Personalized physical exercise recommendation system and method based on big data analysis

Abstract

The application discloses a personalized physical exercise question recommendation system and method based on big data analysis, relates to the technical field of exercise question recommendation, and aims at the problems that the existing method is difficult to identify the thinking blind area of students and does not consider the dynamic influence of group learning progress.The application aligns the time sequence of the historical and review data of individual students and real-time answering data of groups, inputs the individual sequence into a Transform model to output the forgetting probability and performance change gradient of the individual, divides the knowledge point state by using the group sequence features, dynamically allocates the weight by combining the examination time distance parameter, generates a comprehensive landslide index, and finally searches and matches the hierarchical exercise questions from the question bank according to the index and sorts and outputs the questions.The application combines the individual forgetting rule and group cognitive dynamics, accurately depicts the knowledge mastery state and thinking blind area of students, and significantly improves the accuracy and adaptability of the exercise question recommendation.

Description

Technical Field

[0001] This application relates to the technical field of exercise recommendation, and in particular to a personalized physics exercise recommendation system and method based on big data analysis. Background Technology

[0002] Existing personalized physics exercise recommendation technologies are mostly based on students' historical answer accuracy or knowledge mastery, but they still have shortcomings in practical applications. Solving physics problems has distinct thought process characteristics; even if students have mastered the knowledge points, they may still be unable to construct effective problem-solving strategies. Existing methods lack the ability to model the problem-solving process and struggle to identify students' blind spots in thinking, leading to a mismatch between recommended results and actual abilities.

[0003] Meanwhile, existing forgetting models are mostly based on individual static memory patterns, considering only time and review frequency, failing to reflect the dynamic impact of changes in group learning progress on the difficulty of knowledge points. In the teaching process, knowledge points often exhibit periodic hotspot characteristics as the group's cognitive state changes, and related technologies lack effective characterization of this. Furthermore, existing recommendation methods typically rely on post-event adjustments based on answer results, failing to perceive students' cognitive load in real time during problem-solving, making it difficult to optimize recommendation strategies promptly. Additionally, recommendation granularity often remains at the whole question level, failing to identify differences in students' abilities at the sub-question level, and not fully considering differences in learning styles, thus affecting recommendation effectiveness. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, one objective of this application is to propose a personalized physics exercise recommendation system and method based on big data analysis. By integrating individual forgetting patterns and group cognitive dynamics, and combining time series modeling and a multi-dimensional labeling system, it can accurately depict students' knowledge mastery status and cognitive blind spots, thereby improving the accuracy and adaptability of personalized physics exercise recommendations.

[0006] To achieve the above objectives, the first aspect of this application proposes a personalized physics exercise recommendation method based on big data analysis, comprising the following steps: Real-time acquisition of historical answer records and review frequency data of target students, as well as real-time answer performance data of the target class group under the target physics topic; The historical answer record data, the review frequency data, and the group real-time answer performance data are subjected to time series alignment processing to generate an individual knowledge memory intensity sequence and a group knowledge point difficulty fluctuation sequence. The individual knowledge memory strength sequence is input into a pre-trained Transformer model for modeling, and the individual forgetting probability and performance change gradient of each physical knowledge point are output. Using the distribution characteristics of the group error rate in the difficulty fluctuation sequence of the group knowledge points, a hierarchical model of knowledge point status is constructed to divide each of the physical knowledge points into a stable mastery state, a potential forgetting state, a group hotspot state, or a high-risk decline state. Calculate the time span between the current system time and the preset assessment node, and generate assessment time distance parameters; Based on the assessment time distance parameter, a dynamic scheduling function is constructed to perform real-time weight switching calculation on the individual forgetting channel weight corresponding to the individual forgetting probability and the group cognition channel weight corresponding to the knowledge point state hierarchical model, and to generate a comprehensive slippage index for each knowledge point. Based on the comprehensive landslide index and the knowledge point status classification results, a matching hierarchical exercise set is retrieved from the preset exercise bank, sorted and output, and a target personalized physics exercise set is generated.

[0007] In addition, the personalized physics exercise recommendation system and method based on big data analysis proposed in this application may also have the following additional technical features: In one embodiment of this application, the time-series alignment processing of the historical answer record data, the review frequency data, and the group real-time answer performance data includes: Extract answer timestamps and accuracy features from historical answer records, combine them with review interval duration from review frequency data, calculate initial memory feature values ​​using a preset memory decay function, and generate an individual knowledge memory intensity sequence indexed by time. Extract the error rate fluctuation, average answering time deviation, and answering consistency probability of each knowledge point in the real-time answering performance data of the group, and construct the group knowledge point difficulty fluctuation sequence. Linear interpolation was used to align the sampling frequencies of the individual knowledge memory intensity sequence and the group knowledge point difficulty fluctuation sequence to ensure that the two sequences have corresponding data observation points under the same teaching time axis.

[0008] In one embodiment of this application, the specific steps of inputting the individual knowledge memory intensity sequence into a pre-trained Transformer model for modeling include: The individual knowledge memory intensity sequence is converted into a multi-dimensional embedding vector, and positional encoding is added to represent the teaching progress cycle; The memory strength correlation weights at different time points are calculated through the multi-head self-attention mechanism layer of the Transformer model, and the individual memory hidden state vector is output. The hidden state vector is input into the prediction fully connected layer, and the individual forgetting probability of each knowledge point in the next physical cycle is calculated by the logistic regression function. Perform a first-order difference operation on the individual forgetting probability sequence to calculate the change in the forgetting rate of each knowledge point under the current teaching step size, which serves as the performance change gradient.

[0009] In one embodiment of this application, constructing the knowledge point state hierarchy model includes: Calculate the average error rate and standard deviation of each knowledge point in the real-time answer performance data of the group to generate a group cognitive benchmark; If the group error rate of the target knowledge point is lower than the preset threshold and the performance change gradient is stable, it is judged to be in a stable mastery state. If the individual forgetting probability exceeds a preset threshold and the group error rate is within the normal range, it is determined to be a potential forgetting state. If the group error rate increases by a step within the preset sliding window, it is determined to be a group hotspot state. If both the criteria for determining a hotspot state in a group and the criteria for an increased probability of forgetting an individual are met simultaneously, then it is determined to be a high-risk landslide state.

[0010] In one embodiment of this application, the specific steps for generating the comprehensive landslide index for each knowledge point are as follows: Substitute the assessment time distance parameter into the preset S-shaped growth curve function to calculate the group weight adjustment factor that increases non-linearly as the exam approaches. Establish individual forgetting channel weights Group cognitive channel weights The constraint relationship makes ; When the assessment time distance parameter exceeds the preset long-term learning threshold, Set as the primary weight; When the assessment time distance parameter is less than or equal to the pre-exam sprint threshold, the group weight adjustment factor is increased. This enables dynamic switching of weights to the group channel; The probability of individual forgetting and Multiply, the population error rate distribution score and Multiply them and sum them to get the comprehensive landslide index.

[0011] In one embodiment of this application, the specific steps of retrieving a matching hierarchical set of exercises from a preset exercise bank and sorting and outputting it are as follows: Extract knowledge point tags, difficulty level tags, and common thinking error tags from the exercise data; Basic reinforcement exercises are matched with knowledge points in the potential forgetting state; typical and difficult exercises are matched with knowledge points in the group hot state; and comprehensive and variant exercises are matched with knowledge points in the high-risk decline state. Exercises selected by sorting in descending order of comprehensive landslide index; If the error label of a problem matches the historical error characteristics of the target student, the priority of that problem in the personalized problem set will be increased.

[0012] In one embodiment of this application, before obtaining the group's real-time answer performance data, the method further includes: The system monitors the original answer log packets submitted by each student's mobile terminal in real time. The log packets contain unique identifiers for questions, answer option identifiers, answer duration, and answer device environment data. The log packets are subjected to structured parsing and noise reduction to remove invalid records whose response time is below the abnormal threshold; Based on the physics knowledge graph, log packages are mapped to corresponding physics topics, and aggregated to generate real-time group answer performance data.

[0013] In one embodiment of this application, after generating the personalized exercise set, the method further includes: Data on the target students' secondary responses to the exercise set is obtained through a pre-set feedback monitoring loop; Calculate the predicted residual value of the second response data relative to the individual forgetting probability; The residual values ​​are used as backpropagation signals input into the Transformer model, and the model weights are fine-tuned online using the gradient descent algorithm.

[0014] To achieve the above objectives, a second aspect of this application proposes a personalized physics exercise recommendation system based on big data analysis, which is implemented based on the aforementioned personalized physics exercise recommendation method based on big data analysis.

[0015] To achieve the above objectives, a third aspect of this application provides a computer execution device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform steps of a personalized physics exercise recommendation method based on big data analysis.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: 1. This invention constructs the individual knowledge memory intensity sequence and combines it with the Transformer model for time series modeling, which can characterize the dynamic changes in students' knowledge mastery status, achieve accurate prediction of individual forgetting trends, and thus improve the targeting of exercise recommendations.

[0017] 2. This invention introduces the difficulty fluctuation sequence of the group's knowledge points and constructs a hierarchical model of knowledge point status, so that the recommendation results simultaneously reflect the individual's weak links and the group's common difficulties, thereby improving the matching degree between the recommended content and the actual teaching.

[0018] 3. This invention improves personalized learning effectiveness and overall recommendation efficiency by setting a dynamic scheduling mechanism based on assessment time distance parameters and combining it with a multi-dimensional exercise tag system to achieve adaptive adjustment and refined matching of recommendation strategies.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the steps of the personalized physics exercise recommendation method based on big data analysis proposed in this application. Figure 2 This is a structural block diagram of the personalized physics exercise recommendation system based on big data analysis proposed in this application. Detailed Implementation

[0021] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0022] The following description, in conjunction with the accompanying drawings, illustrates a personalized physics exercise recommendation system and method based on big data analysis, according to embodiments of this application.

[0023] like Figure 1 As shown in the figure, the personalized physics exercise recommendation method based on big data analysis in this application embodiment may include the following steps: Real-time acquisition of historical answer records and review frequency data of target students, as well as real-time answer performance data of the target class group under the target physics topic; In this embodiment, historical answer record data refers to the answer records of all physics exercises completed by the target student (e.g., Xiaoming) in the past period of time. Each record includes at least the answer timestamp (accurate to the second), the physics knowledge point corresponding to the question (such as "the application of the law of conservation of momentum"), and the correctness indicator (1 for correct and 0 for incorrect).

[0024] Review frequency data refers to the record of Xiaoming's active review of specific physics knowledge points. For example, if Xiaoming reviews his notes on the "Conservation of Momentum" topic or redoes a classic example problem at a certain time, the system will record the time of this review behavior. The target class group refers to the teaching class Xiaoming is in; in this embodiment, "Grade 11 Class 3" is used as an example. Real-time answer performance data refers to the instant answer feedback data of all students in Grade 11 Class 3 on the "Conservation of Momentum" physics topic within the current teaching cycle (e.g., the most recent week), including the correctness of each student's answers, the time taken to answer, and the interactive behavior during the answering process.

[0025] The aforementioned data can be collected by monitoring the answer logs submitted by students' terminals in real time through the backend server of the online education platform. By acquiring this data in real time, the system can comprehensively capture individual learning trajectories and group learning dynamics, laying a data foundation for subsequent personalized recommendations.

[0026] Time series alignment processing is performed on historical answer record data, review frequency data, and group real-time answer performance data to generate individual knowledge memory intensity sequences and group knowledge point difficulty fluctuation sequences. It should be noted that since historical answer records, review frequency records, and real-time group answer performance data come from different time points and sampling frequencies, they must be aligned onto the same time axis before subsequent fusion analysis can be performed.

[0027] The individual knowledge memory intensity sequence reflects the curve of how well Xiaoming retains specific knowledge points (such as the conservation of momentum) over time. Its theoretical basis is the Ebbinghaus forgetting curve, but it is dynamically adjusted to incorporate the individual's actual review behavior. The group knowledge point difficulty fluctuation sequence reflects the change in the cognitive difficulty of a specific knowledge point for the class as the teaching progresses, quantified by statistical indicators such as group error rate and response time. By generating these two sequences, the system transforms the raw data into a structured, temporal feature representation, providing a standardized input format for subsequent deep learning modeling.

[0028] In one embodiment of this application, time-series alignment processing is performed on historical answer record data, review frequency data, and real-time group answer performance data, including: Extract answer timestamps and accuracy features from historical answer records, combine them with review interval duration from review frequency data, calculate initial memory feature values ​​using a preset memory decay function, and generate an individual knowledge memory intensity sequence indexed by time. Specifically, the preset memory decay function is an improved form based on the Ebbinghaus forgetting curve principle. Its core idea is that after each test or review, memory strength reaches a peak, then decays exponentially over time; when the test or review is repeated, memory strength increases again. This function can simulate the natural forgetting curve of human memory and dynamically adjust it based on individual review behavior, making the estimation of memory strength more accurate. The mathematical expression of the memory decay function is: ; In the formula, The time since the last time you answered a question or reviewed a question. The subsequent memory strength value; The initial memory strength value is set to 1 for correct answers, 0.3 to 0.6 for incorrect answers depending on the difficulty of the question, and 0.7 to 0.9 for review depending on the depth of review. This refers to the time interval since the last time you answered a question or reviewed something. This is an individual memory half-life parameter, dynamically adjusted based on the review interval duration in the review frequency data. The higher the review frequency, the longer the half-life, and the slower the memory decay. Through this function, the system can generate a continuous curve with time as the horizontal axis and memory intensity as the vertical axis, i.e., an individual knowledge memory intensity sequence.

[0029] Extract the error rate fluctuation, average answering time deviation, and answering consistency probability of each knowledge point in the real-time answering performance data of the group, and construct the group knowledge point difficulty fluctuation sequence. In this embodiment, time slicing refers to grouping data according to fixed time windows, such as slicing by "days," to observe the trend of difficulty changes over time. Error rate fluctuation reflects the magnitude of accuracy changes across consecutive time slices; a larger fluctuation indicates a more unstable grasp of the knowledge point by the group. Average answer time deviation reflects the degree of deviation of the group's answer time from the baseline; a significant increase in time usually indicates that the questions are more difficult or that students are confused. Answer consistency probability reflects the degree of consistency in the distribution of answers for a particular knowledge point, and is calculated using the following formula: ; In the formula, For the first The probability of consistency in answers to each knowledge point; The number of times the most frequently selected option for this knowledge point is chosen. This represents the total number of participants who answered questions on this knowledge point. A lower consistency probability indicates greater disagreement among students regarding their understanding of the knowledge point, and thus, a higher level of difficulty. Through these three dimensions, the system can comprehensively depict the group's understanding of the difficulty of grasping a knowledge point and its dynamic changes.

[0030] Linear interpolation was used to align the sampling frequencies of the individual knowledge memory intensity sequence and the group knowledge point difficulty fluctuation sequence to ensure that the two sequences have corresponding data observation points under the same teaching time axis.

[0031] In this embodiment, linear interpolation is used to address the issue of inconsistent sampling frequencies between the two sets of sequences. The time points of the individual knowledge memory intensity sequence depend on the time Xiaoming spends answering questions or reviewing, and these time points are irregular; while the group knowledge point difficulty fluctuation sequence is sampled according to fixed time slices (e.g., daily). If they are not aligned, the two sets of sequences cannot be compared or merged at the same time point. The linear interpolation formula is: ; In the formula, Interpolated time points The value; and For two adjacent known time points in an individual sequence; and This represents the known memory strength value at the corresponding time point. Using this formula, the system can calculate the estimated value of an individual sequence at any time point, achieving time alignment with the group sequence.

[0032] The individual knowledge memory strength sequence is input into the pre-trained Transformer model for modeling, and the individual forgetting probability and performance change gradient of each physical knowledge point are output. Specifically, the pre-trained Transformer model is trained as follows: First, a large number of students' long-term learning records in physics are extracted from the historical database of the online education platform. The time series alignment processing method mentioned above is used to transform the original answer data into a sequence of individual knowledge memory strength indexed by time as input features, and the actual correctness of students' answers in subsequent teaching cycles is used as supervision labels. Then, a multi-layer Transformer encoder model is constructed, which includes an input embedding layer, a position encoding layer, a multi-head self-attention layer, a feedforward neural network layer, and an output layer. The binary classification cross-entropy loss function is used to measure the difference between the predicted value and the true label, and the Adam optimization algorithm is used for iterative optimization on the training dataset. At the same time, an early stopping mechanism is set to prevent overfitting. After training, the model structure and optimized weight parameters are saved as a pre-trained model file. In actual applications, the system directly loads the pre-trained model for forward inference and can be combined with the aforementioned feedback monitoring loop for online fine-tuning and updating, thereby achieving accurate prediction of the probability of individual forgetting.

[0033] In one embodiment of this application, the specific steps for inputting the individual knowledge memory intensity sequence into a pre-trained Transformer model for modeling include: The individual knowledge memory intensity sequence is converted into a multi-dimensional embedding vector, and positional encoding is added to represent the teaching progress cycle; Multidimensional embedding vectors refer to vector representations that map the memory strength value (a scalar) at each time point in a sequence to a high-dimensional space (e.g., 128-dimensional), enabling the model to capture richer feature information. Positional encoding is used to inject the positional information of each element in the sequence into the model, because the Transformer model itself does not possess order awareness. Positional encoding uses sine and cosine functions, and its expression is: ; ; In the formula, For position indices in the sequence (e.g., day 1, day 2); For dimension indexing; is the total dimension of the embedded vector. Positional encoding enables the model to distinguish between "memory strength on day 5" and "memory strength on day 10," thereby capturing the periodic characteristics of teaching progress, such as weekly learning rhythm and periodic review patterns.

[0034] The memory strength correlation weights at different time points are calculated through the multi-head self-attention mechanism layer of the Transformer model, and the individual memory hidden state vector is output. In this embodiment, the multi-head self-attention mechanism can calculate the association strength between any two time points in a sequence, thereby capturing long-range dependencies. For example, the model can learn the association between "review behavior on day 12" and "memory strength on day 20," a long-range dependency that traditional RNN models struggle to capture effectively. The calculation formula for the multi-head self-attention mechanism is as follows: ; In the formula, , , These are the query matrix, key matrix, and value matrix, respectively, which are obtained from the input embedding vector through different linear transformations. The dimension of the key vector is [dimensionality]. Through multiple parallel attention heads (e.g., 8 heads), the model can capture dependencies in the sequence from different perspectives, outputting an individual memory hidden state vector containing global temporal information. This hidden state vector integrates information from all time points in the sequence, providing a rich feature base for subsequent forgetting probability prediction.

[0035] The hidden state vector is input into the prediction fully connected layer, and the individual forgetting probability of each knowledge point in the next physical cycle is calculated by the logistic regression function. The predictive fully connected layer is a linear transformation layer that maps a high-dimensional hidden state vector to a scalar value. The logistic regression function (also known as the sigmoid function) compresses this scalar value to between 0 and 1, giving it a probabilistic meaning. The calculation formula is: ; In the formula, For individual forgetting probability; For each individual's hidden state vector; This is the weight matrix of the fully connected layer; For bias terms; This is a natural constant. The closer the output value is to 1, the higher the probability of forgetting. Through this step, the system achieves end-to-end prediction from the historical memory strength sequence to the future forgetting probability.

[0036] Perform a first-order difference operation on the individual forgetting probability sequence to calculate the change in the forgetting rate of each knowledge point under the current teaching step size, which serves as the performance change gradient.

[0037] In this embodiment, the first-order difference operation is used to calculate the changing trend of the forgetting probability, which captures the dynamic characteristics of the forgetting process. The first-order difference operation calculates the difference in forgetting probabilities between adjacent time points; a positive value indicates an increasing forgetting probability (accelerated forgetting), and a negative value indicates a decreasing forgetting probability (slowed forgetting). This indicator is crucial for judging the state of knowledge points. For example, although the current forgetting probability of a knowledge point is not high, if the forgetting rate is rapidly increasing, there is a tendency for it to evolve into a high-risk state, requiring early intervention. The calculation formula is: ; In the formula, For time points The performance change gradient; This represents the individual's forgetting probability at the current time point. This represents the individual's forgetting probability at the previous time point. A positive gradient of performance change indicates accelerated forgetting, while a negative gradient indicates slowed forgetting.

[0038] By utilizing the distribution characteristics of the group error rate in the difficulty fluctuation sequence of group knowledge points, a hierarchical model of knowledge point status is constructed, which divides each physics knowledge point into a stable mastery state, a potential forgetting state, a group hotspot state, or a high-risk decline state. It should be noted that the role of the knowledge point status stratification model is to qualitatively classify the current state of knowledge points. Individual forgetting probabilities alone can only reflect individual-level memory status, but cannot reflect cognitive patterns at the group level. For example, a certain knowledge point may become a hot topic among the group due to insufficient teacher explanation or changes in exam focus. This kind of group cognitive disturbance will significantly affect individuals' need to review that knowledge point.

[0039] This step analyzes the distribution characteristics of the group's error rate, categorizing knowledge points into four states to provide a basis for subsequent recommendation strategies. The "Stable Mastery" state indicates that the knowledge point has been generally mastered by the group, and individual memory is relatively stable, requiring no additional intervention. The "Potential Forgetting" state indicates that there is a risk of forgetting at the individual level, but it has not yet become a common difficulty for the group; it is suitable for intervention through individualized basic reinforcement exercises. The "Group Hotspot" state indicates that there is a common cognitive difficulty at the group level, becoming a "common difficulty" for the class, requiring focused attention. The "High-Risk Slippage" state indicates that the risk of individual forgetting overlaps with the group's difficulty, making it the state requiring the highest priority for intervention. Through this hierarchical classification, the system can accurately locate the cognitive stage of different knowledge points, laying the foundation for subsequent tiered recommendations.

[0040] In one embodiment of this application, constructing a knowledge point state hierarchy model includes: Calculate the average error rate and standard deviation of each knowledge point in the real-time answer performance data of the group to generate a group cognitive benchmark; In this embodiment, the average error rate reflects the overall mastery of the knowledge point by the group, while the standard deviation reflects the degree of differentiation in cognitive levels within the group. The calculation formulas are as follows: ; ; In the formula, For the first The average error rate of each knowledge point; The total number of students in the class group; For students In the Error indicators for each knowledge point (1 for error, 0 for correct); For the first The standard deviation of the error rate for each knowledge point. The higher the average error rate and the larger the standard deviation, the more difficult the knowledge point is for the group, and the greater the cognitive disagreement.

[0041] For example, the system calculates the answer data of Class 3, Grade 11 on the knowledge point of "conservation of momentum" and obtains the average error rate. Standard deviation This indicates that the class as a whole did not have a good grasp of the knowledge point, and that there were certain differences in understanding among the students.

[0042] If the group error rate of the target knowledge point is lower than the preset threshold and the performance change gradient is stable, it is judged to be in a stable mastery state. In this embodiment, the preset threshold is typically set to 0.3 (i.e., the group accuracy rate exceeds 70%). The criterion for judging the stability of the performance change gradient is that the absolute value of the gradient for three consecutive time steps is less than the preset stability threshold (e.g., 0.05). Knowledge points in this state have the lowest priority in subsequent recommendations, avoiding students wasting time on content they have already mastered. For example, for the knowledge point "Newton's First Law", the group error rate of Class 3, Grade 11 has remained below 0.2 for a long time, and the absolute value of the performance change gradient is always less than 0.03, so the system judges it as a "stable mastery state".

[0043] If the individual forgetting probability exceeds a preset threshold and the group error rate is within the normal range, it is determined to be a potential forgetting state. In this embodiment, the preset threshold is the individual forgetting probability threshold, typically set to 0.6. A population error rate within the normal range means that the population error rate is lower than the sum of the first preset threshold (0.3) and 1.5 times the standard deviation, i.e. This indicates that the knowledge point has not yet posed a general difficulty for the group.

[0044] If the group error rate increases by a step within the preset sliding window, it is determined to be a group hotspot state. In this embodiment, the preset sliding window is typically set to 3 to 5 consecutive teaching cycles (e.g., 3 days). The detection condition for a step increase is: the average group error rate within the current window increases by more than a preset step threshold (e.g., 0.2) compared to the previous window, and the upward trend continues for at least two windows. Knowledge points in this state have a higher recommendation priority during the pre-exam sprint stage.

[0045] For example, the group error rate of Class 3, Grade 11 on the knowledge point of "conservation of momentum" increased from 0.45 on the 10th day to 0.52 on the 11th day, and then to 0.58 on the 12th day. The error rate increased for three consecutive windows, with a cumulative increase of 0.13. The system judged that "conservation of momentum" was in a hot spot state.

[0046] If both the criteria for determining a hotspot state in a group and the criteria for an increased probability of forgetting an individual are met simultaneously, then it is determined to be a high-risk landslide state.

[0047] In this embodiment, the condition for an increased individual forgetting probability is that the individual forgetting probability at the current time point increases by more than a preset increase threshold (e.g., 0.1) compared to the previous time point, or that the individual forgetting probability at two consecutive time points shows an upward trend. Knowledge points in this state have the highest priority in recommendations, and the system will prioritize pushing comprehensive variation exercises for intensive training.

[0048] For example, Xiaoming's individual forgetting probability increases from 0.65 on day 25 to 0.72 on day 30. At the same time, the knowledge point of "conservation of momentum" is judged to be a hot topic in the group. When both conditions are met, the system judges Xiaoming's state of "conservation of momentum" as "high-risk landslide state".

[0049] The system calculates the time span between the current system time and preset assessment nodes, generating an assessment time distance parameter. Preset assessment nodes refer to key teaching assessment dates, such as unit tests, midterm exams, and final exams. The assessment time distance parameter is expressed in days, representing the remaining time until the assessment. This parameter allows the recommendation system to dynamically adjust its recommendation strategy based on "how long until the exam": when the exam is far away, the system focuses more on cultivating individual long-term memory patterns; when the exam is closer, the system focuses more on reinforcing key concepts and challenging areas. By introducing this time dimension, the system achieves timely recommendations that match the teaching pace, avoiding the problem of recommendation strategies becoming disconnected from the teaching cycle.

[0050] Assuming the current system time is May 20th, and the preset midterm exam date is May 30th, the system calculates the time span between the current time and the exam date, obtaining an exam time distance parameter of 10 days.

[0051] Based on the assessment time distance parameter, a dynamic scheduling function is constructed to perform real-time weight switching calculation on the individual forgetting channel weight corresponding to the individual forgetting probability and the group cognition channel weight corresponding to the knowledge point state hierarchical model, and to generate a comprehensive slippage index for each knowledge point. Understandably, the closer to the exam, the greater the impact of group hot topics and difficulties on exam scores, so the weight of the group cognitive channel should be increased; the further away from the exam, the more important it is to cultivate individual long-term memory patterns, so the weight of the individual forgetting channel should be increased.

[0052] In one embodiment of this application, the specific steps for generating the comprehensive landslide index for each knowledge point are as follows: Substituting the assessment time distance parameter into a preset S-shaped growth curve function, the group weight adjustment factor, which increases non-linearly as the exam approaches, is calculated. It should be noted that the S-shaped growth curve function (logistic function) can simulate the psychology of "pre-exam sprint"—when the exam is far away, the group weight adjustment factor increases slowly, and individuals pay less attention to group hot topics; as the exam approaches, the group weight adjustment factor increases rapidly, and individuals begin to focus on group hot topics and difficulties; as the exam nears, the growth tends to level off, and individuals enter the final sprint state. This non-linear change conforms to the psychological patterns of students preparing for exams in teaching practice. Its expression is: ; In the formula, As a group weight adjustment factor; For the assessment time distance parameter (number of days from the assessment); The midpoint parameter of the curve represents the critical point at which the weights begin to change significantly (usually set to 7 days). This is a growth rate parameter that controls the steepness of the weight change (usually set to 0.5). This is the maximum adjustment coefficient for the group weight (usually set to 0.8).

[0053] Establish individual forgetting channel weights Group cognitive channel weights The constraint relationship makes It should be noted that, in the initial state, the individual forgetting channel weights... The value is 0.7, representing the weight of the group's cognitive channel. The value of 0.3 reflects the dominant role of individual factors in the regular learning stage.

[0054] When the assessment time distance parameter exceeds the preset long-term learning threshold, Set as the primary weight; When the assessment time distance parameter exceeds this threshold, the system considers it to be in a long-term learning phase, where the individual's own memory patterns have a greater impact on knowledge acquisition than the group's difficulties; therefore, it maintains... As the primary weight.

[0055] When the assessment time distance parameter is less than or equal to the pre-exam sprint threshold, the group weight adjustment factor is increased. This enables dynamic switching of weights to the group channel; When the assessment time distance parameter is less than or equal to this threshold, the system enters the pre-exam sprint mode, increasing the group weight adjustment factor. The formula for calculating weight switching is: ; ; In the formula, The basic group weight (initial value is 0.3); As a group weight adjustment factor; This is the upper limit of the group weight (usually set to 0.8) to prevent the group weight from overly dominating the recommendation results and causing individual differences to be ignored.

[0056] For example, suppose there are only 2 days left until the exam. Substituting into the S-curve calculation yields ,but , The weighting is somewhat biased towards the group channel.

[0057] The probability of individual forgetting and Multiply, the population error rate distribution score and Multiply and sum to obtain the comprehensive landslide index. The formula for calculating the comprehensive landslide index is: ; In the formula, The comprehensive landslide index; For individual forgetting probability; This section assigns stratified scores to the distribution characteristics of the group error rate. The stratified scores are set as follows: 0.2 for stable mastery, 0.5 for potential forgetting, 0.7 for group hotspots, and 0.9 for high-risk decline. This stratified scoring allows the qualitative classification of knowledge point status to be quantified into continuous values, facilitating weighted integration with individual forgetting probabilities. A higher overall decline index indicates a higher recommendation priority for that knowledge point.

[0058] Based on the comprehensive landslide index and the knowledge point status classification results, the system retrieves matching hierarchical exercise sets from the preset exercise bank, sorts and outputs them, and generates a target personalized physics exercise set.

[0059] The pre-built question bank is a database of physics questions. Each question is labeled with a knowledge point tag (such as "conservation of momentum"), a difficulty level tag (basic, medium, difficult), and a common misconception tag (such as "ignoring the directionality of momentum"). These tags provide the foundation for accurate question matching. Based on the knowledge point classification, the system matches different types of questions. A comprehensive slippage index is used to prioritize the selected questions; questions corresponding to knowledge points with higher comprehensive slippage indices are ranked higher.

[0060] In one embodiment of this application, the specific steps for retrieving a matching hierarchical set of exercises from a preset exercise library and sorting and outputting it are as follows: Extract knowledge point tags, difficulty level tags, and common thinking error tags from the exercise data; It should be noted that the knowledge point tags identify the physics knowledge points tested in the exercises, and the difficulty level tags are divided into three levels: basic, medium, and difficult. These tags are pre-labeled by teachers or algorithms based on factors such as the complexity of the question and the number of reasoning steps required. The thinking error tags identify common erroneous thinking patterns in the exercises (such as "ignoring the directionality of momentum" or "confusing momentum with kinetic energy"), and are used to accurately match an individual's historical error patterns.

[0061] Basic reinforcement exercises are matched for knowledge points in the potential forgetting state, typical and difficult exercises are matched for knowledge points in the group hot state, and comprehensive and variant exercises are matched for knowledge points in the high-risk decline state. The selected exercises are arranged in descending order of comprehensive decline index. The higher the comprehensive decline index of the knowledge point, the higher its corresponding exercises are placed, ensuring that the knowledge points with higher decline risk are presented to the target students first.

[0062] If the error label of a problem matches the historical error characteristics of the target student, the priority of that problem in the personalized problem set will be increased.

[0063] It should be noted that the historical error feature matrix is ​​a data structure that records the past error patterns of individual target students. Each row corresponds to a knowledge point, and each column corresponds to a type of cognitive bias. The values ​​in the matrix represent the frequency or weight of a certain type of error made by the student. When the cognitive bias label of an exercise matches the cognitive bias type with a higher weight under that knowledge point in the historical error feature matrix, the ranking of that exercise in the exercise set is improved (e.g., improved by 3 to 5 places).

[0064] In one embodiment of this application, before obtaining the group's real-time answer performance data, the method further includes: Real-time monitoring of raw answer log packets submitted by each student's mobile terminal. The log packets contain unique question identifiers, answer option identifiers, answer duration, and answer device environment data. The log packets are subjected to structured parsing and noise reduction to remove invalid records whose response time is below the abnormal threshold; Based on the physics knowledge graph, log packages are mapped to corresponding physics topics, and aggregated to generate real-time group answer performance data.

[0065] This step aggregates scattered answer records by topic. A physics knowledge graph is a knowledge network describing the relationships between physics knowledge points, including hierarchical relationships, pre- and post-relationships, and inclusion relationships. The mapping process involves retrieving the knowledge point node to which a question belongs from the knowledge graph based on its unique identifier, thereby determining its corresponding physics topic (e.g., "Mechanics," "Electromagnetism," "Thermodynamics," etc.). Aggregation generation involves summarizing and statistically analyzing the answer data of all students within the same physics topic according to time windows, generating real-time group answer performance data.

[0066] In one embodiment of this application, after generating the personalized exercise set, the method further includes: Data on the target students' secondary responses to the exercise set is obtained through a pre-set feedback monitoring loop; Calculate the predicted residual value of the second response data relative to the individual forgetting probability; The residual values ​​are used as backpropagation signals input into the Transformer model, and the model weights are fine-tuned online using the gradient descent algorithm.

[0067] It should be noted that all the above data was collected with the explicit authorization of the students and their guardians, and in strict accordance with relevant data privacy protection regulations to ensure the security and compliance of user information.

[0068] A personalized physics exercise recommendation system based on big data analysis is implemented using a personalized physics exercise recommendation method based on big data analysis, such as... Figure 2 As shown, it includes a data acquisition module, a feature construction module, a time series prediction module, a state modeling module, a dynamic scheduling module, and a recommendation generation module.

[0069] The data acquisition module acquires real-time historical answer records and review frequency data of individual target students, as well as real-time group answer performance data of the target class on the target physics topic.

[0070] The feature construction module performs time series alignment processing on the historical answer record data, the review frequency data, and the group's real-time answer performance data to construct an individual knowledge memory intensity sequence and a group knowledge point difficulty fluctuation sequence.

[0071] The time-series prediction module inputs the individual knowledge memory intensity sequence into a pre-trained Transformer model for modeling, and outputs the individual forgetting probability and performance change gradient of each physical knowledge point.

[0072] The state modeling module utilizes the distribution characteristics of the group error rate in the difficulty fluctuation sequence of the group's knowledge points to construct a hierarchical model of knowledge point states.

[0073] The dynamic scheduling module calculates the assessment time distance parameter and constructs a dynamic scheduling function based on the assessment time distance parameter. It performs real-time weight switching calculation on the individual forgetting evolution channel weight and the group cognitive disturbance channel weight to generate a comprehensive landslide index.

[0074] The recommended generation module is used to generate and output a target personalized physics exercise set from a preset exercise library based on the comprehensive landslide index and state classification results.

[0075] The modules described above communicate with each other through data interfaces, forming a complete data processing pipeline to jointly realize the personalized physics exercise recommendation function. The specific implementation methods of each module can be found in the corresponding step descriptions in the foregoing method embodiments, and will not be repeated here.

[0076] A computer execution device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any personalized physics exercise recommendation method based on big data analysis.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A personalized physics exercise recommendation method based on big data analysis, characterized in that: Includes the following steps: Real-time acquisition of historical answer records and review frequency data of target students, as well as real-time answer performance data of the target class group under the target physics topic; The historical answer record data, the review frequency data, and the group real-time answer performance data are subjected to time series alignment processing to generate an individual knowledge memory intensity sequence and a group knowledge point difficulty fluctuation sequence. The individual knowledge memory strength sequence is input into a pre-trained Transformer model for modeling, and the individual forgetting probability and performance change gradient of each physical knowledge point are output. Using the distribution characteristics of the group error rate in the difficulty fluctuation sequence of the group knowledge points, a hierarchical model of knowledge point status is constructed to divide each of the physical knowledge points into a stable mastery state, a potential forgetting state, a group hotspot state, or a high-risk decline state. Calculate the time span between the current system time and the preset assessment node, and generate assessment time distance parameters; Based on the assessment time distance parameter, a dynamic scheduling function is constructed to perform real-time weight switching calculation on the individual forgetting channel weight corresponding to the individual forgetting probability and the group cognition channel weight corresponding to the knowledge point state hierarchical model, and to generate a comprehensive slippage index for each knowledge point. Based on the comprehensive landslide index and the knowledge point status classification results, a matching hierarchical exercise set is retrieved from the preset exercise bank, sorted and output, and a target personalized physics exercise set is generated.

2. The personalized physics exercise recommendation method based on big data analysis according to claim 1, characterized in that, The time-series alignment processing of the historical answer record data, the review frequency data, and the group's real-time answer performance data includes: Extract answer timestamps and accuracy features from historical answer records, combine them with review interval duration from review frequency data, calculate initial memory feature values ​​using a preset memory decay function, and generate an individual knowledge memory intensity sequence indexed by time. Extract the error rate fluctuation, average answering time deviation, and answering consistency probability of each knowledge point in the real-time answering performance data of the group, and construct the group knowledge point difficulty fluctuation sequence. Linear interpolation was used to align the sampling frequencies of the individual knowledge memory intensity sequence and the group knowledge point difficulty fluctuation sequence to ensure that the two sequences have corresponding data observation points under the same teaching time axis.

3. The personalized physics exercise recommendation method based on big data analysis according to claim 1, characterized in that, The specific steps for inputting the individual knowledge memory intensity sequence into the pre-trained Transformer model for modeling include: The individual knowledge memory intensity sequence is converted into a multi-dimensional embedding vector, and positional encoding is added to represent the teaching progress cycle; The memory strength correlation weights at different time points are calculated through the multi-head self-attention mechanism layer of the Transformer model, and the individual memory hidden state vector is output. The hidden state vector is input into the prediction fully connected layer, and the individual forgetting probability of each knowledge point in the next physical cycle is calculated by the logistic regression function. Perform a first-order difference operation on the individual forgetting probability sequence to calculate the change in the forgetting rate of each knowledge point under the current teaching step size, which serves as the performance change gradient.

4. The personalized physics exercise recommendation method based on big data analysis according to claim 1, characterized in that, The construction of the knowledge point state hierarchical model includes: Calculate the average error rate and standard deviation of each knowledge point in the real-time answer performance data of the group to generate a group cognitive benchmark; If the group error rate of the target knowledge point is lower than the preset threshold and the performance change gradient is stable, it is judged to be in a stable mastery state. If the individual forgetting probability exceeds a preset threshold and the group error rate is within the normal range, it is determined to be a potential forgetting state. If the group error rate increases by a step within the preset sliding window, it is determined to be a group hotspot state. If both the criteria for determining a hotspot state in a group and the criteria for an increased probability of forgetting an individual are met simultaneously, then it is determined to be a high-risk landslide state.

5. The personalized physics exercise recommendation method based on big data analysis according to claim 1, characterized in that, The specific steps for generating the comprehensive landslide index for each knowledge point are as follows: Substitute the assessment time distance parameter into the preset S-shaped growth curve function to calculate the group weight adjustment factor that increases non-linearly as the exam approaches. Establish individual forgetting channel weights Group cognitive channel weights The constraint relationship makes ; When the assessment time distance parameter exceeds the preset long-term learning threshold, Set as the primary weight; When the assessment time distance parameter is less than or equal to the pre-exam sprint threshold, the group weight adjustment factor is increased. This enables dynamic switching of weights to the group channel; The probability of individual forgetting and Multiply, the population error rate distribution score and Multiply them and sum them to get the comprehensive landslide index.

6. The personalized physics exercise recommendation method based on big data analysis according to claim 1, characterized in that, The specific steps for retrieving and sorting the matching hierarchical exercise set from the preset exercise bank are as follows: Extract knowledge point tags, difficulty level tags, and common thinking error tags from the exercise data; Basic reinforcement exercises are matched with knowledge points in the potential forgetting state; typical and difficult exercises are matched with knowledge points in the group hot state; and comprehensive and variant exercises are matched with knowledge points in the high-risk decline state. Exercises selected by sorting in descending order of comprehensive landslide index; If the error label of a problem matches the historical error characteristics of the target student, the priority of that problem in the personalized problem set will be increased.

7. The personalized physics exercise recommendation method based on big data analysis according to claim 1, characterized in that, Before obtaining real-time group response performance data, the following is also included: The system monitors the original answer log packets submitted by each student's mobile terminal in real time. The log packets contain unique identifiers for questions, answer option identifiers, answer duration, and answer device environment data. The log packets are subjected to structured parsing and noise reduction to remove invalid records whose response time is below the abnormal threshold; Based on the physics knowledge graph, log packages are mapped to corresponding physics topics, and aggregated to generate real-time group answer performance data.

8. The personalized physics exercise recommendation method based on big data analysis according to claim 1, characterized in that, After generating a personalized exercise set, the following is also included: Data on the target students' secondary responses to the exercise set is obtained through a pre-set feedback monitoring loop; Calculate the predicted residual value of the second response data relative to the individual's forgetting probability; The residual values ​​are used as backpropagation signals input into the Transformer model, and the model weights are fine-tuned online using the gradient descent algorithm.

9. A personalized physics exercise recommendation system based on big data analysis, which is implemented based on the personalized physics exercise recommendation method based on big data analysis as described in any one of claims 1-8.

10. A computer execution device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the personalized physics exercise recommendation method based on big data analysis as described in any one of claims 1-8.

Images