Learning path adaptive diagnosis method based on multi-dimensional topology consistency analysis

Abstract

A learning path adaptive diagnosis method based on multi-dimensional topology consistency analysis, comprising: modeling discipline knowledge and its logical relationship into a heterogeneous discipline knowledge graph, extracting topology-aware features through a relationship graph attention network, and offline solidifying the features into a logical feature dictionary in the form of a bit mask to provide a standardized expert benchmark for online real-time diagnosis; mapping the real-time interactive behavior of the learner through bit mask atomic-level or operation into a standardized feature vector aligned with the node index of the discipline knowledge graph, generating an activated subgraph representing the learner's current cognitive field and a dynamic cognitive barycenter vector representing the learner's current learning focus U u ; based on the activated subgraph, the dynamic cognitive barycenter vector U u , the standardized feature vector, generating three-dimensional consistency parallel quantization diagnosis of the learner's cognitive trajectory and the expert benchmark and adaptive intervention decision; converting the quantization diagnosis result into a differentiated interface interaction strategy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology, specifically involving an adaptive diagnostic method for learning paths based on multidimensional topological consistency analysis. Background Technology

[0002] With the digital transformation of education, self-directed learning has become the norm. While knowledge acquisition has become more flexible, the lack of real-time analysis capabilities to rigorously analyze the logical structure of learning paths has led to the following technical problems:

[0003] The discretization of self-directed learning often leads to fragmented knowledge. Although existing research attempts to introduce knowledge graphs, it is difficult to achieve dynamic mapping between fragmented behaviors and deep topological logic in complex interactive scenarios. Current technologies lack the ability to trace the underlying logical chain of the learning path and cannot identify implicit logical gaps caused by leapfrog learning in real time, making it difficult for learners to build a systematic cognitive framework.

[0004] In inquiry-based learning, student paths are highly random, and existing assessment methods mostly rely on lagging score records, making it impossible to quantify the transfer difficulty between steps in the learning path in real time. When students attempt to learn beyond their level, the system struggles to capture topologically discrete and unsupported leaps in learning (i.e., excessively steep cognitive gradients), making it impossible to accurately determine whether the current step exceeds the learner's comprehension.

[0005] Most existing models focus on assessing the amount of knowledge acquired, but cannot accurately identify the risk of compliant behavior but a misunderstanding that has deviated from the core logic. At the same time, due to the lack of efficient computing mechanisms, traditional full graph traversal or path backtracking algorithms will produce significant delays when facing large-scale user concurrency, and cannot meet the engineering requirements of sub-second real-time response in complex subject graph environments.

[0006] In summary, existing online education assessment schemes often focus on outcome-based status recording, lacking the ability to deeply analyze learning paths in terms of logical chain completeness, cognitive gradient smoothness, and logical alignment. Summary of the Invention

[0007] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide an adaptive diagnostic method for learning paths based on multidimensional topological consistency analysis that can deeply analyze the topological relationships of disciplines while taking into account computational efficiency.

[0008] The technical solution adopted to solve the above-mentioned technical problems is: a learning path adaptive diagnosis method based on multidimensional topology consistency analysis, including the following steps:

[0009] Step 1. Model the subject knowledge and its logical relationships into a heterogeneous subject knowledge graph, extract topological awareness features through a relation graph attention network, and solidify them offline as a logical feature dictionary in the form of a bit mask, providing a standardized expert benchmark for online real-time diagnosis;

[0010] Step 2. Map the learner's real-time interactive behavior to a standardized feature vector aligned with the node index of the subject knowledge graph through atomic bitmask OR operation, generating an activation subgraph representing the learner's current cognitive domain and a dynamic cognitive centroid vector U representing the learner's current learning focus. u ;

[0011] Step 3. Based on activation subgraphs and dynamic cognitive centroid vector U u Standardized feature vectors are used to generate parallel quantitative diagnosis and adaptive intervention decisions based on the three-dimensional consistency between learners' cognitive trajectories and expert benchmarks.

[0012] Step 3.1. Execute the multidimensional logical association completeness diagnostic operator to verify the underlying logical support of the current cognitive behavior and identify implicit logical gaps caused by leapfrog learning;

[0013] Step 3.1.1 Using the activation subgraph as the topological boundary constraint, call the target knowledge node v that the learner's current effective behavior sequence points to. t Strong dependency mask M of association pre Perform bitwise operations D list = (M pre AND NOT B u ), B u To enable learners to grasp bit vectors in real time, if D list If the result vector is not empty, then there is an explicit logical break, and D is output directly. list If D is missing a list of knowledge point IDs, execute a circuit breaker and suspend the second-order recursive calculation; list If the result vector is empty, it is considered to be in compliance with direct logic, and a second-order deep tracing will be automatically triggered.

[0014] Step 3.1.2 In the deep source tracing stage, using the activated subgraph as the topological boundary, the target knowledge node v is extracted by recursive diffusion using the transpose adjacency matrix. t The multi-level ancestor node closure compresses long-range logical dependencies into standardized bit vectors, generating the overlap between the learner's actual mastery state and the multi-level closure according to the following formula. :

[0015]

[0016] In the formula, Let be the closure vector. The decay weights are based on topological distance.

[0017] like If the current learning trajectory is ≥ threshold T, it is determined that the current learning trajectory has sufficient logical support, and the current learning progress is maintained; if If the threshold T is less than the threshold value, it is determined that the learner has implicit or long-range logical debt, and execution is performed. AND NOT The calculation locks the list of deeply missing nodes and automatically generates a structured compensation learning path;

[0018] Step 3.2. Execute the cognitive load overflow diagnostic operator to quantify the cognitive transfer difficulty between adjacent learning behaviors;

[0019] Step 3.2.1 Based on the knowledge node pairs involved in adjacent learning behaviors The shortest path hop count generated in a heterogeneous subject knowledge graph At the same time, generate target knowledge points Recent dynamic cognitive focus of learners Semantic spatial distance between After truncating and normalizing the shortest path hop count and normalizing the semantic space distance by mean, hyperbolic tangent compression is applied to eliminate dimensional differences and limit extreme offsets from disrupting the logical framework.

[0020] Step 3.2.2 Construct a cognitive load spillover model The difficulty of migration between adjacent steps is quantified:

[0021]

[0022] In the formula, The preset cognitive bandwidth truncation constant, This represents the average semantic distance between directly related nodes within a heterogeneous subject knowledge graph. The weighting coefficient has a value range of [0,1].

[0023] Step 3.2.3 will Values ​​and security thresholds In comparison, the safety threshold is a preset threshold, or an adaptive threshold determined based on historical learning behavior data, learning outcome data, and the learner's knowledge mastery status. Value ≤ safety threshold If the current step is within the learner's proximal development zone, the current learning flow will be maintained. Value > security threshold If a cognitive overload risk is detected, a logical smoothing instruction is automatically triggered to retrieve and push topological relay knowledge points or auxiliary transitional materials without deviating from the expert path, thereby reducing the pressure of instantaneous information processing.

[0024] Step 3.3 Execute the logical center alignment diagnostic operator to identify inefficient learning states where behavior is compliant but thinking is off track;

[0025] Step 3.3.1 Based on the knowledge points the learner is currently learning Centered on a single node, a contextual closed subgraph is constructed by retrieving the set of adjacent nodes with radius r from a heterogeneous subject knowledge graph. The feature centroid vector, calculated by weighted averaging, serves as the local early warning reference vector. ;

[0026] Step 3.3.2 Generate the learner's cognitive center of gravity and local early warning baseline vector according to the following formula. cosine similarity :

[0027]

[0028] Step 3.3.3 Based on cosine similarity With preset threshold Compare and determine whether the learner is in a state of logical wandering: if The judgment is that the cognitive path is highly aligned with the expert's logic; if The condition is determined to be a logical roaming state. In the contextual closure subgraph, it is highlighted as not yet mastered and related to the local warning reference vector. The core node with the smallest included angle generates logical attraction through visual guidance;

[0029] Step 4. Transform the quantitative diagnostic results into differentiated interface interaction strategies;

[0030] When it is determined that there is a risk of logical gaps, the priority of the page task queue is reconstructed, a knowledge compensation path is generated at the top of the interface, access to higher or subsequent nodes is temporarily locked, and users are guided to complete the repair of missing logic first, so as to ensure the structured completion of the cognitive chain.

[0031] When the current learning behavior is determined to be in an overload warning state, the relay anchor point related to the current node is retrieved and used in a non-intrusive component. Without interrupting the main learning flow, the understanding span is reduced by adding intermediate cognitive steps, thereby achieving dynamic smoothing of the cognitive gradient and reducing the instantaneous processing pressure caused by the large understanding span.

[0032] When it is determined that the current state is in a logical deviation state, the core node is locked, its visual salience parameters are dynamically adjusted, and the interface gravity guidance mechanism is used to generate non-forced visual traction to correct the wandering behavior and guide the user's attention back to the logical main axis defined by the expert.

[0033] As a preferred technical solution, step 1 specifically involves the following steps:

[0034] Step 1.1 Constructing a heterogeneous subject knowledge graph based on subject curriculum textbook content and expert knowledge system V is the set of knowledge nodes, representing atomic knowledge points; E is the set of edges, representing logical connections between nodes; R is the set of heterogeneous relations, containing prerequisite dependencies. Subordinate relationship Other related relationships ;

[0035] Step 1.2 Extract topology-aware features using a relational graph attention network to generate node feature vectors with topology-aware capabilities. ;

[0036] Step 1.3 For each knowledge node in the graph, establish its strong dependencies. Strong dependency mask converted to fixed-length aligned binary bits Based on the globally unique index, a fixed position number is assigned to each knowledge point. The global adjacency matrix is ​​traversed, and all the predecessor dependent nodes of the target node are set from 0 to 1 at the corresponding position.

[0037] Step 1.4 Using the knowledge point address ID as the key, the node feature vector is... With strong dependency mask The associations are encapsulated as key-value pairs and stored.

[0038] As a preferred technical solution, in step 2, the method of mapping the learner's real-time interactive behavior into a standardized feature vector aligned with the node index of the subject knowledge graph through bitmask atomic OR operation is as follows:

[0039] Capture learners' real-time interactive behaviors, parse the associated business tags, extract the corresponding unique knowledge point IDs, and maintain a Boolean bit vector for each learner in memory. The length is consistent with the total number of nodes in the global knowledge graph of the subject. The Boolean bit vector is converted using atomic-level bitwise OR operations. The corresponding knowledge point index position is set from 0 to 1 in real time, and the recorded status is "mastered";

[0040] The method for generating activation subgraphs representing the learner's current cognitive domain is as follows: using Boolean bit vectors To filter the grid, a bitwise AND operation is performed on the row and column dimensions of the global adjacency matrix A, retaining only valid edges where both the start and end nodes are in a mastered state, thus generating an activation subgraph representing the learner's current cognitive domain. ;

[0041] The dynamic cognitive centroid vector representing the learner's current learning focus The generation method is as follows:

[0042] Retrieve the most recent M interaction records using a sliding window, and utilize activation subgraphs. Perform a validity check and generate a valid sequence of actions. Calculate sequences based on subject knowledge graphs Determine the out-degree centrality of each node and generate normalized weight coefficients. Based on weight coefficients Calculate the sequence of subject knowledge graph The feature vectors of each node are linearly weighted and aggregated to generate a dynamic cognitive centroid vector. .

[0043] As a preferred technical solution, the value of M is 5.

[0044] As a preferred technical solution, the attenuation weight based on topological distance , The attenuation coefficient is... This represents the topological distance.

[0045] As a preferred technical solution, the The preferred value is 0.8.

[0046] As a preferred technical solution, the threshold T is preferably 0.6.

[0047] As a preferred technical solution, the preset cognitive bandwidth cutoff constant The preferred value is 5.

[0048] As a preferred technical solution, the security threshold The preferred value is 0.8.

[0049] As a preferred technical solution, the preset threshold The preferred value is 0.8.

[0050] The beneficial effects of this invention are as follows:

[0051] This invention achieves deep quantitative identification of hidden risks such as logical deviations and cognitive gaps by modeling the consistency of learning paths in the spatial dimension, breaking through the limitations of traditional solutions that only focus on the quantity of knowledge points mastered. This significantly improves the granularity of educational diagnosis, enabling the system to capture logical loopholes hidden beneath seemingly compliant behaviors.

[0052] This invention introduces a local benchmark computing mechanism and bitmask filtering technology to narrow the evaluation field from the full spectrum to the local active context, which greatly reduces the computational overhead. While achieving high-precision diagnosis, it also achieves a sub-second response speed and can adapt to heterogeneous spectra from different disciplines, thus possessing strong versatility and engineering scalability.

[0053] This invention achieves a direct mapping from quantitative assessment to user interface interaction. The system can automatically adjust task weights or visual traction based on risk type, enabling real-time correction without disrupting the learning immersion. Simultaneously, by utilizing dynamic axis synthesis technology, it eliminates reliance on manually preset paths, significantly reducing system construction and subsequent maintenance costs. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the adaptive diagnostic method for learning paths based on multidimensional topology consistency analysis, as described in this invention. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0056] Example 1

[0057] This embodiment uses the teaching scenario of sorting algorithms in the computer science fundamentals course "Data Structures and Algorithms" as an example to illustrate the present invention.

[0058] exist Figure 1 In this embodiment, an adaptive diagnostic method for learning paths based on multidimensional topology consistency analysis includes the following steps:

[0059] Step 1. Model the subject knowledge and its logical relationships into a heterogeneous subject knowledge graph. Extract topology-aware features through a relational graph attention network and offline solidify them into a logical feature dictionary in the form of a bitmask. This provides a standardized expert benchmark for online real-time diagnosis. Specifically:

[0060] Step 1.1 Constructing a heterogeneous subject knowledge graph based on subject curriculum textbook content and expert knowledge system V is the set of knowledge nodes, representing atomic knowledge points; E is the set of edges, representing logical connections between nodes; R is the set of heterogeneous relations, containing prerequisite dependencies. Subordinate relationship Other related relationships ;

[0061] This embodiment maps knowledge point IDs to array bases. Bubble Sort Quick sort For example, since quicksort relies on array foundations and swapping logic, adjacent matrix elements in a subject knowledge graph... = 1 and =1. At this time, arrive The standard logical path must pass through Dependencies are transitive. = 1;

[0062] Step 1.2 Extract topology-aware features using a relational graph attention network to generate node feature vectors with topology-aware capabilities. ;

[0063] Step 1.3 For each knowledge node in the graph, establish its strong dependencies. Strong dependency mask converted to fixed-length aligned binary bits Based on a globally unique index, a fixed position number is assigned to each knowledge point. The global adjacency matrix is ​​traversed, and all predecessor dependent nodes of the target node are set from 0 to 1 at their corresponding positions. Then Mpre( ) = [1, 1, 0, 0, ... ], that is and The current bit is set to 1, as Logical admission truth value;

[0064] Step 1.4 Using the knowledge point address ID as the key, the node feature vector is... With strong dependency mask The associations are encapsulated as key-value pairs and stored.

[0065] Step 2. Map the learner's real-time interactive behavior to a standardized feature vector aligned with the node index of the subject knowledge graph through atomic bitmask OR operation, generating an activation subgraph representing the learner's current cognitive domain and a dynamic cognitive centroid vector U representing the learner's current learning focus. u ;

[0066] Step 2.1. Capture learners' real-time interactive behaviors, parse the associated business tags, extract the corresponding unique knowledge point IDs, and maintain a Boolean bit vector for each learner in memory. The length is consistent with the total number of nodes in the global knowledge graph of the subject. The Boolean bit vector is converted using atomic-level bitwise OR operations. The corresponding knowledge point index position is set from 0 to 1 in real time, and the recorded status is "mastered";

[0067] In this embodiment, when a learner triggers the action of watching the instructional video "The Core Recursive Logic of Quick Sort" on the client side, the system aligns the business behavior to a unique knowledge point ID: Locating to the learner's personal bit vector At the 3rd offset index, perform an atomic bitwise OR operation to set that bit to 1, if the learner has previously activated... And skipped Then the current real-time bit vector is = [1, 0, 1, 0, ... ], thus intuitively reflecting the logical gaps in the individual's cognitive space at the underlying binary level;

[0068] Step 2.2. Using Boolean bit vectors To filter the grid, a bitwise AND operation is performed on the row and column dimensions of the global adjacency matrix A, retaining only valid edges where both the start and end nodes are in a mastered state, thus generating an activation subgraph representing the learner's current cognitive domain. ;

[0069] This embodiment is due to [1]=0 (i.e., V2 is not activated and bubble sort is not mastered), the system uses the bit masking mechanism to instantly change the second row of the adjacency matrix A (from (Start) and column 2 (pointing to) Clear all edge data to zero before generating the activation subgraph. In the middle, the originally connected paths { -> -> Due to intermediate nodes Its associated edges were "physically severed" and interrupted, resulting in and Within this slice, it appears as an isolated node or a non-connected state;

[0070] Step 2.3. Retrieve the most recent M=5 interaction records using a sliding window, and utilize the activation subgraph. Perform a validity check and generate a valid sequence of actions. Calculate sequences based on subject knowledge graphs Determine the out-degree centrality of each node and generate normalized weight coefficients. Based on weight coefficients For sequence The feature vectors of each node are linearly weighted and aggregated to generate a dynamic cognitive centroid vector. .

[0071] Step 3. Based on activation subgraphs and dynamic cognitive centroid vector U u Standardized feature vectors are used to generate parallel quantitative diagnosis and adaptive intervention decisions based on the three-dimensional consistency between learners' cognitive trajectories and expert benchmarks.

[0072] Step 3.1. Execute the multidimensional logical association completeness diagnostic operator to verify the underlying logical support of the current cognitive behavior and identify implicit logical gaps caused by leapfrog learning;

[0073] Step 3.1.1 Using the activation subgraph as the topological boundary constraint, call the target knowledge node v that the learner's current effective behavior sequence points to. t Strong dependency mask M of association pre Perform bitwise operations D list = (M pre ( ) AND NOT B u ), B u To enable learners to grasp bit vectors in real time, if D list If the result vector is not empty, then there is an explicit logical break, and D is output directly. list If D is missing a list of knowledge point IDs, execute a circuit breaker and suspend the second-order recursive calculation; list If the result vector is empty, it is considered to be in compliance with direct logic, and a second-order deep tracing will be automatically triggered.

[0074] Step 3.1.2 In the deep source tracing stage, using the activated subgraph as the topological boundary, the target knowledge node v is extracted by recursive diffusion using the transpose adjacency matrix. t The multi-level ancestor node closure compresses long-range logical dependencies into standardized bit vectors, generating the overlap between the learner's actual mastery state and the multi-level closure according to the following formula. :

[0075]

[0076] In the formula, Let be the closure vector. For decay weights based on topological distance, , The attenuation coefficient is... The preferred value is 0.8. Topological distance;

[0077] like If the current learning trajectory is ≥ threshold T, where T = 0.6, it is determined that the current learning trajectory has sufficient logical support, and the current learning progress should be maintained; if If the threshold T is less than the threshold value, it is determined that the learner has implicit or long-range logical debt, and execution is performed. AND NOT The calculation locks the list of deeply missing nodes and automatically generates a structured compensation learning path;

[0078] This embodiment locates nodes using result vectors. If the learning is missing, it is determined that the learner has a "knowledge debt" in the underlying topological structure, and second-order deep tracing is not triggered.

[0079] Step 3.2. Execute the cognitive load overflow diagnostic operator to quantify the cognitive transfer difficulty between adjacent learning behaviors;

[0080] Step 3.2.1 Based on the knowledge node pairs involved in adjacent learning behaviors The shortest path hop count generated in a heterogeneous subject knowledge graph At the same time, generate target knowledge points Recent dynamic cognitive focus of learners Semantic spatial distance between The shortest path hop count is truncated and normalized, and the semantic space distance is mean normalized and then hyperbolic tangent compression is applied to eliminate the dimensional difference between topological distance and semantic distance and limit the excessive influence of abnormal semantic offset on the diagnostic results.

[0081] Step 3.2.2 Construct a cognitive load spillover fusion model The transfer difficulty between adjacent learning steps is quantified:

[0082]

[0083] In the formula, The preset cognitive bandwidth truncation constant is taken in this embodiment. =5, This represents the average semantic distance between directly related nodes within a heterogeneous subject knowledge graph. The weighting coefficient has a value range of [0,1]. This embodiment... ;

[0084] In this embodiment, the bitmask masks the bubble sort algorithm. 2. Enables rapid sorting of nodes leading to the target knowledge in the learner's currently activated subgraph. 3. The effective knowledge dependency path is distorted. After shortest path retrieval, the shortest path hop count between the knowledge node corresponding to the previous learning behavior and the target knowledge node is [not specified]. Reaching the preset cognitive bandwidth truncation limit Therefore, the normalized topological distance term takes a value of 1; at the same time, due to the bubble sort of nodes... 2 is not activated; target node is sorted quickly. 3 and cognitive focus A significant semantic shift occurs between them. The system calculates the semantic space distance and sequentially performs mean normalization and hyperbolic tangent compression to obtain the corresponding semantic distance term. The calculated semantic distance term in this embodiment is greater than 0.6. Since the normalized topological distance term in this embodiment is 1, and the weighting coefficient... = 0.5, therefore the cognitive load overflow satisfies > 0.8.

[0085] Step 3.2.3 will Values ​​and security thresholds In comparison, in this embodiment, the safety threshold... Take 0.8, Value ≤ safety threshold If the cognitive transfer difficulty between adjacent learning behaviors is deemed to be within an acceptable range, then the current learning flow is maintained. Value > security threshold If a cognitive overload risk is detected, a logical smoothing instruction is automatically triggered to retrieve and push topological relay knowledge points or auxiliary transitional materials without deviating from the expert path, thereby reducing the pressure of instantaneous information processing.

[0086] In this embodiment, the result is calculated by the cognitive load overflow diagnostic operator. Greater than the safety threshold Therefore, it is determined that the learner can directly transfer their current cognitive state to the target knowledge node for rapid sorting. At time 3, there is a risk of cognitive load overflow or a tendency to skip levels of learning, which triggers the corresponding logical smoothing instruction.

[0087] Step 3.3 Execute the logical center alignment diagnostic operator to identify inefficient learning states where behavior is compliant but thinking is off track;

[0088] Step 3.3.1 Based on the knowledge points the learner is currently learning Centered on a single node, a contextual closed subgraph is constructed by retrieving the set of adjacent nodes with radius r from a heterogeneous subject knowledge graph. The feature centroid vector, calculated by weighted averaging, serves as the local early warning reference vector. ;

[0089] Step 3.3.2 Generate the learner's cognitive center of gravity and local early warning baseline vector according to the following formula. cosine similarity :

[0090]

[0091] Step 3.3.3 Based on cosine similarity With preset threshold In comparison, in this embodiment, a preset logical threshold is used. =0.8, determine if the learner is in a logical wandering state: if The judgment is that the cognitive path is highly aligned with the expert's logic; if The condition is determined to be a logical roaming state. In the contextual closure subgraph, it is highlighted as not yet mastered and related to the local warning reference vector. The candidate core node with the highest cosine similarity is used to generate logical attraction through visualization;

[0092] This embodiment aggregates and generates individual cognitive center vectors. ,because [1] = 0 (corresponding to bubble sort) This results in the inability of key "exchange logic" feature components to contribute, leading to a decrease in the generated individual cognitive center vector. In high-dimensional space, relative to the expert standard principal axis A significant radial offset (i.e., directional deviation) is observed, and the calculated cosine similarity is below the logical alignment threshold of 0.8. This is identified as a "logical yaw" risk.

[0093] Step 4. Transform the quantitative diagnostic results into differentiated interface interaction strategies;

[0094] When it is determined that there is a risk of logical gaps, the priority of the page task queue is reconstructed, a knowledge compensation path is generated at the top of the interface, access to higher or subsequent nodes is temporarily locked, and users are guided to complete the repair of missing logic first, so as to ensure the structured completion of the cognitive chain.

[0095] When the current learning behavior is determined to be in an overload warning state, the relay anchor point related to the current node is retrieved and used in a non-intrusive component. Without interrupting the main learning flow, the understanding span is reduced by adding intermediate cognitive steps, thereby achieving dynamic smoothing of the cognitive gradient and reducing the instantaneous processing pressure caused by the large understanding span.

[0096] When it is determined that the current state is in a logical deviation state, the core node is locked, its visual salience parameters are dynamically adjusted, and the interface gravity guidance mechanism is used to generate non-forced visual traction to correct the wandering behavior and guide the user's attention back to the logical main axis defined by the expert.

[0097] This embodiment identifies When missing, the compensatory micro-lesson associated with "bubble sort" is forcibly placed at the top of the student learning interface to repair logical gaps from a physical structure perspective. Given that the cognitive load overflow diagnostic operator indicates cognitive overload and the logical center alignment diagnostic operator indicates mental deviation, the system triggers a cognitive re-anchoring action. In this embodiment, the "exchange logic" analysis is prominently highlighted below the video through an interface gravity guidance mechanism, aiming to guide attention back and offset cognitive pressure by strengthening the visual weight of the underlying support.

Claims

1. A learning path adaptive diagnostic method based on multidimensional topological consistency analysis, characterized in that, Includes the following steps: Step 1. Model the subject knowledge and its logical relationships into a heterogeneous subject knowledge graph, extract topological awareness features through a relation graph attention network, and solidify them offline as a logical feature dictionary in the form of a bit mask, providing a standardized expert benchmark for online real-time diagnosis; Step 2. Map the learner's real-time interactive behavior to a standardized feature vector aligned with the node index of the subject knowledge graph through atomic bitmask OR operation, generating an activation subgraph representing the learner's current cognitive domain and a dynamic cognitive centroid vector U representing the learner's current learning focus. u ; Step 3. Based on activation subgraphs and dynamic cognitive centroid vector U u Standardized feature vectors are used to generate parallel quantitative diagnosis and adaptive intervention decisions based on the three-dimensional consistency between learners' cognitive trajectories and expert benchmarks. Step 3.

1. Execute the multidimensional logical association completeness diagnostic operator to verify the underlying logical support of the current cognitive behavior and identify implicit logical gaps caused by leapfrog learning; Step 3.1.1 Using the activation subgraph as the topological boundary constraint, call the target knowledge node v that the learner's current effective behavior sequence points to. t Strong dependency mask M of association pre Perform bitwise operations D list = (M pre AND NOT B u ), B u To enable learners to grasp bit vectors in real time, if D list If the result vector is not empty, then there is an explicit logical break, and D is output directly. list If D is missing a list of knowledge point IDs, execute a circuit breaker and suspend the second-order recursive calculation; list If the result vector is empty, it is considered to be in compliance with direct logic, and a second-order deep tracing will be automatically triggered. Step 3.1.2 In the deep source tracing stage, using the activated subgraph as the topological boundary, the target knowledge node v is extracted by recursive diffusion using the transpose adjacency matrix. t The multi-level ancestor node closure compresses long-range logical dependencies into standardized bit vectors, generating the overlap between the learner's actual mastery state and the multi-level closure according to the following formula. : In the formula, Let be the closure vector. The decay weights are based on topological distance. like If the current learning trajectory is ≥ threshold T, it is determined that the current learning trajectory has sufficient logical support, and the current learning progress is maintained; if If the threshold T is less than the threshold value, it is determined that the learner has implicit or long-range logical debt, and execution is performed. AND NOT The calculation locks the list of deeply missing nodes and automatically generates a structured compensation learning path; Step 3.

2. Execute the cognitive load overflow diagnostic operator to quantify the cognitive transfer difficulty between adjacent learning behaviors; Step 3.2.1 Based on the knowledge node pairs involved in adjacent learning behaviors The shortest path hop count generated in a heterogeneous subject knowledge graph At the same time, generate target knowledge points Recent dynamic cognitive focus of learners Semantic spatial distance between After truncating and normalizing the shortest path hop count and normalizing the semantic space distance by mean, hyperbolic tangent compression is applied to eliminate dimensional differences and limit extreme offsets from disrupting the logical framework. Step 3.2.2 Construct a cognitive load spillover model The difficulty of migration between adjacent steps is quantified: In the formula, The preset cognitive bandwidth truncation constant, This represents the average semantic distance between directly related nodes within a heterogeneous subject knowledge graph. The weighting coefficient has a value range of [0,1]. Step 3.2.3 will Values ​​and security thresholds In comparison, the safety threshold is a preset threshold, or an adaptive threshold determined based on historical learning behavior data, learning outcome data, and the learner's knowledge mastery status. Value ≤ safety threshold If the current step is within the learner's proximal development zone, the current learning flow will be maintained. Value > security threshold If a cognitive overload risk is detected, a logical smoothing instruction is automatically triggered to retrieve and push topological relay knowledge points or auxiliary transitional materials without deviating from the expert path, thereby reducing the pressure of instantaneous information processing. Step 3.3 Execute the logical center alignment diagnostic operator to identify inefficient learning states where behavior is compliant but thinking is off track; Step 3.3.1 Based on the knowledge points the learner is currently learning Centered on a single node, a contextual closed subgraph is constructed by retrieving the set of adjacent nodes with radius r from a heterogeneous subject knowledge graph. The feature centroid vector, calculated by weighted averaging, serves as the local early warning reference vector. ; Step 3.3.2 Generate the learner's cognitive center of gravity and local early warning baseline vector according to the following formula. cosine similarity : Step 3.3.3 Based on cosine similarity With preset threshold Compare and determine whether the learner is in a state of logical wandering: if The judgment is that the cognitive path is highly aligned with the expert's logic; like The condition is determined to be a logical roaming state. In the contextual closure subgraph, it is highlighted as not yet mastered and related to the local warning reference vector. The core node with the smallest included angle generates logical attraction through visual guidance; Step 4. Transform the quantitative diagnostic results into differentiated interface interaction strategies; When it is determined that there is a risk of logical gaps, the priority of the page task queue is reconstructed, a knowledge compensation path is generated at the top of the interface, access to higher or subsequent nodes is temporarily locked, and users are guided to complete the repair of missing logic first, so as to ensure the structured completion of the cognitive chain. When the current learning behavior is determined to be in an overload warning state, the relay anchor point related to the current node is retrieved and used in a non-intrusive component. Without interrupting the main learning flow, the understanding span is reduced by adding intermediate cognitive steps, thereby achieving dynamic smoothing of the cognitive gradient and reducing the instantaneous processing pressure caused by the large understanding span. When it is determined that the current state is in a logical deviation state, the core node is locked, its visual salience parameters are dynamically adjusted, and the interface gravity guidance mechanism is used to generate non-forced visual traction to correct the wandering behavior and guide the user's attention back to the logical main axis defined by the expert.

2. The learning path adaptive diagnosis method based on multidimensional topology consistency analysis according to claim 1, characterized in that, The specific operation of step 1 is as follows: Step 1.1 Constructing a heterogeneous subject knowledge graph based on subject curriculum textbook content and expert knowledge system V is the set of knowledge nodes, representing atomic knowledge points; E is the set of edges, representing logical connections between nodes; R is the set of heterogeneous relations, containing prerequisite dependencies. Subordinate relationship Other related relationships ; Step 1.2 Extract topology-aware features using a relational graph attention network to generate node feature vectors with topology-aware capabilities. ; Step 1.3 For each knowledge node in the graph, establish its strong dependencies. Strong dependency mask converted to fixed-length aligned binary bits Based on the globally unique index, a fixed position number is assigned to each knowledge point. The global adjacency matrix is ​​traversed, and all the predecessor dependent nodes of the target node are set from 0 to 1 at the corresponding position. Step 1.4 Using the knowledge point address ID as the key, the node feature vector is... With strong dependency mask The associations are encapsulated as key-value pairs and stored.

3. The adaptive diagnostic method for learning paths based on multidimensional topological consistency analysis according to claim 1, characterized in that, In step 2, the method of mapping the learner's real-time interactive behavior into a standardized feature vector aligned with the node index of the subject knowledge graph through a bitmask atomic OR operation is as follows: Capture learners' real-time interactive behaviors, parse the associated business tags, extract the corresponding unique knowledge point IDs, and maintain a Boolean bit vector for each learner in memory. The length is consistent with the total number of nodes in the global knowledge graph of the subject. The Boolean bit vector is converted using atomic-level bitwise OR operations. The corresponding knowledge point index position is set from 0 to 1 in real time, and the recorded status is "mastered"; The method for generating activation subgraphs representing the learner's current cognitive domain is as follows: using Boolean bit vectors To filter the grid, a bitwise AND operation is performed on the row and column dimensions of the global adjacency matrix A, retaining only valid edges where both the start and end nodes are in a mastered state, thus generating an activation subgraph representing the learner's current cognitive domain. ; The dynamic cognitive centroid vector representing the learner's current learning focus The generation method is as follows: Retrieve the most recent M interaction records using a sliding window, and utilize activation subgraphs. Perform a validity check and generate a valid sequence of actions. Calculate sequences based on subject knowledge graphs Determine the out-degree centrality of each node and generate normalized weight coefficients. Based on weight coefficients Calculate the sequence of subject knowledge graph The feature vectors of each node are linearly weighted and aggregated to generate a dynamic cognitive centroid vector. .

4. The adaptive diagnostic method for learning paths based on multidimensional topological consistency analysis according to claim 3, characterized in that, The value of M is 5.

5. The adaptive diagnostic method for learning paths based on multidimensional topology consistency analysis according to claim 1, characterized in that, The decay weight based on topological distance , The attenuation coefficient is... This represents the topological distance.

6. The adaptive diagnostic method for learning paths based on multidimensional topology consistency analysis according to claim 1, characterized in that, The The preferred value is 0.

8.

7. The adaptive diagnostic method for learning paths based on multidimensional topology consistency analysis according to claim 1, characterized in that, The threshold T is preferably set to 0.

6.

8. The adaptive diagnostic method for learning paths based on multidimensional topology consistency analysis according to claim 1, characterized in that, The preset cognitive bandwidth cutoff constant The preferred value is 5.

9. The adaptive diagnostic method for learning paths based on multidimensional topological consistency analysis according to claim 1, characterized in that, The security threshold The preferred value is 0.

8.

10. The adaptive diagnostic method for learning paths based on multidimensional topology consistency analysis according to claim 1, characterized in that, The preset threshold The preferred value is 0.8.

Images