Personalized development guidance and resource matching system based on student digital portrait

Abstract

The application discloses a personalized development guidance and resource matching system based on a student digital portrait and relates to the technical field of intelligent education. The system comprises a student digital portrait construction module for constructing a dynamic state vector of a student in a multi-dimensional feature space; wherein the multi-dimensional feature space at least contains an ability dimension, an interest dimension and a learning style dimension; and an educational resource semantic graph module for constructing resource feature vectors of educational resources in the multi-dimensional feature space and a semantic association network among the resources. The application establishes a dynamic balance mechanism of individual development efficiency and group fairness, fundamentally solves the adaptability trap and stereotype reinforcement problem caused by the traditional recommendation system taking the maximum matching degree as a single target, ensures that students from different backgrounds can obtain personalized paths considering the current foundation and development potential, and makes the algorithm a real promoter of educational fairness rather than a replicator of bias.

Description

Technical Field

[0001] This invention relates to the field of smart education technology, specifically to a personalized development guidance and resource matching system based on student digital profiles. Background Technology

[0002] With the deepening of digital transformation in education, personalized recommendation systems based on student digital profiles have been widely used to improve the efficiency of educational resource allocation. Currently, mainstream personalized recommendation methods mainly rely on the analysis of superficial indicators such as student academic performance, behavioral tags, and historical success rates. Through collaborative filtering or tag matching algorithms, they assess the direct matching degree between resources and the student's current state. While these methods can achieve rapid resource matching, their decision-making logic has a fundamental flaw: the system takes maximizing matching degree or success rate as its single optimization objective, lacking the ability to model the dynamic balance of multiple objectives. This leads to an unintentional deep conflict between individual development efficiency and group fairness in the pursuit of precise matching.

[0003] The flaws of this single-objective optimization manifest in several ways: the system tends to keep students on their historical success paths, continuously recommending resources highly aligned with their existing abilities, thus creating an adaptation trap. For example, for a student from an area with limited educational resources but exhibiting high resilience, the system, based on their past academic data, will continuously recommend foundational courses rather than opening new avenues for challenging leaps based on their resilience. This recommendation is not semantically guiding, but rather a path dependency based on historical data, implicitly solidifying the student's initial disadvantages. Simultaneously, the system fails to recognize the potential amplification of biases at the group level: when structural biases already exist in historical data, single-objective optimization unconsciously learns and reinforces these biases, leading to the implicit guidance of specific groups (such as girls or students from underdeveloped areas) to specific types of resources, further compressing their developmental niche.

[0004] Therefore, the core deficiency of existing technologies lies in the lack of a multi-objective dynamic decision-making capability that can establish a quantifiable balance mechanism between individual development benefits and group fairness constraints, and the inability to effectively model and warn of potential adaptation traps and stereotype reinforcement risks during the decision-making process. Summary of the Invention

[0005] The purpose of this invention is to provide a personalized development guidance and resource matching system based on student digital profiles to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a personalized development guidance and resource matching system based on student digital profiles, comprising:

[0007] The student digital profile construction module is used to construct a dynamic state vector of a student in a multi-dimensional feature space; wherein, the multi-dimensional feature space includes at least an ability dimension, an interest dimension, and a learning style dimension.

[0008] The educational resource semantic graph module is used to construct resource feature vectors and semantic association networks between resources in the multi-dimensional feature space.

[0009] A multi-objective dynamic decision engine connects the student digital profile construction module and the educational resource semantic graph module. It is used to generate personalized development paths based on the dynamic state vector, the resource feature vector, and the semantic association network by simulating the motion trajectory in a virtual force field.

[0010] The path stability monitoring unit is connected to the multi-objective dynamic decision engine and is used to calculate the mechanical stability index based on the personalized development path.

[0011] The decision parameter adaptive adjustment unit is connected to the path stability monitoring unit and the multi-objective dynamic decision engine, and is used to trigger and execute adaptive adjustment of the decision model parameters in the multi-objective dynamic decision engine according to the mechanical stability index.

[0012] In a preferred embodiment, the student digital profile construction module specifically includes:

[0013] The data preprocessing submodule is used to clean, align, and normalize multi-source heterogeneous time-series data from academic records, behavioral logs, and psychological assessments.

[0014] The vectorization encoding submodule is used to convert the processed data into fixed-dimensional feature vectors through embedding layers or feature engineering.

[0015] The State Fusion and Update submodule is used to model the feature vector sequence using a temporal neural network model and output the dynamic state vector that integrates historical and current information.

[0016] In a preferred embodiment, the educational resource semantic graph module specifically includes:

[0017] The resource feature extraction submodule is used to parse the text, outline and metadata of resources using natural language processing technology, and extract the knowledge points, skill points, difficulty level and teaching style features involved to form an initial resource feature vector;

[0018] The graph construction submodule is used to construct a semantic association network with resources as nodes and relationships as edges, based on the prerequisite and successive relationships of resources in terms of content and the complementary relationships of teaching objectives.

[0019] The dynamic weight assignment submodule is used to dynamically adjust the weights of edges and the inherent attribute values ​​of resource nodes in the semantic association network based on the historical success rate of resource usage, student evaluation feedback, and real-time popularity information.

[0020] In a preferred embodiment, the multi-objective dynamic decision engine specifically includes:

[0021] The force field calculation submodule is used to obtain the objective function obtained by the adaptive adjustment unit based on the decision parameters, acquire the candidate resource feature vector from the educational resource semantic graph module, and calculate the virtual force.

[0022] The first type of function calculates the attractiveness based on the cosine similarity between the dynamic state vector and the candidate resource feature vector;

[0023] The second type of function calculates the moderating force based on the fairness requirements of the distribution of resources among the student population, wherein the fairness requirements are measured by the Gini coefficient;

[0024] The trajectory simulation submodule is used to treat the dynamic state vector as a point mass in a multi-dimensional feature space, to perform vector synthesis of the attraction and adjustment forces to obtain the resultant force, and to solve the motion equation of the point mass under the action of the resultant force through a numerical integration algorithm, and output the trajectory point sequence.

[0025] The path mapping submodule is used to map the trajectory point sequence back to a specific educational resource node sequence, forming an executable personalized development path.

[0026] In a preferred embodiment, the path stability monitoring unit specifically includes:

[0027] The indicator calculation submodule is used to calculate the mechanical indicators of the personalized development path, including:

[0028] The eccentricity rate reflects whether the path direction is focused on the core capability area; the rate of change of the radius of curvature reflects whether the capability transition between adjacent resource nodes is smooth; and the precession angular velocity reflects whether the long-term development trend has drifted.

[0029] The anomaly diagnosis submodule is used to compare the calculated mechanical indicators with the preset normal threshold range. The default judgment period for the threshold is one week. When any indicator exceeds the threshold for three consecutive periods or undergoes a sudden change, it is judged as path instability and an anomaly diagnosis report is generated.

[0030] In a preferred embodiment, the decision parameter adaptive adjustment unit specifically includes:

[0031] The parameter optimization modeling submodule is used to construct an optimization model for the force field function weights in the multi-objective dynamic decision engine, with the core objective of the mechanical stability index returning to the normal range when the path is determined to be unstable.

[0032] The adaptive parameter tuning submodule is used to solve the optimization model in the historical parameter-effect database using Bayesian optimization or gradient descent algorithm, to obtain a new set of optimal parameter combinations, and provide them as preset values ​​of the objective function to the multi-objective dynamic decision engine.

[0033] The parameter hot update submodule is used to safely update the new optimal parameter combination into the multi-objective dynamic decision engine and monitor the adjustment effect during a subsequent preset observation period.

[0034] In a preferred embodiment, the system further includes an educator collaborative intervention module, which comprises:

[0035] The instruction parsing interface is used to receive instructions from educators to confirm, modify, or reject the personalized development path generated by the system.

[0036] The intervention vectorization submodule is used to quantize the modification or rejection instruction into a directional perturbation vector that can be superimposed on the current student dynamic state vector.

[0037] An influence attenuation controller is used to set an intensity coefficient for the directional disturbance vector that decays exponentially with time or decision cycle, so that its influence on the system decision is reduced to less than 10% of the initial intensity after five consecutive cycles, until normal automatic decision-making is restored.

[0038] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0039] This invention fundamentally solves the adaptation trap and stereotype reinforcement problems caused by traditional recommendation systems that focus solely on maximizing matching degree by establishing a dynamic balance mechanism between individual development efficiency and group fairness. It ensures that students from different backgrounds can obtain personalized paths that take into account their current foundation and development potential, making the algorithm a true promoter of educational equity rather than a copyist of prejudice.

[0040] This invention, through closed-loop adaptive optimization and human-machine collaboration mechanisms, enables the system to continuously adjust its decision-making logic based on student development feedback. While incorporating expert experience, it avoids the solidification of human bias, significantly improving the system's long-term stability and autonomous evolution capabilities. This provides a complete technical guarantee for achieving a balance between personalization and fairness in student development guidance. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0042] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Example 1, please refer to Figure 1 As shown in this embodiment, the personalized development guidance and resource matching system based on student digital profiles includes:

[0045] The student digital profile construction module is used to construct a dynamic state vector of a student in a multi-dimensional feature space; wherein, the multi-dimensional feature space includes at least an ability dimension, an interest dimension, and a learning style dimension.

[0046] The educational resource semantic graph module is used to construct resource feature vectors and semantic association networks between resources in the multi-dimensional feature space.

[0047] A multi-objective dynamic decision engine connects the student digital profile construction module and the educational resource semantic graph module. It is used to generate personalized development paths based on the dynamic state vector, the resource feature vector, and the semantic association network by simulating the motion trajectory in a virtual force field.

[0048] The path stability monitoring unit is connected to the multi-objective dynamic decision engine and is used to calculate the mechanical stability index based on the personalized development path.

[0049] The decision parameter adaptive adjustment unit is connected to the path stability monitoring unit and the multi-objective dynamic decision engine, and is used to trigger and execute adaptive adjustment of the decision model parameters in the multi-objective dynamic decision engine according to the mechanical stability index.

[0050] The educator collaborative intervention module is connected to the multi-objective dynamic decision engine and is used to receive intervention instructions from educators and integrate them into the system decision.

[0051] As an optional implementation, the student digital profile construction module specifically includes:

[0052] The data preprocessing submodule is used to clean, align, and normalize multi-source heterogeneous time-series data from academic records, behavioral logs, and psychological assessments.

[0053] The academic records include grades in each subject, homework completion status, and exam rankings; the behavior log includes online learning time, video viewing progress, and forum interaction frequency.

[0054] Psychological assessments include personality tests, learning motivation questionnaires, etc. These data may be collected at different frequencies (e.g., monthly exam scores once a month, daily behavior logs), and may contain missing or outlier values.

[0055] The data preprocessing submodule fills in missing values ​​using interpolation methods (such as linear interpolation or cubic spline interpolation), removes outliers using the 3σ principle, and synchronizes multi-source data to a unified time axis (e.g., in days) using timestamp alignment.

[0056] Finally, all data were scaled to the [0, 1] interval using the min-max normalization method to eliminate the influence of dimensions.

[0057] The vectorization encoding submodule is used to convert processed data into fixed-dimensional feature vectors through embedding layers or feature engineering.

[0058] For discrete data (such as course selection, gender), one-hot encoding is used. If there are many categories, an embedding layer can be used to map them into low-dimensional dense vectors (for example, mapping 50 courses into a 32-dimensional vector).

[0059] For continuous data (such as scores and duration), it is directly retained as a feature dimension. After processing, the data at each time point is encoded into a 128-dimensional feature vector.

[0060] Definition of multidimensional feature space: The multidimensional feature space consists of the ability dimension, interest dimension, and learning style dimension.

[0061] The ability dimension is a weighted composite of three indicators: standardized academic performance score, knowledge mastery (calculated by the accuracy rate of the assessment), and learning efficiency (the number of knowledge points mastered per unit of time), with weights of 0.5, 0.3, and 0.2, respectively.

[0062] The interest dimension is extracted from three indicators: course selection preference, active forum topics, and self-searched keywords, using the LDA topic model to output an interest intensity vector (20 dimensions).

[0063] The learning style dimension is obtained by k-means clustering three indicators: VARK questionnaire score (visual / auditory / reading / kinesthetic), video viewing behavior pattern (pause frequency, replay frequency), and note-taking behavior characteristics (word count, degree of structure) to obtain a membership vector (4-dimensional) of four style types.

[0064] The three dimensions are initially assumed to be orthogonal. Subsequently, the correlation is tested through principal component analysis. If a strongly correlated dimension is found (correlation coefficient > 0.7), then dimensionality reduction and fusion are performed.

[0065] The State Fusion and Update submodule is used to model the feature vector sequence using a temporal neural network model and output the dynamic state vector that integrates historical and current information.

[0066] A two-layer LSTM structure is used, with 128 hidden units in the first layer and 256 hidden units in the second layer. The dropout rate is set to 0.2 to prevent overfitting.

[0067] The time window length is set to T=30 days. The 128-dimensional feature vectors at each time point are sequentially input into the LSTM. The hidden state at the last time step is used as the dynamic state vector at the current time, with a dimension of 256.

[0068] The model is pre-trained using student data from the entire school over the past three years, and self-supervised learning is performed with the prediction of the state at the next time point as an auxiliary task.

[0069] For students currently enrolled, the state vector is updated incrementally every day at midnight based on the data from the previous 30 days.

[0070] For freshmen, the average status of senior students in the same major is used as the initialization for the first two weeks, and then the status is gradually transitioned to a personalized status after enough data is accumulated.

[0071] Optionally, an attention mechanism can be introduced to assign different weights to historical moments, highlighting the impact of key time points.

[0072] As an optional implementation, the educational resource semantic graph module specifically includes:

[0073] The resource feature extraction submodule is used to parse the text, outline and metadata of resources using natural language processing technology, extract the knowledge points, skill points, difficulty level and teaching style features involved, and form an initial resource feature vector.

[0074] For example, for a Calculus course, the BERT model is used to extract keywords from the course description text, which are then mapped to a predefined knowledge point dictionary (containing 2000 knowledge points) to obtain a knowledge point vector (a 2000-dimensional sparse vector), and then PCA is used to reduce the dimensionality to 100 dimensions.

[0075] The prerequisites are analyzed from the course syllabus to obtain the difficulty level (beginner / intermediate / advanced, coded as 1 / 2 / 3).

[0076] Teaching style characteristics (5 dimensions) such as clarity of explanation and richness of examples were extracted from student evaluations through sentiment analysis.

[0077] These features are concatenated to obtain the initial resource feature vector, which has 108 dimensions (100+3+5).

[0078] The graph construction submodule is used to construct a semantic association network with resources as nodes and relationships as edges, based on the prerequisite and successive relationships of resources in terms of content and the complementary relationships of teaching objectives.

[0079] For example, Calculus is a prerequisite for Advanced Mathematics, and a prerequisite edge is established between the two.

[0080] A matching edge is established between the Calculus video tutorial and the Calculus problem set. All nodes and edges form a directed graph.

[0081] Node attributes include resource feature vectors, and edge attributes include relation type and weight (initial weights are set based on the co-occurrence frequency between resources, ranging from [0, 1]). The graph structure uses an adjacency list for storage, supporting fast adjacency lookup.

[0082] The dynamic weight assignment submodule is used to dynamically adjust the weights of edges and the inherent attribute values ​​of resource nodes in the semantic association network based on the historical success rate of resource usage, student evaluation feedback, and real-time popularity information.

[0083] For example, if the pass rate (success rate) of a course exceeds 80%, the weight of its associated subsequent course edges will be multiplied by a coefficient of 1.2; if a video resource has more than 1,000 clicks in a week, its popularity attribute value will increase by 0.1 (maximum 1.0).

[0084] The evaluation attributes of resource nodes are updated after the student evaluation scores are weighted by time decay. The calculation formula is as follows:

[0085] ;

[0086] in, The original evaluation score. This represents the average score given by students during the current period.

[0087] The multi-objective dynamic decision engine draws on the concept of force fields in physics to model the student's development process as forced motion in a multi-dimensional feature space.

[0088] This system adopts an overdamped dynamic model. The overdamped dynamic model is a simplified model in physics describing the motion of a particle in a viscous medium. Its core feature is the neglect of inertia (assuming mass m=1), considering only the balance between damping force and resultant force (damping coefficient c=1). The purpose of using this model in this invention is to ensure that the trajectory of the student's state vector S monotonically approaches the equilibrium point, avoiding path oscillations caused by inertia. Mathematically, this model is expressed as... The rate of change of state is directly determined by the resultant force F, and its direction always points in the direction of the resultant force to ensure smooth path convergence; where S is the student's dynamic state vector, t is time, and F is the resultant force.

[0089] The mathematical basis of the overdamped dynamic model is that, neglecting the inertial term, the equation of motion of the particle simplifies to the first-order differential equation dS / dt=F(S).

[0090] Compared to the complete Newton's second law equation m·d²S / dt²+c·dS / dt=F(S), the overdamped approximation is applicable to low Reynolds number environments or high-damped scenarios.

[0091] In this invention, the rationale for this approximation lies in the fact that the student development process is characterized by low inertia and high damping—state changes are mainly driven by the current net force, and the continuous influence of historical velocities can be ignored. Mathematically, this model guarantees Lyapunov stability, meaning that the system always has a convergence direction, and the convergence trajectory is monotonic, without oscillations.

[0092] As an optional implementation, the multi-objective dynamic decision engine specifically includes:

[0093] The force field calculation submodule is used to obtain the objective function obtained by the decision parameter adaptive adjustment unit according to the decision parameter, obtain the candidate resource feature vector from the educational resource semantic graph module, and calculate the virtual force. Wherein:

[0094] The first type of function calculates the attractiveness based on the cosine similarity between the dynamic state vector and the candidate resource feature vector.

[0095] Given the student's current state vector S (256 dimensions) and the candidate resource feature vector R (108 dimensions, which needs to be mapped to 256 dimensions through a fully connected layer), the attractiveness... Defined as:

[0096] ;

[0097] Where α is the attraction coefficient (provided by the decision parameter adaptive adjustment unit, initially calibrated to 0.5), and sim(S, R) is the cosine similarity, calculated using the following formula: .

[0098] This force points in the direction of the resource, driving students to move toward that resource.

[0099] The second type of function calculates the moderating force based on the fairness requirements of resource distribution among student groups. The fairness requirements are measured using the Gini coefficient, and the student groups are cross-divided into six groups based on two dimensions: region of origin (East / Central / West) and gender (Male / Female).

[0100] The formula for calculating the proportion of resources used by each group is as follows:

[0101] ;

[0102] in, It represents the number of people in group i who use resource j. It is the total number of people in group i;

[0103] The formula for calculating the Gini coefficient G is:

[0104] ;

[0105] Where n=6 is the size of the group, and μ is the average proportion of all groups using the resource. and Let G represent the usage ratios of group i and group j respectively, and sum them over all group pairs. When G exceeds a preset threshold... When the coefficient of friction is 0.4, a repulsive force is generated. :

[0106] ;

[0107] Here, β is the adjustment coefficient (initially calibrated to 0.3). This force points away from the resource, inhibiting its excessive concentration by a specific group and promoting equitable distribution.

[0108] All group attribute data is used solely for fairness calculations and not for any other purpose, and is anonymized during storage.

[0109] Gini coefficient threshold The Gini coefficient of 0.4 is set in accordance with the internationally accepted standards of the United Nations Development Programme (UNDP) regarding the fairness of resource allocation: a Gini coefficient below 0.2 indicates absolute equality, 0.2-0.3 indicates relatively equal equality, 0.3-0.4 indicates relatively reasonable equality, 0.4-0.5 indicates a large disparity, and above 0.5 indicates a huge disparity. Therefore, using 0.4 as a warning threshold, when the Gini coefficient of the group's resource usage ratio exceeds 0.4, it indicates that significant unfairness has occurred in resource allocation, and the system needs to immediately generate a repulsive force to actively adjust and prevent further expansion of unfairness.

[0110] The trajectory simulation submodule is used to treat the dynamic state vector as a point mass in a multi-dimensional feature space, to perform vector synthesis of the attraction and adjustment forces to obtain the resultant force, and to solve the motion equation of the point mass under the action of the resultant force through a numerical integration algorithm, and output the trajectory point sequence.

[0111] combined efforts ,in, The vector sum of the attractiveness of all candidate resources (K=10 resources closest to the current state can be selected). For the vector sum of all fairness adjustment forces.

[0112] Solving differential equations using the fourth-order Runge-Kutta method Starting from the current state S0, with a step size of Δt = 1 day, the states S1, S2, ..., S30 at N = 30 future time points are iteratively calculated to form a trajectory point sequence.

[0113] Simulation results show that when the step size Δt = 1 day, all trajectories converge to the stable region within 20 steps.

[0114] The path mapping submodule is used to map the trajectory point sequence back to a specific educational resource node sequence, forming an executable personalized development path.

[0115] For each trajectory point Si, find the resource node that is closest to the point in the resource semantic network (based on Euclidean distance of feature vectors, only considering knowledge points and difficulty dimensions, ignoring dynamic attributes such as popularity), and add the node ID to the path sequence.

[0116] If multiple resources are close to each other (distance difference < 0.05), the resource with the best weight (historical success rate) is selected. The final output is an ordered list of resources, such as ["Basic Calculus", "Advanced Calculus", "Linear Algebra", ...].

[0117] The path stability monitoring unit is responsible for evaluating whether the generated path is reasonable and stable, and for promptly identifying potential problems.

[0118] As an optional implementation, the path stability monitoring unit specifically includes:

[0119] The index calculation submodule is used to calculate the mechanical indices of the personalized development path, including eccentricity e, rate of change of radius of curvature ρ', and precession angular velocity ω.

[0120] The calculation methods for each indicator are as follows:

[0121] Eccentricity e: reflects whether the path direction is focused on the core capability area.

[0122] Using the student's initial state vector S0 as the center, calculate the distance di between all path points Si and S0, and the variance. ,in This represents the average distance.

[0123] Eccentricity ;

[0124] If e > 0.7, it indicates that the path is diverging, and students may deviate from their core strengths.

[0125] The rate of change of radius of curvature ρ' reflects the smoothness of capability transitions between adjacent resource nodes. It calculates the radius of curvature of the circle defined by every three consecutive points (Si-1, Si, Si+1) along the calculation path. Then calculate its rate of change. ;

[0126] Average rate of change Exceeding the threshold of 0.5 indicates an excessively large leap in ability, which may lead to learning difficulties.

[0127] Precession angular velocity ω: reflects whether the long-term development trend has drifted. Projecting the path points onto the first two principal component spaces after dimensionality reduction by principal component analysis yields the orientation angles. (Angle with the first principal component), calculate the angle difference Δ between adjacent points. Precession angular velocity ;

[0128] If ω exceeds the threshold of 0.12 rad / step, it indicates that the student's interests or abilities change frequently and lack stability.

[0129] The threshold is determined using the following method: historical data of 1000 students who have completed four years of study are selected, and various mechanical indicators of their development path are calculated. Students who have achieved high-quality employment or pursued further studies after graduation are considered successful samples, and the 90th percentile of the distribution of indicators of the successful samples is taken as the upper limit of the threshold. The threshold can be adjusted according to the characteristics of different disciplines, and the system administrator can modify it through the configuration file.

[0130] The concept of precession angular velocity ω originates from the precession phenomenon in physics, which describes the rate of change of the direction of the rotation axis. In this invention, it is applied by analogy to student development path analysis: the projection of the student's state vector into the first two principal component spaces after PCA dimensionality reduction is regarded as a direction pointer, and the rotation speed of this pointer is the precession angular velocity.

[0131] The magnitude of ω reflects the rate of change in the direction of a student's interests or abilities: when ω is small (below the threshold of 0.12 rad / step), it indicates that the student's development path is stable, with strong focus, which is conducive to deep learning; when ω is too large, it indicates that the student's interests frequently shift or their abilities are unstable, which may lead to insufficient learning depth or scattered resource investment. This threshold of 0.12 rad / step corresponds to a directional change of approximately 7° per step, and was determined using the 90th percentile of 1000 successful samples.

[0132] The abnormality diagnosis submodule is used to compare the calculated mechanical indicators with the preset normal threshold range.

[0133] The default judgment period is one week (7 trajectory points). When any indicator exceeds the threshold for three consecutive periods or undergoes a sudden change (change exceeding 50% within a single period), it is judged as path instability and an abnormal diagnosis report is generated.

[0134] The report includes the type of instability (e.g., divergence, jump, drift), the risk level (divided into low / medium / high based on the degree to which the threshold is exceeded), and recommended measures (e.g., suggesting adjusting the objective function weights, suggesting introducing exploratory resources). This report can be pushed to educators' interfaces or stored in logs.

[0135] The adaptive adjustment unit for decision parameters is responsible for automatically optimizing decision parameters when the path becomes unstable, enabling the system to have self-healing capabilities.

[0136] As an optional implementation, the decision parameter adaptive adjustment unit specifically includes:

[0137] The parameter optimization modeling submodule is used to construct an optimization model of the force field function weights (α, β) in the multi-objective dynamic decision engine when the path is determined to be unstable, with the core objective of the mechanical stability index returning to the normal range.

[0138] The objective function is set as follows:

[0139] ;

[0140] in, e is a preset weighting coefficient (all taken as 1 / 3). ω represents the currently measured eccentricity, the rate of change of the mean radius of curvature, and the precession angular velocity, respectively. The constraints are α∈[0.1, 2.0], β∈[0, 1.0].

[0141] The adaptive parameter tuning submodule is used to solve the optimization model using a Bayesian optimization algorithm based on a historical parameter-effect database. It initially sets 10 sampling points, and each iteration samples the expected improvement (EI) of the objective function. The maximum number of iterations is 50, or the iteration stops when the objective function improvement is less than 1% after five consecutive iterations.

[0142] The historical parameter-effect database is a time-series database used to store past parameter configurations and their corresponding effect metrics. Each record contains the following fields:

[0143] (1) Parameter identification: Parameter combination (α, β) and its application time range;

[0144] (2) Stability index: average eccentricity during application Rate of change of mean radius of curvature Mean precession angular velocity ;

[0145] (3) Performance indicators: average student completion rate, fairness index Path convergence steps. Initial database data was generated via grid search.

[0146] Fifty sets of parameters were uniformly sampled within the range of α∈[0.1, 2.0] and β∈[0, 1.0]. Each set of parameters was applied to the simulation environment and performance metrics were collected within a full semester. Subsequent quarterly updates were made based on actual operational data to provide historical references for Bayesian optimization.

[0147] The historical parameters-effects database is updated quarterly, storing parameter combinations and their corresponding average eccentricity, average completion rate, fairness index, and other indicators.

[0148] After obtaining a new set of optimal parameter combinations (α, β), these are provided as preset values ​​for the objective function to the multi-objective dynamic decision engine.

[0149] The parameter hot update submodule is used to safely update the new optimal parameter combination into the multi-objective dynamic decision engine.

[0150] The update process employs a double-buffer mechanism: this double-buffer mechanism is a parameter update strategy to ensure system stability, comprising two independent storage areas: a primary area and a shadow area. The primary area stores currently used decision parameters, while the shadow area temporarily stores new parameters awaiting verification. The update process consists of three phases:

[0151] (1) Loading phase: Write the new parameter combination (α, β) obtained by the adaptive parameter tuning submodule into the shadow region;

[0152] (2) Observation phase: The system simulates decision-making and monitors the effect indicators based on the shadow area parameters in the next two decision cycles, but it is not actually applied to student recommendations;

[0153] (3) Switching Phase: If the indicator returns to the normal range during the observation period, the shadow area parameters are copied to the primary area, and the new parameters are officially activated; if the indicator does not improve, the shadow area parameters are discarded, and the original parameters are retained. This mechanism can avoid system oscillations caused by parameter mutations and ensure the safety of the update process.

[0154] The educator collaborative intervention module is used to integrate educators' professional judgment into system decision-making.

[0155] As an optional implementation, the educator collaborative intervention module specifically includes:

[0156] The instruction parsing interface is used to receive instructions from educators to confirm, modify, or reject the personalized development path generated by the system.

[0157] Educators can view recommended paths in the system interface and modify them through drag-and-drop, replacement, deletion, or by directly rejecting the entire path. The interface captures these operations in real time and converts them into structured instructions (including information such as operation type, target resource, and original resource).

[0158] The intervention vectorization submodule is used to quantize the modification or rejection instruction into a directional perturbation vector that can be superimposed on the current student dynamic state vector.

[0159] If the educator replaces resource A with resource B, then calculate the difference ΔR between the eigenvectors of resource A and resource B. This is then superimposed as a perturbation vector onto the student's current state vector S0. =S0+λΔR, where λ is the current attenuation coefficient.

[0160] If the educator rejects the entire path, the system randomly selects three resources from the educational resource database that have a relevance greater than 0.6 to the student's historical interests, and calculates their average eigenvector. The generated disturbance ΔR = 0.3 × ( -S0).

[0161] When multiple modification instructions exist within the same decision cycle, perturbations are applied sequentially in chronological order. All perturbed state vectors are checked by a layer to ensure that the values ​​of each dimension remain within the range [0, 1]; if they exceed this range, they are truncated.

[0162] An influence attenuation controller is used to set an intensity coefficient for the directional disturbance vector that decays exponentially with time or decision period, so that its influence on system decision-making gradually weakens until normal automatic decision-making is restored.

[0163] The initial perturbation intensity coefficient is set to 1. Each decision cycle (1 day) is multiplied by a decay factor γ = 0.5, meaning the intensity of the k-th cycle is γ^k = 0.5^k. The system is set to consider the intensity as zero when it falls below a threshold of 0.05, after which no further perturbations are applied.

[0164] For example, after five consecutive cycles, the intensity drops to 0.5^5 = 0.03125, below the threshold, and the system resumes normal automatic decision-making. This approach respects the educator's immediate intervention while preventing long-term solidification of human bias, thus maintaining the system's autonomous evolutionary capacity.

[0165] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A personalized development guidance and resource matching system based on student digital profiles, characterized in that, include: The student digital profile construction module is used to construct a dynamic state vector of a student in a multi-dimensional feature space; wherein, the multi-dimensional feature space includes at least an ability dimension, an interest dimension, and a learning style dimension. The educational resource semantic graph module is used to construct resource feature vectors and semantic association networks between resources in the multi-dimensional feature space. A multi-objective dynamic decision engine connects the student digital profile construction module and the educational resource semantic graph module. It is used to generate personalized development paths based on the dynamic state vector, the resource feature vector, and the semantic association network by simulating the motion trajectory in a virtual force field. The multi-objective dynamic decision engine specifically includes: The force field calculation submodule is used to adaptively adjust the objective function obtained by the decision parameter adaptive adjustment unit, obtain candidate resource feature vectors from the educational resource semantic graph module, and calculate virtual forces. The first type of function calculates the attractiveness based on the cosine similarity between the dynamic state vector and the candidate resource feature vector; The second type of function calculates the moderating force based on the fairness requirements of the distribution of resources among the student population, wherein the fairness requirements are measured by the Gini coefficient; The trajectory simulation submodule is used to treat the dynamic state vector as a point mass in a multi-dimensional feature space, to perform vector synthesis of the attraction and adjustment forces to obtain the resultant force, and to solve the motion equation of the point mass under the action of the resultant force through a numerical integration algorithm, and output the trajectory point sequence. The path mapping submodule is used to map the trajectory point sequence back to a specific educational resource node sequence to form an executable personalized development path. The path stability monitoring unit is connected to the multi-objective dynamic decision engine and is used to calculate the mechanical stability index based on the personalized development path. The path stability monitoring unit specifically includes: The index calculation submodule is used to calculate the mechanical stability index of the personalized development path, including: The eccentricity rate reflects whether the path direction is focused on the core capability area; the rate of change of the radius of curvature reflects whether the capability transition between adjacent resource nodes is smooth; and the precession angular velocity reflects whether the long-term development trend has drifted. The anomaly diagnosis submodule is used to compare the calculated mechanical stability index with the preset normal threshold range. The default judgment period for the threshold is one week. When any index exceeds the threshold for three consecutive periods or a sudden change occurs, it is judged as path instability and an anomaly diagnosis report is generated. The decision parameter adaptive adjustment unit is connected to the path stability monitoring unit and the multi-objective dynamic decision engine, and is used to trigger and execute adaptive adjustment of the decision model parameters in the multi-objective dynamic decision engine according to the mechanical stability index.

2. The personalized development guidance and resource matching system based on student digital profiles according to claim 1, characterized in that: The student digital profile construction module specifically includes: The data preprocessing submodule is used to clean, align, and normalize multi-source heterogeneous time-series data from academic records, behavioral logs, and psychological assessments. The vectorization encoding submodule is used to convert the processed data into fixed-dimensional feature vectors through embedding layers or feature engineering. The State Fusion and Update submodule is used to model the feature vector sequence using a temporal neural network model and output the dynamic state vector that integrates historical and current information.

3. The personalized development guidance and resource matching system based on student digital profiles as described in claim 1, characterized in that: The educational resource semantic graph module specifically includes: The resource feature extraction submodule is used to parse the text, outline and metadata of resources using natural language processing technology, and extract the knowledge points, skill points, difficulty level and teaching style features involved to form an initial resource feature vector; The graph construction submodule is used to construct a semantic association network with resources as nodes and relationships as edges, based on the prerequisite and successive relationships of resources in terms of content and the complementary relationships of teaching objectives. The dynamic weight assignment submodule is used to dynamically adjust the weights of edges and the inherent attribute values ​​of resource nodes in the semantic association network based on the historical success rate of resource usage, student evaluation feedback, and real-time popularity information.

4. The personalized development guidance and resource matching system based on student digital profiles according to claim 1, characterized in that: The adaptive adjustment unit for decision parameters specifically includes: The parameter optimization modeling submodule is used to construct an optimization model for the force field function weights in the multi-objective dynamic decision engine, with the core objective of the mechanical stability index returning to the normal range when the path is determined to be unstable. The adaptive parameter tuning submodule is used to solve the optimization model in the historical parameter-effect database using Bayesian optimization or gradient descent algorithm, to obtain a new set of optimal parameter combinations, and provide them as preset values ​​of the objective function to the multi-objective dynamic decision engine. The parameter hot update submodule is used to safely update the new optimal parameter combination into the multi-objective dynamic decision engine and monitor the adjustment effect during a subsequent preset observation period.

5. The personalized development guidance and resource matching system based on student digital profiles according to claim 1, characterized in that: It also includes an educator collaborative intervention module, which includes: The instruction parsing interface is used to receive instructions from educators to confirm, modify, or reject the personalized development path generated by the system. The intervention vectorization submodule is used to quantize the modification or rejection instruction into a directional perturbation vector that can be superimposed on the current student dynamic state vector. An influence attenuation controller is used to set an intensity coefficient for the directional disturbance vector that decays exponentially with time or decision cycle, so that its influence on the system decision is reduced to less than 10% of the initial intensity after five consecutive cycles, until normal automatic decision-making is restored.

Images