Knowledge graph dynamic construction and optimization method based on group behavior and state propagation

Abstract

The invention discloses a knowledge graph dynamic construction and optimization method based on group behaviors and state propagation, and belongs to the field of knowledge graph construction. Generating an initial knowledge graph based on the course data; based on group learning behavior data mining and recessive relationship supplementation, forming an enhanced knowledge graph; and through a state propagation model, predicting the mastering probability of a non-source knowledge point by taking the real mastering probability of a part of knowledge points as a source, iteratively optimizing the association relationship weight according to the predicted difference degree, and outputting a final map. According to the method, the implicit association is mined and the state propagation model is introduced, so that the association edges of the knowledge graph are enriched, and the association relationship weight is optimized by simulating the diffusion effect of the knowledge mastering state in the graph network, so that the weight of each edge contains the real feedback of the group learning rule; the internal logic and influence intensity between the knowledge points can be described more accurately, and the ability of the knowledge graph to adaptively evolve along with the change of course content and learning groups is improved.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph construction, and in particular relates to a method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation. Background Technology

[0002] Adaptive learning systems are at the heart of modern educational technology, aiming to provide learners with personalized learning paths and content recommendations. The intelligence of these systems rests on their internal knowledge representation models, with knowledge graph technology becoming a mainstream solution. Knowledge graphs connect discrete knowledge points through directed edges, forming a structured semantic network. The most basic relationship is the "prerequisite relation," used to describe the dependency order between knowledge points. However, a high-quality adaptive learning system requires not only a static knowledge structure but also a dynamic model capable of understanding the deep, implicit connections between knowledge points, thus laying the foundation for more accurate learning status assessment and path planning.

[0003] Methods for constructing and optimizing knowledge graphs can be broadly categorized into two types. The first type is the manual construction method based on domain expert experience. This method relies on subject matter experts to pre-define all relationships between knowledge points based on the syllabus and their teaching experience. While this approach produces highly authoritative graphs, it is time-consuming and labor-intensive, and it struggles to exhaustively identify all implicit relationships. Once constructed, the graph remains relatively static and cannot evolve based on real student learning data. The second type is the automated graph construction method based on data mining. This type typically utilizes large amounts of student answer sequence data, employing techniques such as association rule mining or sequence pattern mining to discover frequent co-occurrence or sequential occurrence patterns between knowledge points, and adding these as implicit relationships to the graph. Although this method can automatically discover some relationships, the strength of the discovered relationships (such as confidence and support) is often static, lacking an effective closed-loop feedback mechanism to verify and optimize whether these relationships truly reflect the inherent logic and learning patterns between knowledge points.

[0004] It is evident that expert-based construction methods are highly subjective, lack scalability, and are costly; while automated mining methods often "mine but don't care," lacking verification and iterative optimization of the discovered connections at the level of teaching effectiveness. Both methods ultimately lead to knowledge graphs becoming relatively rigid models, unable to self-calibrate and optimize based on the learning outcomes of a massive number of real users. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a dynamic construction and optimization method for knowledge graphs based on group behavior and state propagation. This method overcomes the limitations of existing knowledge graphs being static and lacking verification. It can automatically and continuously evaluate and optimize the accuracy and effectiveness of the knowledge graph itself using group learning behavior data, transforming the knowledge graph from a pre-set "static map" into a living model that can continuously learn and improve itself through data. This provides a more robust and intelligent infrastructure for upper-level learning applications.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, this invention proposes a method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation, comprising the following steps:

[0008] Step 1: Extract a set of knowledge points based on course materials, establish the relationships between knowledge points, and generate an initial knowledge graph; the relationships include at least prerequisite relationships and explicit relevance relationships; preferably, the knowledge graph is marked with core knowledge points, which are obtained by calculating the network centrality index of each knowledge point in the knowledge graph, and the network centrality index includes at least one of degree centrality, betweenness centrality, and proximity centrality.

[0009] Step 2: Obtain learning behavior data generated by user groups during the learning process, mine implicit relationships between knowledge points based on statistical indicators and confidence levels, and supplement the implicit relationships into the initial knowledge graph to generate an enhanced knowledge graph;

[0010] Step 3, with the goal of optimizing the accuracy of the enhanced knowledge graph, performs an iterative optimization process, which includes:

[0011] Obtain the actual probability of the group mastering each knowledge point;

[0012] Using some knowledge points as source nodes, and based on the current network structure and relational weights of the enhanced knowledge graph, the true group mastery probability of the source nodes is propagated in the knowledge graph through a state propagation model in order to predict the group mastery probability of all or some of the other non-source nodes.

[0013] Calculate the difference between the predicted probability of group mastery and the corresponding actual probability of group mastery;

[0014] Adjust the weights of the relationships in the enhanced knowledge graph based on the degree of difference;

[0015] Step 4: Output the final optimized knowledge graph.

[0016] Furthermore, in step S2, the process of mining implicit relationships between knowledge points based on statistical indicators includes:

[0017] Analyze the learning behavior data and calculate at least one statistical indicator among the following: co-occurrence frequency, mastery correlation, learning sequence proximity, and error association probability among knowledge points;

[0018] When the value of the statistical indicator exceeds the preset indicator threshold, a candidate implicit relationship is established between the corresponding knowledge points, and the value of the statistical indicator is normalized and used as the initial weight of the candidate implicit relationship.

[0019] If the candidate latent relation already has a directed edge in the initial knowledge graph, it is not updated; if the candidate latent relation already has an undirected edge in the initial knowledge graph, the association strength of the undirected edge is updated according to the weighted value of the association strength of the existing undirected edge and the initial weight of the candidate latent relation.

[0020] Furthermore, in step S2, the process of mining implicit relationships between knowledge points based on confidence levels includes:

[0021] Listen for user learning behavior events, including question answering events and note-taking events;

[0022] Extract explicit knowledge point sets and implicit knowledge point sets from the learning behavior events;

[0023] Based on the user's mastery of the implicit knowledge points and the results of the learning behavior events, the strength of evidence for the association between the explicit and implicit knowledge points is calculated.

[0024] A candidate relation database is established to store all candidate associations. The current confidence level and the number of times evidence supports a candidate association are updated in real time based on the strength of evidence. If the current confidence level exceeds the confidence level threshold and the number of times evidence supports a candidate association exceeds the minimum number of times threshold, the candidate association is removed from the candidate relation database and added to the enhanced knowledge graph as a latent relation. If the current confidence level is lower than the forgetting threshold or the number of times evidence supports a candidate association has not been updated for a long time, the candidate association is directly removed from the candidate relation database.

[0025] Furthermore, the calculation process for the strength of evidence includes:

[0026] For question-and-answer events, implicit knowledge points are extracted from the question and answer analysis. Candidate associations are established between these implicit knowledge points and the explicit knowledge points in the question. When a user answers incorrectly and does not know the implicit knowledge points, the evidence is assigned the first strength of evidence. When a user answers correctly and does not know the implicit knowledge points, the evidence is assigned the second strength of evidence. When a user answers correctly and knows the implicit knowledge points, the evidence is assigned the third strength of evidence. In other cases, the evidence strength is 0, and the strength of evidence decreases in the following order: first, second, and third.

[0027] For note events, extract commonly mentioned knowledge point pairs from the note text. If the context of the note contains the knowledge point pair, the correlation between the context of the note and the knowledge point pair is base=1; otherwise, base=0.5. Obtain the lower value of the probability that the user has mastered each knowledge point in the knowledge point pair, take the difference between 1 and the lower value, and multiply it by the base value to get the evidence strength of the candidate correlation relationship corresponding to the knowledge point pair.

[0028] The new confidence level is a weighted average of the strength of evidence and the current confidence level recorded in the candidate relation database.

[0029] Furthermore, the step of removing candidate relationships from the candidate relationship database and adding them as implicit relationships to the enhanced knowledge graph specifically involves:

[0030] If the candidate association is a pre-revision relationship, then add a new directed edge with an association strength of 1; if the directed edge already exists, then do not update.

[0031] If the candidate association is a correlation, a new undirected edge is added with the association strength being the current confidence level. If the undirected edge already exists, the association strength of the undirected edge is updated based on the weighted value of the association strength of the existing undirected edge and the current confidence level of the candidate association.

[0032] Furthermore, the edges corresponding to the relationships in the knowledge graph include both directed and undirected edges, and the priority of directed edges is always higher than that of undirected edges.

[0033] Furthermore, the propagation process of the state propagation model satisfies:

[0034] The influence of knowledge point u on knowledge point v is positively correlated with the association weight of the connection between u and v, decays as the shortest path length from u to v increases, and is positively correlated with the centrality index of knowledge point v in the network, wherein the centrality index includes at least one of betweenness centrality and proximity centrality.

[0035] Furthermore, during propagation prediction, the state propagation model sets the state of the source node to the true group mastery probability, initializes the state of the non-source nodes to the prior estimate of the group mastery probability before simulated propagation, and updates the state of the non-source nodes by predicting the sum of the influence of the source node on the non-source nodes:

[0036] ;

[0037] in, It is a source node to a non-source node The sum of the influence quantities This is the initialization state of a non-source node. The state after the update of the non-source node is used as the predicted probability of the group mastering it.

[0038] In a second aspect, the present invention provides an electronic device, including a memory and a processor;

[0039] The memory is used to store computer programs;

[0040] The processor is configured to implement the above-described method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation when executing the computer program.

[0041] Thirdly, the present invention proposes a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the above-described method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation.

[0042] The beneficial effects of this invention are:

[0043] This invention proposes a dynamic construction and optimization method for knowledge graphs based on group behavior and state propagation. This method first integrates explicit logic from course materials with implicit relationships from historical group behavior data to construct a richer, enhanced knowledge graph as the optimization foundation. Furthermore, a state propagation model is introduced. By simulating the diffusion effect of knowledge mastery states in the knowledge graph network, the predicted results are compared with the actual group mastery probability, thus transforming the weight optimization problem of knowledge graph relationships into a quantifiable mathematical model solution process. This mechanism transforms the knowledge graph from a static database into a model whose internal parameters can be continuously calibrated based on real learning effects, fundamentally overcoming the shortcomings of traditional expert-constructed methods (subjective rigidity) and automated mining methods (lack of validation).

[0044] The optimized knowledge graph contains authentic feedback on the learning patterns of the group in the weight of each edge, enabling a more accurate depiction of the inherent logic and influence strength between knowledge points. This not only provides a more reliable data foundation for personalized learning path planning, adaptive testing, and precise knowledge gap diagnosis, but also allows the knowledge graph itself to adaptively evolve with changes in teaching content and learning groups, significantly improving the accuracy and intelligence of adaptive learning systems based on this graph. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation proposed in this invention.

[0046] Figure 2 This is a flowchart illustrating the process of constructing an initial knowledge graph;

[0047] Figure 3 This is a flowchart illustrating the process of uncovering implicit relationships between knowledge points based on statistical indicators.

[0048] Figure 4 This is a flowchart illustrating the process of uncovering implicit relationships between knowledge points based on confidence levels.

[0049] Figure 5 This is a flowchart illustrating the iterative optimization of a knowledge graph based on a state propagation model.

[0050] Figure 6 This is a schematic diagram of the electronic device terminal structure of the present invention. Detailed Implementation

[0051] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0052] The accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0053] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0054] like Figure 1 As shown, the present invention proposes a method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation, which mainly includes the following steps.

[0055] S1: Extract a set of knowledge points based on course materials, establish the relationships between knowledge points, and form an initial knowledge graph.

[0056] The goal of this step is to establish a structured prototype knowledge graph containing basic attributes and relationships. This knowledge graph is based on prerequisite relationships and explicit related relationships. Its implementation process can be broken down into the following three sub-steps, as follows: Figure 2 As shown:

[0057] S11, Knowledge Point Extraction and Attribute Definition.

[0058] The original course materials, including textbook catalogs, syllabi, and curriculum standards, are input, and natural language processing technology is used to perform structured analysis of the course content. In this structured analysis, a hierarchical knowledge structure of course-chapter-knowledge point is established, breaking down the semester course into the smallest learning units (such as knowledge points), and identifying the core concepts, theorems, and skills within each knowledge point.

[0059] In this embodiment, let course C include n knowledge points: C={ k1, k2,..., k n}; Each knowledge point k i There is a corresponding attribute set: attr(k) i The set is defined as follows: k = {Difficulty, Importance, Learning Time, Prerequisites, ...}. For example, this set should include at least the following attributes: difficulty, importance, learning time, and a list of prerequisites (extracted from the syllabus). For instance, the "Functions" chapter in a math course can be broken down into: k1 = Definition and Representation of Functions, k2 = Properties of Functions (Monotonicity, Parity), k3 = Linear and Quadratic Functions, k4 = Analysis of Function Graphs. Each knowledge point is assigned an attribute, such as "Each knowledge point requires approximately 45 minutes of learning time."

[0060] S12, Construct the relationships between knowledge points and the strength of those relationships.

[0061] In S11, a set of knowledge points was obtained, each with basic attributes. The goal of S12 is to establish the relationships between these knowledge points, including prerequisite relationships and explicit correlations. First, a relationship function f: K×K→R is defined. This function calculates the relationship strength based on the attributes of the knowledge points or external information. The establishment of these relationships combines multiple methods:

[0062] (1) For prior relationships, first infer the association using the attributes of knowledge points: for each knowledge point k i The prerequisite knowledge list is analyzed and matched with knowledge points in set K to establish a directed edge k1→k2 (indicating that k1 is a prerequisite knowledge of k2). The association strength of the prerequisite relationship can be initialized based on the static weight of the teaching logic or the dynamic weight based on the group learning data. This is an initial value that will be optimized in subsequent steps.

[0063] In one specific embodiment of this invention, static weights based on teaching logic refer to inviting subject teaching experts to score each pair of prerequisite relationships, or assigning values ​​according to the intensity of the syllabus's description. For example, "must master" is assigned a value of 1.0, and "recommended to master" is assigned a value of 0.7, etc. Dynamic weights based on group learning data refer to collecting a large amount of historical students' learning sequence data and corresponding answer accuracy rates, defining the weight of k1→k2 as the degree of significant difference in the average accuracy rate of two groups of students (students who learned k1 and then k2 versus students who learned k2 directly without learning k1) on knowledge point k2. The weights obtained by this method are dynamic and can be updated as more data is added, more objectively reflecting the actual teaching effect.

[0064] For example, knowledge point k1 is "definition of function" and k2 is "properties of function". The syllabus states that k1 is a prerequisite for k2, therefore a prerequisite relation edge k1→k2 is established. Meanwhile, subject teaching experts define the association strength as 0.85.

[0065] (2) For explicit correlations, content similarity analysis is used. Content similarity analysis refers to calculating the text content similarity between knowledge points.

[0066] In one specific embodiment of the present invention, the text description based on knowledge points is vectorized using TF-IDF and then the cosine similarity is calculated. For example, the text description similarity of k1 "function definition" and k2 "function property" is calculated. If the result is 0.85, which exceeds the threshold (e.g., 0.5), it is considered that there is a sexual correlation between the two. The calculated cosine similarity value is used as the initial weight of this explicit correlation.

[0067] In addition, those skilled in the art can employ other known methods.

[0068] S13, Construct the initial knowledge graph.

[0069] Each knowledge point in S12 is treated as a graph node, and the relationships are treated as edges of the graph. Prerequisite relationships are directed edges, and related relationships are undirected edges. The relationship strength is used as the weight of the edge to construct a graph model G = (V, E, W).

[0070] In this embodiment, after standardizing the relation matrix, a relation threshold θ (e.g., 0.3) is set, then the edge set E of the graph model is E = {(i, j) | w ij > θ}, weight set W = {w ij | (i, j) ∈ E}. It should be noted that directed edges have higher priority than undirected edges. First, determine the directed edges and their association strength, and then confirm the undirected edges. If a node of an undirected edge coincides with a node of a directed edge, please delete the undirected edge.

[0071] The knowledge graph is mapped to a two-dimensional grid, which is constructed based on the logical relationships between courses, with each grid representing a learning course. During mapping, each node is assigned to the nearest grid cell. A greedy algorithm can be used to sort the nodes in the force-directed layout according to their importance (such as degree), and then place the nodes in the grid cells that are closest to their consecutive positions and are free, until a two-dimensional grid graph with the knowledge graph result is obtained. Connected nodes in the graph are as adjacent as possible in the grid.

[0072] S14, mark the core knowledge points.

[0073] Graph algorithms are used to calculate node centrality metrics to identify core knowledge points. These metrics provide a basis for subsequent optimization, including:

[0074] Degree centrality: measures the number of direct connections between nodes. A node with high degree centrality means that it is directly related to many other knowledge points, and can be identified as a basic concept or core knowledge point. The formula is as follows:

[0075]

[0076] in, It is node v i The degree of , where n is the number of nodes. It is degree centrality.

[0077] Betweenness centrality measures the degree to which a node acts as a bridge, i.e., how many shortest paths pass through that node. Knowledge points with high betweenness centrality play a crucial role in connecting different modules within a knowledge network. Their absence can affect the network's connectivity. It is used to identify key transit knowledge points connecting different knowledge points. The formula is as follows:

[0078]

[0079] in, It is betweenness centrality. It is a path from node s to t that passes through node v. i The number of shortest paths, It is the number of shortest paths from node s to t.

[0080] Proximity centrality: measures the reciprocal of the average distance from a node to other nodes. A knowledge point with high proximity centrality means that its average distance to other knowledge points is short and it is easy to reach. It is used to identify central knowledge points in a knowledge network. The formula is as follows:

[0081]

[0082] in, It is close to centrality. It is node vi to node v j The distance.

[0083] If isolated nodes are found (degree centrality = 0), the knowledge points need to be refined or the content boundaries adjusted. If the relation density is found to be too high (average degree > 10), the relation threshold needs to be increased or related knowledge points need to be merged. For example, if k4 (function graph analysis) has a degree centrality of 0, then k4 in step S1 needs to be subdivided into two knowledge points: "recognition and drawing of function graphs" and "property analysis and application of function graphs".

[0084] S2, based on group behavior data, mines implicit relationships to optimize the initial graph.

[0085] This step aims to overcome the limitations of relying on prior knowledge by discovering deep, undefined but real-world connections between knowledge points from actual group learning behaviors, thereby enriching and optimizing the initial knowledge graph.

[0086] This invention uses group behavior co-occurrence analysis to uncover implicit relationships. Group behavior co-occurrence analysis refers to analyzing massive amounts of students' learning behavior data. If two knowledge points appear frequently in the same learning session (such as a review session), or if students frequently make the same mistakes on two knowledge points when doing exercises, it is considered that there is an implicit relationship. The association index based on student learning data is introduced into the association strength.

[0087] In one specific embodiment of the present invention, the process of obtaining supplementary implicit relationships includes two methods:

[0088] like Figure 3 As shown, Method 1: Discovering implicit relationships based on statistical indicators

[0089] If two knowledge points are frequently learned or practiced together in the same learning context (such as a single learning session), they may be related. This can be assessed using metrics such as co-occurrence frequency, grip relevance, learning sequence proximity, and error association probability.

[0090] For example, regarding co-occurrence frequency, we can statistically analyze the co-occurrence frequency of knowledge point k across all learning sessions for all students over a given period of time. i and k j The proportion of co-occurrences in the total number of sessions indicates the likelihood of a hidden relationship between two knowledge points if their frequency of occurrence in the same historical learning session (co-occurrence frequency) exceeds a threshold.

[0091] For mastery relevance: calculate the mastery score for each student for each knowledge point; for example, the average correct score on questions related to that knowledge point is considered the mastery score; for knowledge point k... i and k jEach student is provided with a pair of scores. The Pearson correlation coefficient is calculated for the score sequences of all students on these two knowledge points. If the correlation coefficient is positive and significant, the likelihood of a hidden relationship between the two is considered to be greater.

[0092] Learning sequence proximity refers to the degree of temporal proximity in which two knowledge points are learned by a history student during the learning process. This is calculated by analyzing students' learning paths and determining the learned knowledge point k. i and k j The time interval between students is calculated by averaging or medianing the time intervals for all students and then performing inverse normalization. The shorter the interval, the higher the proximity. If the interval exceeds a threshold, it is considered that there is a greater possibility of a hidden relationship between the two students.

[0093] For the probability of incorrect association, it is calculated when k is involved in the incorrect answer. i Under the given conditions, answering incorrectly also involves k. j If the probability of a given question is higher than a threshold, then the likelihood of a latent relationship between the two questions is considered to be greater.

[0094] When the value of any of the above indicators or the comprehensive indicator exceeds the preset threshold, a new indicator between knowledge point k is added. i and k j The candidate implicit relationships between them are undirected edges, and the association strength is directly obtained by normalizing these statistical index values. If knowledge point k... i and k j If a directed edge already exists between them, then no update is needed; if knowledge point k... i and k j If an undirected edge already exists, the association strength of the undirected edge is updated based on the weighted value of the association strength of the existing undirected edge and the association strength of the candidate implicit relation.

[0095] like Figure 4 As shown, Method 2: Latent Relationship Discovery Based on Confidence Model

[0096] The goal is to obtain students' work records, notes, and answer explanations from their tablets, grouped by class, grade, or academic year. The aim is to treat the initial knowledge graph as a prior model and use students' work records, notes, and other behavioral data as observational evidence. Through a series of calculations and analyses, the relationships in the initial knowledge graph can be updated, especially to discover and verify those implicit, undefined dependencies.

[0097] Record test data based on a single learning session (such as completing a question or reading a chapter). Notes data The solution involves analyzing the questions and answers, where q is the question ID, ans is the student's answer, correct∈{0,1} indicates whether the answer is correct, and t is the time interval. This is a list of explicit knowledge points marked in the question, Text is the note text, and Loc is the location of the learning area, used to help identify the knowledge points mentioned in the note text. It's about contextual knowledge points. Continuously monitor students' learning behavior, and once it occurs... or The event then proceeds to the implicit knowledge point identification step, as follows:

[0098] (1) Perform NLP analysis on the answer analysis of question q in the test data and extract the set of all knowledge points mentioned therein. By removing the explicit knowledge points marked in the question from the set of all knowledge points, a candidate set of hidden knowledge points is obtained. (That is, the knowledge points mentioned in the answer explanation but not directly marked in the question).

[0099] for Each knowledge point k in c and Each knowledge point k in e Check students' understanding of knowledge point k c The probability of mastery is calculated. If the probability of mastery is lower than the mastery threshold, the student is considered not to have mastered the knowledge point. Based on the combination of the student's correct answer and the probability of mastery, the effect of this action on the candidate relation k is calculated. c →k e Strength of evidence S new For example: if a student answers incorrectly (correct=0), and it is detected that the student has not mastered k. c This is strong evidence that k c It could be k e The prerequisite is to set up. (For example =0.5); If the student answers correctly (correct=1), but it is detected that the student has not mastered k c This is weak evidence because students could answer correctly through guessing or other means. (For example, If the student answers correctly and understands k c This supports the association, but the strength of the evidence is moderate. (For example, ).

[0100] Use confidence level to update rules: , where r is the candidate relation k c →k e , It represents the confidence level before the update.

[0101] Taking a specific set of data as an example, let's say the question q is known. The answer analysis extracts ,so And the student answered the question incorrectly (correct=0), check the student's understanding of k. c The probability of mastering: P(k) c =0.2 (Not mastered). This is because the student has not mastered k. c Furthermore, the incorrect answer is strong evidence, so Snew=0.5 is set for the candidate relation c→Ab; if this is the first time the relation has been discovered, then Confold=0, so Conf new =0.9×0+0.1×0.5=0.05 (assuming λ=0.9); Another student subsequently answered incorrectly and did not understand k. c , then Conf new =0.9×0.05+0.1×0.5=0.095, the confidence level gradually increases; if there is more similar evidence in the future, the confidence level will gradually increase.

[0102] (2) Discovering correlations from note data

[0103] Perform NLP analysis on the notes text to identify all the common knowledge points mentioned in them. .for Each pair of different knowledge points (k) i ,k j Propose a candidate undirected correlation k. i -k j Calculate the contribution of this note to this relationship: if the note context k context ∈{k i ,k j}, set base=1; otherwise base=0.5. Also consider the student's mastery level; if a student hasn't mastered one knowledge point but recorded another, it might indicate they are related); calculate the contribution: Contribution(i,j)=base×(1-min(P(k)). i ),P(k j ))), where base is the base contribution, if the note context is k i or k j If so, then the base is higher.

[0104] Use confidence level to update rules: .

[0105] Taking a specific set of data as an example, suppose a student takes notes while learning knowledge point Ab, and extracts K from the notes. note ={Ab,c,d}, candidate undirected relation pairs include (Ab,c), (Ab,d), and (c,d). This is because the notes record contextual knowledge points. ,so , , Update the confidence levels of each relation.

[0106] (3) Confidence accumulation and verification of candidate relationships

[0107] Establish a candidate relation library C: Store all candidate relations (including directed implicit precondition relations and undirected implicit correlation relations), and record their current confidence Conf(r) and number of support counts Count(r).

[0108] If Conf(r) exceeds the first confidence threshold (e.g., 0.7) and Count(r) exceeds the minimum occurrence threshold, then the relation is considered to have been sufficiently verified, relation r is removed from C, and formally added to the knowledge graph: if it is a prerequisite relation, then set the corresponding directed edge with an association strength of 1; if it is a related relation, then set the corresponding undirected edge with an association strength of Conf(r).

[0109] If Conf(r) is below the second confidence threshold (also known as the forgetting threshold, such as 0.1) or Count(r) has not been updated for a long time, then relation r is directly removed from C.

[0110] By listening or The event continuously generates candidate relations and updates their confidence levels, periodically (e.g., daily) scanning the candidate database for verification. For example, if a candidate relation c→Ab accumulates evidence multiple times, achieving Conf(r)=0.75 and Count(r)=6, it is promoted to a formal prerequisite relation, updating the knowledge graph. It should be noted that this updating also applies to the setting where directed edges have higher priority than undirected edges.

[0111] S3 defines the probability of group mastery and iteratively optimizes the knowledge graph based on the state propagation model.

[0112] The probability of group mastery aims to establish a quantifiable benchmark of mastery based on historical data for each knowledge point. Its input is historical user group learning behavior data, and it calculates the mastery probability for that knowledge point k. i The total number of correct answers to related questions accounts for a percentage of knowledge point k. i The total number of times related questions are answered is taken as the knowledge point k. i The probability of group mastery of a knowledge point is calculated, and the probability of group mastery of all knowledge points constitutes a vector. This step treats the change in the "probability of group mastery" of a knowledge point as an influence that propagates in the knowledge network. By simulating this propagation and comparing the predicted results with the actual data, the accuracy of the relationships in the knowledge graph can be inferred in reverse, thereby optimizing it.

[0113] like Figure 5 As shown, in a specific embodiment of the present invention, the preferred implementation process is as follows:

[0114] S31, obtain the group mastery probability vector and the enhanced knowledge graph, which includes all nodes, edges (including pre-relationships, explicit correlations and implicit relations) and their current weights w(u, v).

[0115] S32 randomly divides all knowledge point nodes in the knowledge graph into two mutually exclusive sets, defined as the source node set and the verification node set. The true group mastery probability of the source nodes is used as the input signal of the state propagation model, and the true group mastery probability of the verification nodes is used to verify the prediction accuracy of the model.

[0116] S33, Construct a prediction model based on state propagation.

[0117] This model takes the group mastery probability of the source node set and the current augmented knowledge graph as input, treats the weights of all edges in the knowledge graph as learnable parameters, and fixes other parameters in the knowledge graph. Essentially, this is a graph-based semi-supervised learning model.

[0118] Define a propagation function that allows the true population probability of the source node to propagate along the network edges based on the current weight w(u, v) of the graph. The propagation function is defined as follows:

[0119] If the verification node is v, a direct neighbor of the source node u, calculate the direct impact of the change in u's mastery over v:

[0120]

[0121]

[0122] in, It is the probability of the group mastering the source knowledge point u. The current weight determines the proportion that influences the spread; It is based on the betweenness centrality of node v. and proximity centrality The adjustment factor is used to amplify the impact on important knowledge points. For coefficients, It is the direct influence of changes in u's mastery of v.

[0123] If the verification node transmits its influence to knowledge point v through intermediate node w, then first calculate the predicted direct influence of knowledge point w under the direct influence of knowledge point u using the same method. Then, this influence is passed to knowledge point v, and the indirect influence of knowledge point v is calculated:

[0124]

[0125] in, It is the propagation attenuation coefficient (e.g., 0.6), which simulates the effect of the influence decreasing with increasing distance.

[0126] Calculate the sum of the direct and indirect influences of all verification nodes on all source nodes. ,according to Update the state of the non-source node v:

[0127]

[0128] in, It is a priori estimate of the probability of mastery by the verification node group before the simulation propagation. It represents the expected mastery level of the node under the assumption that it is not affected by any source node in this simulation. It is based on the macroscopic statistical law of the knowledge point of a large number of students and can be calculated using existing technology.

[0129] In summary, the state propagation function of the prediction model is as follows:

[0130]

[0131] in, It is the propagation function. It is a learnable edge weight. This is the current knowledge graph structure. It is the group mastery probability vector of the source node. It is the sum of the influences of the source node u on the verification node v.

[0132] S34, verify the predicted mastery probability of all knowledge points in the node set (V). The corresponding real group possesses the probability The mean squared error (MSE) is calculated as a loss function through comparison.

[0133] In an accurate knowledge graph, the results of the simulated propagation should be highly consistent with the actual, observed group mastery probability. Therefore, the mean squared error between the two is defined as a loss function to measure the "inaccuracy" of the current graph parameters. In this embodiment, gradient descent or a similar optimization algorithm is used to calculate the gradient of the loss function Loss with respect to the weight w(u, v) of each edge. This gradient indicates the direction and magnitude in which w(u, v) should be adjusted to reduce prediction error.

[0134] Those skilled in the art can also, based on considerations of computational complexity, select only a portion of the source nodes and verification nodes and fine-tune the local weight values.

[0135] S35. After each iteration, a new knowledge graph is generated using the updated weights, and then state propagation prediction is performed to calculate the new loss until the loss converges to the minimum or is below a predetermined threshold.

[0136] S4 outputs a dynamic knowledge graph with globally optimized weights that most accurately simulates the propagation of mastery influence among knowledge points.

[0137] The knowledge graph constructed and optimized by this invention is no longer a static database, but a graph model whose parameters (edge ​​weights) can be continuously learned and improved through data. The weight w(u, v) of each edge has been calibrated and optimized with massive amounts of group learning data, enabling it to most accurately simulate the dynamic influence and propagation law of mastery between knowledge points. This optimized graph lays a solid foundation for subsequent personalized applications.

[0138] It should be noted that the knowledge graph dynamic construction and optimization method based on group behavior and state propagation in the above embodiments can essentially be executed by a computer program. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the method provided in the above embodiments, which includes a memory and a processor;

[0139] The memory is used to store computer programs;

[0140] The processor is configured to implement the knowledge graph dynamic construction and optimization method based on group behavior and state propagation in the above embodiments when executing the computer program.

[0141] From a hardware perspective, such as Figure 6 The diagram shown is a hardware structure diagram provided in this embodiment. In addition to the processor, memory, network interface and non-volatile memory shown in the diagram, any device with data processing capabilities in the embodiment may also include other hardware depending on the actual function of the device with data processing capabilities, which will not be described in detail here.

[0142] When the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium.

[0143] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the method provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize the knowledge graph dynamic construction and optimization method based on group behavior and state propagation in the above embodiments.

[0144] It is understood that the computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0145] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation, characterized in that, Includes the following steps: Step 1: Extract a set of knowledge points based on course materials, establish the relationships between knowledge points, and generate an initial knowledge graph; The association relationships include at least prerequisite relationships and explicit correlation relationships; Step 2: Obtain learning behavior data generated by user groups during the learning process, mine implicit relationships between knowledge points based on statistical indicators and confidence levels, and supplement the implicit relationships into the initial knowledge graph to generate an enhanced knowledge graph; Step 3, with the goal of optimizing the accuracy of the enhanced knowledge graph, performs an iterative optimization process, which includes: Obtain the actual probability of the group mastering each knowledge point; Using some knowledge points as source nodes, and based on the current network structure and relational weights of the enhanced knowledge graph, the true group mastery probability of the source nodes is propagated in the knowledge graph through a state propagation model in order to predict the group mastery probability of all or some of the other non-source nodes. Calculate the difference between the predicted probability of group mastery and the corresponding actual probability of group mastery; Adjust the weights of the relationships in the enhanced knowledge graph based on the degree of difference; Step 4: Output the final optimized knowledge graph.

2. The method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation according to claim 1, characterized in that, In step S2, the process of mining implicit relationships between knowledge points based on statistical indicators includes: Analyze the learning behavior data and calculate at least one statistical indicator among the following: co-occurrence frequency, mastery correlation, learning sequence proximity, and error association probability among knowledge points; When the value of the statistical indicator exceeds the preset indicator threshold, a candidate implicit relationship is established between the corresponding knowledge points, and the value of the statistical indicator is normalized and used as the initial weight of the candidate implicit relationship. If the candidate latent relation already has a directed edge in the initial knowledge graph, it is not updated; if the candidate latent relation already has an undirected edge in the initial knowledge graph, the association strength of the undirected edge is updated according to the weighted value of the association strength of the existing undirected edge and the initial weight of the candidate latent relation.

3. The method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation according to claim 1, characterized in that, In step S2, the process of mining implicit relationships between knowledge points based on confidence levels includes: Listen for user learning behavior events, including question answering events and note-taking events; Extract explicit knowledge point sets and implicit knowledge point sets from the learning behavior events; Based on the user's mastery of the implicit knowledge points and the results of the learning behavior events, the strength of evidence for the association between the explicit and implicit knowledge points is calculated. A candidate relation database is established to store all candidate associations. The current confidence level and the number of times evidence supports a candidate association are updated in real time based on the strength of evidence. If the current confidence level exceeds the confidence level threshold and the number of times evidence supports a candidate association exceeds the minimum number of times threshold, the candidate association is removed from the candidate relation database and added to the enhanced knowledge graph as a latent relation. If the current confidence level is lower than the forgetting threshold or the number of times evidence supports a candidate association has not been updated for a long time, the candidate association is directly removed from the candidate relation database.

4. The method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation according to claim 3, characterized in that, The calculation process for the strength of evidence includes: For question-and-answer events, implicit knowledge points are extracted from the question and answer analysis. Candidate associations are established between these implicit knowledge points and the explicit knowledge points in the question. When a user answers incorrectly and does not know the implicit knowledge points, the evidence is assigned the first strength of evidence. When a user answers correctly and does not know the implicit knowledge points, the evidence is assigned the second strength of evidence. When a user answers correctly and knows the implicit knowledge points, the evidence is assigned the third strength of evidence. In other cases, the evidence strength is 0, and the strength of evidence decreases in the following order: first, second, and third. For note events, extract commonly mentioned knowledge point pairs from the note text. If the context of the note contains the knowledge point pair, the correlation between the context of the note and the knowledge point pair is base=1; otherwise, base=0.

5. Obtain the lower value of the probability that the user has mastered each knowledge point in the knowledge point pair, take the difference between 1 and the lower value, and multiply it by the base value to get the evidence strength of the candidate correlation relationship corresponding to the knowledge point pair. The new confidence level is a weighted average of the strength of evidence and the current confidence level recorded in the candidate relation database.

5. The method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation according to claim 3, characterized in that, The process of removing candidate relationships from the candidate relationship database and adding them as implicit relationships to the enhanced knowledge graph specifically involves: If the candidate association is a pre-revision relationship, then add a new directed edge with an association strength of 1; if the directed edge already exists, then do not update. If the candidate association is a correlation, a new undirected edge is added with the association strength being the current confidence level. If the undirected edge already exists, the association strength of the undirected edge is updated based on the weighted value of the association strength of the existing undirected edge and the current confidence level of the candidate association.

6. The method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation according to claim 1, characterized in that, In a knowledge graph, the edges corresponding to the relationships include both directed and undirected edges, and the priority of directed edges is always higher than that of undirected edges.

7. The method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation according to claim 1, characterized in that, The propagation process of the state propagation model satisfies: The influence of knowledge point u on knowledge point v is positively correlated with the association weight of the connection between u and v, decays as the shortest path length from u to v increases, and is positively correlated with the centrality index of knowledge point v in the network, wherein the centrality index includes at least one of betweenness centrality and proximity centrality.

8. The method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation according to claim 7, characterized in that, In propagation prediction, the state propagation model sets the state of the source node to the true group mastery probability, initializes the state of the non-source nodes to the prior estimate of the group mastery probability before simulated propagation, and then predicts the sum of the influence of the source node on the non-source nodes, updating the state of the non-source nodes accordingly. ； in, It is a source node to a non-source node The sum of the influence quantities This is the initialization state of a non-source node. The state after the update of the non-source node is used as the predicted probability of the group mastering it.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation as described in any one of claims 1 to 8.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the method for dynamic construction and optimization of knowledge graphs based on group behavior and state propagation as described in any one of claims 1 to 8.

Images