A teaching data management method and system based on artificial intelligence

Abstract

The application discloses a kind of based on artificial intelligence's teaching data management method and system, method includes: according to the heterogeneous data stream of real-time generation of multiple-source teaching platform, obtain standardization time series data stream;Based on standardization time series data stream, through the dynamic hierarchical storage strategy based on attribute-based encryption, teaching data is classified and encrypted, and obtain security hierarchical storage topology;According to user query request and teaching scene label, extract target data set from security hierarchical storage topology, and obtain enhanced multi-modal teaching data set;Based on enhanced multi-modal teaching data set, through dynamic subgraph evolution algorithm, the interpretable teaching mode atlas is generated;According to teaching mode atlas and real-time teaching feedback data, through course-learner dual-channel adaptive recommendation model, personalized teaching recommendation strategy is generated. Utilize the embodiment of the application, can improve the utilization efficiency of teaching resources, also can realize personalized, intelligent teaching recommendation and decision-making.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent teaching, and particularly relates to a teaching data management method and system based on artificial intelligence. BACKGROUND

[0002] With the rapid development of information technology, the education field is undergoing profound changes. Based on the wide application of the Internet and multi-source teaching platforms, massive teaching data, including video courses, text materials, behavior logs, test scores, etc., are continuously generated, providing valuable basic resources for personalized education and intelligent teaching. However, in the face of heterogeneous, massive and dynamically changing education data, how to efficiently, safely and intelligently manage and utilize them has become a key technical problem to be solved.

[0003] At present, many teaching data management systems only support single or partial data processing methods, making it difficult to realize cross-platform and cross-modal deep integration. The heterogeneous data formats are numerous, and data security and privacy protection have become another challenge. The diversity and inconsistency of data lead to a sharp increase in processing complexity, and the lack of intelligent analysis models limits the value mining of data, making it difficult to meet the needs of personalized teaching and precise recommendation. SUMMARY

[0004] The purpose of the present application is to provide a teaching data management method and system based on artificial intelligence to solve the problems in the prior art and improve the utilization efficiency of teaching resources, and to realize personalized and intelligent teaching recommendation and decision-making.

[0005] One embodiment of the present application provides a teaching data management method based on artificial intelligence, which comprises:

[0006] According to the real-time generation of heterogeneous data streams from multi-source teaching platforms, the heterogeneous data streams are dynamically analyzed and normalized, the data format features are identified and a dynamic interface adapter is constructed by using a bidirectional long short-term memory network, and standardized time series data streams are obtained;

[0007] Based on the standardized time series data streams, the teaching data is classified and encrypted by using a dynamic hierarchical storage strategy based on attribute-based encryption, a knowledge distillation algorithm is used to compress unstructured data and generate a unique data fingerprint, and a secure hierarchical storage topology is obtained;

[0008] According to user query requests and teaching scene labels, target data sets are extracted from the secure hierarchical storage topology, text, video and behavior log data are jointly embedded and represented by using a multi-modal feature fusion network, and missing features are completed by using an adversarial generative network, and an enhanced multi-modal teaching data set is obtained;

[0009] Based on the enhanced multi-modal teaching data set, a cross-institutional federated knowledge graph is constructed, a differential privacy protected graph attention mechanism is used to mine potential teaching association relationships, and an interpretable teaching mode graph is generated through a dynamic subgraph evolution algorithm;

[0010] According to the teaching mode graph and real-time teaching feedback data, a personalized teaching recommendation strategy is generated through a course-learner dual-channel adaptive recommendation model, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by fusing time series reinforcement learning and meta-path reasoning algorithm.

[0011] Optionally, according to the heterogeneous data stream generated in real time by the multi-source teaching platform, the heterogeneous data stream is dynamically parsed and normalized, the data format features are identified and a dynamic interface adapter is constructed by combining a bidirectional long short-term memory network, and standardized time series data stream is obtained, including:

[0012] Receiving the original data stream from the teaching platform, the learning management system and the online course platform, the time series features of the original data stream are extracted through the bidirectional long short-term memory network, the implicit mode of the data format is identified, and the multi-source data format feature vector is obtained;

[0013] Based on the multi-source data format feature vector, a dynamic interface template library is generated by using a density peak clustering algorithm, and the optimal interface template is dynamically matched according to the real-time data stream characteristics, and the interface adapter initialization parameter is obtained;

[0014] According to the interface adapter initialization parameter, a lightweight data conversion rule engine is constructed, and the heterogeneous data containing SCORM, xAPI and IMS-QTI is converted into a unified JSON-LD format, and an intermediate standardized data block is obtained;

[0015] The intermediate standardized data block is subjected to time series alignment processing, the sliding window mechanism is used to eliminate the cross-platform timestamp deviation, and the continuous and conflict-free standardized time series data stream is output through the event-driven pipeline.

[0016] Optionally, based on the standardized time series data stream, the teaching data is classified and encrypted through a dynamic hierarchical storage strategy based on attribute-based encryption, a knowledge distillation algorithm is used to compress unstructured data and generate a unique data fingerprint, and a secure hierarchical storage topology is obtained, including:

[0017] According to the metadata attributes of the standardized time series data stream, the data sensitivity level label is extracted through the attribute-based encryption algorithm, the CP-ABE strategy is used to define the access control tree, and the dynamic hierarchical encryption strategy is generated;

[0018] Based on the dynamic hierarchical encryption strategy, the teaching video data and the text log data are classified and encrypted, and the hierarchical ciphertext data block is generated.

[0019] For unstructured video data, a knowledge distillation algorithm is used to extract key teaching features through a teacher model and train a lightweight student model to generate compressed feature vectors, outputting a lightweight video feature package that preserves semantics.

[0020] By integrating hierarchical encrypted data blocks with lightweight video feature packets into a hash chain, a unique data fingerprint is generated based on the content features of the hierarchical encrypted data blocks, and a secure hierarchical storage topology with access control labels is constructed.

[0021] Optionally, based on user query requests and teaching scenario tags, the target dataset is extracted from the secure hierarchical storage topology. A multimodal feature fusion network is used to jointly embed and represent text, video, and behavioral log data. An adversarial generative network is then used to complete missing features, resulting in an enhanced multimodal teaching dataset, including:

[0022] Based on the keyword vectors and teaching scenario tags in the user's query request, the target data is retrieved in the secure hierarchical storage topology using a semantic similarity algorithm, and a candidate dataset index table is generated.

[0023] Based on the candidate dataset index table, a multi-head attention mechanism is used to align text TF-IDF features, video optical flow features, and behavioral log temporal features to generate a multimodal joint embedding matrix.

[0024] The multimodal joint embedding matrix is ​​input into the generative adversarial network. The discriminator identifies the location of missing features, the generator reconstructs the missing video keyframes and text semantic segments, and outputs the complete feature tensor.

[0025] Adversarial training optimization is performed on the completed feature tensor, and a gradient penalty strategy is adopted to stabilize the training process, generating an intermediate multimodal dataset with enhanced robustness.

[0026] By using a feature decoupling algorithm to separate modality-specific and common information in intermediate multimodal datasets, an enhanced multimodal teaching dataset with cross-modal association labels is finally output.

[0027] Optionally, the step of constructing a cross-institutional federated knowledge graph based on the enhanced multimodal teaching dataset, mining potential teaching relationships using a differential privacy-preserving graph attention mechanism, and generating an interpretable teaching pattern graph through a dynamic subgraph evolution algorithm includes:

[0028] Based on the entity relation triples in the enhanced multimodal teaching dataset, cross-institutional data is aggregated through a federated learning framework, homomorphic encryption technology is used to protect institutional privacy, and an initial federated knowledge graph is generated.

[0029] Based on the initial federal knowledge graph, a differential privacy protection graph attention mechanism is designed to generate a privacy-safe relationship strength matrix by injecting Laplace noise to disturb the attention weight between nodes;

[0030] According to the relationship strength matrix, a time series graph convolution network is used to capture the evolution law of the teaching mode, and a dynamic community discovery algorithm is used to identify potential teaching association clusters to generate a set of subgraph evolution paths;

[0031] Causal reasoning analysis is performed on the subgraph evolution path, and a Bayesian network is used to calculate the explainability score of the teaching mode to generate an intermediate teaching mode graph with a confidence label;

[0032] The graph neural network distillation technology is used to compress the size of the intermediate teaching mode graph, and finally an explainable teaching mode graph supporting real-time reasoning is output.

[0033] Optionally, according to the teaching mode graph and real-time teaching feedback data, a course-learner dual-channel adaptive recommendation model is used to generate a personalized teaching recommendation strategy, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by fusing time series reinforcement learning and meta-path reasoning algorithm, including:

[0034] According to the knowledge point dependency relationship in the teaching mode graph and the real-time feedback data of the learner, the node weight in the course knowledge topology structure and the time series features of the learner behavior trajectory are extracted, and the topological feature graph of the course channel and the behavior evolution sequence of the learner channel are constructed;

[0035] In the course channel, a time series reinforcement learning algorithm is used to perform dynamic path planning on the topological feature graph, and a sliding window mechanism is used to capture historical teaching effect data to generate a candidate recommendation sequence based on knowledge point association strength;

[0036] In the learner channel, a meta-path reasoning algorithm based on the knowledge graph is used to analyze the knowledge transition mode in the learner behavior evolution sequence to generate a personalized path candidate set that integrates the ability growth curve;

[0037] The candidate recommendation sequence of the course channel and the personalized path candidate set of the learner channel are input into a multi-strategy fusion module, a cross-channel weight distribution matrix is calculated through a dynamic attention mechanism, and a personalized teaching comprehensive recommendation strategy with priority ranking is generated.

[0038] Optionally, according to the metadata attributes of the standardized time series data stream, a data sensitive level label is extracted through an attribute-based encryption algorithm, a CP-ABE strategy is used to define an access control tree, and a dynamic hierarchical encryption strategy is generated, including:

[0039] According to the metadata attribute of the standardized time sequence data stream, the privacy correlation degree of the data field is analyzed through a multi-dimensional sensitive feature extraction network, and a hierarchical clustering algorithm and a self-encoder fusion architecture are used to map the student information, the score record and the classroom behavior log to different sensitive levels, and output a data sensitivity three-dimensional feature vector;

[0040] Based on the data sensitivity three-dimensional feature vector, a dynamic attenuation type decision tree is used to generate a fine-grained access strategy, wherein the course video data is divided into access permissions according to the resolution and content sensitivity, the test data is defined as a decryption condition according to the score distribution interval, and a timestamped initial access control rule set is output;

[0041] According to the context information of the real-time teaching scene, the initial access control rule set is dynamically modified through a strategy evolution engine, when a cross-campus data sharing request is detected, a regional constraint attribute is automatically added and a two-factor authentication condition is introduced, and a spatiotemporal-aware enhanced access control tree is generated;

[0042] The enhanced access control tree and the data sensitivity three-dimensional feature vector are coupled and optimized, an asymmetric bilinear pairing algorithm is used to calculate the strategy node weight, an elliptic curve encryption technology is used to generate a strategy public key matrix, and finally a CP-ABE hierarchical encryption strategy containing a dynamic attenuation factor and a spatiotemporal constraint is output.

[0043] Another embodiment of the application provides an artificial intelligence-based teaching data management system, which comprises:

[0044] The processing module is used for dynamically analyzing and normalizing the heterogeneous data stream generated by the multi-source teaching platform in real time, combining a bidirectional long short-term memory network to identify data format features and constructing a dynamic interface adapter, and obtaining a standardized time sequence data stream;

[0045] The encryption module is used for classifying and encrypting the teaching data through a dynamic hierarchical storage strategy based on attribute-based encryption based on the standardized time sequence data stream, using a knowledge distillation algorithm to compress unstructured data and generate a unique data fingerprint, and obtaining a secure hierarchical storage topology;

[0046] The extraction module is used for extracting a target data set from the secure hierarchical storage topology according to a user query request and a teaching scene label, jointly embedding text, video and behavior log data through a multi-modal feature fusion network, and using an adversarial generative network to complete missing features, and obtaining an enhanced multi-modal teaching data set;

[0047] The mining module is configured to construct a cross-institutional federated knowledge graph based on the enhanced multi-modal teaching data set, mine potential teaching association relationships by using a differential privacy protected graph attention mechanism, and generate an interpretable teaching mode graph by using a dynamic subgraph evolution algorithm.

[0048] The generation module is configured to generate a personalized teaching recommendation strategy by using a course-learner dual-channel adaptive recommendation model according to the teaching mode graph and real-time teaching feedback data, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by fusing a time series reinforcement learning algorithm and a meta-path reasoning algorithm.

[0049] Another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any one of the above embodiments when running.

[0050] Another embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the computer program to execute the method described in any one of the above embodiments.

[0051] Compared with the prior art, the present application provides a teaching data management method based on artificial intelligence, which obtains a standardized time series data stream according to a heterogeneous data stream generated in real time by a multi-source teaching platform, classifies and encrypts teaching data by using a dynamic hierarchical storage strategy based on attribute-based encryption based on the standardized time series data stream, obtains a secure hierarchical storage topology, extracts a target data set from the secure hierarchical storage topology according to a user query request and a teaching scene label, obtains an enhanced multi-modal teaching data set, generates an interpretable teaching mode graph by using a dynamic subgraph evolution algorithm based on the enhanced multi-modal teaching data set, and generates a personalized teaching recommendation strategy by using a course-learner dual-channel adaptive recommendation model according to the teaching mode graph and real-time teaching feedback data, so as to improve the utilization efficiency of teaching resources and realize personalized and intelligent teaching recommendation and decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A hardware structure block diagram of a computer terminal of the teaching data management method based on artificial intelligence provided by the embodiment of the present application is provided.

[0053] Figure 2 A flowchart of the teaching data management method based on artificial intelligence provided by the embodiment of the present application is provided.

[0054] Figure 3 A structure diagram of the teaching data management system based on artificial intelligence provided by the embodiment of the present application is provided. DETAILED DESCRIPTION

[0055] The embodiments described below with reference to the drawings are exemplary and are used only to explain the present application, and cannot be explained as a limitation of the present application.

[0056] The embodiment of the present application first provides an artificial intelligence-based teaching data management method, which can be applied to an electronic device such as a computer terminal, specifically, a general computer, etc.

[0057] The following will be described in detail taking a computer terminal as an example. Figure 1 A hardware structure block diagram of a computer terminal of an artificial intelligence-based teaching data management method provided by the embodiment of the present application is shown in FIG. 1. Figure 1 As shown in the figure, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0058] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions which, when executed, can cause the processor to execute any kind of artificial intelligence-based teaching data management method.

[0059] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0060] The internal memory provides an environment for the running of the computer program in the non-volatile storage medium, which, when executed by the processor, can cause the processor to execute any kind of artificial intelligence-based teaching data management method.

[0061] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 The structure shown in FIG. 1 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0062] It should be appreciated that the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0063] Referring to Figure 2 The embodiments of the present application provide an artificial intelligence-based teaching data management method, which can include the following steps:

[0064] S201, according to the heterogeneous data stream generated by the multi-source teaching platform in real time, the heterogeneous data stream is dynamically parsed and normalized, the data format features are identified and the dynamic interface adapter is constructed in combination with the bidirectional long short-term memory network, and the standardized time sequence data stream is obtained; Specifically, it can include:

[0065] S2011, receiving the original data stream from the teaching platform, the learning management system and the online course platform, extracting the time sequence features of the original data stream through the bidirectional long short-term memory network, identifying the implicit mode of the data format, and obtaining the multi-source data format feature vector;

[0066] Multi-source data access and preprocessing

[0067] ‌Data source definition and protocol analysis: The original data stream of the teaching platform (such as Moodle), the learning management system (such as Canvas) and the online course platform (such as Coursera) is accessed in real time through the API interface (RESTful or WebSocket). For example, the Moodle platform provides SCORM format learning behavior log (such as activity_001.xml), Canvas outputs xAPI statement (such as {"actor":"user123","verb":"completed","object":"quiz_456"}), and Coursera transmits test data through IMS-QTI standard (such as <qti-assessment-item identifier="math_quiz">).

[0068] ‌Data Buffering and Sharding: Use Kafka message queue to shard and buffer raw data streams, set each shard size to 128MB to avoid memory overflow. For example, split xAPI statements into batches by time window (e.g., 5 minutes), assign to different partitions (e.g., partition_1 to partition_8) for parallel processing.

[0069] ‌Noise Filtering and Anomaly Detection: Remove invalid data (e.g., ERROR:invalid session in logs) by regular expression matching, and identify abnormal data points (e.g., records with single-day learning duration exceeding 24 hours) based on Isolation Forest algorithm. For example, set the anomaly score threshold to 0.65, filter out 0.3% of abnormal data.

[0070] ‌Bi-LSTM (Bidirectional Long Short-Term Memory) Time Series Feature Extraction

[0071] ‌Network Architecture Design: Build a Bi-LSTM network with input layer dimensions [batch_size=64, time_steps=30, features=50], where time_steps represents the time window length (30 consecutive data points), and features include data format attributes (e.g., field number, encoding type, nesting level). Set the hidden layer to 128 neurons, with forward and backward LSTMs capturing historical and future dependency relationships, respectively.

[0072] ‌Implicit Pattern Recognition: Extract global time series features through the last layer output of Bi-LSTM. For example, for SCORM data in <sequencing>The label nesting structure is encoded, and the feature vector is output (such as [0.23, 0.87, -0.15] represents the label hierarchical complexity, frequency of occurrence, and timing volatility).

[0073] ‌Dynamic feature mapping: adaptive weight distribution is used for different platform data characteristics. For example, the verb-object combination of xAPI statements (such as answered-question) is mapped to a high-dimensional sparse vector (dimension = 300), which is compressed to a 50-dimensional dense vector through an embedding layer (Embedding Layer) and then input into Bi-LSTM.

[0074] ‌Multi-source data format feature vector generation

[0075] ‌Feature fusion and dimension reduction: the time series features output by Bi-LSTM (dimension = 256) are spliced with static metadata (such as file format, encoding standard), and the multi-source data format feature vector is reduced to 128 dimensions through a dense layer (Dense Layer).

[0076] ‌Incremental update mechanism: when new data formats (such as new H5P interactive content) are detected, an online learning mode is started, and the Bi-LSTM weight is fine-tuned through stochastic gradient descent (SGD) (learning rate = 0.001) to update the feature vector space.

[0077] S2012, based on the multi-source data format feature vector, a density peak clustering algorithm is used to generate a dynamic interface template library, and the optimal interface template is dynamically matched according to the real-time data stream characteristics to obtain the interface adapter initialization parameters;

[0078] Density peak clustering (DPC) template generation

[0079] ‌Local density and distance calculation: the local density ρ and the minimum distance δ of the multi-source data format feature vector (such as 1000 samples) are calculated. For example, ρ_i represents the number of samples within a radius r = 0.5 (Euclidean distance) around feature vector i, and δ_i is the distance from i to the nearest sample with higher density.

[0080] ‌Cluster center selection: draw a decision graph (ρ-δ graph), and manually label samples with density and distance higher than the threshold as cluster centers. For example, select 5 center points with ρ>8 and δ>1.2, which correspond to SCORM 1.2, xAPI 1.0.1, IMS-QTI 2.1, etc.

[0081] ‌Dynamic template library construction: an interface template is generated for each cluster center, which contains field mapping rules (such as mapping the <adlcp:location>Mapping to context.location), type conversion table (e.g. string "90%" → float 0.9), exception handling strategy (e.g. missing field filled with null by default).

[0082] ‌Real-time template matching and parameter initialization‌

[0083] ‌Similarity measure and matching: extract format feature vector (dimension = 128) from real-time data stream, and compute its cosine similarity with all cluster centers in the template library. For example, a xAPI data has similarity 0.93 with xAPI 1.0.1 template and 0.21 with IMS-QTI 2.1 template, then the best matching template is xAPI 1.0.1.

[0084] ‌Adaptive threshold adjustment: if the highest similarity is lower than a dynamic threshold (initial value = 0.85), trigger the new template generation process. For example, when a new data format (e.g. xAPI 2.0 Beta) with similarity = 0.78 is detected, create a temporary template and add it to the template library.

[0085] ‌Interface adapter parameter generation: initialize field mapping rules, data validation regular expression (e.g. ^[A-Za-z0-9_]+$), encoding converter (e.g. UTF-8 → ASCII) according to the matching result. For example, the parameters of xAPI template include verb_mapping: {"completed":"finish", "answered":"respond"} and timestamp_format: "YYYY-MM-DDTHH:mm:ssZ".

[0086] ‌Template library online optimization‌

[0087] ‌Cluster drift detection: periodically (e.g. every hour) recompute ρ and δ of cluster centers, if the original center density drops more than 20%, merge or split the cluster. For example, when ρ of SCORM 1.2 template drops from 15 to 9, trigger the template split into two sub-clusters SCORM 1.2_A and SCORM 1.2_B.

[0088] ‌Template weight decay: set a hotness value (initial value = 100) for each template, increase 5 points for each successful match, decay 10% per day for each unmatch. When the hotness value is lower than 30, mark it as a template to be retired, and need human confirmation for retention.

[0089] S2013, according to the interface adapter initialization parameter, build a lightweight data conversion rule engine, convert heterogeneous data including SCORM, xAPI, IMS-QTI into unified JSON-LD format, and obtain intermediate standardized data blocks;

[0090] ‌Rule engine architecture design‌

[0091] ‌Modular processing pipeline: Build a three-stage pipeline containing Extractor, Transformer, Validator. For example, Extractor parses SCORM XML from a file <adlcp:score>The field, Transformer converts it to a float number, Validator checks the value domain is between 0-1.

[0092] ‌Parallel execution optimization‌: Use multi-thread pool (e.g. 8 threads) to process different data shards in parallel. For example, group xAPI statements by actor, assign each group an independent thread to process, increase throughput to 1000 per second.

[0093] ‌Memory management strategy‌: Set the upper limit of JVM heap memory to 4GB, use zero-copy technology to reduce serialization overhead. For example, use Protocol Buffers to directly map JSON-LD structure, avoid intermediate format conversion.

[0094] ‌Heterogeneous data format conversion‌

[0095] ‌SCORM→JSON-LD conversion‌:

[0096] Field mapping: Convert <adlcp:location>Mapping to {"@context": "http: / / schema.org", "@graph": [{"@type": "Place", "name": "Chapter 3"}]}.

[0097] Flattening of nested structures: Will be <sequencing> <controlmode choice="false" / > Convert to "sequencing.controlMode.choice": false.

[0098] ‌xAPI → JSON-LD Conversion‌:

[0099] Verb Standardization: Abbreviate verb: "http: / / adlnet.gov / expapi / verbs / completed" to "verb": "completed" and add @context link to xAPI official vocabulary.

[0100] Timestamp Formatting: Unify "timestamp": "2023-10-01T12:34:56.789Z" to ISO 8601 extended format.

[0101] ‌IMS-QTI → JSON-LD Conversion‌:

[0102] Question Entity: Convert <qti-assessment-item identifier="math_quiz">Converts to {"@type": "AssessmentItem", "identifier": "math_quiz", "question": "..."}.

[0103] Response processing: Will <responsedeclaration identifier="RESPONSE">Mapping as "response": {"identifier": "RESPONSE", "correctValue": "A"}.

[0104] ‌Interim standardized data block generation‌

[0105] ‌Data consistency check: Verify converted data against JSON Schema to ensure compliance. For example, check if the mandatory field @context exists and timestamp is a valid date.

[0106] ‌Lightweight compression: Apply GZIP compression algorithm (compression level = 6) to reduce storage overhead, reducing the average data volume from 1MB per thousand to 200KB per thousand.

[0107] ‌Batch packaging and indexing: Package every 1000 JSON-LD data into an interim data block (file name format: block_20231001_001.ndjson), and generate an index file (record start timestamp, platform type, data volume).

[0108] S2014, Time series alignment processing of interim standardized data blocks, using sliding window mechanism to eliminate cross-platform timestamp deviation, and through event-driven pipeline output continuous and conflict-free standardized time series data stream.

[0109] ‌Timestamp deviation detection and correction‌

[0110] ‌Global clock synchronization: Based on NTP protocol to calibrate the clock of each platform, the maximum allowed deviation is ±500ms. For records that exceed the threshold (such as Canvas timestamp is 2 seconds slower than Moodle), add deviation marker timestamp_offset: +2000ms.

[0111] ‌Sliding window alignment: Set window size to 5 seconds, sliding step to 2 seconds. For example, sort all events (such as video playback, test submission) in the window 12:00:00-12:00:05 by actual occurrence time to generate an ordered event sequence.

[0112] ‌Causal relationship repair: Through event dependency analysis to correct time series conflicts. For example, if a student submits a test at 12:00:03, but the video playback record shows completion at 12:00:04, automatically adjust the video end time to 12:00:02 to maintain the causal relationship.

[0113] ‌Event-driven pipeline design‌

[0114] ‌State management engine: Maintain a session state machine (e.g., state = in progress, paused, completed) for each learner. When a pause_video event is received, update the state and pause subsequent event processing until a resume_video event is received.

[0115] ‌Conflict detection and resolution: Use Vector Clock algorithm to mark event concurrency. For example, if two platforms report a start_quiz event for the same student at the same time (Vector Clocks are [Moodle:2, Canvas:1] and [Moodle:1, Canvas:2]), sort by platform priority (Moodle > Canvas).

[0116] ‌Watermark generation: Advance global progress based on event timestamps, with a 10-second allowed delay. For example, when processing an event with a timestamp of 12:00:10, assume that all events ≤ 12:00:00 have arrived and trigger window calculations.

[0117] ‌Standardized time-series data stream output

[0118] ‌Data serialization and distribution: Convert time-aligned data to Avro format (Schema includes timestamp, user_id, event_type, payload) and distribute to downstream systems via Kafka topic standardized_events.

[0119] ‌End-to-end latency monitoring: Insert instrumentation at each stage of the pipeline (receiving, converting, aligning) to monitor that the 95th percentile latency does not exceed 800ms. If it does, dynamically scale processing nodes (e.g., from 8 to 12 nodes).

[0120] ‌Data quality closed-loop feedback: Feedback time-series conflict rates (e.g., 0.07%) and field fill rates (e.g., 99.3%) to the interface adapter, triggering rule engine parameter adjustments (e.g., relax timestamp deviation threshold to ±1 second).

[0121] ‌Application scenario example

[0122] ‌Scenario description: A cross-platform online course uses Moodle (SCORM), Canvas (xAPI), and a self-developed quiz system (IMS-QTI) simultaneously, and needs to integrate data in real time to analyze learning behavior.

[0123] ‌Data access: Kafka receives Moodle's activity_001.xml (SCORM 1.2), Canvas's xapi_20231001.json (xAPI 1.0.1), and self-developed system's qti_math.ndjson (IMS-QTI 2.1).

[0124] ‌Feature extraction: Bi-LSTM identifies the SCORM data in <sequencing>Nested schema of tags, output feature vector [0.92, 0.45,...].

[0125] ‌Template matching‌: Density peak clustering matches to SCORM 1.2 template (similarity = 0.95), initializes field mapping rules.

[0126] ‌Rule conversion‌: Converts <adlcp:score> 0.9< / adlcp:score> to {"score": 0.9, "@context": "http: / / schema.org"}.

[0127] ‌Temporal alignment‌: Corrects Canvas timestamp bias (+1.2 seconds), sorts event stream by sliding window 12:00:00-12:00:05.

[0128] ‌Output result‌: Standardized data stream in Avro format is sent to recommendation system, triggering personalized learning path computation.

[0129] This step extracts the temporal features of the data stream (such as timestamp continuity, field arrangement rules) by using Bi-LSTM, identifies implicit format patterns (such as the nested relationship between course video metadata and log behavior data), and converts multi-source data into unified JSON-LD format through a dynamic interface adapter (based on density peak clustering to generate a template library). The timestamps are aligned using a sliding window to eliminate cross-platform time deviations, and finally a standardized temporal data stream is generated, achieving semantic consistency and temporal synchronization of heterogeneous data, solving the problem of mixed formats and temporal misalignment of multi-source teaching data. Through a dynamic interface adapter, automatic data integration is achieved, providing a unified input for subsequent encrypted storage and feature fusion, avoiding the low efficiency and compatibility conflicts caused by manual intervention.

[0130] S202, based on the standardized temporal data stream, the teaching data is classified and encrypted through a dynamic hierarchical storage strategy based on attribute-based encryption, and a knowledge distillation algorithm is used to compress unstructured data and generate a unique data fingerprint, obtaining a secure hierarchical storage topology; specifically, it can include:

[0131] S2021, according to the metadata attributes of the standardized temporal data stream, the data sensitive level label is extracted through the attribute-based encryption algorithm, the CP-ABE strategy is used to define the access control tree, and the dynamic hierarchical encryption strategy is generated;

[0132] ‌Metadata sensitive feature extraction and label generation‌

[0133] ‌Metadata analysis and feature extraction: The metadata of the standardized time-series data stream contains field types (e.g., user_id, course_id), data sources (e.g., platform types, institution codes), and content categories (e.g., videos, logs, test results). The Multi-Dimension Sensitivity Network (MDSN) is used to analyze the privacy correlation of fields. For example, MDSN uses a three-layer fully connected network (input layer dimension = 64, hidden layer dimension = 32, output layer dimension = 3), and the input features include field types (numeric encoding), access frequency (log statistics), and associated entities (e.g., whether associated with student information).

[0134] ‌Sensitivity stratification: Hierarchical clustering algorithm is used to divide the data into three levels of sensitive labels:

[0135] ‌High sensitivity (L3): Student information (e.g., ID number, home address), and raw test data (e.g., test answers).

[0136] ‌Medium sensitivity (L2): Classroom behavior logs (e.g., login time, video viewing duration), and test statistics (e.g., average score, pass rate).

[0137] ‌Low sensitivity (L1): Public course videos (no copyright restrictions), and teaching syllabus texts.

[0138] For example, for the student data of user_id = student_001, MDSN outputs a sensitivity vector [0.92, 0.15, 0.03], indicating that L3 has the highest weight.

[0139] ‌Dynamic label update: The reconstruction error (MSE) of the input data is calculated by the autoencoder (Autoencoder) to serve as a sensitivity correction factor. For example, if a video is frequently accessed (abnormally high access frequency), and the reconstruction error of the autoencoder exceeds the threshold (e.g., 0.1), its sensitive label will be upgraded from L1 to L2.

[0140] ‌CP-ABE access control tree construction

[0141] Policy Definition and Attribute Binding: Ciphertext-Policy Attribute-Based Encryption (CP-ABE) defines access tree structures. For example, for high-sensitive data (L3), set the access policy as (role=teacher AND department=math) OR (role=admin AND time<2024-12-31), i.e., only math department teachers or administrators before the end of 2024 can decrypt.

[0142] Dynamic Decaying Factor Design: Add a decaying weight for the time attribute in the access policy (e.g., time<2024-12-31). For example, use an exponential decay function weight = exp(-λ*(current_time - grant_time)), where λ=0.001 controls the decay rate, and automatically revoke access when the weight is below 0.5.

[0143] Policy Public Key Generation: Use Elliptic Curve Cryptography (ECC) to generate a policy public key matrix. For example, use the secp256k1 curve to generate a public key PK = (x=0x3B9ACA07, y=0x76B5821D), a private key SK = 0x1A2B3C4D, and verify the matching of attributes and policies through Bilinear Pairing.

[0144] Dynamic Hierarchical Encryption Policy Generation

[0145] Policy Optimization and Coupling: Optimize the coupling of access control trees and sensitive labels. For example, add a regional constraint (e.g., location=China) to the access tree of medium-sensitive data (L2) and bind it to the second dimension of the sensitive vector (medium-sensitive weight).

[0146] Space-Time Constraint Injection: When cross-campus data requests are detected, the policy evolution engine automatically adds a two-factor authentication condition (e.g., SMS verification code + fingerprint). For example, when a user accesses from an overseas IP, the policy is updated to (role=teacher AND location=China) + OTP.

[0147] Key Distribution and Synchronization: Based on a distributed key management system (e.g., HashiCorp Vault), generate a unique key shard for each policy node and synchronize it to each storage node through the Gossip protocol, ensuring that at least 3 shards are combined (threshold=3) when decrypting.

[0148] S2022, based on the dynamic hierarchical encryption strategy, the teaching video data and the text log data are classified and encrypted to generate hierarchical ciphertext data blocks;

[0149] ‌Teaching Video Data Encryption‌

[0150] ‌Video Segmentation and Feature Extraction‌: Cut the video into 10-second segments (e.g., video_001_segment_0001.mp4) based on key frames, extract optical flow features (dimension = 512) and color histograms (dimension = 256) using OpenCV, and generate feature vectors as encryption auxiliary information.

[0151] ‌Selective Encryption and Compression‌: Encrypt all frames (AES-256-CBC mode, key length = 256 bits) for high-sensitive videos (e.g., paid courses), and only encrypt I-frames (key frames) for low-sensitive videos (e.g., public courses), while P / B frames remain unencrypted to save computing resources. After encryption, use H.265 encoding compression, which improves compression rate by 40%.

[0152] ‌Dynamic Key Binding‌: Bind video segment keys (e.g., key = 0x5E6F7A8B) with access control tree attributes. For example, if the access policy is role = student AND course_enrolled = true, the key is only visible to students who have enrolled in the course.

[0153] ‌Text Log Data Encryption‌

[0154] ‌Structured Log Encryption‌: Field-level encryption of JSON-formatted logs (e.g., {"user_id": "student_001", "action": "start_video", "timestamp": 1630454400}):

[0155] High-sensitive fields (e.g., user_id): Use CP-ABE encryption, and the ciphertext format is {"encrypted_user_id":"0xA1B2C3D4...", "policy": "(role=admin)"}.

[0156] Low-sensitive fields (e.g., action): Use lightweight encryption algorithms (e.g., ChaCha20-Poly1305, key length = 128 bits).

[0157] Unstructured log processing: For free text logs (such as teacher comments), use Named Entity Recognition (NER) models (such as BERT-Base) to automatically identify sensitive entities (such as student names) and encrypt only the entity part (such as "Zhang San" → "0xE5F6G7H8").

[0158] Log index optimization: Build an inverted index for encrypted logs. Index fields (such as timestamp and action_type) retain plaintext to support fast retrieval, but the index key is signed with HMAC-SHA256 to prevent tampering.

[0159] Hierarchical ciphertext data block generation

[0160] Data block encapsulation: The encrypted video segments and log records are packaged into data blocks according to time windows (e.g., 5 minutes). Each block contains:

[0161] Header information: Block ID (e.g., block_20231001_001), list of encryption algorithms (e.g., ["AES-256", "CP-ABE"]), timestamp range (start=1630454400, end=1630454700).

[0162] Data body: Encrypted video segments and log records.

[0163] Tail check: CRC32 checksum (e.g., 0x78ABCDEF) and digital signature (RSA-2048).

[0164] Hierarchical storage tagging: Tagging data blocks based on sensitivity labels:

[0165] L3 data blocks: marked in red, stored on a local encrypted hard drive (read / write speed ≤200MB / s).

[0166] L2 data blocks: marked in yellow, stored in a distributed file system (such as Ceph, replica count = 3).

[0167] L1 data blocks: marked in green, stored in object storage (such as Amazon S3, compressed storage).

[0168] S2023 employs a knowledge distillation algorithm for unstructured video data, extracts key teaching features through a teacher model, trains a lightweight student model to generate compressed feature vectors, and outputs a lightweight video feature package that preserves semantics.

[0169] Teacher model design and feature extraction

[0170] ‌Multi-modal teacher model architecture‌: The teacher model adopts a hybrid structure of ResNet-101 (image branch) + Transformer (temporal branch):

[0171] Image branch: Input video keyframes (resolution = 224x224), extract spatial features (dimension = 2048) through ResNet-101.

[0172] Temporal branch: Input optical flow sequence (length = 30 frames), extract temporal features (dimension = 512) through Transformer encoder (number of layers = 6, number of heads = 8).

[0173] Feature fusion: Concatenate spatial and temporal features (total dimension = 2560), output teaching semantic features (such as knowledge point correlation, student attention hot area) through fully connected layer (dimension = 1024).

[0174] ‌Key feature distillation‌: Use gradient backpropagation saliency map (Grad-CAM) to identify the feature area that has the greatest impact on teaching effectiveness. For example, the key frame area where the teacher writes the blackboard in the video (coordinates x = 100-200, y = 50-150) is given higher weight.

[0175] ‌Student model training and compression‌

[0176] ‌Lightweight student model design‌: The student model adopts MobileNetV3 (image branch) + Temporal Convolutional Network (temporal branch):

[0177] Image branch: Inverted residual structure of MobileNetV3 (parameters = 2.5M), output feature dimension = 512.

[0178] Temporal branch: Dilated convolution of TCN (number of layers = 4, dilation factor = 2), output feature dimension = 256.

[0179] Feature fusion: Channel attention mechanism (SENet Block) dynamically weights the fused features, and the final output dimension = 768.

[0180] ‌Distillation loss function‌: Define the joint loss of mean square error (MSE) and cosine similarity:

[0181] MSE loss: Restrict the feature distance between student model output and teacher model (such as MSE = 0.05).

[0182] Cosine loss: Ensure feature direction consistency (such as cosθ = 0.92).

[0183] ‌Training optimization strategy: Adam optimizer (learning rate = 0.001, β1 = 0.9, β2 = 0.999), mixed precision training (FP16), batch size = 128.

[0184] ‌Lightweight feature package generation

[0185] ‌Feature compression and encoding: The 768-dimensional feature vector output by the student model is reduced to 128 dimensions using PCA, and then converted to 8-bit integers (e.g., 0.75 → 192) using scalar quantization, increasing the compression rate by 6 times.

[0186] ‌Semantic integrity verification: Randomly sample video clips and manually annotate knowledge coverage (e.g., 0-5 points) to verify the semantic consistency between compressed features and original videos (e.g., average score difference <0.2).

[0187] ‌Feature package packaging: Package compressed features, timestamps, and video IDs into Protobuf format (e.g., feature_pkg_001.bin) and attach an HMAC signature (key = 0x9A8B7C6D) for tamper protection.

[0188] ‌S2024, integrate the hierarchical ciphertext data blocks and lightweight video feature packages into a hash chain, generate a unique data fingerprint based on the content features of the hierarchical ciphertext data blocks, and construct a secure hierarchical storage topology with access control labels.

[0189] ‌Hash chain construction and data association

[0190] ‌Data block hash generation: Calculate the SHA-256 hash value (e.g., hash_block_001 = 0x1A2B3C4D...) for each hierarchical ciphertext data block and embed the hash value of the previous block in the current block header to form an unalterable hash chain. For example, block_002 header contains prev_hash = 0x1A2B3C4D.

[0191] ‌Feature package association: Write the hash value of the lightweight video feature package (e.g., hash_feature_pkg_001 = 0x5E6F7A8B) into the metadata field of the corresponding data block to establish a bidirectional index between video content and encrypted data.

[0192] ‌Cross-block verification: If a data block is modified (e.g., tampered with log records), its hash value will change, causing prev_hash verification failure for all subsequent blocks, triggering an audit alarm.

[0193] ‌Unique data fingerprint generation

[0194] ‌Content Feature Extraction of Hierarchical Cipher Data Blocks: Extract key teaching features (e.g., knowledge point density = 0.8, attention peak time = 120s) from lightweight feature packages and convert them into hexadecimal strings (e.g., "0x8F" represents knowledge point density).

[0195] ‌Fingerprint Synthesis Algorithm: Aggregate multi-dimensional features using Merkle Tree structure:

[0196] Leaf Node: Single Feature Hash (e.g., hash(knowledge_density) = 0xA1B2).

[0197] Intermediate Node: Hash the concatenation of adjacent leaf nodes (e.g., hash(0xA1B2 + 0xC3D4) = 0xE5F6).

[0198] Root Node: Final Data Fingerprint (e.g., root_hash = 0x9A8B).

[0199] ‌Fingerprint Binding and Query: Write data fingerprints to the blockchain (e.g., Hyperledger Fabric) to support reverse retrieval of original data blocks and feature packages through fingerprints.

[0200] ‌Secure Hierarchical Storage Topology Construction

[0201] ‌Storage Level Division:

[0202] ‌Hot Storage Layer (In-memory Cache): Store low-sensitive data (L1) and frequently accessed feature packages, with response time ≤10ms.

[0203] ‌Warm Storage Layer (SSD Array): Store medium-sensitive data (L2) and associated hash chain indexes, with response time ≤100ms.

[0204] ‌Cold Storage Layer (Tape Library): Store high-sensitive data (L3) and full backups, with response time ≥10s.

[0205] ‌Access Control Label Embedding: Label each storage node (e.g., label = {sensitivity:L2, location:Beijing-Campus-3}) and automatically execute policies (e.g., prohibit cross-region access to L3 data) through smart contracts.

[0206] ‌Dynamic Topology Optimization: Based on access pattern analysis (e.g., Q-Learning algorithm), dynamically migrate data blocks to balance load. For example, migrate L2 data blocks with access frequency >1000 within a week from warm storage to hot storage.

[0207] ‌Application Scenario Example

[0208] ‌Scenario: A university needs to securely store math course videos (L2 sensitive) and student exam logs (L3 sensitive).

[0209] ‌Encryption and Compression:

[0210] Video chunk encryption: Use AES-256 to encrypt key frames (I-frames), reducing the volume from 1GB to 600MB after compression.

[0211] Log field encryption: Student ID is encrypted using CP-ABE, with the policy (role=teacher AND course=math101).

[0212] ‌Knowledge Distillation:

[0213] Teacher model extracts the "derivation of calculus formulas" key segment from the video (feature vector = 0.82).

[0214] Student model compresses features to 128 dimensions, generating feature package math101_segment_0001.fpkg.

[0215] ‌Storage Topology Construction:

[0216] Video is stored in warm storage (SSD array), and logs are stored in cold storage (tape library).

[0217] Data fingerprint 0x9A8B is written to the blockchain for the academic department to verify data integrity.

[0218] Additionally, based on the metadata attributes of standardized time series data streams, data sensitivity level tags are extracted through attribute-based encryption algorithms, and CP-ABE policy is used to define access control trees to generate dynamic hierarchical encryption policies, which can include:

[0219] Based on the metadata attributes of standardized time series data streams, the privacy correlation of network analysis data fields is extracted through multi-dimensional sensitive feature extraction, and hierarchical clustering algorithm and self-encoder fusion architecture are used to map student information, score records, and classroom behavior logs to different sensitive levels, outputting a three-dimensional feature vector of data sensitivity;

[0220] ‌Multi-dimensional sensitive feature extraction network design

[0221] ‌Input data definition: Metadata attributes of standardized time series data streams include field types (such as ID number and name in student information), data sources (such as LMS system and exam platform), update frequency (such as monthly update of student information and real-time update of classroom behavior logs), etc. For example, the student information field contains the field name student_id, field type PII (personal identification information), and data source academic system.

[0222] ‌Privacy correlation calculation: Calculate the privacy correlation of each field through a sensitive feature extraction network (network structure: input layer 256-dimension → fully connected layer 128-dimension → ReLU → output layer 3-dimension). Supervised learning is used for network training, and the labeled data includes three types of sensitive labels (high sensitive: such as ID number; medium sensitive: such as grades; low sensitive: such as classroom interaction times). Training parameters: batch_size=256 (process 256 fields per batch), epochs=100 (train for 100 rounds), and the loss function is cross-entropy.

[0223] ‌Hierarchical clustering and autoencoder fusion:

[0224] ‌Hierarchical clustering: Cluster the privacy correlation vector of the field (3-dimension), set the cluster number k=3 (corresponding to high / medium / low sensitive level), and use Ward variance minimization algorithm to calculate the distance between clusters. For example, the student information field is clustered into the high sensitive cluster (cluster center [0.92, 0.85, 0.78]), and the grades field is clustered into the medium sensitive cluster (cluster center [0.65, 0.72, 0.61]).

[0225] ‌Autoencoder: Design an encoder (input 3-dimension → hidden layer 2-dimension → output 3-dimension) and a decoder (symmetric structure), reconstruct the privacy correlation vector through unsupervised learning, and extract the latent sensitive features. Training parameters: learning rate lr=0.001, and the optimizer is Adam.

[0226] ‌Three-dimensional feature vector generation: Concatenate the cluster label (such as high sensitive=2, medium sensitive=1, low sensitive=0) and the autoencoder latent feature (2-dimension) to output a three-dimensional vector (for example, the student information field vector is [2, 0.87, -0.43]).

[0227] ‌Sensitive level mapping and verification

[0228] ‌Sensitive level division rule: Define the sensitive level according to the first dimension value (cluster label) of the three-dimensional feature vector:

[0229] ‌High sensitive level (Level 3): Includes student information, ID number, and home address, and requires dual-factor authentication for access.

[0230] ‌Medium sensitive level (Level 2): Includes grades, test answers, and course assignments, and requires role permission verification (such as teachers or students themselves) for access.

[0231] ‌Low sensitive level (Level 1): Includes classroom behavior logs, interaction times, and resource click rates, allowing anonymous statistical access.

[0232] Data verification and correction: Manual correction of automatic classification results through a human review module (e.g., an administrative interface). For example, if the system mistakenly classifies "class attendance rate" as a medium-sensitive layer, the administrator can manually adjust it to a low-sensitive layer and feed back to the training data set to optimize the model.

[0233] Based on the data sensitivity three-dimensional feature vector, a dynamic attenuation type decision tree is used to generate a fine-grained access strategy, wherein the course video data is divided into access permissions according to resolution and content sensitivity, and the test data is defined as decryption conditions according to the performance distribution interval, and an initial access control rule set with timestamp is output;

[0234] Dynamic attenuation type decision tree construction

[0235] Definition of decision tree node: Design decision conditions according to data sensitivity three-dimensional feature vector:

[0236] Root node: Determine the data sensitivity level (Level 1 / 2 / 3).

[0237] Branch node: Refine conditions for different data types. For example, for course video data, branch conditions include resolution (such as 1080P, 720P) and content sensitivity (such as videos containing student portraits = high sensitivity, teaching animations = low sensitivity).

[0238] Leaf node: Define access permissions (such as readable, editable, prohibited access) and decryption conditions (such as role = teacher and IP belongs to school network).

[0239] Dynamic attenuation mechanism: Add a time attenuation factor (initial weight = 1.0, daily attenuation rate = 0.05) to each access rule, and automatically invalidate when the threshold (such as weight < 0.3) is exceeded. For example, an access rule for final exam papers has an initial validity period of 30 days, and the weight decays to 0.25 after 30 days, triggering automatic revocation.

[0240] Fine-grained access strategy generation

[0241] Course video data permission division:

[0242] Resolution condition: 1080P video only allows teachers and authorized students to access (permission tags r:teacher, s:authorized), and 720P video allows all students to access (tags r:student).

[0243] Content sensitivity condition: Videos containing student portraits need to be checked for privacy agreement signing status (such as field consent_status=True).

[0244] Test data decryption condition:

[0245] ‌Grade interval strategy: Students with scores ≥ 90 can view rankings (decryption condition score ≥ 90), and students with scores < 60 can only view mistake analysis (condition score < 60 & role = student).

[0246] ‌Time window restriction: Test data is allowed to be decrypted 24 hours after the exam ends (timestamp condition timestamp ≥ exam_end + 24h).

[0247] ‌Rule set storage and indexing: Use Redis to store timestamped access control rules, with the key as the data ID (e.g. video_001) and the value as the JSON format rule (e.g. {"resolution": "1080P", "access_roles": ["teacher"], "expire_weight": 0.95}), supporting real-time updating and querying (response time < 10ms).

[0248] According to the context information of the real-time teaching scene, the initial access control rule set is dynamically modified through the strategy evolution engine. When a cross-campus data sharing request is detected, a geographical constraint attribute is automatically added and a two-factor authentication condition is introduced, generating a spatiotemporally aware enhanced access control tree;

[0249] ‌Strategy evolution engine workflow

[0250] ‌Context information collection: Real-time collection of spatiotemporal information (such as the geographical location of the request source IP, the current time period), user behavior (such as frequent access to abnormal data), and device status (such as mobile / PC). For example, detecting a request from a school district (IP belongs to Shanghai, main campus in Beijing), marked as cross-campus access.

[0251] ‌Dynamic rule modification mechanism:

[0252] ‌Regional constraint addition: When the request source is not within the data's school district range, automatically add a geographical location attribute (such as region: Shanghai) and limit access methods (only allow VPN access).

[0253] ‌Two-factor authentication trigger: For requests to access high-sensitive data (such as student information) across school districts, require a short message verification code (field 2fa_code) or biometric identification (such as fingerprint).

[0254] ‌Spatiotemporal aware rule generation: Merge the modified rules into the access control tree, for example, add a branch condition:

[0255] IF Data Sensitivity Level = Level 3 AND Request Origin = Cross-campus THEN Require Two-Factor Authentication AND Restrict Access Time (9:00-18:00).

[0256] ‌Enhanced Access Control Tree Optimization‌

[0257] ‌Conflict Detection and Resolution‌: Use a rule conflict detection algorithm (such as the RETE algorithm) to check for inconsistencies between new rules and existing rules. For example, if an existing rule allows teachers to access all day, and a new rule restricts cross-campus access to 9:00-18:00, an exception rule is generated: Role = teacher AND Origin = cross-campus → Allow access all day.

[0258] ‌Encoding of Temporal and Spatial Constraints‌: Map geographic locations to GeoHash encodings (e.g., Beijing main campus = wx4g) and time periods to minute-level granularity (e.g., 9:00 = 540 minutes). In the access control tree, temporal and spatial conditions are treated as joint decision nodes (e.g., GeoHash IN [wx4g, wx4e] AND time BETWEEN 540 AND 1080).

[0259] Coupling the enhanced access control tree with the three-dimensional feature vector of data sensitivity, using the asymmetric bilinear pairing algorithm to calculate the node weight, generating a public key matrix through elliptic curve encryption technology, and finally outputting the CP-ABE hierarchical encryption strategy containing dynamic attenuation factors and temporal and spatial constraints.

[0260] ‌Policy Coupling and Weight Calculation‌

[0261] ‌Node Weight Definition‌: Each decision node of the access control tree (e.g., "Resolution = 1080P") is associated with a weight, which is calculated by the second dimension of the three-dimensional feature vector of data sensitivity (autoencoder latent feature 1) and the asymmetric bilinear pairing algorithm. For example, using a bilinear mapping function (e.g., e: G1×G2→GT, where G1 / G2 is an elliptic curve group), calculate the node weight: weight = e(H(node condition), data sensitivity vector). H is a hash function that maps node conditions to points on an elliptic curve.

[0262] Dynamic Attenuation Factor Fusion‌: Multiply the time attenuation factor (e.g., 0.95) by the node weight to get the final weight (e.g., 0.95×0.87=0.8265). When the weight is lower than the threshold (0.3), the node condition is automatically disabled.

[0263] ‌CP-ABE Strategy Generation‌

[0264] Elliptic curve parameter selection: Using the SM2 standard (elliptic curve equation y² = x³ + ax + b, where a=0, b=7, prime field p=FFFFFFFE...), a public / private key pair is generated (public key PK=(G, g^α), master private key MSK=α). G is the base point in the public parameters, used to construct the encryption algorithm; g is the generator (possibly with or across groups), used to construct the public key; α is the master private key, the core secret of the system, combined with G and g to generate the public key g^α.

[0265] Policy public key matrix construction: Map each leaf node of the access control tree (e.g., "Allow Teacher Access") to a row of the public key matrix. For example, for the rule Role = teacher AND Resolution = 1080P, generate the public key component: PK_row = g^{α * H(role||resolution)}.

[0266] Policy Distribution and Storage: The final CP-ABE policy includes a dynamic decay factor table (stored in MySQL), spatiotemporal constraints (encoded as JSON rules), and a public key matrix (stored as a binary file), and distributes attribute private keys to authorized users through a key distribution center (KDC).

[0267] Application scenario examples

[0268] Scenario Description: A university needs to encrypt course videos shared across campuses.

[0269] Sensitivity level marker:

[0270] In the video metadata, the resolution is 1080P, and it contains student portraits → the sensitivity 3D vector is [2, 0.91, -0.37], which is classified as a high-sensitivity level.

[0271] Access policy generation:

[0272] Initial rules: Access is allowed for faculty members of this school (no two-factor authentication required), valid for 30 days (initial decay factor = 1.0).

[0273] Cross-campus access triggers policy correction:

[0274] Access request to Shanghai branch detected → Automatically add region constraint (region=Shanghai) and two-factor authentication conditions.

[0275] CP-ABE Strategy Released:

[0276] The attribute private key of teacher A contains role = teacher and region = Beijing, and can decrypt the video of Beijing main campus; the attribute private key of teacher B contains role = teacher and region = Shanghai, and needs to be decrypted after two-factor authentication.

[0277] For sensitive information in standardized data streams (such as student records and test scores), an attribute-based encryption (CP-ABE) policy is defined to define an access control tree (such as "only class teacher decrypts score field"), and an encryption policy is dynamically generated according to the data sensitivity level (such as high-sensitive video and low-sensitive log). For unstructured video data, a knowledge distillation algorithm is used to compress the key features (such as teaching action recognition and knowledge point labeling) extracted by the teacher model (complex network) to a lightweight student model, generate a compressed feature package that retains semantics, and integrate encrypted data and compressed features through a hash chain to generate a unique data fingerprint (such as SHA-3 hash value), construct a hierarchical storage topology with access permission labels, achieve a balance between data security and storage efficiency, optimize storage resource occupation under the premise of ensuring data privacy, reduce storage cost through dynamic hierarchical encryption and feature compression, and at the same time, use unique fingerprint to realize data traceability and tamper-proofing, and meet the compliance requirements of educational data.

[0278] S203, according to the user query request and the teaching scene label, extracting the target data set from the secure hierarchical storage topology structure, through the multi-modal feature fusion network, the text, video, behavior log data are jointly embedded and represented, and the missing features are completed by using the adversarial generation network, and the enhanced multi-modal teaching data set is obtained; specifically, it can include:

[0279] S2031, according to the keyword vector in the user query request and the teaching scene label, through the semantic similarity algorithm, the target data is retrieved in the secure hierarchical storage topology structure, and the candidate data set index table is generated;

[0280] ‌Key word vectorization and teaching scene label analysis‌

[0281] ‌Key word vector generation‌: The keywords (such as "calculus teaching video") in the user query request are converted into 768-dimensional semantic vectors by the pre-trained multilingual BERT model (bert-base-uncased, parameter quantity 110M). For example, "calculus" is encoded as a vector [0.23, -0.45,..., 0.89], and the context (such as "teaching video") is weighted and fused to generate the query vector query_vector.

[0282] ‌Teaching scenario label matching‌: Teaching scenario labels (e.g. "high school math review class") are mapped to 32-dimensional vectors by a label encoder. The label encoder uses a fully connected network (input layer = 64, hidden layer = 32) with input being the TF-IDF weights of the labels (e.g. "review class" weight = 0.8) and scenario type encoding (e.g. "classroom" = 1, "lab" = 2).

[0283] ‌Joint similarity computation‌: The query vector query_vector is concatenated with the scenario label vector scene_vector to form an 800-dimensional joint vector, which is matched with the metadata descriptions of the data blocks in the secure hierarchical storage topology using the cosine similarity algorithm (threshold = 0.7). For example, the metadata vector metadata_vector of the matched video_block_001 has a similarity of 0.82, which is determined as a candidate data.

[0284] ‌Efficient retrieval in secure hierarchical storage

[0285] ‌Distributed inverted index construction‌: An inverted index is constructed for each data block in the secure hierarchical storage topology, with index keys including:

[0286] ‌Content keywords‌ (e.g. "calculus", "derivative") extracted by NLP segmentation tools (e.g. Jieba).

[0287] ‌Sensitive labels‌ (e.g. L1 / L2 / L3) for quick filtering of unauthorized data.

[0288] ‌Time range‌ (e.g. timestamp_start = 1630454400) divided into sliding windows (window size = 1 hour).

[0289] ‌Semantic similarity retrieval acceleration‌: High-dimensional vectors are searched using the Faiss library (Facebook AI Similarity Search) for approximate nearest neighbors (ANN). For example, configure HNSW32 (Hierarchical Navigable Small World) index, search space divided into 32 layers, recall rate > 95%, response time < 50ms.

[0290] ‌Candidate dataset index table generation‌: The retrieval results are sorted in descending order of similarity, and the index table structure is shown in Table 1:

[0291] Table 1

[0292] Data Block ID Similarity Sensitive Tag Storage Location video_block_001 0.92 L2 SSD_Node_3 log_block_045 0.85 L3 Tape_Archive_7

[0293] Index table filters out invalid data blocks via Bloom Filter (false positive rate = 0.1%).

[0294] ‌Dynamic permission verification and result filtering‌

[0295] ‌Access control policy matching: based on user roles (e.g., "student", "teacher") and query context (e.g., IP address, time), call Attribute-Based Encryption (ABE) policy engine to verify permissions. For example, student role can only access data with scene_label = public lecture and sensitivity ≤ L2.

[0296] ‌Data fingerprint integrity check: calculate SHA-256 hash value (e.g., hash = 0x1A2B3C...) for candidate data blocks, compare with stored fingerprints in the blockchain, and remove tampered or damaged data blocks (e.g., hash inconsistency rate > 0.01%).

[0297] ‌Index table optimization and caching: cache high-frequency accessed index table segments (e.g., "Mathematics Foundation Course") in Redis (TTL = 1 hour), use LRU (Least Recently Used) policy to evict low-frequency data, and improve cache hit rate by 40%.

[0298] ‌S2032, based on the candidate dataset index table, uses a multi-head attention mechanism to align text TF-IDF features, video optical flow features, and behavior log time series features, generating a multi-modal joint embedding matrix;

[0299] ‌Multi-modal feature extraction and standardization‌

[0300] ‌Text TF-IDF feature extraction: calculate TF-IDF (Term Frequency-Inverse Document Frequency) weights for teaching texts (e.g., courseware, test questions), and use NLTK library (Python) to generate 300-dimensional sparse vectors. For example, the term "function" has a TF-IDF of 0.65 in mathematics and 0.02 in Chinese.

[0301] ‌Video optical flow feature extraction: use OpenCV's Farneback algorithm to calculate the optical flow (Optical Flow) of consecutive video frames, generating a 512-dimensional dense vector. For example, the optical flow feature of video segment video_segment_001 represents the intensity of blackboard writing actions (peak value = 120px / s).

[0302] ‌Behavior log temporal encoding‌: The student behavior logs (e.g., "click video", "submit homework") are timestamped and fed into an LSTM network (hidden layer = 64) to extract temporal features, outputting a 128-dimensional vector. For example, a student's behavior of repeatedly pausing a video within 10 minutes is encoded as [0.34, -0.56,..., 0.78].

[0303] ‌Feature alignment with multi-head attention‌

[0304] ‌Feature dimension unification and projection‌: The text (300-dimensional), video (512-dimensional), and log (128-dimensional) features are projected to a unified dimension of 512 through fully connected layers (parameter matrices W_text, W_video, W_log), generating text_proj, video_proj, and log_proj.

[0305] ‌Multi-head attention computation‌: An 8-head attention mechanism with a Transformer architecture is used (Head = 8, dimension per head = 64):

[0306] ‌Query‌: The text feature text_proj serves as the Query, focusing on the associated information in the video and log.

[0307] ‌Key-Value‌: The video feature video_proj and the log feature log_proj serve as the Key and Value.

[0308] ‌Attention score‌: The Query × Key^T / sqrt(64) is calculated, and after Softmax normalization, the Value is weighted and aggregated.

[0309] For example, the "derivative" keyword in the text and the optical flow feature of the blackboard derivation formula in the video have an attention weight of 0.91 in the first 3 heads.

[0310] ‌Cross-modal feature fusion‌: The outputs of the 8 heads are concatenated into a 512 × 8 = 4096-dimensional vector, which is reduced to 512-dimensional through a fully connected layer to generate the joint embedding matrix joint_embedding.

[0311] ‌Optimization strategy for feature alignment‌

[0312] ‌Adversarial alignment loss‌: Domain Adversarial Training is introduced, which forces different modal features to be uniformly distributed in the embedding space through a Gradient Reversal Layer. For example, the MMD (Maximum Mean Discrepancy) loss of the text and video features decreases from 0.5 to 0.2.

[0313] ‌Dynamic weight adjustment: automatically adjust attention head weights based on feature quality (e.g., SNR of video features > 30dB). For example, reduce attention weight for low-light videos by 50%, and increase text weight by 30%.

[0314] ‌Memory optimization and parallel computing: use NVIDIA A100 GPU's Tensor Core to accelerate matrix multiplication (FP16 precision), with a batch size of 256, resulting in a 3x speedup in training.

[0315] ‌S2033, input the multi-modal joint embedding matrix into the generative adversarial network, identify the missing feature positions through the discriminator, and reconstruct the missing video key frames and text semantic segments through the generator, output the completed feature tensor;

[0316] ‌Adversarial network architecture design

[0317] ‌Generator structure: U-Net architecture (encoder-decoder), input is a 512-dimensional joint embedding matrix, output is the completed multi-modal feature tensor:

[0318] ‌Encoder: 4 layers of convolution (Kernel=3×3, Stride=2), doubling the number of channels at each layer (512→1024→2048→4096), extracting high-level semantic features.

[0319] ‌Decoder: 4 layers of deconvolution (Kernel=4×4, Stride=2), halving the number of channels at each layer (4096→2048→1024→512), restoring the original resolution.

[0320] ‌Skip connection: concatenate the output of each layer of the encoder with the corresponding layer of the decoder to preserve detailed information (e.g., video edges, text punctuation).

[0321] ‌Discriminator structure: PatchGAN design, input is a 512-dimensional feature tensor, output is the authenticity probability of each local region (Patch):

[0322] ‌Convolution layer: 5 layers of convolution (Kernel=3×3, Stride=2), with the number of channels increasing from 64 to 512.

[0323] ‌Fully connected layer: output a 16×16 authenticity probability matrix, each Patch corresponds to a confidence of 0~1.

[0324] ‌Missing feature recognition and reconstruction

[0325] ‌Missing position detection‌: The discriminator locates abnormal regions in the feature tensor (e.g., missing video keyframes, empty text paragraphs) through self-attention mechanisms. For example, if the activation values of the video feature segment video_feat[120:150] are detected to be below the threshold value 0.1, it is determined that there is a missing.

[0326] ‌Generator reconstruction logic‌:

[0327] ‌Video keyframe completion‌: Based on optical flow features and adjacent frame content, the generator predicts the pixels of the missing frames (e.g., fills in the blurred areas using bilinear interpolation).

[0328] ‌Text semantic recovery‌: Based on TF-IDF context, a GPT-2 small model (parameter size 124M) is used to generate coherent text (e.g., to supplement the missing "limit definition" description).

[0329] ‌Adversarial training strategy‌: The generator and discriminator are trained alternately, and the loss function includes:

[0330] ‌Adversarial loss‌: The cross-entropy loss between the discriminator output and the true label.

[0331] ‌Reconstruction loss‌: The L1 distance between the generated features and the real features (weight = 0.7).

[0332] ‌Perception loss‌: The feature similarity extracted by the VGG-16 network (weight = 0.3).

[0333] ‌Completion effect verification and optimization‌

[0334] ‌Manual evaluation‌: Randomly sample 10% of the completed data and have education experts mark the semantic coherence (e.g., 1-5 points), with an average score of ≥4.0.

[0335] ‌Automated testing‌:

[0336] ‌Video PSNR (Peak Signal-to-Noise Ratio)‌: The PSNR of the completed video and the original video is >30dB.

[0337] ‌Text BLEU score‌: The BLEU-4 score of the generated text and the reference answer is >0.65.

[0338] ‌Dynamic mask training‌: Randomly mask 20%~50% of the input features (simulate different missing rates) to enhance the robustness of the model.

[0339] ‌S2034, adversarial training optimization of the completed feature tensor, using gradient penalty strategy to stabilize the training process, generating robustness-enhanced intermediate multi-modal data set;

[0340] Gradient penalty strategy implementation

[0341] Wasserstein GAN optimization: Adopt the WGAN-GP (Wasserstein GAN with Gradient Penalty) framework, the discriminator output is a continuous value (Critic Score), and the loss function is: Loss = E[D(real)] - E[D(fake)] + λ * GP.

[0342] Where, the gradient penalty term GP = (||∇D(interpolated)||_2 - 1)^2, the interpolated sample interpolated = ε * real + (1-ε) * fake, ε~Uniform(0,1), λ=10.

[0343] Gradient clipping and normalization: Apply gradient clipping (Clip=0.01) after updating the discriminator weights to prevent gradient explosion; the generator uses spectral normalization (Spectral Normalization) to constrain the L2 norm of the weight matrix.

[0344] Improve the stability of adversarial training

[0345] Two-time-step update (TTUR): The discriminator updates 5 times as often as the generator (e.g., the discriminator updates 5 times per batch, and the generator updates 1 time), balancing the training rhythm.

[0346] Dynamic learning rate adjustment: Set the initial learning rate to 2e-4, use the cosine annealing (Cosine Annealing) strategy, period=1000 steps, minimum learning rate=1e-6.

[0347] History data playback: Keep the last 1000 batches of generated samples, replace the new samples in the current batch with a probability of 10%, to alleviate the mode collapse (Mode Collapse).

[0348] Intermediate dataset generation and enhancement

[0349] Feature tensor post-processing: Standardize the completed 512-dimensional feature tensor (mean=0, variance=1) and PCA dimension reduction (retain 95% variance, reduce dimension to 256), to reduce redundant information.

[0350] Data augmentation techniques:

[0351] Jittering: Add Gaussian noise (μ=0, σ=0.1) to the time series features of the behavior logs.

[0352] Elastic Distortion: Simulate camera shake effect on video optical flow features (α=50, σ=5).

[0353] Quality Filtering & Cleaning: Remove poor-quality samples with reconstruction loss >0.3 or discriminator confidence <0.5 to ensure dataset purity.

[0354] S2035: Separate modality-specific and common information in the intermediate multi-modal dataset using the feature decoupling algorithm, finally output an enhanced multi-modal teaching dataset with cross-modal association labels.

[0355] Modality Feature Decoupling Network Design

[0356] Decoupler Architecture: Use a variational autoencoder (VAE) structure, input 256-dimensional intermediate features, output decoupled specific (Specific) and common (Common) features:

[0357] Encoder: 3 layers of fully connected (256→128→64→32), output mean μ and variance σ.

[0358] Sampling Layer: Sample latent vector z = μ + ε * σ from N(μ, σ^2), ε~N(0,1).

[0359] Decoder: 3 layers of fully connected (32→64→128→256), reconstruct input features.

[0360] Decoupling Loss Function:

[0361] Reconstruction Loss: L2 distance between input and decoded output (weight=0.6).

[0362] Mutual Information Minimization: Mutual information (MI) between specific and common features is reduced to 0.05 through adversarial learning.

[0363] Orthogonal Constraint: Covariance matrix of specific feature matrix approaches diagonal matrix (non-diagonal elements <0.01).

[0364] Cross-Modal Association Label Generation

[0365] Association Degree Calculation: Calculate cross-modal cosine similarity on common features (e.g. video-text association degree = 0.75), generate association matrix, example matrix includes information like Table 2:

[0366] Table 2

[0367] Modality Pair Correlation Video-Text 0.82 Text-Log 0.68

[0368] Label Assignment Rules:

[0369] ‌Strong correlation‌ (>0.8): marked as cross_modal_strong, such as formula derivation in video and definition in text.

[0370] ‌Medium correlation‌ (0.5~0.8): marked as cross_modal_medium, such as answer duration in log and video watching times.

[0371] ‌Weak correlation‌ (<0.5): marked as cross_modal_weak, such as chapter title in text and video background music.

[0372] ‌Label propagation optimization‌: use label propagation algorithm (Label Propagation) to correct abnormal labels, iteration times = 100, similarity threshold = 0.7.

[0373] ‌Enhanced dataset packaging and verification‌

[0374] ‌Dataset structure‌: each data sample can contain specific features specific_feature, common features common_feature, cross-modal correlation labels cross_modal_label, metadata metadata, etc.

[0375] ‌Cross-modal consistency test‌: randomly select 1000 samples, manually verify label accuracy >85%.

[0376] ‌Downstream task adaptation‌: provide PyTorch and TensorFlow data loading interfaces (such as Dataset class), support batch loading (Batch Size=128) and distributed training.

[0377] Based on user queries (such as "class math lesson interaction effect analysis") and scene labels (such as "online teaching"), retrieve target data from encrypted storage, align text TF-IDF features (keyword weights), video optical flow features (teacher gesture dynamics), and behavior log time series features (student click frequency) through multi-head attention mechanism, generate multi-modal joint embedding matrix. Use the generator of the generative adversarial network (GAN) to reconstruct missing features (such as filling in the board content in the key frame of the video), the discriminator identifies abnormal data distribution, and the gradient penalty optimizes the generation quality, finally outputs the enhanced dataset containing cross-modal correlation labels, solves the problem of data fragmentation and missing, improves the integrity and correlation of multi-modal data, and enhances the analysis reliability by filling in the missing information through adversarial generation, provides high-quality input for knowledge graph construction, and supports deep mining of cross-modal teaching scenarios.

[0378] S204, based on the enhanced multi-modal teaching data set, a cross-institutional federated knowledge graph is constructed, a differential privacy protected graph attention mechanism is used to mine potential teaching association relationships, and an interpretable teaching mode graph is generated through a dynamic subgraph evolution algorithm; Specifically, it can include:

[0379] S2041, according to the entity relationship triplets in the enhanced multi-modal teaching data set, aggregate cross-institutional data through a federated learning framework, protect institutional privacy using homomorphic encryption technology, and generate an initial federated knowledge graph;

[0380] Entity relationship triplet extraction and federated data alignment

[0381] Cross-institutional entity recognition: Extract entities (such as "knowledge point: derivative", "teacher: Zhang teacher", "course: higher mathematics") from the cross-modal association labels of the enhanced multi-modal teaching data set, and use a pre-trained RoBERTa model (roberta-base, parameter quantity 125M) to perform semantic normalization on entity names. For example, "derivative" and "differential basis" are mapped to the unified entity IDentity_001.

[0382] Federated data alignment protocol: Each participating institution (such as School A, Education Platform B) exchanges the hash value (SHA-256) of the entity ID through a secure multi-party computation (Secure Multi-Party Computation, SMPC) protocol, retains the intersection entities (such as the common knowledge point "limit"), and eliminates non-common data (such as the "machine learning course" unique to institution B). After alignment, a global entity mapping table (global_entity_map.csv) is generated.

[0383] Homomorphic encryption data aggregation: Use the Paillier homomorphic encryption algorithm (key length = 2048 bits), and the institution locally trains the knowledge graph subgraph (such as the subgraph of institution A contains "derivative → calculus → Newton's theorem"), and encrypts the gradient parameters of the relationship triplets (head entity, relationship, tail entity). The federated server aggregates the encrypted gradients, decrypts the global graph after updating, and ensures that the original data does not leave the local. For example, the weight of the relationship "pre-knowledge point" is updated from 0.65 to 0.72.

[0384] Initial federated knowledge graph construction

[0385] Knowledge graph schema definition: Define the schema of the teaching mode graph, including entity types (knowledge points, teachers, courses), relationship types (pre-knowledge points, teaching styles, course dependencies), and constraint rules (such as "teacher → teaches → course" must satisfy "teacher ID ∈ active teacher list").

[0386] ‌Graph database storage optimization: Store triples using Neo4j graph database (version 4.4), node properties include knowledge point difficulty coefficient (difficulty = 0-1), teacher teaching experience (experience = 5 years), edge properties include relationship strength (strength = 0.8) and confidence (confidence = 0.95). To speed up the query, build a combined index (CombinedIndex): entity ID + relationship type.

[0387] ‌Federal model verification and tuning: Evaluate the graph quality through cross-validation (5-fold), calculate the relationship prediction accuracy (Accuracy) and F1 value. If the "knowledge point association" prediction F1 of institution A <0.7, trigger the federal retraining mechanism: adjust the aggregation weight (institution A weight from 0.5 to 0.3), re-encrypt and upload the gradient.

[0388] ‌S2042, based on the initial federal knowledge graph, design a differential privacy protected graph attention mechanism, perturb the attention weight between nodes by injecting Laplace noise, generate a privacy-safe relationship strength matrix;

[0389] ‌‌Design of differential privacy protected graph attention mechanism

[0390] ‌Graph attention network (GAT) modification: On the basis of standard GAT (Graph Attention Network), add differential privacy protection to the calculation process of node attention weight. The weight calculation formula of each layer attention head (Head = 4) is modified as: Attention_Weight = Softmax((Q·K^T) / sqrt(d) + Laplace(0, β)). Where Q and K are query and key vectors (dimension = 64), d = 64 is the scaling factor, and Laplace(0, β) is the Laplace noise (β = 0.1, privacy budget ε = 0.5).

[0391] ‌Noise injection strategy: Dynamically adjust the noise intensity according to the relationship sensitivity level. For example, edges involving student performance (sensitive label = High) inject β = 0.2 noise, and ordinary knowledge point relationships (sensitive label = Low) inject β = 0.05 noise. Noise distribution is generated through inverse transform sampling (Inverse Transform Sampling) to ensure compliance with the strict definition of differential privacy ((ε, δ)-DP, δ = 1e-5).

[0392] Privacy budget allocation: Total privacy budget ε_total=2.0, allocated by network layers (3 layers) and training rounds (10 rounds): ε=0.3 per layer, ε=0.06 per round. Adaptive Budget Allocation algorithm is used: if a layer detects gradient leakage risk (e.g., weight variance >0.1), temporarily increase ε=0.1 for that layer.

[0393] Relationship strength matrix generation and privacy verification

[0394] Multi-hop relationship propagation: Aggregate node features (e.g., knowledge point difficulty, teacher score) through 3-layer GAT to generate node embeddings (dimension=128). Relationship strength matrix R (size=N×N, N=node count) is computed from embedding vector cosine similarity and truncated to [0,1] range. For example, "derivative" and "integral" have strength R[i,j]=0.88.

[0395] Privacy attack simulation test: Evaluate privacy protection effect using Membership Inference Attack. Attackers attempt to infer whether a specific edge exists (e.g., "Student A→Enrolled→Course B") by observing R matrix. Test results show that attack success rate decreases from 32% for non-privacy protection to 8% (β=0.1).

[0396] Matrix sparsification compression: To reduce storage overhead, perform Top-K sparsification (K=10) on R matrix, retaining only the top 10 edges per node. For example, node "Calculus" retains edges "Pre-requisite Knowledge→Derivative (0.92)" and "Related Course→Physics (0.85)".

[0397] S2043, based on the relationship strength matrix, uses a temporal graph convolution network to capture the evolution of teaching patterns, identifies potential teaching association clusters through a dynamic community discovery algorithm, and generates a set of subgraph evolution paths;

[0398] Temporal graph convolution network (T-GCN) modeling

[0399] Time slicing: Divide teaching data into time windows (window size=6 months, sliding step=1 month) by semester, with each window corresponding to a relationship strength matrix R_t (t=1,2,…T). For example, R_1 for the 2023 spring semester includes the new knowledge point "Deep Learning Basics".

[0400] T-GCN architecture: The network includes temporal modules and graph modules:

[0401] Temporal Module: Capture temporal dependencies using Gated Recurrent Unit (GRU, hidden layer=64). Input is a Flatten vector of the historical R_t matrix (dimension=N x N), output is the temporal feature h_t (dimension=64).

[0402] Graph Module: Aggregate neighbor information using Graph Convolutional Network (GCN, layers=2). Each layer formula is: H^{(l+1)}= ReLU(D^{-1 / 2} R_t D^{-1 / 2} H^{(l)} W^{(l)}). Where D is the degree matrix, W is the trainable parameter (size=64 x 64). The final output node embedding Z_t (dimension=64).

[0403] Joint Training Strategy: The temporal module and graph module are fused through Cross-Attention. The loss function includes relationship prediction loss (cross-entropy) and temporal smoothing loss (L2 distance of adjacent Z_t). Learning rate=0.001, batch size=32.

[0404] Dynamic Community Discovery and Evolution Path Generation

[0405] Incremental Community Detection: Use dynamic Louvain algorithm (based on modularity optimization) to detect community structure at each time window t. Modularity threshold Q>0.3 is considered as an effective community. For example, in the spring of 2023, community 1 (math course cluster) and community 2 (programming course cluster) are detected.

[0406] Evolution Path Modeling: Define community evolution events:

[0407] Split: Community A splits into B and C at t+1 (Jaccard similarity <0.4).

[0408] Merge: Communities B and C merge into D at t+1 (overlapping nodes >60%).

[0409] Drift: The core node change rate of community A at t+1 is >30%.

[0410] For example, community 1 splits into "Calculus" and "Linear Algebra" sub-communities in the summer of 2023.

[0411] Evolution Path Storage: Use a graph database to store community evolution chains (such as Community_1 → Split → Community_2&Community_3), edge attributes include timestamp, evolution type, confidence (e.g., split confidence=0.88).

[0412] ‌S2044, Causal inference analysis on subgraph evolution paths, combined with Bayesian network to calculate the explainability score of teaching mode, generate intermediate teaching mode atlas with confidence label;

[0413] Causal inference and Bayesian network modeling‌

[0414] ‌Causal graph construction: Based on events in subgraph evolution paths (such as "community merger" "knowledge point addition"), define causal variables (such as X=teacher turnover rate, Y=course update frequency), learn causal structure through PC algorithm (Peter-Clark). For example, it is found that "teacher turnover rate↑ → course update frequency↑" (causal strength=0.75).

[0415] ‌Bayesian network parameter learning: Use maximum likelihood estimation (MLE) to calculate conditional probability table (CPT). For example:

[0416] P(course update frequency=High | teacher turnover rate=High) = 0.82;

[0417] P(knowledge point correlation strength↑ | course update frequency=High) = 0.67.

[0418] Network structure contains 3 layers of hidden variables (such as "teaching resource input" "student feedback trend"), total number of nodes=15.

[0419] Explainability score calculation: Define scoring rules: Score = 0.6 * causal strength + 0.3 * confidence + 0.1 * evolutionary stability. For example, the causal strength of the path "community merger→ course correlation enhancement" is 0.8, the confidence is 0.91, and the stability is 0.75, the final Score=0.6 0.8 + 0.3 0.91 + 0.1*0.75 = 0.828.

[0420] ‌Intermediate atlas generation and verification‌

[0421] ‌Knowledge fusion with confidence: Attach event causal probability output by Bayesian network (such as "teacher training→ knowledge point update, P=0.78") as attributes to atlas edges. At the same time, add stability label to each community evolution path (such as "Stable=0.9").

[0422] ‌Artificial expert verification: Invite education experts (N=10) to label 100 high Score paths (1-5 points), and require an average score of ≥4.0. For example, the path "increase experimental courses→ improve student performance" scored 4.3.

[0423] ‌Logical consistency check: Verify the logical consistency of the graph via an OWL reasoner (e.g., HermiT). For example, if there exist "Course A → Prerequisite → Course B" and "Course B → Prerequisite → Course A", trigger a contradiction alert and manually correct it.

[0424] ‌S2045, compress the graph scale of the intermediate teaching pattern graph through graph neural network distillation technology, and finally output an interpretable teaching pattern graph supporting real-time inference.

[0425] ‌Graph Neural Network Distillation (GNN Distillation)

[0426] ‌Teacher-student model architecture:

[0427] ‌Teacher model: 3-layer GAT (hidden layer = 256, head number = 8), trained on the complete graph, accuracy = 92%.

[0428] ‌Student model: 2-layer GCN (hidden layer = 128), aiming to approximate the teacher performance on the compressed graph (50% reduction in node number).

[0429] ‌Distillation loss function: Loss = 0.7 * KL divergence (teacher output || student output) + 0.3 * student cross-entropy loss. The KL divergence measures the difference between the output distributions, and the temperature parameter T = 2.0 is used to smooth the probability distribution.

[0430] ‌Progressive pruning strategy: Iterative pruning is used, removing 10% of redundant edges (weight < 0.1) in each round. For example, the initial number of edges = 100,000, after 5 rounds of pruning, the remaining number of edges = 59,000, with an accuracy loss of < 3%.

[0431] ‌Real-time inference optimization and deployment

[0432] ‌Lightweight graph storage: Convert the compressed graph to CSR format (Compressed Sparse Row), reducing storage occupancy by 60%. For example, the original Neo4j storage = 120GB, after compression = 48GB.

[0433] ‌Real-time inference engine: Deploy the student model based on TensorRT, supporting 1000 queries per second (such as "Recommend teaching patterns suitable for teacher A"). The engine response time < 50ms, GPU utilization > 90%.

[0434] ‌Interpretable interface: Generate visual reports for each inference result, including key causal paths (such as "community consolidation → performance improvement") and confidence heat maps (rendered with D3.js). Users can click on nodes to view the details of Bayesian network derivation.

[0435] ‌Application scenario example‌

[0436] ‌Scenario description: An education alliance needs to analyze the evolution of teaching modes across campuses and real-time recommend course optimization strategies.

[0437] ‌Construction of federated knowledge graph‌:

[0438] Aggregate the "Calculus Course Group" of School A and the "AI Experiment Course" of School B to generate interdisciplinary correlation edges (strength = 0.79).

[0439] ‌Privacy-preserving relationship mining‌:

[0440] Detect the causal path of "teacher mobility → course update" (Score = 0.85), and the attack success rate decreases to 7% after injecting noise.

[0441] ‌Real-time reasoning and recommendation‌:

[0442] Input query "How to improve the effectiveness of programming courses", reasoning engine returns "Increase project practice (confidence = 0.91)" and "Related mathematical basis (confidence = 0.87)".

[0443] Based on the enhanced entity relationships in the data set (such as "Knowledge Point A → Pre-requisite Knowledge Point B" and "Teacher X commonly used teaching tool Y"), aggregate multi-institutional data (such as course design in different schools) through a federated learning framework, protect privacy using homomorphic encryption, and construct an initial federated knowledge graph. Use a graph attention mechanism protected by differential privacy (inject Laplace noise) to calculate the correlation strength between nodes (such as knowledge point co-occurrence probability), identify teaching mode evolution paths (such as "group discussion → performance improvement") through dynamic community discovery algorithms, and calculate the explainability score of causal reasoning (such as confidence ≥ 90%) using Bayesian networks. Finally, generate a teaching mode graph that supports real-time updates, reveal the potential laws between teaching strategies and learning effectiveness, achieve safe sharing and knowledge fusion of cross-institutional data, mine explainable correlations of teaching modes through privacy protection techniques, provide basis for education decision-making, and avoid data silos and privacy leakage risks.

[0444] S205, according to the teaching mode graph and real-time teaching feedback data, generate personalized teaching recommendation strategies through a course-learner dual-channel adaptive recommendation model, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by integrating time series reinforcement learning and meta-path reasoning algorithms. Specifically, it can include:

[0445] S2051, according to the knowledge point dependency relationship in the teaching mode map and the real-time feedback data of the learner, extract the node weight in the course knowledge topology structure and the time sequence feature of the learner behavior trajectory, build the topology feature map of the course channel and the behavior evolution sequence of the learner channel;

[0446] ‌Course knowledge topology structure modeling‌

[0447] ‌Node weight calculation‌: Extract knowledge point entities (such as "derivative" and "integral") and their dependency relationships (such as the edge weight of "derivative → integral" is 0.85) from the teaching mode map. Calculate the global importance score of each knowledge point by PageRank algorithm (damping coefficient = 0.85, iteration times = 50) as node weight. For example, the PageRank score of "derivative" is 0.92, and that of "integral" is 0.88.

[0448] ‌Topology feature map construction‌: Combine the node weight of the knowledge point with the edge weight (relationship strength) to construct a directed weighted graph (Directed Weighted Graph). Use NetworkX library (Python) to store the graph structure, node attributes include knowledge point ID, weight, and course label (such as "Advanced Mathematics"), edge attributes include dependency type (predecessor, parallel) and strength value (0-1). For example, the strength of the edge "derivative → integral" is 0.85, and the type is predecessor dependency.

[0449] ‌Time sequence feature extraction‌: Extract behavior sequences from learner real-time feedback data (such as learning duration, test accuracy, interaction frequency) and divide them into time windows (window size = 7 days, step = 1 day). Extract the time sequence feature vector (dimension = 64) of each window by LSTM network (hidden layer = 64, time step = 30), for example, the learning behavior feature of a learner in the tth week of "Calculus" is [0.72, 0.65,..., 0.88].

[0450] ‌Learner behavior evolution sequence construction‌

[0451] ‌Behavior trajectory encoding‌: Encode the video interaction behaviors of learners such as clicking, pausing, and reviewing, as well as text behaviors such as test submission and discussion area speech into multi-modal event sequences. For example, the event "video pause" is encoded as type code 101, timestamp 2023-08-15 14:30:00, and associated knowledge point derivative.

[0452] ‌Knowledge transition pattern recognition: Based on the knowledge point switching records in the event sequence (e.g., "derivative → integral → differential equation"), use the Hidden Markov Model (HMM, state number = 5) to model the learner's knowledge state transition probability. For example, the transition probability from "derivative" to "integral" is 0.78, and to "differential equation" is 0.45.

[0453] ‌Evolution sequence storage: Use MongoDB to store the behavior evolution sequence of learners, the document structure contains learner ID, timestamp sequence, knowledge point path, behavior feature vector. For example, the document learner_001 contains the path ["derivative", "integral", "differential equation"] and the feature vector [[0.72,...], [0.68,...],...].

[0454] ‌S2052, in the course channel, use the time series reinforcement learning algorithm to dynamically plan the topology feature map, capture the historical teaching effect data through the sliding window mechanism, and generate the candidate recommendation sequence based on the knowledge point association strength;

[0455] ‌Time series reinforcement learning model design

[0456] ‌State space definition: Map the node weight and edge strength of the course topology feature map to the state space. State S_t includes the current knowledge point (e.g., "derivative"), the set of adjacent knowledge points (e.g., ["limit", "integral"]), and the historical teaching effect indicators (e.g., the average accuracy of learners under this knowledge point = 75%).

[0457] ‌Action space definition: Action A_t is the next knowledge point that can be jumped from the current knowledge point (e.g., choose "integral" or "limit" from "derivative"), and the action candidate set is filtered according to the edge strength (only keep the edges with strength > 0.6).

[0458] ‌Reward function design: Reward R is dynamically calculated by teaching effect data, including short-term reward (e.g., the test accuracy of learners improves Δ = 5% after choosing the action) and long-term reward (e.g., path completeness score = 0.8). For example, after choosing "integral", the reward R = 0.3*Δ + 0.7*path score = 0.3*5 + 0.7*0.8 = 2.31.

[0459] ‌Dynamic path planning and candidate sequence generation

[0460] ‌Sliding window mechanism: Load historical teaching data (e.g., 1000 learners' path records) by time window (window size = 30 days, step = 7 days), and train reinforcement learning model using DQN (Deep Q-Network) algorithm. Input state S_t (dimension = 64), output Q value (expected return of each action), network structure is fully connected layer (256→128→64), learning rate lr = 0.001, discount factor γ = 0.9.

[0461] ‌Path generation strategy: Use ε-greedy strategy (ε = 0.1) to balance exploration and exploitation. For example, the model has a 90% probability of choosing the action with the highest Q value ("integral"), and a 10% probability of randomly choosing other actions ("limit"). Generate N candidate paths (N = 50) through iterative planning, each path containing a sequence of knowledge points and their cumulative rewards (e.g., path derivative→integral→differential equation, total reward = 18.5).

[0462] ‌Candidate sequence optimization: Sort candidate paths according to path association strength (average edge strength) and teaching effectiveness (average historical reward), and keep Top-10 paths. For example, the path derivative→integral has an association strength of 0.85 and a teaching effectiveness of 8.7, ranking first.

[0463] ‌S2053, in the learner channel, based on the meta-path reasoning algorithm of the knowledge graph, analyzes the knowledge transition patterns in the learner behavior evolution sequence, and generates a personalized path candidate set that integrates the ability growth curve;

[0464] ‌Meta-path reasoning algorithm implementation

[0465] ‌Meta-path definition: Predefine teaching association patterns in the knowledge graph, for example:

[0466] Knowledge point→predecessor→knowledge point (dependency relationship path)

[0467] Learner→learning→knowledge point→belongs to→course (cross-entity path)

[0468] Each meta-path has a length limit of 3 hops (e.g., learner→test→knowledge point→dependency→knowledge point).

[0469] ‌Path instance mining: Extract instances that match the meta-path from the learner behavior sequence. For example, learner A's path derivative→learning→integral→test→differential equation matches the meta-path learner→learning→knowledge point→test→knowledge point.

[0470] ‌Path weight calculation: Use the PathSim algorithm to calculate the semantic similarity of the meta-path. For example, the weight of the path derivative→integral→differential equation is 0.82, and the weight of the path derivative→limit→integral is 0.75.

[0471] ‌Personalized Path Generation

[0472] ‌Ability Growth Curve Modeling: Fit a growth curve based on the learner's historical test scores (e.g., weekly quiz scores). Use Cubic Spline interpolation to generate a continuous curve and extract key indicators: slope (learning speed), curvature (acceleration / deceleration phase), peak (maximum ability value). For example, a learner's slope = 0.15 (slow improvement), curvature = -0.02 (entering bottleneck phase).

[0473] ‌Path Personalization Adaptation: Combine meta-path instances with ability curves and match through a rule engine:

[0474] ‌High-Growth Phase: Recommend long paths (e.g., 3-hop knowledge points) and challenging content (e.g., "differential equations → partial differential equations").

[0475] ‌Bottleneck Phase: Recommend review paths (e.g., "derivatives → integrals") and auxiliary resources (e.g., "problem-solving video tutorials").

[0476] For example, generate a path for a learner in the bottleneck phase: integral → problem reinforcement → differential equations, associated resources include 5 tutorial videos and 3 sets of mock exams.

[0477] ‌Candidate Set Storage and Indexing: Use Elasticsearch to establish a personalized path index, fields include learner ID, meta-path type, adaptation score (e.g., 0.92), and recommendation priority (1-5 levels). Support real-time queries with response time <100ms.

[0478] ‌S2054, input the candidate recommendation sequence of the course channel and the personalized path candidate set of the learner channel into the multi-strategy fusion module, calculate the cross-channel weight distribution matrix through the dynamic attention mechanism, and generate a personalized teaching comprehensive recommendation strategy with priority ranking.

[0479] ‌Multi-strategy Fusion Module Design

[0480] ‌Feature Alignment and Embedding: Map the candidate paths (knowledge point sequences) of the course channel and the personalized paths (meta-path instances) of the learner channel to the same vector space. Use the BERT model (bert-base, 768 dimensions) to encode path description text (e.g., "derivative → integral → differential equation") and generate semantic vectors (e.g., [0.34, -0.56,..., 0.78]).

[0481] Dynamic attention mechanism: Design a Cross-Channel Attention layer to calculate the relevance weight between course paths and learner paths. Attention heads = 4, key / query vector dimension = 64. For example, the attention weight between course path "derivative → integral" and learner path "derivative → exercises → integral" = 0.91.

[0482] Weight distribution matrix generation: Normalize the attention weights by Softmax to generate an M x N matrix (M = number of course candidates, N = number of learner candidates). For example, the weight between course path 1 and learner path 3 = 0.85, and the weight between path 5 = 0.72.

[0483] Integrated recommendation strategy generation

[0484] Multi-objective optimization sorting: Define the sorting objective function: Score = 0.6 * course path intensity + 0.3 * learner adaptation degree + 0.1 * attention weight.

[0485] For example, the course intensity of a certain path = 0.88, the learner adaptation degree = 0.95, and the attention weight = 0.91, then Score = 0.6 * 0.88 + 0.3 * 0.95 + 0.1 * 0.91 = 0.90.

[0486] Real-time priority scheduling: Sort the Top-100 candidate paths according to Score, and dynamically adjust according to the teaching scenario:

[0487] New course import stage: Prioritize high-intensity paths in the course channel (increase the course weight in Score to 0.7).

[0488] Review and consolidation stage: Prioritize high-adaptation paths in the learner channel (increase the learner weight to 0.5).

[0489] Finally output the recommended list with priority, for example: 1. Derivative → Integral (Score = 0.95); 2. Integral → Differential Equation (Score = 0.89).

[0490] Strategy visualization and feedback: Provide a Dashboard for the teacher side to display the associated knowledge points of the recommended path, historical effect comparison (such as the average score of the class using this path increased by 12%), and learner adaptation degree heat map. Support manual adjustment of weights (such as manually setting the priority of a certain path to the top).

[0491] Application scenario example

[0492] Scenario description: An online education platform needs to recommend a learning path for "Advanced Mathematics" to learners.

[0493] Course channel generation candidate:

[0494] Derivative -> integral -> differential equation path is planned through reinforcement learning (course intensity = 0.92, historical promotion effect = 15%).

[0495] Learner channel adaptation:

[0496] The learner is detected to be in the bottleneck period (growth curve slope = 0.05), and a personalized path derivative -> exercise class -> integral is generated (adaptation degree = 0.89).

[0497] Multi-strategy fusion output:

[0498] The weight matrix is calculated by dynamic attention, and after comprehensive scoring, derivative -> exercise class -> integral -> differential equation is recommended (Score = 0.93), and 3 intensive lecture videos and 2 stage tests are inserted.

[0499] Based on the knowledge point topology in the teaching mode graph (such as dependency relationship, difficulty gradient) and the real-time feedback of the learner (such as the correct rate of answering questions, the interaction time), the teaching path is dynamically planned (such as priority recommendation of basic knowledge points) in the course channel using time series reinforcement learning (such as DQN algorithm), and the individual behavior evolution trend is analyzed through meta-path reasoning (such as "learning path A -> ability improvement B") in the learner channel. The dynamic attention mechanism is used to fuse the results of the two channels (such as course weight 60%, learner weight 40%), and a personalized recommendation strategy (such as "recommend calculus primer video for student A") is generated, and the recommendation priority is adjusted in real time to adapt to the learning progress changes, so as to realize precise and dynamic teaching recommendation, balance the requirements of the course system and the individualized needs of the learners through the double-channel model, improve the teaching effect and learning efficiency, and avoid the rigidity and overgeneralization of traditional recommendation systems.

[0500] As can be seen, according to the heterogeneous data stream generated in real time by the multi-source teaching platform, a standardized time series data stream is obtained; based on the standardized time series data stream, the teaching data is classified and encrypted through a dynamic hierarchical storage strategy based on attribute-based encryption, and a secure hierarchical storage topology structure is obtained; according to the user query request and the teaching scene label, the target data set is extracted from the secure hierarchical storage topology structure, and an enhanced multi-modal teaching data set is obtained; based on the enhanced multi-modal teaching data set, an interpretable teaching mode graph is generated through a dynamic subgraph evolution algorithm; according to the teaching mode graph and the real-time teaching feedback data, a personalized teaching recommendation strategy is generated through a course-learner double-channel adaptive recommendation model, so as to improve the utilization efficiency of teaching resources, and realize personalized and intelligent teaching recommendation and decision-making.

[0501] Another embodiment of the present application provides an artificial intelligence-based teaching data management system, which is described with reference to Figure 3 The system can comprise:

[0502] The processing module 301 is configured to dynamically parse and normalize heterogeneous data streams generated in real time by the multi-source teaching platform, identify data format features and construct a dynamic interface adapter by combining a bidirectional long short-term memory network, and obtain standardized time-series data streams.

[0503] The encryption module 302 is configured to classify and encrypt teaching data by a dynamic hierarchical storage strategy based on attribute-based encryption based on the standardized time-series data streams, compress unstructured data and generate a unique data fingerprint by using a knowledge distillation algorithm, and obtain a secure hierarchical storage topology.

[0504] The extraction module 303 is configured to extract a target data set from the secure hierarchical storage topology according to a user query request and a teaching scenario label, jointly embed text, video, and behavior log data by a multi-modal feature fusion network, and complete missing features by using an adversarial generative network, and obtain an enhanced multi-modal teaching data set.

[0505] The mining module 304 is configured to construct a cross-institutional federated knowledge graph based on the enhanced multi-modal teaching data set, mine potential teaching association relationships by using a graph attention mechanism protected by differential privacy, and generate an interpretable teaching mode graph by a dynamic subgraph evolution algorithm.

[0506] The generation module 305 is configured to generate a personalized teaching recommendation strategy by a course-learner dual-channel adaptive recommendation model according to the teaching mode graph and real-time teaching feedback data, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts recommendation weights by fusing time-series reinforcement learning and meta-path reasoning algorithms.

[0507] As can be seen, standardized time-series data streams are obtained from heterogeneous data streams generated in real time by a multi-source teaching platform. Teaching data is classified and encrypted by a dynamic hierarchical storage strategy based on attribute-based encryption based on the standardized time-series data streams, and a secure hierarchical storage topology is obtained. A target data set is extracted from the secure hierarchical storage topology according to a user query request and a teaching scenario label, and an enhanced multi-modal teaching data set is obtained. An interpretable teaching mode graph is generated by a dynamic subgraph evolution algorithm based on the enhanced multi-modal teaching data set. A personalized teaching recommendation strategy is generated by a course-learner dual-channel adaptive recommendation model according to the teaching mode graph and real-time teaching feedback data, thereby improving the utilization efficiency of teaching resources and enabling personalized and intelligent teaching recommendation and decision-making.

[0508] The embodiment of the present application also provides a storage medium, wherein the storage medium stores a computer program, and the computer program is arranged to execute the steps in any of the method embodiments when running.

[0509] Specifically, in the present embodiment, the storage medium can be arranged to store a computer program for executing the following steps:

[0510] S201, according to the heterogeneous data stream generated by the multi-source teaching platform in real time, performing dynamic analysis and normalization processing on the heterogeneous data stream, identifying data format features and constructing a dynamic interface adapter by combining a bidirectional long short-term memory network, to obtain a standardized time series data stream;

[0511] S202, based on the standardized time series data stream, classifying and encrypting teaching data by a dynamic hierarchical storage strategy based on attribute-based encryption, compressing unstructured data by a knowledge distillation algorithm and generating a unique data fingerprint, to obtain a secure hierarchical storage topology;

[0512] S203, according to a user query request and a teaching scene label, extracting a target data set from the secure hierarchical storage topology, jointly embedding text, video and behavior log data by a multi-modal feature fusion network, and utilizing an adversarial generative network to complete missing features, to obtain an enhanced multi-modal teaching data set;

[0513] S204, based on the enhanced multi-modal teaching data set, constructing a cross-institutional federated knowledge graph, mining potential teaching association relationships by a differential privacy protected graph attention mechanism, and generating an interpretable teaching mode graph by a dynamic subgraph evolution algorithm;

[0514] S205, according to the teaching mode graph and real-time teaching feedback data, generating a personalized teaching recommendation strategy by a course-learner dual-channel adaptive recommendation model, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by fusing time series reinforcement learning and meta-path reasoning algorithm.

[0515] It can be seen that, according to the heterogeneous data stream generated in real time by the multi-source teaching platform, the standardized time sequence data stream is obtained; based on the standardized time sequence data stream, the teaching data is classified and encrypted through the dynamic hierarchical storage strategy based on attribute-based encryption, and the safe hierarchical storage topology is obtained; according to the user query request and the teaching scene label, the target data set is extracted from the safe hierarchical storage topology, and the enhanced multi-modal teaching data set is obtained; based on the enhanced multi-modal teaching data set, the interpretable teaching mode graph is generated through the dynamic subgraph evolution algorithm; according to the teaching mode graph and the real-time teaching feedback data, the personalized teaching recommendation strategy is generated through the course-learner dual-channel adaptive recommendation model, so that the utilization efficiency of teaching resources can be improved, and personalized and intelligent teaching recommendation and decision can be realized.

[0516] The embodiment of the application further provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to perform the steps in any of the method embodiments.

[0517] Specifically, the electronic device can further comprise a transmission device connected to the processor and an input-output device connected to the processor.

[0518] Specifically, in the embodiment, the processor can be configured to execute the following steps through the computer program:

[0519] S201, according to the heterogeneous data stream generated in real time by the multi-source teaching platform, the heterogeneous data stream is dynamically parsed and normalized, the data format features are identified and a dynamic interface adapter is constructed in combination with a bidirectional long short-term memory network, and a standardized time sequence data stream is obtained;

[0520] S202, based on the standardized time sequence data stream, the teaching data is classified and encrypted through the dynamic hierarchical storage strategy based on attribute-based encryption, the unstructured data is compressed and a unique data fingerprint is generated using a knowledge distillation algorithm, and a safe hierarchical storage topology is obtained;

[0521] S203, according to the user query request and the teaching scene label, the target data set is extracted from the safe hierarchical storage topology, the text, video and behavior log data are jointly embedded and represented through a multi-modal feature fusion network, and the missing features are completed using an adversarial generative network, and an enhanced multi-modal teaching data set is obtained;

[0522] S204, based on the enhanced multi-modal teaching data set, a cross-institutional federated knowledge graph is constructed, a graph attention mechanism with differential privacy protection is used to mine potential teaching association relationships, and an interpretable teaching mode graph is generated through a dynamic subgraph evolution algorithm;

[0523] S205, generating a personalized teaching recommendation strategy through a course-learner dual-channel adaptive recommendation model according to the teaching mode graph and real-time teaching feedback data, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by fusing the time series reinforcement learning and the meta-path reasoning algorithm.

[0524] It can be seen that the standardized time series data stream is obtained according to the heterogeneous data stream generated by the multi-source teaching platform in real time; the secure hierarchical storage topology is obtained by classifying and encrypting the teaching data through the dynamic hierarchical storage strategy based on attribute-based encryption based on the standardized time series data stream; the enhanced multi-modal teaching data set is obtained by extracting the target data set from the secure hierarchical storage topology according to the user query request and the teaching scene label; the interpretable teaching mode graph is generated through the dynamic subgraph evolution algorithm based on the enhanced multi-modal teaching data set; and the personalized teaching recommendation strategy is generated through the course-learner dual-channel adaptive recommendation model according to the teaching mode graph and real-time teaching feedback data, so as to improve the utilization efficiency of teaching resources and realize personalized and intelligent teaching recommendation and decision-making.

[0525] The above describes the structure, features and effects of the present application in detail according to the embodiments shown in the drawings. The above description is only the preferred embodiments of the present application, but the present application is not limited to the embodiments shown in the drawings. Any changes or modifications made in accordance with the concept of the present application, or equivalent embodiments with equivalent changes, shall be within the scope of the present application.< / sequencing> < / responsedeclaration> < / sequencing> < / adlcp:location> < / adlcp:score> < / adlcp:location> < / sequencing>

Claims

1. A method for managing teaching data based on artificial intelligence, characterized by, The method comprises: According to the real-time heterogeneous data stream generated by the multi-source teaching platform, the heterogeneous data stream is dynamically parsed and normalized, the data format features are identified and the dynamic interface adapter is constructed by combining the bidirectional long short-term memory network, and the standardized time sequence data stream is obtained; Based on the standardized time sequence data stream, the teaching data is classified and encrypted by a dynamic hierarchical storage strategy based on attribute-based encryption, the unstructured data is compressed by a knowledge distillation algorithm, and a unique data fingerprint is generated, to obtain a secure hierarchical storage topology; According to the user query request and the teaching scene label, the target data set is extracted from the secure hierarchical storage topology, the text, video and behavior log data are jointly embedded and represented by a multi-modal feature fusion network, and the missing features are completed by an adversarial generation network, to obtain an enhanced multi-modal teaching data set; Based on the enhanced multi-modal teaching data set, a cross-institutional federated knowledge graph is constructed, a graph attention mechanism with differential privacy protection is used to mine potential teaching association relationships, and an interpretable teaching mode graph is generated by a dynamic subgraph evolution algorithm; According to the teaching mode graph and the real-time teaching feedback data, a personalized teaching recommendation strategy is generated by a course-learner dual-channel adaptive recommendation model, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by fusing time sequence reinforcement learning and meta-path reasoning algorithm.

2. The method of claim 1, wherein, The method comprises: Receiving original data streams from teaching platforms, learning management systems and online course platforms, extracting time sequence features of the original data streams by a bidirectional long short-term memory network, identifying implicit patterns of data formats, and obtaining multi-source data format feature vectors; Based on the multi-source data format feature vectors, a dynamic interface template library is generated by using a density peak clustering algorithm, and the optimal interface template is dynamically matched according to the real-time data stream characteristics, to obtain interface adapter initialization parameters; According to the interface adapter initialization parameters, a lightweight data conversion rule engine is constructed to convert heterogeneous data containing SCORM, xAPI and IMS-QTI into a unified JSON-LD format, to obtain intermediate standardized data blocks; The intermediate standardized data blocks are subjected to time sequence alignment processing, the sliding window mechanism is used to eliminate the cross-platform timestamp deviation, and the continuous and conflict-free standardized time sequence data stream is output through an event-driven pipeline.

3. The method of claim 2, wherein, The method comprises: According to the metadata attributes of the standardized time sequence data stream, the data sensitive level label is extracted by the attribute-based encryption algorithm, the CP-ABE strategy is used to define the access control tree, and the dynamic hierarchical encryption strategy is generated. The teaching video data and the text log data are classified and encrypted based on a dynamic hierarchical encryption strategy to generate hierarchical ciphertext data blocks; The knowledge distillation algorithm is used for the unstructured video data, key teaching features are extracted through a teacher model, a lightweight student model is trained to generate a compressed feature vector, and a lightweight video feature package retaining semantics is output; The hierarchical ciphertext data blocks and the lightweight video feature package are integrated through a hash chain, a unique data fingerprint is generated based on the content features of the hierarchical ciphertext data blocks, and a secure hierarchical storage topology with an access control label is constructed.

4. The method of claim 3, wherein, The target data set is extracted from the secure hierarchical storage topology according to the user query request and the teaching scene label, the text, video and behavior log data are jointly embedded and represented through a multi-modal feature fusion network, and the missing features are completed by using an adversarial generation network to obtain an enhanced multi-modal teaching data set, including: According to the keyword vector in the user query request and the teaching scene label, the target data is searched in the secure hierarchical storage topology through a semantic similarity algorithm to generate a candidate data set index table; Based on the candidate data set index table, a multi-head attention mechanism is used to align the text TF-IDF features, video optical flow features and behavior log time sequence features to generate a multi-modal joint embedding matrix; The multi-modal joint embedding matrix is input into the adversarial generation network, the discriminator identifies the missing feature positions, the generator reconstructs the missing video key frames and text semantic segments, and the completed feature tensor is output; The completed feature tensor is adversarially trained and optimized, a gradient penalty strategy is used to stabilize the training process, and a robust enhanced intermediate multi-modal data set is generated; The modal specificity and commonality information in the intermediate multi-modal data set are separated through a feature decoupling algorithm, and finally an enhanced multi-modal teaching data set with cross-modal correlation labels is output.

5. The method of claim 4, wherein, Based on the enhanced multi-modal teaching data set, a cross-institutional federated knowledge graph is constructed, a differential privacy protected graph attention mechanism is used to mine potential teaching correlation relationships, and an interpretable teaching mode graph is generated through a dynamic subgraph evolution algorithm, including: According to the entity relationship triplets in the enhanced multi-modal teaching data set, cross-institutional data is aggregated through a federated learning framework, a homomorphic encryption technology is used to protect the privacy of the institutions, and an initial federated knowledge graph is generated; Based on the initial federated knowledge graph, a differential privacy protected graph attention mechanism is designed, the attention weights between nodes are disturbed by injecting Laplace noise, and a privacy-safe relationship strength matrix is generated; According to the relationship strength matrix, a time series graph convolution network is used to capture the evolution law of the teaching mode, a dynamic community discovery algorithm is used to identify potential teaching correlation clusters, and a subgraph evolution path set is generated; The subgraph evolution path is analyzed through causal reasoning, the interpretability score of the teaching mode is calculated based on the Bayesian network, and an intermediate teaching mode graph with a confidence label is generated; The graph size of the intermediate teaching mode graph is compressed through graph neural network distillation technology, and finally an interpretable teaching mode graph supporting real-time reasoning is output.

6. The method of claim 5, wherein, The personalized teaching recommendation strategy is generated through a course-learner dual-channel adaptive recommendation model based on the teaching mode graph and real-time teaching feedback data, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts the recommendation weight by fusing the time series reinforcement learning and the meta-path reasoning algorithm, and includes the following steps: According to the knowledge point dependency relationship in the teaching mode graph and the real-time feedback data of the learner, the node weight in the course knowledge topology structure and the time sequence feature of the learner behavior trajectory are extracted, and the topology feature graph of the course channel and the behavior evolution sequence of the learner channel are constructed; In the course channel, a time series reinforcement learning algorithm is used for dynamic path planning of the topology feature graph, a sliding window mechanism is used to capture historical teaching effect data, and a candidate recommendation sequence based on the knowledge point correlation strength is generated; In the learner channel, a meta-path reasoning algorithm based on a knowledge graph is used to analyze the knowledge transition mode in the learner behavior evolution sequence, and a personalized path candidate set integrating the ability growth curve is generated; The candidate recommendation sequence of the course channel and the personalized path candidate set of the learner channel are input into a multi-strategy fusion module, a cross-channel weight distribution matrix is calculated through a dynamic attention mechanism, and a personalized teaching comprehensive recommendation strategy with priority ranking is generated.

7. The method of claim 3, wherein, The data sensitivity level label is extracted through the attribute-based encryption algorithm based on the metadata attribute of the standardized time series data stream, the access control tree is defined by using the CP-ABE strategy, and the dynamic hierarchical encryption strategy is generated, including the following steps: According to the metadata attribute of the standardized time series data stream, the privacy correlation degree of the data field is analyzed by using a multi-dimensional sensitive feature extraction network, and the student information, score record and classroom behavior log are mapped to different sensitive levels by using a hierarchical clustering algorithm and a self-encoder fusion architecture, and a three-dimensional feature vector of data sensitivity is output; Based on the three-dimensional feature vector of data sensitivity, a fine-grained access strategy is generated by using a dynamic decay decision tree, wherein the access permission of course video data is divided according to the resolution and content sensitivity, and the decryption condition is defined according to the score distribution interval of test data, and a timestamped initial access control rule set is output; According to the context information of the real-time teaching scene, the initial access control rule set is dynamically modified by using a strategy evolution engine, when a cross-campus data sharing request is detected, a regional constraint attribute is automatically added and a two-factor authentication condition is introduced, and a time and space aware enhanced access control tree is generated; The enhanced access control tree and the three-dimensional feature vector of data sensitivity are coupled and optimized, the strategy node weight is calculated by using an asymmetric bilinear pairing algorithm, the strategy public key matrix is generated by using an elliptic curve encryption technology, and finally a CP-ABE hierarchical encryption strategy containing a dynamic decay factor and a time and space constraint is output.

8. An artificial intelligence-based teaching data management system, characterized by, The system comprises: A processing module is configured to dynamically analyze and normalize heterogeneous data streams generated in real time by a multi-source teaching platform, identify data format features by using a bidirectional long short-term memory network, and construct a dynamic interface adapter to obtain standardized time series data streams. The encryption module is configured to perform classified encryption on the teaching data based on the standardized time sequence data stream through a dynamic hierarchical storage strategy based on attribute-based encryption, compress unstructured data by using a knowledge distillation algorithm, and generate a unique data fingerprint to obtain a secure hierarchical storage topology; The extraction module is configured to extract a target data set from the secure hierarchical storage topology according to a user query request and a teaching scene label, perform joint embedding representation on text, video, and behavior log data by using a multi-modal feature fusion network, and complete missing features by using an adversarial generative network to obtain an enhanced multi-modal teaching data set; The mining module is configured to construct a cross-institutional federated knowledge graph based on the enhanced multi-modal teaching data set, mine potential teaching association relationships by using a graph attention mechanism protected by differential privacy, and generate an interpretable teaching mode graph by using a dynamic subgraph evolution algorithm. The generation module is configured to generate a personalized teaching recommendation strategy by using a course-learner dual-channel adaptive recommendation model according to the teaching mode graph and real-time teaching feedback data, wherein the course-learner dual-channel adaptive recommendation model dynamically adjusts recommendation weights by fusing time sequence reinforcement learning and meta-path reasoning algorithms.

9. A storage medium, characterized by The storage medium stores a computer program, and the computer program is configured to execute the method in any one of claims 1-7 when running.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method in any one of claims 1-7 by running the computer program.

Images