Adaptive learning path generation method of AI learning machine

Abstract

The invention relates to the technical field of machine learning, in particular to an adaptive learning path generation method of an AI learning machine. According to the method, a subject knowledge graph containing topological logic is constructed, a knowledge point linear index sequence is combined, a continuous error mode in a user answering behavior is captured so as to lock a knowledge mastering fault interval, and a root preposition node causing an error is accurately positioned by utilizing a graph backtracking mechanism; performing time decay weighting operation on historical interaction data in combination with a memory forgetting curve, constructing a closed-loop repair path from a root weak point to a target knowledge point, and meanwhile, introducing an examination period urgency degree and a basic gating evaluation index to dynamically adjust a recommendation priority; the adaptive optimization of the learning path is realized on the premise of conforming to the cognitive law, the defects that the basic weak problem is difficult to radically treat and the time dimension evaluation is lacked in the traditional mode are effectively solved, and the examination preparation efficiency and the knowledge internalization stability are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to an adaptive learning path generation method for an AI learning machine. Background Technology

[0002] The field of machine learning technology involves technologies that enable computer systems to extract patterns from data and make judgments and decisions based on these patterns by constructing mathematical models and computational processes. These include the collection and labeling of sample data, the construction of feature representation methods, the organization of model training processes, the setting of parameter update rules, and the operation mechanism for prediction or classification based on historical data. This technology is widely used in educational informatization scenarios to analyze learning behavior data and form personalized processing procedures.

[0003] Among them, the adaptive learning path generation method of AI learning machine refers to the process in which the AI ​​learning machine summarizes the learning records such as the accuracy rate of the learner's answer, the time spent answering questions, and the distribution of incorrect question types according to the pre-set learning content structure and learning sequence rules, and then selects the next learning unit or review content according to a fixed threshold or predetermined judgment conditions. Usually, the path arrangement for the learning matter is completed by arranging knowledge points according to difficulty level, judging the mastery level based on the results of one or more tests, calling the corresponding knowledge point list in the local machine or server, and generating learning path data including learning content number, presentation order and repetition number in sequence.

[0004] Existing technologies select learning content based solely on pre-set fixed thresholds or predetermined judgment conditions, simply arranging knowledge points according to difficulty levels and determining mastery based on test results. This approach severs the deep topological logical dependencies between knowledge points, making it impossible to accurately pinpoint the chain reaction caused by weak prior knowledge when faced with repeated learning errors by users. Furthermore, the aggregation and processing of answer data lacks consideration of the time dimension of memory forgetting patterns, resulting in a biased assessment of the user's true mastery. Consequently, the generated learning paths are often limited to mechanical repetitive practice of the current incorrect question types, making it difficult to build a closed-loop remedial path that conforms to cognitive logic from the root. This leads to a lack of targeted delivery of learning resources and ineffective waste of preparation time. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an adaptive learning path generation method for AI learning machines.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive learning path generation method for an AI learning machine, comprising the following steps: S1: Obtain the chapter directory and the sequence of knowledge points of the specified subject to construct a subject knowledge graph, mark key knowledge point nodes, and establish a linear index sequence of knowledge points; S2: Respond to the user's login command and display test questions covering key knowledge points on the screen. Determine the initial user's mastery level based on the current and historical question-answering interaction records. S3: Monitor the user's incorrect answer marks for incorrect test questions during continuous learning, extract the corresponding index of the incorrect test questions from the linear index sequence of knowledge points, and determine the user's knowledge gap interval based on the index position; S4: Based on the subject knowledge graph, identify the first knowledge point node that failed to answer the test questions within the knowledge mastery gap interval, collect the second knowledge point node associated with the first knowledge point node, and construct a closed-loop learning path; S5: Based on the closed-loop learning path and the initial user mastery status, calculate the recommendation priority score for each knowledge point node, dynamically refresh the learning path based on the recommendation priority score, and obtain the adaptive learning path generation result.

[0007] The present invention is improved in that the linear index sequence of knowledge points includes a unique sequence position index value assigned to each knowledge point node according to the topological connection order of the subject knowledge graph. The initial user mastery status is specifically obtained by performing a weighted evaluation operation on the user's answer correctness status, answering time and historical answering interaction records for the test questions. The knowledge mastery gap includes the sequence position index value of the first occurrence of the answer error mark extracted from the linear index sequence of knowledge points, and the sequence position index value of the sequence where the consecutive answer error marks end. The closed-loop learning path includes the first knowledge point node that failed to answer the test questions within the knowledge mastery gap and the second knowledge point node that is back-linked through the subject knowledge graph. The adaptive learning path generation result includes a sequence list after dynamically sorting the candidate knowledge point nodes according to the recommendation priority score, and the subsequently linked knowledge point nodes in the subject knowledge graph.

[0008] The present invention is improved in that step S1 is specifically as follows: S101: Obtain the subject chapter directory and knowledge point sequence from the cloud storage, parse the subject chapter directory, extract all knowledge point nodes, define directed edges representing the sequence relationship based on the sequence of knowledge point nodes, and combine knowledge point nodes and directed edges to construct a subject knowledge graph. S102: Based on the subject knowledge graph, configure node attribute data for each knowledge point node, including difficulty coefficient value, examination frequency value, associated question bank and teaching video resources, and mark key knowledge point nodes with reference to the examination frequency value; S103: Assign a unique sequence position index value to each key knowledge point node in the subject knowledge graph according to the topological connection order of the directed edges, arrange all the assigned sequence position index values, and establish a linear index sequence of knowledge points.

[0009] The present invention is improved in that step S2 is specifically as follows: S201: Respond to the user's login command and display test questions covering key knowledge points on the screen; collect the user's answer status and answering time for the test questions; and generate the current answering interaction data. S202: Detect whether there are historical answer interaction records that include the historical answer correctness status and the historical answer duration. If they exist, merge them with the current answer interaction data that includes the current answer correctness status and the current answer duration to form a dataset that includes the historical time series. If they do not exist, directly create a dataset that does not include the historical time series based on the current answer interaction data as the interaction dataset to be evaluated. S203: Identify whether the interactive dataset to be evaluated includes historical time series. If it does, extract the historical correct / incorrect answer status and historical answer duration and match time decay weights. At the same time, perform a comprehensive weighted calculation on the current interactive data by matching real-time response weights. If it does not include historical time series, directly apply standard evaluation weights to the current interactive data to perform a single evaluation calculation and calculate the user's mastery of the knowledge point nodes as the initial user mastery status.

[0010] The present invention is improved in that step S3 is specifically as follows: S301: Monitor the error marks generated by the user's answers to test questions during continuous learning, identify the knowledge point nodes that generated the errors based on the mapping relationship between test questions and knowledge points, extract the sequence position index value corresponding to the knowledge point node from the linear index sequence of knowledge points, and detect the failure status of adjacent nodes of the preceding and following adjacent knowledge point nodes within a preset time window. S302: Summarize the sequence position index values ​​corresponding to multiple answer error markers, compare whether the index values ​​show a monotonically increasing or monotonically decreasing numerical arrangement feature, and whether the adjacent knowledge point nodes in the adjacent node failure state show a coherent numerical arrangement feature, to obtain the order consistency judgment result. S303: When the number of incorrect answer markers exceeds the preset minimum consecutive number threshold and meets the order consistency judgment result, the knowledge point node from the first failure to the knowledge point node where the consecutive failures terminate is delineated, and a knowledge mastery gap interval is generated.

[0011] The present invention is improved in that step S4 is specifically as follows: S401: Based on the subject knowledge graph, identify the first knowledge point node that failed to answer the test question within the knowledge mastery gap interval, trace back the second knowledge point node associated with the first knowledge point node, and obtain the basic associated node set; S402: Referring to the initial user mastery status, extract the mastery value of each knowledge point node in the basic associated node set, filter all knowledge point nodes whose mastery values ​​are lower than the preset basic standard, perform an ascending sorting operation on the filtered knowledge point nodes according to the mastery values, and select the first node in the sorting, i.e. the node with the lowest mastery, as the starting node of the path that needs to be solved first. S403: Call the subject knowledge graph, take the starting node of the path as the starting point, and the first knowledge point node in the knowledge mastery gap interval that failed to be answered as the ending point, retrieve all intermediate transitional knowledge point nodes connected by directed edges between the two, integrate the starting point, intermediate nodes and the ending point into a candidate knowledge point node set, and connect the candidate knowledge point node set to construct a closed-loop learning path.

[0012] The present invention is improved in that step S5 is specifically as follows: S501: Obtain the examination frequency value of the candidate knowledge point node in the closed-loop learning path from the node attribute data, determine it as the importance weight, and extract the current mastery value of the candidate knowledge point node according to the initial user mastery status. Calculate the difference between the preset complete mastery benchmark value and the current mastery value, quantify the unmastery value of the candidate knowledge point node, and perform a product operation on the unmastery value using the importance weight to obtain the remedial needs evaluation index. S502: Obtain the average mastery of all directly related knowledge point nodes of the candidate knowledge point node, and call the Sigmoid function to process the difference between the average mastery and the new knowledge learning basic mastery threshold to obtain the basic gating evaluation index. S503: Collect the number of days remaining until the exam, set the corresponding standard reference time factor and exam urgency adjustment coefficient, calculate the urgency correction index, and calculate the recommendation priority score for each candidate knowledge point based on the remedial needs evaluation index, the basic gating evaluation index, and the urgency correction index. S504: Based on the recommended priority score, each candidate knowledge point is sorted in real time, the mastery value in the updated user mastery status is monitored in real time, compared with the preset breakthrough threshold, and the subsequent related knowledge point nodes are unlocked and re-scored according to the comparison result to obtain the adaptive learning path generation result.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, a subject knowledge graph containing topological logic is constructed and combined with a linear index sequence of knowledge points to capture continuous error patterns in users' answering behavior to pinpoint gaps in knowledge mastery. The graph backtracking mechanism is used to accurately locate the root cause nodes that lead to errors. By combining the memory forgetting curve with time decay weighted calculations on historical interaction data, a closed-loop repair path from the root weakness to the target knowledge point is constructed. At the same time, the urgency of the exam period and basic gating evaluation indicators are introduced to dynamically adjust the recommendation priority, ensuring that the learning path is adaptively optimized under the premise of conforming to cognitive laws. This effectively solves the shortcomings of traditional methods that are difficult to eradicate weak foundations and lack time dimension evaluation, and significantly improves exam preparation efficiency and the stability of knowledge internalization. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a detailed flowchart of step S1 of the present invention; Figure 3 This is a detailed flowchart of step S2 of the present invention; Figure 4 This is a detailed flowchart of step S3 of the present invention; Figure 5 This is a detailed flowchart of step S4 of the present invention; Figure 6 This is a detailed flowchart of step S5 of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1 This invention provides a technical solution: an adaptive learning path generation method for an AI learning machine, comprising the following steps: S1: Obtain the chapter directory and the sequence of knowledge points of the specified subject to construct a subject knowledge graph, mark key knowledge point nodes, and establish a linear index sequence of knowledge points; S2: Respond to the user's login command and display test questions covering key knowledge points on the screen. Determine the initial user's mastery level based on the current and historical question-answering interaction records. S3: Monitor the error marks generated by users for incorrect test questions during continuous learning, extract the corresponding index of the incorrect test questions from the linear index sequence of knowledge points, and determine the gaps in the user's knowledge mastery based on the index position; S4: Based on the subject knowledge graph, identify the first knowledge point node that failed to answer the test questions within the knowledge mastery gap, collect the second knowledge point node associated with the first knowledge point node, and construct a closed-loop learning path; S5: Based on the closed-loop learning path and the initial user mastery status, calculate the recommendation priority score for each knowledge point node, dynamically refresh the learning path based on the recommendation priority score, and obtain the adaptive learning path generation result.

[0017] The linear index sequence of knowledge points includes a unique sequence position index value assigned to each knowledge point node according to the topological connection order of the subject knowledge graph. The initial user mastery status is specifically obtained by performing a weighted evaluation operation on the user's answer status, answer time, and historical answer interaction records for the test questions. The knowledge mastery gap includes the sequence position index value of the first occurrence of the answer error mark extracted from the linear index sequence of knowledge points, and the sequence position index value of the sequence where the consecutive answer error marks end. The closed-loop learning path includes the first knowledge point node that failed to answer the test questions within the knowledge mastery gap and the second knowledge point node that is back-linked through the subject knowledge graph. The adaptive learning path generation result includes a sequence list after dynamically sorting the candidate knowledge point nodes according to the recommendation priority score, and the subsequent related knowledge point nodes in the subject knowledge graph.

[0018] Please see Figure 2 Step S1 is as follows: S101: Obtain the subject chapter directory and knowledge point sequence from the cloud storage, parse the subject chapter directory, extract all knowledge point nodes, define directed edges representing the sequence relationship based on the sequence of knowledge point nodes, and combine knowledge point nodes and directed edges to construct a subject knowledge graph. The AI ​​learning machine sends a data synchronization request to the cloud-based educational resource server via a pre-built encrypted communication protocol, retrieving the high school mathematics subject chapter directory file and its corresponding metadata structure stored in the database. When parsing the subject chapter directory, it uses a built-in machine learning-based text parsing engine to traverse the directory tree structure, identifying each indivisible smallest teaching unit through a natural language processing model, defining it as an independent knowledge point node, and using it as a vertex element of the subject knowledge graph. After extracting all knowledge point nodes, it reads the definition fields regarding the logical relationships between knowledge points from the metadata, identifying the prerequisite and prerequisite logic between the first and second knowledge point nodes (e.g., "a linear equation in one variable" is a prerequisite for "a system of two linear equations"). Based on this logic, it establishes directional connections, i.e., directed edges, between the first and second knowledge point nodes, serving as edge elements of the subject knowledge graph. By traversing all extracted knowledge point nodes and generated directed edges, a network-structured subject knowledge graph is generated using graph database construction tools. This graph is essentially a directed acyclic graph (DAG) composed of a set of nodes (representing specific knowledge content) and a set of directed edges (representing learning order and logical dependencies). For example, after parsing and obtaining two nodes, "function concept" and "function property," and following the logic of learning concepts before properties in the syllabus, directed edges are established from "function concept" to "function property," completing the basic architecture of the graph.

[0019] S102: Based on the subject knowledge graph, configure node attribute data for each knowledge point node, including difficulty coefficient value, examination frequency value, associated question bank and teaching video resources, and mark key knowledge point nodes with reference to the examination frequency value; Each knowledge point node is configured with node attribute data based on a subject knowledge graph. This data includes difficulty coefficient, frequency of examination, associated question banks, and teaching video resources. When configuring the difficulty coefficient, machine learning is applied to analyze the average correct answer rate of all test takers across the network over the past five years. The value obtained by subtracting the average correct answer rate from 1 is determined as the difficulty coefficient. For example, if the statistical data shows that the average correct answer rate for the knowledge point "quadratic function" is 0.4, then 1 - 0.4 = 0.6, and the difficulty coefficient is set to 0.6. When configuring the frequency of examination, the total number of times the knowledge point appeared in the most recent ten major standardized tests is counted. This number is then normalized by dividing it by the number of times the knowledge point appeared most frequently in those ten tests. For example, if "derivative" appeared 8 times and the highest frequency was 10 times, then 8 ÷ 10 = 0.8, and the frequency of examination is set to 0.8. In the process of marking key knowledge point nodes based on the frequency of examination, the arithmetic mean of the frequency of examination for all knowledge point nodes in the graph is calculated, and the standard deviation of all frequency of examination is added to obtain the threshold for determining key nodes. For example, if the calculated mean frequency of observation is 0.5 and the standard deviation is 0.2, then 0.5 + 0.2 = 0.7 is calculated as the threshold. The frequency of observation for each node is compared with 0.7. If the frequency of observation for a node is 0.8, since 0.8 > 0.7, that node is labeled as a key knowledge point.

[0020] S103: Assign a unique sequence position index value to each key knowledge point node in the subject knowledge graph according to the topological connection order of the directed edges, arrange all the assigned sequence position index values, and establish a linear index sequence of knowledge points. A topological sorting algorithm is employed, starting with the initial key node with an in-degree of 0 and sequentially visiting subsequent key nodes along the directions of the directed edges. For each key node visited, a counter is incremented, and the current counter value is used as the sequence position index of that key node. For example, the initial key node "Set" is assigned index value 1001, and its direct successor key node "Function Definition" is assigned index value 1002. After all key nodes are indexed, the corresponding key knowledge point nodes and their index values ​​are extracted and rearranged according to their sequence position index values ​​in ascending order, constructing a unidirectional linear data structure—a linear index sequence of knowledge points—to represent the linear progression of key knowledge points in the subject's logical structure.

[0021] Please see Figure 3 Step S2 is as follows: S201: Respond to the user's login command and display test questions covering key knowledge points on the screen; collect the user's answer status and answering time for the test questions; and generate the current answering interaction data. In response to the user's login command input on the AI ​​Learning Machine V8-09 terminal device, after verifying the user's identity, the "AI 1-on-1 Diagnosis" program is launched. This program selects the key knowledge point node to be learned from the linear index sequence of knowledge points, calls the associated question bank, and renders and presents test questions covering that key knowledge point node on the display screen. A background timer is started, beginning when the test questions are fully loaded and displayed, and stopping when the user submits an answer. The time difference between the two is recorded as the answer duration. The system receives the user's submitted answer options and compares them character-wise with the correct options in the standard answer database. If a match is found, a status code 1 indicating correctness is generated; otherwise, a status code 0 indicating error is generated. This status code represents the correctness of the answer. The correctness of the answer status and the answer duration are packaged to generate the current answer interaction data. For example, if a user takes 120 seconds to answer the question "Definition of the Derivative," and the answer is incorrect, a data packet is generated: Status Code 0, Duration 120 seconds.

[0022] S202: Detect whether there are historical answer interaction records that include the historical answer correctness status and the historical answer duration. If they exist, merge them with the current answer interaction data that includes the current answer correctness status and the current answer duration to form a dataset that includes the historical time series. If they do not exist, directly create a dataset that does not include the historical time series based on the current answer interaction data as the interaction dataset to be evaluated. The system queries the user's personal historical database, using the unique identifier of the current knowledge point node as the index key. If the search results return a non-empty set, indicating the existence of historical records, the system extracts multiple sets of historical answer statuses, their corresponding timestamps, and answer durations. These are then appended and merged with the current answer interaction data in chronological order, constructing a two-dimensional array containing time-series markers to form a dataset including historical time series data. If the search results are empty, indicating the user's first attempt at this knowledge point, the system directly stores the current answer interaction data as the unique element in the array, creating a dataset excluding historical time series data. The final generated set is defined as the interaction dataset to be evaluated.

[0023] S203: Identify whether the interactive dataset to be evaluated includes historical time series. If it does, extract the historical correct / incorrect answer status and historical answer duration and match the time decay weight. At the same time, perform a comprehensive weighted calculation on the current interactive data by matching the real-time response weight. If it does not include, directly apply the standard evaluation weight to the current interactive data to perform a single evaluation calculation and calculate the user's mastery of the knowledge point node as the initial user mastery status. The specific process for setting the time decay weight is as follows: Get the historical generation timestamps corresponding to the historical question-answering interaction records and the current timestamp when the current question-answering interaction data was generated; Calculate the difference between the current timestamp and the historically generated timestamps to obtain the time interval value; Read the preset attenuation coefficient; Calculate the product of the time interval value and the attenuation coefficient; Performs exponentiation with the natural constant as the base and the opposite of the product as the exponent; The numerical result obtained from the exponentiation operation is used as the time decay weight; Extract the historical answer status (correct / incorrect) and duration for each historical question, and perform a matching operation based on time decay weights. Obtain the server generation time (e.g., 1700000000) stored in the historical question interaction record as the historical generation timestamp, and simultaneously obtain the system time at the time the question was completed (e.g., 1700259200) as the current timestamp. Perform a subtraction operation to calculate 1700259200 - 1700000000 = 259200 seconds to obtain the time interval value. Read the preset decay coefficient, which is set based on fitting and training a machine learning model on a large amount of user memory forgetting curve data, selecting an optimized forgetting rate constant as the decay coefficient, for example, set to 0.000005. Calculate the product of the time interval value and the decay coefficient, i.e., 259200 × 0.000005 = 1.296. The mathematical operation library is invoked to perform a power operation with the natural constant e (approximately 2.71828) as the base and the negative of the product, -1.296, as the exponent. The result is e raised to the power of -1.296 ≈ 0.2736. This value of 0.2736 is determined as the time decay weight for this historical record. The answer score (correct / incorrect status) of each historical record is multiplied by its corresponding time decay weight. The score of the current answer interaction data is then multiplied by a preset real-time response weight (set based on the contribution of the current status to mastery, such as 0.8). All product results are summed and divided by the sum of all weights to obtain a weighted average. This value represents the user's mastery of the knowledge point node and serves as the initial user mastery status. If the dataset does not include historical time series, the current correct / incorrect answer status (0 or 1) is directly multiplied by the standard evaluation weight (such as 1.0) to directly obtain the mastery value.

[0024] Please see Figure 4 Step S3 is as follows: S301: Monitor the error marks generated by the user's answers to test questions during continuous learning, identify the knowledge point nodes that caused the error based on the mapping relationship between test questions and knowledge points, extract the sequence position index value corresponding to the knowledge point node from the linear index sequence of knowledge points, and detect the failure status of the adjacent knowledge point nodes before and after the sequence position within a preset time window. The system scans user answer stream data in real time. Once a consecutive sequence of 0 correct / incorrect answers (errors) is detected, a monitoring mechanism is triggered. Based on the unique knowledge point code marked in the test question data, the specific knowledge point node causing the error is identified. This unique knowledge point code is then used to search within the linear index sequence of knowledge points, extracting the corresponding sequence position index value. For example, the index values ​​of three consecutive incorrect nodes are extracted as 50, 51, and 52. Using these index values ​​as the center, the system searches forward to the node with index value 49 and backward to the node with index value 53, examining the status records of these two adjacent nodes within a preset time window (based on the short-term memory phase of the Ebbinghaus forgetting curve, such as the past 24 hours) to determine if there are any records of failed answers, i.e., failed states of adjacent nodes.

[0025] S302: Summarize the sequence position index values ​​corresponding to multiple answer error markers, compare whether the index values ​​show a monotonically increasing or monotonically decreasing numerical arrangement feature, and whether adjacent knowledge point nodes in the failure state of adjacent nodes show a coherent numerical arrangement feature to obtain the order consistency judgment result. The extracted index values ​​(e.g., 50, 51, 52) are placed into an array. The index values ​​are compared to see if they exhibit a monotonically increasing or monotonically decreasing numerical arrangement, and the differences between adjacent elements in the array are calculated. For example, calculating 51-50=1 and 52-51=1, if all differences are 1, it is determined to be monotonically increasing and continuous; if all differences are -1, it is determined to be monotonically decreasing and continuous. Simultaneously, the failure states of adjacent nodes are checked to determine if adjacent knowledge point nodes (e.g., 49 and 53) also exhibit a logically coherent numerical arrangement characteristic with the intermediate node (50-52) (e.g., continuity in the index). When the index values ​​satisfy monotonicity and the states of adjacent nodes verify logical coherence, the "compliant" order consistency judgment result is output.

[0026] S303: When the number of incorrect answer markers exceeds the preset minimum consecutive number threshold and meets the sequential consistency judgment result, the knowledge point node from the first failure to the knowledge point node where the consecutive failures end is delineated to generate a knowledge mastery gap interval. The preset minimum consecutive error threshold is set based on statistical analysis of the probability of random errors made by users in a fatigued state using a machine learning classifier, taking the inflection point value where the probability significantly decreases, for example, set to 3 times. If the current number of consecutive errors detected is 3 or more, and the sequence consistency judgment result is "compliant", the knowledge point node where the first error occurred (index 50) is locked as the starting point, and the knowledge point node where the consecutive failures ended (index 52) is locked as the ending point. These two nodes and all nodes contained in between (50, 51, 52) are defined as a closed set, generating the knowledge mastery gap interval.

[0027] Please see Figure 5Step S4 is as follows: S401: Based on the subject knowledge graph, identify the first knowledge point node that failed to answer the test question within the knowledge mastery gap interval, trace back the second knowledge point node associated with the first knowledge point node, and obtain the basic associated node set; For example, consider a node with index 50 (let's say a "system of two linear equations"). Using this first knowledge point node as the target, initiate the "root weakness" location logic, traversing back along the directed edges in the graph. Query all parent nodes that directly point to this first knowledge point node, i.e., the second knowledge point nodes (let's say a "linear equation"). Continue backtracking until a root node with an in-degree of 0 is found or a preset backtracking depth is reached. Collect all second knowledge point nodes involved in the backtracking path, remove duplicates, and form a basic set of related nodes. For example, the predecessor nodes of node 50 include nodes 48 and 30; these two nodes are included in the basic set of related nodes.

[0028] S402: Referring to the initial user mastery status, extract the mastery value of each knowledge point node in the basic related node set, filter all knowledge point nodes whose mastery value is lower than the preset basic standard, perform an ascending sort operation on the filtered knowledge point nodes according to the mastery value, and select the first node in the sort, i.e. the node with the lowest mastery, as the starting node of the path that needs to be solved first. Extract the corresponding mastery scores from the user profile database constructed using machine learning multidimensional feature extraction. Set a preset baseline standard, based on the mastery percentage corresponding to the passing grade in the subject, for example, 0.6. Filter out all nodes with mastery scores below 0.6. For example, node 48 has a mastery score of 0.3, and node 30 has a mastery score of 0.5. Sort the filtered nodes in ascending order according to their mastery scores, resulting in: node 48 (0.3), node 30 (0.5). Select the node at the top of the list, i.e., node 48 with the lowest mastery score, and define it as the starting node of the path that needs to be solved with the highest priority.

[0029] S403: Call the subject knowledge graph, take the starting node of the path as the starting point, and the first knowledge point node in the knowledge mastery gap interval as the ending point, retrieve all intermediate transitional knowledge point nodes connected by directed edges between the two, integrate the starting point, intermediate nodes and the ending point into a set of candidate knowledge point nodes, and connect the set of candidate knowledge point nodes to construct a closed-loop learning path. Starting from the initial node (node ​​48) and ending at the first knowledge point node (node ​​50) within the knowledge gap interval where the solution failed, a graph path search algorithm (such as breadth-first search, BFS) is used to generate the path. The algorithm initializes an empty queue to store nodes to be visited and adds the initial node 48 to the queue, marking it as visited. Node 48 is taken out of the queue, and all directed edges originating from 48 in the subject knowledge graph are queried to identify the directly adjacent successor node (assumed to be the intermediate node 49). It is determined whether node 49 is the end node 50. If not, 49 is added to the queue and its predecessor is recorded as 48. The loop continues, taking 49 out of the queue and querying its successor node, which is found to contain the end node 50. At this point, the search stops, and the path is reconstructed in reverse based on the recorded predecessor relationship: the predecessor of 50 is 49, and the predecessor of 49 is 48. The starting point (48), intermediate node (49), and end point (50) are integrated in the order of connection to construct a set of candidate knowledge point nodes. These nodes are connected sequentially to generate a closed-loop learning path that follows the logic of "basic consolidation to advanced improvement," which guides users in targeted remedial learning.

[0030] Please see Figure 6 Step S5 is as follows: S501: Obtain the examination frequency values ​​of candidate knowledge point nodes in the closed-loop learning path from the node attribute data, determine them as importance weights, and extract the current mastery values ​​of candidate knowledge point nodes based on the initial user mastery status. Calculate the difference between the preset complete mastery benchmark value and the current mastery value to quantify the unmastery value of candidate knowledge point nodes. Apply the importance weights to the unmastery value and perform a product operation to obtain the remedial needs evaluation index. The frequency values ​​are normalized to floating-point numbers between 0 and 1. For example, the importance weight of candidate node (node ​​48) (denoted as...) The current mastery level is 0.8. Based on the initial user mastery level, the current mastery level value of candidate knowledge point nodes is extracted (denoted as ). For example, a value of 0.3. Set a preset baseline value for complete mastery, which is based on the theoretical state of perfect mastery, with a value of 1.0. Calculate the difference between the baseline value of 1.0 and the current mastery value of 0.3, i.e., 1.0-0.3=0.7. This value quantifies the estimated degree of non-mastery of the candidate knowledge point node.

[0031] S502: Obtain the average mastery of all directly related knowledge point nodes of the candidate knowledge point node, and call the Sigmoid function to process the difference between the average mastery and the new knowledge learning basic mastery threshold to obtain the basic gating evaluation index. Identify the predecessor nodes of candidate nodes, read their mastery levels, and calculate their arithmetic mean, denoted as the predecessor average mastery (denoted as ). For example, if the mastery of the predecessor node is 0.6 and 0.8 respectively, then (0.6 + 0.8) ÷ 2 = 0.7, that is... The threshold is set to 0.7. Simultaneously, a threshold for the basic mastery of new knowledge is set (denoted as...). This threshold is set based on the minimum prerequisite ability score required to learn this type of knowledge point, for example, set to 0.6. If the prior knowledge level is lower than this threshold, it usually leads to a significant decrease in subsequent learning efficiency. and As a key parameter for evaluating basic gating, a comprehensive score calculation is prepared.

[0032] S503: Collect the number of days remaining until the exam, set the corresponding standard reference time factor and exam urgency adjustment coefficient, calculate the urgency correction index, and calculate the recommendation priority score for each candidate knowledge point based on the remedial needs evaluation index, basic gate control evaluation index and urgency correction index. The specific process for setting the exam urgency adjustment coefficient is as follows: Obtain the preset exam importance weights and compare the number of days remaining until the exam with the preset preparation stage division thresholds. When the remaining days are greater than the preset normal preparation threshold, the exam urgency adjustment coefficient will be set to 1. When the remaining days are between the preset normal preparation threshold and the preset emergency sprint threshold, the reduction of the remaining days relative to the preset normal preparation threshold is calculated. Based on the reduction and the importance weight of the exam, a linear weighted calculation is performed to obtain the exam urgency adjustment coefficient. When the remaining days are less than the preset emergency sprint threshold, the reciprocal of the remaining days is exponentially calculated and multiplied by the exam importance weight to obtain the exam urgency adjustment coefficient. The process of setting the exam urgency adjustment coefficient is as follows: First, obtain the preset exam importance weight, which is set according to the exam level, for example, 10 for the college entrance examination. Then, compare the remaining days until the exam (e.g., 20 days) with preset thresholds for different preparation stages. The preset normal preparation threshold is set based on the starting point of a regular review cycle, for example, 60 days; the preset emergency sprint threshold is set based on the extreme time window for pre-exam cramming, for example, 14 days. When the remaining days of 20 fall between the preset normal preparation threshold of 60 and the preset emergency sprint threshold of 14, calculate the reduction in remaining days of 20 relative to the preset normal preparation threshold of 60. A linear weighted calculation is performed based on the reduction amount and the importance weight of the exam, dividing the reduction amount of 40 by the threshold difference ( Obtain the proportion and calculate. Multiply this ratio by the exam weight of 10 and add the base coefficient of 1 to calculate... The exam urgency adjustment coefficient was obtained. It has a score of 9.7. Recommendation priority rating. The calculation formula is as follows: .in, : Represents the recommendation priority score (dimensionless number) of the candidate knowledge point node, with a value range typically [0, +∞). It is derived from the final calculation result of the formula. The larger the value, the more priority the node should be given to recommendation under the current learning state and time constraints. : Indicates candidate knowledge points The importance weight within the subject system, ranging from [0, 1]. It is derived from the examination frequency value obtained and normalized from S501. If the knowledge point is a core high-frequency test point in the syllabus, this value is set close to 1; if it is a peripheral knowledge point, it is set close to 0. In this example, it is set to 0.8. : Indicates the user's current status regarding the candidate knowledge point The mastery level estimate ranges from [0, 1]. It is derived from the weighted average calculated based on historical answer data in S203. A value close to 0 indicates no mastery, while a value close to 1 indicates complete mastery. In this example, it is set to 0.3. : Indicates the degree of lack of mastery of a knowledge point (i.e., the need for remedial learning), with a value range of [0, 1]. Logically The smaller the value (e.g., 0.3), the larger the result of this calculation (0.7), thereby improving the recommendation score and reflecting the remedial principle of "the less you know, the more priority you should give to others". The base of the natural logarithm, approximately 2.71828, is used as the base of the sigmoid function in the denominator calculation to construct nonlinear gated logic. : This represents a coefficient that adjusts the steepness of the Sigmoid function, typically ranging from (0, 10). It originates from the preset algorithm configuration. Its setting is based on controlling the sensitivity of prior mastery differences to the recommendation score. This parameter is obtained by training a machine learning regression model on historical learning behavior data: when the system detects that weak foundational knowledge (i.e., insufficient prior mastery) significantly hinders subsequent learning, the model will train to have a higher... Values ​​(e.g., 5-10) result in severe penalties; conversely, lower values ​​(e.g., 1-3) result in smaller penalties. In this example, the value is set to 5. : Indicates candidate knowledge points The average mastery of all direct predecessor knowledge points, ranging from [0, 1]. It is derived from the arithmetic mean calculation of the mastery of predecessor nodes in S502. In this example, it is set to 0.7. : Represents the threshold of prior knowledge required to learn new knowledge, ranging from [0, 1]. It is derived from preset teaching standards. Its setting is based on the minimum prerequisite score required to achieve a passing or good level: if the subject is logically rigorous (such as mathematics), a higher value is set (such as 0.6-0.8); if the logic is loose, a lower value is set. In this example, it is set to 0.6. This indicates the reserve surplus of the precursor foundation. If the result is positive (surplus), the denominator decreases, the score increases, and learning is encouraged; if the result is negative (deficit), the denominator increases exponentially, the score decreases sharply, and the "pause recommendation due to weak foundation" gating logic is triggered. : Represents the exam urgency adjustment coefficient, with a value range of [1, +∞). It originates from the weighted calculation based on the remaining days and the preparation threshold in this step. Its setting is based on the principle that as the exam approaches (days decrease), conventional logic needs to be broken, and the recommended scores for high-frequency exam points need to be forcibly increased. The more urgent the time, The larger the value, the better. In this example, the result is 9.7. This represents the reciprocal of the number of days remaining until the next important exam, ranging from 0 to 1. It's derived from calculating the reciprocal of the remaining days. The fewer the days, the larger this value, indicating a higher level of urgency. In this example, it's calculated... . : Represents the standard reference time factor, which is a fixed constant. It originates from a preset normalization benchmark. Its setting is based on dimensional normalization using the number of days in a year (365 days) as the benchmark, maintaining the balance of magnitudes between the two terms in the formula. In this example, it is set to . .

[0033] First, calculate the exponent in the denominator: .calculate The denominator is Calculate the numerator: Calculate the score for the first item: Calculate the second urgency correction value: Final recommendation priority score The calculation process is as follows: Assuming the exam date is approaching and the knowledge point is of high importance, the calculated recommendation score is extremely high, prompting this node to be prioritized at the beginning of the learning path.

[0034] S504: Based on the recommendation priority score, each candidate knowledge point is sorted in real time, the mastery value in the updated user mastery status is monitored in real time, compared with the preset breakthrough threshold, and the subsequent related knowledge point nodes are unlocked and re-scored according to the comparison results to obtain the adaptive learning path generation result. Once a user begins learning and completes a test (e.g., watching micro-lessons, doing exercises), the mastery score in the user's mastery status is monitored in real time. This updated mastery score (e.g., 0.85) is compared to a preset threshold. The preset threshold is set based on the standard score for achieving proficiency (usually corresponding to an excellent rate), for example, 0.8. If 0.85 is greater than 0.8, the comparison result is considered a pass, and the node is determined to be "mastered." Based on the connections in the knowledge graph, the subsequent related knowledge point nodes in the closed-loop learning path are unlocked, triggering a recalculation of the scores for subsequent nodes to obtain a new recommendation priority score. This dynamically refreshes the learning path, generating an adaptive learning path based on machine learning strategy optimization.

[0035] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An adaptive learning path generation method for an AI learning machine, characterized in that, Includes the following steps: S1: Obtain the chapter directory and the sequence of knowledge points of the specified subject to construct a subject knowledge graph, mark key knowledge point nodes, and establish a linear index sequence of knowledge points; S2: Respond to the user's login command and display test questions covering key knowledge points on the screen. Determine the initial user's mastery level based on the current and historical question-answering interaction records. S3: Monitor the user's incorrect answer marks for incorrect test questions during continuous learning, extract the corresponding index of the incorrect test questions from the linear index sequence of knowledge points, and determine the user's knowledge gap interval based on the index position; S4: Based on the subject knowledge graph, identify the first knowledge point node that failed to answer the test questions within the knowledge mastery gap interval, collect the second knowledge point node associated with the first knowledge point node, and construct a closed-loop learning path; S5: Based on the closed-loop learning path and the initial user mastery status, calculate the recommendation priority score for each knowledge point node, dynamically refresh the learning path based on the recommendation priority score, and obtain the adaptive learning path generation result.

2. The adaptive learning path generation method for AI learning machines according to claim 1, characterized in that: The linear index sequence of knowledge points includes a unique sequence position index value assigned to each knowledge point node according to the topological connection order of the subject knowledge graph. The initial user mastery status is specifically obtained by performing a weighted evaluation operation on the user's answer status, answer duration, and historical answer interaction records for the test questions. The knowledge mastery gap includes the sequence position index value of the first occurrence of the answer error mark extracted from the linear index sequence of knowledge points, and the sequence position index value of the sequence where the consecutive answer error marks end. The closed-loop learning path includes the first knowledge point node that failed to answer the test questions within the knowledge mastery gap and the second knowledge point node that is back-linked through the subject knowledge graph. The adaptive learning path generation result includes a sequence list after dynamically sorting the candidate knowledge point nodes according to the recommendation priority score, and the subsequently linked knowledge point nodes in the subject knowledge graph.

3. The adaptive learning path generation method for AI learning machines according to claim 1, characterized in that: Step S1 is as follows: S101: Obtain the subject chapter directory and knowledge point sequence from the cloud storage, parse the subject chapter directory, extract all knowledge point nodes, define directed edges representing the sequence relationship based on the sequence of knowledge point nodes, and combine knowledge point nodes and directed edges to construct a subject knowledge graph. S102: Based on the subject knowledge graph, configure node attribute data for each knowledge point node, including difficulty coefficient value, examination frequency value, associated question bank and teaching video resources, and mark key knowledge point nodes with reference to the examination frequency value; S103: Assign a unique sequence position index value to each key knowledge point node in the subject knowledge graph according to the topological connection order of the directed edges, arrange all the assigned sequence position index values, and establish a linear index sequence of knowledge points.

4. The adaptive learning path generation method for AI learning machines according to claim 1, characterized in that: Step S2 is as follows: S201: Respond to the user's login command and display test questions covering key knowledge points on the screen; collect the user's answer status and answering time for the test questions; and generate the current answering interaction data. S202: Detect whether there are historical answer interaction records that include the historical answer correctness status and the historical answer duration. If they exist, merge them with the current answer interaction data that includes the current answer correctness status and the current answer duration to form a dataset that includes the historical time series. If they do not exist, directly create a dataset that does not include the historical time series based on the current answer interaction data as the interaction dataset to be evaluated. S203: Identify whether the interactive dataset to be evaluated includes historical time series. If it does, extract the historical correct / incorrect answer status and historical answer duration and match time decay weights. At the same time, perform a comprehensive weighted calculation on the current interactive data by matching real-time response weights. If it does not include historical time series, directly apply standard evaluation weights to the current interactive data to perform a single evaluation calculation and calculate the user's mastery of the knowledge point nodes as the initial user mastery status.

5. The adaptive learning path generation method for AI learning machines according to claim 1, characterized in that: Step S3 is as follows: S301: Monitor the error marks generated by the user's answers to test questions during continuous learning, identify the knowledge point nodes that generated the errors based on the mapping relationship between test questions and knowledge points, extract the sequence position index value corresponding to the knowledge point node from the linear index sequence of knowledge points, and detect the failure status of adjacent nodes of the preceding and following adjacent knowledge point nodes within a preset time window. S302: Summarize the sequence position index values ​​corresponding to multiple answer error markers, compare whether the index values ​​show a monotonically increasing or monotonically decreasing numerical arrangement feature, and whether the adjacent knowledge point nodes in the adjacent node failure state show a coherent numerical arrangement feature, to obtain the order consistency judgment result. S303: When the number of incorrect answer markers exceeds the preset minimum consecutive number threshold and meets the order consistency judgment result, the knowledge point node from the first failure to the knowledge point node where the consecutive failures terminate is delineated, and a knowledge mastery gap interval is generated.

6. The adaptive learning path generation method for AI learning machines according to claim 1, characterized in that: Step S4 is as follows: S401: Based on the subject knowledge graph, identify the first knowledge point node that failed to answer the test question within the knowledge mastery gap interval, trace back the second knowledge point node associated with the first knowledge point node, and obtain the basic associated node set; S402: Referring to the initial user mastery status, extract the mastery value of each knowledge point node in the basic associated node set, filter all knowledge point nodes whose mastery values ​​are lower than the preset basic standard, perform an ascending sorting operation on the filtered knowledge point nodes according to the mastery values, and select the first node in the sorting, i.e. the node with the lowest mastery, as the starting node of the path that needs to be solved first. S403: Call the subject knowledge graph, take the starting node of the path as the starting point, and the first knowledge point node in the knowledge mastery gap interval that failed to be answered as the ending point, retrieve all intermediate transitional knowledge point nodes connected by directed edges between the two, integrate the starting point, intermediate nodes and the ending point into a candidate knowledge point node set, and connect the candidate knowledge point node set to construct a closed-loop learning path.

7. The adaptive learning path generation method for AI learning machines according to claim 1, characterized in that: Step S5 is as follows: S501: Obtain the examination frequency value of the candidate knowledge point node in the closed-loop learning path from the node attribute data, determine it as the importance weight, and extract the current mastery value of the candidate knowledge point node according to the initial user mastery status. Calculate the difference between the preset complete mastery benchmark value and the current mastery value, quantify the unmastery value of the candidate knowledge point node, and perform a product operation on the unmastery value using the importance weight to obtain the remedial needs evaluation index. S502: Obtain the average mastery of all directly related knowledge point nodes of the candidate knowledge point node, and call the Sigmoid function to process the difference between the average mastery and the new knowledge learning basic mastery threshold to obtain the basic gating evaluation index. S503: Collect the number of days remaining until the exam, set the corresponding standard reference time factor and exam urgency adjustment coefficient, calculate the urgency correction index, and calculate the recommendation priority score for each candidate knowledge point based on the remedial needs evaluation index, the basic gating evaluation index, and the urgency correction index. S504: Based on the recommended priority score, each candidate knowledge point is sorted in real time, the mastery value in the updated user mastery status is monitored in real time, compared with the preset breakthrough threshold, and the subsequent related knowledge point nodes are unlocked and re-scored according to the comparison result to obtain the adaptive learning path generation result.

8. The adaptive learning path generation method for AI learning machines according to claim 4, characterized in that: The process of setting the time decay weight is as follows: Get the historical generation timestamps corresponding to the historical question-answering interaction records and the current timestamp when the current question-answering interaction data was generated; Calculate the difference between the current timestamp and the historically generated timestamps to obtain the time interval value; Read the preset attenuation coefficient; Calculate the product of the time interval value and the attenuation coefficient; Performs exponentiation with the natural constant as the base and the opposite of the product as the exponent; The numerical result obtained from the exponentiation operation is used as the time decay weight.

9. The adaptive learning path generation method for an AI learning machine according to claim 7, characterized in that: The process for setting the examination period urgency adjustment coefficient is as follows: Obtain the preset exam importance weights and compare the number of days remaining until the exam with the preset preparation stage division thresholds. When the remaining days are greater than the preset normal preparation threshold, the exam urgency adjustment coefficient will be set to 1. When the remaining days are between the preset normal preparation threshold and the preset emergency sprint threshold, the reduction of the remaining days relative to the preset normal preparation threshold is calculated. Based on the reduction and the importance weight of the exam, a linear weighted calculation is performed to obtain the exam urgency adjustment coefficient. When the remaining days are less than the preset emergency sprint threshold, the reciprocal of the remaining days is exponentially calculated and multiplied by the exam importance weight to obtain the exam urgency adjustment coefficient.

10. The adaptive learning path generation method for an AI learning machine according to claim 7, characterized in that: For recommendation priority scoring, the formula is as follows: ； in, The score represents the recommendation priority of candidate knowledge point nodes. Indicate candidate knowledge points Its importance and weight within the academic discipline system This indicates the user's current status regarding the candidate knowledge point. The degree of control and valuation This indicates the degree to which the knowledge point has not been mastered. The base of the natural logarithm. The coefficient representing the steepness of the Sigmoid function. Indicate candidate knowledge points Average mastery of all direct precursor knowledge points This represents the threshold of prior knowledge required to learn new knowledge. This represents the adjustment coefficient for the urgency of the exam period. It represents the reciprocal of the number of days remaining until the next important exam. Indicates the standard reference time factor.

Images