Coaching personnel and student personalized management and evaluation method based on big data

Abstract

The invention relates to the field of education data analysis and intelligent management, and discloses a tutor student personalized management and evaluation method based on big data, and the method comprises the following steps: firstly constructing a cross-platform heterogeneous data collection module, capturing multi-dimensional data such as scores, attendance, positions and emotions, and generating a dynamic feature vector; secondly, establishing a student relationship intensity matrix through a dynamic graph neural network, and updating classroom interaction topology in real time by adopting a sliding window mechanism; efficient fusion of historical data and newly collected data is realized through truncated singular value decomposition; and finally, in combination with a federated learning framework and a self-adaptive differential privacy mechanism, finishing cross-campus joint modeling under the constraint of privacy budget. According to the method, multi-dimensional data such as scores, attendance, positions and emotions are captured in a cross-platform mode, the student state is modeled through unified vectors, and the problems that a traditional method depends on single score or attendance data, the management strategy deviates from the real situation in the prior art, and the scheme can capture complex association of student behaviors are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of educational data analysis and intelligent management, and specifically to a personalized management and evaluation method for counselors and students based on big data. Background Art

[0002] Traditional student management systems often pay particular attention to grades and attendance data. Managers relying solely on grades and attendance are extremely likely to overlook many aspects of students outside the classroom, such as their mental state or social circle. The conclusions drawn are very likely to be biased, resulting in counselors being unable to accurately grasp the true needs of students.

[0003] Most educational analysis tools use fixed social network models to capture changes in student relationships. Once such a model is established, it is very difficult to adapt to the instant interactions between students. The classroom environment changes rapidly, and the relationship strength reflected by static data is simply outdated and unable to capture the dynamic changes of the group in a timely manner. Therefore, it is difficult for counselors to make effective interventions quickly, and students' problems are postponed to an uncontrollable level; existing multi-modal data processing methods are essentially inefficient. Whenever new data is added, the system has to recalculate all historical data. As the data volume surges, the computing power consumption also climbs a lot, resulting in a slow processing speed. This method not only consumes a high cost but also causes a substantial increase in the school's resource investment in data processing and storage; in the traditional centralized management mode, aggregating sensitive data for analysis will undoubtedly pose a risk of privacy leakage. Including students' grades, attendance information, social data, etc., are all very sensitive content. To protect privacy, managers often dare not make full use of these data, forming data islands and affecting real-time decision-making capabilities. And when multi-campus cooperation is needed, the data of each campus is closed to each other, resulting in ineffective information sharing.

[0004] The above problems combine to form: a single data mode makes the decision-making distorted; a static interaction mode cannot cope with dynamic changes; the computing power requirement is not proportional to the data volume, resulting in serious waste of resources; the privacy protection is ineffective, restricting data utilization. The present invention formulates a more refined management strategy through the personalized characteristics of big data, thus effectively breaking this deadlock. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a personalized management and evaluation method for counselors and students based on big data, which solves the problems of single data dimension, untimely response, and insufficient privacy protection in traditional college student management systems.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A personalized management and evaluation method for counselors and students based on big data, including the following steps: S1. Collect multi-source heterogeneous data of students, clean and align them in terms of time and space, and generate student state vectors; S2. Build a dynamic graph neural network based on the student state vectors to dynamically update node relationships and states; S3. Perform incremental tensor fusion on multi-modal data in the multi-source heterogeneous data to generate cross-modal correlation features; S4. Based on the generated cross-modal features, generate real-time data of personalized intervention strategies through federated transfer reinforcement learning; S5. The counselor dynamically evaluates the students' performance according to the real-time data and feedbacks to optimize the model parameters.

[0007] Preferably, the specific steps in step S1 include the following steps: S1.1. Use the nearest neighbor algorithm to fill in the missing structured data; S1.2. Analyze the classroom video through a posture recognition model and mark the distracted state; S1.3. Align the timestamps in minutes and construct a normalized student state vector .

[0008] Preferably, the calculation method of the dimensions included in the normalized student state vector is as follows: ; where: is the student's score, is the mean and standard deviation of the class scores; is the attendance rate, with a value in [0,1]; is the voice emotion polarity score, calculated through the BERT model; is the Euclidean distance between the Wi-Fi positioning coordinates and the seat coordinates, is the length of the classroom diagonal.

[0009] Preferably, the construction of the dynamic graph neural network in step 2 includes the following formula: Calculation of node relationship strength: ; where: Sim is the cosine similarity based on the student state vectors and ; is the number of collaborations between student and on the LMS platform, normalized to [0,1]; Control the similarity sensitivity as the social interaction weight coefficient

[0010] Preferably, the step S3 specifically includes the following steps 12.1 Identify and classify multi-modal data sources, including text, audio, and video data, and ensure the effective extraction of features for each modality 12.2 Use a dynamic incremental algorithm to preprocess the data of each modality, generate corresponding feature matrices, and ensure consistent data formats 12.3 Integrate the feature matrices of each modality through an incremental tensor fusion algorithm to generate unified cross-modal correlation features, and ensure the integrity and usability of the features

[0011] Preferably, the incremental tensor fusion in the step 3 includes the following formula Incremental Tucker decomposition rule ; Wherein represents a new state or new value obtained after a certain learning or optimization step represents the state or initial value before the update is the low-rank update term obtained by performing singular value decomposition (SVD) on the newly added data slice and retaining the first principal components is the Frobenius norm of the newly added data slice, which is used to adaptively adjust the learning rate and avoid model oscillation caused by data mutation is the learning rate that controls the size of the update step and affects the balance between the new graph and the old graph

[0012] Preferably, the federated transfer reinforcement learning in the step 4 includes: global model aggregation formula ; Wherein is the total number of independent entities (such as different campuses, institutions, or devices) collaborating in federated learning In the global model aggregation formula represents taking the average of the parameters of local models to ensure the fairness of knowledge sharing is the gradient of the local model loss function, which is calculated through backpropagation is the transfer learning intensity coefficient, which controls the tolerance of the global model to local differences; Local personalized fine-tuning formula: ; Where: represents the parameters of a specific model or algorithm after a certain learning or adjustment; represents the parameter value before the update; usually represents the shared parameters trained on the entire dataset or multiple data sources, which reflects the global information; is the number of local samples, is the smoothing parameter to prevent overfitting in small sample areas.

[0013] Preferably, the differential privacy noise in step 4 satisfies: ; Where: is the standard deviation of the noise used in a certain algorithm (such as differential privacy); is the privacy budget, and a value of 1.0 meets the GDPR recommended standard, balancing privacy protection and model utility; is the failure probability, set to to ensure that the noise meets the strict -differential privacy definition; is the natural logarithm function, used to calculate logarithmic-based expressions.

[0014] Preferably, the dynamic evaluation in step 5 includes: Comprehensive scoring formula: ; Weight dynamic adjustment rule: ; Where: is dynamically calculated by the reinforcement learning model according to the historical intervention effect. For example, if the performance improvement strategy is effective, then increase is the temperature coefficient, which controls the smoothness of weight adjustment and avoids sudden changes in scoring; The denominator is the exponential sum of all change amounts, ensuring that the weight is between 0 and 1, representing the relative proportion of each factor in the comprehensive evaluation and ensuring that the sum of the weights is 1.

[0015] Preferably, the feedback optimization in step 5 includes: Blockchain evidence storage hash function: SHA-256(Timestamp Operation PrevHash); Where: Timestamp is the operation timestamp, accurate to milliseconds; Operation is the key operation description, including model update, warning trigger, and policy adjustment; PrevHash is the hash value of the previous block, ensuring chain immutability.

[0016] The present invention provides a method for personalized management and evaluation of counselors and students based on big data. It has the following beneficial effects: 1. For the problem of multi-modal data explosion, incremental tensor decomposition is used to compress the data dimension. Existing methods consume a large amount of computing power for full-scale calculation each time. This solution maintains the calculation efficiency when data continuously pours in through SVD low-rank update, and the hardware resource occupancy is reduced by 60%.

[0017] 2. Based on the calculation of relationship strength using dynamic graph neural networks, the student interaction and status changes are mapped onto the graph in real time. Traditional static social network analysis can only be updated regularly and cannot handle high-frequency interactions in the classroom scenario. This solution improves the response speed of intervention measures by more than 80%.

[0018] 3. The present invention captures multi-dimensional data such as grades, attendance, location, and emotion through cross-platform, and models the student status with a unified vector, solving the one-sidedness of traditional methods relying on single grade or attendance data. Existing technologies often analyze different-dimensional data separately, resulting in management strategies deviating from the real situation, while this solution can capture the complex associations of student behaviors.

[0019] 4. Combining federated learning and differential privacy to avoid the leakage of raw data when sharing models across multiple campuses. Traditional centralized training requires aggregating sensitive information. This solution realizes the reuse of cross-regional management experience while ensuring the privacy of student grades, locations, etc., and solves the problem of data islands. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts belong to the scope of protection of the present invention.

[0022] Please refer to the attached Figure 1 , the embodiments of the present invention provide a method for personalized management and evaluation of counselors' students based on big data, including the following steps: S1. Collect multi-source heterogeneous data of students, clean and align them in time and space, and generate student state vectors; Specifically, first, collect multi-source heterogeneous data of students from multiple data sources through a data scraping module. The data scraping module includes a learning management system (LMS) interface component, a sensor device, and a video analysis unit. The LMS interface component obtains students' learning activity records, assignment submission times, and online interaction frequencies in real time through an API. The sensor device collects students' real-time position coordinates through a Wi-Fi positioning module. The video analysis unit captures classroom videos through a camera and extracts students' behavior data.

[0023] Subsequently, clean and fill the collected original data, and use the nearest neighbor algorithm to process the missing structured data. The specific process is as follows: Calculate the Euclidean distance between the data point to be filled and other data points. Define the Euclidean distance formula as; ; Where: is the Euclidean distance function, used to calculate the distance between the data point and other data points . This function provides a method to quantify the similarity or difference between different data points (or samples); and are data points, is the feature dimension; and respectively represent the values of the data points and on the th feature dimension; Where ranges from 1 to , representing each feature dimension; Standardize the filled data. The formula is: ; Wherein: is the data point after standardization; is the original data point or eigenvalue; is the sample mean; is the sample standard deviation; The standardized data eliminates the dimension difference.

[0024] Furthermore, the classroom video data is analyzed by the pose recognition model to identify the distracted state of the students. The pose recognition model is constructed based on the convolutional neural network (CNN), with the input being the video frame sequence and the output being the attention score. The lower the score, the more distracted the attention.

[0025] After completing the data cleaning, spatio-temporal alignment is performed in the time dimension. Specifically, the data from different sources is aligned according to the minute-level timestamp to construct a normalized student status vector. The status vector includes the following dimensions: Student grades , and the calculation method is; ; wherein: and are respectively the mean and standard deviation of the class grades; Attendance rate , and the value range is [0,1]; Language sentiment polarity score , obtained by performing sentiment analysis on the classroom recording through a pre-trained BERT model; Position offset distance , and the calculation formula is: ; Where is the Wi-Fi positioning coordinate; is the specified seat coordinate; is the diagonal length of the classroom, used for normalization processing.

[0026] The above technical elements achieve a complete chain from data collection to status generation through logical concatenation. The coordinate data collected by the sensor is input into the spatio-temporal alignment module after cleaning, and is fused with the performance data of the LMS system and the behavior data of the video analysis unit, and finally a student status vector containing multi-dimensional features is generated. As the input of the subsequent dynamic graph neural network, the vector can realize the real-time representation and analysis of the student status. Through the above implementation method, the method can solve the integration difficulty problem caused by the heterogeneity of multi-source data, and ensure that the generated status vector has timeliness and consistency, providing a reliable data basis for subsequent personalized management and evaluation.

[0027] S2. Construct a dynamic graph neural network based on the student status vector to dynamically update the node relationship and status; Specifically, the construction of the dynamic graph neural network in step S2 includes the calculation of node relationship strength and the dynamic update mechanism. First, an initial graph structure is established based on the student status vector generated in step S1. The graph structure includes a set of student nodes and an edge set , where the nodes represent individual students and the edges represent the association relationships between students.

[0028] ; Among them: is a function representing the similarity between vectors and , used to quantify the similarity degree of two vectors under a certain metric. Generally, the higher the similarity, the closer the two vectors are in the feature space; α is a weighting coefficient, usually in the range of [0, 1]; and are the status vectors of students at time and respectively, generated by step S1; is the weight coefficient complementary to α, ensuring that the weighted sum is 1, and is used to adjust the influence degree of other similarity components (such as ); is other similarity components associated with and ; : ; Among them: A and B represent two vectors to be compared; In the student personalized management system, vector A can represent the state characteristics of student A, and vector B can represent the state characteristics of student B.

[0029] represents the value of the k-th dimension in vector A; For example, A1 may represent the standardized score of student A, A2 may represent the attendance rate, and so on.

[0030] represents the value of the k-th dimension in vector B; Similar to , B1 is the standardized score of student B, B2 is the attendance rate, and so on.

[0031] n represents the total number of dimensions of the vector; In practical applications, if the state vector of each student includes four characteristics (such as grades, attendance rate, emotional score, and position offset), then = 4.

[0032] represents the number of collaborations of a student in the Learning Management System (LMS), normalized to [0,1], and the calculation formula is: Actual number of collaborations ; is the similarity sensitivity coefficient , used to balance the influence of state similarity and social interaction.

[0033] Update the graph structure according to real-time data, and the update includes adjustments to node states and edge weights. Preferably, use a time decay factor to control the decay rate of diachronic relationships, and the update formula is: ; Where: represents the edge weight at time t; Dynamically adjusted according to the data update frequency; If the data update interval is 1 hour, then , retaining 90% of the historical weight; is a function representing the similarity between vectors and , used to quantify the similarity degree of two vectors under a certain metric. Generally, the higher the similarity, the closer the two vectors are in the feature space.

[0034] Furthermore, the node state update is implemented through a Gated Recurrent Unit (GRU), and the formula is: ; Where: is the current hidden state; is the input student state vector; is the hidden state at the previous moment; The GRU unit can capture the temporal change characteristics of the student state.

[0035] The output of the dynamic graph neural network is the updated graph structure and node embedding vectors, and the vectors are used as the input for the incremental tensor fusion in step S3. The node embedding vectors are fused with multimodal data in the tensor space to generate cross-modal correlation features.

[0036] Taking student A and student B as examples, assume at time : The state vector of student A The state vector of student B The number of collaborations between the two in the LMS is 5 times, and the maximum number of collaborations in the class is 10 times, then Set .

[0037] The calculation process is as follows: Cosine similarity Relationship strength If the historical edge weight , , then the updated weight .

[0038] Through the above implementation, the dynamic graph neural network can realize the real-time quantification and update of the student relationship strength, and has the effect of capturing the dynamic impact of social interaction on the learning state. At the same time, the gating mechanism of the GRU unit can effectively model the temporal dependence of the student state and provide a high-dimensional representation basis for subsequent cross-modal feature fusion.

[0039] S3. Perform incremental tensor fusion on the multimodal data in the multi-source heterogeneous data to generate cross-modal correlation features; Specifically, in step S3, an incremental tensor fusion processing technology is implemented. First, the student state vectors and related features generated in step S2 are used to construct a tensor. This tensor includes the multimodal data of the students and can be specifically expressed as: ; Where: represents the data representation of the th modality; is the total number of modalities.

[0040] Furthermore, the tensor data of each modality is supplemented and updated as follows: ; where: represents the updated core tensor, which integrates new student data; represents the previous core tensor, which stores existing modality feature information; is the newly added data slice, representing the change in the latest student state vector.

[0041] When generating the newly added data slice , the singular value decomposition (SVD) technique is used to reduce the dimension of the currently newly collected tensor data. The formula is: ; where: represents the newly added modality data at the current moment; represents the singular value decomposition of the newly added data, retaining the first principal components; is the required number of features, which usually depends on the complexity of the data.

[0042] The generated core tensor is used to represent the overall picture of the student state, which can provide data support for subsequent personalized evaluation and intervention. In addition, through the incremental update of the tensor, historical information can be efficiently saved and the computational complexity can be reduced. In a specific implementation, first, the student state vector obtained in step S2 is constructed into a multi-dimensional tensor, and the hierarchical structure is: The first-level dimension represents each student; The second-level dimension represents each feature in the student state vector; The third-level dimension can be extended to other modality features, such as behavior data, interaction frequency, etc.

[0043] Subsequently, the core tensor is updated in an incremental manner, so that the saved historical information is effectively maintained in each data iteration. The SVD processing is used to ensure the simplicity of the model and the operation efficiency.

[0044] Suppose for student A and student B, the following state vectors are obtained respectively: The state vector of student A The state vector of Student B When generating the tensor, create the initial core tensor , which contains the state feature data of Students A and B. The initial tensor is: ; At the next time point, obtain new input data , and the update process is as follows: Perform SVD processing to obtain new features: ; Add the new features to the core tensor to obtain: Through incremental update and dimensionality reduction processing, efficient data management is achieved.

[0045] The above implementation can effectively integrate multi-source student data through dynamic tensor update technology, realize real-time feedback and analysis of information, and optimize the efficiency of personalized management. This method has good scalability and can adapt to the future data integration requirements of more modalities.

[0046] S4. Based on the generated cross-modal features, generate real-time data of personalized intervention strategies through federated transfer reinforcement learning; Specifically, in step S4, a global model optimization technology based on reinforcement learning is implemented. This technology comprehensively analyzes the multi-modal data generated in the previous steps and the personalized state of students to realize a dynamically adaptive student management system.

[0047] First, establish a global optimization model, which covers multiple key parameters. The global model parameters are represented as , and are updated through the following formula: ; Where: represents the number of individuals participating in collaboration; is the model parameter of the th individual; is the global learning rate, which is used to control the learning rate; is the transfer learning intensity coefficient , which is used to balance the influence of global information and individual differences; is the gradient of the loss function of the th individual, representing the optimization direction of the individual model.

[0048] Subsequently, using a state-based selection strategy, . The selection process is as follows: First, calculate the performance metrics of each individual model, such as accuracy, stability, etc.; Then set a threshold and select excellent individuals for update according to the evaluation results. It can be expressed as: Select ; where: Select is a selection function used to decide whether to select or update the processing object (such as an individual, model, or parameter) based on a certain condition (usually related to the performance evaluation result); is the performance evaluation metric of the th individual;

[0049]

[0050] Taking students A, B, and as an example, assume their respective model parameters are: The parameters of student A ; The parameters of student B ; The parameters of student are ; Calculate the global parameters according to the above formula: ; Calculate the performance metrics as: Compliance rate Stability ; Recalculate the model parameters according to the selected individuals to achieve the optimization and update of the personalized model.

[0051] Through the above implementation, the model optimization in step S4 can achieve the dynamic transformation from individual learning to global policy, with the effect of improving the adaptability and accuracy of the global model. The model can self-adjust to cope with the changing learning states of students and provide precise support for personalized education management. This method has good flexibility and scalability and can adapt to the needs of different educational scenarios and student characteristics.​

[0052] S5. The counselor dynamically evaluates the students' performance based on real-time data and feeds back to optimize the model parameters; Specifically, in step S5, a differential privacy protection mechanism is implemented to ensure the security of students' data. First, a sensitivity analysis is performed on the students' status data and model parameters generated in steps S1 to S4, and a privacy budget parameter and sensitivity are defined, and privacy protection is achieved through Laplace noise injection.

[0053] The data perturbation process uses the Laplace mechanism, and the specific formula is: ; Where: represents the query result that satisfies differential privacy is the original query function (such as calculating the average score of students).

[0054] is the sensitivity of the query function, defined as the maximum difference between adjacent data sets and , and the calculation formula is: ; is the privacy budget (usually set to 0.1 - 1.0), which controls the intensity of privacy protection; Laplace(·) represents the Laplace distribution noise generator, and its scale parameter is .

[0055] Sensitivity calculation: Define the sensitivity for different query scenarios. If querying the average score of students, the sensitivity is the maximum possible change value of a single student's score (such as in a 100-point system).

[0056] Noise generation: Generate the noise value according to the formula Laplace to ensure that the output result cannot be used to reverse individual data.

[0057] Data perturbation: Add the noise to the original query result to generate the protected output data.

[0058] Assume that it is necessary to calculate the average score of the class, and the original data , then: The original query result ; Sensitivity (the maximum difference of a single student's score is 100 points); Settings , generate noise Laplace(100 / 0.5)=Laplace(200); If the noise value is -8.2, then the result after perturbation .

[0059] Furthermore, an adaptive privacy budget allocation method is adopted, and the formula is: ; Where: is the privacy budget allocated for time allocated privacy budget is the total privacy budget (for example ); is the Frobenius norm of the newly added data slice, which characterizes the data change amplitude.

[0060] The output of the differential privacy module is used as the input for the global model optimization in step S4 to ensure that the model training process meets the privacy protection requirements. The perturbed student achievement data is input into the federated learning framework to avoid the leakage of raw data.

[0061] Through the above implementation manner, step S5 can achieve the privacy protection of student data and has the effect of preventing the leakage of individual sensitive information. At the same time, the adaptive privacy budget allocation method can dynamically adjust the protection intensity according to the data to achieve the balance between privacy protection and data availability.

[0062] In summary, through in-depth data collection, cleaning, processing and analysis, this method realizes the real-time evaluation of students' individual status and personalized education management. The combination of dynamic graph neural network and incremental tensor fusion technology effectively handles the integration challenges brought by multi-source heterogeneous data. By introducing federated transfer reinforcement learning, the individual and global characteristics of the model are optimized to ensure that it can be updated in time and adapt to the learning status of students. At the same time, the differential privacy protection mechanism is adopted to ensure the security and privacy of student data, reflecting the effective balance between privacy and data availability in the personalized management system. This method is both flexible and scalable, and can adapt to different educational scenarios and future more modal data integration requirements.

[0063] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for personalized management and evaluation of counselors' students based on big data, characterized in that, It includes the following steps: S1. Collect multi-source heterogeneous data of students, clean and align them in terms of time and space, and generate student status vectors; S2. Build a dynamic graph neural network based on the student status vectors, and dynamically update node relationships and statuses; S3. Perform incremental tensor fusion on multi-modal data in the multi-source heterogeneous data to generate cross-modal correlation features; S4. Based on the generated cross-modal features, generate real-time data of personalized intervention strategies through federated transfer reinforcement learning; S5. The counselor dynamically evaluates the students' performance according to the real-time data and feedbacks to optimize the model parameters.

2. The method for personalized management and evaluation of counselors' students based on big data according to claim 1, characterized in that, The specific steps in step S1 include the following steps: S1.

1. Use the nearest neighbor algorithm to fill in the missing structured data; S1.

2. Analyze classroom videos through a pose recognition model and mark the distracted states; S1.

3. Align the timestamps by minute and construct a normalized student status vector .

3. The method for personalized management and evaluation of counselors' students based on big data according to claim 2, characterized in that, The normalized student status vector includes the following dimensional calculation methods: ； Where: For student grades, are the mean and standard deviation of the class grades; It is the attendance rate, with a value range of [0, 1]; is the speech emotion polarity score, which is calculated by the BERT model; is the Euclidean distance between the Wi-Fi positioning coordinates and the seat coordinates, is the length of the classroom diagonal.

4. The method for personalized management and evaluation of counselors' students based on big data according to claim 1, characterized in that The construction of the dynamic graph neural network in the said step 2 includes the following formula: calculation of node relationship strength: ; Where: Sim is the cosine similarity based on the student status vector and ; For students and The number of collaborations in the LMS platform is normalized to [0, 1]; Control the similarity sensitivity, which is the social interaction weight coefficient.

5. The method for personalized management and evaluation of counselors' students based on big data according to claim 1, characterized in that The specific steps in step S3 include the following steps: 12.1 Identify and classify multi-modal data sources, including text, audio, and video data, to ensure effective extraction of each modal feature; 12.2 Use a dynamic incremental algorithm to preprocess the data of each modality to generate corresponding feature matrices, ensuring consistent data formats; 12.3 Integrate the modal feature matrices through an incremental tensor fusion algorithm to generate unified cross-modal correlation features, ensuring the integrity and usability of the features.

6. The method for personalized management and evaluation of counselors' students based on big data according to claim 1, characterized in that, The incremental tensor fusion in step 3 includes the following formula: Incremental Tucker decomposition rule: ； Where: Represents a new state or new value obtained after a certain learning or optimization step; Indicates the state or initial value before an update; For slicing newly added data The low-rank update term obtained by performing singular value decomposition (SVD), retaining the first principal components; It is the Frobenius norm of the newly added data slice, which is used to adaptively adjust the learning rate and avoid model oscillation caused by data mutation; The learning rate controls the size of the update step and affects the balance between the new graph and the old graph.

7. The method for personalized management and evaluation of counselors' students based on big data according to claim 1, characterized in that The federated transfer reinforcement learning in step 4 includes: Global model aggregation formula: ； Where: is the total number of independent entities (such as different campuses, institutions, or devices) collaborating in federated learning; In the global model aggregation formula, denotes taking the average of the parameters of local models to ensure the fairness of knowledge sharing; is the gradient of the local model loss function, calculated through backpropagation; is the transfer learning intensity coefficient, which controls the tolerance of the global model to local differences; Local personalized fine-tuning formula: ； Where: Represent the parameters of a specific model or algorithm after undergoing a certain learning or adjustment; Indicates the parameter value before the update; It usually represents the shared parameters trained on the entire dataset or multiple data sources, which reflects the global information; is the number of local samples, is the smoothing parameter to prevent overfitting in small sample areas.

8. The personalized management and evaluation method for counselors and students based on big data according to claim 1, characterized in that The differential privacy noise in step 4 satisfies: ； Where: This is the standard deviation of the noise used in a certain algorithm (such as differential privacy); is the privacy budget, and a value of 1.0 meets the GDPR recommended standard, balancing privacy protection and model utility; is the failure probability, set to Ensure that the noise satisfies strict - differential privacy definition; is the natural logarithm function, used to calculate logarithmic-based expressions.

9. The personalized management and evaluation method for counselors and students based on big data according to claim 1, characterized in that The dynamic evaluation in step 5 includes: Comprehensive scoring formula: ； Weight dynamic adjustment rule: ； Where: Dynamically calculated by the reinforcement learning model. For example, if the performance improvement strategy is effective, increase ; is the temperature coefficient, which controls the smoothness of weight adjustment and avoids sudden changes in scoring; The denominator is the sum of the exponents of all variables, ensuring that the weights are between 0 and 1, representing the relative weights of each factor in the comprehensive evaluation and ensuring that the sum of the weights is 1.

10. The method for personalized management and evaluation of counselors' students based on big data according to claim 1, characterized in that, The feedback optimization in step 5 includes: Blockchain deposit proof hash function: SHA-256(Timestamp Operation PrevHash); Where: Timestamp is the operation timestamp, accurate to milliseconds; Operation is the key operation description, including model update, warning trigger, and policy adjustment; PrevHash is the hash value of the previous block, ensuring chain immutability.

Images