A Method and System for Constructing Multidimensional Profiles of College Students Based on Dynamic and Static Knowledge Graphs and Large Models

Abstract

This invention belongs to the fields of artificial intelligence and educational big data mining and analysis. Specifically, it is a method and system for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models. The method includes steps such as task initiation and academic path retrieval plan generation, dynamic and static knowledge graph fusion retrieval and cross-subgraph semantic aggregation, data decoupling and cognitive diagnosis, and feature fusion and profile generation. This enables cognitive diagnosis and multi-dimensional profile construction for college students. This invention addresses the problems in existing technologies, such as insufficient integration of dynamic and static data, lack of analysis of college students' learning processes, and lack of objective evaluation indicators.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and educational big data mining and analysis, and in particular relates to a method and system for constructing multi-dimensional profiles of college students based on dynamic and static knowledge graphs and large models. Background Technology

[0002] In the field of educational big data mining and analysis, building accurate student profiles helps universities conduct personalized education for students and gain a comprehensive understanding of their actual academic development. Traditional profile building techniques typically treat student data as a flat record, which has the following drawbacks:

[0003] (1) Lack of effective correlation between static attributes and dynamic behavior data: Existing methods fail to effectively distinguish between "static attributes" (such as major, place of origin, etc.) and "dynamic behaviors" (such as competitions, training, student work, etc.). Traditional evaluation models often adopt independent modeling or simple splicing methods, which not only ignore the potential impact of static attributes on dynamic behaviors, but also fail to update the ability status of college students in real time through dynamic behaviors, resulting in a lack of continuity and integrity in the evaluation results.

[0004] (2) Lack of a time-series causal analysis mechanism: Existing technologies mostly rely on discrete time point data, ignoring the chronological order of college students' dynamic behaviors and their inherent logical relationships. This makes it difficult for evaluation models to distinguish different growth patterns and to accurately uncover college students' deep cognitive abilities and true growth patterns.

[0005] (3) Difficulty in understanding complex search intents and insufficient data association: Traditional search methods based on structured query language or keyword matching have difficulty understanding natural language evaluation intents with complex semantics. At the same time, due to the differences in data structure between static attribute maps and dynamic behavior maps, there is a lack of deep semantic mapping mechanism between the two, which leads to search results being limited to shallow surface matching and unable to achieve in-depth analysis.

[0006] (4) Lack of objective quantitative evaluation indicators: In view of the multidimensional quality characteristics of college students, the evaluation indicators of traditional methods are often not objective enough, and it is difficult to judge whether their knowledge system is vertically focused on a certain professional field or widely distributed in different aspects. Due to the lack of objective indicators based on graph structure, the evaluation results are often highly subjective and lack interpretability.

[0007] Therefore, there is an urgent need for a new method that can synergistically utilize static and dynamic knowledge graphs and integrate the reasoning capabilities of large language models (LLM) to achieve in-depth cognitive diagnosis and multi-dimensional profile construction for college students. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models. This method achieves cognitive diagnosis and multi-dimensional profile construction of college students by designing path retrieval planning, multi-source evidence anchoring retrieval, cognitive diagnosis and reasoning modules, and a multi-dimensional student profile generation module. It effectively solves the problems in the existing technologies mentioned above, such as insufficient integration of dynamic and static data, lack of analysis of the learning process of college students, and lack of objective evaluation indicators.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for constructing a multi-dimensional profile of college students based on knowledge graphs and large models includes:

[0011] S1. Academic Path Retrieval Plan Generation: Obtain the profile construction task instructions, as well as the pre-constructed static college student basic information knowledge graph and dynamic college student behavior knowledge graph; based on the construction task instructions and the predefined pattern structure followed by the static college student basic information knowledge graph and the dynamic college student behavior knowledge graph, generate an academic path retrieval plan to reflect cross-graph query.

[0012] S2. Based on the academic path retrieval plan, the static knowledge graph of basic information of college students and the dynamic knowledge graph of college student behavior are integrated for retrieval and information aggregation to obtain a structured chain of evidence;

[0013] S3. Quantitative Diagnostic Analysis Based on Structured Evidence Chains: The structured evidence chain set is split into a unique entity set and an event sequence; spatial dimension analysis is performed on the unique entity set, outputting entropy parameters and corresponding natural language evaluations; temporal dimension analysis is performed on the event sequence, and graph edit distance is calculated to obtain path deviation analysis results; cognitive diagnostic indicators are generated based on entropy parameters, corresponding natural language evaluations, and path deviation analysis results.

[0014] S4. Based on the attention mechanism, cognitive diagnostic indicators, structured evidence chains, and static features extracted from the static knowledge graph of college students' basic information and processed by Embedding are fused to obtain fused features; based on the fused features, attribution reasoning is performed using a large language model to generate a multi-dimensional college student profile.

[0015] Furthermore, in step S1, generating an academic path retrieval plan to reflect cross-graph query includes:

[0016] The large language model is used to parse the profile and construct task instructions to obtain the evaluation intent;

[0017] The evaluation intent is mapped to the predefined pattern structure followed by the static college student basic information knowledge graph and the dynamic college student behavior knowledge graph, generating meta-path templates.

[0018] Furthermore, the fusion retrieval and information aggregation of static knowledge graphs of basic information about college students and dynamic knowledge graphs of college student behavior include:

[0019] Based on the predefined pattern structure followed by the static college student basic information knowledge graph and the dynamic college student behavior knowledge graph, a static-dynamic fusion subgraph is constructed.

[0020] On the static-dynamic fusion subgraph, execute an academic path retrieval plan based on meta-path templates.

[0021] Furthermore, on the static-dynamic fusion subgraph, executing an academic path retrieval plan based on meta-path templates includes:

[0022] The meta-path template is parsed to obtain path constraints suitable for structured queries and semantic targets suitable for fuzzy queries;

[0023] Based on path constraints, a constrained BFS algorithm based on meta-path templates is executed in the static-dynamic fusion subgraph to obtain explicit evidence;

[0024] Based on semantic objectives, semantic heuristic path exploration is performed in the static-dynamic fusion subgraph to obtain implicit evidence;

[0025] The explicit and implicit evidence are combined to form a structured chain of evidence.

[0026] Furthermore, the process of obtaining explicit evidence by executing a constrained BFS algorithm based on meta-path templates in the static-dynamic fusion subgraph includes:

[0027] Constrained BFS Algorithm Initialization: Constructing the Search Queue This will match the starting type of the metapath template. Push the starting node into the queue, initial depth ;

[0028] Constraint expansion: When the queue is not empty and hour, This indicates the total number of hops in the generated metapath template, and the current node is retrieved. traversal All neighboring nodes ;right Perform type validation and only include the next hop whose type matches the metapath template. Neighboring nodes of the same type are added to the search queue, and the remaining nodes are pruned.

[0029] Forming candidate paths: During the constraint expansion process, traversal paths from the starting node to nodes that conform to the type constraints at each level are collected to form a set of candidate paths;

[0030] Candidate path selection: A time-series causal support strategy is adopted. Based on the time attributes of each node in the candidate path, a time-series causal support score is calculated for each path. Specifically, dynamic nodes are defined by their timestamps, and static nodes are assigned a default value to define their time attributes. The time-series causal support score is calculated by combining a causal direction factor and a time decay factor to simultaneously evaluate the causal logic correctness and temporal correlation strength of the path. Candidate paths are selected by discarding those containing causal logic paradoxes caused by reversed timestamps. The selected paths are ranked according to their time-series causal support scores, and the top K paths with the highest scores are selected as the explicit evidence.

[0031] Furthermore, the process of performing semantic heuristic path exploration in the static-dynamic fusion subgraph to obtain implicit evidence includes:

[0032] The unstructured text corresponding to each node in the static-dynamic fusion subgraph is transformed into semantic vectors; based on the semantic target and the semantic vectors of each node, the semantic similarity between nodes is calculated.

[0033] Based on a preset semantic connectivity threshold, semantic edges are added between node pairs with semantic similarity greater than the preset connectivity threshold to form a semantically enhanced fusion subgraph.

[0034] In the semantically enhanced fusion subgraph, guided by the semantic target, a beam search algorithm is used to explore multi-hop paths. The beam search algorithm uses a heuristic scoring function to evaluate and filter paths. The heuristic scoring function is calculated based on the semantic relevance between the semantic vectors of nodes in the path and the semantic target. The filtered paths are aggregated to form the implicit evidence.

[0035] Furthermore, in step S3, the specific process of performing spatial dimension analysis on the unique entity set and outputting entropy parameters and corresponding natural language evaluation includes:

[0036] For each unique entity set, a semantic association network based on vector similarity is constructed. A breadth-first search (BFS) strategy is used to traverse this semantic association network, identify connected components, and define each connected component as a knowledge cluster. This yields a set of knowledge clusters;

[0037] For a set of knowledge clusters, perform knowledge structure entropy calculation: calculate the entropy value based on the knowledge structure entropy formula, output the quantified entropy result, and provide a natural language evaluation based on the entropy value; the knowledge structure entropy calculation formula is defined as:

[0038] ;

[0039] in, The entropy value. For the first A knowledge cluster (connected component). The percentage of clusters.

[0040] Furthermore, in step S3, the specific process of performing time-dimensional analysis on the event sequence and calculating the graph edit distance to obtain the path deviation analysis results includes:

[0041] Construct a standard path database for training programs, which stores the identifiers of each major and their corresponding standard paths for training programs.

[0042] Based on the target students' major information, the matching standard training program path is retrieved from the database and denoted as... ;

[0043] Actual behavioral sequence of college students Noise cleaning is performed to filter out all irrelevant behavior nodes that do not belong to the professional domain, generating a normalized sequence aligned with the standard path structure. ;

[0044] Using dynamic programming algorithm to Convert to After minimizing the cost, calculate the normalized graph edit distance;

[0045] The minimum cost path is obtained based on the graph edit distance, and the deviation analysis results are calculated.

[0046] Furthermore, the specific steps of step S4 include:

[0047] Feature projection and stitching: Static features, structured evidence chain sets, and cognitive diagnostic indicators are mapped to a unified dimensional space through feature projection to complete feature stitching; at the same time, textual descriptions of key facts and indicators are extracted from static features, structured evidence chain sets, and cognitive diagnostic indicators.

[0048] Attention mechanism fusion: Taking the concatenated features as input, an attention-based perceptual network is used, and after SoftMax normalization, the weight distribution is output. ;in For static data weight distribution, For dynamic data weight distribution, Weight distribution of diagnostic indicators;

[0049] Feature-weighted fusion: utilizing weight distribution The projected features are weighted and fused to obtain the fused features;

[0050] Constructing Mind Chain Cue Keywords: Based on the text description, and according to the importance weight distribution 'a', fill in the following three-layer cue keyword template:

[0051] Fact level: According to The importance of the static characteristics reflected Importance of dynamic features of the reaction, extraction of static features V S Key facts in dynamic features;

[0052] Indicator layer: Based on The importance of the cognitive diagnostic indicators reflected, and the extraction of diagnostic indicator information;

[0053] Prediction layer: Integrates factual information and cognitive indicator diagnostic information to generate a problem task framework for attribution analysis, in order to guide subsequent reasoning;

[0054] Context-enhanced reasoning: The constructed thought chain cue words and fused features are input into the large language model, and attribution reasoning is performed to obtain a multi-dimensional college student profile output.

[0055] A system for constructing multi-dimensional profiles of college students based on dynamic and static knowledge graphs and large models includes:

[0056] The path retrieval planning module is configured to call the large language model interface to generate an academic path retrieval plan based on the given task instructions and knowledge graph pattern definition.

[0057] The multi-source evidence anchoring retrieval module is configured to parse the academic path retrieval plan to extract path constraints and semantic goals; and to execute a constrained BFS algorithm in the dynamic and static fusion knowledge graph based on the path constraints, while performing semantic heuristic path exploration based on the semantic goals to obtain a set of structured evidence chains.

[0058] The cognitive diagnosis and reasoning module is configured to decouple and reconstruct the structured evidence chain set, and apply the structural entropy algorithm and the graph editing distance algorithm to output cognitive diagnosis indicators including entropy value indicators and path deviation indicators.

[0059] The multi-dimensional student profile generation module is configured to integrate static features, structured evidence chain sets, and cognitive diagnostic indicators based on the attention mechanism, and generate multi-dimensional college student profiles by calling a large language model for attribution reasoning.

[0060] The knowledge storage medium stores a static knowledge graph of basic information about college students, a dynamic knowledge graph of college student behavior, and a standard path library for the training plan.

[0061] In summary, by adopting the above technical solution, the present invention has the following beneficial effects:

[0062] (1) Achieving deep integration of static attributes and dynamic behaviors of college students: This invention enhances the data correlation between static attributes and dynamic behaviors of college students by constructing a static-dynamic fusion subgraph architecture and utilizing meta-path planning technology. It effectively solves the problem of data segmentation in existing technologies and realizes the evaluation of college students' comprehensive qualities from a full-cycle perspective;

[0063] (2) It can effectively tap the hidden growth potential of college students: The present invention adopts the GED algorithm, which can accurately calculate the difference between the actual dynamic behavior path of college students and the standard training path. Through analysis, it can effectively identify the growth patterns of different college students and overcome the shortcomings of traditional evaluation methods that rely only on discrete results statistics and ignore the causality of the learning process.

[0064] (3) A quantitative evaluation system based on graph topology was established: This invention innovatively designs a structural entropy algorithm to numerically analyze the structural features of the fused graph. This transforms the distribution characteristics of university students' qualities into calculable objective mathematical indicators, reducing the subjectivity of manual evaluation;

[0065] (4) Improved interpretability of the generated profile: This invention combines the reasoning ability of LLM to construct a complete chain of evidence based on static attributes and dynamic behavior. The final output multidimensional profile not only includes quantitative indicators, but also provides a natural language explanation from data to conclusions, effectively improving the interpretability of the algorithm. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the overall process of the method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models according to the present invention.

[0067] Figure 2 This is a schematic diagram of the path retrieval and planning module;

[0068] Figure 3 This is a schematic diagram of the multi-source evidence anchoring retrieval module;

[0069] Figure 4 This is a schematic diagram of the cognitive diagnosis and reasoning module;

[0070] Figure 5 A schematic diagram of the module for generating multidimensional student profiles for college students;

[0071] Figure 6 This is a schematic diagram of a device for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0073] Example 1: As Figure 1 As shown in the figure, this embodiment provides a method for constructing a multi-dimensional profile of college students based on knowledge graphs and large models, including the following steps:

[0074] S1. Academic Path Retrieval Plan Generation: This step aims to construct task instructions based on the target user's profile and utilize LLM to generate a non-SQL academic path retrieval plan, effectively reducing the search space of full-graph retrieval. For example... Figure 2 As shown, its implementation process includes the following sub-steps: S11. Data acquisition: Acquire the natural language profile construction task instructions, as well as the pre-constructed static college student basic information knowledge graph and dynamic college student behavior knowledge graph. This embodiment introduces [a method / mechanism] to ensure that the paths generated by the subsequent LLM actually exist in the graph.

[0075] S12. LLM Path Planning: Taking the portrait building task instructions as input, the semantic reasoning ability of the Large Language Model (LLM) is used to parse the portrait building task instructions and obtain the evaluation intent;

[0076] S13. Meta-path template generation: Map the evaluation intent to the predefined schema structure followed by the static college student basic information knowledge graph and the dynamic college student behavior knowledge graph to generate a meta-path template, that is, output a set of abstract meta-paths, such as college student -> major -> competition -> award, rather than specific entity IDs;

[0077] S14. Output the results: Output the meta-path template generated in step S13 as the academic path retrieval plan.

[0078] S2. Based on the academic path retrieval plan, perform fusion retrieval and information aggregation on the static university student basic information knowledge graph and the dynamic university student behavior knowledge graph to obtain a structured evidence chain. This step constructs a static-dynamic fusion subgraph based on the predefined pattern structure followed by the static university student basic information knowledge graph and the dynamic university student behavior knowledge graph; then, on the static-dynamic fusion subgraph, execute the academic path retrieval plan based on the meta-path template. For example... Figure 3 As shown, its implementation process includes the following sub-steps:

[0079] S21. Task Distribution: Parse the meta-path template to obtain path constraints suitable for structured queries and semantic targets suitable for fuzzy queries.

[0080] S22. Based on path constraints, execute the constrained BFS algorithm based on meta-path templates in the static-dynamic fusion subgraph to obtain explicit evidence; based on semantic objectives, execute semantic heuristic path exploration in the static-dynamic fusion subgraph to obtain implicit evidence.

[0081] For explicit evidence, this embodiment uses a fusion retrieval executor to execute a constrained BFS algorithm based on meta-path templates in the static-dynamic fusion subgraph to obtain explicit evidence. The input of the fusion retrieval executor is the meta-path template parsed in step S21. ,in Representing the Type constraints for jump nodes, for example: =College student-> =Match-> =Awards).

[0082] The execution steps of the constrained BFS algorithm are as follows:

[0083] a1. Constrained BFS Algorithm Initialization: Constructing the Search Queue This will match the starting type of the metapath template. The starting node, college student A, is pushed into the queue, with an initial depth of... ;

[0084] a2. Constraint-based expansion: When the queue is not empty and hour, Represents the total number of hops in the generated metapath template; retrieves the current node. traversal All neighboring nodes ;right Perform type validation and check. Is it equal to the next-hop type defined in the metapath template? Only nodes that satisfy the type constraints will be added to the next level of the search queue; otherwise, they will be pruned.

[0085] a3. Forming candidate paths: During the constraint expansion process, collect traversal paths from the starting node to nodes that conform to the type constraints at each level to form a set of candidate paths;

[0086] a4. Filtering Candidate Paths to Obtain Explicit Evidence: After obtaining the set of candidate paths, the system needs to filter out logically flawed paths (e.g., "winning an award first, then participating in the competition" or "taking an advanced course in the first year, and a prerequisite course in the fourth year"). To this end, this embodiment uses a temporal causal support strategy to filter candidate paths. Specifically:

[0087] Based on candidate paths The time attribute definition and time attribute function of the middle node Among them, if For dynamic nodes such as behaviors and events, Its timestamp; if Set for static nodes such as major and place of origin. As a precondition that is always valid, it is based on a time attribute function. Calculate each candidate path using the following formula. Temporal causal support score ;

[0088] ;

[0089] in, Candidate paths The total number of nodes; This is a time-decrease factor used to measure the strength of causal relationships; The normalized time interval, The attenuation coefficient; This is a causal direction indicator factor used to determine whether the temporal sequence between two nodes is logically consistent. Its value is determined by the following rules:

[0090] ;

[0091] The shorter the time interval between two events, the greater the likelihood of a causal relationship and the higher the score; if the interval is too large, such as participating in the competition in the first year of university and winning the award in the fourth year, the causal support will decrease.

[0092] Based on the calculation Execute the filtering strategy; if any hop exists in the path... If a jump results in a score of 0, the path is marked as a "logical paradox path" and discarded. For example, the erroneous data path "Winning in 2024 -> Participating in 2023" needs to be filtered out. For the retained valid paths, [following the correct procedure]... Sort the paths from highest to lowest confidence and select the top-K paths as high-confidence path instances. These high-confidence path instances are the explicit evidence.

[0093] For implicit evidence, this embodiment uses a semantic aggregation executor to perform semantic heuristic path exploration in the static-dynamic fusion subgraph; the specific process is as follows:

[0094] b1. Using BERT semantic vectorization, the unstructured text corresponding to each node in the semantic target and the static-dynamic fusion subgraph is converted into semantic vectors. To discover implicit associations, semantic vectors are used as a basis. The cosine similarity formula is used to calculate the nodes within the static-dynamic fused subgraph. and The semantic similarity, cosine similarity is calculated using the following formula:

[0095] ;

[0096] in, For nodes semantic vectors, For nodes semantic vectors, For nodes The magnitude of the high-dimensional semantic vector output after encoding by the BERT model. For nodes The magnitude of the high-dimensional semantic vector output after encoding by the BERT model;

[0097] b2. According to the conditions met Add semantic edges to form a semantically enhanced fusion subgraph; where, To preset the semantic connectivity threshold, For the node and Semantic edges are established between them.

[0098] b3. In the semantically enhanced fusion subgraph, guided by the semantic target, a beam search algorithm is used to explore multi-hop paths; wherein, the beam search algorithm uses a heuristic scoring function to evaluate and filter paths to form the implicit evidence. Specifically:

[0099] Set the current path The end node is The semantic vector corresponding to the semantic target is Each path is scored using a path scoring function, which is:

[0100] ;

[0101] in, For path The cumulative mean of all semantic edge weights. End node With semantic target vector semantic similarity, It serves as a moderating factor, used to balance path quality and target relevance.

[0102] S23. The implicit semantic paths output by the semantic aggregation executor and the explicit temporal paths output by the fusion retrieval executor are converged. If the node sequence overlap of the two paths exceeds 90%, they are retained. A higher approach is to convert the final set of paths into a structured chain of evidence in a unified JSON format.

[0103] S3. Quantitative diagnostic analysis based on structured evidence chains: such as Figure 4 As shown, this embodiment uses a parallel two-branch approach based on structured evidence. The left branch is responsible for stability diagnosis, calculating semantic similarity to build edges -> BFS traversal to divide clusters -> calculating knowledge structure entropy. -> Generate entropy value semantic interpretation; the right branch is responsible for growth diagnosis, loading the standard path of the culture plan -> mapping the actual trajectory -> GED -> analyzing path deviation; details are as follows:

[0104] S31. Data View Decoupling and Reconstructor: Using data view decoupling and reconstruction technology, the structured evidence chain set is split into a unique entity set for spatial analysis and an event sequence for temporal analysis; in this embodiment, the unique entity set refers to the set consisting of the entities themselves, after removing time attributes and duplicate nodes from all path nodes of the structured evidence chain set.

[0105] S32. Parallel analysis of two branches:

[0106] Spatial analysis calculates the entropy value of the knowledge structure: Spatial dimension analysis is performed on a unique entity set, outputting entropy parameters and corresponding natural language evaluations. The calculation process, using the entropy parameters and corresponding natural language evaluations to quantitatively analyze the topological characteristics of university students' knowledge system, includes:

[0107] a1. Construct a semantic association network based on vector similarity for a unique entity set. Use a breadth-first search (BFS) strategy to traverse the semantic association network, identify connected components, and define each connected component as a knowledge cluster. This yields a set of knowledge clusters;

[0108] a2. For the knowledge cluster set, perform knowledge structure entropy calculation: calculate the entropy value based on the knowledge structure entropy value formula, output the quantified entropy value result, and provide a natural language evaluation based on the entropy value magnitude; the knowledge structure entropy value calculation formula is defined as:

[0109] ;

[0110] in, The entropy value. For the first A knowledge cluster (connected component). The percentage of clusters.

[0111] A time-dimensional analysis is performed on the event sequence to calculate the graph edit distance (GED) and obtain the path deviation analysis results. The path deviation analysis results are used to characterize the temporal deviation of the growth trajectory, and the calculation process includes:

[0112] b1. Construct a standard path database for training programs, which stores the identifiers of each major and their corresponding standard paths for training programs;

[0113] b2. Based on the target students' professional information in the profile, retrieve the matching standard training program path from the database, denoted as... ;

[0114] b3. The actual behavioral sequence of college students Noise cleaning is performed to filter out all irrelevant behavior nodes that do not belong to the professional domain, generating a normalized sequence aligned with the standard path structure. ;

[0115] b4. Using dynamic programming algorithm to... Convert to The minimum cost is as follows:

[0116] Define the state transition equation :

[0117]

[0118] in, Represented as The former Each node is converted to The former Minimum cost for each node; Insertion cost represents the cost required to insert an additional action that a college student actually performs but is not present in the standard path into the standard path. The missing cost represents the cost required to delete a node from the standard path, and this cost is related to the node's importance weight coefficient. Related. Set an importance threshold. =0.8, if standard path nodes weight If the node is a core, high-weight node, then set... =5.0; if Then set =1.0); The replacement cost is the nodes in the student's actual path. Replace with a node in the standard path The cost required is calculated based on the semantic distance between the two nodes;

[0119] Calculate graph edit distance : In obtaining Convert to Minimum total cost After that, where n is the actual sequence of students after alignment. The total length, m is the standard path of the training plan. Total length, calculate normalized graph edit distance:

[0120] ;

[0121] in, ) is the first in the standard path The importance weight coefficient of each node;

[0122] b5. The minimum cost path derived during dynamic programming is from... arrive The optimal operation sequence is obtained, and the deviation analysis results are output, such as learning behavior patterns like positive jump or progressive lag.

[0123] S4. Based on the attention mechanism, cognitive diagnostic indicators, structured evidence chains, and static features extracted from static knowledge graphs and processed by embedding are fused to obtain fused features; based on the fused features, a large language model is used for attribution reasoning to generate a multi-dimensional profile of college students. For example... Figure 5 As shown, the specific process of step S4 includes:

[0124] S41. Feature Projection and Stitching: Projecting static features... Structured chain of evidence and cognitive diagnostic indicators Feature stitching is completed by projecting features onto a unified dimensional space; among them, the cognitive diagnostic metric is graph editing distance. and structural entropy constitute.

[0125] S42. Attention Mechanism Fusion: Using the concatenated features as input, an attention-based perceptual network is used. After SoftMax normalization, the output weight distribution is obtained. ;in For static data weight distribution, For dynamic data weight distribution, The weight distribution of diagnostic indicators.

[0126] S43. Perform weighted calculation: using weight distribution The projected features are weighted and fused to obtain the fused features. The calculation formula is as follows: .

[0127] S44. Constructing Mind Chain Cue Words: To address the large model illusion problem and enhance interpretability, based on the importance weight distribution 'a', fill in the following three-layer cue word template to construct a mind chain cue word that follows a logical closed loop of "fact-indicator-prediction":

[0128] Fact level: According to The importance of the reflected static characteristics is relevant to this student's major in software engineering; combined with... The importance of dynamic features of the response, which made it from the school competition to the national competition in 2024; extracting static features V S Key factual information in dynamic features;

[0129] Indicator layer: based on Reflecting the importance of cognitive diagnostic indicators, extracting diagnostic indicator information; knowledge structure entropy. This indicates a highly focused knowledge system; graph edit distance Displays a positive jump;

[0130] Prediction layer: Generates a problem task framework for attribution analysis by integrating factual information and cognitive indicator diagnostic information to guide subsequent reasoning; if the inference is consistent: this student is an intuitively sharp talent, it is recommended to cultivate him in the direction of algorithm research;

[0131] Context-enhanced reasoning: The constructed thought chain prompts and fused features are input into the Large Language Model (LLM) to perform attribution reasoning, resulting in a multi-dimensional profile of college students that includes radar chart data and natural language analysis reports.

[0132] Example 2: Figure 6 As shown, this implementation provides a system for constructing multi-dimensional profiles of college students based on dynamic and static knowledge graphs and large models, including:

[0133] The path retrieval planning module is configured to call the large language model interface to generate an academic path retrieval plan based on the given task instructions and knowledge graph pattern definition.

[0134] The multi-source evidence anchoring retrieval module is configured to parse the academic path retrieval plan to extract path constraints and semantic goals; and to execute a constrained BFS algorithm in the dynamic and static fusion knowledge graph based on the path constraints, while performing semantic heuristic path exploration based on the semantic goals to obtain a set of structured evidence chains.

[0135] The cognitive diagnosis and reasoning module is configured to decouple and reconstruct the structured evidence chain set, and apply the structural entropy algorithm and the graph editing distance algorithm to output cognitive diagnosis indicators including entropy value indicators and path deviation indicators.

[0136] The multi-dimensional student profile generation module is configured to integrate static features, structured evidence chain sets, and cognitive diagnostic indicators based on the attention mechanism, and generate multi-dimensional college student profiles by calling a large language model for attribution reasoning.

[0137] The knowledge storage medium stores a static knowledge graph of basic information about college students, a dynamic knowledge graph of college student behavior, and a standard path library for the training program. The static knowledge graph of basic information about college students consists of relatively stable entity data, including student ID, major, and place of origin. The dynamic knowledge graph of college student behavior consists of streaming data stored in time series, including behavioral event information such as competition awards, project experience, and library borrowing. The standard path library for the training program includes predefined standard growth paths for each major, i.e., expert path data, which are used to provide a benchmark for cognitive diagnosis.

[0138] The above-mentioned method for creating a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models is achieved through the collaborative work of these block modules.

[0139] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models, characterized in that: include: S1. Academic Path Retrieval Plan Generation: Obtain the profile construction task instructions, as well as the pre-built static knowledge graph of basic information of college students and the dynamic knowledge graph of college student behavior; Based on the task instructions for construction, and the predefined pattern structure followed by the static knowledge graph of college students' basic information and the dynamic knowledge graph of college students' behavior, an academic path retrieval plan is generated to reflect cross-graph queries. S2. Based on the academic path retrieval plan, the static knowledge graph of basic information of college students and the dynamic knowledge graph of college student behavior are integrated for retrieval and information aggregation to obtain a structured chain of evidence; Its implementation process includes: Based on the predefined pattern structure followed by the static college student basic information knowledge graph and the dynamic college student behavior knowledge graph, a static-dynamic fusion subgraph is constructed. On the static-dynamic fusion subgraph, an academic path retrieval plan based on meta-path templates is executed, specifically: The meta-path template is parsed to obtain path constraints suitable for structured queries and semantic targets suitable for fuzzy queries; Based on path constraints, a constrained BFS algorithm based on meta-path templates is executed in the static-dynamic fusion subgraph to obtain explicit evidence; Based on semantic objectives, semantic heuristic path exploration is performed in the static-dynamic fusion subgraph to obtain implicit evidence; The explicit and implicit evidence are combined to form a structured chain of evidence; S3. Quantitative Diagnostic Analysis Based on Structured Evidence Chains: The structured evidence chain set is split into a unique entity set and an event sequence; spatial dimension analysis is performed on the unique entity set, outputting entropy parameters and corresponding natural language evaluations; temporal dimension analysis is performed on the event sequence, and graph edit distance is calculated to obtain path deviation analysis results; cognitive diagnostic indicators are generated based on entropy parameters, corresponding natural language evaluations, and path deviation analysis results. S4. Based on the attention mechanism, cognitive diagnostic indicators, structured evidence chains, and static features extracted from the static knowledge graph of college students' basic information are fused to obtain fused features; based on the fused features, attribution reasoning is performed using a large language model to generate a multi-dimensional college student profile.

2. The method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models as described in claim 1, characterized in that, In step S1, generating an academic path retrieval plan to reflect cross-graph query includes: The large language model is used to parse the profile and construct task instructions to obtain the evaluation intent; The evaluation intent is mapped to the predefined pattern structure followed by the static college student basic information knowledge graph and the dynamic college student behavior knowledge graph, generating meta-path templates.

3. The method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models as described in claim 2, characterized in that, The process of obtaining explicit evidence by performing a constrained BFS algorithm based on meta-path templates in a static-dynamic fusion subgraph includes: Constrained BFS Algorithm Initialization: Constructing the Search Queue This will match the starting type of the metapath template. Push the starting node into the queue, initial depth ; Constraint expansion: When the queue is not empty and , The total number of hops in the represented metapath template: retrieve the current node. traversal All neighboring nodes ;right Perform type validation and only include the next hop whose type matches the metapath template. Neighboring nodes of the same type are added to the search queue, and the remaining nodes are pruned. Forming candidate paths: During the constraint expansion process, traversal paths from the starting node to nodes that conform to the type constraints at each level are collected to form a set of candidate paths; Candidate path selection: A time-series causal support strategy is adopted. Based on the time attributes of each node in the candidate path, a time-series causal support score is calculated for each path. Specifically, dynamic nodes are defined by their timestamps, and static nodes are assigned a default value to define their time attributes. The time-series causal support score is calculated by combining a causal direction factor and a time decay factor to simultaneously evaluate the causal logic correctness and temporal correlation strength of the path. Candidate paths are selected by discarding those containing causal logic paradoxes caused by reversed timestamps. The selected paths are ranked according to their time-series causal support scores, and the top K paths with the highest scores are selected as the explicit evidence.

4. The method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models as described in claim 3, characterized in that, The process of performing semantic heuristic path exploration in the static-dynamic fusion subgraph to obtain implicit evidence includes: The unstructured text corresponding to each node in the static-dynamic fusion subgraph is transformed into semantic vectors; based on the semantic target and the semantic vectors of each node, the semantic similarity between nodes is calculated. Based on a preset semantic connectivity threshold, semantic edges are added between node pairs with semantic similarity greater than the preset connectivity threshold to form a semantically enhanced fusion subgraph. In the semantically enhanced fusion subgraph, guided by the semantic target, a beam search algorithm is used to explore multi-hop paths. The beam search algorithm uses a heuristic scoring function to evaluate and filter paths. The heuristic scoring function is calculated based on the semantic relevance between the semantic vectors of nodes in the path and the semantic target. The filtered paths are aggregated to form the implicit evidence.

5. The method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models as described in claim 1, characterized in that, In step S3, the specific process of performing spatial dimension analysis on the unique entity set and outputting entropy parameters and corresponding natural language evaluation includes: For each unique entity set, a semantic association network based on vector similarity is constructed. A breadth-first search strategy is used to traverse this semantic association network, identify connected components, and define each connected component as a knowledge cluster. This yields a set of knowledge clusters; For a set of knowledge clusters, perform knowledge structure entropy calculation: calculate the entropy value based on the knowledge structure entropy formula, output the quantified entropy result, and provide a natural language evaluation based on the entropy value; the knowledge structure entropy calculation formula is defined as: ； in, The entropy value. For the first Each knowledge cluster is a connected component. The percentage of clusters.

6. The method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models as described in claim 1, characterized in that, In step S3, the time dimension analysis of the event sequence is performed to calculate the graph edit distance and obtain the path deviation analysis results. The specific process includes: Construct a standard path database for training programs, which stores the identifiers of each major and their corresponding standard paths for training programs. Based on the target students' major information, the matching standard training program path is retrieved from the database and denoted as... ; Actual behavioral sequence of college students Noise cleaning is performed to filter out all irrelevant behavior nodes that do not belong to the professional domain, generating a normalized sequence aligned with the standard path structure. ; Using dynamic programming algorithm to Convert to After minimizing the cost, calculate the normalized graph edit distance; The minimum cost path is obtained based on the graph edit distance, and the deviation analysis results are calculated.

7. The method for constructing a multi-dimensional profile of college students based on dynamic and static knowledge graphs and large models as described in claim 1, characterized in that, The specific steps of step S4 include: Feature projection and stitching: Static features, structured evidence chain sets, and cognitive diagnostic indicators are mapped to a unified dimensional space through feature projection to complete feature stitching; at the same time, textual descriptions of key facts and indicators are extracted from static features, structured evidence chain sets, and cognitive diagnostic indicators. Attention mechanism fusion: Taking the concatenated features as input, an attention-based perceptual network is used, and after SoftMax normalization, the output weight distribution is obtained. ;in For static data weight distribution, For dynamic data weight distribution, Weight distribution of diagnostic indicators; Feature-weighted fusion: utilizing weight distribution The projected features are weighted and fused to obtain the fused features; Constructing thought chain prompts: Based on the extracted text descriptions, and according to the importance weight distribution 'a', fill in the following three-layer prompt template: Fact level: According to The importance of the static characteristics reflected Importance of dynamic features of the reaction, extraction of static features V S Key facts in dynamic features; Indicator layer: based on The importance of the cognitive diagnostic indicators reflected, and the extraction of diagnostic indicator information; Prediction layer: Integrates factual information and cognitive indicator diagnostic information to generate a problem task framework for attribution analysis, in order to guide subsequent reasoning; Context-enhanced reasoning: The constructed thought chain cue words and fused features are input into the large language model, and attribution reasoning is performed to obtain a multi-dimensional college student profile output.

8. A system for constructing multi-dimensional profiles of college students based on dynamic and static knowledge graphs and large models, characterized in that: include: The path retrieval planning module is configured to call the large language model interface based on the given task instructions and knowledge graph pattern definition to generate an academic path retrieval plan. The multi-source evidence anchoring retrieval module is configured to parse the academic path retrieval plan to extract path constraints and semantic goals; and to execute a constrained BFS algorithm in the dynamic and static fusion knowledge graph based on the path constraints, while performing semantic heuristic path exploration based on the semantic goals to obtain a set of structured evidence chains. The process of obtaining the structured evidence chain set includes: Based on the predefined pattern structure followed by the static college student basic information knowledge graph and the dynamic college student behavior knowledge graph, a static-dynamic fusion subgraph is constructed. On the static-dynamic fusion subgraph, an academic path retrieval plan based on meta-path templates is executed. Specifically: the meta-path templates are parsed to obtain path constraints suitable for structured queries and semantic targets suitable for fuzzy queries; based on the path constraints, a constrained BFS algorithm based on the meta-path templates is executed on the static-dynamic fusion subgraph to obtain explicit evidence; based on the semantic targets, a semantic heuristic path exploration is executed on the static-dynamic fusion subgraph to obtain implicit evidence; the explicit and implicit evidence are summarized to form a structured evidence chain. The cognitive diagnosis and reasoning module is configured to decouple and reconstruct the structured evidence chain set, and apply the structural entropy algorithm and the graph editing distance algorithm to output cognitive diagnosis indicators including entropy value indicators and path deviation indicators. The multi-dimensional student profile generation module is configured to integrate static features, structured evidence chain sets, and cognitive diagnostic indicators based on the attention mechanism, and generate multi-dimensional college student profiles by calling a large language model for attribution reasoning. The knowledge storage medium stores a static knowledge graph of basic information about college students, a dynamic knowledge graph of college student behavior, and a standard path library for the training plan.

Images