is used to continuously accumulate the results of local chain segments into the bearing strength of the entire action chain. After reading the chain segment calculation table, the effectiveness verification module processes the chain segments in ascending order according to the time identifiers corresponding to the chain segments. For the first chain segment, the inheritance probability of the chain segment is divided by the sum of the inheritance probability and deviation of the chain segment to obtain the recursive inheritance probability of the first chain segment, and the deviation of the chain segment is divided by the sum of the inheritance probability and deviation of the chain segment to obtain the recursive deviation of the first chain segment. For the next chain segment, the inheritance probability of the chain segment is multiplied by the recursive inheritance probability of the previous chain segment to obtain the inheritance product value, and then the product value is calculated. The deviation of a chain segment is multiplied by the recursive deviation of the previous chain segment to obtain the deviation product value. The recursive inheritance probability of the next chain segment is obtained by dividing the inheritance product value by the sum of the inheritance product value and the deviation product value. The recursive deviation of the next chain segment is obtained by dividing the deviation product value by the sum of the inheritance product value and the deviation product value. If the sum of the inheritance product value and the deviation product value is zero, the recursive inheritance probability of the next chain segment is written as zero and the recursive deviation of the next chain segment is written as one. After processing, the recursive inheritance probability and recursive deviation of each chain segment are output in chronological order and written into the distributed calculation area for the next processing step to read.

[0077] Next, the proportion of the component change in the subsequent process of each chain segment is taken as the actual distribution, and the KL divergence value is calculated to determine whether the recursive deviation continues to decrease along the direction of the action chain. After the effectiveness verification module reads the subsequent process change, recursive inheritance probability, and recursive deviation of each chain segment, it first divides the number of common components by the sum of the number of common components, the number of the next unique component, and the number of the previous unique component to obtain the actual inheritance proportion. Then, it divides the sum of the number of the next unique component and the number of the previous unique component by this sum to obtain the actual deviation proportion. The actual inheritance proportion and the actual deviation proportion form the actual distribution of the chain segment. Then, the recursive inheritance probability and a subtractive recursive inheritance probability form the inheritance distribution, and the subtractive recursive deviation and the recursive deviation form the deviation distribution. Subsequently, the two distribution values ​​of the actual distribution are multiplied by the two distribution values ​​and divided by the natural logarithm of the corresponding distribution value of the inheritance distribution, and the two terms are added together to obtain the KL divergence value between the inheritance distribution and the actual distribution. The KL divergence value between the deviation distribution and the actual distribution is calculated in the same way. The KL divergence value is used to determine the convergence of deviations. To ensure that the judgment object for continuous convergence of deviations remains unique, this embodiment uses only the KL divergence value between the deviation distribution and the actual distribution corresponding to the preceding and following chain segments as the convergence judgment metric. After calculating the KL divergence value of all chain segments, the KL divergence values ​​corresponding to the deviation distributions of adjacent chain segments are compared in chronological order. When the KL divergence value corresponding to the deviation distribution of the preceding chain segment is greater than that corresponding to the deviation distribution of the following chain segment, the preceding chain segment and the following chain segment are merged into a new chain segment. The common component count, the number of the following unique component, and the number of the following unique component are recalculated for the merged new chain segment. The inheritance probability, deviation, recursive inheritance probability, recursive deviation, and the KL divergence value corresponding to the deviation distribution are recalculated. The above merging and recalculation process is repeated until the KL divergence values ​​corresponding to the deviation distributions of each chain segment decrease or become equal in chronological order. This result is recorded as continuous convergence of deviations. After processing, the recalculated chain segment result is written into the effective judgment area.

[0078] Finally, the inheritance probability of the final chain segment is compared with the deviation amount, and the effective chain is determined by combining the continuous convergence result of the deviation amount. This provides the final conclusion as to whether the entire action chain is a true effective chain. After reading the recalculated recursive inheritance probability of the final chain segment, the recursive deviation amount of the final chain segment, and the continuous convergence result of the deviation amount, the effective verification module first compares the recursive inheritance probability of the final chain segment with the recursive deviation amount of the final chain segment. If the recursive inheritance probability of the final chain segment is greater than the recursive deviation amount of the final chain segment, and the KL divergence value corresponding to the deviation distribution of each chain segment has decreased or equalized in time sequence, then the action chain is marked as an effective chain. If the recursive inheritance probability of the final chain segment is less than or equal to the recursive deviation amount of the final chain segment, or the KL divergence value corresponding to the deviation distribution has not decreased or equalized in time sequence, then the action chain is marked as an ineffective chain. After marking, the action chain identifier, object identifier, event identifier, recursive inheritance probability of the final chain segment, recursive deviation amount of the final chain segment, and effective chain mark are written together into the reading area of ​​the result output module for direct use in the generation of effective results in subsequent processes.

[0079] Through the above processing, the effectiveness verification module further compresses the action chain output by the link generation module into an effectiveness judgment result that can be statistically analyzed segment by segment, recursively calculated, distributedly compared, and convergently verified. This makes it so that whether the change in the previous process truly enters the change in the next process no longer depends on manual interpretation, but is jointly determined by the number of common components, the number of unique components in the next process, the number of unique components in the previous process, the recursive inheritance probability, the recursive deviation, and the KL divergence value corresponding to the deviation distribution.

[0080] In practical applications: When the assessment adjustment of the same student is further incorporated into teaching execution through IEP adjustment, the effectiveness verification module first splits the connection relationship between assessment item factors and IEP item factors, and between IEP item factors and teaching item factors, into two chains. Then, it counts the number of common components, the number of subsequent unique components, and the number of preceding unique components in each chain, calculates the inheritance probability and deviation, and recursively obtains the recursive inheritance probability and recursive deviation of the final chain segment along the time direction. Subsequently, using the actual distribution formed by the subsequent process change of each chain segment as a reference, it calculates the KL divergence value corresponding to the deviation distribution and performs a merge recalculation on adjacent chains that do not satisfy the step-by-step decrease relationship. When the KL divergence value corresponding to the deviation distribution of each chain segment decreases or equals in time order after recalculation, and the recursive inheritance probability of the final chain segment is greater than the recursive deviation of the final chain segment, the action chain is marked as an effective chain, thus indicating that the previous process change has formed a real inheritance in the subsequent process change.

[0081] The result output module is used to determine whether the corresponding process record is an effective process based on the effective chain, and to determine whether the process record that has not formed an effective chain is a formal process, and output the corresponding process effectiveness result.

[0082] In this embodiment, the result output module is used to push the effectiveness chain output by the effectiveness verification module back to the process record layer and process group layer, forming a process effectiveness result that can be directly read by the management center. This allows the platform to further transform the judgment of a single link into process quality results at the object identifier and item identifier dimensions. The processing principle is as follows: First, based on the starting process record, ending process record, and intermediate process record contained in the effectiveness chain, the process is merged according to the object identifier and item identifier to obtain the corresponding effective process. Then, the remaining process records that do not enter any effectiveness chain but belong to the same object identifier and the same item identifier are merged into formal processes. Finally, the number of effective processes and formal processes are counted and the effectiveness ratio is calculated to form the process effectiveness result corresponding to the object identifier and item identifier. To ensure that the hierarchical relationship between process records, process groups, effective processes, and formal processes remains consistent, in this embodiment, the process record is the smallest business record unit, the process group is the set of records merged under the same object identifier and the same item identifier, the effective process is a process group that contains at least one effective chain, and the formal process is a process group that does not contain any effective chain. This implementation process includes the following steps:

[0083] First, the starting process records, ending process records, and intermediate process records corresponding to the effective chain are grouped into the same process group according to object identifier and item identifier. This is used to convert the link-level judgment result into the process-level judgment result. The result output module reads the effective chain identifier, object identifier, item identifier, and all process record identifiers within the chain written by the effective verification module. Then, it groups the starting process records, ending process records, and intermediate process records corresponding to the effective chain according to object identifier and item identifier. Process records with the same object identifier and the same item identifier are written into the same process group, and this process group is marked as an effective process. Subsequently, all process records under the current object identifier and item identifier are read and compared with those already grouped into effective processes. The process records are compared one by one, and process records not belonging to any effective chain are assigned to the corresponding process group. Process groups without any effective chains are marked as formal processes. If there are no effective chains under the same object identifier and the same item identifier, all process records under that object identifier and item identifier are directly merged into formal processes. If the same process record appears in multiple effective chains, it is merged according to the effective chain with the earliest time identifier. The remaining effective chains only retain the link statistics relationship and are not written into the process group repeatedly. After processing, the effective processes and formal processes are output, and the process group identifier, object identifier, item identifier, number of process records and process type are written into the result statistics area for the next processing step to read.

[0084] Secondly, based on the effective processes and formal processes, the number of effective processes and formal processes under the same object identifier is counted separately to generate process effectiveness results that can be directly output. After reading the effective processes and formal processes in the result statistics area, the result output module merges and counts the number of effective processes and formal processes under the object identifier, and then divides the number of effective processes by the sum of the number of effective processes and formal processes to obtain the effectiveness ratio of the corresponding item identifier under the object identifier. If the sum of the number of effective processes and the number of formal processes is zero, the effectiveness ratio is written as zero. Subsequently, the object identifier, item identifier, number of effective processes, number of formal processes, effectiveness ratio, and process type tag are written together into the process effectiveness result. The process type tag is written as the effectiveness priority tag when there are effective processes, and as the formal tag when only formal processes exist. After processing, the process effectiveness results of the corresponding object identifier and item identifier are output, and the process effectiveness results are sent to the data display area and statistical reading area of ​​the management center for direct use in subsequent overview display, assessment statistics, and supervision analysis.

[0085] Through the above processing, the result output module stably writes back the effectiveness chain-level judgment result to the process group, object identifier, and item identifier dimensions. This allows the platform to clearly distinguish which process records have formed real follow-up effects and which process records are still in the formal completion state. A unified output result is generated in terms of the number of effective processes, the number of formal processes, and the percentage of effective processes, facilitating direct reading and aggregation by the subsequent management center. In practical application: when the same teacher generates assessment records, IEP records, and teaching records successively under the same item identifier, and some of these process records have been marked as members of the effectiveness chain by the effectiveness verification module, the result output module first... The starting process record, ending process record, and intermediate process record corresponding to the effective chain are grouped into the same process group according to the teacher's object identifier and the matter identifier, and marked as effective processes. Then, the other process records of the teacher under the same matter identifier that have not entered any effective chain are grouped into formal processes. Subsequently, the number of effective processes and formal processes of the teacher under the matter identifier are counted, the effective ratio is calculated, and a process effectiveness result containing object identifier, matter identifier, number of effective processes, number of formal processes, effective ratio, and process type label is generated, so that the management center can determine whether the teacher's process execution on the matter has truly formed a subsequent effect.

[0086] Working Principle: This solution first collects assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records in a unified manner. At the collection end, it extracts the field changes and fingerprint relationships between records under the same object identifier and the same item identifier, forming a continuously traceable process record. Then, the edge computing node converts the process record into a process tensor, decomposes it into object factors, item factors, and time factors, and forms a process factor sequence in chronological order. Subsequently, it identifies change points in the process factor sequences, calculates the matching relationship between item factors, constructs candidate action chains, and filters out action chains with relationships to continue retention, further development, or further implementation. Finally, it calculates the inheritance probability and deviation of each action chain segment by segment, determines the effective chain based on the convergence of the deviation, and classifies the corresponding process into effective processes and formal processes accordingly, outputting the process effectiveness results corresponding to the object identifier and item identifier. The key point of the entire processing chain is that it does not use the completion of a single link as the criterion for judgment, but rather whether the change in the previous process enters the change in the next process and forms a continuous effect.

[0087] In the context of inclusive education management, if the same student first generates an assessment record, and subsequently generates an IEP adjustment record and a teaching execution record, this solution will first extract the changes in the assessment record relative to the previous record, and then determine whether the changes are carried over in the IEP record and whether they are further implemented in the teaching record. If a continuous chain of action can be formed between the assessment change, IEP adjustment, and teaching execution, and the inheritance relationship is established after effectiveness verification, the corresponding process is marked as an effective process. If a certain assessment, teaching research, or training generates a platform record, but does not subsequently lead to IEP adjustment, teaching change, or implementation, an effective chain cannot be formed, and the corresponding process is marked as a formal process. Thus, the result output by the platform is no longer just the completion status of the process, but the result of whether the process has truly produced a subsequent effect.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An education management platform integrating multi-level data, characterized in that, include: The terminal data acquisition module is used to collect assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records, and output process records with object identifiers, item identifiers, record content, and time identifiers; The edge decoupling module is used to receive process records at the edge computing node, construct a process tensor according to object identifier, event identifier and time identifier, perform tensor decomposition on the process tensor to obtain decomposition factors that respectively represent object continuation, event continuation and time sequence continuation, and output the corresponding process factor sequence. The link generation module is used to perform change point detection and graph matching on the preceding and following process factor sequences. It connects the changed item factors in the preceding process factor sequence with the item factors that are retained, expanded, or implemented in the following process factor sequence using the shortest path to generate the corresponding action chain. The effectiveness verification module is used to independently decompose the execution conditions of the action chain, perform posterior inference and KL divergence calculation, obtain the inheritance probability and deviation of the previous process change on the subsequent process change, and mark the action chain with the inheritance probability higher than the deviation probability and the deviation continuously converging as the effective chain. The results output module is used to determine whether the corresponding process record is an effective process based on the effective chain, and to determine whether the process record that has not formed an effective chain is a formal process, and output the corresponding process effectiveness results.

2. The education management platform integrating multi-level data according to claim 1, characterized in that: The execution of the terminal acquisition module includes: The collected assessment records, IEP records, teaching records, teaching and research records, training records, and supervision records are categorized into corresponding record sets according to object identifiers and item identifiers. When collecting the content of the current record, the previous record with the same object identifier and the same item identifier whose time identifier is immediately preceding it is read, and the same-track record pair consisting of the content of the current record and the content of the previous record is output. Based on the current record content and the previous record content in the same track record pair, the field values ​​are compared item by item according to the same field name. The field names with different field values, the previous field value and the current field value are written into the process record, and the process record with the field change content is output. Based on the process record with field changes, the record content, field changes, object identifier, item identifier, and time identifier are concatenated in a predetermined field order to generate the current record fingerprint. The current record fingerprint and the fingerprint of the previous record corresponding to the previous record content are written together into the process record, and the process record with the previous and next fingerprint relationships is output.

3. The education management platform integrating multi-level data according to claim 2, characterized in that: The execution of the edge decoupling module includes: The received process records are sorted by object identifier, item identifier, and time identifier. Each process record is assigned an object sequence number, item sequence number, and time sequence number. The record content of the process record is converted into a fixed-length numerical vector and written to the corresponding three-dimensional coordinate position using the object sequence number, item sequence number, and time sequence number as three-dimensional coordinates to generate a process tensor. Expand the process tensor into an object expansion matrix along the object index direction, an item expansion matrix along the item index direction, and a time expansion matrix along the time index direction. Calculate the object product matrix obtained by multiplying the object expansion matrix by its transpose, the item product matrix obtained by multiplying the item expansion matrix by its transpose, and the time product matrix obtained by multiplying the time expansion matrix by its transpose. Solve for the eigenvalues ​​and eigenvectors of the object product matrix, item product matrix, and time product matrix respectively. Then, write the first few eigenvectors in each product matrix, arranged in descending order of eigenvalues, as object factor matrices, item factor matrices, and time factor matrices.

4. The education management platform integrating multi-level data according to claim 3, characterized in that: The execution of the edge decoupling module also includes: Based on the object number, item number, and time number corresponding to each process record, retrieve the object factor row vector corresponding to the object number from the object factor matrix, retrieve the item factor row vector corresponding to the item number from the item factor matrix, and retrieve the time factor row vector corresponding to the time number from the time factor matrix. The object factor row vector, the event factor row vector, and the time factor row vector are concatenated in a fixed order to generate the process factor vector corresponding to the process record. Then, the process factor vectors of each process record under the same object identifier are arranged in ascending order by time identifier to output the process factor sequence.

5. The education management platform integrating multi-level data according to claim 4, characterized in that: The execution of the link generation module includes: Arrange the preceding and following process factor sequences in ascending order according to the time identifier under the same object identifier. Calculate the difference vector between the process factor vector at each moment and the process factor vector at the previous moment. Sum the absolute values ​​of each component of the difference vector to obtain the variation value at the corresponding moment. Subtract the variation value of the previous moment from the variation value at the current moment. Take the moment when the variation value changes from decreasing to increasing as the change point position. Output the change point position and the item factor corresponding to the change point position. Based on the item factors corresponding to each change point position in the previous process factor sequence and each item factor in the next process factor sequence, calculate the sum of the absolute values ​​of the factor component differences, the difference of the time identifier, and the difference of the sequence position for each pair. Then, add the sum of the absolute values ​​of the factor component differences, the difference of the time identifier, and the difference of the sequence position to obtain the matching cost value. Delete duplicate matches by retaining only the next item factor with the smallest matching cost value for each previous item factor and only the previous item factor with the smallest matching cost value for each next item factor. Output the one-to-one matching edge between the previous and next item factors.

6. The education management platform integrating multi-level data according to claim 5, characterized in that: The execution of the link generation module also includes: Construct a directed link graph based on the one-to-one matching edges. Treat each preceding event factor and each following event factor as nodes, and each one-to-one matching edge as a directed edge. Add the matching cost of the directed edge to the difference between the time identifiers at both ends of the directed edge to obtain the edge weight. Starting from the preceding event factor with the smallest time identifier, accumulate the edge weights edge by edge in the direction of increasing time identifiers. Select the node sequence with the smallest accumulated edge weight as the candidate action chain and output the candidate action chain. For two adjacent event factors in a candidate action chain, compare whether each component of the latter event factor is equal to the corresponding component of the former event factor, whether a new non-zero component is added in addition to the equal components, and whether the latter event factor reappears at at least one time position afterward. If the corresponding components are equal, it is recorded as "continue to retain". If a new non-zero component is added, it is recorded as "continue to expand". If it reappears afterward, it is recorded as "continue to implement". The candidate action chain containing "continue to retain", "continue to expand", or "continue to implement" is output as the action chain.

7. An education management platform integrating multi-level data according to claim 6, characterized in that: The execution of the practical verification module includes: The action chain is divided into multiple segments according to time identifiers. The number of common components, the number of unique components in the next process change, and the number of unique components in the previous process change are counted in each segment. The inheritance probability of the segment is obtained by dividing the number of common components by the sum of the number of common components, the number of unique components in the next process change, and the number of unique components in the previous process change. The deviation of the segment is obtained by dividing the sum of the number of unique components in the next process change and the number of unique components in the previous process change by the sum of the number of common components, the number of unique components in the next process change, and the number of unique components in the previous process change. The inheritance probability and deviation are recursively calculated according to the time sequence of each chain segment. The inheritance probability of the first chain segment is the inheritance probability of that chain segment divided by the sum of the inheritance probability and the deviation of that chain segment. The deviation of the first chain segment is the deviation of that chain segment divided by the sum of the inheritance probability and the deviation of that chain segment. The inheritance probability of the next chain segment is the product of the inheritance probability of that chain segment and the inheritance probability of the previous chain segment divided by the sum of the product of the product and the deviation of that chain segment and the deviation of the previous chain segment. The deviation of the next chain segment is calculated in reverse order.

8. An education management platform integrating multi-level data according to claim 7, characterized in that: The execution of the practical verification module also includes: The proportion of each component of the subsequent process change in each chain segment is taken as the actual distribution. The KL divergence value between the inheritance distribution composed of inheritance probability and the actual distribution and the deviation distribution composed of deviation amount are calculated respectively. When the KL divergence value of the previous chain segment is greater than the KL divergence value of the next chain segment, the previous chain segment and the next chain segment are merged and recalculated until the KL divergence value of each chain segment decreases or becomes equal in time order, which is used as the deviation amount to continuously converge. The action chain whose inheritance probability of the last chain segment is greater than the deviation and whose deviation continues to converge is marked as the effective chain.

9. An education management platform integrating multi-level data according to claim 8, characterized in that: The execution of the result output module includes: The starting process record, ending process record, and intermediate process record of the effective chain are grouped into the same process group according to the object identifier and the item identifier. The process group is marked as an effective process, and the process record that is not included in any effective chain is marked as a formal process. Based on the effective process and the formal process, count the number of effective processes and the number of formal processes under the same object identifier, calculate the ratio of the number of effective processes to the sum of the number of effective processes and the number of formal processes, generate the process effectiveness results for the corresponding object identifier and item identifier, and output them.

Images