Study performance precise teaching management method based on time sequence behavior modeling

Abstract

The invention provides a learning performance precise teaching management method based on time sequence behavior modeling. The learning performance precise teaching management method comprises the following steps: S1, carrying out multi-modal learning behavior data acquisition; s2, aligning and segmenting the time sequence data; s3, performing knowledge retention rate differential modeling; s4, carrying out time sequence feature extraction; s5, constructing a mixed time sequence model; s6, performing dynamic learning ability evaluation; s7, carrying out adaptive resource recommendation; s8, carrying out cross-correction knowledge diffusion optimization; the method has the following advantages: the method is real-time and accurate; the data is acquired to acquire the intervened closed-loop delay lt; compared with the traditional method, the time is increased by 40 times. And cost optimization: the workload of teachers is reduced by 58% and the resource purchase cost is reduced by 42% through an automation strategy. The scale effect is that federal learning supports ten-thousand-person-level concurrence, and the model updating period is shortened to the hour level from the quarter level. Education fairness: cross-school knowledge diffusion enables the superior rate of weak schools to be improved by 29%.

Description

Technical Field

[0001] The present invention specifically relates to a method for precise teaching management of learning performance based on temporal behavior modeling. Background Art

[0002] The current education field generally adopts periodic assessments, such as monthly exams, midterm exams and standardized teaching models, and relies on teachers' experience for intervention. The existing teaching management system has the following core problems: Lag: Teaching strategies are adjusted only through data at discrete time nodes, such as test scores, and the dynamic changes in the learning process cannot be captured. Homogeneous intervention: Strategies are formulated based on the average performance of the group, ignoring the differences in individual behavior patterns. Resource mismatch: The allocation of teaching resources is out of touch with the actual needs of students, resulting in inefficient investment. Uncontrolled knowledge forgetting: The law of knowledge retention rate decay has not been quantified, and weak links cannot be strengthened in a targeted manner. In summary, this application proposes a precise teaching management method for learning performance based on temporal behavior modeling to solve the above problems. Summary of the Invention

[0003] The purpose of the present invention is to address the deficiencies in the existing technology and provide a precise teaching management method for learning performance based on temporal behavior modeling. The precise teaching management method for learning performance based on temporal behavior modeling can well solve the above problems.

[0004] To achieve the above requirements, the technical solution adopted by the present invention is to provide a method for precise teaching management of learning performance based on temporal behavior modeling, which includes the following steps:

[0005] S1: The steps for collecting multimodal learning behavior data include deploying a multi-dimensional data collection system to cover the physical layer data, cognitive layer data, and physiological layer data of learning behavior;

[0006] S2: The step of aligning and segmenting time series data, using dynamic time warping to flexibly match multi-sensor time series, and finding the minimum distance path through dynamic programming to align the two series into a unified time axis;

[0007] S3: The steps for differential modeling of knowledge retention rate are as follows:

[0008]

[0009] in:

[0010] R(t): knowledge retention rate at time t, 0≤R≤1;

[0011] α: natural forgetting rate, individual difference parameter, fitted by historical data;

[0012] β: nonlinear coefficient of forgetting, β>1 indicates accelerated forgetting, β<1 indicates anti-forgetting;

[0013] γ: intervention intensity coefficient;

[0014] I(t): external intervention input function;

[0015] S4: Steps for extracting temporal features.

[0016] S5: Steps for constructing a hybrid time series model, specifically including:

[0017] Long-range dependencies are handled through the Transformer encoder, with 8 heads and 512 attention dimensions;

[0018] Process local patterns through the TCN branch, with kernel size = 3 and dilation factor = 2;

[0019] S6: Steps for dynamic learning ability assessment, specifically using the following formula:

[0020] VaR α (R) = inf{t: Pr(R(t)<R threshold )≤α};

[0021] in:

[0022] R threshold =0.7, which is the risk threshold;

[0023] VaR α (R) is the α-quantile risk value of the knowledge retention rate R, which represents the minimum time for the student's knowledge retention rate to fall below the safety threshold for the first time under the confidence level (1-α);

[0024] α is the risk tolerance. The smaller α is, the stricter the threshold time for triggering intervention is.

[0025] R(t) is the knowledge retention rate at time t;

[0026] R threshold It is a dynamic risk threshold, set by the teaching objectives and course difficulty;

[0027] Pr( ) is the probability measure;

[0028] S7: The step of performing adaptive resource recommendation, constructing subject knowledge points into a weighted network, where nodes represent knowledge points and edge weights are migration probabilities;

[0029] S8: Steps for optimizing cross-school knowledge diffusion. Each school trains a local model and uploads gradient updates to the server. The server adds Gaussian noise when aggregating gradients and uses the CORAL algorithm to align feature distributions across different schools. The loss function is:

[0030] where ∑ s ,∑ t is the covariance matrix between the source domain and the target domain;

[0031] The global model is synchronized once a week, and the data weight of newly added schools is determined by the course matching degree.

[0032] The advantages of this precise teaching management method for learning performance based on temporal behavior modeling are as follows:

[0033] Real-time and accurate: The closed-loop latency from data collection to intervention is less than 15 minutes, a 40-fold improvement compared to traditional methods. Cost optimization: Automation strategies reduce teacher workload by 58% and resource procurement costs by 42%. Scale: Federated learning supports tens of thousands of concurrent users, shortening model update cycles from quarters to hours. Educational equity: Cross-school knowledge diffusion has increased the excellence rate of disadvantaged schools by 29 percentage points. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The same reference numerals are used in these drawings to represent the same or similar parts. The exemplary embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 The flowchart of a method for precise teaching management of learning performance based on temporal behavior modeling according to one embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0036] In order to make the objectives, technical solutions and advantages of this application clearer, this application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] In the following description, references to "one embodiment," "an embodiment," "an example," "an example," etc. indicate that the embodiment or example described may include certain features, structures, characteristics, properties, elements, or limitations, but not every embodiment or example necessarily includes the certain features, structures, characteristics, properties, elements, or limitations. In addition, repeated use of the phrase "according to one embodiment of the present application" may refer to the same embodiment, but does not necessarily refer to the same embodiment.

[0038] For the sake of simplicity, certain technical features well known to those skilled in the art are omitted in the following description.

[0039] According to one embodiment of the present application, a method for accurate teaching management of learning performance based on temporal behavior modeling is provided, such as Figure 1 As shown, the following steps are included:

[0040] S1: The steps for collecting multimodal learning behavior data include deploying a multi-dimensional data collection system to cover the physical layer data, cognitive layer data, and physiological layer data of learning behavior;

[0041] S2: Align and segment time series data. Dynamic Time Warping (DTW) is used to flexibly match multi-sensor time series. For example, when a student is solving a math problem, the writing time recorded by the smart pen is 32 seconds, while the head posture change time captured by the camera is 28 seconds. DTW uses dynamic programming to find the minimum distance path and align the two sequences to a unified time axis.

[0042] S3: The steps for differential modeling of knowledge retention rate are as follows:

[0043]

[0044] in:

[0045] R(t): knowledge retention rate at time t (0≤R≤1);

[0046] α: natural forgetting rate (individual difference parameter, fitted by historical data);

[0047] β: nonlinear coefficient of forgetting (β>1 indicates accelerated forgetting, β<1 indicates anti-forgetting);

[0048] γ: intervention intensity coefficient (unit: intervention events / hour);

[0049] I(t): external intervention input function (e.g., review reminder intensity ∈ [0, 1]);

[0050] S4: Steps for extracting temporal features.

[0051] S5: Steps for constructing a hybrid time series model, specifically including:

[0052] Long-range dependencies (such as cross-week learning trends) are handled through the Transformer encoder, with 8 heads and 512 attention dimensions;

[0053] Process local patterns (e.g., today’s attention fluctuations) through the TCN branch, with kernel size = 3 and dilation factor = 2;

[0054] S6: Steps for dynamic learning ability assessment, specifically using the following formula:

[0055] VaR α (R) = inf{t: Or(R(t) < R threshold )≤α};

[0056] in:

[0057] R threshold =0.7, which is the risk threshold;

[0058] VaR α (R) is the α-quantile risk value of the knowledge retention rate R, which indicates the minimum time for the student's knowledge retention rate to fall below the safety threshold for the first time under the confidence level (1-α). For example, when α = 0.05, VaR0.05(R) = 72 hours, which means that there is a 95% probability that the retention rate will not fall below the threshold within 72 hours;

[0059] α is the risk tolerance (significance level), ranging from (0, 1), usually set to 0.05 (5%) or 0.01 (1%). The smaller α is, the stricter the threshold time for triggering intervention;

[0060] R(t) is the knowledge retention rate at time t. For example, R(48h) = 0.75 means that the student’s mastery of a certain knowledge point is 75% after 48 hours.

[0061] R threshold It is a dynamic risk threshold, which is set by the teaching objectives and course difficulty. For example, if the course requires a retention rate of no less than 80% within 72 hours, then Rthreshold = 0.8;

[0062] Pr() is a probability measure (based on Monte Carlo simulation). 200 retention rate prediction paths are generated through step 6.2, and the statistics satisfy R(t) <R threshold The path ratio.

[0063] S7: Steps for adaptive resource recommendation: Constructing subject knowledge points into a weighted network (TransE algorithm embedding), with nodes representing knowledge points and edge weights representing transition probabilities. For example, the transition probability from "trigonometric function" to "Fourier series" is 0.72; define the state space S = {R(t), HRV, EDA}, the action space A = {video explanation, variant question, environment adjustment}, and the reward function: R = ω1ΔR + ω2 + ω3, where ω1 = 0.5, ω2 = 0.3, and ω3 = 0.2, trained using inverse reinforcement learning;

[0064] Example:

[0065] Assume that a student's parameters are as follows:

[0066] R(t) decay curve over time: R(24h)=0.85, R(48h)=0.72, R(72h)=0.65

[0067] Set α = 0.1, Rthreshold = 0.7

[0068] VaR 0.1 The calculation process of (R) is:

[0069] A Monte Carlo simulation generates 200 R(t) paths, counting the time ti at which each path first falls below 0.7. The 10% quantile of all ti is calculated (i.e., inf{t:10% path ≤ t}), assuming t = 52 hours. Conclusion: This student has a 10% probability of having a retention rate below 0.7 within 52 hours, triggering intensive intervention.

[0070] S8: Steps for optimizing cross-school knowledge diffusion. Each school trains a local model (with encrypted parameters) and uploads gradient updates to the server. The server adds Gaussian noise (σ = 0.1) when aggregating gradients to satisfy ∈-differential privacy. The CORAL (covariance difference minimization) algorithm is used to align the feature distributions of different schools. The loss function is:

[0071] where ∑ s ,∑ t is the covariance matrix between the source domain and the target domain;

[0072] The global model is synchronized once a week, and the data weight of newly added schools is determined by course matching (cosine similarity).

[0073] According to one embodiment of the present application, the physical layer data in step S1 of the method for precise teaching management of learning performance based on temporal behavior modeling is recorded by a smart pen (such as Livescribe) to record writing trajectory, pen pressure, and problem-solving time with an accuracy of up to 0.1 mm; a camera (such as Tobii Pro) captures eye movement trajectory, and the gaze heat map and pupil diameter change are calculated using the OpenCV algorithm with a sampling frequency of 30 Hz;

[0074] The cognitive layer data is recorded through the learning platform, including click streams (mouse movement trajectory, page dwell time), question interaction logs (such as abandonment rate, number of attempts at variant questions), and natural language processing (BERT model) to analyze the semantic coherence of text-based answer content.

[0075] The physiological layer data is monitored by wearable devices (such as Apple Watch) to include heart rate variability (HRV), skin conductance (EDA), and blood oxygen saturation (SpO2). The data is denoised by Kalman filtering and aligned with the learning behavior timestamp.

[0076] According to one embodiment of the present application, in step S2 of the method for precise teaching management of learning performance based on temporal behavior modeling, data is cut based on a sliding window (window length = 2 days, step length = 12 hours) when segmenting, and each window must meet the following requirements: at least 3 exercises on similar knowledge points (such as quadratic functions), completeness of key events (such as completion of unit tests), and stability of physiological signals (HRV coefficient of variation <0.2).

[0077] According to one embodiment of the present application, in step S3 of the method for precise teaching management of learning performance based on temporal behavior modeling, maximum likelihood estimation (MLE) is used to fit historical data, assuming that R(t) obeys Beta distribution, to construct a likelihood function:

[0078]

[0079] Parameters are optimized by gradient descent method, and the convergence condition is KL divergence < 0.01;

[0080] Given the value of the intervention function I(t): 1 if a review reminder is pushed, 0 otherwise, calculate the change in retention rate under different intervention intensities. For example, when γ = 0.5, the retention rate decay rate decreases by 32%;

[0081] According to the steady-state solution of the differential equation Dynamically adjust the review frequency. If the target retention rate R>0.8, then γ>α(0.8) must be satisfied. -β .

[0082] According to one embodiment of the present application, step S4 of the learning performance precision teaching management method based on temporal behavior modeling specifically includes:

[0083] Time Domain Features: Calculate the skewness and kurtosis of problem-solving time to measure the degree of deviation in behavior distribution. For example, a skewness > 1 indicates that students tend to cram.

[0084] Frequency domain features: Perform Fourier transform on the attention signal (EDA) to extract the energy ratio of delta wave (0.5-4Hz) and beta wave (14-30Hz), reflecting fatigue and alertness;

[0085] Time series pattern mining: Use LSTM autoencoders to compress time series data, extract latent state vectors (dimension = 64), and classify learning style types (such as "impulsive" or "reflective") through clustering (DBSCAN).

[0086] According to one embodiment of the present application, step S5 of the learning performance precision teaching management method based on temporal behavior modeling includes a multi-task learning step specifically including a main task, auxiliary task 1 and auxiliary task 2, wherein:

[0087] Main task: knowledge mastery prediction (MAE loss);

[0088] Auxiliary task 1: Attention decay curve reconstruction (MSE loss);

[0089] Auxiliary task 2: intervention response classification (cross entropy loss);

[0090] In this step, a dynamic weighting mechanism is added to adjust the loss weights λ1, λ2, and λ3 according to the course stage. For example, λ1 = 0.7 for the new course stage and λ1 = 0.3 for the review stage.

[0091] According to one embodiment of the present application, the learning performance precision teaching management method based on temporal behavior modeling further includes the step of generating a personalized intervention strategy, specifically including cognitive intervention, emotional intervention, and environmental intervention;

[0092] Cognitive intervention specifically involves generating a set of variant questions based on knowledge graph vulnerabilities, for example:

[0093] Original question: Solve the equation x²+2x+1=0

[0094] Variation 1: Given (x + a)² = b, find the relationship between a and b.

[0095] Variation 2: Construct a quadratic equation whose roots are the squares of the solutions to the original problem;

[0096] Emotional intervention specifically involves judging emotional states through physiological signal classification (SVM model):

[0097] Anxiety (HRV < 50ms): Push motivational short videos (< 30 seconds, including dopamine-triggering music);

[0098] Boredom (EDA < 0.1 μS): insert interactive gamification exercises;

[0099] Environmental intervention involves calling the IoT device API to adjust screen parameters based on eye tracking data:

[0100] When gaze is distracted: increase contrast (ΔL*=10)

[0101] Long-term staring: Automatically insert a 20-second gaze reminder.

[0102] The above-described embodiments merely represent several implementations of the present invention. While the descriptions are relatively specific and detailed, they are not to be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present invention, and such modifications and improvements fall within the scope of the present invention. Therefore, the scope of the present invention shall be determined by the claims.

Claims

1. A learning performance precision teaching management method based on temporal behavior modeling, characterized by: The steps include: S1: The steps for collecting multimodal learning behavior data include deploying a multi-dimensional data collection system to cover the physical layer data, cognitive layer data, and physiological layer data of learning behavior; S2: The step of aligning and segmenting time series data, using dynamic time warping to flexibly match multi-sensor time series, and finding the minimum distance path through dynamic programming to align the two series into a unified time axis; S3: The steps for differential modeling of knowledge retention rate are as follows: in: R(t): knowledge retention rate at time t, 0≤R≤1; α: natural forgetting rate, individual difference parameter, fitted by historical data; β: nonlinear coefficient of forgetting, β>1 indicates accelerated forgetting, β<1 indicates anti-forgetting; γ: intervention intensity coefficient; I(t): external intervention input function; S4: Steps for extracting temporal features. S5: Steps for constructing a hybrid time series model, specifically including: Long-range dependencies are handled through the Transformer encoder, with 8 heads and 512 attention dimensions; Process local patterns through the TCN branch, with kernel size = 3 and dilation factor = 2; S6: Steps for dynamic learning ability assessment, specifically using the following formula: VaR α (R)=inf{t:Pr(R(t) <R threshold )≤α}; in: R threshold =0.7, which is the risk threshold; VaR α (R) is the α-quantile risk value of the knowledge retention rate R, which represents the minimum time for the student's knowledge retention rate to fall below the safety threshold for the first time under the confidence level (1-α); α is the risk tolerance. The smaller α is, the stricter the threshold time for triggering intervention is. R(t) is the knowledge retention rate at time t; R threshold It is a dynamic risk threshold, set by the teaching objectives and course difficulty; Pr() is the probability measure; S7: The step of performing adaptive resource recommendation, constructing subject knowledge points into a weighted network, where nodes represent knowledge points and edge weights are migration probabilities; S8: Steps for optimizing cross-school knowledge diffusion. Each school trains a local model and uploads gradient updates to the server. The server adds Gaussian noise when aggregating gradients and uses the CORAL algorithm to align feature distributions across different schools. The loss function is: where ∑ s ,∑ t is the covariance matrix between the source domain and the target domain; The global model is synchronized once a week, and the data weight of newly added schools is determined by the course matching degree.

2. The method for precise teaching management of learning performance based on temporal behavior modeling according to claim 1, characterized in that: The physical layer data in step S1 is recorded by the smart pen, including writing trajectory, pen pressure and problem-solving time, with an accuracy of up to 0.1 mm. The camera captures eye movement trajectories and calculates gaze heat maps and pupil diameter changes using the OpenCV algorithm. The sampling frequency is 30Hz. The cognitive layer data records click streams and question interaction logs through the learning platform, and uses natural language processing to analyze the semantic coherence of text-based answer content; The physiological layer data is monitored by wearable devices, including heart rate variability (HRV), skin conductance, and blood oxygen saturation. The data is denoised by Kalman filtering and aligned with the learning behavior timestamp.

3. The learning performance precision teaching management method based on temporal behavior modeling according to claim 1 is characterized in that: When segmenting in step S2, the data is cut based on a sliding window, and each window must meet the following requirements: at least 3 exercises of the same knowledge point, completeness of key events, and stability of physiological signals.

4. The method for precise teaching management of learning performance based on temporal behavior modeling according to claim 1, characterized in that: In step S3, maximum likelihood estimation is used to fit historical data. Assuming that R(t) obeys Beta distribution, the likelihood function is constructed: Parameters are optimized by gradient descent method, and the convergence condition is KL divergence < 0.01; Given the value of the intervention function I(t): 1 if a review reminder is pushed, 0 otherwise, calculate the change in retention rate under different intervention intensities; According to the steady-state solution of the differential equation Dynamically adjust the review frequency. If the target retention rate R>0.8, then γ>α(0.8) must be satisfied. -β .

5. The learning performance precision teaching management method based on temporal behavior modeling according to claim 1 is characterized in that: Step S4 specifically includes: Time domain characteristics: Calculate the skewness and kurtosis of the problem-solving time to measure the degree of deviation of the behavior distribution. For example, a skewness > 1 indicates that students tend to cram; Frequency domain features: Perform Fourier transform on the attention signal to extract the energy ratio of delta waves and beta waves, reflecting fatigue and alertness states; Time series pattern mining: Use LSTM autoencoders to compress time series data, extract latent state vectors, and learn style types through clustering.

6. The method for precise teaching management of learning performance based on temporal behavior modeling according to claim 1, characterized in that: Step S5 includes the steps of multi-task learning, specifically including the main task, auxiliary task 1 and auxiliary task 2, wherein: Main task: knowledge mastery prediction; Auxiliary task 1: Attention decay curve reconstruction; Auxiliary Task 2: Intervention Response Classification; A dynamic weighting mechanism is added in this step to adjust the loss weights λ1, λ2, λ3 according to the course stage.

7. The learning performance precision teaching management method based on temporal behavior modeling according to claim 1 is characterized in that: It also includes the steps of generating personalized intervention strategies, including cognitive intervention, emotional intervention and environmental intervention. Cognitive intervention specifically involves generating a set of variant questions based on knowledge graph vulnerabilities. Emotional intervention specifically involves judging emotional states through physiological signal classification: For anxiety, push inspirational short videos. Boredom: Insert interactive gamified exercises.

Images